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         cfs_spin_lock(&imp->imp_lock);          \
  88         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  89         cfs_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         cfs_spin_lock(&imp->imp_lock);
 106
 107         imp->imp_generation++;
 108         imp->imp_state =  LUSTRE_IMP_NEW;
 109
 110         cfs_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         cfs_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                 cfs_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                 cfs_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         cfs_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         cfs_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         cfs_list_t *tmp, *n;
 248         struct ptlrpc_request *req;
 249         unsigned int timeout = 0;
 250
 251         cfs_spin_lock(&imp->imp_lock);
 252         cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 253                 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
 254                 timeout = max(ptlrpc_inflight_deadline(req, now), timeout);
 255         }
 256         cfs_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         cfs_list_t *tmp, *n;
 269         struct ptlrpc_request *req;
 270         struct l_wait_info lwi;
 271         unsigned int timeout;
 272         int rc;
 273
 274         cfs_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                                   (cfs_atomic_read(&imp->imp_inflight) == 0),
 319                                   &lwi);
 320                 if (rc) {
 321                         const char *cli_tgt = obd2cli_tgt(imp->imp_obd);
 322
 323                         CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 324                                cli_tgt, rc,
 325                                cfs_atomic_read(&imp->imp_inflight));
 326
 327                         cfs_spin_lock(&imp->imp_lock);
 328                         if (cfs_atomic_read(&imp->imp_inflight) == 0) {
 329                                 int count = cfs_atomic_read(&imp->imp_unregistering);
 330
 331                                 /* We know that "unregistering" rpcs only can
 332                                  * survive in sending or delaying lists (they
 333                                  * maybe waiting for long reply unlink in
 334                                  * sluggish nets). Let's check this. If there
 335                                  * is no inflight and unregistering != 0, this
 336                                  * is bug. */
 337                                 LASSERTF(count == 0, "Some RPCs are still "
 338                                          "unregistering: %d\n", count);
 339
 340                                 /* Let's save one loop as soon as inflight have
 341                                  * dropped to zero. No new inflights possible at
 342                                  * this point. */
 343                                 rc = 0;
 344                         } else {
 345                                 cfs_list_for_each_safe(tmp, n,
 346                                                        &imp->imp_sending_list) {
 347                                         req = cfs_list_entry(tmp,
 348                                                              struct ptlrpc_request,
 349                                                              rq_list);
 350                                         DEBUG_REQ(D_ERROR, req,
 351                                                   "still on sending list");
 352                                 }
 353                                 cfs_list_for_each_safe(tmp, n,
 354                                                        &imp->imp_delayed_list) {
 355                                         req = cfs_list_entry(tmp,
 356                                                              struct ptlrpc_request,
 357                                                              rq_list);
 358                                         DEBUG_REQ(D_ERROR, req,
 359                                                   "still on delayed list");
 360                                 }
 361
 362                                 CERROR("%s: RPCs in \"%s\" phase found (%d). "
 363                                        "Network is sluggish? Waiting them "
 364                                        "to error out.\n", cli_tgt,
 365                                        ptlrpc_phase2str(RQ_PHASE_UNREGISTERING),
 366                                        cfs_atomic_read(&imp->
 367                                                        imp_unregistering));
 368                         }
 369                         cfs_spin_unlock(&imp->imp_lock);
 370                   }
 371         } while (rc != 0);
 372
 373         /*
 374          * Let's additionally check that no new rpcs added to import in
 375          * "invalidate" state.
 376          */
 377         LASSERT(cfs_atomic_read(&imp->imp_inflight) == 0);
 378 out:
 379         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 380         sptlrpc_import_flush_all_ctx(imp);
 381
 382         cfs_atomic_dec(&imp->imp_inval_count);
 383         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
 384 }
 385
 386 /* unset imp_invalid */
 387 void ptlrpc_activate_import(struct obd_import *imp)
 388 {
 389         struct obd_device *obd = imp->imp_obd;
 390
 391         cfs_spin_lock(&imp->imp_lock);
 392         imp->imp_invalid = 0;
 393         ptlrpc_activate_timeouts(imp);
 394         cfs_spin_unlock(&imp->imp_lock);
 395         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 396 }
 397
 398 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 399 {
 400         ENTRY;
 401
 402         LASSERT(!imp->imp_dlm_fake);
 403
 404         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 405                 if (!imp->imp_replayable) {
 406                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 407                                "auto-deactivating\n",
 408                                obd2cli_tgt(imp->imp_obd),
 409                                imp->imp_connection->c_remote_uuid.uuid,
 410                                imp->imp_obd->obd_name);
 411                         ptlrpc_deactivate_import(imp);
 412                 }
 413
 414                 CDEBUG(D_HA, "%s: waking up pinger\n",
 415                        obd2cli_tgt(imp->imp_obd));
 416
 417                 cfs_spin_lock(&imp->imp_lock);
 418                 imp->imp_force_verify = 1;
 419                 cfs_spin_unlock(&imp->imp_lock);
 420
 421                 ptlrpc_pinger_wake_up();
 422         }
 423         EXIT;
 424 }
 425
 426 int ptlrpc_reconnect_import(struct obd_import *imp)
 427 {
 428         ptlrpc_set_import_discon(imp, 0);
 429         /* Force a new connect attempt */
 430         ptlrpc_invalidate_import(imp);
 431         /* Do a fresh connect next time by zeroing the handle */
 432         ptlrpc_disconnect_import(imp, 1);
 433         /* Wait for all invalidate calls to finish */
