kernel/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 "../include/obd_support.h"
  44 #include "../include/lustre_ha.h"
  45 #include "../include/lustre_net.h"
  46 #include "../include/lustre_import.h"
  47 #include "../include/lustre_export.h"
  48 #include "../include/obd.h"
  49 #include "../include/obd_cksum.h"
  50 #include "../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         switch (state) {
  67         case LUSTRE_IMP_CLOSED:
  68         case LUSTRE_IMP_NEW:
  69         case LUSTRE_IMP_DISCON:
  70         case LUSTRE_IMP_CONNECTING:
  71                 break;
  72         case LUSTRE_IMP_REPLAY_WAIT:
  73                 imp->imp_replay_state = LUSTRE_IMP_REPLAY_LOCKS;
  74                 break;
  75         default:
  76                 imp->imp_replay_state = LUSTRE_IMP_REPLAY;
  77         }
  78
  79         imp->imp_state = state;
  80         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  81         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  82                 ktime_get_real_seconds();
  83         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  84                 IMP_STATE_HIST_LEN;
  85 }
  86
  87 /* A CLOSED import should remain so. */
  88 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  89 do {                                                                           \
  90         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  91                 CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",   \
  92                        imp, obd2cli_tgt(imp->imp_obd),                         \
  93                        ptlrpc_import_state_name(imp->imp_state),               \
  94                        ptlrpc_import_state_name(state));                       \
  95                 __import_set_state(imp, state);                                \
  96         }                                                                      \
  97 } while (0)
  98
  99 #define IMPORT_SET_STATE(imp, state)                                    \
 100 do {                                                                    \
 101         spin_lock(&imp->imp_lock);                                      \
 102         IMPORT_SET_STATE_NOLOCK(imp, state);                            \
 103         spin_unlock(&imp->imp_lock);                                    \
 104 } while (0)
 105
 106 static int ptlrpc_connect_interpret(const struct lu_env *env,
 107                                     struct ptlrpc_request *request,
 108                                     void *data, int rc);
 109 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 110
 111 /* Only this function is allowed to change the import state when it is
 112  * CLOSED. I would rather refcount the import and free it after
 113  * disconnection like we do with exports. To do that, the client_obd
 114  * will need to save the peer info somewhere other than in the import,
 115  * though. */
 116 int ptlrpc_init_import(struct obd_import *imp)
 117 {
 118         spin_lock(&imp->imp_lock);
 119
 120         imp->imp_generation++;
 121         imp->imp_state = LUSTRE_IMP_NEW;
 122
 123         spin_unlock(&imp->imp_lock);
 124
 125         return 0;
 126 }
 127 EXPORT_SYMBOL(ptlrpc_init_import);
 128
 129 #define UUID_STR "_UUID"
 130 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
 131                       int *uuid_len)
 132 {
 133         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 134                 ? uuid : uuid + strlen(prefix);
 135
 136         *uuid_len = strlen(*uuid_start);
 137
 138         if (*uuid_len < strlen(UUID_STR))
 139                 return;
 140
 141         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 142                     UUID_STR, strlen(UUID_STR)))
 143                 *uuid_len -= strlen(UUID_STR);
 144 }
 145
 146 /**
 147  * Returns true if import was FULL, false if import was already not
 148  * connected.
 149  * @imp - import to be disconnected
 150  * @conn_cnt - connection count (epoch) of the request that timed out
 151  *           and caused the disconnection.  In some cases, multiple
 152  *           inflight requests can fail to a single target (e.g. OST
 153  *           bulk requests) and if one has already caused a reconnection
 154  *           (increasing the import->conn_cnt) the older failure should
 155  *           not also cause a reconnection.  If zero it forces a reconnect.
 156  */
 157 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 158 {
 159         int rc = 0;
 160
 161         spin_lock(&imp->imp_lock);
 162
 163         if (imp->imp_state == LUSTRE_IMP_FULL &&
 164             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 165                 char *target_start;
 166                 int   target_len;
 167
 168                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 169                           &target_start, &target_len);
 170
 171                 if (imp->imp_replayable) {
 172                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was lost; in progress operations using this service will wait for recovery to complete\n",
 173                                       imp->imp_obd->obd_name, target_len, target_start,
 174                                       libcfs_nid2str(imp->imp_connection->c_peer.nid));
 175                 } else {
 176                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to %.*s (at %s) was lost; in progress operations using this service will fail\n",
 177                                            imp->imp_obd->obd_name,
 178                                            target_len, target_start,
 179                                            libcfs_nid2str(imp->imp_connection->c_peer.nid));
 180                 }
 181                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 182                 spin_unlock(&imp->imp_lock);
 183
 184                 if (obd_dump_on_timeout)
 185                         libcfs_debug_dumplog();
 186
 187                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 188                 rc = 1;
 189         } else {
 190                 spin_unlock(&imp->imp_lock);
 191                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 192                        imp->imp_client->cli_name, imp,
 193                        (imp->imp_state == LUSTRE_IMP_FULL &&
 194                         imp->imp_conn_cnt > conn_cnt) ?
 195                        "reconnected" : "not connected", imp->imp_conn_cnt,
 196                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 197         }
 198
 199         return rc;
 200 }
 201
 202 /*
 203  * This acts as a barrier; all existing requests are rejected, and
 204  * no new requests will be accepted until the import is valid again.
 205  */
 206 void ptlrpc_deactivate_import(struct obd_import *imp)
 207 {
 208         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 209
 210         spin_lock(&imp->imp_lock);
 211         imp->imp_invalid = 1;
 212         imp->imp_generation++;
 213         spin_unlock(&imp->imp_lock);
 214
 215         ptlrpc_abort_inflight(imp);
 216         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 217 }
 218 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 219
 220 static unsigned int
 221 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time64_t now)
 222 {
 223         long dl;
 224
 225         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 226               (req->rq_phase == RQ_PHASE_BULK) ||
 227               (req->rq_phase == RQ_PHASE_NEW)))
 228                 return 0;
 229
 230         if (req->rq_timedout)
 231                 return 0;
 232
 233         if (req->rq_phase == RQ_PHASE_NEW)
 234                 dl = req->rq_sent;
 235         else
 236                 dl = req->rq_deadline;
 237
 238         if (dl <= now)
 239                 return 0;
 240
 241         return dl - now;
 242 }
 243
 244 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 245 {
 246         time64_t now = ktime_get_real_seconds();
 247         struct list_head *tmp, *n;
 248         struct ptlrpc_request *req;
 249         unsigned int timeout = 0;
 250
 251         spin_lock(&imp->imp_lock);
 252         list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 253                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
 254                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 255         }
 256         spin_unlock(&imp->imp_lock);
 257         return timeout;
 258 }
 259
 260 /**
 261  * This function will invalidate the import, if necessary, then block
 262  * for all the RPC completions, and finally notify the obd to
 263  * invalidate its state (ie cancel locks, clear pending requests,
 264  * etc).
