lustre/ptlrpc/import.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
   5  *   Author: Mike Shaver <shaver@clusterfs.com>
   6  *
   7  *   This file is part of the Lustre file system, http://www.lustre.org
   8  *   Lustre is a trademark of Cluster File Systems, Inc.
   9  *
  10  *   You may have signed or agreed to another license before downloading
  11  *   this software.  If so, you are bound by the terms and conditions
  12  *   of that agreement, and the following does not apply to you.  See the
  13  *   LICENSE file included with this distribution for more information.
  14  *
  15  *   If you did not agree to a different license, then this copy of Lustre
  16  *   is open source software; you can redistribute it and/or modify it
  17  *   under the terms of version 2 of the GNU General Public License as
  18  *   published by the Free Software Foundation.
  19  *
  20  *   In either case, Lustre is distributed in the hope that it will be
  21  *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
  22  *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  23  *   license text for more details.
  24  */
  25
  26 #define DEBUG_SUBSYSTEM S_RPC
  27 #ifndef __KERNEL__
  28 # include <liblustre.h>
  29 #endif
  30
  31 #include <obd_support.h>
  32 #include <lustre_ha.h>
  33 #include <lustre_net.h>
  34 #include <lustre_import.h>
  35 #include <lustre_export.h>
  36 #include <obd.h>
  37 #include <obd_class.h>
  38 #include <lustre_ver.h>
  39
  40 #include "ptlrpc_internal.h"
  41
  42 struct ptlrpc_connect_async_args {
  43          __u64 pcaa_peer_committed;
  44         int pcaa_initial_connect;
  45 };
  46
  47 /* A CLOSED import should remain so. */
  48 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  49 do {                                                                           \
  50         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  51                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  52                       imp, obd2cli_tgt(imp->imp_obd),                          \
  53                       ptlrpc_import_state_name(imp->imp_state),                \
  54                       ptlrpc_import_state_name(state));                        \
  55                imp->imp_state = state;                                         \
  56         }                                                                      \
  57 } while(0)
  58
  59 #define IMPORT_SET_STATE(imp, state)            \
  60 do {                                            \
  61         spin_lock(&imp->imp_lock);              \
  62         IMPORT_SET_STATE_NOLOCK(imp, state);    \
  63         spin_unlock(&imp->imp_lock);            \
  64 } while(0)
  65
  66
  67 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  68                                     void * data, int rc);
  69 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  70
  71 /* Only this function is allowed to change the import state when it is
  72  * CLOSED. I would rather refcount the import and free it after
  73  * disconnection like we do with exports. To do that, the client_obd
  74  * will need to save the peer info somewhere other than in the import,
  75  * though. */
  76 int ptlrpc_init_import(struct obd_import *imp)
  77 {
  78         spin_lock(&imp->imp_lock);
  79
  80         imp->imp_generation++;
  81         imp->imp_state =  LUSTRE_IMP_NEW;
  82
  83         spin_unlock(&imp->imp_lock);
  84
  85         return 0;
  86 }
  87 EXPORT_SYMBOL(ptlrpc_init_import);
  88
  89 #define UUID_STR "_UUID"
  90 static void deuuidify(char *uuid, const char *prefix, char **uuid_start,
  91                       int *uuid_len)
  92 {
  93         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
  94                 ? uuid : uuid + strlen(prefix);
  95
  96         *uuid_len = strlen(*uuid_start);
  97
  98         if (*uuid_len < strlen(UUID_STR))
  99                 return;
 100
 101         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 102                     UUID_STR, strlen(UUID_STR)))
 103                 *uuid_len -= strlen(UUID_STR);
 104 }
 105
 106 /* Returns true if import was FULL, false if import was already not
 107  * connected.
 108  * @imp - import to be disconnected
 109  * @conn_cnt - connection count (epoch) of the request that timed out
 110  *             and caused the disconnection.  In some cases, multiple
 111  *             inflight requests can fail to a single target (e.g. OST
 112  *             bulk requests) and if one has already caused a reconnection
 113  *             (increasing the import->conn_cnt) the older failure should
 114  *             not also cause a reconnection.  If zero it forces a reconnect.