 434         if (cfs_atomic_read(&imp->imp_inval_count) > 0) {
 435                 int rc;
 436                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 437                 rc = l_wait_event(imp->imp_recovery_waitq,
 438                                   (cfs_atomic_read(&imp->imp_inval_count) == 0),
 439                                   &lwi);
 440                 if (rc)
 441                         CERROR("Interrupted, inval=%d\n",
 442                                cfs_atomic_read(&imp->imp_inval_count));
 443         }
 444
 445         /* Allow reconnect attempts */
 446         imp->imp_obd->obd_no_recov = 0;
 447         /* Remove 'invalid' flag */
 448         ptlrpc_activate_import(imp);
 449         /* Attempt a new connect */
 450         ptlrpc_recover_import(imp, NULL);
 451         return 0;
 452 }
 453
 454 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 455
 456 static int import_select_connection(struct obd_import *imp)
 457 {
 458         struct obd_import_conn *imp_conn = NULL, *conn;
 459         struct obd_export *dlmexp;
 460         int tried_all = 1;
 461         ENTRY;
 462
 463         cfs_spin_lock(&imp->imp_lock);
 464
 465         if (cfs_list_empty(&imp->imp_conn_list)) {
 466                 CERROR("%s: no connections available\n",
 467                         imp->imp_obd->obd_name);
 468                 cfs_spin_unlock(&imp->imp_lock);
 469                 RETURN(-EINVAL);
 470         }
 471
 472         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 473                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 474                        imp->imp_obd->obd_name,
 475                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 476                        conn->oic_last_attempt);
 477                 /* Don't thrash connections */
 478                 if (cfs_time_before_64(cfs_time_current_64(),
 479                                      conn->oic_last_attempt +
 480                                      cfs_time_seconds(CONNECTION_SWITCH_MIN))) {
 481                         continue;
 482                 }
 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             !imp->imp_recon_bk /* not retrying */) {
 519                 if (at_get(&imp->imp_at.iat_net_latency) <
 520                     CONNECTION_SWITCH_MAX) {
 521                         at_add(&imp->imp_at.iat_net_latency,
 522                                at_get(&imp->imp_at.iat_net_latency) +
 523                                CONNECTION_SWITCH_INC);
 524                 }
 525                 LASSERT(imp_conn->oic_last_attempt);
 526                 CWARN("%s: tried all connections, increasing latency to %ds\n",
 527                       imp->imp_obd->obd_name,
 528                       at_get(&imp->imp_at.iat_net_latency));
 529         }
 530
 531         imp_conn->oic_last_attempt = cfs_time_current_64();
 532
 533         /* switch connection, don't mind if it's same as the current one */
 534         if (imp->imp_connection)
 535                 ptlrpc_connection_put(imp->imp_connection);
 536         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 537
 538         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 539         LASSERT(dlmexp != NULL);
 540         if (dlmexp->exp_connection)
 541                 ptlrpc_connection_put(dlmexp->exp_connection);
 542         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 543         class_export_put(dlmexp);
 544
 545         if (imp->imp_conn_current != imp_conn) {
 546                 if (imp->imp_conn_current)
 547                         CDEBUG(D_HA, "Changing connection for %s to %s/%s\n",
 548                                imp->imp_obd->obd_name, imp_conn->oic_uuid.uuid,
 549                                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 550                 imp->imp_conn_current = imp_conn;
 551         }
 552
 553         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 554                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 555                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 556
 557         cfs_spin_unlock(&imp->imp_lock);
 558
 559         RETURN(0);
 560 }
 561
 562 /*
 563  * must be called under imp_lock
 564  */
 565 static int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 566 {
 567         struct ptlrpc_request *req;
 568         cfs_list_t *tmp;
 569
 570         if (cfs_list_empty(&imp->imp_replay_list))
 571                 return 0;
 572         tmp = imp->imp_replay_list.next;
 573         req = cfs_list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 574         *transno = req->rq_transno;
 575         if (req->rq_transno == 0) {
 576                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 577                 LBUG();
 578         }
 579
 580         return 1;
 581 }
 582
 583 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 584 {
 585         struct obd_device *obd = imp->imp_obd;
 586         int initial_connect = 0;
 587         int set_transno = 0;
 588         __u64 committed_before_reconnect = 0;
 589         struct ptlrpc_request *request;
 590         char *bufs[] = { NULL,
 591                          obd2cli_tgt(imp->imp_obd),
 592                          obd->obd_uuid.uuid,
 593                          (char *)&imp->imp_dlm_handle,
 594                          (char *)&imp->imp_connect_data };
 595         struct ptlrpc_connect_async_args *aa;
 596         int rc;
 597         ENTRY;
 598
 599         cfs_spin_lock(&imp->imp_lock);
 600         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 601                 cfs_spin_unlock(&imp->imp_lock);
 602                 CERROR("can't connect to a closed import\n");
 603                 RETURN(-EINVAL);
 604         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 605                 cfs_spin_unlock(&imp->imp_lock);
 606                 CERROR("already connected\n");
 607                 RETURN(0);
 608         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 609                 cfs_spin_unlock(&imp->imp_lock);
 610                 CERROR("already connecting\n");
 611                 RETURN(-EALREADY);
 612         }
 613
 614         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 615
 616         imp->imp_conn_cnt++;
 617         imp->imp_resend_replay = 0;
 618
 619         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 620                 initial_connect = 1;
 621         else
 622                 committed_before_reconnect = imp->imp_peer_committed_transno;
 623
 624         set_transno = ptlrpc_first_transno(imp,
 625                                            &imp->imp_connect_data.ocd_transno);
 626         cfs_spin_unlock(&imp->imp_lock);
 627
 628         if (new_uuid) {
 629                 struct obd_uuid uuid;
 630
 631                 obd_str2uuid(&uuid, new_uuid);
 632                 rc = import_set_conn_priority(imp, &uuid);
 633                 if (rc)
 634                         GOTO(out, rc);
 635         }
 636
 637         rc = import_select_connection(imp);
