lustre/ptlrpc/import.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  *
  39  * lustre/ptlrpc/import.c
  40  *
  41  * Author: Mike Shaver <shaver@clusterfs.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_RPC
  45 #ifndef __KERNEL__
  46 # include <liblustre.h>
  47 #endif
  48
  49 #include <obd_support.h>
  50 #include <lustre_ha.h>
  51 #include <lustre_net.h>
  52 #include <lustre_import.h>
  53 #include <lustre_export.h>
  54 #include <obd.h>
  55 #include <obd_cksum.h>
  56 #include <obd_class.h>
  57
  58 #include "ptlrpc_internal.h"
  59
  60 struct ptlrpc_connect_async_args {
  61          __u64 pcaa_peer_committed;
  62         int pcaa_initial_connect;
  63 };
  64
  65 /**
  66  * Updates import \a imp current state to provided \a state value
  67  * Helper function. Must be called under imp_lock.
  68  */
  69 static void __import_set_state(struct obd_import *imp,
  70                                enum lustre_imp_state state)
  71 {
  72         imp->imp_state = state;
  73         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  74         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  75                 cfs_time_current_sec();
  76         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  77                 IMP_STATE_HIST_LEN;
  78 }
  79
  80 /* A CLOSED import should remain so. */
  81 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  82 do {                                                                           \
  83         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  84                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  85                       imp, obd2cli_tgt(imp->imp_obd),                          \
  86                       ptlrpc_import_state_name(imp->imp_state),                \
  87                       ptlrpc_import_state_name(state));                        \
  88                __import_set_state(imp, state);                                 \
  89         }                                                                      \
  90 } while(0)
  91
  92 #define IMPORT_SET_STATE(imp, state)            \
  93 do {                                            \
  94         cfs_spin_lock(&imp->imp_lock);          \
  95         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  96         cfs_spin_unlock(&imp->imp_lock);        \
  97 } while(0)
  98
  99
 100 static int ptlrpc_connect_interpret(const struct lu_env *env,
 101                                     struct ptlrpc_request *request,
 102                                     void * data, int rc);
 103 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 104
 105 /* Only this function is allowed to change the import state when it is
 106  * CLOSED. I would rather refcount the import and free it after
 107  * disconnection like we do with exports. To do that, the client_obd
 108  * will need to save the peer info somewhere other than in the import,
 109  * though. */
 110 int ptlrpc_init_import(struct obd_import *imp)
 111 {
 112         cfs_spin_lock(&imp->imp_lock);
 113
 114         imp->imp_generation++;
 115         imp->imp_state =  LUSTRE_IMP_NEW;
 116
 117         cfs_spin_unlock(&imp->imp_lock);
 118
 119         return 0;
 120 }
 121 EXPORT_SYMBOL(ptlrpc_init_import);
 122
 123 #define UUID_STR "_UUID"
 124 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
 125                       int *uuid_len)
 126 {
 127         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 128                 ? uuid : uuid + strlen(prefix);
 129
 130         *uuid_len = strlen(*uuid_start);
 131
 132         if (*uuid_len < strlen(UUID_STR))
 133                 return;
 134
 135         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 136                     UUID_STR, strlen(UUID_STR)))
 137                 *uuid_len -= strlen(UUID_STR);
 138 }
 139
 140 /**
 141  * Returns true if import was FULL, false if import was already not
 142  * connected.
 143  * @imp - import to be disconnected
 144  * @conn_cnt - connection count (epoch) of the request that timed out
 145  *             and caused the disconnection.  In some cases, multiple
 146  *             inflight requests can fail to a single target (e.g. OST
 147  *             bulk requests) and if one has already caused a reconnection
 148  *             (increasing the import->conn_cnt) the older failure should
 149  *             not also cause a reconnection.  If zero it forces a reconnect.
 150  */
 151 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 152 {
 153         int rc = 0;
 154
 155         cfs_spin_lock(&imp->imp_lock);
 156
 157         if (imp->imp_state == LUSTRE_IMP_FULL &&
 158             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 159                 char *target_start;
 160                 int   target_len;
 161
 162                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 163                           &target_start, &target_len);
 164
 165                 if (imp->imp_replayable) {
 166                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 167                                "%s was lost; in progress operations using this "
 168                                "service will wait for recovery to complete.\n",
 169                                imp->imp_obd->obd_name, target_len, target_start,
 170                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 171                 } else {
 172                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 173                                "%.*s via nid %s was lost; in progress "
 174                                "operations using this service will fail.\n",
 175                                imp->imp_obd->obd_name,
 176                                target_len, target_start,
 177                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 178                 }
 179                 ptlrpc_deactivate_timeouts(imp);
 180                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 181                 cfs_spin_unlock(&imp->imp_lock);
 182
 183                 if (obd_dump_on_timeout)
 184                         libcfs_debug_dumplog();
 185
 186                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 187                 rc = 1;
 188         } else {
 189                 cfs_spin_unlock(&imp->imp_lock);
 190                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 191                        imp->imp_client->cli_name, imp,
 192                        (imp->imp_state == LUSTRE_IMP_FULL &&
 193                         imp->imp_conn_cnt > conn_cnt) ?
 194                        "reconnected" : "not connected", imp->imp_conn_cnt,
 195                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 196         }
 197
 198         return rc;
 199 }
 200
 201 /* Must be called with imp_lock held! */
 202 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 203 {
 204         ENTRY;
 205         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 206
 207         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 208         imp->imp_invalid = 1;
 209         imp->imp_generation++;
 210         cfs_spin_unlock(&imp->imp_lock);
 211
 212         ptlrpc_abort_inflight(imp);
 213         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 214
 215         EXIT;
 216 }
 217
 218 /*
 219  * This acts as a barrier; all existing requests are rejected, and
 220  * no new requests will be accepted until the import is valid again.