 115  */
 116 int ptlrpc_set_import_discon(struct obd_import *imp, __u32 conn_cnt)
 117 {
 118         int rc = 0;
 119
 120         spin_lock(&imp->imp_lock);
 121
 122         if (imp->imp_state == LUSTRE_IMP_FULL &&
 123             (conn_cnt == 0 || conn_cnt == imp->imp_conn_cnt)) {
 124                 char *target_start;
 125                 int   target_len;
 126
 127                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 128                           &target_start, &target_len);
 129
 130                 LCONSOLE_ERROR("%s: Connection to service %.*s via nid %s was "
 131                                "lost; in progress operations using this "
 132                                "service will %s.\n", imp->imp_obd->obd_name,
 133                                target_len, target_start,
 134                                libcfs_nid2str(imp->imp_connection->c_peer.nid),
 135                                imp->imp_replayable ?
 136                                       "wait for recovery to complete" : "fail");
 137
 138                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 139                 spin_unlock(&imp->imp_lock);
 140
 141                 if (obd_dump_on_timeout)
 142                         libcfs_debug_dumplog();
 143
 144                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 145                 rc = 1;
 146         } else {
 147                 spin_unlock(&imp->imp_lock);
 148                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 149                        imp->imp_client->cli_name, imp,
 150                        (imp->imp_state == LUSTRE_IMP_FULL &&
 151                         imp->imp_conn_cnt > conn_cnt) ?
 152                        "reconnected" : "not connected", imp->imp_conn_cnt,
 153                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 154         }
 155
 156         return rc;
 157 }
 158
 159 /*
 160  * This acts as a barrier; all existing requests are rejected, and
 161  * no new requests will be accepted until the import is valid again.
 162  */
 163 void ptlrpc_deactivate_import(struct obd_import *imp)
 164 {
 165         ENTRY;
 166
 167         spin_lock(&imp->imp_lock);
 168         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 169         imp->imp_invalid = 1;
 170         imp->imp_generation++;
 171         spin_unlock(&imp->imp_lock);
 172
 173         ptlrpc_abort_inflight(imp);
 174         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 175 }
 176
 177 /*
 178  * This function will invalidate the import, if necessary, then block
 179  * for all the RPC completions, and finally notify the obd to
 180  * invalidate its state (ie cancel locks, clear pending requests,
 181  * etc).
 182  */
 183 void ptlrpc_invalidate_import(struct obd_import *imp)
 184 {
 185         struct l_wait_info lwi;
 186         int rc;
 187
 188         if (!imp->imp_invalid)
 189                 ptlrpc_deactivate_import(imp);
 190
 191         LASSERT(imp->imp_invalid);
 192
 193         /* wait for all requests to error out and call completion callbacks */
 194         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 195                                    HZ, NULL, NULL);
 196         rc = l_wait_event(imp->imp_recovery_waitq,
 197                           (atomic_read(&imp->imp_inflight) == 0),
 198                           &lwi);
 199
 200         if (rc)
 201                 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 202                        obd2cli_tgt(imp->imp_obd), rc,
 203                        atomic_read(&imp->imp_inflight));
 204
 205         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 206         sptlrpc_import_flush_all_ctx(imp);
 207 }
 208
 209 /* unset imp_invalid */
 210 void ptlrpc_activate_import(struct obd_import *imp)
 211 {
 212         struct obd_device *obd = imp->imp_obd;
 213
 214         spin_lock(&imp->imp_lock);
 215         imp->imp_invalid = 0;
 216         spin_unlock(&imp->imp_lock);
 217
 218         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 219 }
 220
 221 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 222 {
 223         ENTRY;
 224
 225         LASSERT(!imp->imp_dlm_fake);
 226
 227         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 228                 if (!imp->imp_replayable) {
 229                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 230                                "auto-deactivating\n",
 231                                obd2cli_tgt(imp->imp_obd),
 232                                imp->imp_connection->c_remote_uuid.uuid,
 233                                imp->imp_obd->obd_name);
 234                         ptlrpc_deactivate_import(imp);
 235                 }
 236
 237                 CDEBUG(D_HA, "%s: waking up pinger\n",
 238                        obd2cli_tgt(imp->imp_obd));
 239
 240                 spin_lock(&imp->imp_lock);
 241                 imp->imp_force_verify = 1;
 242                 spin_unlock(&imp->imp_lock);
 243
 244                 ptlrpc_pinger_wake_up();
 245         }
 246         EXIT;
 247 }
 248
 249 static int import_select_connection(struct obd_import *imp)
 250 {
 251         struct obd_import_conn *imp_conn = NULL, *conn;
 252         struct obd_export *dlmexp;
 253         ENTRY;
 254
 255         spin_lock(&imp->imp_lock);
 256
 257         if (list_empty(&imp->imp_conn_list)) {
 258                 CERROR("%s: no connections available\n",
 259                         imp->imp_obd->obd_name);
 260                 spin_unlock(&imp->imp_lock);
