drivers/staging/lustre/lustre/ptlrpc/import.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/ptlrpc/import.c
  37  *
  38  * Author: Mike Shaver <shaver@clusterfs.com>
  39  */
  40
  41 #define DEBUG_SUBSYSTEM S_RPC
  42
  43 #include <obd_support.h>
  44 #include <lustre_ha.h>
  45 #include <lustre_net.h>
  46 #include <lustre_import.h>
  47 #include <lustre_export.h>
  48 #include <obd.h>
  49 #include <obd_cksum.h>
  50 #include <obd_class.h>
  51
  52 #include "ptlrpc_internal.h"
  53
  54 struct ptlrpc_connect_async_args {
  55          __u64 pcaa_peer_committed;
  56         int pcaa_initial_connect;
  57 };
  58
  59 /**
  60  * Updates import \a imp current state to provided \a state value
  61  * Helper function. Must be called under imp_lock.
  62  */
  63 static void __import_set_state(struct obd_import *imp,
  64                                enum lustre_imp_state state)
  65 {
  66         imp->imp_state = state;
  67         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  68         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  69                 cfs_time_current_sec();
  70         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  71                 IMP_STATE_HIST_LEN;
  72 }
  73
  74 /* A CLOSED import should remain so. */
  75 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                 \
  76 do {                                                                       \
  77         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                           \
  78                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  79                       imp, obd2cli_tgt(imp->imp_obd),                     \
  80                       ptlrpc_import_state_name(imp->imp_state),         \
  81                       ptlrpc_import_state_name(state));                 \
  82                __import_set_state(imp, state);                           \
  83         }                                                                     \
  84 } while(0)
  85
  86 #define IMPORT_SET_STATE(imp, state)                                    \
  87 do {                                                                    \
  88         spin_lock(&imp->imp_lock);                                      \
  89         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
  90         spin_unlock(&imp->imp_lock);                                    \
  91 } while(0)
  92
  93
  94 static int ptlrpc_connect_interpret(const struct lu_env *env,
  95                                     struct ptlrpc_request *request,
  96                                     void * data, int rc);
  97 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  98
  99 /* Only this function is allowed to change the import state when it is
 100  * CLOSED. I would rather refcount the import and free it after
 101  * disconnection like we do with exports. To do that, the client_obd
 102  * will need to save the peer info somewhere other than in the import,
 103  * though. */
 104 int ptlrpc_init_import(struct obd_import *imp)
 105 {
 106         spin_lock(&imp->imp_lock);
 107
 108         imp->imp_generation++;
 109         imp->imp_state =  LUSTRE_IMP_NEW;
 110
 111         spin_unlock(&imp->imp_lock);
 112
 113         return 0;
 114 }
 115 EXPORT_SYMBOL(ptlrpc_init_import);
 116
 117 #define UUID_STR "_UUID"
 118 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 119 {
 120         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 121                 ? uuid : uuid + strlen(prefix);
 122
 123         *uuid_len = strlen(*uuid_start);
 124
 125         if (*uuid_len < strlen(UUID_STR))
 126                 return;
 127
 128         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 129                     UUID_STR, strlen(UUID_STR)))
 130                 *uuid_len -= strlen(UUID_STR);
 131 }
 132 EXPORT_SYMBOL(deuuidify);
 133
 134 /**
 135  * Returns true if import was FULL, false if import was already not
 136  * connected.
 137  * @imp - import to be disconnected
 138  * @conn_cnt - connection count (epoch) of the request that timed out
 139  *           and caused the disconnection.  In some cases, multiple
 140  *           inflight requests can fail to a single target (e.g. OST
 141  *           bulk requests) and if one has already caused a reconnection
 142  *           (increasing the import->conn_cnt) the older failure should
 143  *           not also cause a reconnection.  If zero it forces a reconnect.
 144  */
 145 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 146 {
 147         int rc = 0;
 148
 149         spin_lock(&imp->imp_lock);
 150
 151         if (imp->imp_state == LUSTRE_IMP_FULL &&
 152             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 153                 char *target_start;
 154                 int   target_len;
 155
 156                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 157                           &target_start, &target_len);
 158
 159                 if (imp->imp_replayable) {
 160                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 161                                "lost; in progress operations using this "
 162                                "service will wait for recovery to complete\n",
 163                                imp->imp_obd->obd_name, target_len, target_start,
 164                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 165                 } else {
 166                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 167                                "%.*s (at %s) was lost; in progress "
 168                                "operations using this service will fail\n",
 169                                imp->imp_obd->obd_name,
 170                                target_len, target_start,
 171                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 172                 }
 173                 ptlrpc_deactivate_timeouts(imp);
 174                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 175                 spin_unlock(&imp->imp_lock);
 176
 177                 if (obd_dump_on_timeout)
 178                         libcfs_debug_dumplog();
 179
 180                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 181                 rc = 1;
 182         } else {
 183                 spin_unlock(&imp->imp_lock);
 184                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 185                        imp->imp_client->cli_name, imp,
 186                        (imp->imp_state == LUSTRE_IMP_FULL &&
 187                         imp->imp_conn_cnt > conn_cnt) ?
 188                        "reconnected" : "not connected", imp->imp_conn_cnt,
 189                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 190         }
 191
 192         return rc;
 193 }
 194
 195 /* Must be called with imp_lock held! */
 196 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 197 {
 198         LASSERT(spin_is_locked(&imp->imp_lock));
 199
 200         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 201         imp->imp_invalid = 1;
 202         imp->imp_generation++;
 203         spin_unlock(&imp->imp_lock);
 204
 205         ptlrpc_abort_inflight(imp);
 206         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 207
 208         EXIT;
 209 }
 210
 211 /*
 212  * This acts as a barrier; all existing requests are rejected, and
 213  * no new requests will be accepted until the import is valid again.