 221  */
 222 void ptlrpc_deactivate_import(struct obd_import *imp)
 223 {
 224         cfs_spin_lock(&imp->imp_lock);
 225         ptlrpc_deactivate_and_unlock_import(imp);
 226 }
 227
 228 static unsigned int
 229 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 230 {
 231         long dl;
 232
 233         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 234               (req->rq_phase == RQ_PHASE_BULK) ||
 235               (req->rq_phase == RQ_PHASE_NEW)))
 236                 return 0;
 237
 238         if (req->rq_timedout)
 239                 return 0;
 240
 241         if (req->rq_phase == RQ_PHASE_NEW)
 242                 dl = req->rq_sent;
 243         else
 244                 dl = req->rq_deadline;
 245
 246         if (dl <= now)
 247                 return 0;
 248
 249         return dl - now;
 250 }
 251
 252 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 253 {
 254         time_t now = cfs_time_current_sec();
 255         cfs_list_t *tmp, *n;
 256         struct ptlrpc_request *req;
 257         unsigned int timeout = 0;
 258
 259         cfs_spin_lock(&imp->imp_lock);
 260         cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 261                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
 262                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 263         }
 264         cfs_spin_unlock(&imp->imp_lock);
 265         return timeout;
 266 }
 267
 268 /**
 269  * This function will invalidate the import, if necessary, then block
 270  * for all the RPC completions, and finally notify the obd to
 271  * invalidate its state (ie cancel locks, clear pending requests,
 272  * etc).
 273  */
 274 void ptlrpc_invalidate_import(struct obd_import *imp)
 275 {
 276         cfs_list_t *tmp, *n;
 277         struct ptlrpc_request *req;
 278         struct l_wait_info lwi;
 279         unsigned int timeout;
 280         int rc;
 281
 282         cfs_atomic_inc(&imp->imp_inval_count);
 283
 284         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 285                 ptlrpc_deactivate_import(imp);
 286
 287         LASSERT(imp->imp_invalid);
 288
 289         /* Wait forever until inflight == 0. We really can't do it another
 290          * way because in some cases we need to wait for very long reply
 291          * unlink. We can't do anything before that because there is really
 292          * no guarantee that some rdma transfer is not in progress right now. */
 293         do {
 294                 /* Calculate max timeout for waiting on rpcs to error
 295                  * out. Use obd_timeout if calculated value is smaller
 296                  * than it. */
 297                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 298                         timeout = ptlrpc_inflight_timeout(imp);
 299                         timeout += timeout / 3;
 300
 301                         if (timeout == 0)
 302                                 timeout = obd_timeout;
 303                 } else {
 304                         /* decrease the interval to increase race condition */
 305                         timeout = 1;
 306                 }
 307
 308                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 309                        timeout);
 310
 311                 /* Wait for all requests to error out and call completion
 312                  * callbacks. Cap it at obd_timeout -- these should all
 313                  * have been locally cancelled by ptlrpc_abort_inflight. */
 314                 lwi = LWI_TIMEOUT_INTERVAL(
 315                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 316                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 317                         NULL, NULL);
 318                 rc = l_wait_event(imp->imp_recovery_waitq,
 319                                   (cfs_atomic_read(&imp->imp_inflight) == 0),
 320                                   &lwi);
 321                 if (rc) {
 322                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 323
 324                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 325                                cli_tgt, rc,
 326                                cfs_atomic_read(&imp->imp_inflight));
 327
 328                         cfs_spin_lock(&imp->imp_lock);
 329                         if (cfs_atomic_read(&imp->imp_inflight) == 0) {
 330                                 int count = cfs_atomic_read(&imp->imp_unregistering);
 331
 332                                 /* We know that "unregistering" rpcs only can
 333                                  * survive in sending or delaying lists (they
 334                                  * maybe waiting for long reply unlink in
 335                                  * sluggish nets). Let's check this. If there
 336                                  * is no inflight and unregistering != 0, this
 337                                  * is bug. */
 338                                 LASSERTF(count == 0, "Some RPCs are still "
 339                                          "unregistering: %d\n", count);
 340
 341                                 /* Let's save one loop as soon as inflight have
 342                                  * dropped to zero. No new inflights possible at
 343                                  * this point. */
 344                                 rc = 0;
 345                         } else {
 346                                 cfs_list_for_each_safe(tmp, n,
 347                                                        &imp->imp_sending_list) {
 348                                         req = cfs_list_entry(tmp,
 349                                                              struct ptlrpc_request,
 350                                                              rq_list);
 351                                         DEBUG_REQ(D_ERROR, req,
 352                                                   "still on sending list");
 353                                 }
 354                                 cfs_list_for_each_safe(tmp, n,
 355                                                        &imp->imp_delayed_list) {
 356                                         req = cfs_list_entry(tmp,
 357                                                              struct ptlrpc_request,
 358                                                              rq_list);
 359                                         DEBUG_REQ(D_ERROR, req,
 360                                                   "still on delayed list");
 361                                 }
 362
 363                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 364                                        "Network is sluggish? Waiting them "
 365                                        "to error out.\n", cli_tgt,
 366                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 367                                        cfs_atomic_read(&imp->
 368                                                        imp_unregistering));
 369                         }
 370                         cfs_spin_unlock(&imp->imp_lock);
 371                   }
 372         } while (rc != 0);
 373
 374         /*
 375          * Let's additionally check that no new rpcs added to import in
 376          * "invalidate" state.
