lustre/ptlrpc/import.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  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 static void __import_set_state(struct obd_import *imp,
  63                                enum lustre_imp_state state)
  64 {
  65         imp->imp_state = state;
  66         imp->imp_state_hist[imp->imp_state_hist_idx].ish_state = state;
  67         imp->imp_state_hist[imp->imp_state_hist_idx].ish_time =
  68                 cfs_time_current_sec();
  69         imp->imp_state_hist_idx = (imp->imp_state_hist_idx + 1) %
  70                 IMP_STATE_HIST_LEN;
  71 }
  72
  73 /* A CLOSED import should remain so. */
  74 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  75 do {                                                                           \
  76         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  77                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  78                       imp, obd2cli_tgt(imp->imp_obd),                          \
  79                       ptlrpc_import_state_name(imp->imp_state),                \
  80                       ptlrpc_import_state_name(state));                        \
  81                __import_set_state(imp, state);                                 \
  82         }                                                                      \
  83 } while(0)
  84
  85 #define IMPORT_SET_STATE(imp, state)            \
  86 do {                                            \
  87         spin_lock(&imp->imp_lock);              \
  88         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  89         spin_unlock(&imp->imp_lock);            \
  90 } while(0)
  91
  92
  93 static int ptlrpc_connect_interpret(const struct lu_env *env,
  94                                     struct ptlrpc_request *request,
  95                                     void * data, int rc);
  96 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  97
  98 /* Only this function is allowed to change the import state when it is
  99  * CLOSED. I would rather refcount the import and free it after
 100  * disconnection like we do with exports. To do that, the client_obd
 101  * will need to save the peer info somewhere other than in the import,
 102  * though. */
 103 int ptlrpc_init_import(struct obd_import *imp)
 104 {
 105         spin_lock(&imp->imp_lock);
 106
 107         imp->imp_generation++;
 108         imp->imp_state =  LUSTRE_IMP_NEW;
 109
 110         spin_unlock(&imp->imp_lock);
 111
 112         return 0;
 113 }
 114 EXPORT_SYMBOL(ptlrpc_init_import);
 115
 116 #define UUID_STR "_UUID"
 117 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
 118                       int *uuid_len)
 119 {
 120         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
 121                 ? uuid : uuid + strlen(prefix);
 122
 123         *uuid_len = strlen(*uuid_start);
 124
 125         if (*uuid_len < strlen(UUID_STR))
 126                 return;
 127
 128         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 129                     UUID_STR, strlen(UUID_STR)))
 130                 *uuid_len -= strlen(UUID_STR);
 131 }
 132
 133 /* Returns true if import was FULL, false if import was already not
 134  * connected.
 135  * @imp - import to be disconnected
 136  * @conn_cnt - connection count (epoch) of the request that timed out
 137  *             and caused the disconnection.  In some cases, multiple
 138  *             inflight requests can fail to a single target (e.g. OST
 139  *             bulk requests) and if one has already caused a reconnection
 140  *             (increasing the import->conn_cnt) the older failure should
 141  *             not also cause a reconnection.  If zero it forces a reconnect.
 142  */
 143 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 144 {
 145         int rc = 0;
 146
 147         spin_lock(&imp->imp_lock);
 148
 149         if (imp->imp_state == LUSTRE_IMP_FULL &&
 150             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 151                 char *target_start;
 152                 int   target_len;
 153
 154                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 155                           &target_start, &target_len);
 156
 157                 if (imp->imp_replayable) {
 158                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 159                                "%s was lost; in progress operations using this "
 160                                "service will wait for recovery to complete.\n",
 161                                imp->imp_obd->obd_name, target_len, target_start,
 162                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 163                 } else {
 164                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 165                                "%.*s via nid %s was lost; in progress "
 166                                "operations using this service will fail.\n",
 167                                imp->imp_obd->obd_name,
 168                                target_len, target_start,
 169                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 170                 }
 171                 ptlrpc_deactivate_timeouts(imp);
 172                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 173                 spin_unlock(&imp->imp_lock);
 174
 175                 if (obd_dump_on_timeout)
 176                         libcfs_debug_dumplog();
 177
 178                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 179                 rc = 1;
 180         } else {
 181                 spin_unlock(&imp->imp_lock);
 182                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 183                        imp->imp_client->cli_name, imp,
 184                        (imp->imp_state == LUSTRE_IMP_FULL &&
 185                         imp->imp_conn_cnt > conn_cnt) ?
 186                        "reconnected" : "not connected", imp->imp_conn_cnt,
 187                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 188         }
 189
 190         return rc;
 191 }
 192
 193 /* Must be called with imp_lock held! */
 194 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 195 {
 196         ENTRY;
 197         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 198
 199         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 200         imp->imp_invalid = 1;
 201         imp->imp_generation++;
 202         spin_unlock(&imp->imp_lock);
 203
 204         ptlrpc_abort_inflight(imp);
 205         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 206
 207         EXIT;
 208 }
 209
 210 /*
 211  * This acts as a barrier; all existing requests are rejected, and
 212  * no new requests will be accepted until the import is valid again.
 213  */
 214 void ptlrpc_deactivate_import(struct obd_import *imp)
 215 {
 216         spin_lock(&imp->imp_lock);
 217         ptlrpc_deactivate_and_unlock_import(imp);
 218 }
 219
 220 static unsigned int
 221 ptlrpc_inflight_deadline(struct ptlrpc_request *req, time_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         time_t now = cfs_time_current_sec();
 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         /*
 277          * If this is an invalid MGC connection, then don't bother
 278          * waiting for imp_inflight to drop to 0.
