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 #ifndef __KERNEL__
  25 # include <liblustre.h>
  26 #endif
  27
  28 #include <linux/obd_support.h>
  29 #include <linux/lustre_ha.h>
  30 #include <linux/lustre_net.h>
  31 #include <linux/lustre_import.h>
  32 #include <linux/lustre_export.h>
  33 #include <linux/obd.h>
  34 #include <linux/obd_class.h>
  35 #include <linux/lustre_sec.h>
  36
  37 #include "ptlrpc_internal.h"
  38
  39 struct ptlrpc_connect_async_args {
  40          __u64 pcaa_peer_committed;
  41         int pcaa_initial_connect;
  42 };
  43
  44 /* A CLOSED import should remain so. */
  45 #define IMPORT_SET_STATE_NOLOCK(imp, state)                                    \
  46 do {                                                                           \
  47         if (imp->imp_state != LUSTRE_IMP_CLOSED) {                             \
  48                CDEBUG(D_HA, "%p %s: changing import state from %s to %s\n",    \
  49                       imp, imp->imp_target_uuid.uuid,                          \
  50                       ptlrpc_import_state_name(imp->imp_state),                \
  51                       ptlrpc_import_state_name(state));                        \
  52                imp->imp_state = state;                                         \
  53         }                                                                      \
  54 } while(0)
  55
  56 #define IMPORT_SET_STATE(imp, state)                    \
  57 do {                                                    \
  58         unsigned long flags;                            \
  59                                                         \
  60         spin_lock_irqsave(&imp->imp_lock, flags);       \
  61         IMPORT_SET_STATE_NOLOCK(imp, state);            \
  62         spin_unlock_irqrestore(&imp->imp_lock, flags);  \
  63 } while(0)
  64
  65
  66 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
  67                                     void * data, int rc);
  68 int ptlrpc_import_recovery_state_machine(struct obd_import *imp);
  69
  70 /* Only this function is allowed to change the import state when it is
  71  * CLOSED. I would rather refcount the import and free it after
  72  * disconnection like we do with exports. To do that, the client_obd
  73  * will need to save the peer info somewhere other than in the import,
  74  * though. */
  75 int ptlrpc_init_import(struct obd_import *imp)
  76 {
  77         unsigned long flags;
  78
  79         spin_lock_irqsave(&imp->imp_lock, flags);
  80
  81         imp->imp_generation++;
  82         imp->imp_state =  LUSTRE_IMP_NEW;
  83
  84         spin_unlock_irqrestore(&imp->imp_lock, flags);
  85
  86         return 0;
  87 }
  88
  89 /* Returns true if import was FULL, false if import was already not
  90  * connected.
  91  */
  92 int ptlrpc_set_import_discon(struct obd_import *imp)
  93 {
  94         unsigned long flags;
  95         int rc = 0;
  96
  97         spin_lock_irqsave(&imp->imp_lock, flags);
  98
  99         if (imp->imp_state == LUSTRE_IMP_FULL) {
 100                 CWARN("%s: connection lost to %s@%s\n",
 101                       imp->imp_obd->obd_name,
 102                       imp->imp_target_uuid.uuid,
 103                       imp->imp_connection->c_remote_uuid.uuid);
 104                 ptlrpc_deactivate_timeouts();
 105                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 106                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 107                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 108                 rc = 1;
 109         } else {
 110                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 111                 CDEBUG(D_HA, "%p %s: import already not connected: %s\n",
 112                        imp,imp->imp_client->cli_name,
 113                        ptlrpc_import_state_name(imp->imp_state));
 114         }
 115
 116         return rc;
 117 }
 118
 119 /*
 120  * This acts as a barrier; all existing requests are rejected, and
 121  * no new requests will be accepted until the import is valid again.