 261                 RETURN(-EINVAL);
 262         }
 263
 264         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 265                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 266                        imp->imp_obd->obd_name,
 267                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 268                        conn->oic_last_attempt);
 269                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 270                 if (cfs_time_before_64(conn->oic_last_attempt +
 271                                        RECONNECT_INTERVAL * HZ,
 272                                        cfs_time_current_64())) {
 273                         /* If we have never tried this connection since the
 274                            the last successful attempt, go with this one */
 275                         if (cfs_time_before_64(conn->oic_last_attempt,
 276                                                imp->imp_last_success_conn)) {
 277                                 imp_conn = conn;
 278                                 break;
 279                         }
 280
 281                         /* Both of these connections have already been tried
 282                            since the last successful connection; just choose the
 283                            least recently used */
 284                         if (!imp_conn)
 285                                 imp_conn = conn;
 286                         else if (cfs_time_before_64(conn->oic_last_attempt,
 287                                                     imp_conn->oic_last_attempt))
 288                                 imp_conn = conn;
 289                 }
 290         }
 291
 292         /* if not found, simply choose the current one */
 293         if (!imp_conn) {
 294                 LASSERT(imp->imp_conn_current);
 295                 imp_conn = imp->imp_conn_current;
 296         }
 297         LASSERT(imp_conn->oic_conn);
 298
 299         imp_conn->oic_last_attempt = cfs_time_current_64();
 300
 301         /* switch connection, don't mind if it's same as the current one */
 302         if (imp->imp_connection)
 303                 ptlrpc_put_connection(imp->imp_connection);
 304         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 305
 306         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 307         LASSERT(dlmexp != NULL);
 308         if (dlmexp->exp_connection)
 309                 ptlrpc_put_connection(dlmexp->exp_connection);
 310         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 311         class_export_put(dlmexp);
 312
 313         if (imp->imp_conn_current != imp_conn) {
 314                 if (imp->imp_conn_current)
 315                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 316                                       imp->imp_obd->obd_name,
 317                                       imp_conn->oic_uuid.uuid,
 318                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 319                 imp->imp_conn_current = imp_conn;
 320         }
 321
 322         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 323                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 324                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 325
 326         spin_unlock(&imp->imp_lock);
 327
 328         RETURN(0);
 329 }
 330
 331 /*
 332  * must be called under imp_lock
 333  */
 334 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 335 {
 336         struct ptlrpc_request *req;
 337         struct list_head *tmp;
 338
 339         if (list_empty(&imp->imp_replay_list))
 340                 return 0;
 341         tmp = imp->imp_replay_list.next;
 342         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 343         *transno = req->rq_transno;
 344         return 1;
 345 }
 346
 347 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 348 {
 349         struct obd_device *obd = imp->imp_obd;
 350         int initial_connect = 0;
 351         int set_transno = 0;
 352         int rc;
 353         __u64 committed_before_reconnect = 0;
 354         struct ptlrpc_request *request;
 355         int size[] = { sizeof(struct ptlrpc_body),
 356                        sizeof(imp->imp_obd->u.cli.cl_target_uuid),
 357                        sizeof(obd->obd_uuid),
 358                        sizeof(imp->imp_dlm_handle),
 359                        sizeof(imp->imp_connect_data) };
 360         char *tmp[] = { NULL,
 361                         obd2cli_tgt(imp->imp_obd),
 362                         obd->obd_uuid.uuid,
 363                         (char *)&imp->imp_dlm_handle,
 364                         (char *)&imp->imp_connect_data };
 365         struct ptlrpc_connect_async_args *aa;
 366         ENTRY;
 367
 368         spin_lock(&imp->imp_lock);
 369         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 370                 spin_unlock(&imp->imp_lock);
 371                 CERROR("can't connect to a closed import\n");
 372                 RETURN(-EINVAL);
 373         } else if (imp->imp_state == LUSTRE_IMP_FULL &&
 374                    imp->imp_force_reconnect == 0) {
 375                 spin_unlock(&imp->imp_lock);
 376                 CERROR("already connected\n");
 377                 RETURN(0);
 378         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 379                 spin_unlock(&imp->imp_lock);
 380                 CERROR("already connecting\n");
 381                 RETURN(-EALREADY);
 382         }
 383
 384         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 385
 386         imp->imp_conn_cnt++;
