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_measured(&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         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
 671
 672         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 673                            &obd->obd_uuid, &imp->imp_connect_data, NULL);
 674         if (rc)
 675                 GOTO(out, rc);
 676
 677         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 678         if (request == NULL)
 679                 GOTO(out, rc = -ENOMEM);
 680
 681         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 682                                       imp->imp_connect_op, bufs, NULL);
 683         if (rc) {
 684                 ptlrpc_request_free(request);
 685                 GOTO(out, rc);
 686         }
 687
 688         /* Report the rpc service time to the server so that it knows how long
 689          * to wait for clients to join recovery */
 690         lustre_msg_set_service_time(request->rq_reqmsg,
 691                                     at_timeout2est(request->rq_timeout));
 692
 693         /* The amount of time we give the server to process the connect req.
 694          * import_select_connection will increase the net latency on
 695          * repeated reconnect attempts to cover slow networks.
 696          * We override/ignore the server rpc completion estimate here,
 697          * which may be large if this is a reconnect attempt */
 698         request->rq_timeout = INITIAL_CONNECT_TIMEOUT;
 699         lustre_msg_set_timeout(request->rq_reqmsg, request->rq_timeout);
 700
 701 #ifndef __KERNEL__
 702         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 703 #endif
 704         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 705
 706         request->rq_no_resend = request->rq_no_delay = 1;
 707         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 708         /* Allow a slightly larger reply for future growth compatibility */
 709         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 710                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 711         ptlrpc_request_set_replen(request);
 712         request->rq_interpret_reply = ptlrpc_connect_interpret;
 713
 714         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 715         aa = ptlrpc_req_async_args(request);
 716         memset(aa, 0, sizeof *aa);
 717
 718         aa->pcaa_peer_committed = committed_before_reconnect;
 719         aa->pcaa_initial_connect = initial_connect;
 720
 721         if (aa->pcaa_initial_connect) {
 722                 cfs_spin_lock(&imp->imp_lock);
 723                 imp->imp_replayable = 1;
 724                 cfs_spin_unlock(&imp->imp_lock);
 725                 lustre_msg_add_op_flags(request->rq_reqmsg,
 726                                         MSG_CONNECT_INITIAL);
 727         }
 728
 729         if (set_transno)
 730                 lustre_msg_add_op_flags(request->rq_reqmsg,
 731                                         MSG_CONNECT_TRANSNO);
 732
 733         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request (timeout %d)",
 734                   request->rq_timeout);
 735         ptlrpcd_add_req(request, PSCOPE_OTHER);
 736         rc = 0;
 737 out:
 738         if (rc != 0) {
 739                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 740         }
 741
 742         RETURN(rc);
 743 }
 744 EXPORT_SYMBOL(ptlrpc_connect_import);
 745
 746 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 747 {
 748 #ifdef __KERNEL__
 749         struct obd_import_conn *imp_conn;
 750 #endif
 751         int wake_pinger = 0;
 752
 753         ENTRY;
 754
 755         cfs_spin_lock(&imp->imp_lock);
 756         if (cfs_list_empty(&imp->imp_conn_list))
 757                 GOTO(unlock, 0);
 758
 759 #ifdef __KERNEL__
 760         imp_conn = cfs_list_entry(imp->imp_conn_list.prev,
 761                                   struct obd_import_conn,
 762                                   oic_item);
 763
 764         /* XXX: When the failover node is the primary node, it is possible
 765          * to have two identical connections in imp_conn_list. We must
 766          * compare not conn's pointers but NIDs, otherwise we can defeat
 767          * connection throttling. (See bug 14774.) */
 768         if (imp->imp_conn_current->oic_conn->c_peer.nid !=
 769                                 imp_conn->oic_conn->c_peer.nid) {
 770                 ptlrpc_ping_import_soon(imp);
 771                 wake_pinger = 1;
 772         }
 773 #else
 774         /* liblustre has no pinger thread, so we wakeup pinger anyway */
 775         wake_pinger = 1;
 776 #endif
 777
 778  unlock:
 779         cfs_spin_unlock(&imp->imp_lock);
 780
 781         if (wake_pinger)
 782                 ptlrpc_pinger_wake_up();
 783
 784         EXIT;
 785 }
 786
 787 static int ptlrpc_busy_reconnect(int rc)
 788 {
 789         return (rc == -EBUSY) || (rc == -EAGAIN);
 790 }
 791
 792
 793 static int ptlrpc_connect_interpret(const struct lu_env *env,