 638         if (rc)
 639                 GOTO(out, rc);
 640
 641         /* last in connection list */
 642         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 643                 if (imp->imp_initial_recov_bk && initial_connect) {
 644                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 645                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 646                         /* Don't retry if connect fails */
 647                         rc = 0;
 648                         obd_set_info_async(obd->obd_self_export,
 649                                            sizeof(KEY_INIT_RECOV),
 650                                            KEY_INIT_RECOV,
 651                                            sizeof(rc), &rc, NULL);
 652                 }
 653                 if (imp->imp_recon_bk) {
 654                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 655                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 656                         cfs_spin_lock(&imp->imp_lock);
 657                         imp->imp_last_recon = 1;
 658                         cfs_spin_unlock(&imp->imp_lock);
 659                 }
 660         }
 661
 662         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 663         if (rc)
 664                 GOTO(out, rc);
 665
 666         /* Reset connect flags to the originally requested flags, in case
 667          * the server is updated on-the-fly we will get the new features. */
 668         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 669         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 670
 671         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 672                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 673         if (rc)
 674                 GOTO(out, rc);
 675
 676         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 677         if (request == NULL)
 678                 GOTO(out, rc = -ENOMEM);
 679
 680         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 681                                       imp->imp_connect_op, bufs, NULL);
 682         if (rc) {
 683                 ptlrpc_request_free(request);
 684                 GOTO(out, rc);
 685         }
 686
 687         /* Report the rpc service time to the server so that it knows how long
 688          * to wait for clients to join recovery */
 689         lustre_msg_set_service_time(request->rq_reqmsg,
 690                                     at_timeout2est(request->rq_timeout));
 691
 692         /* The amount of time we give the server to process the connect req.
 693          * import_select_connection will increase the net latency on
 694          * repeated reconnect attempts to cover slow networks.
 695          * We override/ignore the server rpc completion estimate here,
 696          * which may be large if this is a reconnect attempt */
 697         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 698         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 699
 700 #ifndef __KERNEL__
 701         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 702 #endif
 703         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 704
 705         request->rq_no_resend = request->rq_no_delay = 1;
 706         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 707         /* Allow a slightly larger reply for future growth compatibility */
 708         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 709                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 710         ptlrpc_request_set_replen(request);
 711         request->rq_interpret_reply = ptlrpc_connect_interpret;
 712
 713         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 714         aa = ptlrpc_req_async_args(request);
 715         memset(aa, 0, sizeof *aa);
 716
 717         aa->pcaa_peer_committed = committed_before_reconnect;
 718         aa->pcaa_initial_connect = initial_connect;
 719
 720         if (aa->pcaa_initial_connect) {
 721                 cfs_spin_lock(&imp->imp_lock);
 722                 imp->imp_replayable = 1;
 723                 cfs_spin_unlock(&imp->imp_lock);
 724                 lustre_msg_add_op_flags(request->rq_reqmsg,
 725                                         MSG_CONNECT_INITIAL);
 726         }
 727
 728         if (set_transno)
 729                 lustre_msg_add_op_flags(request->rq_reqmsg,
 730                                         MSG_CONNECT_TRANSNO);
 731
 732         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 733         ptlrpcd_add_req(request, PSCOPE_OTHER);
 734         rc = 0;
 735 out:
 736         if (rc != 0) {
 737                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 738         }
 739
 740         RETURN(rc);
 741 }
 742 EXPORT_SYMBOL(ptlrpc_connect_import);
 743
 744 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 745 {
 746 #ifdef __KERNEL__
 747         struct obd_import_conn *imp_conn;
 748 #endif
 749         int wake_pinger = 0;
 750
 751         ENTRY;
 752
 753         cfs_spin_lock(&imp->imp_lock);
 754         if (cfs_list_empty(&imp->imp_conn_list))
 755                 GOTO(unlock, 0);
 756
 757 #ifdef __KERNEL__
 758         imp_conn = cfs_list_entry(imp->imp_conn_list.prev,
 759                                   struct obd_import_conn,
 760                                   oic_item);
 761
 762         /* XXX: When the failover node is the primary node, it is possible
 763          * to have two identical connections in imp_conn_list. We must
 764          * compare not conn's pointers but NIDs, otherwise we can defeat
 765          * connection throttling. (See bug 14774.) */
 766         if (imp->imp_conn_current->oic_conn->c_peer.nid !=
 767                                 imp_conn->oic_conn->c_peer.nid) {
 768                 ptlrpc_ping_import_soon(imp);
 769                 wake_pinger = 1;
 770         }
 771 #else
 772         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 773         wake_pinger = 1;
 774 #endif
 775
 776  unlock:
 777         cfs_spin_unlock(&imp->imp_lock);
 778
 779         if (wake_pinger)
 780                 ptlrpc_pinger_wake_up();
 781
 782         EXIT;
 783 }
 784
 785 static int ptlrpc_busy_reconnect(int rc)
 786 {
 787         return (rc == -EBUSY) || (rc == -EAGAIN);
 788 }
 789
 790
 791 static int ptlrpc_connect_interpret(const struct lu_env *env,
 792                                     struct ptlrpc_request *request,
 793                                     void *data, int rc)
 794 {
 795         struct ptlrpc_connect_async_args *aa = data;
 796         struct obd_import *imp = request->rq_import;
 797         struct client_obd *cli = &imp->imp_obd->u.cli;
 798         struct lustre_handle old_hdl;
 799         __u64 old_connect_flags;
 800         int msg_flags;
 801         ENTRY;
 802
 803         cfs_spin_lock(&imp->imp_lock);
 804         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 805                 cfs_spin_unlock(&imp->imp_lock);
 806                 RETURN(0);
 807         }
 808
 809         if (rc) {
 810                 /* if this reconnect to busy export - not need select new target