 387         imp->imp_resend_replay = 0;
 388
 389         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 390                 initial_connect = 1;
 391         else
 392                 committed_before_reconnect = imp->imp_peer_committed_transno;
 393
 394         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 395         spin_unlock(&imp->imp_lock);
 396
 397         if (new_uuid) {
 398                 struct obd_uuid uuid;
 399
 400                 obd_str2uuid(&uuid, new_uuid);
 401                 rc = import_set_conn_priority(imp, &uuid);
 402                 if (rc)
 403                         GOTO(out, rc);
 404         }
 405
 406         rc = import_select_connection(imp);
 407         if (rc)
 408                 GOTO(out, rc);
 409
 410         /* last in connection list */
 411         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 412                 if (imp->imp_initial_recov_bk && initial_connect) {
 413                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 414                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 415                         /* Don't retry if connect fails */
 416                         rc = 0;
 417                         obd_set_info_async(obd->obd_self_export,
 418                                            strlen(KEY_INIT_RECOV),
 419                                            KEY_INIT_RECOV,
 420                                            sizeof(rc), &rc, NULL);
 421                 }
 422                 if (imp->imp_recon_bk) {
 423                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 424                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 425                         imp->imp_last_recon = 1;
 426                 }
 427         }
 428
 429         /* Reset connect flags to the originally requested flags, in case
 430          * the server is updated on-the-fly we will get the new features. */
 431         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 432         rc = obd_reconnect(imp->imp_obd->obd_self_export, obd,
 433                            &obd->obd_uuid, &imp->imp_connect_data);
 434         if (rc)
 435                 GOTO(out, rc);
 436
 437         request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, imp->imp_connect_op,
 438                                   5, size, tmp);
 439         if (!request)
 440                 GOTO(out, rc = -ENOMEM);
 441
 442 #ifndef __KERNEL__
 443         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 444 #endif
 445         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 446
 447         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 448         /* Allow a slightly larger reply for future growth compatibility */
 449         size[REPLY_REC_OFF] = sizeof(struct obd_connect_data) +
 450                               16 * sizeof(__u64);
 451         ptlrpc_req_set_repsize(request, 2, size);
 452         request->rq_interpret_reply = ptlrpc_connect_interpret;
 453
 454         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 455         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 456         memset(aa, 0, sizeof *aa);
 457
 458         aa->pcaa_peer_committed = committed_before_reconnect;
 459         aa->pcaa_initial_connect = initial_connect;
 460
 461         if (aa->pcaa_initial_connect) {
 462                 imp->imp_replayable = 1;
 463                 /* On an initial connect, we don't know which one of a
 464                    failover server pair is up.  Don't wait long. */
 465                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 466                 lustre_msg_add_op_flags(request->rq_reqmsg,
 467                                         MSG_CONNECT_INITIAL);
 468
 469         }
 470
 471         if (set_transno)
 472                 lustre_msg_add_op_flags(request->rq_reqmsg,
 473                                         MSG_CONNECT_TRANSNO);
 474
 475         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 476         ptlrpcd_add_req(request);
 477         rc = 0;
 478 out:
 479         if (rc != 0) {
 480                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 481         }
 482
 483         RETURN(rc);
 484 }
 485 EXPORT_SYMBOL(ptlrpc_connect_import);
 486
 487 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 488 {
 489         struct obd_import_conn *imp_conn;
 490         int wake_pinger = 0;
 491
 492         ENTRY;
 493
 494         spin_lock(&imp->imp_lock);
 495         if (list_empty(&imp->imp_conn_list))
 496                 GOTO(unlock, 0);
 497
 498         imp_conn = list_entry(imp->imp_conn_list.prev,
 499                               struct obd_import_conn,
 500                               oic_item);
 501
 502         if (imp->imp_conn_current != imp_conn) {
 503                 ptlrpc_ping_import_soon(imp);
 504                 wake_pinger = 1;
 505         }
 506
 507  unlock:
 508         spin_unlock(&imp->imp_lock);
 509
 510         if (wake_pinger)
 511                 ptlrpc_pinger_wake_up();
 512
 513         EXIT;
 514 }
 515
 516 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 517                                     void * data, int rc)
 518 {
 519         struct ptlrpc_connect_async_args *aa = data;
 520         struct obd_import *imp = request->rq_import;
 521         struct client_obd *cli = &imp->imp_obd->u.cli;
 522         struct lustre_handle old_hdl;
 523         int msg_flags;
 524         ENTRY;
 525
 526         spin_lock(&imp->imp_lock);
 527         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 528                 spin_unlock(&imp->imp_lock);
