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  * Copyright (c) 2011, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  *
  38  * lustre/ptlrpc/import.c
  39  *
  40  * Author: Mike Shaver <shaver@clusterfs.com>
  41  */
  42
  43 #define DEBUG_SUBSYSTEM S_RPC
  44 #ifndef __KERNEL__
  45 # include <liblustre.h>
  46 #endif
  47
  48 #include <obd_support.h>
  49 #include <lustre_ha.h>
  50 #include <lustre_net.h>
  51 #include <lustre_import.h>
  52 #include <lustre_export.h>
  53 #include <obd.h>
  54 #include <obd_cksum.h>
  55 #include <obd_class.h>
  56
  57 #include "ptlrpc_internal.h"
  58
  59 struct ptlrpc_connect_async_args {
  60          __u64 pcaa_peer_committed;
  61         int pcaa_initial_connect;
  62 };
  63
  64 /**
  65  * Updates import \a imp current state to provided \a state value
  66  * Helper function. Must be called under imp_lock.
  67  */
  68 static void __import_set_state(struct obd_import *imp,
  69                                enum lustre_imp_state state)
  70 {
  71         imp->imp_state = state;
  72         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  73         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  74                 cfs_time_current_sec();
  75         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  76                 IMP_STATE_HIST_LEN;
  77 }
  78
  79 /* A CLOSED import should remain so. */
  80 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  81 do {                                                                           \
  82         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  83                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  84                       imp, obd2cli_tgt(imp->imp_obd),                          \
  85                       ptlrpc_import_state_name(imp->imp_state),                \
  86                       ptlrpc_import_state_name(state));                        \
  87                __import_set_state(imp, state);                                 \
  88         }                                                                      \
  89 } while(0)
  90
  91 #define IMPORT_SET_STATE(imp, state)            \
  92 do {                                            \
  93         cfs_spin_lock(&imp->imp_lock);          \
  94         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  95         cfs_spin_unlock(&imp->imp_lock);        \
  96 } while(0)
  97
  98
  99 static int ptlrpc_connect_interpret(const struct lu_env *env,
 100                                     struct ptlrpc_request *request,
 101                                     void * data, int rc);
 102 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 103
 104 /* Only this function is allowed to change the import state when it is
 105  * CLOSED. I would rather refcount the import and free it after
 106  * disconnection like we do with exports. To do that, the client_obd
 107  * will need to save the peer info somewhere other than in the import,
 108  * though. */
 109 int ptlrpc_init_import(struct obd_import *imp)
 110 {
 111         cfs_spin_lock(&imp->imp_lock);
 112
 113         imp->imp_generation++;
 114         imp->imp_state =  LUSTRE_IMP_NEW;
 115
 116         cfs_spin_unlock(&imp->imp_lock);
 117
 118         return 0;
 119 }
 120 EXPORT_SYMBOL(ptlrpc_init_import);
 121
 122 #define UUID_STR "_UUID"
 123 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
 124                       int *uuid_len)
 125 {
 126         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 127                 ? uuid : uuid + strlen(prefix);
 128
 129         *uuid_len = strlen(*uuid_start);
 130
 131         if (*uuid_len < strlen(UUID_STR))
 132                 return;
 133
 134         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 135                     UUID_STR, strlen(UUID_STR)))
 136                 *uuid_len -= strlen(UUID_STR);
 137 }
 138
 139 /**
 140  * Returns true if import was FULL, false if import was already not
 141  * connected.
 142  * @imp - import to be disconnected
 143  * @conn_cnt - connection count (epoch) of the request that timed out
 144  *             and caused the disconnection.  In some cases, multiple
 145  *             inflight requests can fail to a single target (e.g. OST
 146  *             bulk requests) and if one has already caused a reconnection
 147  *             (increasing the import->conn_cnt) the older failure should
 148  *             not also cause a reconnection.  If zero it forces a reconnect.
 149  */
 150 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 151 {
 152         int rc = 0;
 153
 154         cfs_spin_lock(&imp->imp_lock);
 155
 156         if (imp->imp_state == LUSTRE_IMP_FULL &&
 157             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 158                 char *target_start;
 159                 int   target_len;
 160
 161                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 162                           &target_start, &target_len);
 163
 164                 if (imp->imp_replayable) {
 165                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 166                                "%s was lost; in progress operations using this "
 167                                "service will wait for recovery to complete.\n",
 168                                imp->imp_obd->obd_name, target_len, target_start,
 169                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 170                 } else {
 171                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 172                                "%.*s via nid %s was lost; in progress "
 173                                "operations using this service will fail.\n",
 174                                imp->imp_obd->obd_name,
 175                                target_len, target_start,
 176                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 177                 }
 178                 ptlrpc_deactivate_timeouts(imp);
 179                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 180                 cfs_spin_unlock(&imp->imp_lock);
 181
 182                 if (obd_dump_on_timeout)
 183                         libcfs_debug_dumplog();
 184
 185                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 186                 rc = 1;
 187         } else {
 188                 cfs_spin_unlock(&imp->imp_lock);
 189                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 190                        imp->imp_client->cli_name, imp,
 191                        (imp->imp_state == LUSTRE_IMP_FULL &&
 192                         imp->imp_conn_cnt > conn_cnt) ?
 193                        "reconnected" : "not connected", imp->imp_conn_cnt,
 194                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 195         }
 196
 197         return rc;
 198 }
 199
 200 /* Must be called with imp_lock held! */
 201 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 202 {
 203         ENTRY;
 204         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 205
 206         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 207         imp->imp_invalid = 1;
 208         imp->imp_generation++;
 209         cfs_spin_unlock(&imp->imp_lock);
 210
 211         ptlrpc_abort_inflight(imp);
 212         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 213
 214         EXIT;
 215 }
 216
 217 /*
 218  * This acts as a barrier; all existing requests are rejected, and
 219  * no new requests will be accepted until the import is valid again.