 122  */
 123 void ptlrpc_deactivate_import(struct obd_import *imp)
 124 {
 125         unsigned long flags;
 126         ENTRY;
 127
 128         spin_lock_irqsave(&imp->imp_lock, flags);
 129         CDEBUG(D_HA, "setting import %s INVALID\n",
 130                imp->imp_target_uuid.uuid);
 131         imp->imp_invalid = 1;
 132         imp->imp_generation++;
 133         spin_unlock_irqrestore(&imp->imp_lock, flags);
 134
 135         ptlrpc_abort_inflight(imp);
 136         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 137 }
 138
 139 /*
 140  * This function will invalidate the import, if necessary, then block
 141  * for all the RPC completions, and finally notify the obd to
 142  * invalidate its state (ie cancel locks, clear pending requests,
 143  * etc).
 144  *
 145  * in_rpc: true if this is called while processing an rpc, like
 146  *    CONNECT. It will allow for one RPC to be inflight while
 147  *    waiting for requests to complete. Ugly, yes, but I don't see an
 148  *    cleaner way right now.
 149  */
 150 void ptlrpc_invalidate_import(struct obd_import *imp, int in_rpc)
 151 {
 152         struct l_wait_info lwi;
 153         unsigned long timeout;
 154         int inflight = 0;
 155         int rc;
 156
 157         if (!imp->imp_invalid)
 158                 ptlrpc_deactivate_import(imp);
 159
 160         LASSERT(imp->imp_invalid);
 161
 162         if (in_rpc)
 163                 inflight = 1;
 164
 165         /* wait for all requests to error out and call completion
 166            callbacks */
 167         if (imp->imp_server_timeout)
 168                 timeout = obd_timeout / 2;
 169         else
 170                 timeout = obd_timeout;
 171         timeout = MAX(timeout * HZ, 1);
 172         lwi = LWI_TIMEOUT_INTR(timeout, NULL, NULL, NULL);
 173         rc = l_wait_event(imp->imp_recovery_waitq,
 174                           (atomic_read(&imp->imp_inflight) == inflight),
 175                           &lwi);
 176
 177         if (rc)
 178                 CERROR("%s: rc = %d waiting for callback (%d != %d)\n",
 179                        imp->imp_target_uuid.uuid, rc,
 180                        atomic_read(&imp->imp_inflight), !!in_rpc);
 181
 182         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 183 }
 184
 185 void ptlrpc_activate_import(struct obd_import *imp)
 186 {
 187         struct obd_device *obd = imp->imp_obd;
 188         unsigned long flags;
 189
 190         spin_lock_irqsave(&imp->imp_lock, flags);
 191         imp->imp_invalid = 0;
 192         spin_unlock_irqrestore(&imp->imp_lock, flags);
 193
 194         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 195         ptlrpc_activate_timeouts();
 196 }
 197
 198 void ptlrpc_fail_import(struct obd_import *imp, int generation)
 199 {
 200         ENTRY;
 201
 202         LASSERT (!imp->imp_dlm_fake);
 203
 204         if (ptlrpc_set_import_discon(imp)) {
 205                 unsigned long flags;
 206
 207                 if (!imp->imp_replayable) {
 208                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 209                                "auto-deactivating\n",
 210                                imp->imp_target_uuid.uuid,
 211                                imp->imp_connection->c_remote_uuid.uuid,
 212                                imp->imp_obd->obd_name);
 213                         ptlrpc_deactivate_import(imp);
 214                 }
 215
 216                 CDEBUG(D_HA, "%s: waking up pinger\n",
 217                        imp->imp_target_uuid.uuid);
 218
 219                 spin_lock_irqsave(&imp->imp_lock, flags);
 220                 imp->imp_force_verify = 1;
 221                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 222
 223                 ptlrpc_pinger_wake_up();
 224         }
 225         EXIT;
 226 }
 227
 228 #define ATTEMPT_TOO_SOON(last)  \
 229         ((last) && ((long)(jiffies - (last)) <= (long)(obd_timeout * 2 * HZ)))
 230
 231 static int import_select_connection(struct obd_import *imp)
 232 {
 233         struct obd_import_conn *imp_conn, *tmp;
 234         struct obd_export *dlmexp;
 235         int found = 0;
 236         ENTRY;
 237
 238         spin_lock(&imp->imp_lock);
 239
 240         if (list_empty(&imp->imp_conn_list)) {
 241                 CERROR("no available connections on imp %p@%s\n",
 242                         imp, imp->imp_obd->obd_name);
 243                 spin_unlock(&imp->imp_lock);
 244                 RETURN(-EINVAL);
 245         }
 246