 794                                     struct ptlrpc_request *request,
 795                                     void *data, int rc)
 796 {
 797         struct ptlrpc_connect_async_args *aa = data;
 798         struct obd_import *imp = request->rq_import;
 799         struct client_obd *cli = &imp->imp_obd->u.cli;
 800         struct lustre_handle old_hdl;
 801         __u64 old_connect_flags;
 802         int msg_flags;
 803         ENTRY;
 804
 805         cfs_spin_lock(&imp->imp_lock);
 806         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 807                 cfs_spin_unlock(&imp->imp_lock);
 808                 RETURN(0);
 809         }
 810
 811         if (rc) {
 812                 /* if this reconnect to busy export - not need select new target
 813                  * for connecting*/
 814                 imp->imp_force_reconnect = ptlrpc_busy_reconnect(rc);
 815                 cfs_spin_unlock(&imp->imp_lock);
 816                 GOTO(out, rc);
 817         }
 818
 819         LASSERT(imp->imp_conn_current);
 820
 821         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 822
 823         /* All imports are pingable */
 824         imp->imp_pingable = 1;
 825         imp->imp_force_reconnect = 0;
 826
 827         if (aa->pcaa_initial_connect) {
 828                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 829                         imp->imp_replayable = 1;
 830                         cfs_spin_unlock(&imp->imp_lock);
 831                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 832                                obd2cli_tgt(imp->imp_obd));
 833                 } else {
 834                         imp->imp_replayable = 0;
 835                         cfs_spin_unlock(&imp->imp_lock);
 836                 }
 837
 838                 /* if applies, adjust the imp->imp_msg_magic here
 839                  * according to reply flags */
 840
 841                 imp->imp_remote_handle =
 842                                 *lustre_msg_get_handle(request->rq_repmsg);
 843
 844                 /* Initial connects are allowed for clients with non-random
 845                  * uuids when servers are in recovery.  Simply signal the
 846                  * servers replay is complete and wait in REPLAY_WAIT. */
 847                 if (msg_flags & MSG_CONNECT_RECOVERING) {
 848                         CDEBUG(D_HA, "connect to %s during recovery\n",
 849                                obd2cli_tgt(imp->imp_obd));
 850                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 851                 } else {
 852                         IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 853                         ptlrpc_activate_import(imp);
 854                 }
 855
 856                 GOTO(finish, rc = 0);
 857         } else {
 858                 cfs_spin_unlock(&imp->imp_lock);
 859         }
 860
 861         /* Determine what recovery state to move the import to. */
 862         if (MSG_CONNECT_RECONNECT & msg_flags) {
 863                 memset(&old_hdl, 0, sizeof(old_hdl));
 864                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 865                             sizeof (old_hdl))) {
 866                         CERROR("%s@%s didn't like our handle "LPX64
 867                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 868                                imp->imp_connection->c_remote_uuid.uuid,
 869                                imp->imp_dlm_handle.cookie);
 870                         GOTO(out, rc = -ENOTCONN);
 871                 }
 872
 873                 if (memcmp(&imp->imp_remote_handle,
 874                            lustre_msg_get_handle(request->rq_repmsg),
 875                            sizeof(imp->imp_remote_handle))) {
 876                         int level = msg_flags & MSG_CONNECT_RECOVERING ?
 877                                 D_HA : D_WARNING;
 878
 879                         /* Bug 16611/14775: if server handle have changed,
 880                          * that means some sort of disconnection happened.
 881                          * If the server is not in recovery, that also means it
 882                          * already erased all of our state because of previous
 883                          * eviction. If it is in recovery - we are safe to
 884                          * participate since we can reestablish all of our state
 885                          * with server again */
 886                         CDEBUG(level,"%s@%s changed server handle from "
 887                                      LPX64" to "LPX64"%s\n",
 888                                      obd2cli_tgt(imp->imp_obd),
 889                                      imp->imp_connection->c_remote_uuid.uuid,
 890                                      imp->imp_remote_handle.cookie,
 891                                      lustre_msg_get_handle(request->rq_repmsg)->
 892                                                                         cookie,
 893                                      (MSG_CONNECT_RECOVERING & msg_flags) ?
