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 #define UUID_STR "_UUID"
  93 static void deuuidify(char *uuid, const char *prefix, char **uuid_start, int *uuid_len)
  94 {
  95         *uuid_start = !prefix || strncmp(uuid, prefix, strlen(prefix))
  96                 ? uuid : uuid + strlen(prefix);
  97
  98         *uuid_len = strlen(*uuid_start);
  99
 100         if (*uuid_len < strlen(UUID_STR))
 101                 return;
 102
 103         if (!strncmp(*uuid_start + *uuid_len - strlen(UUID_STR),
 104                     UUID_STR, strlen(UUID_STR)))
 105                 *uuid_len -= strlen(UUID_STR);
 106 }
 107
 108 /* Returns true if import was FULL, false if import was already not
 109  * connected.
 110  */
 111 int ptlrpc_set_import_discon(struct obd_import *imp)
 112 {
 113         unsigned long flags;
 114         int rc = 0;
 115
 116         spin_lock_irqsave(&imp->imp_lock, flags);
 117
 118         if (imp->imp_state == LUSTRE_IMP_FULL) {
 119                 char nidbuf[PTL_NALFMT_SIZE];
 120                 char *target_start;
 121                 int   target_len;
 122
 123                 deuuidify(imp->imp_target_uuid.uuid, NULL,
 124                           &target_start, &target_len);
 125
 126                 LCONSOLE_ERROR("Connection to service %.*s via nid %s was "
 127                                "lost; in progress operations using this "
 128                                "service will %s.\n",
 129                                target_len, target_start,
 130                                ptlrpc_peernid2str(&imp->imp_connection->c_peer,
 131                                                   nidbuf),
 132                                imp->imp_replayable
 133                                ? "wait for recovery to complete"
 134                                : "fail");
 135
 136                 CWARN("%s: connection lost to %s@%s\n",
 137                       imp->imp_obd->obd_name,
 138                       imp->imp_target_uuid.uuid,
 139                       imp->imp_connection->c_remote_uuid.uuid);
 140                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 141                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 142                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 143                 rc = 1;
 144         } else {
 145                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 146                 CDEBUG(D_HA, "%p %s: import already not connected: %s\n",
 147                        imp,imp->imp_client->cli_name,
 148                        ptlrpc_import_state_name(imp->imp_state));
 149         }
 150
 151         return rc;
 152 }
 153
 154 /*
 155  * This acts as a barrier; all existing requests are rejected, and
 156  * no new requests will be accepted until the import is valid again.
 157  */
 158 void ptlrpc_deactivate_import(struct obd_import *imp)
 159 {
 160         unsigned long flags;
 161         ENTRY;
 162
 163         spin_lock_irqsave(&imp->imp_lock, flags);
 164         CDEBUG(D_HA, "setting import %s INVALID\n", imp->imp_target_uuid.uuid);
 165         imp->imp_invalid = 1;
 166         imp->imp_generation++;
 167         spin_unlock_irqrestore(&imp->imp_lock, flags);
 168
 169         ptlrpc_abort_inflight(imp);
 170         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 171 }
 172
 173 /*
 174  * This function will invalidate the import, if necessary, then block
 175  * for all the RPC completions, and finally notify the obd to
 176  * invalidate its state (ie cancel locks, clear pending requests,
 177  * etc).
 178  */
 179 void ptlrpc_invalidate_import(struct obd_import *imp)
 180 {
 181         struct l_wait_info lwi;
 182         int rc;
 183
 184         if (!imp->imp_invalid)
 185                 ptlrpc_deactivate_import(imp);
 186
 187         LASSERT(imp->imp_invalid);
 188
 189         /* wait for all requests to error out and call completion callbacks */
 190         lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), NULL,
 191                                NULL, NULL);
 192         rc = l_wait_event(imp->imp_recovery_waitq,
 193                           (atomic_read(&imp->imp_inflight) == 0),
 194                           &lwi);
 195
 196         if (rc)
 197                 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 198                        imp->imp_target_uuid.uuid, rc,
 199                        atomic_read(&imp->imp_inflight));
 200
 201         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 202 }
 203
 204 void ptlrpc_activate_import(struct obd_import *imp)
 205 {
 206         struct obd_device *obd = imp->imp_obd;
 207         unsigned long flags;
 208
 209         spin_lock_irqsave(&imp->imp_lock, flags);
 210         imp->imp_invalid = 0;
 211         spin_unlock_irqrestore(&imp->imp_lock, flags);
 212
 213         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 214 }
 215
 216 void ptlrpc_fail_import(struct obd_import *imp, int generation)
 217 {
 218         ENTRY;
 219
 220         LASSERT (!imp->imp_dlm_fake);
 221
 222         if (ptlrpc_set_import_discon(imp)) {
 223                 unsigned long flags;
 224
 225                 if (!imp->imp_replayable) {
 226                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 227                                "auto-deactivating\n",
 228                                imp->imp_target_uuid.uuid,
 229                                imp->imp_connection->c_remote_uuid.uuid,
 230                                imp->imp_obd->obd_name);