 279          */
 280         if (imp->imp_invalid && imp->imp_recon_bk &&!imp->imp_obd->obd_no_recov)
 281                 goto out;
 282
 283         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 284                 ptlrpc_deactivate_import(imp);
 285
 286         LASSERT(imp->imp_invalid);
 287
 288         /* Wait forever until inflight == 0. We really can't do it another
 289          * way because in some cases we need to wait for very long reply
 290          * unlink. We can't do anything before that because there is really
 291          * no guarantee that some rdma transfer is not in progress right now. */
 292         do {
 293                 /* Calculate max timeout for waiting on rpcs to error
 294                  * out. Use obd_timeout if calculated value is smaller
 295                  * than it. */
 296                 if (!OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
 297                         timeout = ptlrpc_inflight_timeout(imp);
 298                         timeout += timeout / 3;
 299
 300                         if (timeout == 0)
 301                                 timeout = obd_timeout;
 302                 } else {
 303                         /* decrease the interval to increase race condition */
 304                         timeout = 1;
 305                 }
 306
 307                 CDEBUG(D_RPCTRACE,"Sleeping %d sec for inflight to error out\n",
 308                        timeout);
 309
 310                 /* Wait for all requests to error out and call completion
 311                  * callbacks. Cap it at obd_timeout -- these should all
 312                  * have been locally cancelled by ptlrpc_abort_inflight. */
 313                 lwi = LWI_TIMEOUT_INTERVAL(
 314                         cfs_timeout_cap(cfs_time_seconds(timeout)),
 315                         (timeout > 1)?cfs_time_seconds(1):cfs_time_seconds(1)/2,
 316                         NULL, NULL);
 317                 rc = l_wait_event(imp->imp_recovery_waitq,
 318                                 (atomic_read(&imp->imp_inflight) == 0), &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, atomic_read(&imp->imp_inflight));
 324
 325                         spin_lock(&imp->imp_lock);
 326                         if (atomic_read(&imp->imp_inflight) == 0) {
 327                                 int count = atomic_read(&imp->imp_unregistering);
 328
 329                                 /* We know that "unregistering" rpcs only can
 330                                  * survive in sending or delaying lists (they
 331                                  * maybe waiting for long reply unlink in
 332                                  * sluggish nets). Let's check this. If there
 333                                  * is no inflight and unregistering != 0, this
 334                                  * is bug. */
 335                                 LASSERTF(count == 0, "Some RPCs are still "
 336                                          "unregistering: %d\n", count);
 337
 338                                 /* Let's save one loop as soon as inflight have
 339                                  * dropped to zero. No new inflights possible at
 340                                  * this point. */
 341                                 rc = 0;
 342                         } else {
 343                                 list_for_each_safe(tmp, n,
 344                                                    &imp->imp_sending_list) {
 345                                         req = list_entry(tmp,
 346                                                          struct ptlrpc_request,
 347                                                          rq_list);
 348                                         DEBUG_REQ(D_ERROR, req,
 349                                                   "still on sending list");
 350                                 }
 351                                 list_for_each_safe(tmp, n,
 352                                                    &imp->imp_delayed_list) {
 353                                         req = list_entry(tmp,
 354                                                          struct ptlrpc_request,
 355                                                          rq_list);
 356                                         DEBUG_REQ(D_ERROR, req,
 357                                                   "still on delayed list");
 358                                 }
 359
 360                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 361                                        "Network is sluggish? Waiting them "
 362                                        "to error out.\n", cli_tgt,
 363                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 364                                        atomic_read(&imp->imp_unregistering));
 365                         }
 366                         spin_unlock(&imp->imp_lock);
 367                   }
 368         } while (rc != 0);
 369
 370         /*
 371          * Let's additionally check that no new rpcs added to import in
 372          * "invalidate" state.
 373          */
 374         LASSERT(atomic_read(&imp->imp_inflight) == 0);
 375 out:
 376         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 377         sptlrpc_import_flush_all_ctx(imp);
 378
 379         atomic_dec(&imp->imp_inval_count);
 380         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
 381 }
 382
 383 /* unset imp_invalid */
 384 void ptlrpc_activate_import(struct obd_import *imp)
 385 {
 386         struct obd_device *obd = imp->imp_obd;
 387
 388         spin_lock(&imp->imp_lock);
 389         imp->imp_invalid = 0;
 390         ptlrpc_activate_timeouts(imp);
 391         spin_unlock(&imp->imp_lock);
 392         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 393 }
 394
 395 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 396 {
 397         ENTRY;
 398
 399         LASSERT(!imp->imp_dlm_fake);
 400
 401         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 402                 if (!imp->imp_replayable) {
 403                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 404                                "auto-deactivating\n",
 405                                obd2cli_tgt(imp->imp_obd),
 406                                imp->imp_connection->c_remote_uuid.uuid,
 407                                imp->imp_obd->obd_name);
 408                         ptlrpc_deactivate_import(imp);
 409                 }
 410
 411                 CDEBUG(D_HA, "%s: waking up pinger\n",
 412                        obd2cli_tgt(imp->imp_obd));
 413
 414                 spin_lock(&imp->imp_lock);
 415                 imp->imp_force_verify = 1;
 416                 spin_unlock(&imp->imp_lock);
 417
 418                 ptlrpc_pinger_wake_up();
 419         }
 420         EXIT;
 421 }
 422
 423 int ptlrpc_reconnect_import(struct obd_import *imp)
 424 {
 425         ptlrpc_set_import_discon(imp, 0);
 426         /* Force a new connect attempt */
 427         ptlrpc_invalidate_import(imp);
 428         /* Do a fresh connect next time by zeroing the handle */
 429         ptlrpc_disconnect_import(imp, 1);
 430         /* Wait for all invalidate calls to finish */
 431         if (atomic_read(&imp->imp_inval_count) > 0) {
 432                 int rc;
 433                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 434                 rc = l_wait_event(imp->imp_recovery_waitq,
 435                                   (atomic_read(&imp->imp_inval_count) == 0),
 436                                   &lwi);
 437                 if (rc)
 438                         CERROR("Interrupted, inval=%d\n",
 439                                atomic_read(&imp->imp_inval_count));
 440         }
 441
 442         /* Allow reconnect attempts */
 443         imp->imp_obd->obd_no_recov = 0;
 444         /* Remove 'invalid' flag */
 445         ptlrpc_activate_import(imp);
 446         /* Attempt a new connect */
 447         ptlrpc_recover_import(imp, NULL);
 448         return 0;
 449 }
 450
 451 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 452
 453 static int import_select_connection(struct obd_import *imp)
 454 {
 455         struct obd_import_conn *imp_conn = NULL, *conn;
 456         struct obd_export *dlmexp;
 457         int tried_all = 1;
 458         ENTRY;
 459
 460         spin_lock(&imp->imp_lock);
 461
 462         if (list_empty(&imp->imp_conn_list)) {
 463                 CERROR("%s: no connections available\n",
 464                         imp->imp_obd->obd_name);
 465                 spin_unlock(&imp->imp_lock);
 466                 RETURN(-EINVAL);
 467         }
 468
 469         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 470                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 471                        imp->imp_obd->obd_name,
 472                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 473                        conn->oic_last_attempt);
 474                 /* Don't thrash connections */
 475                 if (cfs_time_before_64(cfs_time_current_64(),
 476                                      conn->oic_last_attempt +
 477                                      cfs_time_seconds(CONNECTION_SWITCH_MIN))) {
 478                         continue;