 377          */
 378         LASSERT(cfs_atomic_read(&imp->imp_inflight) == 0);
 379         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 380         sptlrpc_import_flush_all_ctx(imp);
 381
 382         cfs_atomic_dec(&imp->imp_inval_count);
 383         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
 384 }
 385
 386 /* unset imp_invalid */
 387 void ptlrpc_activate_import(struct obd_import *imp)
 388 {
 389         struct obd_device *obd = imp->imp_obd;
 390
 391         cfs_spin_lock(&imp->imp_lock);
 392         imp->imp_invalid = 0;
 393         ptlrpc_activate_timeouts(imp);
 394         cfs_spin_unlock(&imp->imp_lock);
 395         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 396 }
 397
 398 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 399 {
 400         ENTRY;
 401
 402         LASSERT(!imp->imp_dlm_fake);
 403
 404         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 405                 if (!imp->imp_replayable) {
 406                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 407                                "auto-deactivating\n",
 408                                obd2cli_tgt(imp->imp_obd),
 409                                imp->imp_connection->c_remote_uuid.uuid,
 410                                imp->imp_obd->obd_name);
 411                         ptlrpc_deactivate_import(imp);
 412                 }
 413
 414                 CDEBUG(D_HA, "%s: waking up pinger\n",
 415                        obd2cli_tgt(imp->imp_obd));
 416
 417                 cfs_spin_lock(&imp->imp_lock);
 418                 imp->imp_force_verify = 1;
 419                 cfs_spin_unlock(&imp->imp_lock);
 420
 421                 ptlrpc_pinger_wake_up();
 422         }
 423         EXIT;
 424 }
 425
 426 int ptlrpc_reconnect_import(struct obd_import *imp)
 427 {
 428         ptlrpc_set_import_discon(imp, 0);
 429         /* Force a new connect attempt */
 430         ptlrpc_invalidate_import(imp);
 431         /* Do a fresh connect next time by zeroing the handle */
 432         ptlrpc_disconnect_import(imp, 1);
 433         /* Wait for all invalidate calls to finish */
 434         if (cfs_atomic_read(&imp->imp_inval_count) > 0) {
 435                 int rc;
 436                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 437                 rc = l_wait_event(imp->imp_recovery_waitq,
 438                                   (cfs_atomic_read(&imp->imp_inval_count) == 0),
 439                                   &lwi);
 440                 if (rc)
 441                         CERROR("Interrupted, inval=%d\n",
 442                                cfs_atomic_read(&imp->imp_inval_count));
 443         }
 444
 445         /* Allow reconnect attempts */
 446         imp->imp_obd->obd_no_recov = 0;
 447         /* Remove 'invalid' flag */
 448         ptlrpc_activate_import(imp);
 449         /* Attempt a new connect */
 450         ptlrpc_recover_import(imp, NULL, 0);
 451         return 0;
 452 }
 453 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 454
 455 /**
 456  * Connection on import \a imp is changed to another one (if more than one is
 457  * present). We typically chose connection that we have not tried to connect to
 458  * the longest
 459  */
 460 static int import_select_connection(struct obd_import *imp)
 461 {
 462         struct obd_import_conn *imp_conn = NULL, *conn;
 463         struct obd_export *dlmexp;
 464         int tried_all = 1;
 465         ENTRY;
 466
 467         cfs_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                 cfs_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                 if (at_get(&imp->imp_at.iat_net_latency) <
 517                     CONNECTION_SWITCH_MAX) {
 518                         at_measured(&imp->imp_at.iat_net_latency,
 519                                     at_get(&imp->imp_at.iat_net_latency) +
 520                                     CONNECTION_SWITCH_INC);
 521                 }
 522                 LASSERT(imp_conn->oic_last_attempt);
 523                 CWARN("%s: tried all connections, increasing latency to %ds\n",
 524                       imp->imp_obd->obd_name,
 525                       at_get(&imp->imp_at.iat_net_latency));
 526         }
 527
 528         imp_conn->oic_last_attempt = cfs_time_current_64();
 529
 530         /* switch connection, don't mind if it's same as the current one */
 531         if (imp->imp_connection)
 532                 ptlrpc_connection_put(imp->imp_connection);
 533         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 534
 535         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 536         LASSERT(dlmexp != NULL);
 537         if (dlmexp->exp_connection)
 538                 ptlrpc_connection_put(dlmexp->exp_connection);
 539         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 540         class_export_put(dlmexp);
 541
 542         if (imp->imp_conn_current != imp_conn) {
 543                 if (imp->imp_conn_current)
 544                         CDEBUG(D_HA, "Changing connection for %s to %s/%s\n",
 545                                imp->imp_obd->obd_name, imp_conn->oic_uuid.uuid,
 546                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 547                 imp->imp_conn_current = imp_conn;
 548         }
 549
 550         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 551                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 552                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 553
 554         cfs_spin_unlock(&imp->imp_lock);
 555
 556         RETURN(0);
 557 }
 558
 559 /*
 560  * must be called under imp_lock
 561  */
 562 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 563 {
 564         struct ptlrpc_request *req;
 565         cfs_list_t *tmp;
 566
 567         if (cfs_list_empty(&imp->imp_replay_list))
 568                 return 0;
 569         tmp = imp->imp_replay_list.next;
 570         req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 571         *transno = req->rq_transno;
 572         if (req->rq_transno == 0) {
 573                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 574                 LBUG();
 575         }
 576
 577         return 1;
 578 }
 579
 580 /**
 581  * Attempt to (re)connect import \a imp. This includes all preparations,
 582  * initializing CONNECT RPC request and passing it to ptlrpcd for
 583  * actual sending.
 584  * Returns 0 on success or error code.
 585  */
 586 int ptlrpc_connect_import(struct obd_import *imp)
 587 {
 588         struct obd_device *obd = imp->imp_obd;
 589         int initial_connect = 0;
 590         int set_transno = 0;
 591         __u64 committed_before_reconnect = 0;
 592         struct ptlrpc_request *request;
 593         char *bufs[] = { NULL,
 594                          obd2cli_tgt(imp->imp_obd),
 595                          obd->obd_uuid.uuid,
 596                          (char *)&imp->imp_dlm_handle,
 597                          (char *)&imp->imp_connect_data };
 598         struct ptlrpc_connect_async_args *aa;
 599         int rc;
 600         ENTRY;
 601
 602         cfs_spin_lock(&imp->imp_lock);
 603         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 604                 cfs_spin_unlock(&imp->imp_lock);
 605                 CERROR("can't connect to a closed import\n");
 606                 RETURN(-EINVAL);
 607         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 608                 cfs_spin_unlock(&imp->imp_lock);
 609                 CERROR("already connected\n");
 610                 RETURN(0);
 611         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 612                 cfs_spin_unlock(&imp->imp_lock);
 613                 CERROR("already connecting\n");
 614                 RETURN(-EALREADY);
 615         }
 616
 617         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 618
 619         imp->imp_conn_cnt++;
 620         imp->imp_resend_replay = 0;
 621
 622         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 623                 initial_connect = 1;
 624         else
 625                 committed_before_reconnect = imp->imp_peer_committed_transno;
 626
 627         set_transno = ptlrpc_first_transno(imp,
 628                                            &imp->imp_connect_data.ocd_transno);
 629         cfs_spin_unlock(&imp->imp_lock);
 630
 631         rc = import_select_connection(imp);
 632         if (rc)
 633                 GOTO(out, rc);
 634
 635         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 636         if (rc)
 637                 GOTO(out, rc);
 638
 639         /* Reset connect flags to the originally requested flags, in case
 640          * the server is updated on-the-fly we will get the new features. */
 641         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 642         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 643         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 644
 645         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 646                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 647         if (rc)
 648                 GOTO(out, rc);
 649
 650         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 651         if (request == NULL)
 652                 GOTO(out, rc = -ENOMEM);
 653
 654         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 655                                       imp->imp_connect_op, bufs, NULL);
 656         if (rc) {
 657                 ptlrpc_request_free(request);
 658                 GOTO(out, rc);
 659         }
 660
 661         /* Report the rpc service time to the server so that it knows how long
 662          * to wait for clients to join recovery */
 663         lustre_msg_set_service_time(request->rq_reqmsg,
 664                                     at_timeout2est(request->rq_timeout));
 665
 666         /* The amount of time we give the server to process the connect req.
