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);
 451         return 0;
 452 }
 453
 454 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 455
 456 /**
 457  * Connection on import \a imp is changed to another one (if more than one is
 458  * present). We typically chose connection that we have not tried to connect to
 459  * the longest
 460  */
 461 static int import_select_connection(struct obd_import *imp)
 462 {
 463         struct obd_import_conn *imp_conn = NULL, *conn;
 464         struct obd_export *dlmexp;
 465         int tried_all = 1;
 466         ENTRY;
 467
 468         cfs_spin_lock(&imp->imp_lock);
 469
 470         if (cfs_list_empty(&imp->imp_conn_list)) {
 471                 CERROR("%s: no connections available\n",
 472                         imp->imp_obd->obd_name);
 473                 cfs_spin_unlock(&imp->imp_lock);
 474                 RETURN(-EINVAL);
 475         }
 476
 477         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 478                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 479                        imp->imp_obd->obd_name,
 480                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 481                        conn->oic_last_attempt);
 482                 /* Don't thrash connections */
 483                 if (cfs_time_before_64(cfs_time_current_64(),
 484                                      conn->oic_last_attempt +
 485                                      cfs_time_seconds(CONNECTION_SWITCH_MIN))) {
 486                         continue;
 487                 }
 488
 489                 /* If we have not tried this connection since
 490                    the last successful attempt, go with this one */
 491                 if ((conn->oic_last_attempt == 0) ||
 492                     cfs_time_beforeq_64(conn->oic_last_attempt,
 493                                        imp->imp_last_success_conn)) {
 494                         imp_conn = conn;
 495                         tried_all = 0;
 496                         break;
 497                 }
 498
 499                 /* If all of the connections have already been tried
 500                    since the last successful connection; just choose the
 501                    least recently used */
 502                 if (!imp_conn)
 503                         imp_conn = conn;
 504                 else if (cfs_time_before_64(conn->oic_last_attempt,
 505                                             imp_conn->oic_last_attempt))
 506                         imp_conn = conn;
 507         }
 508
 509         /* if not found, simply choose the current one */
 510         if (!imp_conn || imp->imp_force_reconnect) {
 511                 LASSERT(imp->imp_conn_current);
 512                 imp_conn = imp->imp_conn_current;
 513                 tried_all = 0;
 514         }
 515         LASSERT(imp_conn->oic_conn);
 516
 517         /* If we've tried everything, and we're back to the beginning of the
 518            list, increase our timeout and try again. It will be reset when
 519            we do finally connect. (FIXME: really we should wait for all network
 520            state associated with the last connection attempt to drain before
 521            trying to reconnect on it.) */
 522         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 523                 if (at_get(&imp->imp_at.iat_net_latency) <
 524                     CONNECTION_SWITCH_MAX) {
 525                         at_measured(&imp->imp_at.iat_net_latency,
 526                                     at_get(&imp->imp_at.iat_net_latency) +
 527                                     CONNECTION_SWITCH_INC);
 528                 }
 529                 LASSERT(imp_conn->oic_last_attempt);
 530                 CWARN("%s: tried all connections, increasing latency to %ds\n",
 531                       imp->imp_obd->obd_name,
 532                       at_get(&imp->imp_at.iat_net_latency));
 533         }
 534
 535         imp_conn->oic_last_attempt = cfs_time_current_64();
 536
 537         /* switch connection, don't mind if it's same as the current one */
 538         if (imp->imp_connection)
 539                 ptlrpc_connection_put(imp->imp_connection);
 540         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 541
 542         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 543         LASSERT(dlmexp != NULL);
 544         if (dlmexp->exp_connection)
 545                 ptlrpc_connection_put(dlmexp->exp_connection);
 546         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 547         class_export_put(dlmexp);
 548
 549         if (imp->imp_conn_current != imp_conn) {
 550                 if (imp->imp_conn_current)
 551                         CDEBUG(D_HA, "Changing connection for %s to %s/%s\n",
 552                                imp->imp_obd->obd_name, imp_conn->oic_uuid.uuid,
 553                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 554                 imp->imp_conn_current = imp_conn;
 555         }
 556
 557         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 558                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 559                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 560
 561         cfs_spin_unlock(&imp->imp_lock);
 562
 563         RETURN(0);
 564 }
 565
 566 /*
 567  * must be called under imp_lock
 568  */
 569 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 570 {
 571         struct ptlrpc_request *req;
 572         cfs_list_t *tmp;
 573
 574         if (cfs_list_empty(&imp->imp_replay_list))
 575                 return 0;
 576         tmp = imp->imp_replay_list.next;
 577         req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 578         *transno = req->rq_transno;
 579         if (req->rq_transno == 0) {
 580                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 581                 LBUG();
 582         }
 583
 584         return 1;
 585 }
 586
 587 /**
 588  * Attempt to (re)connect import \a imp. This includes all preparations,
 589  * initializing CONNECT RPC request and passing it to ptlrpcd for
 590  * actual sending.