 479                 }
 480
 481                 /* If we have not tried this connection since
 482                    the last successful attempt, go with this one */
 483                 if ((conn->oic_last_attempt == 0) ||
 484                     cfs_time_beforeq_64(conn->oic_last_attempt,
 485                                        imp->imp_last_success_conn)) {
 486                         imp_conn = conn;
 487                         tried_all = 0;
 488                         break;
 489                 }
 490
 491                 /* If all of the connections have already been tried
 492                    since the last successful connection; just choose the
 493                    least recently used */
 494                 if (!imp_conn)
 495                         imp_conn = conn;
 496                 else if (cfs_time_before_64(conn->oic_last_attempt,
 497                                             imp_conn->oic_last_attempt))
 498                         imp_conn = conn;
 499         }
 500
 501         /* if not found, simply choose the current one */
 502         if (!imp_conn || imp->imp_force_reconnect) {
 503                 LASSERT(imp->imp_conn_current);
 504                 imp_conn = imp->imp_conn_current;
 505                 tried_all = 0;
 506         }
 507         LASSERT(imp_conn->oic_conn);
 508
 509         /* If we've tried everything, and we're back to the beginning of the
 510            list, increase our timeout and try again. It will be reset when
 511            we do finally connect. (FIXME: really we should wait for all network
 512            state associated with the last connection attempt to drain before
 513            trying to reconnect on it.) */
 514         if (tried_all && (imp->imp_conn_list.next == &imp_conn->oic_item) &&
 515             !imp->imp_recon_bk /* not retrying */) {
 516                 if (at_get(&imp->imp_at.iat_net_latency) <
 517                     CONNECTION_SWITCH_MAX) {
 518                         at_add(&imp->imp_at.iat_net_latency,
 519                                at_get(&imp->imp_at.iat_net_latency) +
 520                                CONNECTION_SWITCH_INC);
 521                 }
 522                 LASSERT(imp_conn->oic_last_attempt);
 523                 CWARN("%s: tried all connections, increasing latency to %ds\n",
 524                       imp->imp_obd->obd_name,
 525                       at_get(&imp->imp_at.iat_net_latency));
 526         }
 527
 528         imp_conn->oic_last_attempt = cfs_time_current_64();
 529
 530         /* switch connection, don't mind if it's same as the current one */
 531         if (imp->imp_connection)
 532                 ptlrpc_connection_put(imp->imp_connection);
 533         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 534
 535         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 536         LASSERT(dlmexp != NULL);
 537         if (dlmexp->exp_connection)
 538                 ptlrpc_connection_put(dlmexp->exp_connection);
 539         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 540         class_export_put(dlmexp);
 541
 542         if (imp->imp_conn_current != imp_conn) {
 543                 if (imp->imp_conn_current)
 544                         CDEBUG(D_HA, "Changing connection for %s to %s/%s\n",
 545                                imp->imp_obd->obd_name, imp_conn->oic_uuid.uuid,
 546                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 547                 imp->imp_conn_current = imp_conn;
 548         }
 549
 550         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 551                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 552                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 553
 554         spin_unlock(&imp->imp_lock);
 555
 556         RETURN(0);
 557 }
 558
 559 /*
 560  * must be called under imp_lock
 561  */
 562 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 563 {
 564         struct ptlrpc_request *req;
 565         struct list_head *tmp;
 566
 567         if (list_empty(&imp->imp_replay_list))
 568                 return 0;
 569         tmp = imp->imp_replay_list.next;
 570         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 571         *transno = req->rq_transno;
 572         if (req->rq_transno == 0) {
 573                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 574                 LBUG();
 575         }
 576
 577         return 1;
 578 }
 579
 580 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 581 {
 582         struct obd_device *obd = imp->imp_obd;
 583         int initial_connect = 0;
 584         int set_transno = 0;
 585         __u64 committed_before_reconnect = 0;
 586         struct ptlrpc_request *request;
 587         char *bufs[] = { NULL,
 588                          obd2cli_tgt(imp->imp_obd),
 589                          obd->obd_uuid.uuid,
 590                          (char *)&imp->imp_dlm_handle,
 591                          (char *)&imp->imp_connect_data };
 592         struct ptlrpc_connect_async_args *aa;
 593         int rc;
 594         ENTRY;
 595
 596         spin_lock(&imp->imp_lock);
 597         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 598                 spin_unlock(&imp->imp_lock);
 599                 CERROR("can't connect to a closed import\n");
 600                 RETURN(-EINVAL);
 601         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 602                 spin_unlock(&imp->imp_lock);
 603                 CERROR("already connected\n");
 604                 RETURN(0);
 605         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 606                 spin_unlock(&imp->imp_lock);
 607                 CERROR("already connecting\n");
 608                 RETURN(-EALREADY);
 609         }
 610
 611         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 612
 613         imp->imp_conn_cnt++;
 614         imp->imp_resend_replay = 0;
 615
 616         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 617                 initial_connect = 1;
 618         else
 619                 committed_before_reconnect = imp->imp_peer_committed_transno;
 620
 621         set_transno = ptlrpc_first_transno(imp,
 622                                            &imp->imp_connect_data.ocd_transno);
 623         spin_unlock(&imp->imp_lock);
 624
 625         if (new_uuid) {
 626                 struct obd_uuid uuid;
 627
 628                 obd_str2uuid(&uuid, new_uuid);
 629                 rc = import_set_conn_priority(imp, &uuid);
 630                 if (rc)
 631                         GOTO(out, rc);
 632         }
 633
 634         rc = import_select_connection(imp);
 635         if (rc)
 636                 GOTO(out, rc);
 637
 638         /* last in connection list */
 639         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 640                 if (imp->imp_initial_recov_bk && initial_connect) {
 641                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 642                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 643                         /* Don't retry if connect fails */
 644                         rc = 0;
 645                         obd_set_info_async(obd->obd_self_export,
 646                                            sizeof(KEY_INIT_RECOV),
 647                                            KEY_INIT_RECOV,
 648                                            sizeof(rc), &rc, NULL);
 649                 }
 650                 if (imp->imp_recon_bk) {
 651                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 652                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 653                         spin_lock(&imp->imp_lock);
 654                         imp->imp_last_recon = 1;
 655                         spin_unlock(&imp->imp_lock);
 656                 }
 657         }
 658
 659         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 660         if (rc)
 661                 GOTO(out, rc);
 662
 663         /* Reset connect flags to the originally requested flags, in case
 664          * the server is updated on-the-fly we will get the new features. */
 665         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 666         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 667
 668         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 669                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 670         if (rc)
 671                 GOTO(out, rc);
 672
 673         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 674         if (request == NULL)
 675                 GOTO(out, rc = -ENOMEM);
 676
 677         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 678                                       imp->imp_connect_op, bufs, NULL);
 679         if (rc) {
 680                 ptlrpc_request_free(request);
 681                 GOTO(out, rc);
 682         }
 683
 684         /* Report the rpc service time to the server so that it knows how long
 685          * to wait for clients to join recovery */
 686         lustre_msg_set_service_time(request->rq_reqmsg,
 687                                     at_timeout2est(request->rq_timeout));
 688
 689         /* The amount of time we give the server to process the connect req.