 811                  * for connecting*/
 812                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 813                 cfs_spin_unlock(&imp->imp_lock);
 814                 GOTO(out, rc);
 815         }
 816
 817         LASSERT(imp->imp_conn_current);
 818
 819         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 820
 821         /* All imports are pingable */
 822         imp->imp_pingable = 1;
 823         imp->imp_force_reconnect = 0;
 824
 825         if (aa->pcaa_initial_connect) {
 826                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 827                         imp->imp_replayable = 1;
 828                         cfs_spin_unlock(&imp->imp_lock);
 829                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 830                                obd2cli_tgt(imp->imp_obd));
 831                 } else {
 832                         imp->imp_replayable = 0;
 833                         cfs_spin_unlock(&imp->imp_lock);
 834                 }
 835
 836                 /* if applies, adjust the imp->imp_msg_magic here
 837                  * according to reply flags */
 838
 839                 imp->imp_remote_handle =
 840                                 *lustre_msg_get_handle(request->rq_repmsg);
 841
 842                 /* Initial connects are allowed for clients with non-random
 843                  * uuids when servers are in recovery.  Simply signal the
 844                  * servers replay is complete and wait in REPLAY_WAIT. */
 845                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 846                         CDEBUG(D_HA, "connect to %s during recovery\n",
 847                                obd2cli_tgt(imp->imp_obd));
 848                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 849                 } else {
 850                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 851                         ptlrpc_activate_import(imp);
 852                 }
 853
 854                 GOTO(finish, rc = 0);
 855         } else {
 856                 cfs_spin_unlock(&imp->imp_lock);
 857         }
 858
 859         /* Determine what recovery state to move the import to. */
 860         if (MSG_CONNECT_RECONNECT & msg_flags) {
 861                 memset(&old_hdl, 0, sizeof(old_hdl));
 862                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 863                             sizeof (old_hdl))) {
 864                         CERROR("%s@%s didn't like our handle "LPX64
 865                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 866                                imp->imp_connection->c_remote_uuid.uuid,
 867                                imp->imp_dlm_handle.cookie);
 868                         GOTO(out, rc = -ENOTCONN);
 869                 }
 870
 871                 if (memcmp(&imp->imp_remote_handle,
 872                            lustre_msg_get_handle(request->rq_repmsg),
 873                            sizeof(imp->imp_remote_handle))) {
 874                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 875                                 D_HA : D_WARNING;
 876
 877                         /* Bug 16611/14775: if server handle have changed,
 878                          * that means some sort of disconnection happened.
 879                          * If the server is not in recovery, that also means it
 880                          * already erased all of our state because of previous
 881                          * eviction. If it is in recovery - we are safe to
 882                          * participate since we can reestablish all of our state
 883                          * with server again */
 884                         CDEBUG(level,"%s@%s changed server handle from "
 885                                      LPX64" to "LPX64"%s\n",
 886                                      obd2cli_tgt(imp->imp_obd),
 887                                      imp->imp_connection->c_remote_uuid.uuid,
 888                                      imp->imp_remote_handle.cookie,
 889                                      lustre_msg_get_handle(request->rq_repmsg)->
 890                                                                         cookie,
 891                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
 892                                          " but is still in recovery" : "");
 893
 894                         imp->imp_remote_handle =
 895                                      *lustre_msg_get_handle(request->rq_repmsg);
 896
 897                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 898                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 899                                 GOTO(finish, rc = 0);
 900                         }
 901
 902                 } else {
 903                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 904                                obd2cli_tgt(imp->imp_obd),
 905                                imp->imp_connection->c_remote_uuid.uuid);
 906                 }
 907
 908                 if (imp->imp_invalid) {
 909                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 910                                "marking evicted\n", imp->imp_obd->obd_name);
 911                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 912                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 913                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 914                                imp->imp_obd->obd_name,
 915                                obd2cli_tgt(imp->imp_obd));
 916
 917                         cfs_spin_lock(&imp->imp_lock);
 918                         imp->imp_resend_replay = 1;
 919                         cfs_spin_unlock(&imp->imp_lock);
 920
 921                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 922                 } else {
 923                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 924                 }
 925         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 926                 LASSERT(imp->imp_replayable);
 927                 imp->imp_remote_handle =
 928                                 *lustre_msg_get_handle(request->rq_repmsg);
 929                 imp->imp_last_replay_transno = 0;
 930                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 931         } else {
 932                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 933                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 934                 imp->imp_remote_handle =
 935                                 *lustre_msg_get_handle(request->rq_repmsg);
 936                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 937         }
 938
 939         /* Sanity checks for a reconnected import. */
 940         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 941                 CERROR("imp_replayable flag does not match server "
 942                        "after reconnect. We should LBUG right here.\n");
 943         }
 944
 945         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 946             aa->pcaa_peer_committed) {
