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, Whamcloud, Inc.
  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 #ifndef __KERNEL__
  43 # include <liblustre.h>
  44 #endif
  45
  46 #include <obd_support.h>
  47 #include <lustre_ha.h>
  48 #include <lustre_net.h>
  49 #include <lustre_import.h>
  50 #include <lustre_export.h>
  51 #include <obd.h>
  52 #include <obd_cksum.h>
  53 #include <obd_class.h>
  54
  55 #include "ptlrpc_internal.h"
  56
  57 struct ptlrpc_connect_async_args {
  58          __u64 pcaa_peer_committed;
  59         int pcaa_initial_connect;
  60 };
  61
  62 /**
  63  * Updates import \a imp current state to provided \a state value
  64  * Helper function. Must be called under imp_lock.
  65  */
  66 static void __import_set_state(struct obd_import *imp,
  67                                enum lustre_imp_state state)
  68 {
  69         imp->imp_state = state;
  70         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  71         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  72                 cfs_time_current_sec();
  73         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  74                 IMP_STATE_HIST_LEN;
  75 }
  76
  77 /* A CLOSED import should remain so. */
  78 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  79 do {                                                                           \
  80         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  81                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  82                       imp, obd2cli_tgt(imp->imp_obd),                          \
  83                       ptlrpc_import_state_name(imp->imp_state),                \
  84                       ptlrpc_import_state_name(state));                        \
  85                __import_set_state(imp, state);                                 \
  86         }                                                                      \
  87 } while(0)
  88
  89 #define IMPORT_SET_STATE(imp, state)            \
  90 do {                                            \
  91         cfs_spin_lock(&imp->imp_lock);          \
  92         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  93         cfs_spin_unlock(&imp->imp_lock);        \
  94 } while(0)
  95
  96
  97 static int ptlrpc_connect_interpret(const struct lu_env *env,
  98                                     struct ptlrpc_request *request,
  99                                     void * data, int rc);
 100 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
 101
 102 /* Only this function is allowed to change the import state when it is
 103  * CLOSED. I would rather refcount the import and free it after
 104  * disconnection like we do with exports. To do that, the client_obd
 105  * will need to save the peer info somewhere other than in the import,
 106  * though. */
 107 int ptlrpc_init_import(struct obd_import *imp)
 108 {
 109         cfs_spin_lock(&imp->imp_lock);
 110
 111         imp->imp_generation++;
 112         imp->imp_state =  LUSTRE_IMP_NEW;
 113
 114         cfs_spin_unlock(&imp->imp_lock);
 115
 116         return 0;
 117 }
 118 EXPORT_SYMBOL(ptlrpc_init_import);
 119
 120 #define UUID_STR "_UUID"
 121 void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
 122 {
 123         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 124                 ? uuid : uuid + strlen(prefix);
 125
 126         *uuid_len = strlen(*uuid_start);
 127
 128         if (*uuid_len < strlen(UUID_STR))
 129                 return;
 130
 131         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 132                     UUID_STR, strlen(UUID_STR)))
 133                 *uuid_len -= strlen(UUID_STR);
 134 }
 135 EXPORT_SYMBOL(deuuidify);
 136
 137 /**
 138  * Returns true if import was FULL, false if import was already not
 139  * connected.
 140  * @imp - import to be disconnected
 141  * @conn_cnt - connection count (epoch) of the request that timed out
 142  *             and caused the disconnection.  In some cases, multiple
 143  *             inflight requests can fail to a single target (e.g. OST
 144  *             bulk requests) and if one has already caused a reconnection
 145  *             (increasing the import->conn_cnt) the older failure should
 146  *             not also cause a reconnection.  If zero it forces a reconnect.
 147  */
 148 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 149 {
 150         int rc = 0;
 151
 152         cfs_spin_lock(&imp->imp_lock);
 153
 154         if (imp->imp_state == LUSTRE_IMP_FULL &&
 155             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 156                 char *target_start;
 157                 int   target_len;
 158
 159                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 160                           &target_start, &target_len);
 161
 162                 if (imp->imp_replayable) {
 163                         LCONSOLE_WARN("%s: Connection to %.*s (at %s) was "
 164                                "lost; in progress operations using this "
 165                                "service will wait for recovery to complete\n",
 166                                imp->imp_obd->obd_name, target_len, target_start,
 167                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 168                 } else {
 169                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to "
 170                                "%.*s (at %s) was lost; in progress "
 171                                "operations using this service will fail\n",
 172                                imp->imp_obd->obd_name,
 173                                target_len, target_start,
 174                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 175                 }
 176                 ptlrpc_deactivate_timeouts(imp);
 177                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 178                 cfs_spin_unlock(&imp->imp_lock);
 179
 180                 if (obd_dump_on_timeout)
 181                         libcfs_debug_dumplog();
 182
 183                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 184                 rc = 1;
 185         } else {
 186                 cfs_spin_unlock(&imp->imp_lock);
 187                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 188                        imp->imp_client->cli_name, imp,
 189                        (imp->imp_state == LUSTRE_IMP_FULL &&
 190                         imp->imp_conn_cnt > conn_cnt) ?
 191                        "reconnected" : "not connected", imp->imp_conn_cnt,
 192                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 193         }
 194
 195         return rc;
 196 }
 197
 198 /* Must be called with imp_lock held! */
 199 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 200 {
 201         ENTRY;
 202         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 203
 204         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 205         imp->imp_invalid = 1;
 206         imp->imp_generation++;
 207         cfs_spin_unlock(&imp->imp_lock);
 208
 209         ptlrpc_abort_inflight(imp);
 210         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 211
 212         EXIT;
 213 }
 214
 215 /*
 216  * This acts as a barrier; all existing requests are rejected, and
 217  * no new requests will be accepted until the import is valid again.
 218  */
 219 void ptlrpc_deactivate_import(struct obd_import *imp)
 220 {
 221         cfs_spin_lock(&imp->imp_lock);
 222         ptlrpc_deactivate_and_unlock_import(imp);
 223 }
 224 EXPORT_SYMBOL(ptlrpc_deactivate_import);
 225
 226 static unsigned int
 227 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_t now)
 228 {
 229         long dl;
 230
 231         if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
 232               (req->rq_phase == RQ_PHASE_BULK) ||
 233               (req->rq_phase == RQ_PHASE_NEW)))
 234                 return 0;
 235
 236         if (req->rq_timedout)
 237                 return 0;
 238
 239         if (req->rq_phase == RQ_PHASE_NEW)
 240                 dl = req->rq_sent;
 241         else
 242                 dl = req->rq_deadline;
 243
 244         if (dl <= now)
 245                 return 0;
 246
 247         return dl - now;
 248 }
 249
 250 static unsigned int ptlrpc_inflight_timeout(struct obd_import *imp)
 251 {
 252         time_t now = cfs_time_current_sec();
 253         cfs_list_t *tmp, *n;
 254         struct ptlrpc_request *req;
 255         unsigned int timeout = 0;
 256
 257         cfs_spin_lock(&imp->imp_lock);
 258         cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 259                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
 260                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 261         }
 262         cfs_spin_unlock(&imp->imp_lock);
 263         return timeout;
 264 }
 265
 266 /**
 267  * This function will invalidate the import, if necessary, then block
 268  * for all the RPC completions, and finally notify the obd to
 269  * invalidate its state (ie cancel locks, clear pending requests,
 270  * etc).