 265  */
 266 void ptlrpc_invalidate_import(struct obd_import *imp)
 267 {
 268         struct list_head *tmp, *n;
 269         struct ptlrpc_request *req;
 270         struct l_wait_info lwi;
 271         unsigned int timeout;
 272         int rc;
 273
 274         atomic_inc(&imp->imp_inval_count);
 275
 276         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 277                 ptlrpc_deactivate_import(imp);
 278
 279         CFS_FAIL_TIMEOUT(OBD_FAIL_MGS_CONNECT_NET, 3 * cfs_fail_val / 2);
 280         LASSERT(imp->imp_invalid);
 281
 282         /* Wait forever until inflight == 0. We really can't do it another
 283          * way because in some cases we need to wait for very long reply
 284          * unlink. We can't do anything before that because there is really
 285          * no guarantee that some rdma transfer is not in progress right now. */
 286         do {
 287                 /* Calculate max timeout for waiting on rpcs to error
 288                  * out. Use obd_timeout if calculated value is smaller
 289                  * than it. */
 290                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 291                         timeout = ptlrpc_inflight_timeout(imp);
 292                         timeout += timeout / 3;
 293
 294                         if (timeout == 0)
 295                                 timeout = obd_timeout;
 296                 } else {
 297                         /* decrease the interval to increase race condition */
 298                         timeout = 1;
 299                 }
 300
 301                 CDEBUG(D_RPCTRACE,
 302                        "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 unregistering: %d\n",
 333                                          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). Network is sluggish? Waiting them to error out.\n",
 358                                        cli_tgt,
 359                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 360                                        atomic_read(&imp->
 361                                                    imp_unregistering));
 362                         }
 363                         spin_unlock(&imp->imp_lock);
 364                 }
 365         } while (rc != 0);
 366
 367         /*
 368          * Let's additionally check that no new rpcs added to import in
 369          * "invalidate" state.
 370          */
 371         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 372         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 373         sptlrpc_import_flush_all_ctx(imp);
 374
 375         atomic_dec(&imp->imp_inval_count);
 376         wake_up_all(&imp->imp_recovery_waitq);
 377 }
 378 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 379
 380 /* unset imp_invalid */
 381 void ptlrpc_activate_import(struct obd_import *imp)
 382 {
 383         struct obd_device *obd = imp->imp_obd;
 384
 385         spin_lock(&imp->imp_lock);
 386         if (imp->imp_deactive != 0) {
 387                 spin_unlock(&imp->imp_lock);
 388                 return;
 389         }
 390
 391         imp->imp_invalid = 0;
 392         spin_unlock(&imp->imp_lock);
 393         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 394 }
 395 EXPORT_SYMBOL(ptlrpc_activate_import);
 396
 397 static void ptlrpc_pinger_force(struct obd_import *imp)
 398 {
 399         CDEBUG(D_HA, "%s: waking up pinger s:%s\n", obd2cli_tgt(imp->imp_obd),
 400                ptlrpc_import_state_name(imp->imp_state));
 401
 402         spin_lock(&imp->imp_lock);
 403         imp->imp_force_verify = 1;
 404         spin_unlock(&imp->imp_lock);
 405
 406         if (imp->imp_state != LUSTRE_IMP_CONNECTING)
 407                 ptlrpc_pinger_wake_up();
 408 }
 409
 410 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 411 {
 412         LASSERT(!imp->imp_dlm_fake);
 413
 414         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 415                 if (!imp->imp_replayable) {
 416                         CDEBUG(D_HA, "import %s@%s for %s not replayable, auto-deactivating\n",
 417                                obd2cli_tgt(imp->imp_obd),
 418                                imp->imp_connection->c_remote_uuid.uuid,
 419                                imp->imp_obd->obd_name);
 420                         ptlrpc_deactivate_import(imp);
 421                 }
 422
 423                 ptlrpc_pinger_force(imp);
 424         }
 425 }
 426 EXPORT_SYMBOL(ptlrpc_fail_import);
 427
 428 int ptlrpc_reconnect_import(struct obd_import *imp)
 429 {
 430 #ifdef ENABLE_PINGER
 431         struct l_wait_info lwi;
 432         int secs = cfs_time_seconds(obd_timeout);
 433         int rc;
 434
 435         ptlrpc_pinger_force(imp);
 436
 437         CDEBUG(D_HA, "%s: recovery started, waiting %u seconds\n",
 438                obd2cli_tgt(imp->imp_obd), secs);
 439
 440         lwi = LWI_TIMEOUT(secs, NULL, NULL);
 441         rc = l_wait_event(imp->imp_recovery_waitq,
 442                           !ptlrpc_import_in_recovery(imp), &lwi);
 443         CDEBUG(D_HA, "%s: recovery finished s:%s\n", obd2cli_tgt(imp->imp_obd),
 444                ptlrpc_import_state_name(imp->imp_state));
 445         return rc;
 446 #else
 447         ptlrpc_set_import_discon(imp, 0);
 448         /* Force a new connect attempt */
 449         ptlrpc_invalidate_import(imp);
 450         /* Do a fresh connect next time by zeroing the handle */
 451         ptlrpc_disconnect_import(imp, 1);
 452         /* Wait for all invalidate calls to finish */
 453         if (atomic_read(&imp->imp_inval_count) > 0) {
 454                 int rc;
 455                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 456
 457                 rc = l_wait_event(imp->imp_recovery_waitq,
 458                                   (atomic_read(&imp->imp_inval_count) == 0),
 459                                   &lwi);
 460                 if (rc)
 461                         CERROR("Interrupted, inval=%d\n",
 462                                atomic_read(&imp->imp_inval_count));
 463         }
 464
 465         /* Allow reconnect attempts */
 466         imp->imp_obd->obd_no_recov = 0;
 467         /* Remove 'invalid' flag */
 468         ptlrpc_activate_import(imp);
 469         /* Attempt a new connect */
 470         ptlrpc_recover_import(imp, NULL, 0);
 471         return 0;
 472 #endif
 473 }
 474 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 475
 476 /**
 477  * Connection on import \a imp is changed to another one (if more than one is
 478  * present). We typically chose connection that we have not tried to connect to
 479  * the longest
 480  */
 481 static int import_select_connection(struct obd_import *imp)
 482 {
 483         struct obd_import_conn *imp_conn = NULL, *conn;
 484         struct obd_export *dlmexp;
 485         char *target_start;
 486         int target_len, tried_all = 1;
 487
 488         spin_lock(&imp->imp_lock);
 489
 490         if (list_empty(&imp->imp_conn_list)) {
 491                 CERROR("%s: no connections available\n",
 492                        imp->imp_obd->obd_name);
 493                 spin_unlock(&imp->imp_lock);
 494                 return -EINVAL;
 495         }
 496
 497         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 498                 CDEBUG(D_HA, "%s: connect to NID %s last attempt %llu\n",
 499                        imp->imp_obd->obd_name,
 500                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 501                        conn->oic_last_attempt);
 502
 503                 /* If we have not tried this connection since
 504                    the last successful attempt, go with this one */
 505                 if ((conn->oic_last_attempt == 0) ||
 506                     cfs_time_beforeq_64(conn->oic_last_attempt,
 507                                        imp->imp_last_success_conn)) {
 508                         imp_conn = conn;
 509                         tried_all = 0;
 510                         break;
 511                 }
 512
 513                 /* If all of the connections have already been tried
 514                    since the last successful connection; just choose the
 515                    least recently used */
 516                 if (!imp_conn)