 247         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 248                 if (!ATTEMPT_TOO_SOON(imp_conn->oic_last_attempt)) {
 249                         found = 1;
 250                         break;
 251                 }
 252         }
 253
 254         /* if not found, simply choose the current one */
 255         if (!found) {
 256                 CWARN("obd %s imp 0x%p: all connections have been "
 257                       "tried recently\n", imp->imp_obd->obd_name, imp);
 258                 LASSERT(imp->imp_conn_current);
 259                 imp_conn = imp->imp_conn_current;
 260         }
 261         LASSERT(imp_conn->oic_conn);
 262
 263         imp_conn->oic_last_attempt = jiffies;
 264
 265         /* move the items ahead of the selected one to list tail */
 266         while (1) {
 267                 tmp= list_entry(imp->imp_conn_list.next,
 268                                 struct obd_import_conn, oic_item);
 269                 if (tmp == imp_conn)
 270                         break;
 271                 list_del(&tmp->oic_item);
 272                 list_add_tail(&tmp->oic_item, &imp->imp_conn_list);
 273         }
 274
 275         /* switch connection if we chose a new one */
 276         if (imp->imp_connection != imp_conn->oic_conn) {
 277                 if (imp->imp_connection) {
 278                         ptlrpcs_sec_invalidate_cache(imp->imp_sec);
 279                         ptlrpc_put_connection(imp->imp_connection);
 280                 }
 281                 imp->imp_connection =
 282                         ptlrpc_connection_addref(imp_conn->oic_conn);
 283         }
 284
 285         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 286         LASSERT(dlmexp != NULL);
 287         if (dlmexp->exp_connection)
 288                 ptlrpc_put_connection(imp->imp_connection);
 289         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 290         class_export_put(dlmexp);
 291
 292         imp->imp_conn_current = imp_conn;
 293         CWARN("obd %s imp 0x%p: select conn %s\n",
 294                imp->imp_obd->obd_name, imp,
 295                imp_conn->oic_uuid.uuid);
 296         spin_unlock(&imp->imp_lock);
 297
 298         RETURN(0);
 299 }
 300
 301
 302
 303 int ptlrpc_connect_import(struct obd_import *imp, char * new_uuid)
 304 {
 305         struct obd_device *obd = imp->imp_obd;
 306         int initial_connect = 0;
 307         int rc;
 308         __u64 committed_before_reconnect = 0;
 309         struct ptlrpc_request *request;
 310         int size[] = {0,
 311                       sizeof(imp->imp_target_uuid),
 312                       sizeof(obd->obd_uuid),
 313                       sizeof(imp->imp_dlm_handle),
 314                       sizeof(imp->imp_connect_flags),
 315                       sizeof(imp->imp_connect_data)};
 316         char *tmp[] = {NULL,
 317                        imp->imp_target_uuid.uuid,
 318                        obd->obd_uuid.uuid,
 319                        (char *)&imp->imp_dlm_handle,
 320                        (char *)&imp->imp_connect_flags, /* XXX: make this portable! */
 321                        (char*) &imp->imp_connect_data};
 322         int repsize = sizeof(struct obd_connect_data);
 323
 324         struct ptlrpc_connect_async_args *aa;
 325         unsigned long flags;
 326
 327         spin_lock_irqsave(&imp->imp_lock, flags);
 328         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 329                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 330                 CERROR("can't connect to a closed import\n");
 331                 RETURN(-EINVAL);
 332         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 333                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 334                 CERROR("already connected\n");
 335                 RETURN(0);
 336         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 337                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 338                 CERROR("already connecting\n");
 339                 RETURN(-EALREADY);
 340         }
 341
 342         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 343
 344         imp->imp_resend_replay = 0;
 345
 346         if (imp->imp_remote_handle.cookie == 0) {
 347                 initial_connect = 1;
 348         } else {
 349                 committed_before_reconnect = imp->imp_peer_committed_transno;;
 350                 imp->imp_conn_cnt++;
 351         }
 352
 353