 220  */
 221 void ptlrpc_deactivate_import(struct obd_import *imp)
 222 {
 223         cfs_spin_lock(&imp->imp_lock);
 224         ptlrpc_deactivate_and_unlock_import(imp);
 225 }
 226
 227 static unsigned int
 228 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 229 {
 230         long dl;
 231
 232         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 233               (req->rq_phase == RQ_PHASE_BULK) ||
 234               (req->rq_phase == RQ_PHASE_NEW)))
 235                 return 0;
 236
 237         if (req->rq_timedout)
 238                 return 0;
 239
 240         if (req->rq_phase == RQ_PHASE_NEW)
 241                 dl = req->rq_sent;
 242         else
 243                 dl = req->rq_deadline;
 244
 245         if (dl <= now)
 246                 return 0;
 247
 248         return dl - now;
 249 }
 250
 251 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 252 {
 253         time_t now = cfs_time_current_sec();
 254         cfs_list_t *tmp, *n;
 255         struct ptlrpc_request *req;
 256         unsigned int timeout = 0;
 257
 258         cfs_spin_lock(&imp->imp_lock);
 259         cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 260                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
 261                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 262         }
 263         cfs_spin_unlock(&imp->imp_lock);
 264         return timeout;
 265 }
 266
 267 /**
 268  * This function will invalidate the import, if necessary, then block
 269  * for all the RPC completions, and finally notify the obd to
 270  * invalidate its state (ie cancel locks, clear pending requests,
 271  * etc).
 272  */
 273 void ptlrpc_invalidate_import(struct obd_import *imp)
 274 {
 275         cfs_list_t *tmp, *n;
 276         struct ptlrpc_request *req;
 277         struct l_wait_info lwi;
 278         unsigned int timeout;
 279         int rc;
 280
 281         cfs_atomic_inc(&imp->imp_inval_count);
 282
 283         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 284                 ptlrpc_deactivate_import(imp);
 285
 286         LASSERT(imp->imp_invalid);
 287
 288         /* Wait forever until inflight == 0. We really can't do it another
 289          * way because in some cases we need to wait for very long reply
 290          * unlink. We can't do anything before that because there is really
 291          * no guarantee that some rdma transfer is not in progress right now. */
 292         do {
 293                 /* Calculate max timeout for waiting on rpcs to error
 294                  * out. Use obd_timeout if calculated value is smaller
 295                  * than it. */
 296                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 297                         timeout = ptlrpc_inflight_timeout(imp);
 298                         timeout += timeout / 3;
 299
 300                         if (timeout == 0)
 301                                 timeout = obd_timeout;
 302                 } else {
 303                         /* decrease the interval to increase race condition */
 304                         timeout = 1;
 305                 }
 306
 307                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 308                        timeout);
 309
 310                 /* Wait for all requests to error out and call completion
 311                  * callbacks. Cap it at obd_timeout -- these should all
 312                  * have been locally cancelled by ptlrpc_abort_inflight. */
 313                 lwi = LWI_TIMEOUT_INTERVAL(
 314                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 315                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 316                         NULL, NULL);
 317                 rc = l_wait_event(imp->imp_recovery_waitq,
 318                                   (cfs_atomic_read(&imp->imp_inflight) == 0),
 319                                   &lwi);
 320                 if (rc) {
 321                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 322
 323                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 324                                cli_tgt, rc,
 325                                cfs_atomic_read(&imp->imp_inflight));
 326
 327                         cfs_spin_lock(&imp->imp_lock);
 328                         if (cfs_atomic_read(&imp->imp_inflight) == 0) {
 329                                 int count = cfs_atomic_read(&imp->imp_unregistering);
 330
 331                                 /* We know that "unregistering" rpcs only can
 332                                  * survive in sending or delaying lists (they
 333                                  * maybe waiting for long reply unlink in
 334                                  * sluggish nets). Let's check this. If there
 335                                  * is no inflight and unregistering != 0, this
 336                                  * is bug. */
 337                                 LASSERTF(count == 0, "Some RPCs are still "
 338                                          "unregistering: %d\n", count);
 339
 340                                 /* Let's save one loop as soon as inflight have
 341                                  * dropped to zero. No new inflights possible at
 342                                  * this point. */
 343                                 rc = 0;
 344                         } else {
 345                                 cfs_list_for_each_safe(tmp, n,
 346                                                        &imp->imp_sending_list) {
 347                                         req = cfs_list_entry(tmp,
 348                                                              struct ptlrpc_request,
 349                                                              rq_list);
 350                                         DEBUG_REQ(D_ERROR, req,
 351                                                   "still on sending list");
 352                                 }
 353                                 cfs_list_for_each_safe(tmp, n,
 354                                                        &imp->imp_delayed_list) {
 355                                         req = cfs_list_entry(tmp,
 356                                                              struct ptlrpc_request,
 357                                                              rq_list);
 358                                         DEBUG_REQ(D_ERROR, req,
 359                                                   "still on delayed list");
 360                                 }
 361
 362                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 363                                        "Network is sluggish? Waiting them "
 364                                        "to error out.\n", cli_tgt,
 365                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 366                                        cfs_atomic_read(&imp->
 367                                                        imp_unregistering));
 368                         }
 369                         cfs_spin_unlock(&imp->imp_lock);
 370                   }
 371         } while (rc != 0);
 372
 373         /*
 374          * Let's additionally check that no new rpcs added to import in
 375          * "invalidate" state.