 894                                          " but is still in recovery" : "");
 895
 896                         imp->imp_remote_handle =
 897                                      *lustre_msg_get_handle(request->rq_repmsg);
 898
 899                         if (!(MSG_CONNECT_RECOVERING & msg_flags)) {
 900                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 901                                 GOTO(finish, rc = 0);
 902                         }
 903
 904                 } else {
 905                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 906                                obd2cli_tgt(imp->imp_obd),
 907                                imp->imp_connection->c_remote_uuid.uuid);
 908                 }
 909
 910                 if (imp->imp_invalid) {
 911                         CDEBUG(D_HA, "%s: reconnected but import is invalid; "
 912                                "marking evicted\n", imp->imp_obd->obd_name);
 913                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 914                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 915                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 916                                imp->imp_obd->obd_name,
 917                                obd2cli_tgt(imp->imp_obd));
 918
 919                         cfs_spin_lock(&imp->imp_lock);
 920                         imp->imp_resend_replay = 1;
 921                         cfs_spin_unlock(&imp->imp_lock);
 922
 923                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 924                 } else {
 925                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 926                 }
 927         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 928                 LASSERT(imp->imp_replayable);
 929                 imp->imp_remote_handle =
 930                                 *lustre_msg_get_handle(request->rq_repmsg);
 931                 imp->imp_last_replay_transno = 0;
 932                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 933         } else {
 934                 DEBUG_REQ(D_HA, request, "%s: evicting (reconnect/recover flags"
 935                           " not set: %x)", imp->imp_obd->obd_name, msg_flags);
 936                 imp->imp_remote_handle =
 937                                 *lustre_msg_get_handle(request->rq_repmsg);
 938                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 939         }
 940
 941         /* Sanity checks for a reconnected import. */
 942         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 943                 CERROR("imp_replayable flag does not match server "
 944                        "after reconnect. We should LBUG right here.\n");
 945         }
 946
 947         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 948             aa->pcaa_peer_committed) {
 949                 CERROR("%s went back in time (transno "LPD64
 950                        " was previously committed, server now claims "LPD64
 951                        ")!  See https://bugzilla.lustre.org/show_bug.cgi?"
 952                        "id=9646\n",
 953                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 954                        lustre_msg_get_last_committed(request->rq_repmsg));
 955         }
 956
 957 finish:
 958         rc = ptlrpc_import_recovery_state_machine(imp);
 959         if (rc != 0) {
 960                 if (rc == -ENOTCONN) {
 961                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 962                                "invalidating and reconnecting\n",
 963                                obd2cli_tgt(imp->imp_obd),
 964                                imp->imp_connection->c_remote_uuid.uuid);
 965                         ptlrpc_connect_import(imp, NULL);
 966                         RETURN(0);
 967                 }
 968         } else {
 969                 struct obd_connect_data *ocd;
 970                 struct obd_export *exp;
 971                 int ret;
 972                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 973                                            RCL_SERVER);
 974                 /* server replied obd_connect_data is always bigger */
 975                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 976                                                    &RMF_CONNECT_DATA, ret);
 977
 978                 cfs_spin_lock(&imp->imp_lock);
 979                 cfs_list_del(&imp->imp_conn_current->oic_item);
 980                 cfs_list_add(&imp->imp_conn_current->oic_item,
 981                              &imp->imp_conn_list);
 982                 imp->imp_last_success_conn =
 983                         imp->imp_conn_current->oic_last_attempt;
 984
 985                 if (ocd == NULL) {
 986                         cfs_spin_unlock(&imp->imp_lock);
 987                         CERROR("Wrong connect data from server\n");
 988                         rc = -EPROTO;
 989                         GOTO(out, rc);
 990                 }
 991
 992                 imp->imp_connect_data = *ocd;
 993
 994                 exp = class_conn2export(&imp->imp_dlm_handle);
 995                 cfs_spin_unlock(&imp->imp_lock);
 996
 997                 /* check that server granted subset of flags we asked for. */
 998                 LASSERTF((ocd->ocd_connect_flags &
 999                           imp->imp_connect_flags_orig) ==
1000                          ocd->ocd_connect_flags, LPX64" != "LPX64,
1001                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
1002
1003                 if (!exp) {
1004                         /* This could happen if export is cleaned during the
1005                            connect attempt */
1006                         CERROR("Missing export for %s\n",
1007                                imp->imp_obd->obd_name);
1008                         GOTO(out, rc = -ENODEV);
1009                 }