 517                         imp_conn = conn;
 518                 else if (cfs_time_before_64(conn->oic_last_attempt,
 519                                             imp_conn->oic_last_attempt))
 520                         imp_conn = conn;
 521         }
 522
 523         /* if not found, simply choose the current one */
 524         if (!imp_conn || imp->imp_force_reconnect) {
 525                 LASSERT(imp->imp_conn_current);
 526                 imp_conn = imp->imp_conn_current;
 527                 tried_all = 0;
 528         }
 529         LASSERT(imp_conn->oic_conn);
 530
 531         /* If we've tried everything, and we're back to the beginning of the
 532            list, increase our timeout and try again. It will be reset when
 533            we do finally connect. (FIXME: really we should wait for all network
 534            state associated with the last connection attempt to drain before
 535            trying to reconnect on it.) */
 536         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 537                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 538
 539                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 540                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 541                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 542                                 at_reset(at, CONNECTION_SWITCH_MAX);
 543                 }
 544                 LASSERT(imp_conn->oic_last_attempt);
 545                 CDEBUG(D_HA, "%s: tried all connections, increasing latency to %ds\n",
 546                        imp->imp_obd->obd_name, at_get(at));
 547         }
 548
 549         imp_conn->oic_last_attempt = cfs_time_current_64();
 550
 551         /* switch connection, don't mind if it's same as the current one */
 552         ptlrpc_connection_put(imp->imp_connection);
 553         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 554
 555         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 556         LASSERT(dlmexp != NULL);
 557         ptlrpc_connection_put(dlmexp->exp_connection);
 558         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 559         class_export_put(dlmexp);
 560
 561         if (imp->imp_conn_current != imp_conn) {
 562                 if (imp->imp_conn_current) {
 563                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 564                                   &target_start, &target_len);
 565
 566                         CDEBUG(D_HA, "%s: Connection changing to %.*s (at %s)\n",
 567                                imp->imp_obd->obd_name,
 568                                target_len, target_start,
 569                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 570                 }
 571
 572                 imp->imp_conn_current = imp_conn;
 573         }
 574
 575         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 576                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 577                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 578
 579         spin_unlock(&imp->imp_lock);
 580
 581         return 0;
 582 }
 583
 584 /*
 585  * must be called under imp_lock
 586  */
 587 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 588 {
 589         struct ptlrpc_request *req;
 590         struct list_head *tmp;
 591
 592         /* The requests in committed_list always have smaller transnos than
 593          * the requests in replay_list */
 594         if (!list_empty(&imp->imp_committed_list)) {
 595                 tmp = imp->imp_committed_list.next;
 596                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 597                 *transno = req->rq_transno;
 598                 if (req->rq_transno == 0) {
 599                         DEBUG_REQ(D_ERROR, req,
 600                                   "zero transno in committed_list");
 601                         LBUG();
 602                 }
 603                 return 1;
 604         }
 605         if (!list_empty(&imp->imp_replay_list)) {
 606                 tmp = imp->imp_replay_list.next;
 607                 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 608                 *transno = req->rq_transno;
 609                 if (req->rq_transno == 0) {
 610                         DEBUG_REQ(D_ERROR, req, "zero transno in replay_list");
 611                         LBUG();
 612                 }
 613                 return 1;
 614         }
 615         return 0;
 616 }
 617
 618 /**
 619  * Attempt to (re)connect import \a imp. This includes all preparations,
 620  * initializing CONNECT RPC request and passing it to ptlrpcd for
 621  * actual sending.
 622  * Returns 0 on success or error code.
 623  */
 624 int ptlrpc_connect_import(struct obd_import *imp)
 625 {
 626         struct obd_device *obd = imp->imp_obd;
 627         int initial_connect = 0;
 628         int set_transno = 0;
 629         __u64 committed_before_reconnect = 0;
 630         struct ptlrpc_request *request;
 631         char *bufs[] = { NULL,
 632                          obd2cli_tgt(imp->imp_obd),
 633                          obd->obd_uuid.uuid,
 634                          (char *)&imp->imp_dlm_handle,
 635                          (char *)&imp->imp_connect_data };
 636         struct ptlrpc_connect_async_args *aa;
 637         int rc;
 638
 639         spin_lock(&imp->imp_lock);
 640         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 641                 spin_unlock(&imp->imp_lock);
 642                 CERROR("can't connect to a closed import\n");
 643                 return -EINVAL;
 644         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 645                 spin_unlock(&imp->imp_lock);
 646                 CERROR("already connected\n");
 647                 return 0;
 648         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 649                 spin_unlock(&imp->imp_lock);
 650                 CERROR("already connecting\n");
 651                 return -EALREADY;
 652         }
 653
 654         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 655
 656         imp->imp_conn_cnt++;
 657         imp->imp_resend_replay = 0;
 658
 659         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 660                 initial_connect = 1;
 661         else
 662                 committed_before_reconnect = imp->imp_peer_committed_transno;
 663
 664         set_transno = ptlrpc_first_transno(imp,
 665                                            &imp->imp_connect_data.ocd_transno);
 666         spin_unlock(&imp->imp_lock);
 667
 668         rc = import_select_connection(imp);
 669         if (rc)
 670                 goto out;
 671
 672         rc = sptlrpc_import_sec_adapt(imp, NULL, NULL);
 673         if (rc)
 674                 goto out;
 675
 676         /* Reset connect flags to the originally requested flags, in case
 677          * the server is updated on-the-fly we will get the new features. */
 678         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 679         /* Reset ocd_version each time so the server knows the exact versions */
 680         imp->imp_connect_data.ocd_version = LUSTRE_VERSION_CODE;
 681         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 682         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 683
 684         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 685                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 686         if (rc)
 687                 goto out;
 688
 689         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 690         if (request == NULL) {
 691                 rc = -ENOMEM;
 692                 goto out;
 693         }
 694
 695         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 696                                       imp->imp_connect_op, bufs, NULL);
 697         if (rc) {
 698                 ptlrpc_request_free(request);
 699                 goto out;
 700         }
 701
 702         /* Report the rpc service time to the server so that it knows how long
 703          * to wait for clients to join recovery */
 704         lustre_msg_set_service_time(request->rq_reqmsg,
 705                                     at_timeout2est(request->rq_timeout));
 706
 707         /* The amount of time we give the server to process the connect req.