 376          */
 377         LASSERT(cfs_atomic_read(&imp->imp_inflight) == 0);
 378         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 379         sptlrpc_import_flush_all_ctx(imp);
 380
 381         cfs_atomic_dec(&imp->imp_inval_count);
 382         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
 383 }
 384
 385 /* unset imp_invalid */
 386 void ptlrpc_activate_import(struct obd_import *imp)
 387 {
 388         struct obd_device *obd = imp->imp_obd;
 389
 390         cfs_spin_lock(&imp->imp_lock);
 391         imp->imp_invalid = 0;
 392         ptlrpc_activate_timeouts(imp);
 393         cfs_spin_unlock(&imp->imp_lock);
 394         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 395 }
 396
 397 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 398 {
 399         ENTRY;
 400
 401         LASSERT(!imp->imp_dlm_fake);
 402
 403         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 404                 if (!imp->imp_replayable) {
 405                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 406                                "auto-deactivating\n",
 407                                obd2cli_tgt(imp->imp_obd),
 408                                imp->imp_connection->c_remote_uuid.uuid,
 409                                imp->imp_obd->obd_name);
 410                         ptlrpc_deactivate_import(imp);
 411                 }
 412
 413                 CDEBUG(D_HA, "%s: waking up pinger\n",
 414                        obd2cli_tgt(imp->imp_obd));
 415
 416                 cfs_spin_lock(&imp->imp_lock);
 417                 imp->imp_force_verify = 1;
 418                 cfs_spin_unlock(&imp->imp_lock);
 419
 420                 ptlrpc_pinger_wake_up();
 421         }
 422         EXIT;
 423 }
 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         int tried_all = 1;
 464         ENTRY;
 465
 466         cfs_spin_lock(&imp->imp_lock);
 467
 468         if (cfs_list_empty(&imp->imp_conn_list)) {
 469                 CERROR("%s: no connections available\n",
 470                         imp->imp_obd->obd_name);
 471                 cfs_spin_unlock(&imp->imp_lock);
 472                 RETURN(-EINVAL);
 473         }
 474
 475         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 476                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 477                        imp->imp_obd->obd_name,
 478                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 479                        conn->oic_last_attempt);
 480
 481                 /* If we have not tried this connection since
 482                    the last successful attempt, go with this one */
 483                 if ((conn->oic_last_attempt == 0) ||
 484                     cfs_time_beforeq_64(conn->oic_last_attempt,
 485                                        imp->imp_last_success_conn)) {
 486                         imp_conn = conn;
 487                         tried_all = 0;
 488                         break;
 489                 }
 490
 491                 /* If all of the connections have already been tried
 492                    since the last successful connection; just choose the
 493                    least recently used */
 494                 if (!imp_conn)
 495                         imp_conn = conn;
 496                 else if (cfs_time_before_64(conn->oic_last_attempt,
 497                                             imp_conn->oic_last_attempt))
 498                         imp_conn = conn;
 499         }
 500
 501         /* if not found, simply choose the current one */
 502         if (!imp_conn || imp->imp_force_reconnect) {
 503                 LASSERT(imp->imp_conn_current);
 504                 imp_conn = imp->imp_conn_current;
 505                 tried_all = 0;
 506         }
 507         LASSERT(imp_conn->oic_conn);
 508
 509         /* If we've tried everything, and we're back to the beginning of the
 510            list, increase our timeout and try again. It will be reset when
 511            we do finally connect. (FIXME: really we should wait for all network
 512            state associated with the last connection attempt to drain before
 513            trying to reconnect on it.) */
 514         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 515                 if (at_get(&imp->imp_at.iat_net_latency) <
 516                     CONNECTION_SWITCH_MAX) {
 517                         at_measured(&imp->imp_at.iat_net_latency,
 518                                     at_get(&imp->imp_at.iat_net_latency) +
 519                                     CONNECTION_SWITCH_INC);
 520                 }
 521                 LASSERT(imp_conn->oic_last_attempt);
 522                 CWARN("%s: tried all connections, increasing latency to %ds\n",
 523                       imp->imp_obd->obd_name,
 524                       at_get(&imp->imp_at.iat_net_latency));
 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                         CDEBUG(D_HA, "Changing connection for %s to %s/%s\n",
 544                                imp->imp_obd->obd_name, imp_conn->oic_uuid.uuid,
 545                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 546                 imp->imp_conn_current = imp_conn;
 547         }
 548
 549         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 550                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 551                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 552
 553         cfs_spin_unlock(&imp->imp_lock);
 554
 555         RETURN(0);
 556 }
 557
 558 /*
 559  * must be called under imp_lock
 560  */
 561 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 562 {
 563         struct ptlrpc_request *req;
 564         cfs_list_t *tmp;
 565
 566         if (cfs_list_empty(&imp->imp_replay_list))
 567                 return 0;
 568         tmp = imp->imp_replay_list.next;
 569         req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 570         *transno = req->rq_transno;
 571         if (req->rq_transno == 0) {
 572                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 573                 LBUG();
 574         }
 575
 576         return 1;
 577 }
 578
 579 /**
 580  * Attempt to (re)connect import \a imp. This includes all preparations,
 581  * initializing CONNECT RPC request and passing it to ptlrpcd for
 582  * actual sending.
 583  * Returns 0 on success or error code.
 584  */
 585 int ptlrpc_connect_import(struct obd_import *imp)
 586 {
 587         struct obd_device *obd = imp->imp_obd;
 588         int initial_connect = 0;
 589         int set_transno = 0;
 590         __u64 committed_before_reconnect = 0;
 591         struct ptlrpc_request *request;
 592         char *bufs[] = { NULL,
 593                          obd2cli_tgt(imp->imp_obd),
 594                          obd->obd_uuid.uuid,
 595                          (char *)&imp->imp_dlm_handle,
 596                          (char *)&imp->imp_connect_data };
 597         struct ptlrpc_connect_async_args *aa;
 598         int rc;
 599         ENTRY;
 600
 601         cfs_spin_lock(&imp->imp_lock);
 602         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 603                 cfs_spin_unlock(&imp->imp_lock);
 604                 CERROR("can't connect to a closed import\n");
 605                 RETURN(-EINVAL);
 606         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 607                 cfs_spin_unlock(&imp->imp_lock);
 608                 CERROR("already connected\n");
 609                 RETURN(0);
 610         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 611                 cfs_spin_unlock(&imp->imp_lock);
 612                 CERROR("already connecting\n");
 613                 RETURN(-EALREADY);
 614         }
 615
 616         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 617
 618         imp->imp_conn_cnt++;
 619         imp->imp_resend_replay = 0;
 620
 621         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 622                 initial_connect = 1;
 623         else
 624                 committed_before_reconnect = imp->imp_peer_committed_transno;
 625
 626         set_transno = ptlrpc_first_transno(imp,
 627                                            &imp->imp_connect_data.ocd_transno);
 628         cfs_spin_unlock(&imp->imp_lock);
 629
 630         rc = import_select_connection(imp);
 631         if (rc)
 632                 GOTO(out, rc);
 633
 634         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 635         if (rc)
 636                 GOTO(out, rc);
 637
 638         /* Reset connect flags to the originally requested flags, in case
 639          * the server is updated on-the-fly we will get the new features. */
 640         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 641         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 642         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 643
 644         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 645                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 646         if (rc)
 647                 GOTO(out, rc);
 648
 649         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 650         if (request == NULL)
 651                 GOTO(out, rc = -ENOMEM);
 652
 653         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 654                                       imp->imp_connect_op, bufs, NULL);
 655         if (rc) {
 656                 ptlrpc_request_free(request);
 657                 GOTO(out, rc);
 658         }
 659
 660         /* Report the rpc service time to the server so that it knows how long
 661          * to wait for clients to join recovery */
 662         lustre_msg_set_service_time(request->rq_reqmsg,
 663                                     at_timeout2est(request->rq_timeout));
 664
 665         /* The amount of time we give the server to process the connect req.