1010                 old_connect_flags = exp->exp_connect_flags;
1011                 exp->exp_connect_flags = ocd->ocd_connect_flags;
1012                 imp->imp_obd->obd_self_export->exp_connect_flags =
1013                                                         ocd->ocd_connect_flags;
1014                 class_export_put(exp);
1015
1016                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
1017
1018                 if (!ocd->ocd_ibits_known &&
1019                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
1020                         CERROR("Inodebits aware server returned zero compatible"
1021                                " bits?\n");
1022
1023                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1024                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
1025                                         LUSTRE_VERSION_OFFSET_WARN ||
1026                      ocd->ocd_version < LUSTRE_VERSION_CODE -
1027                                         LUSTRE_VERSION_OFFSET_WARN)) {
1028                         /* Sigh, some compilers do not like #ifdef in the middle
1029                            of macro arguments */
1030 #ifdef __KERNEL__
1031                         const char *older =
1032                                 "older. Consider upgrading this client";
1033 #else
1034                         const char *older =
1035                                 "older. Consider recompiling this application";
1036 #endif
1037                         const char *newer = "newer than client version";
1038
1039                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
1040                                       "is much %s (%s)\n",
1041                                       obd2cli_tgt(imp->imp_obd),
1042                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1043                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1044                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1045                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
1046                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
1047                                       newer : older, LUSTRE_VERSION_STRING);
1048                 }
1049
1050                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
1051                         /* We sent to the server ocd_cksum_types with bits set
1052                          * for algorithms we understand. The server masked off
1053                          * the checksum types it doesn't support */
1054                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
1055                                 LCONSOLE_WARN("The negotiation of the checksum "
1056                                               "alogrithm to use with server %s "
1057                                               "failed (%x/%x), disabling "
1058                                               "checksums\n",
1059                                               obd2cli_tgt(imp->imp_obd),
1060                                               ocd->ocd_cksum_types,
1061                                               OBD_CKSUM_ALL);
1062                                 cli->cl_checksum = 0;
1063                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1064                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
1065                         } else {
1066                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
1067
1068                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
1069                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
1070                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
1071                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
1072                                 else
1073                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1074                         }
1075                 } else {
1076                         /* The server does not support OBD_CONNECT_CKSUM.
1077                          * Enforce CRC32 for backward compatibility*/
1078                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
1079                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
1080                 }
1081
1082                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
1083                         cli->cl_max_pages_per_rpc =
1084                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
1085                 }
1086
1087                 /* Reset ns_connect_flags only for initial connect. It might be
1088                  * changed in while using FS and if we reset it in reconnect
1089                  * this leads to losing user settings done before such as
1090                  * disable lru_resize, etc. */
1091                 if (old_connect_flags != exp->exp_connect_flags ||
1092                     aa->pcaa_initial_connect) {
1093                         CDEBUG(D_HA, "%s: Resetting ns_connect_flags to server "
1094                                "flags: "LPX64"\n", imp->imp_obd->obd_name,
1095                               ocd->ocd_connect_flags);
1096                         imp->imp_obd->obd_namespace->ns_connect_flags =
1097                                 ocd->ocd_connect_flags;
1098                         imp->imp_obd->obd_namespace->ns_orig_connect_flags =
1099                                 ocd->ocd_connect_flags;
1100                 }
1101
1102                 if ((ocd->ocd_connect_flags & OBD_CONNECT_AT) &&
1103                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1104                         /* We need a per-message support flag, because
1105                            a. we don't know if the incoming connect reply
1106                               supports AT or not (in reply_in_callback)
1107                               until we unpack it.