 591  * Returns 0 on success or error code.
 592  */
 593 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 594 {
 595         struct obd_device *obd = imp->imp_obd;
 596         int initial_connect = 0;
 597         int set_transno = 0;
 598         __u64 committed_before_reconnect = 0;
 599         struct ptlrpc_request *request;
 600         char *bufs[] = { NULL,
 601                          obd2cli_tgt(imp->imp_obd),
 602                          obd->obd_uuid.uuid,
 603                          (char *)&imp->imp_dlm_handle,
 604                          (char *)&imp->imp_connect_data };
 605         struct ptlrpc_connect_async_args *aa;
 606         int rc;
 607         ENTRY;
 608
 609         cfs_spin_lock(&imp->imp_lock);
 610         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 611                 cfs_spin_unlock(&imp->imp_lock);
 612                 CERROR("can't connect to a closed import\n");
 613                 RETURN(-EINVAL);
 614         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 615                 cfs_spin_unlock(&imp->imp_lock);
 616                 CERROR("already connected\n");
 617                 RETURN(0);
 618         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 619                 cfs_spin_unlock(&imp->imp_lock);
 620                 CERROR("already connecting\n");
 621                 RETURN(-EALREADY);
 622         }
 623
 624         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 625
 626         imp->imp_conn_cnt++;
 627         imp->imp_resend_replay = 0;
 628
 629         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 630                 initial_connect = 1;
 631         else
 632                 committed_before_reconnect = imp->imp_peer_committed_transno;
 633
 634         set_transno = ptlrpc_first_transno(imp,
 635                                            &imp->imp_connect_data.ocd_transno);
 636         cfs_spin_unlock(&imp->imp_lock);
 637
 638         if (new_uuid) {
 639                 struct obd_uuid uuid;
 640
 641                 obd_str2uuid(&uuid, new_uuid);
 642                 rc = import_set_conn_priority(imp, &uuid);
 643                 if (rc)
 644                         GOTO(out, rc);
 645         }
 646
 647         rc = import_select_connection(imp);
 648         if (rc)
 649                 GOTO(out, rc);
 650
 651         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 652         if (rc)
 653                 GOTO(out, rc);
 654
 655         /* Reset connect flags to the originally requested flags, in case
 656          * the server is updated on-the-fly we will get the new features. */
 657         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 658         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 659         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 660
 661         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 662                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 663         if (rc)
 664                 GOTO(out, rc);
 665
 666         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 667         if (request == NULL)
 668                 GOTO(out, rc = -ENOMEM);
 669
 670         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 671                                       imp->imp_connect_op, bufs, NULL);
 672         if (rc) {
 673                 ptlrpc_request_free(request);
 674                 GOTO(out, rc);
 675         }
 676
 677         /* Report the rpc service time to the server so that it knows how long
 678          * to wait for clients to join recovery */
 679         lustre_msg_set_service_time(request->rq_reqmsg,
 680                                     at_timeout2est(request->rq_timeout));
 681
 682         /* The amount of time we give the server to process the connect req.
 683          * import_select_connection will increase the net latency on
 684          * repeated reconnect attempts to cover slow networks.
 685          * We override/ignore the server rpc completion estimate here,
 686          * which may be large if this is a reconnect attempt */
 687         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 688         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 689
 690 #ifndef __KERNEL__
 691         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 692 #endif
 693         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 694
 695         request->rq_no_resend = request->rq_no_delay = 1;
 696         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 697         /* Allow a slightly larger reply for future growth compatibility */
 698         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 699                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 700         ptlrpc_request_set_replen(request);
 701         request->rq_interpret_reply = ptlrpc_connect_interpret;
 702
 703         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 704         aa = ptlrpc_req_async_args(request);
 705         memset(aa, 0, sizeof *aa);
 706
 707         aa->pcaa_peer_committed = committed_before_reconnect;
 708         aa->pcaa_initial_connect = initial_connect;
 709
 710         if (aa->pcaa_initial_connect) {
 711                 cfs_spin_lock(&imp->imp_lock);
 712                 imp->imp_replayable = 1;
 713                 cfs_spin_unlock(&imp->imp_lock);
 714                 lustre_msg_add_op_flags(request->rq_reqmsg,
 715                                         MSG_CONNECT_INITIAL);
 716         }
 717
 718         if (set_transno)
 719                 lustre_msg_add_op_flags(request->rq_reqmsg,
 720                                         MSG_CONNECT_TRANSNO);
 721
 722         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 723                   request->rq_timeout);
 724         ptlrpcd_add_req(request, PSCOPE_OTHER);
 725         rc = 0;
 726 out:
 727         if (rc != 0) {
 728                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 729         }