 529                 RETURN(0);
 530         }
 531         imp->imp_force_reconnect = 0;
 532         spin_unlock(&imp->imp_lock);
 533
 534         if (rc)
 535                 GOTO(out, rc);
 536
 537         rc = sptlrpc_cli_install_rvs_ctx(imp, request->rq_cli_ctx);
 538         if (rc)
 539                 GOTO(out, rc);
 540
 541         LASSERT(imp->imp_conn_current);
 542
 543         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 544
 545         /* All imports are pingable */
 546         imp->imp_pingable = 1;
 547
 548         if (aa->pcaa_initial_connect) {
 549                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 550                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 551                                obd2cli_tgt(imp->imp_obd));
 552                         imp->imp_replayable = 1;
 553                 } else {
 554                         imp->imp_replayable = 0;
 555                 }
 556
 557                 if (msg_flags & MSG_CONNECT_NEXT_VER) {
 558                         imp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
 559                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v2\n",
 560                                obd2cli_tgt(imp->imp_obd));
 561                 } else {
 562                         CDEBUG(D_RPCTRACE, "connect to %s with lustre_msg_v1\n",
 563                                obd2cli_tgt(imp->imp_obd));
 564                 }
 565
 566                 imp->imp_remote_handle =
 567                                 *lustre_msg_get_handle(request->rq_repmsg);
 568
 569                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 570                 GOTO(finish, rc = 0);
 571         }
 572
 573         /* Determine what recovery state to move the import to. */
 574         if (MSG_CONNECT_RECONNECT & msg_flags) {
 575                 memset(&old_hdl, 0, sizeof(old_hdl));
 576                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 577                             sizeof (old_hdl))) {
 578                         CERROR("%s@%s didn't like our handle "LPX64
 579                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 580                                imp->imp_connection->c_remote_uuid.uuid,
 581                                imp->imp_dlm_handle.cookie);
 582                         GOTO(out, rc = -ENOTCONN);
 583                 }
 584
 585                 if (memcmp(&imp->imp_remote_handle,
 586                            lustre_msg_get_handle(request->rq_repmsg),
 587                            sizeof(imp->imp_remote_handle))) {
 588                         int level = D_ERROR;
 589                         /* The MGC does this all the time */
 590                         if (strcmp(imp->imp_obd->obd_type->typ_name,
 591                                    LUSTRE_MGC_NAME) == 0) {
 592                                 level = D_CONFIG;
 593                         }
 594                         CDEBUG(level,
 595                                "%s@%s changed handle from "LPX64" to "LPX64
 596                                "; copying, but this may foreshadow disaster\n",
 597                                obd2cli_tgt(imp->imp_obd),
 598                                imp->imp_connection->c_remote_uuid.uuid,
 599                                imp->imp_remote_handle.cookie,
 600                                lustre_msg_get_handle(request->rq_repmsg)->
 601                                         cookie);
 602                         imp->imp_remote_handle =
 603                                      *lustre_msg_get_handle(request->rq_repmsg);
 604                 } else {
 605                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 606                                obd2cli_tgt(imp->imp_obd),
 607                                imp->imp_connection->c_remote_uuid.uuid);
 608                 }
 609
 610                 if (imp->imp_invalid) {
 611                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 612                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 613                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 614                                imp->imp_obd->obd_name,
 615                                obd2cli_tgt(imp->imp_obd));
 616                         imp->imp_resend_replay = 1;
 617                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 618                 } else {
 619                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 620                 }
 621         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 622                 LASSERT(imp->imp_replayable);
 623                 imp->imp_remote_handle =
 624                                 *lustre_msg_get_handle(request->rq_repmsg);
 625                 imp->imp_last_replay_transno = 0;
 626                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 627         } else {
 628                 imp->imp_remote_handle =
 629                                 *lustre_msg_get_handle(request->rq_repmsg);
 630                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 631         }
 632
 633         /* Sanity checks for a reconnected import. */
 634         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 635                 CERROR("imp_replayable flag does not match server "
 636                        "after reconnect. We should LBUG right here.\n");
 637         }
 638
 639         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 640             aa->pcaa_peer_committed) {