 271  */
 272 void ptlrpc_invalidate_import(struct obd_import *imp)
 273 {
 274         cfs_list_t *tmp, *n;
 275         struct ptlrpc_request *req;
 276         struct l_wait_info lwi;
 277         unsigned int timeout;
 278         int rc;
 279
 280         cfs_atomic_inc(&imp->imp_inval_count);
 281
 282         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 283                 ptlrpc_deactivate_import(imp);
 284
 285         LASSERT(imp->imp_invalid);
 286
 287         /* Wait forever until inflight == 0. We really can't do it another
 288          * way because in some cases we need to wait for very long reply
 289          * unlink. We can't do anything before that because there is really
 290          * no guarantee that some rdma transfer is not in progress right now. */
 291         do {
 292                 /* Calculate max timeout for waiting on rpcs to error
 293                  * out. Use obd_timeout if calculated value is smaller
 294                  * than it. */
 295                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 296                         timeout = ptlrpc_inflight_timeout(imp);
 297                         timeout += timeout / 3;
 298
 299                         if (timeout == 0)
 300                                 timeout = obd_timeout;
 301                 } else {
 302                         /* decrease the interval to increase race condition */
 303                         timeout = 1;
 304                 }
 305
 306                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 307                        timeout);
 308
 309                 /* Wait for all requests to error out and call completion
 310                  * callbacks. Cap it at obd_timeout -- these should all
 311                  * have been locally cancelled by ptlrpc_abort_inflight. */
 312                 lwi = LWI_TIMEOUT_INTERVAL(
 313                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 314                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 315                         NULL, NULL);
 316                 rc = l_wait_event(imp->imp_recovery_waitq,
 317                                   (cfs_atomic_read(&imp->imp_inflight) == 0),
 318                                   &lwi);
 319                 if (rc) {
 320                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 321
 322                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 323                                cli_tgt, rc,
 324                                cfs_atomic_read(&imp->imp_inflight));
 325
 326                         cfs_spin_lock(&imp->imp_lock);
 327                         if (cfs_atomic_read(&imp->imp_inflight) == 0) {
 328                                 int count = cfs_atomic_read(&imp->imp_unregistering);
 329
 330                                 /* We know that "unregistering" rpcs only can
 331                                  * survive in sending or delaying lists (they
 332                                  * maybe waiting for long reply unlink in
 333                                  * sluggish nets). Let's check this. If there
 334                                  * is no inflight and unregistering != 0, this
 335                                  * is bug. */
 336                                 LASSERTF(count == 0, "Some RPCs are still "
 337                                          "unregistering: %d\n", count);
 338
 339                                 /* Let's save one loop as soon as inflight have
 340                                  * dropped to zero. No new inflights possible at
 341                                  * this point. */
 342                                 rc = 0;
 343                         } else {
 344                                 cfs_list_for_each_safe(tmp, n,
 345                                                        &imp->imp_sending_list) {
 346                                         req = cfs_list_entry(tmp,
 347                                                              struct ptlrpc_request,
 348                                                              rq_list);
 349                                         DEBUG_REQ(D_ERROR, req,
 350                                                   "still on sending list");
 351                                 }
 352                                 cfs_list_for_each_safe(tmp, n,
 353                                                        &imp->imp_delayed_list) {
 354                                         req = cfs_list_entry(tmp,
 355                                                              struct ptlrpc_request,
 356                                                              rq_list);
 357                                         DEBUG_REQ(D_ERROR, req,
 358                                                   "still on delayed list");
 359                                 }
 360
 361                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 362                                        "Network is sluggish? Waiting them "
 363                                        "to error out.\n", cli_tgt,
 364                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 365                                        cfs_atomic_read(&imp->
 366                                                        imp_unregistering));
 367                         }
 368                         cfs_spin_unlock(&imp->imp_lock);
 369                   }
 370         } while (rc != 0);
 371
 372         /*
 373          * Let's additionally check that no new rpcs added to import in
 374          * "invalidate" state.
 375          */
 376         LASSERT(cfs_atomic_read(&imp->imp_inflight) == 0);
 377         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 378         sptlrpc_import_flush_all_ctx(imp);
 379
 380         cfs_atomic_dec(&imp->imp_inval_count);
 381         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
 382 }
 383 EXPORT_SYMBOL(ptlrpc_invalidate_import);
 384
 385 /* unset imp_invalid */
 386 void ptlrpc_activate_import(struct obd_import *imp)
 387 {
 388         struct obd_device *obd = imp->imp_obd;
 389
 390         cfs_spin_lock(&imp->imp_lock);
 391         imp->imp_invalid = 0;
 392         ptlrpc_activate_timeouts(imp);
 393         cfs_spin_unlock(&imp->imp_lock);
 394         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 395 }
 396 EXPORT_SYMBOL(ptlrpc_activate_import);
 397
 398 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 399 {
 400         ENTRY;
 401
 402         LASSERT(!imp->imp_dlm_fake);
 403
 404         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 405                 if (!imp->imp_replayable) {
 406                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 407                                "auto-deactivating\n",
 408                                obd2cli_tgt(imp->imp_obd),
 409                                imp->imp_connection->c_remote_uuid.uuid,
 410                                imp->imp_obd->obd_name);
 411                         ptlrpc_deactivate_import(imp);
 412                 }
 413
 414                 CDEBUG(D_HA, "%s: waking up pinger\n",
 415                        obd2cli_tgt(imp->imp_obd));
 416
 417                 cfs_spin_lock(&imp->imp_lock);
 418                 imp->imp_force_verify = 1;
 419                 cfs_spin_unlock(&imp->imp_lock);
 420
 421                 ptlrpc_pinger_wake_up();
 422         }
 423         EXIT;
 424 }
 425 EXPORT_SYMBOL(ptlrpc_fail_import);
 426
 427 int ptlrpc_reconnect_import(struct obd_import *imp)
 428 {
 429         ptlrpc_set_import_discon(imp, 0);
 430         /* Force a new connect attempt */
 431         ptlrpc_invalidate_import(imp);
 432         /* Do a fresh connect next time by zeroing the handle */
 433         ptlrpc_disconnect_import(imp, 1);
 434         /* Wait for all invalidate calls to finish */
 435         if (cfs_atomic_read(&imp->imp_inval_count) > 0) {
 436                 int rc;
 437                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 438                 rc = l_wait_event(imp->imp_recovery_waitq,
 439                                   (cfs_atomic_read(&imp->imp_inval_count) == 0),
 440                                   &lwi);
 441                 if (rc)
 442                         CERROR("Interrupted, inval=%d\n",
 443                                cfs_atomic_read(&imp->imp_inval_count));
 444         }
 445
 446         /* Allow reconnect attempts */
 447         imp->imp_obd->obd_no_recov = 0;
 448         /* Remove 'invalid' flag */
 449         ptlrpc_activate_import(imp);
 450         /* Attempt a new connect */
 451         ptlrpc_recover_import(imp, NULL, 0);
 452         return 0;
 453 }
 454 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 455
 456 /**
 457  * Connection on import \a imp is changed to another one (if more than one is
 458  * present). We typically chose connection that we have not tried to connect to
 459  * the longest
 460  */
 461 static int import_select_connection(struct obd_import *imp)
 462 {
 463         struct obd_import_conn *imp_conn = NULL, *conn;