 214  */
 215 void ptlrpc_deactivate_import(struct obd_import *imp)
 216 {
 217         spin_lock(&imp->imp_lock);
 218         ptlrpc_deactivate_and_unlock_import(imp);
 219 }
 220 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 221
 222 static unsigned int
 223 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 224 {
 225         long dl;
 226
 227         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 228               (req->rq_phase == RQ_PHASE_BULK) ||
 229               (req->rq_phase == RQ_PHASE_NEW)))
 230                 return 0;
 231
 232         if (req->rq_timedout)
 233                 return 0;
 234
 235         if (req->rq_phase == RQ_PHASE_NEW)
 236                 dl = req->rq_sent;
 237         else
 238                 dl = req->rq_deadline;
 239
 240         if (dl <= now)
 241                 return 0;
 242
 243         return dl - now;
 244 }
 245
 246 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 247 {
 248         time_t now = cfs_time_current_sec();
 249         struct list_head *tmp, *n;
 250         struct ptlrpc_request *req;
 251         unsigned int timeout = 0;
 252
 253         spin_lock(&imp->imp_lock);
 254         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 255                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 256                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 257         }
 258         spin_unlock(&imp->imp_lock);
 259         return timeout;
 260 }
 261
 262 /**
 263  * This function will invalidate the import, if necessary, then block
 264  * for all the RPC completions, and finally notify the obd to
 265  * invalidate its state (ie cancel locks, clear pending requests,
 266  * etc).
 267  */
 268 void ptlrpc_invalidate_import(struct obd_import *imp)
 269 {
 270         struct list_head *tmp, *n;
 271         struct ptlrpc_request *req;
 272         struct l_wait_info lwi;
 273         unsigned int timeout;
 274         int rc;
 275
 276         atomic_inc(&imp->imp_inval_count);
 277
 278         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 279                 ptlrpc_deactivate_import(imp);
 280
 281         LASSERT(imp->imp_invalid);
 282
 283         /* Wait forever until inflight == 0. We really can't do it another
 284          * way because in some cases we need to wait for very long reply
 285          * unlink. We can't do anything before that because there is really
 286          * no guarantee that some rdma transfer is not in progress right now. */
 287         do {
 288                 /* Calculate max timeout for waiting on rpcs to error
 289                  * out. Use obd_timeout if calculated value is smaller
 290                  * than it. */
 291                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 292                         timeout = ptlrpc_inflight_timeout(imp);
 293                         timeout += timeout / 3;
 294
 295                         if (timeout == 0)
 296                                 timeout = obd_timeout;
 297                 } else {
 298                         /* decrease the interval to increase race condition */
 299                         timeout = 1;
 300                 }
 301
 302                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 303                        timeout);
 304
 305                 /* Wait for all requests to error out and call completion
 306                  * callbacks. Cap it at obd_timeout -- these should all
 307                  * have been locally cancelled by ptlrpc_abort_inflight. */
 308                 lwi = LWI_TIMEOUT_INTERVAL(
 309                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 310                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 311                         NULL, NULL);
 312                 rc = l_wait_event(imp->imp_recovery_waitq,
 313                                   (atomic_read(&imp->imp_inflight) == 0),
 314                                   &lwi);
 315                 if (rc) {
 316                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 317
 318                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 319                                cli_tgt, rc,
 320                                atomic_read(&imp->imp_inflight));
 321
 322                         spin_lock(&imp->imp_lock);
 323                         if (atomic_read(&imp->imp_inflight) == 0) {
 324                                 int count = atomic_read(&imp->imp_unregistering);
 325
 326                                 /* We know that "unregistering" rpcs only can
 327                                  * survive in sending or delaying lists (they
 328                                  * maybe waiting for long reply unlink in
 329                                  * sluggish nets). Let's check this. If there
 330                                  * is no inflight and unregistering != 0, this
 331                                  * is bug. */
 332                                 LASSERTF(count == 0, "Some RPCs are still "
 333                                          "unregistering: %d\n", count);
 334
 335                                 /* Let's save one loop as soon as inflight have
 336                                  * dropped to zero. No new inflights possible at
 337                                  * this point. */
 338                                 rc = 0;
 339                         } else {
 340                                 list_for_each_safe(tmp, n,
 341                                                        &imp->imp_sending_list) {
 342                                         req = list_entry(tmp,
 343                                                              struct ptlrpc_request,
 344                                                              rq_list);
 345                                         DEBUG_REQ(D_ERROR, req,
 346                                                   "still on sending list");
 347                                 }
 348                                 list_for_each_safe(tmp, n,
 349                                                        &imp->imp_delayed_list) {
 350                                         req = list_entry(tmp,
 351                                                              struct ptlrpc_request,
 352                                                              rq_list);
 353                                         DEBUG_REQ(D_ERROR, req,
 354                                                   "still on delayed list");
 355                                 }
 356
 357                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 358                                        "Network is sluggish? Waiting them "
 359                                        "to error out.\n", cli_tgt,
 360                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 361                                        atomic_read(&imp->
 362                                                        imp_unregistering));
 363                         }
 364                         spin_unlock(&imp->imp_lock);
 365                   }
 366         } while (rc != 0);
 367
 368         /*
 369          * Let's additionally check that no new rpcs added to import in
 370          * "invalidate" state.