 708          * import_select_connection will increase the net latency on
 709          * repeated reconnect attempts to cover slow networks.
 710          * We override/ignore the server rpc completion estimate here,
 711          * which may be large if this is a reconnect attempt */
 712         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 713         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 714
 715         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 716
 717         request->rq_no_resend = request->rq_no_delay = 1;
 718         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 719         /* Allow a slightly larger reply for future growth compatibility */
 720         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 721                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 722         ptlrpc_request_set_replen(request);
 723         request->rq_interpret_reply = ptlrpc_connect_interpret;
 724
 725         CLASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
 726         aa = ptlrpc_req_async_args(request);
 727         memset(aa, 0, sizeof(*aa));
 728
 729         aa->pcaa_peer_committed = committed_before_reconnect;
 730         aa->pcaa_initial_connect = initial_connect;
 731
 732         if (aa->pcaa_initial_connect) {
 733                 spin_lock(&imp->imp_lock);
 734                 imp->imp_replayable = 1;
 735                 spin_unlock(&imp->imp_lock);
 736                 lustre_msg_add_op_flags(request->rq_reqmsg,
 737                                         MSG_CONNECT_INITIAL);
 738         }
 739
 740         if (set_transno)
 741                 lustre_msg_add_op_flags(request->rq_reqmsg,
 742                                         MSG_CONNECT_TRANSNO);
 743
 744         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 745                   request->rq_timeout);
 746         ptlrpcd_add_req(request);
 747         rc = 0;
 748 out:
 749         if (rc != 0)
 750                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 751
 752         return rc;
 753 }
 754 EXPORT_SYMBOL(ptlrpc_connect_import);
 755
 756 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 757 {
 758         int force_verify;
 759
 760         spin_lock(&imp->imp_lock);
 761         force_verify = imp->imp_force_verify != 0;
 762         spin_unlock(&imp->imp_lock);
 763
 764         if (force_verify)
 765                 ptlrpc_pinger_wake_up();
 766 }
 767
 768 static int ptlrpc_busy_reconnect(int rc)
 769 {
 770         return (rc == -EBUSY) || (rc == -EAGAIN);
 771 }
 772
 773 /**
 774  * interpret_reply callback for connect RPCs.
 775  * Looks into returned status of connect operation and decides
 776  * what to do with the import - i.e enter recovery, promote it to
 777  * full state for normal operations of disconnect it due to an error.
 778  */
 779 static int ptlrpc_connect_interpret(const struct lu_env *env,
 780                                     struct ptlrpc_request *request,
 781                                     void *data, int rc)
 782 {
 783         struct ptlrpc_connect_async_args *aa = data;
 784         struct obd_import *imp = request->rq_import;
 785         struct client_obd *cli = &imp->imp_obd->u.cli;
 786         struct lustre_handle old_hdl;
 787         __u64 old_connect_flags;
 788         int msg_flags;
 789         struct obd_connect_data *ocd;
 790         struct obd_export *exp;
 791         int ret;
 792
 793         spin_lock(&imp->imp_lock);
 794         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 795                 imp->imp_connect_tried = 1;
 796                 spin_unlock(&imp->imp_lock);
 797                 return 0;
 798         }
 799
 800         if (rc) {
 801                 /* if this reconnect to busy export - not need select new target
 802                  * for connecting*/
 803                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 804                 spin_unlock(&imp->imp_lock);
 805                 ptlrpc_maybe_ping_import_soon(imp);
 806                 goto out;
 807         }
 808         spin_unlock(&imp->imp_lock);
 809
 810         LASSERT(imp->imp_conn_current);
 811
 812         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 813
 814         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 815                                    RCL_SERVER);
 816         /* server replied obd_connect_data is always bigger */
 817         ocd = req_capsule_server_sized_get(&request->rq_pill,
 818                                            &RMF_CONNECT_DATA, ret);
 819
 820         if (ocd == NULL) {
 821                 CERROR("%s: no connect data from server\n",
 822                        imp->imp_obd->obd_name);
 823                 rc = -EPROTO;
 824                 goto out;
 825         }
 826
 827         spin_lock(&imp->imp_lock);
 828
 829         /* All imports are pingable */
 830         imp->imp_pingable = 1;
 831         imp->imp_force_reconnect = 0;
 832         imp->imp_force_verify = 0;
 833
 834         imp->imp_connect_data = *ocd;
 835
 836         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
 837                imp->imp_obd->obd_name, ocd->ocd_instance);
 838         exp = class_conn2export(&imp->imp_dlm_handle);
 839
 840         spin_unlock(&imp->imp_lock);
 841
 842         /* check that server granted subset of flags we asked for. */
 843         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
 844             ocd->ocd_connect_flags) {
 845                 CERROR("%s: Server didn't granted asked subset of flags: asked=%#llx grranted=%#llx\n",
 846                        imp->imp_obd->obd_name, imp->imp_connect_flags_orig,
 847                        ocd->ocd_connect_flags);
 848                 rc = -EPROTO;
 849                 goto out;
 850         }
 851
 852         if (!exp) {
 853                 /* This could happen if export is cleaned during the
 854                    connect attempt */
 855                 CERROR("%s: missing export after connect\n",
 856                        imp->imp_obd->obd_name);
 857                 rc = -ENODEV;
 858                 goto out;
 859         }
 860         old_connect_flags = exp_connect_flags(exp);
 861         exp->exp_connect_data = *ocd;
 862         imp->imp_obd->obd_self_export->exp_connect_data = *ocd;
 863         class_export_put(exp);
 864
 865         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 866
 867         if (aa->pcaa_initial_connect) {
 868                 spin_lock(&imp->imp_lock);
 869                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 870                         imp->imp_replayable = 1;
 871                         spin_unlock(&imp->imp_lock);
 872                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 873                                obd2cli_tgt(imp->imp_obd));
 874                 } else {
 875                         imp->imp_replayable = 0;
 876                         spin_unlock(&imp->imp_lock);
 877                 }
 878
 879                 /* if applies, adjust the imp->imp_msg_magic here
 880                  * according to reply flags */
 881
 882                 imp->imp_remote_handle =
 883                                 *lustre_msg_get_handle(request->rq_repmsg);
 884
 885                 /* Initial connects are allowed for clients with non-random
 886                  * uuids when servers are in recovery.  Simply signal the
 887                  * servers replay is complete and wait in REPLAY_WAIT. */
 888                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 889                         CDEBUG(D_HA, "connect to %s during recovery\n",
 890                                obd2cli_tgt(imp->imp_obd));
 891                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 892                 } else {
 893                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 894                         ptlrpc_activate_import(imp);
 895                 }
 896
 897                 rc = 0;
 898                 goto finish;
 899         }
 900
 901         /* Determine what recovery state to move the import to. */
 902         if (msg_flags & MSG_CONNECT_RECONNECT) {
 903                 memset(&old_hdl, 0, sizeof(old_hdl));
 904                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 905                             sizeof(old_hdl))) {
 906                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due bad handle %#llx\n",
 907                                       obd2cli_tgt(imp->imp_obd),
 908                                       imp->imp_connection->c_remote_uuid.uuid,
 909                                       imp->imp_dlm_handle.cookie);
 910                         rc = -ENOTCONN;
 911                         goto out;
 912                 }
 913
 914                 if (memcmp(&imp->imp_remote_handle,
 915                            lustre_msg_get_handle(request->rq_repmsg),
 916                            sizeof(imp->imp_remote_handle))) {
 917                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 918                                 D_HA : D_WARNING;
 919
 920                         /* Bug 16611/14775: if server handle have changed,
 921                          * that means some sort of disconnection happened.