 667          * import_select_connection will increase the net latency on
 668          * repeated reconnect attempts to cover slow networks.
 669          * We override/ignore the server rpc completion estimate here,
 670          * which may be large if this is a reconnect attempt */
 671         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 672         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 673
 674 #ifndef __KERNEL__
 675         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 676 #endif
 677         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 678
 679         request->rq_no_resend = request->rq_no_delay = 1;
 680         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 681         /* Allow a slightly larger reply for future growth compatibility */
 682         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 683                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 684         ptlrpc_request_set_replen(request);
 685         request->rq_interpret_reply = ptlrpc_connect_interpret;
 686
 687         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 688         aa = ptlrpc_req_async_args(request);
 689         memset(aa, 0, sizeof *aa);
 690
 691         aa->pcaa_peer_committed = committed_before_reconnect;
 692         aa->pcaa_initial_connect = initial_connect;
 693
 694         if (aa->pcaa_initial_connect) {
 695                 cfs_spin_lock(&imp->imp_lock);
 696                 imp->imp_replayable = 1;
 697                 cfs_spin_unlock(&imp->imp_lock);
 698                 lustre_msg_add_op_flags(request->rq_reqmsg,
 699                                         MSG_CONNECT_INITIAL);
 700         }
 701
 702         if (set_transno)
 703                 lustre_msg_add_op_flags(request->rq_reqmsg,
 704                                         MSG_CONNECT_TRANSNO);
 705
 706         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 707                   request->rq_timeout);
 708         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 709         rc = 0;
 710 out:
 711         if (rc != 0) {
 712                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 713         }
 714
 715         RETURN(rc);
 716 }
 717 EXPORT_SYMBOL(ptlrpc_connect_import);
 718
 719 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 720 {
 721 #ifdef __KERNEL__
 722         int force_verify;
 723
 724         cfs_spin_lock(&imp->imp_lock);
 725         force_verify = imp->imp_force_verify != 0;
 726         cfs_spin_unlock(&imp->imp_lock);
 727
 728         if (force_verify)
 729                 ptlrpc_pinger_wake_up();
 730 #else
 731         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 732         ptlrpc_pinger_wake_up();
 733 #endif
 734 }
 735
 736 static int ptlrpc_busy_reconnect(int rc)
 737 {
 738         return (rc == -EBUSY) || (rc == -EAGAIN);
 739 }
 740
 741 /**
 742  * interpret_reply callback for connect RPCs.
 743  * Looks into returned status of connect operation and decides
 744  * what to do with the import - i.e enter recovery, promote it to
 745  * full state for normal operations of disconnect it due to an error.
 746  */
 747 static int ptlrpc_connect_interpret(const struct lu_env *env,
 748                                     struct ptlrpc_request *request,
 749                                     void *data, int rc)
 750 {
 751         struct ptlrpc_connect_async_args *aa = data;
 752         struct obd_import *imp = request->rq_import;
 753         struct client_obd *cli = &imp->imp_obd->u.cli;
 754         struct lustre_handle old_hdl;
 755         __u64 old_connect_flags;
 756         int msg_flags;
 757         ENTRY;
 758
 759         cfs_spin_lock(&imp->imp_lock);
 760         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 761                 cfs_spin_unlock(&imp->imp_lock);
 762                 RETURN(0);
 763         }
 764
 765         if (rc) {
 766                 /* if this reconnect to busy export - not need select new target
 767                  * for connecting*/
 768                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 769                 cfs_spin_unlock(&imp->imp_lock);
 770                 GOTO(out, rc);
 771         }
 772
 773         LASSERT(imp->imp_conn_current);
 774
 775         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 776
 777         /* All imports are pingable */
 778         imp->imp_pingable = 1;
 779         imp->imp_force_reconnect = 0;
 780         imp->imp_force_verify = 0;
 781
 782         if (aa->pcaa_initial_connect) {
 783                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 784                         imp->imp_replayable = 1;
 785                         cfs_spin_unlock(&imp->imp_lock);
 786                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 787                                obd2cli_tgt(imp->imp_obd));
 788                 } else {
 789                         imp->imp_replayable = 0;
 790                         cfs_spin_unlock(&imp->imp_lock);
 791                 }
 792
 793                 /* if applies, adjust the imp->imp_msg_magic here
 794                  * according to reply flags */
 795
 796                 imp->imp_remote_handle =
 797                                 *lustre_msg_get_handle(request->rq_repmsg);
 798
 799                 /* Initial connects are allowed for clients with non-random
 800                  * uuids when servers are in recovery.  Simply signal the
 801                  * servers replay is complete and wait in REPLAY_WAIT. */
 802                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 803                         CDEBUG(D_HA, "connect to %s during recovery\n",
 804                                obd2cli_tgt(imp->imp_obd));
 805                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 806                 } else {
 807                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 808                         ptlrpc_activate_import(imp);
 809                 }
 810
 811                 GOTO(finish, rc = 0);
 812         } else {
 813                 cfs_spin_unlock(&imp->imp_lock);
 814         }
 815
 816         /* Determine what recovery state to move the import to. */
 817         if (MSG_CONNECT_RECONNECT & msg_flags) {
 818                 memset(&old_hdl, 0, sizeof(old_hdl));
 819                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 820                             sizeof (old_hdl))) {
 821                         CERROR("%s@%s didn't like our handle "LPX64
 822                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 823                                imp->imp_connection->c_remote_uuid.uuid,
 824                                imp->imp_dlm_handle.cookie);
 825                         GOTO(out, rc = -ENOTCONN);
 826                 }
 827
 828                 if (memcmp(&imp->imp_remote_handle,
 829                            lustre_msg_get_handle(request->rq_repmsg),
 830                            sizeof(imp->imp_remote_handle))) {
 831                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 832                                 D_HA : D_WARNING;
 833
 834                         /* Bug 16611/14775: if server handle have changed,
 835                          * that means some sort of disconnection happened.