1108                            b. failovered server means export and flags are gone
1109                               (in ptlrpc_send_reply).
1110                            Can only be set when we know AT is supported at
1111                            both ends */
1112                         imp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1113                 else
1114                         imp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1115
1116                 if ((ocd->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1117                     (imp->imp_msg_magic == LUSTRE_MSG_MAGIC_V2))
1118                         imp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1119                 else
1120                         imp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1121
1122                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
1123                         (cli->cl_max_pages_per_rpc > 0));
1124         }
1125
1126 out:
1127         if (rc != 0) {
1128                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
1129                 cfs_spin_lock(&imp->imp_lock);
1130                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
1131                     (request->rq_import_generation == imp->imp_generation))
1132                         ptlrpc_deactivate_and_unlock_import(imp);
1133                 else
1134                         cfs_spin_unlock(&imp->imp_lock);
1135
1136                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
1137                     (rc == -EACCES)) {
1138                         /*
1139                          * Give up trying to reconnect
1140                          * EACCES means client has no permission for connection
1141                          */
1142                         imp->imp_obd->obd_no_recov = 1;
1143                         ptlrpc_deactivate_import(imp);
1144                 }
1145
1146                 if (rc == -EPROTO) {
1147                         struct obd_connect_data *ocd;
1148
1149                         /* reply message might not be ready */
1150                         if (request->rq_repmsg == NULL)
1151                                 RETURN(-EPROTO);
1152
1153                         ocd = req_capsule_server_get(&request->rq_pill,
1154                                                      &RMF_CONNECT_DATA);
1155                         if (ocd &&
1156                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
1157                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
1158                            /* Actually servers are only supposed to refuse
1159                               connection from liblustre clients, so we should
1160                               never see this from VFS context */
1161                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
1162                                         "(%d.%d.%d.%d)"
1163                                         " refused connection from this client "
1164                                         "with an incompatible version (%s).  "
1165                                         "Client must be recompiled\n",
1166                                         obd2cli_tgt(imp->imp_obd),
1167                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
1168                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
1169                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
1170                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
1171                                         LUSTRE_VERSION_STRING);
1172                                 ptlrpc_deactivate_import(imp);
1173                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
1174                         }
1175                         RETURN(-EPROTO);
1176                 }
1177
1178                 ptlrpc_maybe_ping_import_soon(imp);
1179
1180                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
1181                        obd2cli_tgt(imp->imp_obd),
1182                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
1183         }
1184
1185         cfs_spin_lock(&imp->imp_lock);
1186         imp->imp_last_recon = 0;
1187         cfs_spin_unlock(&imp->imp_lock);
1188
1189         cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1190         RETURN(rc);
1191 }
1192
1193 static int completed_replay_interpret(const struct lu_env *env,
1194                                       struct ptlrpc_request *req,
1195                                       void * data, int rc)
1196 {
1197         ENTRY;
1198         cfs_atomic_dec(&req->rq_import->imp_replay_inflight);
1199         if (req->rq_status == 0 &&
1200             !req->rq_import->imp_vbr_failed) {
1201                 ptlrpc_import_recovery_state_machine(req->rq_import);
1202         } else {
1203                 if (req->rq_import->imp_vbr_failed) {
1204                         CDEBUG(D_WARNING,
1205                                "%s: version recovery fails, reconnecting\n",
1206                                req->rq_import->imp_obd->obd_name);
1207                 } else {
1208                         CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
1209                                      "reconnecting\n",
1210                                req->rq_import->imp_obd->obd_name,
1211                                req->rq_status);
1212                 }
1213                 ptlrpc_connect_import(req->rq_import, NULL);
1214         }
1215
1216         RETURN(0);
1217 }
1218
