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 Lustre, http://www.lustre.org.
   8  *
   9  *   Lustre is free software; you can redistribute it and/or
  10  *   modify it under the terms of version 2 of the GNU General Public
  11  *   License as published by the Free Software Foundation.
  12  *
  13  *   Lustre is distributed in the hope that it will be useful,
  14  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  *   GNU General Public License for more details.
  17  *
  18  *   You should have received a copy of the GNU General Public License
  19  *   along with Lustre; if not, write to the Free Software
  20  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_RPC
  24 #ifdef __KERNEL__
  25 # include <linux/config.h>
  26 # include <linux/module.h>
  27 # include <linux/kmod.h>
  28 #else
  29 # include <liblustre.h>
  30 #endif
  31
  32 #include <linux/obd_support.h>
  33 #include <linux/lustre_ha.h>
  34 #include <linux/lustre_net.h>
  35 #include <linux/lustre_import.h>
  36 #include <linux/lustre_export.h>
  37 #include <linux/obd.h>
  38 #include <linux/obd_class.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, imp->imp_target_uuid.uuid,                          \
  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         unsigned long flags;                            \
  62                                                         \
  63         spin_lock_irqsave(&imp->imp_lock, flags);       \
  64         IMPORT_SET_STATE_NOLOCK(imp, state);            \
  65         spin_unlock_irqrestore(&imp->imp_lock, flags);  \
  66 } while(0)
  67
  68
  69 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  70                                     void * data, int rc);
  71 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  72
  73 /* Only this function is allowed to change the import state when it is
  74  * CLOSED. I would rather refcount the import and free it after
  75  * disconnection like we do with exports. To do that, the client_obd
  76  * will need to save the peer info somewhere other than in the import,
  77  * though. */
  78 int ptlrpc_init_import(struct obd_import *imp)
  79 {
  80         unsigned long flags;
  81
  82         spin_lock_irqsave(&imp->imp_lock, flags);
  83
  84         imp->imp_generation++;
  85         imp->imp_state =  LUSTRE_IMP_NEW;
  86
  87         spin_unlock_irqrestore(&imp->imp_lock, flags);
  88
  89         return 0;
  90 }
  91
  92 /* Returns true if import was FULL, false if import was already not
  93  * connected.
  94  */
  95 int ptlrpc_set_import_discon(struct obd_import *imp)
  96 {
  97         unsigned long flags;
  98         int rc = 0;
  99
 100         spin_lock_irqsave(&imp->imp_lock, flags);
 101
 102         if (imp->imp_state == LUSTRE_IMP_FULL) {
 103                 CERROR("%s: connection lost to %s@%s\n",
 104                        imp->imp_obd->obd_name,
 105                        imp->imp_target_uuid.uuid,
 106                        imp->imp_connection->c_remote_uuid.uuid);
 107                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 108                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 109                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 110                 rc = 1;
 111         } else {
 112                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 113                 CDEBUG(D_HA, "%p %s: import already not connected: %s\n",
 114                        imp,imp->imp_client->cli_name,
 115                        ptlrpc_import_state_name(imp->imp_state));
 116         }
 117
 118         return rc;
 119 }
 120
 121 /*
 122  * This acts as a barrier; all existing requests are rejected, and
 123  * no new requests will be accepted until the import is valid again.
 124  */
 125 void ptlrpc_deactivate_import(struct obd_import *imp)
 126 {
 127         unsigned long flags;
 128         ENTRY;
 129
 130         spin_lock_irqsave(&imp->imp_lock, flags);
 131         CDEBUG(D_HA, "setting import %s INVALID\n",
 132                imp->imp_target_uuid.uuid);
 133         imp->imp_invalid = 1;
 134         imp->imp_generation++;
 135         spin_unlock_irqrestore(&imp->imp_lock, flags);
 136
 137         ptlrpc_abort_inflight(imp);
 138         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 139 }
 140
 141 /*
 142  * This function will invalidate the import, if necessary, then block
 143  * for all the RPC completions, and finally notify the obd to
 144  * invalidate its state (ie cancel locks, clear pending requests,
 145  * etc).
 146  *
 147  * in_rpc: true if this is called while processing an rpc, like
 148  *    CONNECT. It will allow for one RPC to be inflight while
 149  *    waiting for requests to complete. Ugly, yes, but I don't see an
 150  *    cleaner way right now.