 690          * import_select_connection will increase the net latency on
 691          * repeated reconnect attempts to cover slow networks.
 692          * We override/ignore the server rpc completion estimate here,
 693          * which may be large if this is a reconnect attempt */
 694         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 695         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 696
 697 #ifndef __KERNEL__
 698         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 699 #endif
 700         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 701
 702         request->rq_no_resend = request->rq_no_delay = 1;
 703         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 704         /* Allow a slightly larger reply for future growth compatibility */
 705         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 706                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 707         ptlrpc_request_set_replen(request);
 708         request->rq_interpret_reply = ptlrpc_connect_interpret;
 709
 710         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 711         aa = ptlrpc_req_async_args(request);
 712         memset(aa, 0, sizeof *aa);
 713
 714         aa->pcaa_peer_committed = committed_before_reconnect;
 715         aa->pcaa_initial_connect = initial_connect;
 716
 717         if (aa->pcaa_initial_connect) {
 718                 spin_lock(&imp->imp_lock);
 719                 imp->imp_replayable = 1;
 720                 spin_unlock(&imp->imp_lock);
 721                 lustre_msg_add_op_flags(request->rq_reqmsg,
 722                                         MSG_CONNECT_INITIAL);
 723         }
 724
 725         if (set_transno)
 726                 lustre_msg_add_op_flags(request->rq_reqmsg,
 727                                         MSG_CONNECT_TRANSNO);
 728
 729         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 730         ptlrpcd_add_req(request, PSCOPE_OTHER);
 731         rc = 0;
 732 out:
 733         if (rc != 0) {
 734                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 735         }
 736
 737         RETURN(rc);
 738 }
 739 EXPORT_SYMBOL(ptlrpc_connect_import);
 740
 741 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 742 {
 743 #ifdef __KERNEL__
 744         struct obd_import_conn *imp_conn;
 745 #endif
 746         int wake_pinger = 0;
 747
 748         ENTRY;
 749
 750         spin_lock(&imp->imp_lock);
 751         if (list_empty(&imp->imp_conn_list))
 752                 GOTO(unlock, 0);
 753
 754 #ifdef __KERNEL__
 755         imp_conn = list_entry(imp->imp_conn_list.prev,
 756                               struct obd_import_conn,
 757                               oic_item);
 758
 759         /* XXX: When the failover node is the primary node, it is possible
 760          * to have two identical connections in imp_conn_list. We must
 761          * compare not conn's pointers but NIDs, otherwise we can defeat
 762          * connection throttling. (See bug 14774.) */
 763         if (imp->imp_conn_current->oic_conn->c_peer.nid !=
 764                                 imp_conn->oic_conn->c_peer.nid) {
 765                 ptlrpc_ping_import_soon(imp);
 766                 wake_pinger = 1;
 767         }
 768 #else
 769         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 770         wake_pinger = 1;
 771 #endif
 772
 773  unlock:
 774         spin_unlock(&imp->imp_lock);
 775
 776         if (wake_pinger)
 777                 ptlrpc_pinger_wake_up();
 778
 779         EXIT;
 780 }
 781
 782 static int ptlrpc_busy_reconnect(int rc)
 783 {
 784         return (rc == -EBUSY) || (rc == -EAGAIN);
 785 }
 786
 787
 788 static int ptlrpc_connect_interpret(const struct lu_env *env,
 789                                     struct ptlrpc_request *request,
 790                                     void *data, int rc)
 791 {
 792         struct ptlrpc_connect_async_args *aa = data;
 793         struct obd_import *imp = request->rq_import;
 794         struct client_obd *cli = &imp->imp_obd->u.cli;
 795         struct lustre_handle old_hdl;
 796         __u64 old_connect_flags;
 797         int msg_flags;
 798         ENTRY;
 799
 800         spin_lock(&imp->imp_lock);
 801         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 802                 spin_unlock(&imp->imp_lock);
 803                 RETURN(0);
 804         }
 805
 806         if (rc) {
 807                 /* if this reconnect to busy export - not need select new target
 808                  * for connecting*/
 809                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 810                 spin_unlock(&imp->imp_lock);
 811                 GOTO(out, rc);
 812         }
 813
 814         LASSERT(imp->imp_conn_current);
 815
 816         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 817
 818         /* All imports are pingable */
 819         imp->imp_pingable = 1;
 820         imp->imp_force_reconnect = 0;
 821
 822         if (aa->pcaa_initial_connect) {
 823                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 824                         imp->imp_replayable = 1;
 825                         spin_unlock(&imp->imp_lock);
 826                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 827                                obd2cli_tgt(imp->imp_obd));
 828                 } else {
 829                         imp->imp_replayable = 0;
 830                         spin_unlock(&imp->imp_lock);
 831                 }
 832
 833                 /* if applies, adjust the imp->imp_msg_magic here
 834                  * according to reply flags */
 835
 836                 imp->imp_remote_handle =
 837                                 *lustre_msg_get_handle(request->rq_repmsg);
 838
 839                 /* Initial connects are allowed for clients with non-random
 840                  * uuids when servers are in recovery.  Simply signal the
 841                  * servers replay is complete and wait in REPLAY_WAIT. */
 842                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 843                         CDEBUG(D_HA, "connect to %s during recovery\n",
 844                                obd2cli_tgt(imp->imp_obd));
 845                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 846                 } else {
 847                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 848                         ptlrpc_activate_import(imp);
 849                 }
 850
 851                 GOTO(finish, rc = 0);
 852         } else {
 853                 spin_unlock(&imp->imp_lock);
 854         }
 855
 856         /* Determine what recovery state to move the import to. */
 857         if (MSG_CONNECT_RECONNECT & msg_flags) {
 858                 memset(&old_hdl, 0, sizeof(old_hdl));
 859                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 860                             sizeof (old_hdl))) {
 861                         CERROR("%s@%s didn't like our handle "LPX64
 862                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 863                                imp->imp_connection->c_remote_uuid.uuid,
 864                                imp->imp_dlm_handle.cookie);
 865                         GOTO(out, rc = -ENOTCONN);
 866                 }
 867
 868                 if (memcmp(&imp->imp_remote_handle,
 869                            lustre_msg_get_handle(request->rq_repmsg),
 870                            sizeof(imp->imp_remote_handle))) {
 871                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 872                                 D_HA : D_WARNING;
 873
 874                         /* Bug 16611/14775: if server handle have changed,
 875                          * that means some sort of disconnection happened.
 876                          * If the server is not in recovery, that also means it
 877                          * already erased all of our state because of previous
 878                          * eviction. If it is in recovery - we are safe to
 879                          * participate since we can reestablish all of our state
 880                          * with server again */
 881                         CDEBUG(level,"%s@%s changed server handle from "