1219 static int signal_completed_replay(struct obd_import *imp)
1220 {
1221         struct ptlrpc_request *req;
1222         ENTRY;
1223
1224         LASSERT(cfs_atomic_read(&imp->imp_replay_inflight) == 0);
1225         cfs_atomic_inc(&imp->imp_replay_inflight);
1226
1227         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
1228                                         OBD_PING);
1229         if (req == NULL) {
1230                 cfs_atomic_dec(&imp->imp_replay_inflight);
1231                 RETURN(-ENOMEM);
1232         }
1233
1234         ptlrpc_request_set_replen(req);
1235         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
1236         lustre_msg_add_flags(req->rq_reqmsg,
1237                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
1238         if (AT_OFF)
1239                 req->rq_timeout *= 3;
1240         req->rq_interpret_reply = completed_replay_interpret;
1241
1242         ptlrpcd_add_req(req, PSCOPE_OTHER);
1243         RETURN(0);
1244 }
1245
1246 #ifdef __KERNEL__
1247 static int ptlrpc_invalidate_import_thread(void *data)
1248 {
1249         struct obd_import *imp = data;
1250
1251         ENTRY;
1252
1253         cfs_daemonize_ctxt("ll_imp_inval");
1254
1255         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1256                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1257                imp->imp_connection->c_remote_uuid.uuid);
1258
1259         ptlrpc_invalidate_import(imp);
1260
1261         if (obd_dump_on_eviction) {
1262                 CERROR("dump the log upon eviction\n");
1263                 libcfs_debug_dumplog();
1264         }
1265
1266         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1267         ptlrpc_import_recovery_state_machine(imp);
1268
1269         class_import_put(imp);
1270         RETURN(0);
1271 }
1272 #endif
1273
1274 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1275 {
1276         int rc = 0;
1277         int inflight;
1278         char *target_start;
1279         int target_len;
1280
1281         ENTRY;
1282         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1283                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1284                           &target_start, &target_len);
1285                 /* Don't care about MGC eviction */
1286                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1287                            LUSTRE_MGC_NAME) != 0) {
1288                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1289                                            "%.*s; in progress operations using "
1290                                            "this service will fail.\n",
1291                                            target_len, target_start);
1292                 }
1293                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1294                        obd2cli_tgt(imp->imp_obd),
1295                        imp->imp_connection->c_remote_uuid.uuid);
1296                 /* reset vbr_failed flag upon eviction */
1297                 cfs_spin_lock(&imp->imp_lock);
1298                 imp->imp_vbr_failed = 0;
1299                 cfs_spin_unlock(&imp->imp_lock);
1300
1301 #ifdef __KERNEL__
1302                 /* bug 17802:  XXX client_disconnect_export vs connect request
1303                  * race. if client will evicted at this time, we start
1304                  * invalidate thread without reference to import and import can
1305                  * be freed at same time. */
1306                 class_import_get(imp);
1307                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1308                                        CLONE_VM | CLONE_FILES);
1309                 if (rc < 0) {
1310                         class_import_put(imp);
1311                         CERROR("error starting invalidate thread: %d\n", rc);
1312                 } else {
1313                         rc = 0;
1314                 }
1315                 RETURN(rc);
1316 #else
1317                 ptlrpc_invalidate_import(imp);
1318
1319                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1320 #endif
1321         }
1322
1323         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1324                 CDEBUG(D_HA, "replay requested by %s\n",
1325                        obd2cli_tgt(imp->imp_obd));
1326                 rc = ptlrpc_replay_next(imp, &inflight);
1327                 if (inflight == 0 &&
1328                     cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1329                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1330                         rc = ldlm_replay_locks(imp);
1331                         if (rc)
1332                                 GOTO(out, rc);
1333                 }
1334                 rc = 0;
1335         }
1336
1337         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1338                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1339                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1340                         rc = signal_completed_replay(imp);
1341                         if (rc)
1342                                 GOTO(out, rc);
1343                 }
1344
1345         }
1346
1347         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1348                 if (cfs_atomic_read(&imp->imp_replay_inflight) == 0) {
1349                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1350                 }
1351         }
1352
1353         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1354                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1355                        obd2cli_tgt(imp->imp_obd),