 666          * import_select_connection will increase the net latency on
 667          * repeated reconnect attempts to cover slow networks.
 668          * We override/ignore the server rpc completion estimate here,
 669          * which may be large if this is a reconnect attempt */
 670         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 671         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 672
 673 #ifndef __KERNEL__
 674         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 675 #endif
 676         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 677
 678         request->rq_no_resend = request->rq_no_delay = 1;
 679         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 680         /* Allow a slightly larger reply for future growth compatibility */
 681         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 682                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 683         ptlrpc_request_set_replen(request);
 684         request->rq_interpret_reply = ptlrpc_connect_interpret;
 685
 686         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 687         aa = ptlrpc_req_async_args(request);
 688         memset(aa, 0, sizeof *aa);
 689
 690         aa->pcaa_peer_committed = committed_before_reconnect;
 691         aa->pcaa_initial_connect = initial_connect;
 692
 693         if (aa->pcaa_initial_connect) {
 694                 cfs_spin_lock(&imp->imp_lock);
 695                 imp->imp_replayable = 1;
 696                 cfs_spin_unlock(&imp->imp_lock);
 697                 lustre_msg_add_op_flags(request->rq_reqmsg,
 698                                         MSG_CONNECT_INITIAL);
 699         }
 700
 701         if (set_transno)
 702                 lustre_msg_add_op_flags(request->rq_reqmsg,
 703                                         MSG_CONNECT_TRANSNO);
 704
 705         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 706                   request->rq_timeout);
 707         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 708         rc = 0;
 709 out:
 710         if (rc != 0) {
 711                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 712         }
 713
 714         RETURN(rc);
 715 }
 716 EXPORT_SYMBOL(ptlrpc_connect_import);
 717
 718 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 719 {
 720 #ifdef __KERNEL__
 721         int force_verify;
 722
 723         cfs_spin_lock(&imp->imp_lock);
 724         force_verify = imp->imp_force_verify != 0;
 725         cfs_spin_unlock(&imp->imp_lock);
 726
 727         if (force_verify)
 728                 ptlrpc_pinger_wake_up();
 729 #else
 730         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 731         ptlrpc_pinger_wake_up();
 732 #endif
 733 }
 734
 735 static int ptlrpc_busy_reconnect(int rc)
 736 {
 737         return (rc == -EBUSY) || (rc == -EAGAIN);
 738 }
 739
 740 /**
 741  * interpret_reply callback for connect RPCs.
 742  * Looks into returned status of connect operation and decides
 743  * what to do with the import - i.e enter recovery, promote it to
 744  * full state for normal operations of disconnect it due to an error.
 745  */
 746 static int ptlrpc_connect_interpret(const struct lu_env *env,
 747                                     struct ptlrpc_request *request,
 748                                     void *data, int rc)
 749 {
 750         struct ptlrpc_connect_async_args *aa = data;
 751         struct obd_import *imp = request->rq_import;
 752         struct client_obd *cli = &imp->imp_obd->u.cli;
 753         struct lustre_handle old_hdl;
 754         __u64 old_connect_flags;
 755         int msg_flags;
 756         ENTRY;
 757
 758         cfs_spin_lock(&imp->imp_lock);
 759         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 760                 cfs_spin_unlock(&imp->imp_lock);
 761                 RETURN(0);
 762         }
 763
 764         if (rc) {
 765                 /* if this reconnect to busy export - not need select new target
 766                  * for connecting*/
 767                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 768                 cfs_spin_unlock(&imp->imp_lock);
 769                 GOTO(out, rc);
 770         }
 771
 772         LASSERT(imp->imp_conn_current);
 773
 774         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 775
 776         /* All imports are pingable */
 777         imp->imp_pingable = 1;
 778         imp->imp_force_reconnect = 0;
 779         imp->imp_force_verify = 0;
 780
 781         if (aa->pcaa_initial_connect) {
 782                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 783                         imp->imp_replayable = 1;
 784                         cfs_spin_unlock(&imp->imp_lock);
 785                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 786                                obd2cli_tgt(imp->imp_obd));
 787                 } else {
 788                         imp->imp_replayable = 0;
 789                         cfs_spin_unlock(&imp->imp_lock);
 790                 }
 791
 792                 /* if applies, adjust the imp->imp_msg_magic here
 793                  * according to reply flags */
 794
 795                 imp->imp_remote_handle =
 796                                 *lustre_msg_get_handle(request->rq_repmsg);
 797
 798                 /* Initial connects are allowed for clients with non-random
 799                  * uuids when servers are in recovery.  Simply signal the
 800                  * servers replay is complete and wait in REPLAY_WAIT. */
 801                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 802                         CDEBUG(D_HA, "connect to %s during recovery\n",
 803                                obd2cli_tgt(imp->imp_obd));
 804                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 805                 } else {
 806                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 807                         ptlrpc_activate_import(imp);
 808                 }
 809
 810                 GOTO(finish, rc = 0);
 811         } else {
 812                 cfs_spin_unlock(&imp->imp_lock);
 813         }
 814
 815         /* Determine what recovery state to move the import to. */
 816         if (MSG_CONNECT_RECONNECT & msg_flags) {
 817                 memset(&old_hdl, 0, sizeof(old_hdl));
 818                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 819                             sizeof (old_hdl))) {
 820                         CERROR("%s@%s didn't like our handle "LPX64
 821                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 822                                imp->imp_connection->c_remote_uuid.uuid,
 823                                imp->imp_dlm_handle.cookie);
 824                         GOTO(out, rc = -ENOTCONN);
 825                 }
 826
 827                 if (memcmp(&imp->imp_remote_handle,
 828                            lustre_msg_get_handle(request->rq_repmsg),
 829                            sizeof(imp->imp_remote_handle))) {
 830                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 831                                 D_HA : D_WARNING;
 832
 833                         /* Bug 16611/14775: if server handle have changed,
 834                          * that means some sort of disconnection happened.