 947                 CERROR("%s went back in time (transno "LPD64
 948                        " was previously committed, server now claims "LPD64
 949                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 950                        "id=9646\n",
 951                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 952                        lustre_msg_get_last_committed(request->rq_repmsg));
 953         }
 954
 955 finish:
 956         rc = ptlrpc_import_recovery_state_machine(imp);
 957         if (rc != 0) {
 958                 if (rc == -ENOTCONN) {
 959                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 960                                "invalidating and reconnecting\n",
 961                                obd2cli_tgt(imp->imp_obd),
 962                                imp->imp_connection->c_remote_uuid.uuid);
 963                         ptlrpc_connect_import(imp, NULL);
 964                         RETURN(0);
 965                 }
 966         } else {
 967                 struct obd_connect_data *ocd;
 968                 struct obd_export *exp;
 969                 int ret;
 970                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 971                                            RCL_SERVER);
 972                 /* server replied obd_connect_data is always bigger */
 973                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 974                                                    &RMF_CONNECT_DATA, ret);
 975
 976                 cfs_spin_lock(&imp->imp_lock);
 977                 cfs_list_del(&imp->imp_conn_current->oic_item);
 978                 cfs_list_add(&imp->imp_conn_current->oic_item,
 979                              &imp->imp_conn_list);
 980                 imp->imp_last_success_conn =
 981                         imp->imp_conn_current->oic_last_attempt;
 982
 983                 if (ocd == NULL) {
 984                         cfs_spin_unlock(&imp->imp_lock);
 985                         CERROR("Wrong connect data from server\n");
 986                         rc = -EPROTO;
 987                         GOTO(out, rc);
 988                 }
 989
 990                 imp->imp_connect_data = *ocd;
 991
 992                 exp = class_conn2export(&imp->imp_dlm_handle);
 993                 cfs_spin_unlock(&imp->imp_lock);
 994
 995                 /* check that server granted subset of flags we asked for. */
 996                 LASSERTF((ocd->ocd_connect_flags &
 997                           imp->imp_connect_flags_orig) ==
 998                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 999                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
1000
1001                 if (!exp) {
1002                         /* This could happen if export is cleaned during the
1003                            connect attempt */
1004                         CERROR("Missing export for %s\n",
1005                                imp->imp_obd->obd_name);
1006                         GOTO(out, rc = -ENODEV);
1007                 }
1008                 old_connect_flags = exp->exp_connect_flags;
1009                 exp->exp_connect_flags = ocd->ocd_connect_flags;
1010                 imp->imp_obd->obd_self_export->exp_connect_flags =
1011                                                         ocd->ocd_connect_flags;
1012                 class_export_put(exp);
1013
1014                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1015
1016                 if (!ocd->ocd_ibits_known &&
1017                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
1018                         CERROR("Inodebits aware server returned zero compatible"
1019                                " bits?\n");
1020
1021                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1022                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1023                                         LUSTRE_VERSION_OFFSET_WARN ||
1024                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1025                                         LUSTRE_VERSION_OFFSET_WARN)) {
1026                         /* Sigh, some compilers do not like #ifdef in the middle
1027                            of macro arguments */
1028 #ifdef __KERNEL__
1029                         const char *older =
1030                                 "older. Consider upgrading this client";
1031 #else
1032                         const char *older =
1033                                 "older. Consider recompiling this application";
1034 #endif
1035                         const char *newer = "newer than client version";
1036
1037                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1038                                       "is much %s (%s)\n",
1039                                       obd2cli_tgt(imp->imp_obd),
1040                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1041                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1042                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1043                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1044                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1045                                       newer : older, LUSTRE_VERSION_STRING);
1046                 }
1047
1048                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1049                         /* We sent to the server ocd_cksum_types with bits set
1050                          * for algorithms we understand. The server masked off
1051                          * the checksum types it doesn't support */
1052                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
1053                                 LCONSOLE_WARN("The negotiation of the checksum "
1054                                               "alogrithm to use with server %s "
1055                                               "failed (%x/%x), disabling "
1056                                               "checksums\n",
1057                                               obd2cli_tgt(imp->imp_obd),
1058                                               ocd->ocd_cksum_types,
1059                                               OBD_CKSUM_ALL);
1060                                 cli->cl_checksum = 0;
1061                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1062                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1063                         } else {
1064                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1065
1066                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
1067                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
1068                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
1069                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
1070                                 else
1071                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1072                         }
1073                 } else {
1074                         /* The server does not support OBD_CONNECT_CKSUM.