 730
 731         RETURN(rc);
 732 }
 733 EXPORT_SYMBOL(ptlrpc_connect_import);
 734
 735 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 736 {
 737 #ifdef __KERNEL__
 738         struct obd_import_conn *imp_conn;
 739 #endif
 740         int wake_pinger = 0;
 741
 742         ENTRY;
 743
 744         cfs_spin_lock(&imp->imp_lock);
 745         if (cfs_list_empty(&imp->imp_conn_list))
 746                 GOTO(unlock, 0);
 747
 748 #ifdef __KERNEL__
 749         imp_conn = cfs_list_entry(imp->imp_conn_list.prev,
 750                                   struct obd_import_conn,
 751                                   oic_item);
 752
 753         /* XXX: When the failover node is the primary node, it is possible
 754          * to have two identical connections in imp_conn_list. We must
 755          * compare not conn's pointers but NIDs, otherwise we can defeat
 756          * connection throttling. (See bug 14774.) */
 757         if (imp->imp_conn_current->oic_conn->c_peer.nid !=
 758                                 imp_conn->oic_conn->c_peer.nid) {
 759                 ptlrpc_ping_import_soon(imp);
 760                 wake_pinger = 1;
 761         }
 762 #else
 763         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 764         wake_pinger = 1;
 765 #endif
 766
 767  unlock:
 768         cfs_spin_unlock(&imp->imp_lock);
 769
 770         if (wake_pinger)
 771                 ptlrpc_pinger_wake_up();
 772
 773         EXIT;
 774 }
 775
 776 static int ptlrpc_busy_reconnect(int rc)
 777 {
 778         return (rc == -EBUSY) || (rc == -EAGAIN);
 779 }
 780
 781 /**
 782  * interpret_reply callback for connect RPCs.
 783  * Looks into returned status of connect operation and decides
 784  * what to do with the import - i.e enter recovery, promote it to
 785  * full state for normal operations of disconnect it due to an error.
 786  */
 787 static int ptlrpc_connect_interpret(const struct lu_env *env,
 788                                     struct ptlrpc_request *request,
 789                                     void *data, int rc)
 790 {
 791         struct ptlrpc_connect_async_args *aa = data;
 792         struct obd_import *imp = request->rq_import;
 793         struct client_obd *cli = &imp->imp_obd->u.cli;
 794         struct lustre_handle old_hdl;
 795         __u64 old_connect_flags;
 796         int msg_flags;
 797         ENTRY;
 798
 799         cfs_spin_lock(&imp->imp_lock);
 800         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 801                 cfs_spin_unlock(&imp->imp_lock);
 802                 RETURN(0);
 803         }
 804
 805         if (rc) {
 806                 /* if this reconnect to busy export - not need select new target
 807                  * for connecting*/
 808                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 809                 cfs_spin_unlock(&imp->imp_lock);
 810                 GOTO(out, rc);
 811         }
 812
 813         LASSERT(imp->imp_conn_current);
 814
 815         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 816
 817         /* All imports are pingable */
 818         imp->imp_pingable = 1;
 819         imp->imp_force_reconnect = 0;
 820
 821         if (aa->pcaa_initial_connect) {
 822                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 823                         imp->imp_replayable = 1;
 824                         cfs_spin_unlock(&imp->imp_lock);
 825                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 826                                obd2cli_tgt(imp->imp_obd));
 827                 } else {
 828                         imp->imp_replayable = 0;
 829                         cfs_spin_unlock(&imp->imp_lock);
 830                 }
 831
 832                 /* if applies, adjust the imp->imp_msg_magic here
 833                  * according to reply flags */
 834
 835                 imp->imp_remote_handle =
 836                                 *lustre_msg_get_handle(request->rq_repmsg);
 837
 838                 /* Initial connects are allowed for clients with non-random
 839                  * uuids when servers are in recovery.  Simply signal the
 840                  * servers replay is complete and wait in REPLAY_WAIT. */
 841                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 842                         CDEBUG(D_HA, "connect to %s during recovery\n",
 843                                obd2cli_tgt(imp->imp_obd));
 844                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 845                 } else {
 846                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 847                         ptlrpc_activate_import(imp);
 848                 }
 849
 850                 GOTO(finish, rc = 0);
 851         } else {
 852                 cfs_spin_unlock(&imp->imp_lock);
 853         }
 854
 855         /* Determine what recovery state to move the import to. */
 856         if (MSG_CONNECT_RECONNECT & msg_flags) {
 857                 memset(&old_hdl, 0, sizeof(old_hdl));
 858                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 859                             sizeof (old_hdl))) {
 860                         CERROR("%s@%s didn't like our handle "LPX64
 861                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 862                                imp->imp_connection->c_remote_uuid.uuid,
 863                                imp->imp_dlm_handle.cookie);
 864                         GOTO(out, rc = -ENOTCONN);
 865                 }
 866
 867                 if (memcmp(&imp->imp_remote_handle,
 868                            lustre_msg_get_handle(request->rq_repmsg),
 869                            sizeof(imp->imp_remote_handle))) {
 870                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 871                                 D_HA : D_WARNING;
 872
 873                         /* Bug 16611/14775: if server handle have changed,
 874                          * that means some sort of disconnection happened.