 371          */
 372         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 373         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 374         sptlrpc_import_flush_all_ctx(imp);
 375
 376         atomic_dec(&imp->imp_inval_count);
 377         wake_up_all(&imp->imp_recovery_waitq);
 378 }
 379 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 380
 381 /* unset imp_invalid */
 382 void ptlrpc_activate_import(struct obd_import *imp)
 383 {
 384         struct obd_device *obd = imp->imp_obd;
 385
 386         spin_lock(&imp->imp_lock);
 387         imp->imp_invalid = 0;
 388         ptlrpc_activate_timeouts(imp);
 389         spin_unlock(&imp->imp_lock);
 390         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 391 }
 392 EXPORT_SYMBOL(ptlrpc_activate_import);
 393
 394 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 395 {
 396         LASSERT(!imp->imp_dlm_fake);
 397
 398         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 399                 if (!imp->imp_replayable) {
 400                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 401                                "auto-deactivating\n",
 402                                obd2cli_tgt(imp->imp_obd),
 403                                imp->imp_connection->c_remote_uuid.uuid,
 404                                imp->imp_obd->obd_name);
 405                         ptlrpc_deactivate_import(imp);
 406                 }
 407
 408                 CDEBUG(D_HA, "%s: waking up pinger\n",
 409                        obd2cli_tgt(imp->imp_obd));
 410
 411                 spin_lock(&imp->imp_lock);
 412                 imp->imp_force_verify = 1;
 413                 spin_unlock(&imp->imp_lock);
 414
 415                 ptlrpc_pinger_wake_up();
 416         }
 417         EXIT;
 418 }
 419 EXPORT_SYMBOL(ptlrpc_fail_import);
 420
 421 int ptlrpc_reconnect_import(struct obd_import *imp)
 422 {
 423         ptlrpc_set_import_discon(imp, 0);
 424         /* Force a new connect attempt */
 425         ptlrpc_invalidate_import(imp);
 426         /* Do a fresh connect next time by zeroing the handle */
 427         ptlrpc_disconnect_import(imp, 1);
 428         /* Wait for all invalidate calls to finish */
 429         if (atomic_read(&imp->imp_inval_count) > 0) {
 430                 int rc;
 431                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 432                 rc = l_wait_event(imp->imp_recovery_waitq,
 433                                   (atomic_read(&imp->imp_inval_count) == 0),
 434                                   &lwi);
 435                 if (rc)
 436                         CERROR("Interrupted, inval=%d\n",
 437                                atomic_read(&imp->imp_inval_count));
 438         }
 439
 440         /* Allow reconnect attempts */
 441         imp->imp_obd->obd_no_recov = 0;
 442         /* Remove 'invalid' flag */
 443         ptlrpc_activate_import(imp);
 444         /* Attempt a new connect */
 445         ptlrpc_recover_import(imp, NULL, 0);
 446         return 0;
 447 }
 448 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 449
 450 /**
 451  * Connection on import \a imp is changed to another one (if more than one is
 452  * present). We typically chose connection that we have not tried to connect to
 453  * the longest
 454  */
 455 static int import_select_connection(struct obd_import *imp)
 456 {
 457         struct obd_import_conn *imp_conn = NULL, *conn;
 458         struct obd_export *dlmexp;
 459         char *target_start;
 460         int target_len, tried_all = 1;
 461
 462         spin_lock(&imp->imp_lock);
 463
 464         if (list_empty(&imp->imp_conn_list)) {
 465                 CERROR("%s: no connections available\n",
 466                        imp->imp_obd->obd_name);
 467                 spin_unlock(&imp->imp_lock);
 468                 RETURN(-EINVAL);
 469         }
 470
 471         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 472                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 473                        imp->imp_obd->obd_name,
 474                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 475                        conn->oic_last_attempt);
 476
 477                 /* If we have not tried this connection since
 478                    the last successful attempt, go with this one */
 479                 if ((conn->oic_last_attempt == 0) ||
 480                     cfs_time_beforeq_64(conn->oic_last_attempt,
 481                                        imp->imp_last_success_conn)) {
 482                         imp_conn = conn;
 483                         tried_all = 0;
 484                         break;
 485                 }
 486
 487                 /* If all of the connections have already been tried
 488                    since the last successful connection; just choose the
 489                    least recently used */
 490                 if (!imp_conn)
 491                         imp_conn = conn;
 492                 else if (cfs_time_before_64(conn->oic_last_attempt,
 493                                             imp_conn->oic_last_attempt))
 494                         imp_conn = conn;
 495         }
 496
 497         /* if not found, simply choose the current one */
 498         if (!imp_conn || imp->imp_force_reconnect) {
 499                 LASSERT(imp->imp_conn_current);
 500                 imp_conn = imp->imp_conn_current;
 501                 tried_all = 0;
 502         }
 503         LASSERT(imp_conn->oic_conn);
 504
 505         /* If we've tried everything, and we're back to the beginning of the
 506            list, increase our timeout and try again. It will be reset when
 507            we do finally connect. (FIXME: really we should wait for all network
 508            state associated with the last connection attempt to drain before
 509            trying to reconnect on it.) */
 510         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 511                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 512                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 513                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 514                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 515                                 at_reset(at, CONNECTION_SWITCH_MAX);
 516                 }
 517                 LASSERT(imp_conn->oic_last_attempt);
 518                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 519                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 520         }
 521
 522         imp_conn->oic_last_attempt = cfs_time_current_64();
 523
 524         /* switch connection, don't mind if it's same as the current one */
 525         if (imp->imp_connection)
 526                 ptlrpc_connection_put(imp->imp_connection);
 527         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 528
 529         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 530         LASSERT(dlmexp != NULL);
 531         if (dlmexp->exp_connection)
 532                 ptlrpc_connection_put(dlmexp->exp_connection);
 533         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 534         class_export_put(dlmexp);
 535
 536         if (imp->imp_conn_current != imp_conn) {
 537                 if (imp->imp_conn_current) {
 538                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 539                                   &target_start, &target_len);
 540
 541                         CDEBUG(D_HA, "%s: Connection changing to"
 542                                " %.*s (at %s)\n",
 543                                imp->imp_obd->obd_name,
 544                                target_len, target_start,
 545                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 546                 }
 547
 548                 imp->imp_conn_current = imp_conn;
 549         }
 550
 551         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 552                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 553                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 554
 555         spin_unlock(&imp->imp_lock);
 556
 557         RETURN(0);
 558 }
 559
 560 /*
 561  * must be called under imp_lock
 562  */
 563 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 564 {
 565         struct ptlrpc_request *req;
 566         struct list_head *tmp;
 567
 568         if (list_empty(&imp->imp_replay_list))
 569                 return 0;
 570         tmp = imp->imp_replay_list.next;
 571         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 572         *transno = req->rq_transno;
 573         if (req->rq_transno == 0) {
 574                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 575                 LBUG();
 576         }
 577
 578         return 1;
 579 }
 580
 581 /**
 582  * Attempt to (re)connect import \a imp. This includes all preparations,
 583  * initializing CONNECT RPC request and passing it to ptlrpcd for
 584  * actual sending.
 585  * Returns 0 on success or error code.