 836                          * If the server is not in recovery, that also means it
 837                          * already erased all of our state because of previous
 838                          * eviction. If it is in recovery - we are safe to
 839                          * participate since we can reestablish all of our state
 840                          * with server again */
 841                         CDEBUG(level,"%s@%s changed server handle from "
 842                                      LPX64" to "LPX64"%s\n",
 843                                      obd2cli_tgt(imp->imp_obd),
 844                                      imp->imp_connection->c_remote_uuid.uuid,
 845                                      imp->imp_remote_handle.cookie,
 846                                      lustre_msg_get_handle(request->rq_repmsg)->
 847                                                                         cookie,
 848                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
 849                                          " but is still in recovery" : "");
 850
 851                         imp->imp_remote_handle =
 852                                      *lustre_msg_get_handle(request->rq_repmsg);
 853
 854                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 855                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 856                                 GOTO(finish, rc = 0);
 857                         }
 858
 859                 } else {
 860                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 861                                obd2cli_tgt(imp->imp_obd),
 862                                imp->imp_connection->c_remote_uuid.uuid);
 863                 }
 864
 865                 if (imp->imp_invalid) {
 866                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 867                                "marking evicted\n", imp->imp_obd->obd_name);
 868                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 869                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 870                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 871                                imp->imp_obd->obd_name,
 872                                obd2cli_tgt(imp->imp_obd));
 873
 874                         cfs_spin_lock(&imp->imp_lock);
 875                         imp->imp_resend_replay = 1;
 876                         cfs_spin_unlock(&imp->imp_lock);
 877
 878                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 879                 } else {
 880                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 881                 }
 882         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 883                 LASSERT(imp->imp_replayable);
 884                 imp->imp_remote_handle =
 885                                 *lustre_msg_get_handle(request->rq_repmsg);
 886                 imp->imp_last_replay_transno = 0;
 887                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 888         } else {
 889                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 890                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 891                 imp->imp_remote_handle =
 892                                 *lustre_msg_get_handle(request->rq_repmsg);
 893                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 894         }
 895
 896         /* Sanity checks for a reconnected import. */
 897         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 898                 CERROR("imp_replayable flag does not match server "
 899                        "after reconnect. We should LBUG right here.\n");
 900         }
 901
 902         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 903             lustre_msg_get_last_committed(request->rq_repmsg) <
 904             aa->pcaa_peer_committed) {
 905                 CERROR("%s went back in time (transno "LPD64
 906                        " was previously committed, server now claims "LPD64
 907                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 908                        "id=9646\n",
 909                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 910                        lustre_msg_get_last_committed(request->rq_repmsg));
 911         }
 912
 913 finish:
 914         rc = ptlrpc_import_recovery_state_machine(imp);
 915         if (rc != 0) {
 916                 if (rc == -ENOTCONN) {
 917                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 918                                "invalidating and reconnecting\n",
 919                                obd2cli_tgt(imp->imp_obd),
 920                                imp->imp_connection->c_remote_uuid.uuid);
 921                         ptlrpc_connect_import(imp);
 922                         RETURN(0);
 923                 }
 924         } else {
 925                 struct obd_connect_data *ocd;
 926                 struct obd_export *exp;
 927                 int ret;
 928                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 929                                            RCL_SERVER);
 930                 /* server replied obd_connect_data is always bigger */
 931                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 932                                                    &RMF_CONNECT_DATA, ret);
 933
 934                 cfs_spin_lock(&imp->imp_lock);
 935                 cfs_list_del(&imp->imp_conn_current->oic_item);
 936                 cfs_list_add(&imp->imp_conn_current->oic_item,
 937                              &imp->imp_conn_list);
 938                 imp->imp_last_success_conn =
 939                         imp->imp_conn_current->oic_last_attempt;
 940
 941                 if (ocd == NULL) {
 942                         cfs_spin_unlock(&imp->imp_lock);
 943                         CERROR("Wrong connect data from server\n");
 944                         rc = -EPROTO;
 945                         GOTO(out, rc);
 946                 }
 947
 948                 imp->imp_connect_data = *ocd;
 949                 CDEBUG(D_HA, "obd %s to target with inst %u\n",
 950                        imp->imp_obd->obd_name, ocd->ocd_instance);
 951
 952                 exp = class_conn2export(&imp->imp_dlm_handle);
 953                 cfs_spin_unlock(&imp->imp_lock);
 954
 955                 /* check that server granted subset of flags we asked for. */
 956                 LASSERTF((ocd->ocd_connect_flags &
 957                           imp->imp_connect_flags_orig) ==
 958                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 959                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 960
 961                 if (!exp) {
 962                         /* This could happen if export is cleaned during the
 963                            connect attempt */
 964                         CERROR("Missing export for %s\n",
 965                                imp->imp_obd->obd_name);
 966                         GOTO(out, rc = -ENODEV);
 967                 }
 968                 old_connect_flags = exp->exp_connect_flags;
 969                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 970                 imp->imp_obd->obd_self_export->exp_connect_flags =
 971                                                         ocd->ocd_connect_flags;
 972                 class_export_put(exp);
 973
 974                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 975
 976                 if (!ocd->ocd_ibits_known &&
 977                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 978                         CERROR("Inodebits aware server returned zero compatible"
 979                                " bits?\n");
 980
 981                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 982                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 983                                         LUSTRE_VERSION_OFFSET_WARN ||
 984                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 985                                         LUSTRE_VERSION_OFFSET_WARN)) {
 986                         /* Sigh, some compilers do not like #ifdef in the middle
 987                            of macro arguments */
 988 #ifdef __KERNEL__
 989                         const char *older = "older. Consider upgrading server "
 990                                             "or downgrading client";
 991 #else
 992                         const char *older = "older. Consider recompiling this "
 993                                             "application";
 994 #endif
 995                         const char *newer = "newer than client version. "
 996                                             "Consider upgrading client";
 997
 998                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 999                                       "is much %s (%s)\n",
1000                                       obd2cli_tgt(imp->imp_obd),
1001                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1002                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1003                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1004                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1005                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1006                                       newer : older, LUSTRE_VERSION_STRING);
1007                 }
1008
1009                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1010                         /* We sent to the server ocd_cksum_types with bits set
1011                          * for algorithms we understand. The server masked off
1012                          * the checksum types it doesn't support */
1013                         if ((ocd->ocd_cksum_types & cksum_types_supported()) == 0) {
1014                                 LCONSOLE_WARN("The negotiation of the checksum "
1015                                               "alogrithm to use with server %s "
1016                                               "failed (%x/%x), disabling "
1017                                               "checksums\n",
1018                                               obd2cli_tgt(imp->imp_obd),
1019                                               ocd->ocd_cksum_types,
1020                                               cksum_types_supported());
1021                                 cli->cl_checksum = 0;
1022                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1023                         } else {
1024                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1025                         }
1026                 } else {
1027                         /* The server does not support OBD_CONNECT_CKSUM.