 835                          * If the server is not in recovery, that also means it
 836                          * already erased all of our state because of previous
 837                          * eviction. If it is in recovery - we are safe to
 838                          * participate since we can reestablish all of our state
 839                          * with server again */
 840                         CDEBUG(level,"%s@%s changed server handle from "
 841                                      LPX64" to "LPX64"%s\n",
 842                                      obd2cli_tgt(imp->imp_obd),
 843                                      imp->imp_connection->c_remote_uuid.uuid,
 844                                      imp->imp_remote_handle.cookie,
 845                                      lustre_msg_get_handle(request->rq_repmsg)->
 846                                                                         cookie,
 847                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
 848                                          " but is still in recovery" : "");
 849
 850                         imp->imp_remote_handle =
 851                                      *lustre_msg_get_handle(request->rq_repmsg);
 852
 853                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 854                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 855                                 GOTO(finish, rc = 0);
 856                         }
 857
 858                 } else {
 859                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 860                                obd2cli_tgt(imp->imp_obd),
 861                                imp->imp_connection->c_remote_uuid.uuid);
 862                 }
 863
 864                 if (imp->imp_invalid) {
 865                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 866                                "marking evicted\n", imp->imp_obd->obd_name);
 867                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 868                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 869                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 870                                imp->imp_obd->obd_name,
 871                                obd2cli_tgt(imp->imp_obd));
 872
 873                         cfs_spin_lock(&imp->imp_lock);
 874                         imp->imp_resend_replay = 1;
 875                         cfs_spin_unlock(&imp->imp_lock);
 876
 877                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 878                 } else {
 879                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 880                 }
 881         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 882                 LASSERT(imp->imp_replayable);
 883                 imp->imp_remote_handle =
 884                                 *lustre_msg_get_handle(request->rq_repmsg);
 885                 imp->imp_last_replay_transno = 0;
 886                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 887         } else {
 888                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 889                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 890                 imp->imp_remote_handle =
 891                                 *lustre_msg_get_handle(request->rq_repmsg);
 892                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 893         }
 894
 895         /* Sanity checks for a reconnected import. */
 896         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 897                 CERROR("imp_replayable flag does not match server "
 898                        "after reconnect. We should LBUG right here.\n");
 899         }
 900
 901         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 902             lustre_msg_get_last_committed(request->rq_repmsg) <
 903             aa->pcaa_peer_committed) {
 904                 CERROR("%s went back in time (transno "LPD64
 905                        " was previously committed, server now claims "LPD64
 906                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 907                        "id=9646\n",
 908                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 909                        lustre_msg_get_last_committed(request->rq_repmsg));
 910         }
 911
 912 finish:
 913         rc = ptlrpc_import_recovery_state_machine(imp);
 914         if (rc != 0) {
 915                 if (rc == -ENOTCONN) {
 916                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 917                                "invalidating and reconnecting\n",
 918                                obd2cli_tgt(imp->imp_obd),
 919                                imp->imp_connection->c_remote_uuid.uuid);
 920                         ptlrpc_connect_import(imp);
 921                         RETURN(0);
 922                 }
 923         } else {
 924                 struct obd_connect_data *ocd;
 925                 struct obd_export *exp;
 926                 int ret;
 927                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 928                                            RCL_SERVER);
 929                 /* server replied obd_connect_data is always bigger */
 930                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 931                                                    &RMF_CONNECT_DATA, ret);
 932
 933                 cfs_spin_lock(&imp->imp_lock);
 934                 cfs_list_del(&imp->imp_conn_current->oic_item);
 935                 cfs_list_add(&imp->imp_conn_current->oic_item,
 936                              &imp->imp_conn_list);
 937                 imp->imp_last_success_conn =
 938                         imp->imp_conn_current->oic_last_attempt;
 939
 940                 if (ocd == NULL) {
 941                         cfs_spin_unlock(&imp->imp_lock);
 942                         CERROR("Wrong connect data from server\n");
 943                         rc = -EPROTO;
 944                         GOTO(out, rc);
 945                 }
 946
 947                 imp->imp_connect_data = *ocd;
 948                 CDEBUG(D_HA, "obd %s to target with inst %u\n",
 949                        imp->imp_obd->obd_name, ocd->ocd_instance);
 950
 951                 exp = class_conn2export(&imp->imp_dlm_handle);
 952                 cfs_spin_unlock(&imp->imp_lock);
 953
 954                 /* check that server granted subset of flags we asked for. */
 955                 LASSERTF((ocd->ocd_connect_flags &
 956                           imp->imp_connect_flags_orig) ==
 957                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 958                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 959
 960                 if (!exp) {
 961                         /* This could happen if export is cleaned during the
 962                            connect attempt */
 963                         CERROR("Missing export for %s\n",
 964                                imp->imp_obd->obd_name);
 965                         GOTO(out, rc = -ENODEV);
 966                 }
 967                 old_connect_flags = exp->exp_connect_flags;
 968                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 969                 imp->imp_obd->obd_self_export->exp_connect_flags =
 970                                                         ocd->ocd_connect_flags;
 971                 class_export_put(exp);
 972
 973                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 974
 975                 if (!ocd->ocd_ibits_known &&
 976                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 977                         CERROR("Inodebits aware server returned zero compatible"
 978                                " bits?\n");
 979
 980                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 981                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 982                                         LUSTRE_VERSION_OFFSET_WARN ||
 983                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 984                                         LUSTRE_VERSION_OFFSET_WARN)) {
 985                         /* Sigh, some compilers do not like #ifdef in the middle
 986                            of macro arguments */
 987 #ifdef __KERNEL__
 988                         const char *older = "older. Consider upgrading server "
 989                                             "or downgrading client";
 990 #else
 991                         const char *older = "older. Consider recompiling this "
 992                                             "application";
 993 #endif
 994                         const char *newer = "newer than client version. "
 995                                             "Consider upgrading client";
 996
 997                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 998                                       "is much %s (%s)\n",
 999                                       obd2cli_tgt(imp->imp_obd),
1000                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1001                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1002                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1003                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1004                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1005                                       newer : older, LUSTRE_VERSION_STRING);
1006                 }
1007
1008                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1009                         /* We sent to the server ocd_cksum_types with bits set
1010                          * for algorithms we understand. The server masked off
1011                          * the checksum types it doesn't support */
1012                         if ((ocd->ocd_cksum_types & cksum_types_supported()) == 0) {
1013                                 LCONSOLE_WARN("The negotiation of the checksum "
1014                                               "alogrithm to use with server %s "
1015                                               "failed (%x/%x), disabling "
1016                                               "checksums\n",
1017                                               obd2cli_tgt(imp->imp_obd),
1018                                               ocd->ocd_cksum_types,
1019                                               cksum_types_supported());
1020                                 cli->cl_checksum = 0;
1021                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1022                         } else {
1023                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1024                         }
1025                 } else {
1026                         /* The server does not support OBD_CONNECT_CKSUM.