 922                          * If the server is not in recovery, that also means it
 923                          * already erased all of our state because of previous
 924                          * eviction. If it is in recovery - we are safe to
 925                          * participate since we can reestablish all of our state
 926                          * with server again */
 927                         if ((msg_flags & MSG_CONNECT_RECOVERING)) {
 928                                 CDEBUG(level, "%s@%s changed server handle from %#llx to %#llx but is still in recovery\n",
 929                                        obd2cli_tgt(imp->imp_obd),
 930                                        imp->imp_connection->c_remote_uuid.uuid,
 931                                        imp->imp_remote_handle.cookie,
 932                                        lustre_msg_get_handle(
 933                                        request->rq_repmsg)->cookie);
 934                         } else {
 935                                 LCONSOLE_WARN("Evicted from %s (at %s) after server handle changed from %#llx to %#llx\n",
 936                                               obd2cli_tgt(imp->imp_obd),
 937                                               imp->imp_connection-> \
 938                                               c_remote_uuid.uuid,
 939                                               imp->imp_remote_handle.cookie,
 940                                               lustre_msg_get_handle(
 941                                                       request->rq_repmsg)->cookie);
 942                         }
 943
 944                         imp->imp_remote_handle =
 945                                      *lustre_msg_get_handle(request->rq_repmsg);
 946
 947                         if (!(msg_flags & MSG_CONNECT_RECOVERING)) {
 948                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 949                                 rc = 0;
 950                                 goto finish;
 951                         }
 952
 953                 } else {
 954                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 955                                obd2cli_tgt(imp->imp_obd),
 956                                imp->imp_connection->c_remote_uuid.uuid);
 957                 }
 958
 959                 if (imp->imp_invalid) {
 960                         CDEBUG(D_HA, "%s: reconnected but import is invalid; marking evicted\n",
 961                                imp->imp_obd->obd_name);
 962                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 963                 } else if (msg_flags & MSG_CONNECT_RECOVERING) {
 964                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 965                                imp->imp_obd->obd_name,
 966                                obd2cli_tgt(imp->imp_obd));
 967
 968                         spin_lock(&imp->imp_lock);
 969                         imp->imp_resend_replay = 1;
 970                         spin_unlock(&imp->imp_lock);
 971
 972                         IMPORT_SET_STATE(imp, imp->imp_replay_state);
 973                 } else {
 974                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 975                 }
 976         } else if ((msg_flags & MSG_CONNECT_RECOVERING) && !imp->imp_invalid) {
 977                 LASSERT(imp->imp_replayable);
 978                 imp->imp_remote_handle =
 979                                 *lustre_msg_get_handle(request->rq_repmsg);
 980                 imp->imp_last_replay_transno = 0;
 981                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 982         } else {
 983                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags not set: %x)",
 984                           imp->imp_obd->obd_name, msg_flags);
 985                 imp->imp_remote_handle =
 986                                 *lustre_msg_get_handle(request->rq_repmsg);
 987                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 988         }
 989
 990         /* Sanity checks for a reconnected import. */
 991         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 992                 CERROR("imp_replayable flag does not match server after reconnect. We should LBUG right here.\n");
 993         }
 994
 995         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 996             lustre_msg_get_last_committed(request->rq_repmsg) <
 997             aa->pcaa_peer_committed) {
 998                 CERROR("%s went back in time (transno %lld was previously committed, server now claims %lld)!  See https://bugzilla.lustre.org/show_bug.cgi?id=9646\n",
 999                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
1000                        lustre_msg_get_last_committed(request->rq_repmsg));
1001         }
1002
1003 finish:
1004         rc = ptlrpc_import_recovery_state_machine(imp);
1005         if (rc != 0) {
1006                 if (rc == -ENOTCONN) {
1007                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery; invalidating and reconnecting\n",
1008                                obd2cli_tgt(imp->imp_obd),
1009                                imp->imp_connection->c_remote_uuid.uuid);
1010                         ptlrpc_connect_import(imp);
1011                         imp->imp_connect_tried = 1;
1012                         return 0;
1013                 }
1014         } else {
1015
1016                 spin_lock(&imp->imp_lock);
1017                 list_del(&imp->imp_conn_current->oic_item);
1018                 list_add(&imp->imp_conn_current->oic_item,
1019                              &imp->imp_conn_list);
1020                 imp->imp_last_success_conn =
1021                         imp->imp_conn_current->oic_last_attempt;
1022
1023                 spin_unlock(&imp->imp_lock);
1024
1025                 if ((imp->imp_connect_flags_orig & OBD_CONNECT_IBITS) &&
1026                     !(ocd->ocd_connect_flags & OBD_CONNECT_IBITS)) {
1027                         LCONSOLE_WARN("%s: MDS %s does not support ibits lock, either very old or invalid: requested %llx, replied %llx\n",
1028                                       imp->imp_obd->obd_name,
1029                                       imp->imp_connection->c_remote_uuid.uuid,
1030                                       imp->imp_connect_flags_orig,
1031                                       ocd->ocd_connect_flags);
1032                         rc = -EPROTO;
1033                         goto out;
1034                 }
1035
1036                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1037                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1038                                         LUSTRE_VERSION_OFFSET_WARN ||
1039                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1040                                         LUSTRE_VERSION_OFFSET_WARN)) {
1041                         /* Sigh, some compilers do not like #ifdef in the middle
1042                            of macro arguments */
1043                         const char *older = "older. Consider upgrading server or downgrading client"
1044                                 ;
1045                         const char *newer = "newer than client version. Consider upgrading client"
1046                                             ;
1047
1048                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) is much %s (%s)\n",
1049                                       obd2cli_tgt(imp->imp_obd),
1050                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1051                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1052                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1053                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1054                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1055                                       newer : older, LUSTRE_VERSION_STRING);
1056                 }
1057
1058 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 2, 50, 0)
1059                 /* Check if server has LU-1252 fix applied to not always swab
1060                  * the IR MNE entries. Do this only once per connection.  This
1061                  * fixup is version-limited, because we don't want to carry the