 875                          * If the server is not in recovery, that also means it
 876                          * already erased all of our state because of previous
 877                          * eviction. If it is in recovery - we are safe to
 878                          * participate since we can reestablish all of our state
 879                          * with server again */
 880                         CDEBUG(level,"%s@%s changed server handle from "
 881                                      LPX64" to "LPX64"%s\n",
 882                                      obd2cli_tgt(imp->imp_obd),
 883                                      imp->imp_connection->c_remote_uuid.uuid,
 884                                      imp->imp_remote_handle.cookie,
 885                                      lustre_msg_get_handle(request->rq_repmsg)->
 886                                                                         cookie,
 887                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
 888                                          " but is still in recovery" : "");
 889
 890                         imp->imp_remote_handle =
 891                                      *lustre_msg_get_handle(request->rq_repmsg);
 892
 893                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 894                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 895                                 GOTO(finish, rc = 0);
 896                         }
 897
 898                 } else {
 899                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 900                                obd2cli_tgt(imp->imp_obd),
 901                                imp->imp_connection->c_remote_uuid.uuid);
 902                 }
 903
 904                 if (imp->imp_invalid) {
 905                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 906                                "marking evicted\n", imp->imp_obd->obd_name);
 907                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 908                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 909                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 910                                imp->imp_obd->obd_name,
 911                                obd2cli_tgt(imp->imp_obd));
 912
 913                         cfs_spin_lock(&imp->imp_lock);
 914                         imp->imp_resend_replay = 1;
 915                         cfs_spin_unlock(&imp->imp_lock);
 916
 917                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 918                 } else {
 919                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 920                 }
 921         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 922                 LASSERT(imp->imp_replayable);
 923                 imp->imp_remote_handle =
 924                                 *lustre_msg_get_handle(request->rq_repmsg);
 925                 imp->imp_last_replay_transno = 0;
 926                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 927         } else {
 928                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 929                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 930                 imp->imp_remote_handle =
 931                                 *lustre_msg_get_handle(request->rq_repmsg);
 932                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 933         }
 934
 935         /* Sanity checks for a reconnected import. */
 936         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 937                 CERROR("imp_replayable flag does not match server "
 938                        "after reconnect. We should LBUG right here.\n");
 939         }
 940
 941         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 942             lustre_msg_get_last_committed(request->rq_repmsg) <
 943             aa->pcaa_peer_committed) {
 944                 CERROR("%s went back in time (transno "LPD64
 945                        " was previously committed, server now claims "LPD64
 946                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 947                        "id=9646\n",
 948                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 949                        lustre_msg_get_last_committed(request->rq_repmsg));
 950         }
 951
 952 finish:
 953         rc = ptlrpc_import_recovery_state_machine(imp);
 954         if (rc != 0) {
 955                 if (rc == -ENOTCONN) {
 956                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 957                                "invalidating and reconnecting\n",
 958                                obd2cli_tgt(imp->imp_obd),
 959                                imp->imp_connection->c_remote_uuid.uuid);
 960                         ptlrpc_connect_import(imp, NULL);
 961                         RETURN(0);
 962                 }
 963         } else {
 964                 struct obd_connect_data *ocd;
 965                 struct obd_export *exp;
 966                 int ret;
 967                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 968                                            RCL_SERVER);
 969                 /* server replied obd_connect_data is always bigger */
 970                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 971                                                    &RMF_CONNECT_DATA, ret);
 972
 973                 cfs_spin_lock(&imp->imp_lock);
 974                 cfs_list_del(&imp->imp_conn_current->oic_item);
 975                 cfs_list_add(&imp->imp_conn_current->oic_item,
 976                              &imp->imp_conn_list);
 977                 imp->imp_last_success_conn =
 978                         imp->imp_conn_current->oic_last_attempt;
 979
 980                 if (ocd == NULL) {
 981                         cfs_spin_unlock(&imp->imp_lock);
 982                         CERROR("Wrong connect data from server\n");
 983                         rc = -EPROTO;
 984                         GOTO(out, rc);
 985                 }
 986
 987                 imp->imp_connect_data = *ocd;
 988
 989                 exp = class_conn2export(&imp->imp_dlm_handle);
 990                 cfs_spin_unlock(&imp->imp_lock);
 991
 992                 /* check that server granted subset of flags we asked for. */
 993                 LASSERTF((ocd->ocd_connect_flags &
 994                           imp->imp_connect_flags_orig) ==
 995                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 996                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 997
 998                 if (!exp) {
 999                         /* This could happen if export is cleaned during the
1000                            connect attempt */
1001                         CERROR("Missing export for %s\n",
1002                                imp->imp_obd->obd_name);
1003                         GOTO(out, rc = -ENODEV);
1004                 }
1005                 old_connect_flags = exp->exp_connect_flags;
1006                 exp->exp_connect_flags = ocd->ocd_connect_flags;
1007                 imp->imp_obd->obd_self_export->exp_connect_flags =
1008                                                         ocd->ocd_connect_flags;
1009                 class_export_put(exp);
1010
1011                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1012
1013                 if (!ocd->ocd_ibits_known &&
1014                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
1015                         CERROR("Inodebits aware server returned zero compatible"
1016                                " bits?\n");
1017
1018                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1019                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1020                                         LUSTRE_VERSION_OFFSET_WARN ||
1021                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1022                                         LUSTRE_VERSION_OFFSET_WARN)) {
1023                         /* Sigh, some compilers do not like #ifdef in the middle
1024                            of macro arguments */
1025 #ifdef __KERNEL__
1026                         const char *older =
1027                                 "older. Consider upgrading this client";
1028 #else
1029                         const char *older =
1030                                 "older. Consider recompiling this application";