1028                          * Enforce CRC32 for backward compatibility*/
1029                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1030                 }
1031                 cli->cl_cksum_type =cksum_type_select(cli->cl_supp_cksum_types);
1032
1033                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1034                         cli->cl_max_pages_per_rpc =
1035                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1036                 else if (imp->imp_connect_op == MDS_CONNECT ||
1037                          imp->imp_connect_op == MGS_CONNECT)
1038                         cli->cl_max_pages_per_rpc = 1;
1039
1040                 /* Reset ns_connect_flags only for initial connect. It might be
1041                  * changed in while using FS and if we reset it in reconnect
1042                  * this leads to losing user settings done before such as
1043                  * disable lru_resize, etc. */
1044                 if (old_connect_flags != exp->exp_connect_flags ||
1045                     aa->pcaa_initial_connect) {
1046                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1047                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1048                               ocd->ocd_connect_flags);
1049                         imp->imp_obd->obd_namespace->ns_connect_flags =
1050                                 ocd->ocd_connect_flags;
1051                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1052                                 ocd->ocd_connect_flags;
1053                 }
1054
1055                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1056                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1057                         /* We need a per-message support flag, because
1058                            a. we don't know if the incoming connect reply
1059                               supports AT or not (in reply_in_callback)
1060                               until we unpack it.
1061                            b. failovered server means export and flags are gone
1062                               (in ptlrpc_send_reply).
1063                            Can only be set when we know AT is supported at
1064                            both ends */
1065                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1066                 else
1067                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1068
1069                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1070                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1071                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1072                 else
1073                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1074
1075                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1076                         (cli->cl_max_pages_per_rpc > 0));
1077         }
1078
1079 out:
1080         if (rc != 0) {
1081                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1082                 if (rc == -EACCES) {
1083                         /*
1084                          * Give up trying to reconnect
1085                          * EACCES means client has no permission for connection
1086                          */
1087                         imp->imp_obd->obd_no_recov = 1;
1088                         ptlrpc_deactivate_import(imp);
1089                 }
1090
1091                 if (rc == -EPROTO) {
1092                         struct obd_connect_data *ocd;
1093
1094                         /* reply message might not be ready */
1095                         if (request->rq_repmsg == NULL)
1096                                 RETURN(-EPROTO);
1097
1098                         ocd = req_capsule_server_get(&request->rq_pill,
1099                                                      &RMF_CONNECT_DATA);
1100                         if (ocd &&
1101                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1102                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1103                            /* Actually servers are only supposed to refuse
1104                               connection from liblustre clients, so we should
1105                               never see this from VFS context */
1106                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1107                                         "(%d.%d.%d.%d)"
1108                                         " refused connection from this client "
1109                                         "with an incompatible version (%s).  "
1110                                         "Client must be recompiled\n",
1111                                         obd2cli_tgt(imp->imp_obd),
1112                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1113                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1114                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1115                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1116                                         LUSTRE_VERSION_STRING);
1117                                 ptlrpc_deactivate_import(imp);
1118                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1119                         }
1120                         RETURN(-EPROTO);
1121                 }
1122
1123                 ptlrpc_maybe_ping_import_soon(imp);
1124
1125                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1126                        obd2cli_tgt(imp->imp_obd),
1127                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1128         }
1129
1130         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1131         RETURN(rc);
1132 }
1133
1134 /**
1135  * interpret callback for "completed replay" RPCs.
1136  * \see signal_completed_replay
1137  */
1138 static int completed_replay_interpret(const struct lu_env *env,
1139                                       struct ptlrpc_request *req,
1140                                       void * data, int rc)
1141 {
1142         ENTRY;
1143         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1144         if (req->rq_status == 0 &&
1145             !req->rq_import->imp_vbr_failed) {
1146                 ptlrpc_import_recovery_state_machine(req->rq_import);
1147         } else {
1148                 if (req->rq_import->imp_vbr_failed) {
1149                         CDEBUG(D_WARNING,
1150                                "%s: version recovery fails, reconnecting\n",
1151                                req->rq_import->imp_obd->obd_name);
1152                 } else {
1153                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1154                                      "reconnecting\n",
1155                                req->rq_import->imp_obd->obd_name,
1156                                req->rq_status);
1157                 }
1158                 ptlrpc_connect_import(req->rq_import);
1159         }
1160
1161         RETURN(0);
1162 }
1163
1164 /**
1165  * Let server know that we have no requests to replay anymore.