1075                          * Enforce CRC32 for backward compatibility*/
1076                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1077                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1078                 }
1079
1080                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
1081                         cli->cl_max_pages_per_rpc =
1082                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1083                 }
1084
1085                 /* Reset ns_connect_flags only for initial connect. It might be
1086                  * changed in while using FS and if we reset it in reconnect
1087                  * this leads to losing user settings done before such as
1088                  * disable lru_resize, etc. */
1089                 if (old_connect_flags != exp->exp_connect_flags ||
1090                     aa->pcaa_initial_connect) {
1091                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1092                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1093                               ocd->ocd_connect_flags);
1094                         imp->imp_obd->obd_namespace->ns_connect_flags =
1095                                 ocd->ocd_connect_flags;
1096                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1097                                 ocd->ocd_connect_flags;
1098                 }
1099
1100                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1101                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1102                         /* We need a per-message support flag, because
1103                            a. we don't know if the incoming connect reply
1104                               supports AT or not (in reply_in_callback)
1105                               until we unpack it.
1106                            b. failovered server means export and flags are gone
1107                               (in ptlrpc_send_reply).
1108                            Can only be set when we know AT is supported at
1109                            both ends */
1110                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1111                 else
1112                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1113
1114                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1115                         (cli->cl_max_pages_per_rpc > 0));
1116         }
1117
1118 out:
1119         if (rc != 0) {
1120                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1121                 cfs_spin_lock(&imp->imp_lock);
1122                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
1123                     (request->rq_import_generation == imp->imp_generation))
1124                         ptlrpc_deactivate_and_unlock_import(imp);
1125                 else
1126                         cfs_spin_unlock(&imp->imp_lock);
1127
1128                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
1129                     (rc == -EACCES)) {
1130                         /*
1131                          * Give up trying to reconnect
1132                          * EACCES means client has no permission for connection
1133                          */
1134                         imp->imp_obd->obd_no_recov = 1;
1135                         ptlrpc_deactivate_import(imp);
1136                 }
1137
1138                 if (rc == -EPROTO) {
1139                         struct obd_connect_data *ocd;
1140
1141                         /* reply message might not be ready */
1142                         if (request->rq_repmsg == NULL)
1143                                 RETURN(-EPROTO);
1144
1145                         ocd = req_capsule_server_get(&request->rq_pill,
1146                                                      &RMF_CONNECT_DATA);
1147                         if (ocd &&
1148                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1149                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1150                            /* Actually servers are only supposed to refuse
1151                               connection from liblustre clients, so we should
1152                               never see this from VFS context */
1153                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1154                                         "(%d.%d.%d.%d)"
1155                                         " refused connection from this client "
1156                                         "with an incompatible version (%s).  "
1157                                         "Client must be recompiled\n",
1158                                         obd2cli_tgt(imp->imp_obd),
1159                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1160                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1161                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1162                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1163                                         LUSTRE_VERSION_STRING);
1164                                 ptlrpc_deactivate_import(imp);
1165                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1166                         }
1167                         RETURN(-EPROTO);
1168                 }
1169
1170                 ptlrpc_maybe_ping_import_soon(imp);
1171
1172                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1173                        obd2cli_tgt(imp->imp_obd),
1174                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1175         }
1176
1177         cfs_spin_lock(&imp->imp_lock);
1178         imp->imp_last_recon = 0;
1179         cfs_spin_unlock(&imp->imp_lock);
1180
1181         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1182         RETURN(rc);
1183 }
1184
1185 static int completed_replay_interpret(const struct lu_env *env,
1186                                       struct ptlrpc_request *req,
1187                                       void * data, int rc)
1188 {
1189         ENTRY;
1190         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1191         if (req->rq_status == 0 &&
1192             !req->rq_import->imp_vbr_failed) {
1193                 ptlrpc_import_recovery_state_machine(req->rq_import);
1194         } else {
1195                 if (req->rq_import->imp_vbr_failed) {
1196                         CDEBUG(D_WARNING,
1197                                "%s: version recovery fails, reconnecting\n",
1198                                req->rq_import->imp_obd->obd_name);
1199                         cfs_spin_lock(&req->rq_import->imp_lock);
1200                         req->rq_import->imp_vbr_failed = 0;
1201                         cfs_spin_unlock(&req->rq_import->imp_lock);
1202                 } else {
1203                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1204                                      "reconnecting\n",
1205                                req->rq_import->imp_obd->obd_name,
1206                                req->rq_status);
1207                 }
1208                 ptlrpc_connect_import(req->rq_import, NULL);
1209         }
1210
1211         RETURN(0);
1212 }
1213
1214 static int signal_completed_replay(struct obd_import *imp)
1215 {
1216         struct ptlrpc_request *req;
1217         ENTRY;
1218
1219         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1220         cfs_atomic_inc(&imp->imp_replay_inflight);
1221
1222         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1223                                         OBD_PING);
1224         if (req == NULL) {
1225                 cfs_atomic_dec(&imp->imp_replay_inflight);
1226                 RETURN(-ENOMEM);
1227         }
1228
1229         ptlrpc_request_set_replen(req);
1230         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1231         lustre_msg_add_flags(req->rq_reqmsg,
1232                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1233         if (AT_OFF)
1234                 req->rq_timeout *= 3;
1235         req->rq_interpret_reply = completed_replay_interpret;
1236
1237         ptlrpcd_add_req(req, PSCOPE_OTHER);
1238         RETURN(0);
1239 }
1240
1241 #ifdef __KERNEL__
1242 static int ptlrpc_invalidate_import_thread(void *data)
1243 {
1244         struct obd_import *imp = data;
1245
1246         ENTRY;
1247
1248         cfs_daemonize_ctxt("ll_imp_inval");
1249
1250         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1251                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1252                imp->imp_connection->c_remote_uuid.uuid);
1253
1254         ptlrpc_invalidate_import(imp);
1255
1256         if (obd_dump_on_eviction) {
1257                 CERROR("dump the log upon eviction\n");
1258                 libcfs_debug_dumplog();
1259         }
1260
1261         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1262         ptlrpc_import_recovery_state_machine(imp);
1263
1264         class_import_put(imp);
1265         RETURN(0);
1266 }