1031 #endif
1032                         const char *newer = "newer than client version";
1033
1034                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1035                                       "is much %s (%s)\n",
1036                                       obd2cli_tgt(imp->imp_obd),
1037                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1038                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1039                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1040                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1041                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1042                                       newer : older, LUSTRE_VERSION_STRING);
1043                 }
1044
1045                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1046                         /* We sent to the server ocd_cksum_types with bits set
1047                          * for algorithms we understand. The server masked off
1048                          * the checksum types it doesn't support */
1049                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
1050                                 LCONSOLE_WARN("The negotiation of the checksum "
1051                                               "alogrithm to use with server %s "
1052                                               "failed (%x/%x), disabling "
1053                                               "checksums\n",
1054                                               obd2cli_tgt(imp->imp_obd),
1055                                               ocd->ocd_cksum_types,
1056                                               OBD_CKSUM_ALL);
1057                                 cli->cl_checksum = 0;
1058                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1059                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1060                         } else {
1061                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1062
1063                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
1064                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
1065                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
1066                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
1067                                 else
1068                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1069                         }
1070                 } else {
1071                         /* The server does not support OBD_CONNECT_CKSUM.
1072                          * Enforce CRC32 for backward compatibility*/
1073                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1074                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1075                 }
1076
1077                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1078                         cli->cl_max_pages_per_rpc =
1079                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1080                 else if (imp->imp_connect_op == MDS_CONNECT ||
1081                          imp->imp_connect_op == MGS_CONNECT)
1082                         cli->cl_max_pages_per_rpc = 1;
1083
1084                 /* Reset ns_connect_flags only for initial connect. It might be
1085                  * changed in while using FS and if we reset it in reconnect
1086                  * this leads to losing user settings done before such as
1087                  * disable lru_resize, etc. */
1088                 if (old_connect_flags != exp->exp_connect_flags ||
1089                     aa->pcaa_initial_connect) {
1090                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1091                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1092                               ocd->ocd_connect_flags);
1093                         imp->imp_obd->obd_namespace->ns_connect_flags =
1094                                 ocd->ocd_connect_flags;
1095                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1096                                 ocd->ocd_connect_flags;
1097                 }
1098
1099                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1100                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1101                         /* We need a per-message support flag, because
1102                            a. we don't know if the incoming connect reply
1103                               supports AT or not (in reply_in_callback)
1104                               until we unpack it.
1105                            b. failovered server means export and flags are gone
1106                               (in ptlrpc_send_reply).
1107                            Can only be set when we know AT is supported at
1108                            both ends */
1109                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1110                 else
1111                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1112
1113                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1114                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1115                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1116                 else
1117                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1118
1119                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1120                         (cli->cl_max_pages_per_rpc > 0));
1121         }
1122
1123 out:
1124         if (rc != 0) {
1125                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1126                 if (rc == -EACCES) {
1127                         /*
1128                          * Give up trying to reconnect
1129                          * EACCES means client has no permission for connection
1130                          */
1131                         imp->imp_obd->obd_no_recov = 1;
1132                         ptlrpc_deactivate_import(imp);
1133                 }
1134
1135                 if (rc == -EPROTO) {
1136                         struct obd_connect_data *ocd;
1137
1138                         /* reply message might not be ready */
1139                         if (request->rq_repmsg == NULL)
1140                                 RETURN(-EPROTO);
1141
1142                         ocd = req_capsule_server_get(&request->rq_pill,
1143                                                      &RMF_CONNECT_DATA);
1144                         if (ocd &&
1145                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1146                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1147                            /* Actually servers are only supposed to refuse
1148                               connection from liblustre clients, so we should
1149                               never see this from VFS context */
1150                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1151                                         "(%d.%d.%d.%d)"
1152                                         " refused connection from this client "
1153                                         "with an incompatible version (%s).  "
1154                                         "Client must be recompiled\n",
1155                                         obd2cli_tgt(imp->imp_obd),
1156                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1157                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1158                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1159                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1160                                         LUSTRE_VERSION_STRING);
1161                                 ptlrpc_deactivate_import(imp);
1162                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1163                         }
1164                         RETURN(-EPROTO);
1165                 }
1166
1167                 ptlrpc_maybe_ping_import_soon(imp);
1168
1169                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1170                        obd2cli_tgt(imp->imp_obd),
1171                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1172         }
1173
1174         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1175         RETURN(rc);
1176 }
1177
1178 /**
1179  * interpret callback for "completed replay" RPCs.