 586  */
 587 int ptlrpc_connect_import(struct obd_import *imp)
 588 {
 589         struct obd_device *obd = imp->imp_obd;
 590         int initial_connect = 0;
 591         int set_transno = 0;
 592         __u64 committed_before_reconnect = 0;
 593         struct ptlrpc_request *request;
 594         char *bufs[] = { NULL,
 595                          obd2cli_tgt(imp->imp_obd),
 596                          obd->obd_uuid.uuid,
 597                          (char *)&imp->imp_dlm_handle,
 598                          (char *)&imp->imp_connect_data };
 599         struct ptlrpc_connect_async_args *aa;
 600         int rc;
 601
 602         spin_lock(&imp->imp_lock);
 603         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 604                 spin_unlock(&imp->imp_lock);
 605                 CERROR("can't connect to a closed import\n");
 606                 RETURN(-EINVAL);
 607         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 608                 spin_unlock(&imp->imp_lock);
 609                 CERROR("already connected\n");
 610                 RETURN(0);
 611         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 612                 spin_unlock(&imp->imp_lock);
 613                 CERROR("already connecting\n");
 614                 RETURN(-EALREADY);
 615         }
 616
 617         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 618
 619         imp->imp_conn_cnt++;
 620         imp->imp_resend_replay = 0;
 621
 622         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 623                 initial_connect = 1;
 624         else
 625                 committed_before_reconnect = imp->imp_peer_committed_transno;
 626
 627         set_transno = ptlrpc_first_transno(imp,
 628                                            &imp->imp_connect_data.ocd_transno);
 629         spin_unlock(&imp->imp_lock);
 630
 631         rc = import_select_connection(imp);
 632         if (rc)
 633                 GOTO(out, rc);
 634
 635         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 636         if (rc)
 637                 GOTO(out, rc);
 638
 639         /* Reset connect flags to the originally requested flags, in case
 640          * the server is updated on-the-fly we will get the new features. */
 641         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 642         /* Reset ocd_version each time so the server knows the exact versions */
 643         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 644         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 645         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 646
 647         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 648                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 649         if (rc)
 650                 GOTO(out, rc);
 651
 652         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 653         if (request == NULL)
 654                 GOTO(out, rc = -ENOMEM);
 655
 656         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 657                                       imp->imp_connect_op, bufs, NULL);
 658         if (rc) {
 659                 ptlrpc_request_free(request);
 660                 GOTO(out, rc);
 661         }
 662
 663         /* Report the rpc service time to the server so that it knows how long
 664          * to wait for clients to join recovery */
 665         lustre_msg_set_service_time(request->rq_reqmsg,
 666                                     at_timeout2est(request->rq_timeout));
 667
 668         /* The amount of time we give the server to process the connect req.
 669          * import_select_connection will increase the net latency on
 670          * repeated reconnect attempts to cover slow networks.
 671          * We override/ignore the server rpc completion estimate here,
 672          * which may be large if this is a reconnect attempt */
 673         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 674         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 675
 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                 spin_lock(&imp->imp_lock);
 695                 imp->imp_replayable = 1;
 696                 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         int force_verify;
 721
 722         spin_lock(&imp->imp_lock);
 723         force_verify = imp->imp_force_verify != 0;
 724         spin_unlock(&imp->imp_lock);
 725
 726         if (force_verify)
 727                 ptlrpc_pinger_wake_up();
 728 }
 729
 730 static int ptlrpc_busy_reconnect(int rc)
 731 {
 732         return (rc == -EBUSY) || (rc == -EAGAIN);
 733 }
 734
 735 /**
 736  * interpret_reply callback for connect RPCs.
 737  * Looks into returned status of connect operation and decides
 738  * what to do with the import - i.e enter recovery, promote it to
 739  * full state for normal operations of disconnect it due to an error.
 740  */
 741 static int ptlrpc_connect_interpret(const struct lu_env *env,
 742                                     struct ptlrpc_request *request,
 743                                     void *data, int rc)
 744 {
 745         struct ptlrpc_connect_async_args *aa = data;
 746         struct obd_import *imp = request->rq_import;
 747         struct client_obd *cli = &imp->imp_obd->u.cli;
 748         struct lustre_handle old_hdl;
 749         __u64 old_connect_flags;
 750         int msg_flags;
 751         struct obd_connect_data *ocd;
 752         struct obd_export *exp;
 753         int ret;
 754
 755         spin_lock(&imp->imp_lock);
 756         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 757                 imp->imp_connect_tried = 1;
 758                 spin_unlock(&imp->imp_lock);
 759                 RETURN(0);
 760         }
 761
 762         if (rc) {
 763                 /* if this reconnect to busy export - not need select new target
 764                  * for connecting*/
 765                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 766                 spin_unlock(&imp->imp_lock);
 767                 ptlrpc_maybe_ping_import_soon(imp);
 768                 GOTO(out, rc);
 769         }
 770         spin_unlock(&imp->imp_lock);
 771
 772         LASSERT(imp->imp_conn_current);
 773
 774         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 775
 776         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 777                                    RCL_SERVER);
 778         /* server replied obd_connect_data is always bigger */
 779         ocd = req_capsule_server_sized_get(&request->rq_pill,
 780                                            &RMF_CONNECT_DATA, ret);
 781
 782         if (ocd == NULL) {
 783                 CERROR("%s: no connect data from server\n",
 784                        imp->imp_obd->obd_name);
 785                 rc = -EPROTO;
 786                 GOTO(out, rc);
 787         }
 788
 789         spin_lock(&imp->imp_lock);
 790
 791         /* All imports are pingable */
 792         imp->imp_pingable = 1;
 793         imp->imp_force_reconnect = 0;
 794         imp->imp_force_verify = 0;
 795
 796         imp->imp_connect_data = *ocd;
 797
 798         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
 799                imp->imp_obd->obd_name, ocd->ocd_instance);
 800         exp = class_conn2export(&imp->imp_dlm_handle);
 801
 802         spin_unlock(&imp->imp_lock);
 803
 804         /* check that server granted subset of flags we asked for. */
 805         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
 806             ocd->ocd_connect_flags) {
 807                 CERROR("%s: Server didn't granted asked subset of flags: "
 808                        "asked="LPX64" grranted="LPX64"\n",
 809                        imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
 810                        ocd->ocd_connect_flags);
 811                 GOTO(out, rc = -EPROTO);
 812         }
 813
 814         if (!exp) {
 815                 /* This could happen if export is cleaned during the
 816                    connect attempt */
 817                 CERROR("%s: missing export after connect\n",
 818                        imp->imp_obd->obd_name);
 819                 GOTO(out, rc = -ENODEV);
 820         }
 821         old_connect_flags = exp_connect_flags(exp);
 822         exp->exp_connect_data = *ocd;
 823         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
 824         class_export_put(exp);
 825
 826         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 827
 828         if (aa->pcaa_initial_connect) {
 829                 spin_lock(&imp->imp_lock);
 830                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 831                         imp->imp_replayable = 1;
 832                         spin_unlock(&imp->imp_lock);
 833                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 834                                obd2cli_tgt(imp->imp_obd));
 835                 } else {
 836                         imp->imp_replayable = 0;
 837                         spin_unlock(&imp->imp_lock);
 838                 }
 839
 840                 /* if applies, adjust the imp->imp_msg_magic here
 841                  * according to reply flags */
 842
 843                 imp->imp_remote_handle =
 844                                 *lustre_msg_get_handle(request->rq_repmsg);
 845
 846                 /* Initial connects are allowed for clients with non-random
 847                  * uuids when servers are in recovery.  Simply signal the
 848                  * servers replay is complete and wait in REPLAY_WAIT. */
 849                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 850                         CDEBUG(D_HA, "connect to %s during recovery\n",
 851                                obd2cli_tgt(imp->imp_obd));
 852                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 853                 } else {
 854                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 855                         ptlrpc_activate_import(imp);
 856                 }
 857
 858                 GOTO(finish, rc = 0);
 859         }
 860
 861         /* Determine what recovery state to move the import to. */
 862         if (MSG_CONNECT_RECONNECT & msg_flags) {
 863                 memset(&old_hdl, 0, sizeof(old_hdl));
 864                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 865                             sizeof (old_hdl))) {
 866                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
 867                                       "bad handle "LPX64"\n",
 868                                       obd2cli_tgt(imp->imp_obd),
 869                                       imp->imp_connection->c_remote_uuid.uuid,
 870                                       imp->imp_dlm_handle.cookie);
 871                         GOTO(out, rc = -ENOTCONN);
 872                 }
 873
 874                 if (memcmp(&imp->imp_remote_handle,
 875                            lustre_msg_get_handle(request->rq_repmsg),
 876                            sizeof(imp->imp_remote_handle))) {
 877                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 878                                 D_HA : D_WARNING;
 879
 880                         /* Bug 16611/14775: if server handle have changed,
 881                          * that means some sort of disconnection happened.