 354         spin_unlock_irqrestore(&imp->imp_lock, flags);
 355
 356         if (new_uuid) {
 357                 struct obd_uuid uuid;
 358
 359                 obd_str2uuid(&uuid, new_uuid);
 360
 361                 rc = import_set_conn_priority(imp, &uuid);
 362                 if (rc)
 363                         GOTO(out, rc);
 364         }
 365         rc = import_select_connection(imp);
 366         if (rc)
 367                 GOTO(out, rc);
 368
 369         LASSERT(imp->imp_sec);
 370
 371         size[0] = lustre_secdesc_size();
 372         request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION,
 373                                   imp->imp_connect_op, 6, size, tmp);
 374         if (!request)
 375                 GOTO(out, rc = -ENOMEM);
 376
 377         lustre_pack_secdesc(request, size[0]);
 378
 379 #ifndef __KERNEL__
 380         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 381 #endif
 382         if (obd->u.cli.cl_async) {
 383                 lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_ASYNC);
 384         }
 385
 386         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 387         request->rq_replen = lustre_msg_size(1, &repsize);
 388         request->rq_interpret_reply = ptlrpc_connect_interpret;
 389
 390         LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
 391         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 392         memset(aa, 0, sizeof *aa);
 393
 394         aa->pcaa_peer_committed = committed_before_reconnect;
 395         aa->pcaa_initial_connect = initial_connect;
 396
 397         if (aa->pcaa_initial_connect) {
 398                 lustre_msg_add_op_flags(request->rq_reqmsg,
 399                                         MSG_CONNECT_INITIAL);
 400                 imp->imp_replayable = 1;
 401         }
 402
 403         imp->imp_reqs_replayed = imp->imp_locks_replayed = 0;
 404
 405         ptlrpcd_add_req(request);
 406         rc = 0;
 407         imp->imp_connect_start = jiffies;
 408 out:
 409         if (rc != 0) {
 410                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 411         }
 412
 413         RETURN(rc);
 414 }
 415
 416 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 417                                     void *data, int rc)
 418 {
 419         struct ptlrpc_connect_async_args *aa = data;
 420         struct obd_import *imp = request->rq_import;
 421         struct lustre_handle old_hdl;
 422         unsigned long flags;
 423         int msg_flags;
 424         ENTRY;
 425
 426         spin_lock_irqsave(&imp->imp_lock, flags);
 427         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 428                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 429                 RETURN(0);
 430         }
 431         spin_unlock_irqrestore(&imp->imp_lock, flags);
 432
 433         if (rc)
 434                 GOTO(out, rc);
 435         LASSERT(imp->imp_conn_current);
 436         imp->imp_conn_current->oic_last_attempt = 0;
 437 /*
 438         remote_flag = lustre_msg_buf(request->rq_repmsg, 0, sizeof(int));
 439         LASSERT(remote_flag != NULL);
 440         imp->imp_obd->u.cli.cl_remote = *remote_flag;
 441 */
 442         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 443
 444         if (aa->pcaa_initial_connect) {
 445                 struct obd_connect_data *conn_data;
 446
 447                 conn_data = lustre_swab_repbuf(request, 0, sizeof(*conn_data),
 448                                                lustre_swab_connect);
 449                 LASSERT(conn_data);
 450                 imp->imp_connect_data.ocd_connect_flags =
 451                                         conn_data->ocd_connect_flags;
 452
 453                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 454                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 455                                imp->imp_target_uuid.uuid);
 456                         imp->imp_pingable = imp->imp_replayable = 1;
 457                 } else {
 458                         imp->imp_replayable = 0;
 459                 }
 460                 LASSERTF(imp->imp_conn_cnt < request->rq_repmsg->conn_cnt,
 461                          "imp conn_cnt %d req conn_cnt %d",
 462                          imp->imp_conn_cnt, request->rq_repmsg->conn_cnt);
 463                 imp->imp_conn_cnt = request->rq_repmsg->conn_cnt;
 464                 imp->imp_remote_handle = request->rq_repmsg->handle;