 151  */
 152 void ptlrpc_invalidate_import(struct obd_import *imp, int in_rpc)
 153 {
 154         struct l_wait_info lwi;
 155         unsigned long timeout;
 156         int inflight = 0;
 157         int rc;
 158
 159         if (!imp->imp_invalid)
 160                 ptlrpc_deactivate_import(imp);
 161
 162         LASSERT(imp->imp_invalid);
 163
 164         if (in_rpc)
 165                 inflight = 1;
 166
 167         /* wait for all requests to error out and call completion
 168            callbacks */
 169         if (imp->imp_server_timeout)
 170                 timeout = obd_timeout / 2;
 171         else
 172                 timeout = obd_timeout;
 173         timeout = MAX(timeout * HZ, 1);
 174         lwi = LWI_TIMEOUT_INTR(timeout, NULL, NULL, NULL);
 175         rc = l_wait_event(imp->imp_recovery_waitq,
 176                           (atomic_read(&imp->imp_inflight) == inflight),
 177                           &lwi);
 178
 179         if (rc)
 180                 CERROR("%s: rc = %d waiting for callback (%d != %d)\n",
 181                        imp->imp_target_uuid.uuid, rc,
 182                        atomic_read(&imp->imp_inflight), inflight);
 183
 184         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 185 }
 186
 187 void ptlrpc_activate_import(struct obd_import *imp)
 188 {
 189         struct obd_device *obd = imp->imp_obd;
 190         unsigned long flags;
 191
 192         spin_lock_irqsave(&imp->imp_lock, flags);
 193         imp->imp_invalid = 0;
 194         spin_unlock_irqrestore(&imp->imp_lock, flags);
 195
 196         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 197 }
 198
 199 void ptlrpc_fail_import(struct obd_import *imp, int generation)
 200 {
 201         ENTRY;
 202
 203         LASSERT (!imp->imp_dlm_fake);
 204
 205         if (ptlrpc_set_import_discon(imp)) {
 206                 unsigned long flags;
 207
 208                 if (!imp->imp_replayable) {
 209                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 210                                "auto-deactivating\n",
 211                                imp->imp_target_uuid.uuid,
 212                                imp->imp_connection->c_remote_uuid.uuid,
 213                                imp->imp_obd->obd_name);
 214                         ptlrpc_deactivate_import(imp);
 215                 }
 216
 217                 CDEBUG(D_HA, "%s: waking up pinger\n",
 218                        imp->imp_target_uuid.uuid);
 219
 220                 spin_lock_irqsave(&imp->imp_lock, flags);
 221                 imp->imp_force_verify = 1;
 222                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 223
 224                 ptlrpc_pinger_wake_up();
 225         }
 226         EXIT;
 227 }
 228
 229 int ptlrpc_connect_import(struct obd_import *imp, char * new_uuid)
 230 {
 231         struct obd_device *obd = imp->imp_obd;
 232         int initial_connect = 0;
 233         int rc;
 234         __u64 committed_before_reconnect = 0;
 235         struct ptlrpc_request *request;
 236         int size[] = {sizeof(imp->imp_target_uuid),
 237                                  sizeof(obd->obd_uuid),
 238                                  sizeof(imp->imp_dlm_handle),
 239                                  sizeof(unsigned long)};
 240         char *tmp[] = {imp->imp_target_uuid.uuid,
 241                        obd->obd_uuid.uuid,
 242                        (char *)&imp->imp_dlm_handle,
 243                        (char *)&imp->imp_connect_flags}; /* XXX: make this portable! */
 244         struct ptlrpc_connect_async_args *aa;
 245         unsigned long flags;
 246
 247         spin_lock_irqsave(&imp->imp_lock, flags);
 248         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 249                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 250                 CERROR("can't connect to a closed import\n");
 251                 RETURN(-EINVAL);
 252         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 253                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 254                 CERROR("already connected\n");
 255                 RETURN(0);
 256         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 257                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 258                 CERROR("already connecting\n");
 259                 RETURN(-EALREADY);
 260         }
 261
 262         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 263
 264         imp->imp_resend_replay = 0;
 265
 266         if (imp->imp_remote_handle.cookie == 0) {
 267                 initial_connect = 1;
 268         } else {
 269                 committed_before_reconnect = imp->imp_peer_committed_transno;;
 270                 imp->imp_conn_cnt++;
 271         }
 272
 273
 274         spin_unlock_irqrestore(&imp->imp_lock, flags);
 275
 276         if (new_uuid) {
 277                 struct ptlrpc_connection *conn;
 278                 struct obd_uuid uuid;
 279                 struct obd_export *dlmexp;
 280
 281                 obd_str2uuid(&uuid, new_uuid);
 282
 283                 conn = ptlrpc_uuid_to_connection(&uuid);
 284                 if (!conn)
 285                         GOTO(out, rc = -ENOENT);
 286
 287                 CDEBUG(D_HA, "switching import %s/%s from %s to %s\n",
 288                        imp->imp_target_uuid.uuid, imp->imp_obd->obd_name,
 289                        imp->imp_connection->c_remote_uuid.uuid,