 882                          * If the server is not in recovery, that also means it
 883                          * already erased all of our state because of previous
 884                          * eviction. If it is in recovery - we are safe to
 885                          * participate since we can reestablish all of our state
 886                          * with server again */
 887                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
 888                                 CDEBUG(level,"%s@%s changed server handle from "
 889                                        LPX64" to "LPX64
 890                                        " but is still in recovery\n",
 891                                        obd2cli_tgt(imp->imp_obd),
 892                                        imp->imp_connection->c_remote_uuid.uuid,
 893                                        imp->imp_remote_handle.cookie,
 894                                        lustre_msg_get_handle(
 895                                        request->rq_repmsg)->cookie);
 896                         } else {
 897                                 LCONSOLE_WARN("Evicted from %s (at %s) "
 898                                               "after server handle changed from "
 899                                               LPX64" to "LPX64"\n",
 900                                               obd2cli_tgt(imp->imp_obd),
 901                                               imp->imp_connection-> \
 902                                               c_remote_uuid.uuid,
 903                                               imp->imp_remote_handle.cookie,
 904                                               lustre_msg_get_handle(
 905                                               request->rq_repmsg)->cookie);
 906                         }
 907
 908
 909                         imp->imp_remote_handle =
 910                                      *lustre_msg_get_handle(request->rq_repmsg);
 911
 912                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 913                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 914                                 GOTO(finish, rc = 0);
 915                         }
 916
 917                 } else {
 918                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 919                                obd2cli_tgt(imp->imp_obd),
 920                                imp->imp_connection->c_remote_uuid.uuid);
 921                 }
 922
 923                 if (imp->imp_invalid) {
 924                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 925                                "marking evicted\n", imp->imp_obd->obd_name);
 926                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 927                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 928                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 929                                imp->imp_obd->obd_name,
 930                                obd2cli_tgt(imp->imp_obd));
 931
 932                         spin_lock(&imp->imp_lock);
 933                         imp->imp_resend_replay = 1;
 934                         spin_unlock(&imp->imp_lock);
 935
 936                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 937                 } else {
 938                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 939                 }
 940         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 941                 LASSERT(imp->imp_replayable);
 942                 imp->imp_remote_handle =
 943                                 *lustre_msg_get_handle(request->rq_repmsg);
 944                 imp->imp_last_replay_transno = 0;
 945                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 946         } else {
 947                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 948                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 949                 imp->imp_remote_handle =
 950                                 *lustre_msg_get_handle(request->rq_repmsg);
 951                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 952         }
 953
 954         /* Sanity checks for a reconnected import. */
 955         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 956                 CERROR("imp_replayable flag does not match server "
 957                        "after reconnect. We should LBUG right here.\n");
 958         }
 959
 960         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 961             lustre_msg_get_last_committed(request->rq_repmsg) <
 962             aa->pcaa_peer_committed) {
 963                 CERROR("%s went back in time (transno "LPD64
 964                        " was previously committed, server now claims "LPD64
 965                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 966                        "id=9646\n",
 967                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 968                        lustre_msg_get_last_committed(request->rq_repmsg));
 969         }
 970
 971 finish:
 972         rc = ptlrpc_import_recovery_state_machine(imp);
 973         if (rc != 0) {
 974                 if (rc == -ENOTCONN) {
 975                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 976                                "invalidating and reconnecting\n",
 977                                obd2cli_tgt(imp->imp_obd),
 978                                imp->imp_connection->c_remote_uuid.uuid);
 979                         ptlrpc_connect_import(imp);
 980                         imp->imp_connect_tried = 1;
 981                         RETURN(0);
 982                 }
 983         } else {
 984
 985                 spin_lock(&imp->imp_lock);
 986                 list_del(&imp->imp_conn_current->oic_item);
 987                 list_add(&imp->imp_conn_current->oic_item,
 988                              &imp->imp_conn_list);
 989                 imp->imp_last_success_conn =
 990                         imp->imp_conn_current->oic_last_attempt;
 991
 992                 spin_unlock(&imp->imp_lock);
 993
 994                 if (!ocd->ocd_ibits_known &&
 995                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 996                         CERROR("Inodebits aware server returned zero compatible"
 997                                " bits?\n");
 998
 999                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1000                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1001                                         LUSTRE_VERSION_OFFSET_WARN ||
1002                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1003                                         LUSTRE_VERSION_OFFSET_WARN)) {
1004                         /* Sigh, some compilers do not like #ifdef in the middle
1005                            of macro arguments */
1006                         const char *older = "older. Consider upgrading server "
1007                                             "or downgrading client";
1008                         const char *newer = "newer than client version. "
1009                                             "Consider upgrading client";
1010
1011                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1012                                       "is much %s (%s)\n",
1013                                       obd2cli_tgt(imp->imp_obd),
1014                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1015                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1016                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1017                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1018                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1019                                       newer : older, LUSTRE_VERSION_STRING);
1020                 }
1021
1022 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1023                 /* Check if server has LU-1252 fix applied to not always swab
1024                  * the IR MNE entries. Do this only once per connection.  This
1025                  * fixup is version-limited, because we don't want to carry the
1026                  * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1027                  * need interop with unpatched 2.2 servers.  For newer servers,
1028                  * the client will do MNE swabbing only as needed.  LU-1644 */
1029                 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1030                              !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1031                              OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
1032                              OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
1033                              OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
1034                              strcmp(imp->imp_obd->obd_type->typ_name,