 465                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 466                 ptlrpc_pinger_sending_on_import(imp);
 467                 GOTO(finish, rc = 0);
 468         }
 469
 470         /* Determine what recovery state to move the import to. */
 471         if (MSG_CONNECT_RECONNECT & msg_flags) {
 472                 memset(&old_hdl, 0, sizeof(old_hdl));
 473                 if (!memcmp(&old_hdl, &request->rq_repmsg->handle,
 474                             sizeof (old_hdl))) {
 475                         CERROR("%s@%s didn't like our handle "LPX64
 476                                ", failed\n", imp->imp_target_uuid.uuid,
 477                                imp->imp_connection->c_remote_uuid.uuid,
 478                                imp->imp_dlm_handle.cookie);
 479                         GOTO(out, rc = -ENOTCONN);
 480                 }
 481
 482                 if (memcmp(&imp->imp_remote_handle, &request->rq_repmsg->handle,
 483                            sizeof(imp->imp_remote_handle))) {
 484                         CERROR("%s@%s changed handle from "LPX64" to "LPX64
 485                                "; copying, but this may foreshadow disaster\n",
 486                                imp->imp_target_uuid.uuid,
 487                                imp->imp_connection->c_remote_uuid.uuid,
 488                                imp->imp_remote_handle.cookie,
 489                                request->rq_repmsg->handle.cookie);
 490                         imp->imp_remote_handle = request->rq_repmsg->handle;
 491                 } else {
 492                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 493                                imp->imp_target_uuid.uuid,
 494                                imp->imp_connection->c_remote_uuid.uuid);
 495                 }
 496
 497                 if (imp->imp_invalid) {
 498                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 499                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 500                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 501                                imp->imp_obd->obd_name,
 502                                imp->imp_target_uuid.uuid);
 503                         imp->imp_resend_replay = 1;
 504                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 505                 } else {
 506                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 507                 }
 508         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 509                 LASSERT(imp->imp_replayable);
 510                 imp->imp_remote_handle = request->rq_repmsg->handle;
 511                 imp->imp_last_replay_transno = 0;
 512                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 513         } else {
 514                 CDEBUG(D_HA, "oops! we get evicted from %s\n", imp->imp_target_uuid.uuid);
 515                 imp->imp_remote_handle = request->rq_repmsg->handle;
 516                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 517         }
 518
 519         /* Sanity checks for a reconnected import. */
 520         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 521                 CERROR("imp_replayable flag does not match server "
 522                        "after reconnect. We should LBUG right here.\n");
 523         }
 524
 525         if (request->rq_repmsg->last_committed < aa->pcaa_peer_committed) {
 526                 CERROR("%s went back in time (transno "LPD64
 527                        " was previously committed, server now claims "LPD64
 528                        ")! is shared storage not coherent?\n",
 529                        imp->imp_target_uuid.uuid,
 530                        aa->pcaa_peer_committed,
 531                        request->rq_repmsg->last_committed);
 532         }
 533
 534 finish:
 535         rc = ptlrpc_import_recovery_state_machine(imp);
 536         if (rc != 0) {
 537                 if (rc == -ENOTCONN) {
 538                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 539                                "invalidating and reconnecting\n",
 540                                imp->imp_target_uuid.uuid,
 541                                imp->imp_connection->c_remote_uuid.uuid);
 542                         ptlrpc_connect_import(imp, NULL);
 543                         RETURN(0);
 544                 }
 545         }
 546  out:
 547         if (rc != 0) {
 548                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 549                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov)