1356                        imp->imp_connection->c_remote_uuid.uuid);
1357
1358                 rc = ptlrpc_resend(imp);
1359                 if (rc)
1360                         GOTO(out, rc);
1361                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1362                 ptlrpc_activate_import(imp);
1363
1364                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1365                           &target_start, &target_len);
1366                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1367                               "using nid %s.\n", imp->imp_obd->obd_name,
1368                               target_len, target_start,
1369                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1370         }
1371
1372         if (imp->imp_state == LUSTRE_IMP_FULL) {
1373                 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1374                 ptlrpc_wake_delayed(imp);
1375         }
1376
1377 out:
1378         RETURN(rc);
1379 }
1380
1381 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1382 {
1383         struct ptlrpc_request *req;
1384         int rq_opc, rc = 0;
1385         int nowait = imp->imp_obd->obd_force;
1386         ENTRY;
1387
1388         if (nowait)
1389                 GOTO(set_state, rc);
1390
1391         switch (imp->imp_connect_op) {
1392         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1393         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1394         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1395         default:
1396                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1397                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1398                 RETURN(-EINVAL);
1399         }
1400
1401         if (ptlrpc_import_in_recovery(imp)) {
1402                 struct l_wait_info lwi;
1403                 cfs_duration_t timeout;
1404
1405
1406                 if (AT_OFF) {
1407                         if (imp->imp_server_timeout)
1408                                 timeout = cfs_time_seconds(obd_timeout / 2);
1409                         else
1410                                 timeout = cfs_time_seconds(obd_timeout);
1411                 } else {
1412                         int idx = import_at_get_index(imp,
1413                                 imp->imp_client->cli_request_portal);
1414                         timeout = cfs_time_seconds(
1415                                 at_get(&imp->imp_at.iat_service_estimate[idx]));
1416                 }
1417
1418                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1419                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1420                 rc = l_wait_event(imp->imp_recovery_waitq,
1421                                   !ptlrpc_import_in_recovery(imp), &lwi);
1422
1423         }
1424
1425         cfs_spin_lock(&imp->imp_lock);
1426         if (imp->imp_state != LUSTRE_IMP_FULL)
1427                 GOTO(out, 0);
1428
1429         cfs_spin_unlock(&imp->imp_lock);
1430
1431         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1432                                         LUSTRE_OBD_VERSION, rq_opc);
1433         if (req) {
1434                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1435                  * it fails.  We can get through the above with a down server
1436                  * if the client doesn't know the server is gone yet. */
1437                 req->rq_no_resend = 1;
1438
1439 #ifndef CRAY_XT3
1440                 /* We want client umounts to happen quickly, no matter the
1441                    server state... */
1442                 req->rq_timeout = min_t(int, req->rq_timeout,
1443                                         INITIAL_CONNECT_TIMEOUT);
1444 #else
1445                 /* ... but we always want liblustre clients to nicely
1446                    disconnect, so only use the adaptive value. */
1447                 if (AT_OFF)
1448                         req->rq_timeout = obd_timeout / 3;
1449 #endif
1450
1451                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1452                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1453                 ptlrpc_request_set_replen(req);
1454                 rc = ptlrpc_queue_wait(req);
1455                 ptlrpc_req_finished(req);
1456         }
1457
1458 set_state:
1459         cfs_spin_lock(&imp->imp_lock);
1460 out:
1461         if (noclose)
1462                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1463         else
1464                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1465         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1466         /* Try all connections in the future - bz 12758 */
1467         imp->imp_last_recon = 0;
1468         cfs_spin_unlock(&imp->imp_lock);
1469
1470         RETURN(rc);
1471 }
1472
1473 void ptlrpc_cleanup_imp(struct obd_import *imp)
1474 {
1475         ENTRY;
1476
1477         cfs_spin_lock(&imp->imp_lock);
1478         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1479         imp->imp_generation++;
1480         cfs_spin_unlock(&imp->imp_lock);
1481         ptlrpc_abort_inflight(imp);
1482
1483         EXIT;
1484 }
1485
1486 /* Adaptive Timeout utils */
1487 extern unsigned int at_min, at_max, at_history;
1488
1489 /* Bin into timeslices using AT_BINS bins.
1490    This gives us a max of the last binlimit*AT_BINS secs without the storage,
1491    but still smoothing out a return to normalcy from a slow response.