1035                                     LUSTRE_MGC_NAME) == 0))
1036                         imp->imp_need_mne_swab = 1;
1037                 else /* clear if server was upgraded since last connect */
1038                         imp->imp_need_mne_swab = 0;
1039 #else
1040 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1041 #endif
1042
1043                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1044                         /* We sent to the server ocd_cksum_types with bits set
1045                          * for algorithms we understand. The server masked off
1046                          * the checksum types it doesn't support */
1047                         if ((ocd->ocd_cksum_types &
1048                              cksum_types_supported_client()) == 0) {
1049                                 LCONSOLE_WARN("The negotiation of the checksum "
1050                                               "alogrithm to use with server %s "
1051                                               "failed (%x/%x), disabling "
1052                                               "checksums\n",
1053                                               obd2cli_tgt(imp->imp_obd),
1054                                               ocd->ocd_cksum_types,
1055                                               cksum_types_supported_client());
1056                                 cli->cl_checksum = 0;
1057                                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1058                         } else {
1059                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1060                         }
1061                 } else {
1062                         /* The server does not support OBD_CONNECT_CKSUM.
1063                          * Enforce ADLER for backward compatibility*/
1064                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1065                 }
1066                 cli->cl_cksum_type =cksum_type_select(cli->cl_supp_cksum_types);
1067
1068                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1069                         cli->cl_max_pages_per_rpc =
1070                                 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
1071                                     cli->cl_max_pages_per_rpc);
1072                 else if (imp->imp_connect_op == MDS_CONNECT ||
1073                          imp->imp_connect_op == MGS_CONNECT)
1074                         cli->cl_max_pages_per_rpc = 1;
1075
1076                 /* Reset ns_connect_flags only for initial connect. It might be
1077                  * changed in while using FS and if we reset it in reconnect
1078                  * this leads to losing user settings done before such as
1079                  * disable lru_resize, etc. */
1080                 if (old_connect_flags != exp_connect_flags(exp) ||
1081                     aa->pcaa_initial_connect) {
1082                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1083                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1084                               ocd->ocd_connect_flags);
1085                         imp->imp_obd->obd_namespace->ns_connect_flags =
1086                                 ocd->ocd_connect_flags;
1087                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1088                                 ocd->ocd_connect_flags;
1089                 }
1090
1091                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1092                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1093                         /* We need a per-message support flag, because
1094                            a. we don't know if the incoming connect reply
1095                               supports AT or not (in reply_in_callback)
1096                               until we unpack it.
1097                            b. failovered server means export and flags are gone
1098                               (in ptlrpc_send_reply).
1099                            Can only be set when we know AT is supported at
1100                            both ends */
1101                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1102                 else
1103                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1104
1105                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1106                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1107                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1108                 else
1109                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1110
1111                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1112                         (cli->cl_max_pages_per_rpc > 0));
1113         }
1114
1115 out:
1116         imp->imp_connect_tried = 1;
1117
1118         if (rc != 0) {
1119                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1120                 if (rc == -EACCES) {
1121                         /*
1122                          * Give up trying to reconnect
1123                          * EACCES means client has no permission for connection
1124                          */
1125                         imp->imp_obd->obd_no_recov = 1;
1126                         ptlrpc_deactivate_import(imp);
1127                 }
1128
1129                 if (rc == -EPROTO) {
1130                         struct obd_connect_data *ocd;
1131
1132                         /* reply message might not be ready */
1133                         if (request->rq_repmsg == NULL)
1134                                 RETURN(-EPROTO);
1135
1136                         ocd = req_capsule_server_get(&request->rq_pill,
1137                                                      &RMF_CONNECT_DATA);
1138                         if (ocd &&
1139                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1140                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1141                            /* Actually servers are only supposed to refuse
1142                               connection from liblustre clients, so we should
1143                               never see this from VFS context */
1144                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1145                                         "(%d.%d.%d.%d)"
1146                                         " refused connection from this client "
1147                                         "with an incompatible version (%s).  "
1148                                         "Client must be recompiled\n",
1149                                         obd2cli_tgt(imp->imp_obd),
1150                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1151                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1152                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1153                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1154                                         LUSTRE_VERSION_STRING);
1155                                 ptlrpc_deactivate_import(imp);
1156                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1157                         }
1158                         RETURN(-EPROTO);
1159                 }
1160
1161                 ptlrpc_maybe_ping_import_soon(imp);
1162
1163                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1164                        obd2cli_tgt(imp->imp_obd),
1165                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1166         }
1167
1168         wake_up_all(&imp->imp_recovery_waitq);
1169         RETURN(rc);
1170 }
1171
1172 /**
1173  * interpret callback for "completed replay" RPCs.
1174  * \see signal_completed_replay
1175  */
1176 static int completed_replay_interpret(const struct lu_env *env,
1177                                       struct ptlrpc_request *req,
1178                                       void * data, int rc)
1179 {
1180         atomic_dec(&req->rq_import->imp_replay_inflight);
1181         if (req->rq_status == 0 &&
1182             !req->rq_import->imp_vbr_failed) {
1183                 ptlrpc_import_recovery_state_machine(req->rq_import);
1184         } else {
1185                 if (req->rq_import->imp_vbr_failed) {
1186                         CDEBUG(D_WARNING,
1187                                "%s: version recovery fails, reconnecting\n",
1188                                req->rq_import->imp_obd->obd_name);
1189                 } else {
1190                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1191                                      "reconnecting\n",
1192                                req->rq_import->imp_obd->obd_name,
1193                                req->rq_status);
1194                 }
1195                 ptlrpc_connect_import(req->rq_import);
1196         }
1197
1198         RETURN(0);
1199 }
1200
1201 /**
1202  * Let server know that we have no requests to replay anymore.