1027                          * Enforce CRC32 for backward compatibility*/
1028                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1029                 }
1030                 cli->cl_cksum_type =cksum_type_select(cli->cl_supp_cksum_types);
1031
1032                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1033                         cli->cl_max_pages_per_rpc =
1034                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1035                 else if (imp->imp_connect_op == MDS_CONNECT ||
1036                          imp->imp_connect_op == MGS_CONNECT)
1037                         cli->cl_max_pages_per_rpc = 1;
1038
1039                 /* Reset ns_connect_flags only for initial connect. It might be
1040                  * changed in while using FS and if we reset it in reconnect
1041                  * this leads to losing user settings done before such as
1042                  * disable lru_resize, etc. */
1043                 if (old_connect_flags != exp->exp_connect_flags ||
1044                     aa->pcaa_initial_connect) {
1045                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1046                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1047                               ocd->ocd_connect_flags);
1048                         imp->imp_obd->obd_namespace->ns_connect_flags =
1049                                 ocd->ocd_connect_flags;
1050                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1051                                 ocd->ocd_connect_flags;
1052                 }
1053
1054                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1055                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1056                         /* We need a per-message support flag, because
1057                            a. we don't know if the incoming connect reply
1058                               supports AT or not (in reply_in_callback)
1059                               until we unpack it.
1060                            b. failovered server means export and flags are gone
1061                               (in ptlrpc_send_reply).
1062                            Can only be set when we know AT is supported at
1063                            both ends */
1064                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1065                 else
1066                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1067
1068                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1069                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1070                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1071                 else
1072                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1073
1074                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1075                         (cli->cl_max_pages_per_rpc > 0));
1076         }
1077
1078 out:
1079         if (rc != 0) {
1080                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1081                 if (rc == -EACCES) {
1082                         /*
1083                          * Give up trying to reconnect
1084                          * EACCES means client has no permission for connection
1085                          */
1086                         imp->imp_obd->obd_no_recov = 1;
1087                         ptlrpc_deactivate_import(imp);
1088                 }
1089
1090                 if (rc == -EPROTO) {
1091                         struct obd_connect_data *ocd;
1092
1093                         /* reply message might not be ready */
1094                         if (request->rq_repmsg == NULL)
1095                                 RETURN(-EPROTO);
1096
1097                         ocd = req_capsule_server_get(&request->rq_pill,
1098                                                      &RMF_CONNECT_DATA);
1099                         if (ocd &&
1100                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1101                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1102                            /* Actually servers are only supposed to refuse
1103                               connection from liblustre clients, so we should
1104                               never see this from VFS context */
1105                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1106                                         "(%d.%d.%d.%d)"
1107                                         " refused connection from this client "
1108                                         "with an incompatible version (%s).  "
1109                                         "Client must be recompiled\n",
1110                                         obd2cli_tgt(imp->imp_obd),
1111                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1112                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1113                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1114                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1115                                         LUSTRE_VERSION_STRING);
1116                                 ptlrpc_deactivate_import(imp);
1117                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1118                         }
1119                         RETURN(-EPROTO);
1120                 }
1121
1122                 ptlrpc_maybe_ping_import_soon(imp);
1123
1124                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1125                        obd2cli_tgt(imp->imp_obd),
1126                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1127         }
1128
1129         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1130         RETURN(rc);
1131 }
1132
1133 /**
1134  * interpret callback for "completed replay" RPCs.
1135  * \see signal_completed_replay
1136  */
1137 static int completed_replay_interpret(const struct lu_env *env,
1138                                       struct ptlrpc_request *req,
1139                                       void * data, int rc)
1140 {
1141         ENTRY;
1142         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1143         if (req->rq_status == 0 &&
1144             !req->rq_import->imp_vbr_failed) {
1145                 ptlrpc_import_recovery_state_machine(req->rq_import);
1146         } else {
1147                 if (req->rq_import->imp_vbr_failed) {
1148                         CDEBUG(D_WARNING,
1149                                "%s: version recovery fails, reconnecting\n",
1150                                req->rq_import->imp_obd->obd_name);
1151                 } else {
1152                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1153                                      "reconnecting\n",
1154                                req->rq_import->imp_obd->obd_name,
1155                                req->rq_status);
1156                 }
1157                 ptlrpc_connect_import(req->rq_import);
1158         }
1159
1160         RETURN(0);
1161 }
1162
1163 /**
1164  * Let server know that we have no requests to replay anymore.