 464         struct obd_export *dlmexp;
 465         char *target_start;
 466         int target_len, tried_all = 1;
 467         ENTRY;
 468
 469         cfs_spin_lock(&imp->imp_lock);
 470
 471         if (cfs_list_empty(&imp->imp_conn_list)) {
 472                 CERROR("%s: no connections available\n",
 473                         imp->imp_obd->obd_name);
 474                 cfs_spin_unlock(&imp->imp_lock);
 475                 RETURN(-EINVAL);
 476         }
 477
 478         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 479                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 480                        imp->imp_obd->obd_name,
 481                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 482                        conn->oic_last_attempt);
 483
 484                 /* If we have not tried this connection since
 485                    the last successful attempt, go with this one */
 486                 if ((conn->oic_last_attempt == 0) ||
 487                     cfs_time_beforeq_64(conn->oic_last_attempt,
 488                                        imp->imp_last_success_conn)) {
 489                         imp_conn = conn;
 490                         tried_all = 0;
 491                         break;
 492                 }
 493
 494                 /* If all of the connections have already been tried
 495                    since the last successful connection; just choose the
 496                    least recently used */
 497                 if (!imp_conn)
 498                         imp_conn = conn;
 499                 else if (cfs_time_before_64(conn->oic_last_attempt,
 500                                             imp_conn->oic_last_attempt))
 501                         imp_conn = conn;
 502         }
 503
 504         /* if not found, simply choose the current one */
 505         if (!imp_conn || imp->imp_force_reconnect) {
 506                 LASSERT(imp->imp_conn_current);
 507                 imp_conn = imp->imp_conn_current;
 508                 tried_all = 0;
 509         }
 510         LASSERT(imp_conn->oic_conn);
 511
 512         /* If we've tried everything, and we're back to the beginning of the
 513            list, increase our timeout and try again. It will be reset when
 514            we do finally connect. (FIXME: really we should wait for all network
 515            state associated with the last connection attempt to drain before
 516            trying to reconnect on it.) */
 517         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item)) {
 518                 struct adaptive_timeout *at = &imp->imp_at.iat_net_latency;
 519                 if (at_get(at) < CONNECTION_SWITCH_MAX) {
 520                         at_measured(at, at_get(at) + CONNECTION_SWITCH_INC);
 521                         if (at_get(at) > CONNECTION_SWITCH_MAX)
 522                                 at_reset(at, CONNECTION_SWITCH_MAX);
 523                 }
 524                 LASSERT(imp_conn->oic_last_attempt);
 525                 CDEBUG(D_HA, "%s: tried all connections, increasing latency "
 526                         "to %ds\n", imp->imp_obd->obd_name, at_get(at));
 527         }
 528
 529         imp_conn->oic_last_attempt = cfs_time_current_64();
 530
 531         /* switch connection, don't mind if it's same as the current one */
 532         if (imp->imp_connection)
 533                 ptlrpc_connection_put(imp->imp_connection);
 534         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 535
 536         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 537         LASSERT(dlmexp != NULL);
 538         if (dlmexp->exp_connection)
 539                 ptlrpc_connection_put(dlmexp->exp_connection);
 540         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 541         class_export_put(dlmexp);
 542
 543         if (imp->imp_conn_current != imp_conn) {
 544                 if (imp->imp_conn_current) {
 545                         deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 546                                   &target_start, &target_len);
 547
 548                         CDEBUG(D_HA, "%s: Connection changing to"
 549                                " %.*s (at %s)\n",
 550                                imp->imp_obd->obd_name,
 551                                target_len, target_start,
 552                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 553                 }
 554
 555                 imp->imp_conn_current = imp_conn;
 556         }
 557
 558         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 559                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 560                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 561
 562         cfs_spin_unlock(&imp->imp_lock);
 563
 564         RETURN(0);
 565 }
 566
 567 /*
 568  * must be called under imp_lock
 569  */
 570 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 571 {
 572         struct ptlrpc_request *req;
 573         cfs_list_t *tmp;
 574
 575         if (cfs_list_empty(&imp->imp_replay_list))
 576                 return 0;
 577         tmp = imp->imp_replay_list.next;
 578         req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 579         *transno = req->rq_transno;
 580         if (req->rq_transno == 0) {
 581                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 582                 LBUG();
 583         }
 584
 585         return 1;
 586 }
 587
 588 /**
 589  * Attempt to (re)connect import \a imp. This includes all preparations,
 590  * initializing CONNECT RPC request and passing it to ptlrpcd for
 591  * actual sending.
 592  * Returns 0 on success or error code.
 593  */
 594 int ptlrpc_connect_import(struct obd_import *imp)
 595 {
 596         struct obd_device *obd = imp->imp_obd;
 597         int initial_connect = 0;
 598         int set_transno = 0;
 599         __u64 committed_before_reconnect = 0;
 600         struct ptlrpc_request *request;
 601         char *bufs[] = { NULL,
 602                          obd2cli_tgt(imp->imp_obd),
 603                          obd->obd_uuid.uuid,
 604                          (char *)&imp->imp_dlm_handle,
 605                          (char *)&imp->imp_connect_data };
 606         struct ptlrpc_connect_async_args *aa;
 607         int rc;
 608         ENTRY;
 609
 610         cfs_spin_lock(&imp->imp_lock);
 611         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 612                 cfs_spin_unlock(&imp->imp_lock);
 613                 CERROR("can't connect to a closed import\n");
 614                 RETURN(-EINVAL);
 615         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 616                 cfs_spin_unlock(&imp->imp_lock);
 617                 CERROR("already connected\n");
 618                 RETURN(0);
 619         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 620                 cfs_spin_unlock(&imp->imp_lock);
 621                 CERROR("already connecting\n");
 622                 RETURN(-EALREADY);
 623         }
 624
 625         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 626
 627         imp->imp_conn_cnt++;
 628         imp->imp_resend_replay = 0;
 629
 630         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 631                 initial_connect = 1;
 632         else
 633                 committed_before_reconnect = imp->imp_peer_committed_transno;
 634
 635         set_transno = ptlrpc_first_transno(imp,
 636                                            &imp->imp_connect_data.ocd_transno);
 637         cfs_spin_unlock(&imp->imp_lock);
 638
 639         rc = import_select_connection(imp);
 640         if (rc)
 641                 GOTO(out, rc);
 642
 643         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 644         if (rc)
 645                 GOTO(out, rc);
 646
 647         /* Reset connect flags to the originally requested flags, in case
 648          * the server is updated on-the-fly we will get the new features. */
 649         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 650         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 651         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 652
 653         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 654                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 655         if (rc)
 656                 GOTO(out, rc);
 657
 658         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 659         if (request == NULL)
 660                 GOTO(out, rc = -ENOMEM);
 661
 662         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 663                                       imp->imp_connect_op, bufs, NULL);
 664         if (rc) {
 665                 ptlrpc_request_free(request);
 666                 GOTO(out, rc);
 667         }
 668
 669         /* Report the rpc service time to the server so that it knows how long
 670          * to wait for clients to join recovery */
 671         lustre_msg_set_service_time(request->rq_reqmsg,
 672                                     at_timeout2est(request->rq_timeout));
 673
 674         /* The amount of time we give the server to process the connect req.