1203  * Achieved by just sending a PING request
1204  */
1205 static int signal_completed_replay(struct obd_import *imp)
1206 {
1207         struct ptlrpc_request *req;
1208
1209         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1210                 RETURN(0);
1211
1212         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1213         atomic_inc(&imp->imp_replay_inflight);
1214
1215         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1216                                         OBD_PING);
1217         if (req == NULL) {
1218                 atomic_dec(&imp->imp_replay_inflight);
1219                 RETURN(-ENOMEM);
1220         }
1221
1222         ptlrpc_request_set_replen(req);
1223         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1224         lustre_msg_add_flags(req->rq_reqmsg,
1225                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1226         if (AT_OFF)
1227                 req->rq_timeout *= 3;
1228         req->rq_interpret_reply = completed_replay_interpret;
1229
1230         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1231         RETURN(0);
1232 }
1233
1234 /**
1235  * In kernel code all import invalidation happens in its own
1236  * separate thread, so that whatever application happened to encounter
1237  * a problem could still be killed or otherwise continue
1238  */
1239 static int ptlrpc_invalidate_import_thread(void *data)
1240 {
1241         struct obd_import *imp = data;
1242
1243         unshare_fs_struct();
1244
1245         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1246                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1247                imp->imp_connection->c_remote_uuid.uuid);
1248
1249         ptlrpc_invalidate_import(imp);
1250
1251         if (obd_dump_on_eviction) {
1252                 CERROR("dump the log upon eviction\n");
1253                 libcfs_debug_dumplog();
1254         }
1255
1256         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1257         ptlrpc_import_recovery_state_machine(imp);
1258
1259         class_import_put(imp);
1260         RETURN(0);
1261 }
1262
1263 /**
1264  * This is the state machine for client-side recovery on import.
1265  *
1266  * Typicaly we have two possibly paths. If we came to server and it is not
1267  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1268  * state and reconnect from scratch.
1269  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1270  * We go through our list of requests to replay and send them to server one by
1271  * one.
1272  * After sending all request from the list we change import state to
1273  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1274  * and also all the locks we don't yet have and wait for server to grant us.
1275  * After that we send a special "replay completed" request and change import
1276  * state to IMP_REPLAY_WAIT.
1277  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1278  * state and resend all requests from sending list.
1279  * After that we promote import to FULL state and send all delayed requests
1280  * and import is fully operational after that.
1281  *
1282  */
1283 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1284 {
1285         int rc = 0;
1286         int inflight;
1287         char *target_start;
1288         int target_len;
1289
1290         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1291                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1292                           &target_start, &target_len);
1293                 /* Don't care about MGC eviction */
1294                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1295                            LUSTRE_MGC_NAME) != 0) {
1296                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted "
1297                                            "by %.*s; in progress operations "
1298                                            "using this service will fail.\n",
1299                                            imp->imp_obd->obd_name, target_len,
1300                                            target_start);
1301                 }
1302                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1303                        obd2cli_tgt(imp->imp_obd),
1304                        imp->imp_connection->c_remote_uuid.uuid);
1305                 /* reset vbr_failed flag upon eviction */
1306                 spin_lock(&imp->imp_lock);
1307                 imp->imp_vbr_failed = 0;
1308                 spin_unlock(&imp->imp_lock);
1309
1310                 {
1311                 task_t *task;
1312                 /* bug 17802:  XXX client_disconnect_export vs connect request
1313                  * race. if client will evicted at this time, we start
1314                  * invalidate thread without reference to import and import can
1315                  * be freed at same time. */
1316                 class_import_get(imp);
1317                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1318                                      "ll_imp_inval");
1319                 if (IS_ERR(task)) {
1320                         class_import_put(imp);
1321                         CERROR("error starting invalidate thread: %d\n", rc);
1322                         rc = PTR_ERR(task);
1323                 } else {
1324                         rc = 0;
1325                 }
1326                 RETURN(rc);
1327                 }
1328         }
1329
1330         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1331                 CDEBUG(D_HA, "replay requested by %s\n",
1332                        obd2cli_tgt(imp->imp_obd));
1333                 rc = ptlrpc_replay_next(imp, &inflight);
1334                 if (inflight == 0 &&
1335                     atomic_read(&imp->imp_replay_inflight) == 0) {
1336                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1337                         rc = ldlm_replay_locks(imp);
1338                         if (rc)
1339                                 GOTO(out, rc);
1340                 }
1341                 rc = 0;
1342         }
1343
1344         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1345                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1346                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1347                         rc = signal_completed_replay(imp);
1348                         if (rc)
1349                                 GOTO(out, rc);
1350                 }
1351
1352         }
1353
1354         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1355                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1356                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1357                 }
1358         }
1359
1360         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1361                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1362                        obd2cli_tgt(imp->imp_obd),
1363                        imp->imp_connection->c_remote_uuid.uuid);
1364
1365                 rc = ptlrpc_resend(imp);
1366                 if (rc)
1367                         GOTO(out, rc);
1368                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1369                 ptlrpc_activate_import(imp);
1370
1371                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1372                           &target_start, &target_len);
1373                 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1374                               imp->imp_obd->obd_name,
1375                               target_len, target_start,
1376                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1377         }
1378
1379         if (imp->imp_state == LUSTRE_IMP_FULL) {
1380                 wake_up_all(&imp->imp_recovery_waitq);
1381                 ptlrpc_wake_delayed(imp);
1382         }
1383
1384 out:
1385         RETURN(rc);
1386 }
1387
1388 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1389 {
1390         struct ptlrpc_request *req;
1391         int rq_opc, rc = 0;
1392         int nowait = imp->imp_obd->obd_force;
1393
1394         if (nowait)
1395                 GOTO(set_state, rc);
1396
1397         switch (imp->imp_connect_op) {