1180  * \see signal_completed_replay
1181  */
1182 static int completed_replay_interpret(const struct lu_env *env,
1183                                       struct ptlrpc_request *req,
1184                                       void * data, int rc)
1185 {
1186         ENTRY;
1187         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1188         if (req->rq_status == 0 &&
1189             !req->rq_import->imp_vbr_failed) {
1190                 ptlrpc_import_recovery_state_machine(req->rq_import);
1191         } else {
1192                 if (req->rq_import->imp_vbr_failed) {
1193                         CDEBUG(D_WARNING,
1194                                "%s: version recovery fails, reconnecting\n",
1195                                req->rq_import->imp_obd->obd_name);
1196                 } else {
1197                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1198                                      "reconnecting\n",
1199                                req->rq_import->imp_obd->obd_name,
1200                                req->rq_status);
1201                 }
1202                 ptlrpc_connect_import(req->rq_import, NULL);
1203         }
1204
1205         RETURN(0);
1206 }
1207
1208 /**
1209  * Let server know that we have no requests to replay anymore.
1210  * Achieved by just sending a PING request
1211  */
1212 static int signal_completed_replay(struct obd_import *imp)
1213 {
1214         struct ptlrpc_request *req;
1215         ENTRY;
1216
1217         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1218                 RETURN(0);
1219
1220         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1221         cfs_atomic_inc(&imp->imp_replay_inflight);
1222
1223         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1224                                         OBD_PING);
1225         if (req == NULL) {
1226                 cfs_atomic_dec(&imp->imp_replay_inflight);
1227                 RETURN(-ENOMEM);
1228         }
1229
1230         ptlrpc_request_set_replen(req);
1231         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1232         lustre_msg_add_flags(req->rq_reqmsg,
1233                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1234         if (AT_OFF)
1235                 req->rq_timeout *= 3;
1236         req->rq_interpret_reply = completed_replay_interpret;
1237
1238         ptlrpcd_add_req(req, PSCOPE_OTHER);
1239         RETURN(0);
1240 }
1241
1242 #ifdef __KERNEL__
1243 /**
1244  * In kernel code all import invalidation happens in its own
1245  * separate thread, so that whatever application happened to encounter
1246  * a problem could still be killed or otherwise continue
1247  */
1248 static int ptlrpc_invalidate_import_thread(void *data)
1249 {
1250         struct obd_import *imp = data;
1251
1252         ENTRY;
1253
1254         cfs_daemonize_ctxt("ll_imp_inval");
1255
1256         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1257                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1258                imp->imp_connection->c_remote_uuid.uuid);
1259
1260         ptlrpc_invalidate_import(imp);
1261
1262         if (obd_dump_on_eviction) {
1263                 CERROR("dump the log upon eviction\n");
1264                 libcfs_debug_dumplog();
1265         }
1266
1267         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1268         ptlrpc_import_recovery_state_machine(imp);
1269
1270         class_import_put(imp);
1271         RETURN(0);
1272 }
1273 #endif
1274
1275 /**
1276  * This is the state machine for client-side recovery on import.
1277  *
1278  * Typicaly we have two possibly paths. If we came to server and it is not
1279  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1280  * state and reconnect from scratch.
1281  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1282  * We go through our list of requests to replay and send them to server one by
1283  * one.
1284  * After sending all request from the list we change import state to
1285  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1286  * and also all the locks we don't yet have and wait for server to grant us.
1287  * After that we send a special "replay completed" request and change import
1288  * state to IMP_REPLAY_WAIT.
1289  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1290  * state and resend all requests from sending list.
1291  * After that we promote import to FULL state and send all delayed requests
1292  * and import is fully operational after that.