 550                         ptlrpc_deactivate_import(imp);
 551                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 552                        imp->imp_target_uuid.uuid,
 553                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 554         }
 555
 556         wake_up(&imp->imp_recovery_waitq);
 557         RETURN(rc);
 558 }
 559
 560 static int completed_replay_interpret(struct ptlrpc_request *req,
 561                                       void *data, int rc)
 562 {
 563         atomic_dec(&req->rq_import->imp_replay_inflight);
 564         if (req->rq_status == 0) {
 565                 ptlrpc_import_recovery_state_machine(req->rq_import);
 566         } else {
 567                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 568                        "reconnecting\n",
 569                        req->rq_import->imp_obd->obd_name, req->rq_status);
 570                 ptlrpc_connect_import(req->rq_import, NULL);
 571         }
 572
 573         RETURN(0);
 574 }
 575
 576 static int signal_completed_replay(struct obd_import *imp)
 577  {
 578         struct ptlrpc_request *req;
 579         ENTRY;
 580
 581         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 582         atomic_inc(&imp->imp_replay_inflight);
 583
 584         req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, OBD_PING, 0, NULL, NULL);
 585         if (!req) {
 586                 atomic_dec(&imp->imp_replay_inflight);
 587                 RETURN(-ENOMEM);
 588         }
 589
 590         req->rq_replen = lustre_msg_size(0, NULL);
 591         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 592         req->rq_reqmsg->flags |= MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE;
 593         req->rq_timeout *= 3;
 594         req->rq_interpret_reply = completed_replay_interpret;
 595
 596         ptlrpcd_add_req(req);
 597         RETURN(0);
 598 }
 599
 600 #ifdef __KERNEL__
 601 static int ptlrpc_invalidate_import_thread(void *data)
 602 {
 603         struct obd_import *imp = data;
 604         unsigned long flags;
 605
 606         ENTRY;
 607
 608         lock_kernel();
 609         ptlrpc_daemonize();
 610
 611         SIGNAL_MASK_LOCK(current, flags);
 612         sigfillset(&current->blocked);
 613         RECALC_SIGPENDING;
 614         SIGNAL_MASK_UNLOCK(current, flags);
 615         THREAD_NAME(current->comm, sizeof(current->comm), "ll_imp_inval");
 616         unlock_kernel();
 617
 618         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 619                imp->imp_obd->obd_name, imp->imp_target_uuid.uuid,
 620                imp->imp_connection->c_remote_uuid.uuid);
 621
 622         ptlrpc_invalidate_import(imp, 0);
 623         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 624
 625         ptlrpc_import_recovery_state_machine(imp);
 626
 627         RETURN(0);
 628 }
 629 #endif
 630
 631 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 632 {
 633         int rc = 0;
 634         int inflight;
 635
 636         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 637                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 638                        imp->imp_target_uuid.uuid,
 639                        imp->imp_connection->c_remote_uuid.uuid);
 640
 641 #ifdef __KERNEL__
 642                 rc = kernel_thread(ptlrpc_invalidate_import_thread, imp,
 643                                    CLONE_VM | CLONE_FILES);
 644                 if (rc < 0)
 645                         CERROR("error starting invalidate thread: %d\n", rc);
 646                 RETURN(rc < 0 ? rc : 0);
 647 #else
 648                 ptlrpc_invalidate_import(imp, 1);
 649
 650                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 651 #endif
 652         }
 653
 654         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 655                 CDEBUG(D_HA, "replay requested by %s\n",
 656                        imp->imp_target_uuid.uuid);
 657                 rc = ptlrpc_replay_next(imp, &inflight);
 658                 if (inflight == 0 &&
 659                     atomic_read(&imp->imp_replay_inflight) == 0) {
 660                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 661                         rc = ldlm_replay_locks(imp);
 662                         if (rc)
 663                                 GOTO(out, rc);
 664                 }
 665                 rc = 0;
 666         }
 667