 641                 CERROR("%s went back in time (transno "LPD64
 642                        " was previously committed, server now claims "LPD64
 643                        ")!  See https://bugzilla.clusterfs.com/"
 644                        "long_list.cgi?buglist=9646\n",
 645                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 646                        lustre_msg_get_last_committed(request->rq_repmsg));
 647         }
 648
 649 finish:
 650         rc = ptlrpc_import_recovery_state_machine(imp);
 651         if (rc != 0) {
 652                 if (rc == -ENOTCONN) {
 653                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 654                                "invalidating and reconnecting\n",
 655                                obd2cli_tgt(imp->imp_obd),
 656                                imp->imp_connection->c_remote_uuid.uuid);
 657                         ptlrpc_connect_import(imp, NULL);
 658                         RETURN(0);
 659                 }
 660         } else {
 661                 struct obd_connect_data *ocd;
 662                 struct obd_export *exp;
 663
 664                 ocd = lustre_swab_repbuf(request, REPLY_REC_OFF, sizeof(*ocd),
 665                                          lustre_swab_connect);
 666
 667                 spin_lock(&imp->imp_lock);
 668                 list_del(&imp->imp_conn_current->oic_item);
 669                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 670                 imp->imp_last_success_conn =
 671                         imp->imp_conn_current->oic_last_attempt;
 672
 673                 if (ocd == NULL) {
 674                         spin_unlock(&imp->imp_lock);
 675                         CERROR("Wrong connect data from server\n");
 676                         rc = -EPROTO;
 677                         GOTO(out, rc);
 678                 }
 679
 680                 imp->imp_connect_data = *ocd;
 681
 682                 exp = class_conn2export(&imp->imp_dlm_handle);
 683                 spin_unlock(&imp->imp_lock);
 684
 685                 /* check that server granted subset of flags we asked for. */
 686                 LASSERTF((ocd->ocd_connect_flags &
 687                           imp->imp_connect_flags_orig) ==
 688                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 689                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 690
 691                 if (!exp) {
 692                         /* This could happen if export is cleaned during the
 693                            connect attempt */
 694                         spin_unlock(&imp->imp_lock);
 695                         CERROR("Missing export for %s\n",
 696                                imp->imp_obd->obd_name);
 697                         GOTO(out, rc = -ENODEV);
 698                 }
 699                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 700                 class_export_put(exp);
 701
 702                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 703
 704                 if (!ocd->ocd_ibits_known &&
 705                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 706                         CERROR("Inodebits aware server returned zero compatible"
 707                                " bits?\n");
 708
 709                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 710                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 711                     LUSTRE_VERSION_OFFSET_WARN)) {
 712                         /* Sigh, some compilers do not like #ifdef in the middle
 713                            of macro arguments */
 714 #ifdef __KERNEL__
 715                         const char *action = "upgrading this client";
 716 #else
 717                         const char *action = "recompiling this application";
 718 #endif
 719
 720                         CWARN("Server %s version (%d.%d.%d.%d) is much newer. "
 721                               "Consider %s (%s).\n",
 722                               obd2cli_tgt(imp->imp_obd),
 723                               OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 724                               OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 725                               OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 726                               OBD_OCD_VERSION_FIX(ocd->ocd_version),
 727                               action, LUSTRE_VERSION_STRING);
 728                 }
 729
 730                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 731                         cli->cl_max_pages_per_rpc = ocd->ocd_brw_size >> PAGE_SHIFT;
 732                 } else {
 733                         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 734                                                         (int)(1024 * 1024 >> PAGE_SHIFT));
 735                 }
 736
 737                 LASSERT(cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES &&
 738                         cli->cl_max_pages_per_rpc > 0);
 739
 740         }
 741
 742  out:
 743         if (rc != 0) {
 744                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 745                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov)
 746                         ptlrpc_deactivate_import(imp);
 747
 748                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 749                     (rc == -EACCES)) {
 750                         /*
 751                          * Give up trying to reconnect
 752                          * EACCES means client has no permission for connection