 290                        conn->c_remote_uuid.uuid);
 291
 292                 /* Switch the import's connection and the DLM export's
 293                  * connection (which are almost certainly the same, but we
 294                  * keep distinct refs just to make things clearer. I think. */
 295                 if (imp->imp_connection)
 296                         ptlrpc_put_connection(imp->imp_connection);
 297                 /* We hand off the ref from ptlrpc_get_connection. */
 298                 imp->imp_connection = conn;
 299
 300                 dlmexp = class_conn2export(&imp->imp_dlm_handle);
 301
 302                 LASSERT(dlmexp != NULL);
 303
 304                 if (dlmexp->exp_connection)
 305                         ptlrpc_put_connection(dlmexp->exp_connection);
 306                 dlmexp->exp_connection = ptlrpc_connection_addref(conn);
 307                 class_export_put(dlmexp);
 308
 309         }
 310
 311         request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION,
 312                                   imp->imp_connect_op, 4, size, tmp);
 313         if (!request)
 314                 GOTO(out, rc = -ENOMEM);
 315
 316 #ifndef __KERNEL__
 317         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 318 #endif
 319
 320         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 321         request->rq_replen = lustre_msg_size(0, NULL);
 322         request->rq_interpret_reply = ptlrpc_connect_interpret;
 323
 324         LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
 325         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 326         memset(aa, 0, sizeof *aa);
 327
 328         aa->pcaa_peer_committed = committed_before_reconnect;
 329         aa->pcaa_initial_connect = initial_connect;
 330
 331         if (aa->pcaa_initial_connect) {
 332                 lustre_msg_add_op_flags(request->rq_reqmsg,
 333                                         MSG_CONNECT_INITIAL);
 334                 imp->imp_replayable = 1;
 335         }
 336
 337         ptlrpcd_add_req(request);
 338         rc = 0;
 339         imp->imp_connect_start = jiffies;
 340 out:
 341         if (rc != 0) {
 342                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 343         }
 344
 345         RETURN(rc);
 346 }
 347
 348 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 349                                     void * data, int rc)
 350 {
 351         struct ptlrpc_connect_async_args *aa = data;
 352         struct obd_import *imp = request->rq_import;
 353         struct lustre_handle old_hdl;
 354         unsigned long flags;
 355         int msg_flags;
 356         ENTRY;
 357
 358         spin_lock_irqsave(&imp->imp_lock, flags);
 359         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 360                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 361                 RETURN(0);
 362         }
 363         spin_unlock_irqrestore(&imp->imp_lock, flags);
 364
 365         if (rc)
 366                 GOTO(out, rc);
 367
 368         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 369
 370         if (aa->pcaa_initial_connect) {
 371                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 372                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 373                                imp->imp_target_uuid.uuid);
 374                         imp->imp_pingable = imp->imp_replayable = 1;
 375                 } else {
 376                         imp->imp_replayable = 0;
 377                 }
 378                 LASSERTF(imp->imp_conn_cnt < request->rq_repmsg->conn_cnt,
 379                          "imp conn_cnt %d req conn_cnt %d",
 380                          imp->imp_conn_cnt, request->rq_repmsg->conn_cnt);
 381                 imp->imp_conn_cnt = request->rq_repmsg->conn_cnt;
 382                 imp->imp_remote_handle = request->rq_repmsg->handle;
 383                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 384                 GOTO(finish, rc = 0);
 385         }
 386
 387         /* Determine what recovery state to move the import to. */
 388         if (MSG_CONNECT_RECONNECT & msg_flags) {
 389                 memset(&old_hdl, 0, sizeof(old_hdl));
 390                 if (!memcmp(&old_hdl, &request->rq_repmsg->handle,
 391                             sizeof (old_hdl))) {
 392                         CERROR("%s@%s didn't like our handle "LPX64
 393                                ", failed\n", imp->imp_target_uuid.uuid,
 394                                imp->imp_connection->c_remote_uuid.uuid,
 395                                imp->imp_dlm_handle.cookie);
 396                         GOTO(out, rc = -ENOTCONN);
 397                 }
 398
 399                 if (memcmp(&imp->imp_remote_handle, &request->rq_repmsg->handle,
 400                            sizeof(imp->imp_remote_handle))) {
 401                         CERROR("%s@%s changed handle from "LPX64" to "LPX64
 402                                "; copying, but this may foreshadow disaster\n",
 403                                imp->imp_target_uuid.uuid,
 404                                imp->imp_connection->c_remote_uuid.uuid,
 405                                imp->imp_remote_handle.cookie,
 406                                request->rq_repmsg->handle.cookie);
 407                         imp->imp_remote_handle = request->rq_repmsg->handle;