1165  * Achieved by just sending a PING request
1166  */
1167 static int signal_completed_replay(struct obd_import *imp)
1168 {
1169         struct ptlrpc_request *req;
1170         ENTRY;
1171
1172         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1173                 RETURN(0);
1174
1175         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1176         cfs_atomic_inc(&imp->imp_replay_inflight);
1177
1178         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1179                                         OBD_PING);
1180         if (req == NULL) {
1181                 cfs_atomic_dec(&imp->imp_replay_inflight);
1182                 RETURN(-ENOMEM);
1183         }
1184
1185         ptlrpc_request_set_replen(req);
1186         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1187         lustre_msg_add_flags(req->rq_reqmsg,
1188                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1189         if (AT_OFF)
1190                 req->rq_timeout *= 3;
1191         req->rq_interpret_reply = completed_replay_interpret;
1192
1193         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1194         RETURN(0);
1195 }
1196
1197 #ifdef __KERNEL__
1198 /**
1199  * In kernel code all import invalidation happens in its own
1200  * separate thread, so that whatever application happened to encounter
1201  * a problem could still be killed or otherwise continue
1202  */
1203 static int ptlrpc_invalidate_import_thread(void *data)
1204 {
1205         struct obd_import *imp = data;
1206
1207         ENTRY;
1208
1209         cfs_daemonize_ctxt("ll_imp_inval");
1210
1211         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1212                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1213                imp->imp_connection->c_remote_uuid.uuid);
1214
1215         ptlrpc_invalidate_import(imp);
1216
1217         if (obd_dump_on_eviction) {
1218                 CERROR("dump the log upon eviction\n");
1219                 libcfs_debug_dumplog();
1220         }
1221
1222         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1223         ptlrpc_import_recovery_state_machine(imp);
1224
1225         class_import_put(imp);
1226         RETURN(0);
1227 }
1228 #endif
1229
1230 /**
1231  * This is the state machine for client-side recovery on import.
1232  *
1233  * Typicaly we have two possibly paths. If we came to server and it is not
1234  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1235  * state and reconnect from scratch.
1236  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1237  * We go through our list of requests to replay and send them to server one by
1238  * one.
1239  * After sending all request from the list we change import state to
1240  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1241  * and also all the locks we don't yet have and wait for server to grant us.
1242  * After that we send a special "replay completed" request and change import
1243  * state to IMP_REPLAY_WAIT.
1244  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1245  * state and resend all requests from sending list.
1246  * After that we promote import to FULL state and send all delayed requests
1247  * and import is fully operational after that.
1248  *
1249  */
1250 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1251 {
1252         int rc = 0;
1253         int inflight;
1254         char *target_start;
1255         int target_len;
1256
1257         ENTRY;
1258         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1259                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1260                           &target_start, &target_len);
1261                 /* Don't care about MGC eviction */
1262                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1263                            LUSTRE_MGC_NAME) != 0) {
1264                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1265                                            "%.*s; in progress operations using "
1266                                            "this service will fail.\n",
1267                                            target_len, target_start);
1268                 }
1269                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1270                        obd2cli_tgt(imp->imp_obd),
1271                        imp->imp_connection->c_remote_uuid.uuid);
1272                 /* reset vbr_failed flag upon eviction */
1273                 cfs_spin_lock(&imp->imp_lock);
1274                 imp->imp_vbr_failed = 0;
1275                 cfs_spin_unlock(&imp->imp_lock);
1276
1277 #ifdef __KERNEL__
1278                 /* bug 17802:  XXX client_disconnect_export vs connect request
1279                  * race. if client will evicted at this time, we start
1280                  * invalidate thread without reference to import and import can
1281                  * be freed at same time. */
1282                 class_import_get(imp);
1283                 rc = cfs_create_thread(ptlrpc_invalidate_import_thread, imp,
1284                                        CFS_DAEMON_FLAGS);
1285                 if (rc < 0) {
1286                         class_import_put(imp);
1287                         CERROR("error starting invalidate thread: %d\n", rc);
1288                 } else {
1289                         rc = 0;
1290                 }
1291                 RETURN(rc);
1292 #else
1293                 ptlrpc_invalidate_import(imp);
1294
1295                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1296 #endif
1297         }
1298
1299         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1300                 CDEBUG(D_HA, "replay requested by %s\n",
1301                        obd2cli_tgt(imp->imp_obd));
1302                 rc = ptlrpc_replay_next(imp, &inflight);
1303                 if (inflight == 0 &&
1304                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1305                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1306                         rc = ldlm_replay_locks(imp);
1307                         if (rc)
1308                                 GOTO(out, rc);
1309                 }
1310                 rc = 0;
1311         }
1312
1313         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1314                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1315                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1316                         rc = signal_completed_replay(imp);
1317                         if (rc)
1318                                 GOTO(out, rc);
1319                 }
1320
1321         }
1322
1323         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1324                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1325                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1326                 }
1327         }
1328
1329         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1330                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1331                        obd2cli_tgt(imp->imp_obd),
1332                        imp->imp_connection->c_remote_uuid.uuid);
1333
1334                 rc = ptlrpc_resend(imp);
1335                 if (rc)
1336                         GOTO(out, rc);
1337                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1338                 ptlrpc_activate_import(imp);
1339
1340                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1341                           &target_start, &target_len);
1342                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1343                               "using nid %s.\n", imp->imp_obd->obd_name,
1344                               target_len, target_start,
1345                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1346         }
1347
1348         if (imp->imp_state == LUSTRE_IMP_FULL) {
1349                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1350                 ptlrpc_wake_delayed(imp);
1351         }
1352
1353 out:
1354         RETURN(rc);
1355 }
1356
1357 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1358 {
1359         struct ptlrpc_request *req;
1360         int rq_opc, rc = 0;
1361         int nowait = imp->imp_obd->obd_force;
1362         ENTRY;
1363
1364         if (nowait)
1365                 GOTO(set_state, rc);
1366
1367         switch (imp->imp_connect_op) {