 882                                      LPX64" to "LPX64"%s\n",
 883                                      obd2cli_tgt(imp->imp_obd),
 884                                      imp->imp_connection->c_remote_uuid.uuid,
 885                                      imp->imp_remote_handle.cookie,
 886                                      lustre_msg_get_handle(request->rq_repmsg)->
 887                                                                         cookie,
 888                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
 889                                          " but is still in recovery" : "");
 890
 891                         imp->imp_remote_handle =
 892                                      *lustre_msg_get_handle(request->rq_repmsg);
 893
 894                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 895                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 896                                 GOTO(finish, rc = 0);
 897                         }
 898
 899                 } else {
 900                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 901                                obd2cli_tgt(imp->imp_obd),
 902                                imp->imp_connection->c_remote_uuid.uuid);
 903                 }
 904
 905                 if (imp->imp_invalid) {
 906                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 907                                "marking evicted\n", imp->imp_obd->obd_name);
 908                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 909                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 910                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 911                                imp->imp_obd->obd_name,
 912                                obd2cli_tgt(imp->imp_obd));
 913
 914                         spin_lock(&imp->imp_lock);
 915                         imp->imp_resend_replay = 1;
 916                         spin_unlock(&imp->imp_lock);
 917
 918                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 919                 } else {
 920                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 921                 }
 922         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 923                 LASSERT(imp->imp_replayable);
 924                 imp->imp_remote_handle =
 925                                 *lustre_msg_get_handle(request->rq_repmsg);
 926                 imp->imp_last_replay_transno = 0;
 927                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 928         } else {
 929                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 930                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 931                 imp->imp_remote_handle =
 932                                 *lustre_msg_get_handle(request->rq_repmsg);
 933                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 934         }
 935
 936         /* Sanity checks for a reconnected import. */
 937         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 938                 CERROR("imp_replayable flag does not match server "
 939                        "after reconnect. We should LBUG right here.\n");
 940         }
 941
 942         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 943             aa->pcaa_peer_committed) {
 944                 CERROR("%s went back in time (transno "LPD64
 945                        " was previously committed, server now claims "LPD64
 946                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 947                        "id=9646\n",
 948                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 949                        lustre_msg_get_last_committed(request->rq_repmsg));
 950         }
 951
 952 finish:
 953         rc = ptlrpc_import_recovery_state_machine(imp);
 954         if (rc != 0) {
 955                 if (rc == -ENOTCONN) {
 956                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 957                                "invalidating and reconnecting\n",
 958                                obd2cli_tgt(imp->imp_obd),
 959                                imp->imp_connection->c_remote_uuid.uuid);
 960                         ptlrpc_connect_import(imp, NULL);
 961                         RETURN(0);
 962                 }
 963         } else {
 964                 struct obd_connect_data *ocd;
 965                 struct obd_export *exp;
 966                 int ret;
 967                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 968                                            RCL_SERVER);
 969                 /* server replied obd_connect_data is always bigger */
 970                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 971                                                    &RMF_CONNECT_DATA, ret);
 972
 973                 spin_lock(&imp->imp_lock);
 974                 list_del(&imp->imp_conn_current->oic_item);
 975                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 976                 imp->imp_last_success_conn =
 977                         imp->imp_conn_current->oic_last_attempt;
 978
 979                 if (ocd == NULL) {
 980                         spin_unlock(&imp->imp_lock);
 981                         CERROR("Wrong connect data from server\n");
 982                         rc = -EPROTO;
 983                         GOTO(out, rc);
 984                 }
 985
 986                 imp->imp_connect_data = *ocd;
 987
 988                 exp = class_conn2export(&imp->imp_dlm_handle);
 989                 spin_unlock(&imp->imp_lock);
 990
 991                 /* check that server granted subset of flags we asked for. */
 992                 LASSERTF((ocd->ocd_connect_flags &
 993                           imp->imp_connect_flags_orig) ==
 994                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 995                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 996
 997                 if (!exp) {
 998                         /* This could happen if export is cleaned during the
 999                            connect attempt */
1000                         CERROR("Missing export for %s\n",
1001                                imp->imp_obd->obd_name);
1002                         GOTO(out, rc = -ENODEV);
1003                 }
1004                 old_connect_flags = exp->exp_connect_flags;
1005                 exp->exp_connect_flags = ocd->ocd_connect_flags;
1006                 imp->imp_obd->obd_self_export->exp_connect_flags =
1007                                                         ocd->ocd_connect_flags;
1008                 class_export_put(exp);
1009
1010                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1011
1012                 if (!ocd->ocd_ibits_known &&
1013                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
1014                         CERROR("Inodebits aware server returned zero compatible"
1015                                " bits?\n");
1016
1017                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1018                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1019                                         LUSTRE_VERSION_OFFSET_WARN ||
1020                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1021                                         LUSTRE_VERSION_OFFSET_WARN)) {
1022                         /* Sigh, some compilers do not like #ifdef in the middle
1023                            of macro arguments */
1024 #ifdef __KERNEL__
1025                         const char *older =
1026                                 "older. Consider upgrading this client";
1027 #else
1028                         const char *older =
1029                                 "older. Consider recompiling this application";
1030 #endif
1031                         const char *newer = "newer than client version";
1032
1033                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1034                                       "is much %s (%s)\n",
1035                                       obd2cli_tgt(imp->imp_obd),