1267 #endif
1268
1269 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1270 {
1271         int rc = 0;
1272         int inflight;
1273         char *target_start;
1274         int target_len;
1275
1276         ENTRY;
1277         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1278                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1279                           &target_start, &target_len);
1280                 /* Don't care about MGC eviction */
1281                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1282                            LUSTRE_MGC_NAME) != 0) {
1283                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1284                                            "%.*s; in progress operations using "
1285                                            "this service will fail.\n",
1286                                            target_len, target_start);
1287                 }
1288                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1289                        obd2cli_tgt(imp->imp_obd),
1290                        imp->imp_connection->c_remote_uuid.uuid);
1291
1292 #ifdef __KERNEL__
1293                 /* bug 17802:  XXX client_disconnect_export vs connect request
1294                  * race. if client will evicted at this time, we start
1295                  * invalidate thread without reference to import and import can
1296                  * be freed at same time. */
1297                 class_import_get(imp);
1298                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1299                                        CLONE_VM | CLONE_FILES);
1300                 if (rc < 0) {
1301                         class_import_put(imp);
1302                         CERROR("error starting invalidate thread: %d\n", rc);
1303                 } else {
1304                         rc = 0;
1305                 }
1306                 RETURN(rc);
1307 #else
1308                 ptlrpc_invalidate_import(imp);
1309
1310                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1311 #endif
1312         }
1313
1314         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1315                 CDEBUG(D_HA, "replay requested by %s\n",
1316                        obd2cli_tgt(imp->imp_obd));
1317                 rc = ptlrpc_replay_next(imp, &inflight);
1318                 if (inflight == 0 &&
1319                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1320                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1321                         rc = ldlm_replay_locks(imp);
1322                         if (rc)
1323                                 GOTO(out, rc);
1324                 }
1325                 rc = 0;
1326         }
1327
1328         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1329                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1330                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1331                         rc = signal_completed_replay(imp);
1332                         if (rc)
1333                                 GOTO(out, rc);
1334                 }
1335
1336         }
1337
1338         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1339                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1340                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1341                 }
1342         }
1343
1344         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1345                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1346                        obd2cli_tgt(imp->imp_obd),
1347                        imp->imp_connection->c_remote_uuid.uuid);
1348
1349                 rc = ptlrpc_resend(imp);
1350                 if (rc)
1351                         GOTO(out, rc);
1352                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1353                 ptlrpc_activate_import(imp);
1354
1355                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1356                           &target_start, &target_len);
1357                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1358                               "using nid %s.\n", imp->imp_obd->obd_name,
1359                               target_len, target_start,
1360                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1361         }
1362
1363         if (imp->imp_state == LUSTRE_IMP_FULL) {
1364                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1365                 ptlrpc_wake_delayed(imp);
1366         }
1367
1368 out:
1369         RETURN(rc);
1370 }
1371
1372 static int back_to_sleep(void *unused)
1373 {
1374         return 0;
1375 }
1376
1377 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1378 {
1379         struct ptlrpc_request *req;
1380         int rq_opc, rc = 0;
1381         int nowait = imp->imp_obd->obd_force;
1382         ENTRY;
1383
1384         if (nowait)
1385                 GOTO(set_state, rc);
1386
1387         switch (imp->imp_connect_op) {
1388         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1389         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1390         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1391         default:
1392                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1393                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1394                 RETURN(-EINVAL);
1395         }
1396
1397         if (ptlrpc_import_in_recovery(imp)) {
1398                 struct l_wait_info lwi;
1399                 cfs_duration_t timeout;
1400
1401
1402                 if (AT_OFF) {
1403                         if (imp->imp_server_timeout)
1404                                 timeout = cfs_time_seconds(obd_timeout / 2);
1405                         else
1406                                 timeout = cfs_time_seconds(obd_timeout);
1407                 } else {
1408                         int idx = import_at_get_index(imp,
1409                                 imp->imp_client->cli_request_portal);
1410                         timeout = cfs_time_seconds(
1411                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1412                 }
1413
1414                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1415                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1416                 rc = l_wait_event(imp->imp_recovery_waitq,
1417                                   !ptlrpc_import_in_recovery(imp), &lwi);
1418
1419         }
1420
1421         cfs_spin_lock(&imp->imp_lock);
1422         if (imp->imp_state != LUSTRE_IMP_FULL)
1423                 GOTO(out, 0);
1424
1425         cfs_spin_unlock(&imp->imp_lock);
1426
1427         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1428                                         LUSTRE_OBD_VERSION, rq_opc);
1429         if (req) {
1430                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1431                  * it fails.  We can get through the above with a down server
1432                  * if the client doesn't know the server is gone yet. */
1433                 req->rq_no_resend = 1;
1434
1435 #ifndef CRAY_XT3
1436                 /* We want client umounts to happen quickly, no matter the
1437                    server state... */
1438                 req->rq_timeout = min_t(int, req->rq_timeout,
1439                                         INITIAL_CONNECT_TIMEOUT);
1440 #else
1441                 /* ... but we always want liblustre clients to nicely
1442                    disconnect, so only use the adaptive value. */
1443                 if (AT_OFF)
1444                         req->rq_timeout = obd_timeout / 3;
1445 #endif
1446
1447                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1448                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1449                 ptlrpc_request_set_replen(req);
1450                 rc = ptlrpc_queue_wait(req);
1451                 ptlrpc_req_finished(req);
1452         }
1453
1454 set_state:
1455         cfs_spin_lock(&imp->imp_lock);
1456 out:
1457         if (noclose)
1458                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1459         else
1460                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1461         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1462         /* Try all connections in the future - bz 12758 */
1463         imp->imp_last_recon = 0;
1464         cfs_spin_unlock(&imp->imp_lock);
1465
1466         RETURN(rc);
1467 }
1468
1469 void ptlrpc_cleanup_imp(struct obd_import *imp)
1470 {
1471         ENTRY;
1472
1473         cfs_spin_lock(&imp->imp_lock);
1474         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1475         imp->imp_generation++;
1476         cfs_spin_unlock(&imp->imp_lock);
1477         ptlrpc_abort_inflight(imp);
1478
1479         EXIT;
1480 }
1481
1482 /* Adaptive Timeout utils */
1483 extern unsigned int at_min, at_max, at_history;
1484
1485 /* Bin into timeslices using AT_BINS bins.