1368         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1369         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1370         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1371         default:
1372                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1373                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1374                 RETURN(-EINVAL);
1375         }
1376
1377         if (ptlrpc_import_in_recovery(imp)) {
1378                 struct l_wait_info lwi;
1379                 cfs_duration_t timeout;
1380
1381
1382                 if (AT_OFF) {
1383                         if (imp->imp_server_timeout)
1384                                 timeout = cfs_time_seconds(obd_timeout / 2);
1385                         else
1386                                 timeout = cfs_time_seconds(obd_timeout);
1387                 } else {
1388                         int idx = import_at_get_index(imp,
1389                                 imp->imp_client->cli_request_portal);
1390                         timeout = cfs_time_seconds(
1391                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1392                 }
1393
1394                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1395                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1396                 rc = l_wait_event(imp->imp_recovery_waitq,
1397                                   !ptlrpc_import_in_recovery(imp), &lwi);
1398
1399         }
1400
1401         cfs_spin_lock(&imp->imp_lock);
1402         if (imp->imp_state != LUSTRE_IMP_FULL)
1403                 GOTO(out, 0);
1404
1405         cfs_spin_unlock(&imp->imp_lock);
1406
1407         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1408                                         LUSTRE_OBD_VERSION, rq_opc);
1409         if (req) {
1410                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1411                  * it fails.  We can get through the above with a down server
1412                  * if the client doesn't know the server is gone yet. */
1413                 req->rq_no_resend = 1;
1414
1415 #ifndef CRAY_XT3
1416                 /* We want client umounts to happen quickly, no matter the
1417                    server state... */
1418                 req->rq_timeout = min_t(int, req->rq_timeout,
1419                                         INITIAL_CONNECT_TIMEOUT);
1420 #else
1421                 /* ... but we always want liblustre clients to nicely
1422                    disconnect, so only use the adaptive value. */
1423                 if (AT_OFF)
1424                         req->rq_timeout = obd_timeout / 3;
1425 #endif
1426
1427                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1428                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1429                 ptlrpc_request_set_replen(req);
1430                 rc = ptlrpc_queue_wait(req);
1431                 ptlrpc_req_finished(req);
1432         }
1433
1434 set_state:
1435         cfs_spin_lock(&imp->imp_lock);
1436 out:
1437         if (noclose)
1438                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1439         else
1440                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1441         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1442         cfs_spin_unlock(&imp->imp_lock);
1443
1444         RETURN(rc);
1445 }
1446
1447 void ptlrpc_cleanup_imp(struct obd_import *imp)
1448 {
1449         ENTRY;
1450
1451         cfs_spin_lock(&imp->imp_lock);
1452         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1453         imp->imp_generation++;
1454         cfs_spin_unlock(&imp->imp_lock);
1455         ptlrpc_abort_inflight(imp);
1456
1457         EXIT;
1458 }
1459
1460 /* Adaptive Timeout utils */
1461 extern unsigned int at_min, at_max, at_history;
1462
1463 /* Bin into timeslices using AT_BINS bins.
1464    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1465    but still smoothing out a return to normalcy from a slow response.
1466    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1467 int at_measured(struct adaptive_timeout *at, unsigned int val)
1468 {
1469         unsigned int old = at->at_current;
1470         time_t now = cfs_time_current_sec();
1471         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1472
1473         LASSERT(at);
1474         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1475                val, at, now - at->at_binstart, at->at_current,
1476                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1477
1478         if (val == 0)
1479                 /* 0's don't count, because we never want our timeout to
1480                    drop to 0, and because 0 could mean an error */
1481                 return 0;
1482
1483         cfs_spin_lock(&at->at_lock);
1484
1485         if (unlikely(at->at_binstart == 0)) {
1486                 /* Special case to remove default from history */
1487                 at->at_current = val;
1488                 at->at_worst_ever = val;
1489                 at->at_worst_time = now;
1490                 at->at_hist[0] = val;
1491                 at->at_binstart = now;
1492         } else if (now - at->at_binstart < binlimit ) {
1493                 /* in bin 0 */
1494                 at->at_hist[0] = max(val, at->at_hist[0]);
1495                 at->at_current = max(val, at->at_current);
1496         } else {
1497                 int i, shift;
1498                 unsigned int maxv = val;
1499                 /* move bins over */
1500                 shift = (now - at->at_binstart) / binlimit;
1501                 LASSERT(shift > 0);
1502                 for(i = AT_BINS - 1; i >= 0; i--) {
1503                         if (i >= shift) {
1504                                 at->at_hist[i] = at->at_hist[i - shift];
1505                                 maxv = max(maxv, at->at_hist[i]);
1506                         } else {
1507                                 at->at_hist[i] = 0;
1508                         }
1509                 }
1510                 at->at_hist[0] = val;
1511                 at->at_current = maxv;
1512                 at->at_binstart += shift * binlimit;
1513         }
1514
1515         if (at->at_current > at->at_worst_ever) {
1516                 at->at_worst_ever = at->at_current;
1517                 at->at_worst_time = now;
1518         }
1519
1520         if (at->at_flags & AT_FLG_NOHIST)
1521                 /* Only keep last reported val; keeping the rest of the history
1522                    for proc only */
1523                 at->at_current = val;
1524
1525         if (at_max > 0)
1526                 at->at_current =  min(at->at_current, at_max);
1527         at->at_current =  max(at->at_current, at_min);
1528
1529         if (at->at_current != old)
1530                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1531                        "(val=%u) hist %u %u %u %u\n", at,
1532                        old, at->at_current, at->at_current - old, val,
1533                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1534                        at->at_hist[3]);
1535
1536         /* if we changed, report the old value */
1537         old = (at->at_current != old) ? old : 0;
1538
1539         cfs_spin_unlock(&at->at_lock);
1540         return old;
1541 }
1542
1543 /* Find the imp_at index for a given portal; assign if space available */
1544 int import_at_get_index(struct obd_import *imp, int portal)
1545 {
1546         struct imp_at *at = &imp->imp_at;
1547         int i;
1548
1549         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1550                 if (at->iat_portal[i] == portal)
1551                         return i;
1552                 if (at->iat_portal[i] == 0)
1553                         /* unused */
1554                         break;
1555         }
1556
1557         /* Not found in list, add it under a lock */
1558         cfs_spin_lock(&imp->imp_lock);
1559
1560         /* Check unused under lock */
1561         for (; i < IMP_AT_MAX_PORTALS; i++) {
1562                 if (at->iat_portal[i] == portal)
1563                         goto out;
1564                 if (at->iat_portal[i] == 0)
1565                         /* unused */
1566                         break;
1567         }
1568
1569         /* Not enough portals? */
1570         LASSERT(i < IMP_AT_MAX_PORTALS);
1571
1572         at->iat_portal[i] = portal;
1573 out:
1574         cfs_spin_unlock(&imp->imp_lock);
1575         return i;
1576 }