1036                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1037                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1038                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1039                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1040                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1041                                       newer : older, LUSTRE_VERSION_STRING);
1042                 }
1043
1044                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1045                         /* We sent to the server ocd_cksum_types with bits set
1046                          * for algorithms we understand. The server masked off
1047                          * the checksum types it doesn't support */
1048                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
1049                                 LCONSOLE_WARN("The negotiation of the checksum "
1050                                               "alogrithm to use with server %s "
1051                                               "failed (%x/%x), disabling "
1052                                               "checksums\n",
1053                                               obd2cli_tgt(imp->imp_obd),
1054                                               ocd->ocd_cksum_types,
1055                                               OBD_CKSUM_ALL);
1056                                 cli->cl_checksum = 0;
1057                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1058                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1059                         } else {
1060                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1061
1062                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
1063                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
1064                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
1065                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
1066                                 else
1067                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1068                         }
1069                 } else {
1070                         /* The server does not support OBD_CONNECT_CKSUM.
1071                          * Enforce CRC32 for backward compatibility*/
1072                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1073                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1074                 }
1075
1076                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
1077                         cli->cl_max_pages_per_rpc =
1078                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1079                 }
1080
1081                 /* Reset ns_connect_flags only for initial connect. It might be
1082                  * changed in while using FS and if we reset it in reconnect
1083                  * this leads to losing user settings done before such as
1084                  * disable lru_resize, etc. */
1085                 if (old_connect_flags != exp->exp_connect_flags ||
1086                     aa->pcaa_initial_connect) {
1087                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1088                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1089                               ocd->ocd_connect_flags);
1090                         imp->imp_obd->obd_namespace->ns_connect_flags =
1091                                 ocd->ocd_connect_flags;
1092                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1093                                 ocd->ocd_connect_flags;
1094                 }
1095
1096                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1097                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1098                         /* We need a per-message support flag, because
1099                            a. we don't know if the incoming connect reply
1100                               supports AT or not (in reply_in_callback)
1101                               until we unpack it.
1102                            b. failovered server means export and flags are gone
1103                               (in ptlrpc_send_reply).
1104                            Can only be set when we know AT is supported at
1105                            both ends */
1106                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1107                 else
1108                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1109
1110                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1111                         (cli->cl_max_pages_per_rpc > 0));
1112         }
1113
1114 out:
1115         if (rc != 0) {
1116                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1117                 spin_lock(&imp->imp_lock);
1118                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
1119                     (request->rq_import_generation == imp->imp_generation))
1120                         ptlrpc_deactivate_and_unlock_import(imp);
1121                 else
1122                         spin_unlock(&imp->imp_lock);
1123
1124                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
1125                     (rc == -EACCES)) {
1126                         /*
1127                          * Give up trying to reconnect
1128                          * EACCES means client has no permission for connection
1129                          */
1130                         imp->imp_obd->obd_no_recov = 1;
1131                         ptlrpc_deactivate_import(imp);
1132                 }
1133
1134                 if (rc == -EPROTO) {
1135                         struct obd_connect_data *ocd;
1136
1137                         /* reply message might not be ready */
1138                         if (request->rq_repmsg == NULL)
1139                                 RETURN(-EPROTO);
1140
1141                         ocd = req_capsule_server_get(&request->rq_pill,
1142                                                      &RMF_CONNECT_DATA);
1143                         if (ocd &&
1144                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1145                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1146                            /* Actually servers are only supposed to refuse
1147                               connection from liblustre clients, so we should
1148                               never see this from VFS context */
1149                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1150                                         "(%d.%d.%d.%d)"
1151                                         " refused connection from this client "
1152                                         "with an incompatible version (%s).  "
1153                                         "Client must be recompiled\n",
1154                                         obd2cli_tgt(imp->imp_obd),
1155                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1156                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1157                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1158                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1159                                         LUSTRE_VERSION_STRING);
1160                                 ptlrpc_deactivate_import(imp);
1161                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1162                         }
1163                         RETURN(-EPROTO);
1164                 }
1165
1166                 ptlrpc_maybe_ping_import_soon(imp);
1167
1168                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1169                        obd2cli_tgt(imp->imp_obd),
1170                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1171         }
1172
1173         spin_lock(&imp->imp_lock);
1174         imp->imp_last_recon = 0;
1175         spin_unlock(&imp->imp_lock);
1176
1177         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1178         RETURN(rc);
1179 }
1180
1181 static int completed_replay_interpret(const struct lu_env *env,
1182                                       struct ptlrpc_request *req,
1183                                       void * data, int rc)
1184 {
1185         ENTRY;
1186         atomic_dec(&req->rq_import->imp_replay_inflight);