 675          * import_select_connection will increase the net latency on
 676          * repeated reconnect attempts to cover slow networks.
 677          * We override/ignore the server rpc completion estimate here,
 678          * which may be large if this is a reconnect attempt */
 679         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 680         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 681
 682 #ifndef __KERNEL__
 683         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 684 #endif
 685         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 686
 687         request->rq_no_resend = request->rq_no_delay = 1;
 688         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 689         /* Allow a slightly larger reply for future growth compatibility */
 690         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 691                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 692         ptlrpc_request_set_replen(request);
 693         request->rq_interpret_reply = ptlrpc_connect_interpret;
 694
 695         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 696         aa = ptlrpc_req_async_args(request);
 697         memset(aa, 0, sizeof *aa);
 698
 699         aa->pcaa_peer_committed = committed_before_reconnect;
 700         aa->pcaa_initial_connect = initial_connect;
 701
 702         if (aa->pcaa_initial_connect) {
 703                 cfs_spin_lock(&imp->imp_lock);
 704                 imp->imp_replayable = 1;
 705                 cfs_spin_unlock(&imp->imp_lock);
 706                 lustre_msg_add_op_flags(request->rq_reqmsg,
 707                                         MSG_CONNECT_INITIAL);
 708         }
 709
 710         if (set_transno)
 711                 lustre_msg_add_op_flags(request->rq_reqmsg,
 712                                         MSG_CONNECT_TRANSNO);
 713
 714         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 715                   request->rq_timeout);
 716         ptlrpcd_add_req(request, PDL_POLICY_ROUND, -1);
 717         rc = 0;
 718 out:
 719         if (rc != 0) {
 720                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 721         }
 722
 723         RETURN(rc);
 724 }
 725 EXPORT_SYMBOL(ptlrpc_connect_import);
 726
 727 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 728 {
 729 #ifdef __KERNEL__
 730         int force_verify;
 731
 732         cfs_spin_lock(&imp->imp_lock);
 733         force_verify = imp->imp_force_verify != 0;
 734         cfs_spin_unlock(&imp->imp_lock);
 735
 736         if (force_verify)
 737                 ptlrpc_pinger_wake_up();
 738 #else
 739         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 740         ptlrpc_pinger_wake_up();
 741 #endif
 742 }
 743
 744 static int ptlrpc_busy_reconnect(int rc)
 745 {
 746         return (rc == -EBUSY) || (rc == -EAGAIN);
 747 }
 748
 749 /**
 750  * interpret_reply callback for connect RPCs.
 751  * Looks into returned status of connect operation and decides
 752  * what to do with the import - i.e enter recovery, promote it to
 753  * full state for normal operations of disconnect it due to an error.
 754  */
 755 static int ptlrpc_connect_interpret(const struct lu_env *env,
 756                                     struct ptlrpc_request *request,
 757                                     void *data, int rc)
 758 {
 759         struct ptlrpc_connect_async_args *aa = data;
 760         struct obd_import *imp = request->rq_import;
 761         struct client_obd *cli = &imp->imp_obd->u.cli;
 762         struct lustre_handle old_hdl;
 763         __u64 old_connect_flags;
 764         int msg_flags;
 765         struct obd_connect_data *ocd;
 766         struct obd_export *exp;
 767         int ret;
 768         ENTRY;
 769
 770         cfs_spin_lock(&imp->imp_lock);
 771         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 772                 cfs_spin_unlock(&imp->imp_lock);
 773                 RETURN(0);
 774         }
 775
 776         if (rc) {
 777                 /* if this reconnect to busy export - not need select new target
 778                  * for connecting*/
 779                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 780                 cfs_spin_unlock(&imp->imp_lock);
 781                 ptlrpc_maybe_ping_import_soon(imp);
 782                 GOTO(out, rc);
 783         }
 784         cfs_spin_unlock(&imp->imp_lock);
 785
 786         LASSERT(imp->imp_conn_current);
 787
 788         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 789
 790         ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 791                                    RCL_SERVER);
 792         /* server replied obd_connect_data is always bigger */
 793         ocd = req_capsule_server_sized_get(&request->rq_pill,
 794                                            &RMF_CONNECT_DATA, ret);
 795
 796         if (ocd == NULL) {
 797                 CERROR("%s: no connect data from server\n",
 798                        imp->imp_obd->obd_name);
 799                 rc = -EPROTO;
 800                 GOTO(out, rc);
 801         }
 802
 803         cfs_spin_lock(&imp->imp_lock);
 804
 805         /* All imports are pingable */
 806         imp->imp_pingable = 1;
 807         imp->imp_force_reconnect = 0;
 808         imp->imp_force_verify = 0;
 809
 810         imp->imp_connect_data = *ocd;
 811
 812         CDEBUG(D_HA, "%s: connect to target with instance %u\n",
 813                imp->imp_obd->obd_name, ocd->ocd_instance);
 814         exp = class_conn2export(&imp->imp_dlm_handle);
 815
 816         cfs_spin_unlock(&imp->imp_lock);
 817
 818         /* check that server granted subset of flags we asked for. */
 819         if ((ocd->ocd_connect_flags & imp->imp_connect_flags_orig) !=
 820             ocd->ocd_connect_flags) {
 821                 CERROR("%s: Server didn't granted asked subset of flags: "
 822                        "asked="LPX64" grranted="LPX64"\n",
 823                        imp->imp_obd->obd_name,imp->imp_connect_flags_orig,
 824                        ocd->ocd_connect_flags);
 825                 GOTO(out, rc = -EPROTO);
 826         }
 827
 828         if (!exp) {
 829                 /* This could happen if export is cleaned during the
 830                    connect attempt */
 831                 CERROR("%s: missing export after connect\n",
 832                        imp->imp_obd->obd_name);
 833                 GOTO(out, rc = -ENODEV);
 834         }
 835         old_connect_flags = exp->exp_connect_flags;
 836         exp->exp_connect_flags = ocd->ocd_connect_flags;
 837         imp->imp_obd->obd_self_export->exp_connect_flags =
 838                                                 ocd->ocd_connect_flags;
 839         class_export_put(exp);
 840
 841         obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 842
 843         if (aa->pcaa_initial_connect) {
 844                 cfs_spin_lock(&imp->imp_lock);
 845                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 846                         imp->imp_replayable = 1;
 847                         cfs_spin_unlock(&imp->imp_lock);
 848                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 849                                obd2cli_tgt(imp->imp_obd));
 850                 } else {
 851                         imp->imp_replayable = 0;
 852                         cfs_spin_unlock(&imp->imp_lock);
 853                 }
 854
 855                 /* if applies, adjust the imp->imp_msg_magic here
 856                  * according to reply flags */
 857
 858                 imp->imp_remote_handle =
 859                                 *lustre_msg_get_handle(request->rq_repmsg);
 860
 861                 /* Initial connects are allowed for clients with non-random
 862                  * uuids when servers are in recovery.  Simply signal the
 863                  * servers replay is complete and wait in REPLAY_WAIT. */
 864                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 865                         CDEBUG(D_HA, "connect to %s during recovery\n",
 866                                obd2cli_tgt(imp->imp_obd));
 867                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 868                 } else {
 869                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 870                         ptlrpc_activate_import(imp);
 871                 }
 872
 873                 GOTO(finish, rc = 0);
 874         }
 875
 876         /* Determine what recovery state to move the import to. */
 877         if (MSG_CONNECT_RECONNECT & msg_flags) {
 878                 memset(&old_hdl, 0, sizeof(old_hdl));
 879                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 880                             sizeof (old_hdl))) {