1398         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1399         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1400         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1401         default:
1402                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1403                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1404                 RETURN(-EINVAL);
1405         }
1406
1407         if (ptlrpc_import_in_recovery(imp)) {
1408                 struct l_wait_info lwi;
1409                 cfs_duration_t timeout;
1410
1411
1412                 if (AT_OFF) {
1413                         if (imp->imp_server_timeout)
1414                                 timeout = cfs_time_seconds(obd_timeout / 2);
1415                         else
1416                                 timeout = cfs_time_seconds(obd_timeout);
1417                 } else {
1418                         int idx = import_at_get_index(imp,
1419                                 imp->imp_client->cli_request_portal);
1420                         timeout = cfs_time_seconds(
1421                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1422                 }
1423
1424                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1425                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1426                 rc = l_wait_event(imp->imp_recovery_waitq,
1427                                   !ptlrpc_import_in_recovery(imp), &lwi);
1428
1429         }
1430
1431         spin_lock(&imp->imp_lock);
1432         if (imp->imp_state != LUSTRE_IMP_FULL)
1433                 GOTO(out, 0);
1434
1435         spin_unlock(&imp->imp_lock);
1436
1437         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1438                                         LUSTRE_OBD_VERSION, rq_opc);
1439         if (req) {
1440                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1441                  * it fails.  We can get through the above with a down server
1442                  * if the client doesn't know the server is gone yet. */
1443                 req->rq_no_resend = 1;
1444
1445                 /* We want client umounts to happen quickly, no matter the
1446                    server state... */
1447                 req->rq_timeout = min_t(int, req->rq_timeout,
1448                                         INITIAL_CONNECT_TIMEOUT);
1449
1450                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1451                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1452                 ptlrpc_request_set_replen(req);
1453                 rc = ptlrpc_queue_wait(req);
1454                 ptlrpc_req_finished(req);
1455         }
1456
1457 set_state:
1458         spin_lock(&imp->imp_lock);
1459 out:
1460         if (noclose)
1461                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1462         else
1463                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1464         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1465         spin_unlock(&imp->imp_lock);
1466
1467         RETURN(rc);
1468 }
1469 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1470
1471 void ptlrpc_cleanup_imp(struct obd_import *imp)
1472 {
1473         spin_lock(&imp->imp_lock);
1474         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1475         imp->imp_generation++;
1476         spin_unlock(&imp->imp_lock);
1477         ptlrpc_abort_inflight(imp);
1478
1479         EXIT;
1480 }
1481 EXPORT_SYMBOL(ptlrpc_cleanup_imp);
1482
1483 /* Adaptive Timeout utils */
1484 extern unsigned int at_min, at_max, at_history;
1485
1486 /* Bin into timeslices using AT_BINS bins.
1487    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1488    but still smoothing out a return to normalcy from a slow response.
1489    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1490 int at_measured(struct adaptive_timeout *at, unsigned int val)
1491 {
1492         unsigned int old = at->at_current;
1493         time_t now = cfs_time_current_sec();
1494         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1495
1496         LASSERT(at);
1497         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1498                val, at, now - at->at_binstart, at->at_current,
1499                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1500
1501         if (val == 0)
1502                 /* 0's don't count, because we never want our timeout to
1503                    drop to 0, and because 0 could mean an error */
1504                 return 0;
1505
1506         spin_lock(&at->at_lock);
1507
1508         if (unlikely(at->at_binstart == 0)) {
1509                 /* Special case to remove default from history */
1510                 at->at_current = val;
1511                 at->at_worst_ever = val;
1512                 at->at_worst_time = now;
1513                 at->at_hist[0] = val;
1514                 at->at_binstart = now;
1515         } else if (now - at->at_binstart < binlimit ) {
1516                 /* in bin 0 */
1517                 at->at_hist[0] = max(val, at->at_hist[0]);
1518                 at->at_current = max(val, at->at_current);
1519         } else {
1520                 int i, shift;
1521                 unsigned int maxv = val;
1522                 /* move bins over */
1523                 shift = (now - at->at_binstart) / binlimit;
1524                 LASSERT(shift > 0);
1525                 for(i = AT_BINS - 1; i >= 0; i--) {
1526                         if (i >= shift) {
1527                                 at->at_hist[i] = at->at_hist[i - shift];
1528                                 maxv = max(maxv, at->at_hist[i]);
1529                         } else {
1530                                 at->at_hist[i] = 0;
1531                         }
1532                 }
1533                 at->at_hist[0] = val;
1534                 at->at_current = maxv;
1535                 at->at_binstart += shift * binlimit;
1536         }
1537
1538         if (at->at_current > at->at_worst_ever) {
1539                 at->at_worst_ever = at->at_current;
1540                 at->at_worst_time = now;
1541         }
1542
1543         if (at->at_flags & AT_FLG_NOHIST)
1544                 /* Only keep last reported val; keeping the rest of the history
1545                    for proc only */
1546                 at->at_current = val;
1547
1548         if (at_max > 0)
1549                 at->at_current =  min(at->at_current, at_max);
1550         at->at_current =  max(at->at_current, at_min);
1551
1552         if (at->at_current != old)
1553                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1554                        "(val=%u) hist %u %u %u %u\n", at,
1555                        old, at->at_current, at->at_current - old, val,
1556                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1557                        at->at_hist[3]);
1558
1559         /* if we changed, report the old value */
1560         old = (at->at_current != old) ? old : 0;
1561
1562         spin_unlock(&at->at_lock);
1563         return old;
1564 }
1565
1566 /* Find the imp_at index for a given portal; assign if space available */
1567 int import_at_get_index(struct obd_import *imp, int portal)
1568 {
1569         struct imp_at *at = &imp->imp_at;
1570         int i;
1571
1572         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1573                 if (at->iat_portal[i] == portal)
1574                         return i;
1575                 if (at->iat_portal[i] == 0)
1576                         /* unused */
1577                         break;
1578         }
1579
1580         /* Not found in list, add it under a lock */
1581         spin_lock(&imp->imp_lock);
1582
1583         /* Check unused under lock */
1584         for (; i < IMP_AT_MAX_PORTALS; i++) {
1585                 if (at->iat_portal[i] == portal)
1586                         goto out;
1587                 if (at->iat_portal[i] == 0)
1588                         /* unused */
1589                         break;
1590         }
1591
1592         /* Not enough portals? */
1593         LASSERT(i < IMP_AT_MAX_PORTALS);
1594
1595         at->iat_portal[i] = portal;
1596 out:
1597         spin_unlock(&imp->imp_lock);
1598         return i;
1599 }