1062                  * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1063                  * need interop with unpatched 2.2 servers.  For newer servers,
1064                  * the client will do MNE swabbing only as needed.  LU-1644 */
1065                 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1066                              !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1067                              OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
1068                              OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
1069                              OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
1070                              strcmp(imp->imp_obd->obd_type->typ_name,
1071                                     LUSTRE_MGC_NAME) == 0))
1072                         imp->imp_need_mne_swab = 1;
1073                 else /* clear if server was upgraded since last connect */
1074                         imp->imp_need_mne_swab = 0;
1075 #else
1076 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and imp_need_mne_swab"
1077 #endif
1078
1079                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1080                         /* We sent to the server ocd_cksum_types with bits set
1081                          * for algorithms we understand. The server masked off
1082                          * the checksum types it doesn't support */
1083                         if ((ocd->ocd_cksum_types &
1084                              cksum_types_supported_client()) == 0) {
1085                                 LCONSOLE_WARN("The negotiation of the checksum algorithm to use with server %s failed (%x/%x), disabling checksums\n",
1086                                               obd2cli_tgt(imp->imp_obd),
1087                                               ocd->ocd_cksum_types,
1088                                               cksum_types_supported_client());
1089                                 cli->cl_checksum = 0;
1090                                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1091                         } else {
1092                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1093                         }
1094                 } else {
1095                         /* The server does not support OBD_CONNECT_CKSUM.
1096                          * Enforce ADLER for backward compatibility*/
1097                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1098                 }
1099                 cli->cl_cksum_type = cksum_type_select(cli->cl_supp_cksum_types);
1100
1101                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1102                         cli->cl_max_pages_per_rpc =
1103                                 min(ocd->ocd_brw_size >> PAGE_CACHE_SHIFT,
1104                                     cli->cl_max_pages_per_rpc);
1105                 else if (imp->imp_connect_op == MDS_CONNECT ||
1106                          imp->imp_connect_op == MGS_CONNECT)
1107                         cli->cl_max_pages_per_rpc = 1;
1108
1109                 /* Reset ns_connect_flags only for initial connect. It might be
1110                  * changed in while using FS and if we reset it in reconnect
1111                  * this leads to losing user settings done before such as
1112                  * disable lru_resize, etc. */
1113                 if (old_connect_flags != exp_connect_flags(exp) ||
1114                     aa->pcaa_initial_connect) {
1115                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server flags: %#llx\n",
1116                                imp->imp_obd->obd_name, ocd->ocd_connect_flags);
1117                         imp->imp_obd->obd_namespace->ns_connect_flags =
1118                                 ocd->ocd_connect_flags;
1119                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1120                                 ocd->ocd_connect_flags;
1121                 }
1122
1123                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1124                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1125                         /* We need a per-message support flag, because
1126                            a. we don't know if the incoming connect reply
1127                               supports AT or not (in reply_in_callback)
1128                               until we unpack it.
1129                            b. failovered server means export and flags are gone
1130                               (in ptlrpc_send_reply).
1131                            Can only be set when we know AT is supported at
1132                            both ends */
1133                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1134                 else
1135                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1136
1137                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1138                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1139                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1140                 else
1141                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1142
1143                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1144                         (cli->cl_max_pages_per_rpc > 0));
1145         }
1146
1147 out:
1148         imp->imp_connect_tried = 1;
1149
1150         if (rc != 0) {
1151                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1152                 if (rc == -EACCES) {
1153                         /*
1154                          * Give up trying to reconnect
1155                          * EACCES means client has no permission for connection
1156                          */
1157                         imp->imp_obd->obd_no_recov = 1;
1158                         ptlrpc_deactivate_import(imp);
1159                 }
1160
1161                 if (rc == -EPROTO) {
1162                         struct obd_connect_data *ocd;
1163
1164                         /* reply message might not be ready */
1165                         if (request->rq_repmsg == NULL)
1166                                 return -EPROTO;
1167
1168                         ocd = req_capsule_server_get(&request->rq_pill,
1169                                                      &RMF_CONNECT_DATA);
1170                         if (ocd &&
1171                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1172                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1173                                 /*
1174                                  * Actually servers are only supposed to refuse
1175                                  * connection from liblustre clients, so we
1176                                  * should never see this from VFS context
1177                                  */
1178                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version (%d.%d.%d.%d) refused connection from this client with an incompatible version (%s).  Client must be recompiled\n",
1179                                                    obd2cli_tgt(imp->imp_obd),
1180                                                    OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1181                                                    OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1182                                                    OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1183                                                    OBD_OCD_VERSION_FIX(ocd->ocd_version),
1184                                                    LUSTRE_VERSION_STRING);
1185                                 ptlrpc_deactivate_import(imp);
1186                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1187                         }
1188                         return -EPROTO;
1189                 }
1190
1191                 ptlrpc_maybe_ping_import_soon(imp);
1192
1193                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1194                        obd2cli_tgt(imp->imp_obd),
1195                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1196         }
1197
1198         wake_up_all(&imp->imp_recovery_waitq);
1199         return rc;
1200 }
1201
1202 /**
1203  * interpret callback for "completed replay" RPCs.