 753                          */
 754                         imp->imp_obd->obd_no_recov = 1;
 755                         ptlrpc_deactivate_import(imp);
 756                 }
 757
 758                 if (rc == -EPROTO) {
 759                         struct obd_connect_data *ocd;
 760
 761                         /* reply message might not be ready */
 762                         if (request->rq_repmsg != NULL)
 763                                 RETURN(-EPROTO);
 764
 765                         ocd = lustre_swab_repbuf(request, REPLY_REC_OFF,
 766                                                  sizeof *ocd,
 767                                                  lustre_swab_connect);
 768                         if (ocd &&
 769                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 770                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 771                            /* Actually servers are only supposed to refuse
 772                               connection from liblustre clients, so we should
 773                               never see this from VFS context */
 774                                 CERROR("Server %s version (%d.%d.%d.%d) "
 775                                        "refused connection from this client "
 776                                        "as too old version (%s).  Client must "
 777                                        "be recompiled\n",
 778                                       obd2cli_tgt(imp->imp_obd),
 779                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 780                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 781                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 782                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 783                                       LUSTRE_VERSION_STRING);
 784                                 ptlrpc_deactivate_import(imp);
 785                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 786                         }
 787                         RETURN(-EPROTO);
 788                 }
 789
 790                 ptlrpc_maybe_ping_import_soon(imp);
 791
 792                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 793                        obd2cli_tgt(imp->imp_obd),
 794                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 795         }
 796
 797         imp->imp_last_recon = 0;
 798
 799         cfs_waitq_signal(&imp->imp_recovery_waitq);
 800         RETURN(rc);
 801 }
 802
 803 static int completed_replay_interpret(struct ptlrpc_request *req,
 804                                     void * data, int rc)
 805 {
 806         ENTRY;
 807         atomic_dec(&req->rq_import->imp_replay_inflight);
 808         if (req->rq_status == 0) {
 809                 ptlrpc_import_recovery_state_machine(req->rq_import);
 810         } else {
 811                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 812                        "reconnecting\n",
 813                        req->rq_import->imp_obd->obd_name, req->rq_status);
 814                 ptlrpc_connect_import(req->rq_import, NULL);
 815         }
 816
 817         RETURN(0);
 818 }
 819
 820 static int signal_completed_replay(struct obd_import *imp)
 821 {
 822         struct ptlrpc_request *req;
 823         ENTRY;
 824
 825         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 826         atomic_inc(&imp->imp_replay_inflight);
 827
 828         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 1, NULL, NULL);
 829         if (!req) {
 830                 atomic_dec(&imp->imp_replay_inflight);
 831                 RETURN(-ENOMEM);
 832         }
 833
 834         ptlrpc_req_set_repsize(req, 1, NULL);
 835         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 836         lustre_msg_add_flags(req->rq_reqmsg,
 837                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
 838         req->rq_timeout *= 3;
 839         req->rq_interpret_reply = completed_replay_interpret;
 840
 841         ptlrpcd_add_req(req);
 842         RETURN(0);
 843 }
 844
 845 #ifdef __KERNEL__
 846 static int ptlrpc_invalidate_import_thread(void *data)
 847 {
 848         struct obd_import *imp = data;
 849
 850         ENTRY;
 851
 852         ptlrpc_daemonize("ll_imp_inval");
 853
 854         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 855                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
 856                imp->imp_connection->c_remote_uuid.uuid);
 857
 858         ptlrpc_invalidate_import(imp);
 859
 860         if (obd_dump_on_eviction) {
 861                 CERROR("dump the log upon eviction\n");
 862                 libcfs_debug_dumplog();
 863         }
 864
 865         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 866         ptlrpc_import_recovery_state_machine(imp);
 867
 868         RETURN(0);
 869 }
 870 #endif
 871
 872 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 873 {
 874         int rc = 0;
 875         int inflight;
 876         char *target_start;
 877         int target_len;
 878
 879         ENTRY;
 880         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 881                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 882                           &target_start, &target_len);
 883                 LCONSOLE_ERROR("This client was evicted by %.*s; in progress "
 884                                "operations using this service will fail.\n",
 885                                target_len, target_start);