1486    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1487    but still smoothing out a return to normalcy from a slow response.
1488    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1489 int at_add(struct adaptive_timeout *at, unsigned int val)
1490 {
1491         unsigned int old = at->at_current;
1492         time_t now = cfs_time_current_sec();
1493         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1494
1495         LASSERT(at);
1496         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1497                val, at, now - at->at_binstart, at->at_current,
1498                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1499
1500         if (val == 0)
1501                 /* 0's don't count, because we never want our timeout to
1502                    drop to 0, and because 0 could mean an error */
1503                 return 0;
1504
1505         cfs_spin_lock(&at->at_lock);
1506
1507         if (unlikely(at->at_binstart == 0)) {
1508                 /* Special case to remove default from history */
1509                 at->at_current = val;
1510                 at->at_worst_ever = val;
1511                 at->at_worst_time = now;
1512                 at->at_hist[0] = val;
1513                 at->at_binstart = now;
1514         } else if (now - at->at_binstart < binlimit ) {
1515                 /* in bin 0 */
1516                 at->at_hist[0] = max(val, at->at_hist[0]);
1517                 at->at_current = max(val, at->at_current);
1518         } else {
1519                 int i, shift;
1520                 unsigned int maxv = val;
1521                 /* move bins over */
1522                 shift = (now - at->at_binstart) / binlimit;
1523                 LASSERT(shift > 0);
1524                 for(i = AT_BINS - 1; i >= 0; i--) {
1525                         if (i >= shift) {
1526                                 at->at_hist[i] = at->at_hist[i - shift];
1527                                 maxv = max(maxv, at->at_hist[i]);
1528                         } else {
1529                                 at->at_hist[i] = 0;
1530                         }
1531                 }
1532                 at->at_hist[0] = val;
1533                 at->at_current = maxv;
1534                 at->at_binstart += shift * binlimit;
1535         }
1536
1537         if (at->at_current > at->at_worst_ever) {
1538                 at->at_worst_ever = at->at_current;
1539                 at->at_worst_time = now;
1540         }
1541
1542         if (at->at_flags & AT_FLG_NOHIST)
1543                 /* Only keep last reported val; keeping the rest of the history
1544                    for proc only */
1545                 at->at_current = val;
1546
1547         if (at_max > 0)
1548                 at->at_current =  min(at->at_current, at_max);
1549         at->at_current =  max(at->at_current, at_min);
1550
1551         if (at->at_current != old)
1552                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1553                        "(val=%u) hist %u %u %u %u\n", at,
1554                        old, at->at_current, at->at_current - old, val,
1555                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1556                        at->at_hist[3]);
1557
1558         /* if we changed, report the old value */
1559         old = (at->at_current != old) ? old : 0;
1560
1561         cfs_spin_unlock(&at->at_lock);
1562         return old;
1563 }
1564
1565 /* Find the imp_at index for a given portal; assign if space available */
1566 int import_at_get_index(struct obd_import *imp, int portal)
1567 {
1568         struct imp_at *at = &imp->imp_at;
1569         int i;
1570
1571         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1572                 if (at->iat_portal[i] == portal)
1573                         return i;
1574                 if (at->iat_portal[i] == 0)
1575                         /* unused */
1576                         break;
1577         }
1578
1579         /* Not found in list, add it under a lock */
1580         cfs_spin_lock(&imp->imp_lock);
1581
1582         /* Check unused under lock */
1583         for (; i < IMP_AT_MAX_PORTALS; i++) {
1584                 if (at->iat_portal[i] == portal)
1585                         goto out;
1586                 if (at->iat_portal[i] == 0)
1587                         /* unused */
1588                         break;
1589         }
1590
1591         /* Not enough portals? */
1592         LASSERT(i < IMP_AT_MAX_PORTALS);
1593
1594         at->iat_portal[i] = portal;
1595 out:
1596         cfs_spin_unlock(&imp->imp_lock);
1597         return i;
1598 }