 408                 } else {
 409                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 410                                imp->imp_target_uuid.uuid,
 411                                imp->imp_connection->c_remote_uuid.uuid);
 412                 }
 413
 414                 if (imp->imp_invalid) {
 415                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 416                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 417                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 418                                imp->imp_obd->obd_name,
 419                                imp->imp_target_uuid.uuid);
 420                         imp->imp_resend_replay = 1;
 421                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 422                 } else {
 423                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 424                 }
 425         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 426                 LASSERT(imp->imp_replayable);
 427                 imp->imp_remote_handle = request->rq_repmsg->handle;
 428                 imp->imp_last_replay_transno = 0;
 429                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 430         } else {
 431                 CDEBUG(D_HA, "oops! we get evicted from %s\n", imp->imp_target_uuid.uuid);
 432                 imp->imp_remote_handle = request->rq_repmsg->handle;
 433                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 434         }
 435
 436         /* Sanity checks for a reconnected import. */
 437         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 438                 CERROR("imp_replayable flag does not match server "
 439                        "after reconnect. We should LBUG right here.\n");
 440         }
 441
 442         if (request->rq_repmsg->last_committed < aa->pcaa_peer_committed) {
 443                 CERROR("%s went back in time (transno "LPD64
 444                        " was previously committed, server now claims "LPD64
 445                        ")! is shared storage not coherent?\n",
 446                        imp->imp_target_uuid.uuid,
 447                        aa->pcaa_peer_committed,
 448                        request->rq_repmsg->last_committed);
 449         }
 450
 451 finish:
 452         rc = ptlrpc_import_recovery_state_machine(imp);
 453         if (rc != 0) {
 454                 if (rc == -ENOTCONN) {
 455                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 456                                "invalidating and reconnecting\n",
 457                                imp->imp_target_uuid.uuid,
 458                                imp->imp_connection->c_remote_uuid.uuid);
 459                         ptlrpc_connect_import(imp, NULL);
 460                         RETURN(0);
 461                 }
 462         }
 463  out:
 464         if (rc != 0) {
 465                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 466                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov) {
 467                         ptlrpc_deactivate_import(imp);
 468                 }
 469                 if (rc == -ETIMEDOUT && (jiffies - imp->imp_connect_start) > HZ) {
 470                         CDEBUG(D_ERROR, "recovery of %s on %s failed (timeout)\n",
 471                                imp->imp_target_uuid.uuid,
 472                                (char *)imp->imp_connection->c_remote_uuid.uuid);
 473                         ptlrpc_connect_import(imp, NULL);
 474                         RETURN(0);
 475                 }
 476                 CDEBUG(D_ERROR, "recovery of %s on %s failed (%d)\n",
 477                        imp->imp_target_uuid.uuid,
 478                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 479         }
 480
 481         wake_up(&imp->imp_recovery_waitq);
 482         RETURN(rc);
 483 }
 484
 485 static int completed_replay_interpret(struct ptlrpc_request *req,
 486                                     void * data, int rc)
 487 {
 488         atomic_dec(&req->rq_import->imp_replay_inflight);
 489         if (req->rq_status == 0) {
 490                 ptlrpc_import_recovery_state_machine(req->rq_import);
 491         } else {
 492                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 493                        "reconnecting\n",
 494                        req->rq_import->imp_obd->obd_name, req->rq_status);
 495                 ptlrpc_connect_import(req->rq_import, NULL);
 496         }
 497
 498         RETURN(0);
 499 }
 500
 501 static int signal_completed_replay(struct obd_import *imp)
 502  {
 503         struct ptlrpc_request *req;
 504         ENTRY;
 505
 506         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 507         atomic_inc(&imp->imp_replay_inflight);
 508
 509         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 0, NULL, NULL);
 510         if (!req)
 511                 RETURN(-ENOMEM);
 512
 513         req->rq_replen = lustre_msg_size(0, NULL);
 514         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 515         req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
 516         req->rq_timeout *= 3;
 517         req->rq_interpret_reply = completed_replay_interpret;
 518
 519         ptlrpcd_add_req(req);
 520         RETURN(0);
 521 }
 522
 523 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 524 {
 525         int rc = 0;
 526         int inflight;
 527