 881                         LCONSOLE_WARN("Reconnect to %s (at @%s) failed due "
 882                                       "bad handle "LPX64"\n",
 883                                       obd2cli_tgt(imp->imp_obd),
 884                                       imp->imp_connection->c_remote_uuid.uuid,
 885                                       imp->imp_dlm_handle.cookie);
 886                         GOTO(out, rc = -ENOTCONN);
 887                 }
 888
 889                 if (memcmp(&imp->imp_remote_handle,
 890                            lustre_msg_get_handle(request->rq_repmsg),
 891                            sizeof(imp->imp_remote_handle))) {
 892                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 893                                 D_HA : D_WARNING;
 894
 895                         /* Bug 16611/14775: if server handle have changed,
 896                          * that means some sort of disconnection happened.
 897                          * If the server is not in recovery, that also means it
 898                          * already erased all of our state because of previous
 899                          * eviction. If it is in recovery - we are safe to
 900                          * participate since we can reestablish all of our state
 901                          * with server again */
 902                         if ((MSG_CONNECT_RECOVERING & msg_flags)) {
 903                                 CDEBUG(level,"%s@%s changed server handle from "
 904                                        LPX64" to "LPX64
 905                                        " but is still in recovery\n",
 906                                        obd2cli_tgt(imp->imp_obd),
 907                                        imp->imp_connection->c_remote_uuid.uuid,
 908                                        imp->imp_remote_handle.cookie,
 909                                        lustre_msg_get_handle(
 910                                        request->rq_repmsg)->cookie);
 911                         } else {
 912                                 LCONSOLE_WARN("Evicted from %s (at %s) "
 913                                               "after server handle changed from "
 914                                               LPX64" to "LPX64"\n",
 915                                               obd2cli_tgt(imp->imp_obd),
 916                                               imp->imp_connection-> \
 917                                               c_remote_uuid.uuid,
 918                                               imp->imp_remote_handle.cookie,
 919                                               lustre_msg_get_handle(
 920                                               request->rq_repmsg)->cookie);
 921                         }
 922
 923
 924                         imp->imp_remote_handle =
 925                                      *lustre_msg_get_handle(request->rq_repmsg);
 926
 927                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 928                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 929                                 GOTO(finish, rc = 0);
 930                         }
 931
 932                 } else {
 933                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 934                                obd2cli_tgt(imp->imp_obd),
 935                                imp->imp_connection->c_remote_uuid.uuid);
 936                 }
 937
 938                 if (imp->imp_invalid) {
 939                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 940                                "marking evicted\n", imp->imp_obd->obd_name);
 941                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 942                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 943                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 944                                imp->imp_obd->obd_name,
 945                                obd2cli_tgt(imp->imp_obd));
 946
 947                         cfs_spin_lock(&imp->imp_lock);
 948                         imp->imp_resend_replay = 1;
 949                         cfs_spin_unlock(&imp->imp_lock);
 950
 951                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 952                 } else {
 953                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 954                 }
 955         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 956                 LASSERT(imp->imp_replayable);
 957                 imp->imp_remote_handle =
 958                                 *lustre_msg_get_handle(request->rq_repmsg);
 959                 imp->imp_last_replay_transno = 0;
 960                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 961         } else {
 962                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 963                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 964                 imp->imp_remote_handle =
 965                                 *lustre_msg_get_handle(request->rq_repmsg);
 966                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 967         }
 968
 969         /* Sanity checks for a reconnected import. */
 970         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 971                 CERROR("imp_replayable flag does not match server "
 972                        "after reconnect. We should LBUG right here.\n");
 973         }
 974
 975         if (lustre_msg_get_last_committed(request->rq_repmsg) > 0 &&
 976             lustre_msg_get_last_committed(request->rq_repmsg) <
 977             aa->pcaa_peer_committed) {
 978                 CERROR("%s went back in time (transno "LPD64
 979                        " was previously committed, server now claims "LPD64
 980                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 981                        "id=9646\n",
 982                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 983                        lustre_msg_get_last_committed(request->rq_repmsg));
 984         }
 985
 986 finish:
 987         rc = ptlrpc_import_recovery_state_machine(imp);
 988         if (rc != 0) {
 989                 if (rc == -ENOTCONN) {
 990                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 991                                "invalidating and reconnecting\n",
 992                                obd2cli_tgt(imp->imp_obd),
 993                                imp->imp_connection->c_remote_uuid.uuid);
 994                         ptlrpc_connect_import(imp);
 995                         RETURN(0);
 996                 }
 997         } else {
 998
 999                 cfs_spin_lock(&imp->imp_lock);
1000                 cfs_list_del(&imp->imp_conn_current->oic_item);
1001                 cfs_list_add(&imp->imp_conn_current->oic_item,
1002                              &imp->imp_conn_list);
1003                 imp->imp_last_success_conn =
1004                         imp->imp_conn_current->oic_last_attempt;
1005
1006                 cfs_spin_unlock(&imp->imp_lock);
1007
1008                 if (!ocd->ocd_ibits_known &&
1009                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
1010                         CERROR("Inodebits aware server returned zero compatible"
1011                                " bits?\n");
1012
1013                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1014                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1015                                         LUSTRE_VERSION_OFFSET_WARN ||
1016                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1017                                         LUSTRE_VERSION_OFFSET_WARN)) {
1018                         /* Sigh, some compilers do not like #ifdef in the middle
1019                            of macro arguments */
1020 #ifdef __KERNEL__
1021                         const char *older = "older. Consider upgrading server "
1022                                             "or downgrading client";
1023 #else
1024                         const char *older = "older. Consider recompiling this "
1025                                             "application";
1026 #endif
1027                         const char *newer = "newer than client version. "
1028                                             "Consider upgrading client";
1029
1030                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1031                                       "is much %s (%s)\n",
1032                                       obd2cli_tgt(imp->imp_obd),
1033                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1034                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1035                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1036                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1037                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1038                                       newer : older, LUSTRE_VERSION_STRING);