1492    (E.g. remember the maximum latency in each minute of the last 4 minutes.) */
1493 int at_measured(struct adaptive_timeout *at, unsigned int val)
1494 {
1495         unsigned int old = at->at_current;
1496         time_t now = cfs_time_current_sec();
1497         time_t binlimit = max_t(time_t, at_history / AT_BINS, 1);
1498
1499         LASSERT(at);
1500         CDEBUG(D_OTHER, "add %u to %p time=%lu v=%u (%u %u %u %u)\n",
1501                val, at, now - at->at_binstart, at->at_current,
1502                at->at_hist[0], at->at_hist[1], at->at_hist[2], at->at_hist[3]);
1503
1504         if (val == 0)
1505                 /* 0's don't count, because we never want our timeout to
1506                    drop to 0, and because 0 could mean an error */
1507                 return 0;
1508
1509         cfs_spin_lock(&at->at_lock);
1510
1511         if (unlikely(at->at_binstart == 0)) {
1512                 /* Special case to remove default from history */
1513                 at->at_current = val;
1514                 at->at_worst_ever = val;
1515                 at->at_worst_time = now;
1516                 at->at_hist[0] = val;
1517                 at->at_binstart = now;
1518         } else if (now - at->at_binstart < binlimit ) {
1519                 /* in bin 0 */
1520                 at->at_hist[0] = max(val, at->at_hist[0]);
1521                 at->at_current = max(val, at->at_current);
1522         } else {
1523                 int i, shift;
1524                 unsigned int maxv = val;
1525                 /* move bins over */
1526                 shift = (now - at->at_binstart) / binlimit;
1527                 LASSERT(shift > 0);
1528                 for(i = AT_BINS - 1; i >= 0; i--) {
1529                         if (i >= shift) {
1530                                 at->at_hist[i] = at->at_hist[i - shift];
1531                                 maxv = max(maxv, at->at_hist[i]);
1532                         } else {
1533                                 at->at_hist[i] = 0;
1534                         }
1535                 }
1536                 at->at_hist[0] = val;
1537                 at->at_current = maxv;
1538                 at->at_binstart += shift * binlimit;
1539         }
1540
1541         if (at->at_current > at->at_worst_ever) {
1542                 at->at_worst_ever = at->at_current;
1543                 at->at_worst_time = now;
1544         }
1545
1546         if (at->at_flags & AT_FLG_NOHIST)
1547                 /* Only keep last reported val; keeping the rest of the history
1548                    for proc only */
1549                 at->at_current = val;
1550
1551         if (at_max > 0)
1552                 at->at_current =  min(at->at_current, at_max);
1553         at->at_current =  max(at->at_current, at_min);
1554
1555         if (at->at_current != old)
1556                 CDEBUG(D_OTHER, "AT %p change: old=%u new=%u delta=%d "
1557                        "(val=%u) hist %u %u %u %u\n", at,
1558                        old, at->at_current, at->at_current - old, val,
1559                        at->at_hist[0], at->at_hist[1], at->at_hist[2],
1560                        at->at_hist[3]);
1561
1562         /* if we changed, report the old value */
1563         old = (at->at_current != old) ? old : 0;
1564
1565         cfs_spin_unlock(&at->at_lock);
1566         return old;
1567 }
1568
1569 /* Find the imp_at index for a given portal; assign if space available */
1570 int import_at_get_index(struct obd_import *imp, int portal)
1571 {
1572         struct imp_at *at = &imp->imp_at;
1573         int i;
1574
1575         for (i = 0; i < IMP_AT_MAX_PORTALS; i++) {
1576                 if (at->iat_portal[i] == portal)
1577                         return i;
1578                 if (at->iat_portal[i] == 0)
1579                         /* unused */
1580                         break;
1581         }
1582
1583         /* Not found in list, add it under a lock */
1584         cfs_spin_lock(&imp->imp_lock);
1585
1586         /* Check unused under lock */
1587         for (; i < IMP_AT_MAX_PORTALS; i++) {
1588                 if (at->iat_portal[i] == portal)
1589                         goto out;
1590                 if (at->iat_portal[i] == 0)
1591                         /* unused */
1592                         break;
1593         }
1594
1595         /* Not enough portals? */
1596         LASSERT(i < IMP_AT_MAX_PORTALS);
1597
1598         at->iat_portal[i] = portal;
1599 out:
1600         cfs_spin_unlock(&imp->imp_lock);
1601         return i;
1602 }