1293  *
1294  */
1295 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1296 {
1297         int rc = 0;
1298         int inflight;
1299         char *target_start;
1300         int target_len;
1301
1302         ENTRY;
1303         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1304                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1305                           &target_start, &target_len);
1306                 /* Don't care about MGC eviction */
1307                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1308                            LUSTRE_MGC_NAME) != 0) {
1309                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1310                                            "%.*s; in progress operations using "
1311                                            "this service will fail.\n",
1312                                            target_len, target_start);
1313                 }
1314                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1315                        obd2cli_tgt(imp->imp_obd),
1316                        imp->imp_connection->c_remote_uuid.uuid);
1317                 /* reset vbr_failed flag upon eviction */
1318                 cfs_spin_lock(&imp->imp_lock);
1319                 imp->imp_vbr_failed = 0;
1320                 cfs_spin_unlock(&imp->imp_lock);
1321
1322 #ifdef __KERNEL__
1323                 /* bug 17802:  XXX client_disconnect_export vs connect request
1324                  * race. if client will evicted at this time, we start
1325                  * invalidate thread without reference to import and import can
1326                  * be freed at same time. */
1327                 class_import_get(imp);
1328                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1329                                        CLONE_VM | CLONE_FILES);
1330                 if (rc < 0) {
1331                         class_import_put(imp);
1332                         CERROR("error starting invalidate thread: %d\n", rc);
1333                 } else {
1334                         rc = 0;
1335                 }
1336                 RETURN(rc);
1337 #else
1338                 ptlrpc_invalidate_import(imp);
1339
1340                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1341 #endif
1342         }
1343
1344         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1345                 CDEBUG(D_HA, "replay requested by %s\n",
1346                        obd2cli_tgt(imp->imp_obd));
1347                 rc = ptlrpc_replay_next(imp, &inflight);
1348                 if (inflight == 0 &&
1349                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1350                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1351                         rc = ldlm_replay_locks(imp);
1352                         if (rc)
1353                                 GOTO(out, rc);
1354                 }
1355                 rc = 0;
1356         }
1357
1358         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1359                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1360                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1361                         rc = signal_completed_replay(imp);
1362                         if (rc)
1363                                 GOTO(out, rc);
1364                 }
1365
1366         }
1367
1368         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1369                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1370                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1371                 }
1372         }
1373
1374         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1375                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1376                        obd2cli_tgt(imp->imp_obd),
1377                        imp->imp_connection->c_remote_uuid.uuid);
1378
1379                 rc = ptlrpc_resend(imp);
1380                 if (rc)
1381                         GOTO(out, rc);
1382                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1383                 ptlrpc_activate_import(imp);
1384
1385                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1386                           &target_start, &target_len);
1387                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1388                               "using nid %s.\n", imp->imp_obd->obd_name,
1389                               target_len, target_start,
1390                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1391         }
1392
1393         if (imp->imp_state == LUSTRE_IMP_FULL) {
1394                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1395                 ptlrpc_wake_delayed(imp);
1396         }
1397
1398 out:
1399         RETURN(rc);
1400 }
1401
1402 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1403 {
1404         struct ptlrpc_request *req;
1405         int rq_opc, rc = 0;
1406         int nowait = imp->imp_obd->obd_force;
1407         ENTRY;
1408
1409         if (nowait)
1410                 GOTO(set_state, rc);
1411
1412         switch (imp->imp_connect_op) {
1413         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1414         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1415         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1416         default:
1417                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1418                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1419                 RETURN(-EINVAL);
1420         }
1421
1422         if (ptlrpc_import_in_recovery(imp)) {
1423                 struct l_wait_info lwi;
1424                 cfs_duration_t timeout;
1425
1426
1427                 if (AT_OFF) {
1428                         if (imp->imp_server_timeout)
1429                                 timeout = cfs_time_seconds(obd_timeout / 2);
1430                         else
1431                                 timeout = cfs_time_seconds(obd_timeout);
1432                 } else {
1433                         int idx = import_at_get_index(imp,
1434                                 imp->imp_client->cli_request_portal);
1435                         timeout = cfs_time_seconds(
1436                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1437                 }
1438
1439                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1440                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1441                 rc = l_wait_event(imp->imp_recovery_waitq,
1442                                   !ptlrpc_import_in_recovery(imp), &lwi);
1443
1444         }
1445
1446         cfs_spin_lock(&imp->imp_lock);
1447         if (imp->imp_state != LUSTRE_IMP_FULL)
1448                 GOTO(out, 0);
1449
1450         cfs_spin_unlock(&imp->imp_lock);
1451
1452         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1453                                         LUSTRE_OBD_VERSION, rq_opc);
1454         if (req) {
1455                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1456                  * it fails.  We can get through the above with a down server
1457                  * if the client doesn't know the server is gone yet. */
1458                 req->rq_no_resend = 1;
1459
1460 #ifndef CRAY_XT3
1461                 /* We want client umounts to happen quickly, no matter the
1462                    server state... */
1463                 req->rq_timeout = min_t(int, req->rq_timeout,
1464                                         INITIAL_CONNECT_TIMEOUT);
1465 #else
1466                 /* ... but we always want liblustre clients to nicely
1467                    disconnect, so only use the adaptive value. */
1468                 if (AT_OFF)
1469                         req->rq_timeout = obd_timeout / 3;
1470 #endif
1471
1472                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1473                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1474                 ptlrpc_request_set_replen(req);
1475                 rc = ptlrpc_queue_wait(req);
1476                 ptlrpc_req_finished(req);
1477         }
1478
1479 set_state:
1480         cfs_spin_lock(&imp->imp_lock);
1481 out:
1482         if (noclose)
1483                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1484         else
1485                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1486         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1487         cfs_spin_unlock(&imp->imp_lock);
1488
1489         RETURN(rc);
1490 }
1491
1492 void ptlrpc_cleanup_imp(struct obd_import *imp)
1493 {
1494         ENTRY;
1495
1496         cfs_spin_lock(&imp->imp_lock);
1497         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1498         imp->imp_generation++;
1499         cfs_spin_unlock(&imp->imp_lock);
1500         ptlrpc_abort_inflight(imp);
1501
1502         EXIT;
1503 }
1504
1505 /* Adaptive Timeout utils */
1506 extern unsigned int at_min, at_max, at_history;
1507
1508 /* Bin into timeslices using AT_BINS bins.