1187         if (req->rq_status == 0 &&
1188             !req->rq_import->imp_vbr_failed) {
1189                 ptlrpc_import_recovery_state_machine(req->rq_import);
1190         } else {
1191                 if (req->rq_import->imp_vbr_failed) {
1192                         CDEBUG(D_WARNING,
1193                                "%s: version recovery fails, reconnecting\n",
1194                                req->rq_import->imp_obd->obd_name);
1195                         spin_lock(&req->rq_import->imp_lock);
1196                         req->rq_import->imp_vbr_failed = 0;
1197                         spin_unlock(&req->rq_import->imp_lock);
1198                 } else {
1199                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1200                                      "reconnecting\n",
1201                                req->rq_import->imp_obd->obd_name,
1202                                req->rq_status);
1203                 }
1204                 ptlrpc_connect_import(req->rq_import, NULL);
1205         }
1206
1207         RETURN(0);
1208 }
1209
1210 static int signal_completed_replay(struct obd_import *imp)
1211 {
1212         struct ptlrpc_request *req;
1213         ENTRY;
1214
1215         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
1216         atomic_inc(&imp->imp_replay_inflight);
1217
1218         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1219                                         OBD_PING);
1220         if (req == NULL) {
1221                 atomic_dec(&imp->imp_replay_inflight);
1222                 RETURN(-ENOMEM);
1223         }
1224
1225         ptlrpc_request_set_replen(req);
1226         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1227         lustre_msg_add_flags(req->rq_reqmsg,
1228                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1229         if (AT_OFF)
1230                 req->rq_timeout *= 3;
1231         req->rq_interpret_reply = completed_replay_interpret;
1232
1233         ptlrpcd_add_req(req, PSCOPE_OTHER);
1234         RETURN(0);
1235 }
1236
1237 #ifdef __KERNEL__
1238 static int ptlrpc_invalidate_import_thread(void *data)
1239 {
1240         struct obd_import *imp = data;
1241
1242         ENTRY;
1243
1244         cfs_daemonize_ctxt("ll_imp_inval");
1245
1246         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1247                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1248                imp->imp_connection->c_remote_uuid.uuid);
1249
1250         ptlrpc_invalidate_import(imp);
1251
1252         if (obd_dump_on_eviction) {
1253                 CERROR("dump the log upon eviction\n");
1254                 libcfs_debug_dumplog();
1255         }
1256
1257         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1258         ptlrpc_import_recovery_state_machine(imp);
1259
1260         class_import_put(imp);
1261         RETURN(0);
1262 }
1263 #endif
1264
1265 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1266 {
1267         int rc = 0;
1268         int inflight;
1269         char *target_start;
1270         int target_len;
1271
1272         ENTRY;
1273         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1274                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1275                           &target_start, &target_len);
1276                 /* Don't care about MGC eviction */
1277                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1278                            LUSTRE_MGC_NAME) != 0) {
1279                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1280                                            "%.*s; in progress operations using "
1281                                            "this service will fail.\n",
1282                                            target_len, target_start);
1283                 }
1284                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1285                        obd2cli_tgt(imp->imp_obd),
1286                        imp->imp_connection->c_remote_uuid.uuid);
1287
1288 #ifdef __KERNEL__
1289                 /* bug 17802:  XXX client_disconnect_export vs connect request
1290                  * race. if client will evicted at this time, we start
1291                  * invalidate thread without reference to import and import can
1292                  * be freed at same time. */
1293                 class_import_get(imp);
1294                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1295                                        CLONE_VM | CLONE_FILES);
1296                 if (rc < 0) {
1297                         class_import_put(imp);
1298                         CERROR("error starting invalidate thread: %d\n", rc);
1299                 } else {
1300                         rc = 0;
1301                 }
1302                 RETURN(rc);
1303 #else
1304                 ptlrpc_invalidate_import(imp);
1305
1306                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1307 #endif
1308         }
1309
1310         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1311                 CDEBUG(D_HA, "replay requested by %s\n",
1312                        obd2cli_tgt(imp->imp_obd));
1313                 rc = ptlrpc_replay_next(imp, &inflight);
1314                 if (inflight == 0 &&
1315                     atomic_read(&imp->imp_replay_inflight) == 0) {
1316                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1317                         rc = ldlm_replay_locks(imp);
1318                         if (rc)
1319                                 GOTO(out, rc);
1320                 }
1321                 rc = 0;
1322         }
1323
1324         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1325                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1326                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1327                         rc = signal_completed_replay(imp);
1328                         if (rc)
1329                                 GOTO(out, rc);
1330                 }
1331
1332         }
1333
1334         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1335                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1336                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1337                 }
1338         }
1339
1340         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1341                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1342                        obd2cli_tgt(imp->imp_obd),
1343                        imp->imp_connection->c_remote_uuid.uuid);
1344
1345                 rc = ptlrpc_resend(imp);
1346                 if (rc)
1347                         GOTO(out, rc);
1348                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1349                 ptlrpc_activate_import(imp);
1350
1351                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1352                           &target_start, &target_len);
1353                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1354                               "using nid %s.\n", imp->imp_obd->obd_name,
1355                               target_len, target_start,
1356                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1357         }
1358
1359         if (imp->imp_state == LUSTRE_IMP_FULL) {
1360                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1361                 ptlrpc_wake_delayed(imp);
1362         }
1363
1364 out:
1365         RETURN(rc);
1366 }
1367
1368 static int back_to_sleep(void *unused)
1369 {
1370         return 0;
1371 }
1372
1373 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1374 {
1375         struct ptlrpc_request *req;
1376         int rq_opc, rc = 0;
1377         int nowait = imp->imp_obd->obd_force;
1378         ENTRY;
1379
1380         if (nowait)
1381                 GOTO(set_state, rc);
1382
1383         switch (imp->imp_connect_op) {
1384         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1385         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1386         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1387         default:
1388                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1389                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1390                 RETURN(-EINVAL);