1204  * \see signal_completed_replay
1205  */
1206 static int completed_replay_interpret(const struct lu_env *env,
1207                                       struct ptlrpc_request *req,
1208                                       void *data, int rc)
1209 {
1210         atomic_dec(&req->rq_import->imp_replay_inflight);
1211         if (req->rq_status == 0 &&
1212             !req->rq_import->imp_vbr_failed) {
1213                 ptlrpc_import_recovery_state_machine(req->rq_import);
1214         } else {
1215                 if (req->rq_import->imp_vbr_failed) {
1216                         CDEBUG(D_WARNING,
1217                                "%s: version recovery fails, reconnecting\n",
1218                                req->rq_import->imp_obd->obd_name);
1219                 } else {
1220                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, reconnecting\n",
1221                                req->rq_import->imp_obd->obd_name,
1222                                req->rq_status);
1223                 }
1224                 ptlrpc_connect_import(req->rq_import);
1225         }
1226
1227         return 0;
1228 }
1229
1230 /**
1231  * Let server know that we have no requests to replay anymore.
1232  * Achieved by just sending a PING request
1233  */
1234 static int signal_completed_replay(struct obd_import *imp)
1235 {
1236         struct ptlrpc_request *req;
1237
1238         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1239                 return 0;
1240
1241         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1242         atomic_inc(&imp->imp_replay_inflight);
1243
1244         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1245                                         OBD_PING);
1246         if (req == NULL) {
1247                 atomic_dec(&imp->imp_replay_inflight);
1248                 return -ENOMEM;
1249         }
1250
1251         ptlrpc_request_set_replen(req);
1252         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1253         lustre_msg_add_flags(req->rq_reqmsg,
1254                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1255         if (AT_OFF)
1256                 req->rq_timeout *= 3;
1257         req->rq_interpret_reply = completed_replay_interpret;
1258
1259         ptlrpcd_add_req(req);
1260         return 0;
1261 }
1262
1263 /**
1264  * In kernel code all import invalidation happens in its own
1265  * separate thread, so that whatever application happened to encounter
1266  * a problem could still be killed or otherwise continue
1267  */
1268 static int ptlrpc_invalidate_import_thread(void *data)
1269 {
1270         struct obd_import *imp = data;
1271
1272         unshare_fs_struct();
1273
1274         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1275                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1276                imp->imp_connection->c_remote_uuid.uuid);
1277
1278         ptlrpc_invalidate_import(imp);
1279
1280         if (obd_dump_on_eviction) {
1281                 CERROR("dump the log upon eviction\n");
1282                 libcfs_debug_dumplog();
1283         }
1284
1285         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1286         ptlrpc_import_recovery_state_machine(imp);
1287
1288         class_import_put(imp);
1289         return 0;
1290 }
1291
1292 /**
1293  * This is the state machine for client-side recovery on import.
1294  *
1295  * Typically we have two possibly paths. If we came to server and it is not
1296  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1297  * state and reconnect from scratch.
1298  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1299  * We go through our list of requests to replay and send them to server one by
1300  * one.
1301  * After sending all request from the list we change import state to
1302  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1303  * and also all the locks we don't yet have and wait for server to grant us.
1304  * After that we send a special "replay completed" request and change import
1305  * state to IMP_REPLAY_WAIT.
1306  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1307  * state and resend all requests from sending list.
1308  * After that we promote import to FULL state and send all delayed requests
1309  * and import is fully operational after that.
1310  *
1311  */
1312 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1313 {
1314         int rc = 0;
1315         int inflight;
1316         char *target_start;
1317         int target_len;
1318
1319         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1320                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1321                           &target_start, &target_len);
1322                 /* Don't care about MGC eviction */
1323                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1324                            LUSTRE_MGC_NAME) != 0) {
1325                         LCONSOLE_ERROR_MSG(0x167, "%s: This client was evicted by %.*s; in progress operations using this service will fail.\n",
1326                                            imp->imp_obd->obd_name, target_len,
1327                                            target_start);
1328                 }
1329                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1330                        obd2cli_tgt(imp->imp_obd),
1331                        imp->imp_connection->c_remote_uuid.uuid);
1332                 /* reset vbr_failed flag upon eviction */
1333                 spin_lock(&imp->imp_lock);
1334                 imp->imp_vbr_failed = 0;
1335                 spin_unlock(&imp->imp_lock);
1336
1337                 {
1338                 struct task_struct *task;
1339                 /* bug 17802:  XXX client_disconnect_export vs connect request
1340                  * race. if client will evicted at this time, we start
1341                  * invalidate thread without reference to import and import can
1342                  * be freed at same time. */
1343                 class_import_get(imp);
1344                 task = kthread_run(ptlrpc_invalidate_import_thread, imp,
1345                                      "ll_imp_inval");
1346                 if (IS_ERR(task)) {
1347                         class_import_put(imp);
1348                         CERROR("error starting invalidate thread: %d\n", rc);
1349                         rc = PTR_ERR(task);
1350                 } else {
1351                         rc = 0;
1352                 }
1353                 return rc;
1354                 }
1355         }
1356
1357         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1358                 CDEBUG(D_HA, "replay requested by %s\n",
1359                        obd2cli_tgt(imp->imp_obd));
1360                 rc = ptlrpc_replay_next(imp, &inflight);
1361                 if (inflight == 0 &&
1362                     atomic_read(&imp->imp_replay_inflight) == 0) {
1363                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1364                         rc = ldlm_replay_locks(imp);
1365                         if (rc)
1366                                 goto out;
1367                 }
1368                 rc = 0;
1369         }
1370
1371         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1372                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1373                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1374                         rc = signal_completed_replay(imp);
1375                         if (rc)
1376                                 goto out;
1377                 }
1378
1379         }
1380
1381         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1382                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1383                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1384                 }
1385         }
1386
1387         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1388                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1389                        obd2cli_tgt(imp->imp_obd),
1390                        imp->imp_connection->c_remote_uuid.uuid);
1391
1392                 rc = ptlrpc_resend(imp);
1393                 if (rc)
1394                         goto out;
1395                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1396                 ptlrpc_activate_import(imp);
1397
1398                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1399                           &target_start, &target_len);
1400                 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1401                               imp->imp_obd->obd_name,
1402                               target_len, target_start,
1403                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1404         }