1166  * Achieved by just sending a PING request
1167  */
1168 static int signal_completed_replay(struct obd_import *imp)
1169 {
1170         struct ptlrpc_request *req;
1171         ENTRY;
1172
1173         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1174                 RETURN(0);
1175
1176         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1177         cfs_atomic_inc(&imp->imp_replay_inflight);
1178
1179         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1180                                         OBD_PING);
1181         if (req == NULL) {
1182                 cfs_atomic_dec(&imp->imp_replay_inflight);
1183                 RETURN(-ENOMEM);
1184         }
1185
1186         ptlrpc_request_set_replen(req);
1187         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1188         lustre_msg_add_flags(req->rq_reqmsg,
1189                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1190         if (AT_OFF)
1191                 req->rq_timeout *= 3;
1192         req->rq_interpret_reply = completed_replay_interpret;
1193
1194         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1195         RETURN(0);
1196 }
1197
1198 #ifdef __KERNEL__
1199 /**
1200  * In kernel code all import invalidation happens in its own
1201  * separate thread, so that whatever application happened to encounter
1202  * a problem could still be killed or otherwise continue
1203  */
1204 static int ptlrpc_invalidate_import_thread(void *data)
1205 {
1206         struct obd_import *imp = data;
1207
1208         ENTRY;
1209
1210         cfs_daemonize_ctxt("ll_imp_inval");
1211
1212         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1213                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1214                imp->imp_connection->c_remote_uuid.uuid);
1215
1216         ptlrpc_invalidate_import(imp);
1217
1218         if (obd_dump_on_eviction) {
1219                 CERROR("dump the log upon eviction\n");
1220                 libcfs_debug_dumplog();
1221         }
1222
1223         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1224         ptlrpc_import_recovery_state_machine(imp);
1225
1226         class_import_put(imp);
1227         RETURN(0);
1228 }
1229 #endif
1230
1231 /**
1232  * This is the state machine for client-side recovery on import.
1233  *
1234  * Typicaly we have two possibly paths. If we came to server and it is not
1235  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1236  * state and reconnect from scratch.
1237  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1238  * We go through our list of requests to replay and send them to server one by
1239  * one.
1240  * After sending all request from the list we change import state to
1241  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1242  * and also all the locks we don't yet have and wait for server to grant us.
1243  * After that we send a special "replay completed" request and change import
1244  * state to IMP_REPLAY_WAIT.
1245  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1246  * state and resend all requests from sending list.
1247  * After that we promote import to FULL state and send all delayed requests
1248  * and import is fully operational after that.
1249  *
1250  */
1251 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1252 {
1253         int rc = 0;
1254         int inflight;
1255         char *target_start;
1256         int target_len;
1257
1258         ENTRY;
1259         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1260                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1261                           &target_start, &target_len);
1262                 /* Don't care about MGC eviction */
1263                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1264                            LUSTRE_MGC_NAME) != 0) {
1265                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1266                                            "%.*s; in progress operations using "
1267                                            "this service will fail.\n",
1268                                            target_len, target_start);
1269                 }
1270                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1271                        obd2cli_tgt(imp->imp_obd),
1272                        imp->imp_connection->c_remote_uuid.uuid);
1273                 /* reset vbr_failed flag upon eviction */
1274                 cfs_spin_lock(&imp->imp_lock);
1275                 imp->imp_vbr_failed = 0;
1276                 cfs_spin_unlock(&imp->imp_lock);
1277
1278 #ifdef __KERNEL__
1279                 /* bug 17802:  XXX client_disconnect_export vs connect request
1280                  * race. if client will evicted at this time, we start
1281                  * invalidate thread without reference to import and import can
1282                  * be freed at same time. */
1283                 class_import_get(imp);
1284                 rc = cfs_create_thread(ptlrpc_invalidate_import_thread, imp,
1285                                        CFS_DAEMON_FLAGS);
1286                 if (rc < 0) {
1287                         class_import_put(imp);
1288                         CERROR("error starting invalidate thread: %d\n", rc);
1289                 } else {
1290                         rc = 0;
1291                 }
1292                 RETURN(rc);
1293 #else
1294                 ptlrpc_invalidate_import(imp);
1295
1296                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1297 #endif
1298         }
1299
1300         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1301                 CDEBUG(D_HA, "replay requested by %s\n",
1302                        obd2cli_tgt(imp->imp_obd));
1303                 rc = ptlrpc_replay_next(imp, &inflight);
1304                 if (inflight == 0 &&
1305                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1306                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1307                         rc = ldlm_replay_locks(imp);
1308                         if (rc)
1309                                 GOTO(out, rc);
1310                 }
1311                 rc = 0;
1312         }
1313
1314         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1315                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1316                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1317                         rc = signal_completed_replay(imp);
1318                         if (rc)
1319                                 GOTO(out, rc);
1320                 }
1321
1322         }
1323
1324         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1325                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1326                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1327                 }
1328         }
1329
1330         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1331                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1332                        obd2cli_tgt(imp->imp_obd),
1333                        imp->imp_connection->c_remote_uuid.uuid);
1334
1335                 rc = ptlrpc_resend(imp);
1336                 if (rc)
1337                         GOTO(out, rc);
1338                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1339                 ptlrpc_activate_import(imp);
1340
1341                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1342                           &target_start, &target_len);
1343                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1344                               "using nid %s.\n", imp->imp_obd->obd_name,
1345                               target_len, target_start,
1346                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1347         }
1348
1349         if (imp->imp_state == LUSTRE_IMP_FULL) {
1350                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1351                 ptlrpc_wake_delayed(imp);
1352         }
1353
1354 out:
1355         RETURN(rc);
1356 }
1357
1358 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1359 {
1360         struct ptlrpc_request *req;
1361         int rq_opc, rc = 0;
1362         int nowait = imp->imp_obd->obd_force;
1363         ENTRY;
1364
1365         if (nowait)
1366                 GOTO(set_state, rc);
1367
1368         switch (imp->imp_connect_op) {