1039                 }
1040
1041 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1042                 /* Check if server has LU-1252 fix applied to not always swab
1043                  * the IR MNE entries. Do this only once per connection.  This
1044                  * fixup is version-limited, because we don't want to carry the
1045                  * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we
1046                  * need interop with unpatched 2.2 servers.  For newer servers,
1047                  * the client will do MNE swabbing only as needed.  LU-1644 */
1048                 if (unlikely((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1049                              !(ocd->ocd_connect_flags & OBD_CONNECT_MNE_SWAB) &&
1050                              OBD_OCD_VERSION_MAJOR(ocd->ocd_version) == 2 &&
1051                              OBD_OCD_VERSION_MINOR(ocd->ocd_version) == 2 &&
1052                              OBD_OCD_VERSION_PATCH(ocd->ocd_version) < 55 &&
1053                              strcmp(imp->imp_obd->obd_type->typ_name,
1054                                     LUSTRE_MGC_NAME) == 0))
1055                         imp->imp_need_mne_swab = 1;
1056                 else /* clear if server was upgraded since last connect */
1057                         imp->imp_need_mne_swab = 0;
1058 #else
1059 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1060 #endif
1061
1062                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1063                         /* We sent to the server ocd_cksum_types with bits set
1064                          * for algorithms we understand. The server masked off
1065                          * the checksum types it doesn't support */
1066                         if ((ocd->ocd_cksum_types &
1067                              cksum_types_supported_client()) == 0) {
1068                                 LCONSOLE_WARN("The negotiation of the checksum "
1069                                               "alogrithm to use with server %s "
1070                                               "failed (%x/%x), disabling "
1071                                               "checksums\n",
1072                                               obd2cli_tgt(imp->imp_obd),
1073                                               ocd->ocd_cksum_types,
1074                                               cksum_types_supported_client());
1075                                 cli->cl_checksum = 0;
1076                                 cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1077                         } else {
1078                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1079                         }
1080                 } else {
1081                         /* The server does not support OBD_CONNECT_CKSUM.
1082                          * Enforce ADLER for backward compatibility*/
1083                         cli->cl_supp_cksum_types = OBD_CKSUM_ADLER;
1084                 }
1085                 cli->cl_cksum_type =cksum_type_select(cli->cl_supp_cksum_types);
1086
1087                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE)
1088                         cli->cl_max_pages_per_rpc =
1089                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1090                 else if (imp->imp_connect_op == MDS_CONNECT ||
1091                          imp->imp_connect_op == MGS_CONNECT)
1092                         cli->cl_max_pages_per_rpc = 1;
1093
1094                 /* Reset ns_connect_flags only for initial connect. It might be
1095                  * changed in while using FS and if we reset it in reconnect
1096                  * this leads to losing user settings done before such as
1097                  * disable lru_resize, etc. */
1098                 if (old_connect_flags != exp->exp_connect_flags ||
1099                     aa->pcaa_initial_connect) {
1100                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1101                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1102                               ocd->ocd_connect_flags);
1103                         imp->imp_obd->obd_namespace->ns_connect_flags =
1104                                 ocd->ocd_connect_flags;
1105                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1106                                 ocd->ocd_connect_flags;
1107                 }
1108
1109                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1110                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1111                         /* We need a per-message support flag, because
1112                            a. we don't know if the incoming connect reply
1113                               supports AT or not (in reply_in_callback)
1114                               until we unpack it.
1115                            b. failovered server means export and flags are gone
1116                               (in ptlrpc_send_reply).
1117                            Can only be set when we know AT is supported at
1118                            both ends */
1119                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1120                 else
1121                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1122
1123                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1124                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1125                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1126                 else
1127                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1128
1129                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1130                         (cli->cl_max_pages_per_rpc > 0));
1131         }
1132
1133 out:
1134         if (rc != 0) {
1135                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1136                 if (rc == -EACCES) {
1137                         /*
1138                          * Give up trying to reconnect
1139                          * EACCES means client has no permission for connection
1140                          */
1141                         imp->imp_obd->obd_no_recov = 1;
1142                         ptlrpc_deactivate_import(imp);
1143                 }
1144
1145                 if (rc == -EPROTO) {
1146                         struct obd_connect_data *ocd;
1147
1148                         /* reply message might not be ready */
1149                         if (request->rq_repmsg == NULL)
1150                                 RETURN(-EPROTO);
1151
1152                         ocd = req_capsule_server_get(&request->rq_pill,
1153                                                      &RMF_CONNECT_DATA);
1154                         if (ocd &&
1155                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1156                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1157                            /* Actually servers are only supposed to refuse
1158                               connection from liblustre clients, so we should
1159                               never see this from VFS context */
1160                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1161                                         "(%d.%d.%d.%d)"
1162                                         " refused connection from this client "
1163                                         "with an incompatible version (%s).  "
1164                                         "Client must be recompiled\n",
1165                                         obd2cli_tgt(imp->imp_obd),
1166                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1167                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1168                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1169                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1170                                         LUSTRE_VERSION_STRING);
1171                                 ptlrpc_deactivate_import(imp);
1172                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1173                         }
1174                         RETURN(-EPROTO);
1175                 }
1176
1177                 ptlrpc_maybe_ping_import_soon(imp);
1178
1179                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1180                        obd2cli_tgt(imp->imp_obd),
1181                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1182         }
1183
1184         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1185         RETURN(rc);
1186 }
1187
1188 /**
1189  * interpret callback for "completed replay" RPCs.