 528         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 529                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 530                        imp->imp_target_uuid.uuid,
 531                        imp->imp_connection->c_remote_uuid.uuid);
 532
 533                 ptlrpc_invalidate_import(imp, 1);
 534
 535                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 536         }
 537
 538         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 539                 CDEBUG(D_HA, "replay requested by %s\n",
 540                        imp->imp_target_uuid.uuid);
 541                 rc = ptlrpc_replay_next(imp, &inflight);
 542                 if (inflight == 0 &&
 543                     atomic_read(&imp->imp_replay_inflight) == 0) {
 544                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 545                         rc = ldlm_replay_locks(imp);
 546                         if (rc)
 547                                 GOTO(out, rc);
 548                 }
 549                 rc = 0;
 550         }
 551
 552         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 553                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 554                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
 555                         rc = signal_completed_replay(imp);
 556                         if (rc)
 557                                 GOTO(out, rc);
 558                 }
 559
 560         }
 561
 562         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
 563                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 564                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 565                 }
 566         }
 567
 568         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
 569                 CDEBUG(D_HA, "reconnected to %s@%s\n",
 570                        imp->imp_target_uuid.uuid,
 571                        imp->imp_connection->c_remote_uuid.uuid);
 572
 573                 rc = ptlrpc_resend(imp);
 574                 if (rc)
 575                         GOTO(out, rc);
 576                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 577                 ptlrpc_activate_import(imp);
 578                 CERROR("%s: connection restored to %s@%s\n",
 579                        imp->imp_obd->obd_name,
 580                        imp->imp_target_uuid.uuid,
 581                        imp->imp_connection->c_remote_uuid.uuid);
 582         }
 583
 584         if (imp->imp_state == LUSTRE_IMP_FULL) {
 585                 wake_up(&imp->imp_recovery_waitq);
 586                 ptlrpc_wake_delayed(imp);
 587         }
 588
 589  out:
 590         RETURN(rc);
 591 }
 592
 593 static int back_to_sleep(void *unused)
 594 {
 595         return 0;
 596 }
 597
 598 int ptlrpc_disconnect_import(struct obd_import *imp)
 599 {
 600         struct ptlrpc_request *request;
 601         int rq_opc;
 602         int rc = 0;
 603         unsigned long flags;
 604         ENTRY;
 605
 606         switch (imp->imp_connect_op) {
 607         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
 608         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
 609         case MGMT_CONNECT:rq_opc = MGMT_DISCONNECT;break;
 610         default:
 611                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
 612                        imp->imp_target_uuid.uuid, imp->imp_connect_op);
 613                 RETURN(-EINVAL);
 614         }
 615
 616
 617         if (ptlrpc_import_in_recovery(imp)) {
 618                 struct l_wait_info lwi;
 619                 unsigned long timeout;
 620                 if (imp->imp_server_timeout)
 621                         timeout = obd_timeout / 2;
 622                 else
 623                         timeout = obd_timeout;
 624                 timeout = MAX(timeout * HZ, 1);
 625                 lwi = LWI_TIMEOUT_INTR(obd_timeout, back_to_sleep, NULL, NULL);
 626                 rc = l_wait_event(imp->imp_recovery_waitq,
 627                                   !ptlrpc_import_in_recovery(imp), &lwi);
 628
 629         }
 630
 631         spin_lock_irqsave(&imp->imp_lock, flags);
 632         if (imp->imp_state != LUSTRE_IMP_FULL) {
 633                 GOTO(out, 0);
 634         }
 635         spin_unlock_irqrestore(&imp->imp_lock, flags);
 636
 637         request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, rq_opc,
 638                                   0, NULL, NULL);
 639         if (request) {
 640                 /* For non-replayable connections, don't attempt
 641                    reconnect if this fails */
 642                 if (!imp->imp_replayable) {
 643                         request->rq_no_resend = 1;
 644                         IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
 645                         request->rq_send_state =  LUSTRE_IMP_CONNECTING;
 646                 }
 647                 request->rq_replen = lustre_msg_size(0, NULL);
 648                 rc = ptlrpc_queue_wait(request);
 649                 ptlrpc_req_finished(request);
 650         }
 651
 652         spin_lock_irqsave(&imp->imp_lock, flags);
 653 out:
 654         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
 655         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
 656         imp->imp_conn_cnt = 0;
 657         spin_unlock_irqrestore(&imp->imp_lock, flags);
 658
 659         RETURN(rc);
 660 }
 661