1405
1406         if (imp->imp_state == LUSTRE_IMP_FULL) {
1407                 wake_up_all(&imp->imp_recovery_waitq);
1408                 ptlrpc_wake_delayed(imp);
1409         }
1410
1411 out:
1412         return rc;
1413 }
1414
1415 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1416 {
1417         struct ptlrpc_request *req;
1418         int rq_opc, rc = 0;
1419
1420         if (imp->imp_obd->obd_force)
1421                 goto set_state;
1422
1423         switch (imp->imp_connect_op) {
1424         case OST_CONNECT:
1425                 rq_opc = OST_DISCONNECT;
1426                 break;
1427         case MDS_CONNECT:
1428                 rq_opc = MDS_DISCONNECT;
1429                 break;
1430         case MGS_CONNECT:
1431                 rq_opc = MGS_DISCONNECT;
1432                 break;
1433         default:
1434                 rc = -EINVAL;
1435                 CERROR("%s: don't know how to disconnect from %s (connect_op %d): rc = %d\n",
1436                        imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1437                        imp->imp_connect_op, rc);
1438                 return rc;
1439         }
1440
1441         if (ptlrpc_import_in_recovery(imp)) {
1442                 struct l_wait_info lwi;
1443                 long timeout;
1444
1445                 if (AT_OFF) {
1446                         if (imp->imp_server_timeout)
1447                                 timeout = cfs_time_seconds(obd_timeout / 2);
1448                         else
1449                                 timeout = cfs_time_seconds(obd_timeout);
1450                 } else {
1451                         int idx = import_at_get_index(imp,
1452                                 imp->imp_client->cli_request_portal);
1453                         timeout = cfs_time_seconds(
1454                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1455                 }
1456
1457                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1458                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1459                 rc = l_wait_event(imp->imp_recovery_waitq,
1460                                   !ptlrpc_import_in_recovery(imp), &lwi);
1461
1462         }
1463
1464         spin_lock(&imp->imp_lock);
1465         if (imp->imp_state != LUSTRE_IMP_FULL)
1466                 goto out;
1467         spin_unlock(&imp->imp_lock);
1468
1469         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1470                                         LUSTRE_OBD_VERSION, rq_opc);
1471         if (req) {
1472                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1473                  * it fails.  We can get through the above with a down server
1474                  * if the client doesn't know the server is gone yet. */
1475                 req->rq_no_resend = 1;
1476
1477                 /* We want client umounts to happen quickly, no matter the
1478                    server state... */
1479                 req->rq_timeout = min_t(int, req->rq_timeout,
1480                                         INITIAL_CONNECT_TIMEOUT);
1481
1482                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1483                 req->rq_send_state = LUSTRE_IMP_CONNECTING;
1484                 ptlrpc_request_set_replen(req);
1485                 rc = ptlrpc_queue_wait(req);
1486                 ptlrpc_req_finished(req);
1487         }
1488
1489 set_state:
1490         spin_lock(&imp->imp_lock);
1491 out:
1492         if (noclose)
1493                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1494         else
1495                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1496         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1497         spin_unlock(&imp->imp_lock);
1498
1499         if (rc == -ETIMEDOUT || rc == -ENOTCONN || rc == -ESHUTDOWN)
1500                 rc = 0;
1501
1502         return rc;
1503 }
1504 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1505
1506 /* Adaptive Timeout utils */
1507 extern unsigned int at_min, at_max, at_history;
1508
1509 /* Bin into timeslices using AT_BINS bins.
1510    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1511    but still smoothing out a return to normalcy from a slow response.
1512    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1513 int at_measured(struct adaptive_timeout *at, unsigned int val)
1514 {
1515         unsigned int old = at->at_current;
1516         time64_t now = ktime_get_real_seconds();
1517         long binlimit = max_t(long, at_history / AT_BINS, 1);
1518
1519         LASSERT(at);
1520         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1521                val, at, (long)(now - at->at_binstart), at->at_current,
1522                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1523
1524         if (val == 0)
1525                 /* 0's don't count, because we never want our timeout to
1526                    drop to 0, and because 0 could mean an error */
1527                 return 0;
1528
1529         spin_lock(&at->at_lock);
1530
1531         if (unlikely(at->at_binstart == 0)) {
1532                 /* Special case to remove default from history */
1533                 at->at_current = val;
1534                 at->at_worst_ever = val;
1535                 at->at_worst_time = now;
1536                 at->at_hist[0] = val;
1537                 at->at_binstart = now;
1538         } else if (now - at->at_binstart < binlimit) {
1539                 /* in bin 0 */
1540                 at->at_hist[0] = max(val, at->at_hist[0]);
1541                 at->at_current = max(val, at->at_current);
1542         } else {
1543                 int i, shift;
1544                 unsigned int maxv = val;
1545                 /* move bins over */
1546                 shift = (u32)(now - at->at_binstart) / binlimit;
1547                 LASSERT(shift > 0);
1548                 for (i = AT_BINS - 1; i >= 0; i--) {
1549                         if (i >= shift) {
1550                                 at->at_hist[i] = at->at_hist[i - shift];
1551                                 maxv = max(maxv, at->at_hist[i]);
1552                         } else {
1553                                 at->at_hist[i] = 0;
1554                         }
1555                 }
1556                 at->at_hist[0] = val;
1557                 at->at_current = maxv;
1558                 at->at_binstart += shift * binlimit;
1559         }
1560
1561         if (at->at_current > at->at_worst_ever) {
1562                 at->at_worst_ever = at->at_current;
1563                 at->at_worst_time = now;
1564         }
1565
1566         if (at->at_flags & AT_FLG_NOHIST)
1567                 /* Only keep last reported val; keeping the rest of the history
1568                    for proc only */
1569                 at->at_current = val;
1570
1571         if (at_max > 0)
1572                 at->at_current =  min(at->at_current, at_max);
1573         at->at_current =  max(at->at_current, at_min);
1574
1575         if (at->at_current != old)
1576                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d (val=%u) hist %u %u %u %u\n",
1577                        at,
1578                        old, at->at_current, at->at_current - old, val,
1579                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1580                        at->at_hist[3]);
1581
1582         /* if we changed, report the old value */
1583         old = (at->at_current != old) ? old : 0;
1584
1585         spin_unlock(&at->at_lock);
1586         return old;
1587 }
1588
1589 /* Find the imp_at index for a given portal; assign if space available */
1590 int import_at_get_index(struct obd_import *imp, int portal)
1591 {
1592         struct imp_at *at = &imp->imp_at;
1593         int i;
1594
1595         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1596                 if (at->iat_portal[i] == portal)
1597                         return i;
1598                 if (at->iat_portal[i] == 0)
1599                         /* unused */
1600                         break;
1601         }
1602
1603         /* Not found in list, add it under a lock */
1604         spin_lock(&imp->imp_lock);
1605
1606         /* Check unused under lock */
1607         for (; i < IMP_AT_MAX_PORTALS; i++) {
1608                 if (at->iat_portal[i] == portal)
1609                         goto out;
1610                 if (at->iat_portal[i] == 0)
1611                         /* unused */
1612                         break;
1613         }
1614
1615         /* Not enough portals? */
1616         LASSERT(i < IMP_AT_MAX_PORTALS);
1617
1618         at->iat_portal[i] = portal;
1619 out:
1620         spin_unlock(&imp->imp_lock);
1621         return i;
1622 }