1190  * \see signal_completed_replay
1191  */
1192 static int completed_replay_interpret(const struct lu_env *env,
1193                                       struct ptlrpc_request *req,
1194                                       void * data, int rc)
1195 {
1196         ENTRY;
1197         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1198         if (req->rq_status == 0 &&
1199             !req->rq_import->imp_vbr_failed) {
1200                 ptlrpc_import_recovery_state_machine(req->rq_import);
1201         } else {
1202                 if (req->rq_import->imp_vbr_failed) {
1203                         CDEBUG(D_WARNING,
1204                                "%s: version recovery fails, reconnecting\n",
1205                                req->rq_import->imp_obd->obd_name);
1206                 } else {
1207                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1208                                      "reconnecting\n",
1209                                req->rq_import->imp_obd->obd_name,
1210                                req->rq_status);
1211                 }
1212                 ptlrpc_connect_import(req->rq_import);
1213         }
1214
1215         RETURN(0);
1216 }
1217
1218 /**
1219  * Let server know that we have no requests to replay anymore.
1220  * Achieved by just sending a PING request
1221  */
1222 static int signal_completed_replay(struct obd_import *imp)
1223 {
1224         struct ptlrpc_request *req;
1225         ENTRY;
1226
1227         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_FINISH_REPLAY)))
1228                 RETURN(0);
1229
1230         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1231         cfs_atomic_inc(&imp->imp_replay_inflight);
1232
1233         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1234                                         OBD_PING);
1235         if (req == NULL) {
1236                 cfs_atomic_dec(&imp->imp_replay_inflight);
1237                 RETURN(-ENOMEM);
1238         }
1239
1240         ptlrpc_request_set_replen(req);
1241         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1242         lustre_msg_add_flags(req->rq_reqmsg,
1243                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1244         if (AT_OFF)
1245                 req->rq_timeout *= 3;
1246         req->rq_interpret_reply = completed_replay_interpret;
1247
1248         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
1249         RETURN(0);
1250 }
1251
1252 #ifdef __KERNEL__
1253 /**
1254  * In kernel code all import invalidation happens in its own
1255  * separate thread, so that whatever application happened to encounter
1256  * a problem could still be killed or otherwise continue
1257  */
1258 static int ptlrpc_invalidate_import_thread(void *data)
1259 {
1260         struct obd_import *imp = data;
1261
1262         ENTRY;
1263
1264         cfs_daemonize_ctxt("ll_imp_inval");
1265
1266         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1267                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1268                imp->imp_connection->c_remote_uuid.uuid);
1269
1270         ptlrpc_invalidate_import(imp);
1271
1272         if (obd_dump_on_eviction) {
1273                 CERROR("dump the log upon eviction\n");
1274                 libcfs_debug_dumplog();
1275         }
1276
1277         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1278         ptlrpc_import_recovery_state_machine(imp);
1279
1280         class_import_put(imp);
1281         RETURN(0);
1282 }
1283 #endif
1284
1285 /**
1286  * This is the state machine for client-side recovery on import.
1287  *
1288  * Typicaly we have two possibly paths. If we came to server and it is not
1289  * in recovery, we just enter IMP_EVICTED state, invalidate our import
1290  * state and reconnect from scratch.
1291  * If we came to server that is in recovery, we enter IMP_REPLAY import state.
1292  * We go through our list of requests to replay and send them to server one by
1293  * one.
1294  * After sending all request from the list we change import state to
1295  * IMP_REPLAY_LOCKS and re-request all the locks we believe we have from server
1296  * and also all the locks we don't yet have and wait for server to grant us.
1297  * After that we send a special "replay completed" request and change import
1298  * state to IMP_REPLAY_WAIT.
1299  * Upon receiving reply to that "replay completed" RPC we enter IMP_RECOVER
1300  * state and resend all requests from sending list.
1301  * After that we promote import to FULL state and send all delayed requests
1302  * and import is fully operational after that.
1303  *
1304  */
1305 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1306 {
1307         int rc = 0;
1308         int inflight;
1309         char *target_start;
1310         int target_len;
1311
1312         ENTRY;
1313         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1314                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1315                           &target_start, &target_len);
1316                 /* Don't care about MGC eviction */
1317                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1318                            LUSTRE_MGC_NAME) != 0) {
1319                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1320                                            "%.*s; in progress operations using "
1321                                            "this service will fail.\n",
1322                                            target_len, target_start);
1323                 }
1324                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1325                        obd2cli_tgt(imp->imp_obd),
1326                        imp->imp_connection->c_remote_uuid.uuid);
1327                 /* reset vbr_failed flag upon eviction */
1328                 cfs_spin_lock(&imp->imp_lock);
1329                 imp->imp_vbr_failed = 0;
1330                 cfs_spin_unlock(&imp->imp_lock);
1331
1332 #ifdef __KERNEL__
1333                 /* bug 17802:  XXX client_disconnect_export vs connect request
1334                  * race. if client will evicted at this time, we start
1335                  * invalidate thread without reference to import and import can
1336                  * be freed at same time. */
1337                 class_import_get(imp);
1338                 rc = cfs_create_thread(ptlrpc_invalidate_import_thread, imp,
1339                                        CFS_DAEMON_FLAGS);
1340                 if (rc < 0) {
1341                         class_import_put(imp);
1342                         CERROR("error starting invalidate thread: %d\n", rc);
1343                 } else {
1344                         rc = 0;
1345                 }
1346                 RETURN(rc);
1347 #else
1348                 ptlrpc_invalidate_import(imp);
1349
1350                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1351 #endif
1352         }
1353
1354         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1355                 CDEBUG(D_HA, "replay requested by %s\n",
1356                        obd2cli_tgt(imp->imp_obd));
1357                 rc = ptlrpc_replay_next(imp, &inflight);
1358                 if (inflight == 0 &&
1359                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1360                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1361                         rc = ldlm_replay_locks(imp);
1362                         if (rc)
1363                                 GOTO(out, rc);
1364                 }
1365                 rc = 0;
1366         }
1367
1368         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1369                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1370                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1371                         rc = signal_completed_replay(imp);
1372                         if (rc)
1373                                 GOTO(out, rc);
1374                 }
1375
1376         }
1377
1378         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1379                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1380                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1381                 }
1382         }
1383
1384         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1385                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1386                        obd2cli_tgt(imp->imp_obd),
1387                        imp->imp_connection->c_remote_uuid.uuid);
1388
1389                 rc = ptlrpc_resend(imp);
1390                 if (rc)
1391                         GOTO(out, rc);
1392                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1393                 ptlrpc_activate_import(imp);
1394
1395                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1396                           &target_start, &target_len);
1397                 LCONSOLE_INFO("%s: Connection restored to %.*s (at %s)\n",
1398                               imp->imp_obd->obd_name,