 886                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 887                        obd2cli_tgt(imp->imp_obd),
 888                        imp->imp_connection->c_remote_uuid.uuid);
 889
 890 #ifdef __KERNEL__
 891                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
 892                                        CLONE_VM | CLONE_FILES);
 893                 if (rc < 0)
 894                         CERROR("error starting invalidate thread: %d\n", rc);
 895                 else
 896                         rc = 0;
 897                 RETURN(rc);
 898 #else
 899                 ptlrpc_invalidate_import(imp);
 900
 901                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 902 #endif
 903         }
 904
 905         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 906                 CDEBUG(D_HA, "replay requested by %s\n",
 907                        obd2cli_tgt(imp->imp_obd));
 908                 rc = ptlrpc_replay_next(imp, &inflight);
 909                 if (inflight == 0 &&
 910                     atomic_read(&imp->imp_replay_inflight) == 0) {
 911                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 912                         rc = ldlm_replay_locks(imp);
 913                         if (rc)
 914                                 GOTO(out, rc);
 915                 }
 916                 rc = 0;
 917         }
 918
 919         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 920                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 921                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
 922                         rc = signal_completed_replay(imp);
 923                         if (rc)
 924                                 GOTO(out, rc);
 925                 }
 926
 927         }
 928
 929         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
 930                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 931                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 932                 }
 933         }
 934
 935         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
 936                 CDEBUG(D_HA, "reconnected to %s@%s\n",
 937                        obd2cli_tgt(imp->imp_obd),
 938                        imp->imp_connection->c_remote_uuid.uuid);
 939
 940                 rc = ptlrpc_resend(imp);
 941                 if (rc)
 942                         GOTO(out, rc);
 943                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 944                 ptlrpc_activate_import(imp);
 945
 946                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
 947                           &target_start, &target_len);
 948                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
 949                               "using nid %s.\n", imp->imp_obd->obd_name,
 950                               target_len, target_start,
 951                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
 952         }
 953
 954         if (imp->imp_state == LUSTRE_IMP_FULL) {
 955                 cfs_waitq_signal(&imp->imp_recovery_waitq);
 956                 ptlrpc_wake_delayed(imp);
 957         }
 958
 959  out:
 960         RETURN(rc);
 961 }
 962
 963 static int back_to_sleep(void *unused)
 964 {
 965         return 0;
 966 }
 967
 968 int ptlrpc_disconnect_import(struct obd_import *imp)
 969 {
 970         struct ptlrpc_request *req;
 971         int rq_opc, rc = 0;
 972         ENTRY;
 973
 974         switch (imp->imp_connect_op) {
 975         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
 976         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
 977         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
 978         default:
 979                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
 980                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
 981                 RETURN(-EINVAL);
 982         }
 983
 984         if (ptlrpc_import_in_recovery(imp)) {
 985                 struct l_wait_info lwi;
 986                 cfs_duration_t timeout = cfs_time_seconds(obd_timeout);
 987
 988                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
 989                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
 990                 rc = l_wait_event(imp->imp_recovery_waitq,
 991                                   !ptlrpc_import_in_recovery(imp), &lwi);
 992
 993         }
 994
 995         spin_lock(&imp->imp_lock);
 996         if (imp->imp_state != LUSTRE_IMP_FULL)
 997                 GOTO(out, 0);
 998
 999         spin_unlock(&imp->imp_lock);
1000
1001         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, rq_opc, 1, NULL, NULL);
1002         if (req) {
1003                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1004                  * it fails.  We can get through the above with a down server
1005                  * if the client doesn't know the server is gone yet. */
1006                 req->rq_no_resend = 1;
1007                 req->rq_timeout = 5;
1008                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1009                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1010                 ptlrpc_req_set_repsize(req, 1, NULL);
1011                 rc = ptlrpc_queue_wait(req);
1012                 ptlrpc_req_finished(req);
1013         }
1014
1015         spin_lock(&imp->imp_lock);
1016 out:
1017         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1018         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1019         spin_unlock(&imp->imp_lock);
1020
1021         RETURN(rc);
1022 }
1023