1391         }
1392
1393         if (ptlrpc_import_in_recovery(imp)) {
1394                 struct l_wait_info lwi;
1395                 cfs_duration_t timeout;
1396
1397
1398                 if (AT_OFF) {
1399                         if (imp->imp_server_timeout)
1400                                 timeout = cfs_time_seconds(obd_timeout / 2);
1401                         else
1402                                 timeout = cfs_time_seconds(obd_timeout);
1403                 } else {
1404                         int idx = import_at_get_index(imp,
1405                                 imp->imp_client->cli_request_portal);
1406                         timeout = cfs_time_seconds(
1407                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1408                 }
1409
1410                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1411                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1412                 rc = l_wait_event(imp->imp_recovery_waitq,
1413                                   !ptlrpc_import_in_recovery(imp), &lwi);
1414
1415         }
1416
1417         spin_lock(&imp->imp_lock);
1418         if (imp->imp_state != LUSTRE_IMP_FULL)
1419                 GOTO(out, 0);
1420
1421         spin_unlock(&imp->imp_lock);
1422
1423         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1424                                         LUSTRE_OBD_VERSION, rq_opc);
1425         if (req) {
1426                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1427                  * it fails.  We can get through the above with a down server
1428                  * if the client doesn't know the server is gone yet. */
1429                 req->rq_no_resend = 1;
1430
1431 #ifndef CRAY_XT3
1432                 /* We want client umounts to happen quickly, no matter the
1433                    server state... */
1434                 req->rq_timeout = min_t(int, req->rq_timeout,
1435                                         INITIAL_CONNECT_TIMEOUT);
1436 #else
1437                 /* ... but we always want liblustre clients to nicely
1438                    disconnect, so only use the adaptive value. */
1439                 if (AT_OFF)
1440                         req->rq_timeout = obd_timeout / 3;
1441 #endif
1442
1443                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1444                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1445                 ptlrpc_request_set_replen(req);
1446                 rc = ptlrpc_queue_wait(req);
1447                 ptlrpc_req_finished(req);
1448         }
1449
1450 set_state:
1451         spin_lock(&imp->imp_lock);
1452 out:
1453         if (noclose)
1454                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1455         else
1456                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1457         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1458         /* Try all connections in the future - bz 12758 */
1459         imp->imp_last_recon = 0;
1460         spin_unlock(&imp->imp_lock);
1461
1462         RETURN(rc);
1463 }
1464
1465 void ptlrpc_cleanup_imp(struct obd_import *imp)
1466 {
1467         ENTRY;
1468
1469         spin_lock(&imp->imp_lock);
1470         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1471         imp->imp_generation++;
1472         spin_unlock(&imp->imp_lock);
1473         ptlrpc_abort_inflight(imp);
1474
1475         EXIT;
1476 }
1477
1478 /* Adaptive Timeout utils */
1479 extern unsigned int at_min, at_max, at_history;
1480
1481 /* Bin into timeslices using AT_BINS bins.
1482    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1483    but still smoothing out a return to normalcy from a slow response.
1484    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1485 int at_add(struct adaptive_timeout *at, unsigned int val)
1486 {
1487         unsigned int old = at->at_current;
1488         time_t now = cfs_time_current_sec();
1489         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1490
1491         LASSERT(at);
1492         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1493                val, at, now - at->at_binstart, at->at_current,
1494                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1495
1496         if (val == 0)
1497                 /* 0's don't count, because we never want our timeout to
1498                    drop to 0, and because 0 could mean an error */
1499                 return 0;
1500
1501         spin_lock(&at->at_lock);
1502
1503         if (unlikely(at->at_binstart == 0)) {
1504                 /* Special case to remove default from history */
1505                 at->at_current = val;
1506                 at->at_worst_ever = val;
1507                 at->at_worst_time = now;
1508                 at->at_hist[0] = val;
1509                 at->at_binstart = now;
1510         } else if (now - at->at_binstart < binlimit ) {
1511                 /* in bin 0 */
1512                 at->at_hist[0] = max(val, at->at_hist[0]);
1513                 at->at_current = max(val, at->at_current);
1514         } else {
1515                 int i, shift;
1516                 unsigned int maxv = val;
1517                 /* move bins over */
1518                 shift = (now - at->at_binstart) / binlimit;
1519                 LASSERT(shift > 0);
1520                 for(i = AT_BINS - 1; i >= 0; i--) {
1521                         if (i >= shift) {
1522                                 at->at_hist[i] = at->at_hist[i - shift];
1523                                 maxv = max(maxv, at->at_hist[i]);
1524                         } else {
1525                                 at->at_hist[i] = 0;
1526                         }
1527                 }
1528                 at->at_hist[0] = val;
1529                 at->at_current = maxv;
1530                 at->at_binstart += shift * binlimit;
1531         }
1532
1533         if (at->at_current > at->at_worst_ever) {
1534                 at->at_worst_ever = at->at_current;
1535                 at->at_worst_time = now;
1536         }
1537
1538         if (at->at_flags & AT_FLG_NOHIST)
1539                 /* Only keep last reported val; keeping the rest of the history
1540                    for proc only */
1541                 at->at_current = val;
1542
1543         if (at_max > 0)
1544                 at->at_current =  min(at->at_current, at_max);
1545         at->at_current =  max(at->at_current, at_min);
1546
1547         if (at->at_current != old)
1548                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1549                        "(val=%u) hist %u %u %u %u\n", at,
1550                        old, at->at_current, at->at_current - old, val,
1551                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1552                        at->at_hist[3]);
1553
1554         /* if we changed, report the old value */
1555         old = (at->at_current != old) ? old : 0;
1556
1557         spin_unlock(&at->at_lock);
1558         return old;
1559 }
1560
1561 /* Find the imp_at index for a given portal; assign if space available */
1562 int import_at_get_index(struct obd_import *imp, int portal)
1563 {
1564         struct imp_at *at = &imp->imp_at;
1565         int i;
1566
1567         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1568                 if (at->iat_portal[i] == portal)
1569                         return i;
1570                 if (at->iat_portal[i] == 0)
1571                         /* unused */
1572                         break;
1573         }
1574
1575         /* Not found in list, add it under a lock */
1576         spin_lock(&imp->imp_lock);
1577
1578         /* Check unused under lock */
1579         for (; i < IMP_AT_MAX_PORTALS; i++) {
1580                 if (at->iat_portal[i] == portal)
1581                         goto out;
1582                 if (at->iat_portal[i] == 0)
1583                         /* unused */
1584                         break;
1585         }
1586
1587         /* Not enough portals? */
1588         LASSERT(i < IMP_AT_MAX_PORTALS);
1589
1590         at->iat_portal[i] = portal;
1591 out:
1592         spin_unlock(&imp->imp_lock);
1593         return i;
1594 }