1369         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1370         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1371         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1372         default:
1373                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1374                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1375                 RETURN(-EINVAL);
1376         }
1377
1378         if (ptlrpc_import_in_recovery(imp)) {
1379                 struct l_wait_info lwi;
1380                 cfs_duration_t timeout;
1381
1382
1383                 if (AT_OFF) {
1384                         if (imp->imp_server_timeout)
1385                                 timeout = cfs_time_seconds(obd_timeout / 2);
1386                         else
1387                                 timeout = cfs_time_seconds(obd_timeout);
1388                 } else {
1389                         int idx = import_at_get_index(imp,
1390                                 imp->imp_client->cli_request_portal);
1391                         timeout = cfs_time_seconds(
1392                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1393                 }
1394
1395                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1396                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1397                 rc = l_wait_event(imp->imp_recovery_waitq,
1398                                   !ptlrpc_import_in_recovery(imp), &lwi);
1399
1400         }
1401
1402         cfs_spin_lock(&imp->imp_lock);
1403         if (imp->imp_state != LUSTRE_IMP_FULL)
1404                 GOTO(out, 0);
1405
1406         cfs_spin_unlock(&imp->imp_lock);
1407
1408         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1409                                         LUSTRE_OBD_VERSION, rq_opc);
1410         if (req) {
1411                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1412                  * it fails.  We can get through the above with a down server
1413                  * if the client doesn't know the server is gone yet. */
1414                 req->rq_no_resend = 1;
1415
1416 #ifndef CRAY_XT3
1417                 /* We want client umounts to happen quickly, no matter the
1418                    server state... */
1419                 req->rq_timeout = min_t(int, req->rq_timeout,
1420                                         INITIAL_CONNECT_TIMEOUT);
1421 #else
1422                 /* ... but we always want liblustre clients to nicely
1423                    disconnect, so only use the adaptive value. */
1424                 if (AT_OFF)
1425                         req->rq_timeout = obd_timeout / 3;
1426 #endif
1427
1428                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1429                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1430                 ptlrpc_request_set_replen(req);
1431                 rc = ptlrpc_queue_wait(req);
1432                 ptlrpc_req_finished(req);
1433         }
1434
1435 set_state:
1436         cfs_spin_lock(&imp->imp_lock);
1437 out:
1438         if (noclose)
1439                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1440         else
1441                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1442         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1443         cfs_spin_unlock(&imp->imp_lock);
1444
1445         RETURN(rc);
1446 }
1447
1448 void ptlrpc_cleanup_imp(struct obd_import *imp)
1449 {
1450         ENTRY;
1451
1452         cfs_spin_lock(&imp->imp_lock);
1453         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1454         imp->imp_generation++;
1455         cfs_spin_unlock(&imp->imp_lock);
1456         ptlrpc_abort_inflight(imp);
1457
1458         EXIT;
1459 }
1460
1461 /* Adaptive Timeout utils */
1462 extern unsigned int at_min, at_max, at_history;
1463
1464 /* Bin into timeslices using AT_BINS bins.
1465    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1466    but still smoothing out a return to normalcy from a slow response.
1467    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1468 int at_measured(struct adaptive_timeout *at, unsigned int val)
1469 {
1470         unsigned int old = at->at_current;
1471         time_t now = cfs_time_current_sec();
1472         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1473
1474         LASSERT(at);
1475         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1476                val, at, now - at->at_binstart, at->at_current,
1477                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1478
1479         if (val == 0)
1480                 /* 0's don't count, because we never want our timeout to
1481                    drop to 0, and because 0 could mean an error */
1482                 return 0;
1483
1484         cfs_spin_lock(&at->at_lock);
1485
1486         if (unlikely(at->at_binstart == 0)) {
1487                 /* Special case to remove default from history */
1488                 at->at_current = val;
1489                 at->at_worst_ever = val;
1490                 at->at_worst_time = now;
1491                 at->at_hist[0] = val;
1492                 at->at_binstart = now;
1493         } else if (now - at->at_binstart < binlimit ) {
1494                 /* in bin 0 */
1495                 at->at_hist[0] = max(val, at->at_hist[0]);
1496                 at->at_current = max(val, at->at_current);
1497         } else {
1498                 int i, shift;
1499                 unsigned int maxv = val;
1500                 /* move bins over */
1501                 shift = (now - at->at_binstart) / binlimit;
1502                 LASSERT(shift > 0);
1503                 for(i = AT_BINS - 1; i >= 0; i--) {
1504                         if (i >= shift) {
1505                                 at->at_hist[i] = at->at_hist[i - shift];
1506                                 maxv = max(maxv, at->at_hist[i]);
1507                         } else {
1508                                 at->at_hist[i] = 0;
1509                         }
1510                 }
1511                 at->at_hist[0] = val;
1512                 at->at_current = maxv;
1513                 at->at_binstart += shift * binlimit;
1514         }
1515
1516         if (at->at_current > at->at_worst_ever) {
1517                 at->at_worst_ever = at->at_current;
1518                 at->at_worst_time = now;
1519         }
1520
1521         if (at->at_flags & AT_FLG_NOHIST)
1522                 /* Only keep last reported val; keeping the rest of the history
1523                    for proc only */
1524                 at->at_current = val;
1525
1526         if (at_max > 0)
1527                 at->at_current =  min(at->at_current, at_max);
1528         at->at_current =  max(at->at_current, at_min);
1529
1530         if (at->at_current != old)
1531                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1532                        "(val=%u) hist %u %u %u %u\n", at,
1533                        old, at->at_current, at->at_current - old, val,
1534                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1535                        at->at_hist[3]);
1536
1537         /* if we changed, report the old value */
1538         old = (at->at_current != old) ? old : 0;
1539
1540         cfs_spin_unlock(&at->at_lock);
1541         return old;
1542 }
1543
1544 /* Find the imp_at index for a given portal; assign if space available */
1545 int import_at_get_index(struct obd_import *imp, int portal)
1546 {
1547         struct imp_at *at = &imp->imp_at;
1548         int i;
1549
1550         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1551                 if (at->iat_portal[i] == portal)
1552                         return i;
1553                 if (at->iat_portal[i] == 0)
1554                         /* unused */
1555                         break;
1556         }
1557
1558         /* Not found in list, add it under a lock */
1559         cfs_spin_lock(&imp->imp_lock);
1560
1561         /* Check unused under lock */
1562         for (; i < IMP_AT_MAX_PORTALS; i++) {
1563                 if (at->iat_portal[i] == portal)
1564                         goto out;
1565                 if (at->iat_portal[i] == 0)
1566                         /* unused */
1567                         break;
1568         }
1569
1570         /* Not enough portals? */
1571         LASSERT(i < IMP_AT_MAX_PORTALS);
1572
1573         at->iat_portal[i] = portal;
1574 out:
1575         cfs_spin_unlock(&imp->imp_lock);
1576         return i;
1577 }