 668         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 669                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 670                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
 671                         rc = signal_completed_replay(imp);
 672                         if (rc)
 673                                 GOTO(out, rc);
 674                 }
 675
 676         }
 677
 678         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
 679                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 680                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 681                 }
 682         }
 683
 684         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
 685                 CDEBUG(D_HA, "reconnected to %s@%s\n",
 686                        imp->imp_target_uuid.uuid,
 687                        imp->imp_connection->c_remote_uuid.uuid);
 688
 689                 rc = ptlrpc_resend(imp);
 690                 if (rc)
 691                         GOTO(out, rc);
 692                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 693                 ptlrpc_activate_import(imp);
 694                 CWARN("%s: connection restored to %s@%s, "
 695                        "%d/%d req/lock replayed\n",
 696                       imp->imp_obd->obd_name,
 697                       imp->imp_target_uuid.uuid,
 698                       imp->imp_connection->c_remote_uuid.uuid,
 699                       imp->imp_reqs_replayed,
 700                       imp->imp_locks_replayed);
 701         }
 702
 703         if (imp->imp_state == LUSTRE_IMP_FULL) {
 704                 wake_up(&imp->imp_recovery_waitq);
 705                 ptlrpc_wake_delayed(imp);
 706         }
 707
 708  out:
 709         RETURN(rc);
 710 }
 711
 712 static int back_to_sleep(void *unused)
 713 {
 714         return 0;
 715 }
 716
 717 int ptlrpc_disconnect_import(struct obd_import *imp)
 718 {
 719         struct ptlrpc_request *request;
 720         int rq_opc;
 721         int rc = 0;
 722         unsigned long flags;
 723         ENTRY;
 724
 725         switch (imp->imp_connect_op) {
 726         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
 727         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
 728         case MGMT_CONNECT: rq_opc = MGMT_DISCONNECT; break;
 729         default:
 730                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
 731                        imp->imp_target_uuid.uuid, imp->imp_connect_op);
 732                 RETURN(-EINVAL);
 733         }
 734
 735
 736         if (ptlrpc_import_in_recovery(imp)) {
 737                 struct l_wait_info lwi;
 738                 unsigned long timeout;
 739                 if (imp->imp_server_timeout)
 740                         timeout = obd_timeout / 2;
 741                 else
 742                         timeout = obd_timeout;
 743                 timeout = MAX(timeout * HZ, 1);
 744                 lwi = LWI_TIMEOUT_INTR(obd_timeout, back_to_sleep, NULL, NULL);
 745                 rc = l_wait_event(imp->imp_recovery_waitq,
 746                                   !ptlrpc_import_in_recovery(imp), &lwi);
 747
 748         }
 749
 750         spin_lock_irqsave(&imp->imp_lock, flags);
 751         if (imp->imp_state != LUSTRE_IMP_FULL) {
 752                 GOTO(out, 0);
 753         }
 754         spin_unlock_irqrestore(&imp->imp_lock, flags);
 755
 756         request = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, rq_opc,
 757                                   0, NULL, NULL);
 758         if (request) {
 759                 /* For non-replayable connections, don't attempt
 760                    reconnect if this fails */
 761                 if (!imp->imp_replayable) {
 762                         request->rq_no_resend = 1;
 763                         IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
 764                         request->rq_send_state =  LUSTRE_IMP_CONNECTING;
 765                 }
 766                 request->rq_replen = lustre_msg_size(0, NULL);
 767                 rc = ptlrpc_queue_wait(request);
 768                 ptlrpc_req_finished(request);
 769         }
 770
 771         spin_lock_irqsave(&imp->imp_lock, flags);
 772 out:
 773         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
 774         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
 775         imp->imp_conn_cnt = 0;
 776         spin_unlock_irqrestore(&imp->imp_lock, flags);
 777
 778         RETURN(rc);
 779 }
 780