1399                               target_len, target_start,
1400                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1401         }
1402
1403         if (imp->imp_state == LUSTRE_IMP_FULL) {
1404                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1405                 ptlrpc_wake_delayed(imp);
1406         }
1407
1408 out:
1409         RETURN(rc);
1410 }
1411
1412 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1413 {
1414         struct ptlrpc_request *req;
1415         int rq_opc, rc = 0;
1416         int nowait = imp->imp_obd->obd_force;
1417         ENTRY;
1418
1419         if (nowait)
1420                 GOTO(set_state, rc);
1421
1422         switch (imp->imp_connect_op) {
1423         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1424         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1425         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1426         default:
1427                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1428                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1429                 RETURN(-EINVAL);
1430         }
1431
1432         if (ptlrpc_import_in_recovery(imp)) {
1433                 struct l_wait_info lwi;
1434                 cfs_duration_t timeout;
1435
1436
1437                 if (AT_OFF) {
1438                         if (imp->imp_server_timeout)
1439                                 timeout = cfs_time_seconds(obd_timeout / 2);
1440                         else
1441                                 timeout = cfs_time_seconds(obd_timeout);
1442                 } else {
1443                         int idx = import_at_get_index(imp,
1444                                 imp->imp_client->cli_request_portal);
1445                         timeout = cfs_time_seconds(
1446                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1447                 }
1448
1449                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1450                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1451                 rc = l_wait_event(imp->imp_recovery_waitq,
1452                                   !ptlrpc_import_in_recovery(imp), &lwi);
1453
1454         }
1455
1456         cfs_spin_lock(&imp->imp_lock);
1457         if (imp->imp_state != LUSTRE_IMP_FULL)
1458                 GOTO(out, 0);
1459
1460         cfs_spin_unlock(&imp->imp_lock);
1461
1462         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1463                                         LUSTRE_OBD_VERSION, rq_opc);
1464         if (req) {
1465                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1466                  * it fails.  We can get through the above with a down server
1467                  * if the client doesn't know the server is gone yet. */
1468                 req->rq_no_resend = 1;
1469
1470                 /* We want client umounts to happen quickly, no matter the
1471                    server state... */
1472                 req->rq_timeout = min_t(int, req->rq_timeout,
1473                                         INITIAL_CONNECT_TIMEOUT);
1474
1475                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1476                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1477                 ptlrpc_request_set_replen(req);
1478                 rc = ptlrpc_queue_wait(req);
1479                 ptlrpc_req_finished(req);
1480         }
1481
1482 set_state:
1483         cfs_spin_lock(&imp->imp_lock);
1484 out:
1485         if (noclose)
1486                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1487         else
1488                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1489         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1490         cfs_spin_unlock(&imp->imp_lock);
1491
1492         RETURN(rc);
1493 }
1494 EXPORT_SYMBOL(ptlrpc_disconnect_import);
1495
1496 void ptlrpc_cleanup_imp(struct obd_import *imp)
1497 {
1498         ENTRY;
1499
1500         cfs_spin_lock(&imp->imp_lock);
1501         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1502         imp->imp_generation++;
1503         cfs_spin_unlock(&imp->imp_lock);
1504         ptlrpc_abort_inflight(imp);
1505
1506         EXIT;
1507 }
1508 EXPORT_SYMBOL(ptlrpc_cleanup_imp);
1509
1510 /* Adaptive Timeout utils */
1511 extern unsigned int at_min, at_max, at_history;
1512
1513 /* Bin into timeslices using AT_BINS bins.
1514    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1515    but still smoothing out a return to normalcy from a slow response.
1516    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1517 int at_measured(struct adaptive_timeout *at, unsigned int val)
1518 {
1519         unsigned int old = at->at_current;
1520         time_t now = cfs_time_current_sec();
1521         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1522
1523         LASSERT(at);
1524         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1525                val, at, now - at->at_binstart, at->at_current,
1526                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1527
1528         if (val == 0)
1529                 /* 0's don't count, because we never want our timeout to
1530                    drop to 0, and because 0 could mean an error */
1531                 return 0;
1532
1533         cfs_spin_lock(&at->at_lock);
1534
1535         if (unlikely(at->at_binstart == 0)) {
1536                 /* Special case to remove default from history */
1537                 at->at_current = val;
1538                 at->at_worst_ever = val;
1539                 at->at_worst_time = now;
1540                 at->at_hist[0] = val;
1541                 at->at_binstart = now;
1542         } else if (now - at->at_binstart < binlimit ) {
1543                 /* in bin 0 */
1544                 at->at_hist[0] = max(val, at->at_hist[0]);
1545                 at->at_current = max(val, at->at_current);
1546         } else {
1547                 int i, shift;
1548                 unsigned int maxv = val;
1549                 /* move bins over */
1550                 shift = (now - at->at_binstart) / binlimit;
1551                 LASSERT(shift > 0);
1552                 for(i = AT_BINS - 1; i >= 0; i--) {
1553                         if (i >= shift) {
1554                                 at->at_hist[i] = at->at_hist[i - shift];
1555                                 maxv = max(maxv, at->at_hist[i]);
1556                         } else {
1557                                 at->at_hist[i] = 0;
1558                         }
1559                 }
1560                 at->at_hist[0] = val;
1561                 at->at_current = maxv;
1562                 at->at_binstart += shift * binlimit;
1563         }
1564
1565         if (at->at_current > at->at_worst_ever) {
1566                 at->at_worst_ever = at->at_current;
1567                 at->at_worst_time = now;
1568         }
1569
1570         if (at->at_flags & AT_FLG_NOHIST)
1571                 /* Only keep last reported val; keeping the rest of the history
1572                    for proc only */
1573                 at->at_current = val;
1574
1575         if (at_max > 0)
1576                 at->at_current =  min(at->at_current, at_max);
1577         at->at_current =  max(at->at_current, at_min);
1578
1579         if (at->at_current != old)
1580                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1581                        "(val=%u) hist %u %u %u %u\n", at,
1582                        old, at->at_current, at->at_current - old, val,
1583                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1584                        at->at_hist[3]);
1585
1586         /* if we changed, report the old value */
1587         old = (at->at_current != old) ? old : 0;
1588
1589         cfs_spin_unlock(&at->at_lock);
1590         return old;
1591 }
1592
1593 /* Find the imp_at index for a given portal; assign if space available */
1594 int import_at_get_index(struct obd_import *imp, int portal)
1595 {
1596         struct imp_at *at = &imp->imp_at;
1597         int i;
1598
1599         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1600                 if (at->iat_portal[i] == portal)
1601                         return i;
1602                 if (at->iat_portal[i] == 0)
1603                         /* unused */
1604                         break;
1605         }
1606
1607         /* Not found in list, add it under a lock */
1608         cfs_spin_lock(&imp->imp_lock);
1609
1610         /* Check unused under lock */
1611         for (; i < IMP_AT_MAX_PORTALS; i++) {
1612                 if (at->iat_portal[i] == portal)
1613                         goto out;
1614                 if (at->iat_portal[i] == 0)
1615                         /* unused */
1616                         break;
1617         }
1618
1619         /* Not enough portals? */
1620         LASSERT(i < IMP_AT_MAX_PORTALS);
1621
1622         at->iat_portal[i] = portal;
1623 out:
1624         cfs_spin_unlock(&imp->imp_lock);
1625         return i;
1626 }