1509    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1510    but still smoothing out a return to normalcy from a slow response.
1511    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1512 int at_measured(struct adaptive_timeout *at, unsigned int val)
1513 {
1514         unsigned int old = at->at_current;
1515         time_t now = cfs_time_current_sec();
1516         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1517
1518         LASSERT(at);
1519         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1520                val, at, now - at->at_binstart, at->at_current,
1521                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1522
1523         if (val == 0)
1524                 /* 0's don't count, because we never want our timeout to
1525                    drop to 0, and because 0 could mean an error */
1526                 return 0;
1527
1528         cfs_spin_lock(&at->at_lock);
1529
1530         if (unlikely(at->at_binstart == 0)) {
1531                 /* Special case to remove default from history */
1532                 at->at_current = val;
1533                 at->at_worst_ever = val;
1534                 at->at_worst_time = now;
1535                 at->at_hist[0] = val;
1536                 at->at_binstart = now;
1537         } else if (now - at->at_binstart < binlimit ) {
1538                 /* in bin 0 */
1539                 at->at_hist[0] = max(val, at->at_hist[0]);
1540                 at->at_current = max(val, at->at_current);
1541         } else {
1542                 int i, shift;
1543                 unsigned int maxv = val;
1544                 /* move bins over */
1545                 shift = (now - at->at_binstart) / binlimit;
1546                 LASSERT(shift > 0);
1547                 for(i = AT_BINS - 1; i >= 0; i--) {
1548                         if (i >= shift) {
1549                                 at->at_hist[i] = at->at_hist[i - shift];
1550                                 maxv = max(maxv, at->at_hist[i]);
1551                         } else {
1552                                 at->at_hist[i] = 0;
1553                         }
1554                 }
1555                 at->at_hist[0] = val;
1556                 at->at_current = maxv;
1557                 at->at_binstart += shift * binlimit;
1558         }
1559
1560         if (at->at_current > at->at_worst_ever) {
1561                 at->at_worst_ever = at->at_current;
1562                 at->at_worst_time = now;
1563         }
1564
1565         if (at->at_flags & AT_FLG_NOHIST)
1566                 /* Only keep last reported val; keeping the rest of the history
1567                    for proc only */
1568                 at->at_current = val;
1569
1570         if (at_max > 0)
1571                 at->at_current =  min(at->at_current, at_max);
1572         at->at_current =  max(at->at_current, at_min);
1573
1574         if (at->at_current != old)
1575                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1576                        "(val=%u) hist %u %u %u %u\n", at,
1577                        old, at->at_current, at->at_current - old, val,
1578                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1579                        at->at_hist[3]);
1580
1581         /* if we changed, report the old value */
1582         old = (at->at_current != old) ? old : 0;
1583
1584         cfs_spin_unlock(&at->at_lock);
1585         return old;
1586 }
1587
1588 /* Find the imp_at index for a given portal; assign if space available */
1589 int import_at_get_index(struct obd_import *imp, int portal)
1590 {
1591         struct imp_at *at = &imp->imp_at;
1592         int i;
1593
1594         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1595                 if (at->iat_portal[i] == portal)
1596                         return i;
1597                 if (at->iat_portal[i] == 0)
1598                         /* unused */
1599                         break;
1600         }
1601
1602         /* Not found in list, add it under a lock */
1603         cfs_spin_lock(&imp->imp_lock);
1604
1605         /* Check unused under lock */
1606         for (; i < IMP_AT_MAX_PORTALS; i++) {
1607                 if (at->iat_portal[i] == portal)
1608                         goto out;
1609                 if (at->iat_portal[i] == 0)
1610                         /* unused */
1611                         break;
1612         }
1613
1614         /* Not enough portals? */
1615         LASSERT(i < IMP_AT_MAX_PORTALS);
1616
1617         at->iat_portal[i] = portal;
1618 out:
1619         cfs_spin_unlock(&imp->imp_lock);
1620         return i;
1621 }