 231                         ptlrpc_deactivate_import(imp);
 232                 }
 233
 234                 CDEBUG(D_HA, "%s: waking up pinger\n",
 235                        imp->imp_target_uuid.uuid);
 236
 237                 spin_lock_irqsave(&imp->imp_lock, flags);
 238                 imp->imp_force_verify = 1;
 239                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 240
 241                 ptlrpc_pinger_wake_up();
 242         }
 243         EXIT;
 244 }
 245
 246 #define ATTEMPT_TOO_SOON(last)  \
 247         ((last) && ((long)(jiffies - (last)) <= (long)(obd_timeout * 2 * HZ)))
 248
 249 static int import_select_connection(struct obd_import *imp)
 250 {
 251         struct obd_import_conn *imp_conn, *tmp;
 252         struct obd_export *dlmexp;
 253         int found = 0;
 254         ENTRY;
 255
 256         spin_lock(&imp->imp_lock);
 257
 258         if (list_empty(&imp->imp_conn_list)) {
 259                 CERROR("%s: no connections available\n",
 260                         imp->imp_obd->obd_name);
 261                 spin_unlock(&imp->imp_lock);
 262                 RETURN(-EINVAL);
 263         }
 264
 265         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 266                 if (!imp_conn->oic_last_attempt ||
 267                     time_after(jiffies, imp_conn->oic_last_attempt +
 268                                obd_timeout * 2 * HZ)) {
 269                         found = 1;
 270                         break;
 271                 }
 272         }
 273
 274         /* if not found, simply choose the current one */
 275         if (!found) {
 276                 CWARN("%s: continuing with current connection\n",
 277                       imp->imp_obd->obd_name);
 278                 LASSERT(imp->imp_conn_current);
 279                 imp_conn = imp->imp_conn_current;
 280         }
 281         LASSERT(imp_conn->oic_conn);
 282
 283         imp_conn->oic_last_attempt = jiffies;
 284
 285         /* move the items ahead of the selected one to list tail */
 286         while (1) {
 287                 tmp= list_entry(imp->imp_conn_list.next,
 288                                 struct obd_import_conn, oic_item);
 289                 if (tmp == imp_conn)
 290                         break;
 291                 list_del(&tmp->oic_item);
 292                 list_add_tail(&tmp->oic_item, &imp->imp_conn_list);
 293         }
 294
 295         /* switch connection, don't mind if it's same as the current one */
 296         if (imp->imp_connection)
 297                 ptlrpc_put_connection(imp->imp_connection);
 298         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 299
 300         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 301         LASSERT(dlmexp != NULL);
 302         if (dlmexp->exp_connection)
 303                 ptlrpc_put_connection(dlmexp->exp_connection);
 304         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 305         class_export_put(dlmexp);
 306
 307         imp->imp_conn_current = imp_conn;
 308         CWARN("%s: Using connection %s\n",
 309                imp->imp_obd->obd_name,
 310                imp_conn->oic_uuid.uuid);
 311         spin_unlock(&imp->imp_lock);
 312
 313         RETURN(0);
 314 }
 315
 316 int ptlrpc_connect_import(struct obd_import *imp, char * new_uuid)
 317 {
 318         struct obd_device *obd = imp->imp_obd;
 319         int initial_connect = 0;
 320         int rc;
 321         __u64 committed_before_reconnect = 0;
 322         struct ptlrpc_request *request;
 323         int size[] = {sizeof(imp->imp_target_uuid),
 324                       sizeof(obd->obd_uuid),
 325                       sizeof(imp->imp_dlm_handle),
 326                       sizeof(imp->imp_connect_data)};
 327         char *tmp[] = {imp->imp_target_uuid.uuid,
 328                        obd->obd_uuid.uuid,
 329                        (char *)&imp->imp_dlm_handle,
 330                        (char *)&imp->imp_connect_data};
 331         struct ptlrpc_connect_async_args *aa;
 332         unsigned long flags;
 333
 334         spin_lock_irqsave(&imp->imp_lock, flags);
 335         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 336                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 337                 CERROR("can't connect to a closed import\n");
 338                 RETURN(-EINVAL);
 339         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 340                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 341                 CERROR("already connected\n");
 342                 RETURN(0);
 343         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 344                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 345                 CERROR("already connecting\n");
 346                 RETURN(-EALREADY);
 347         }
 348
 349         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 350
 351         imp->imp_conn_cnt++;
 352         imp->imp_resend_replay = 0;
 353
 354         if (imp->imp_remote_handle.cookie == 0) {
 355                 initial_connect = 1;
 356         } else {
 357                 committed_before_reconnect = imp->imp_peer_committed_transno;
 358         }
 359
 360         spin_unlock_irqrestore(&imp->imp_lock, flags);
 361
 362         if (new_uuid) {
 363                 struct obd_uuid uuid;
 364
 365                 obd_str2uuid(&uuid, new_uuid);
 366                 rc = import_set_conn_priority(imp, &uuid);
 367                 if (rc)
 368                         GOTO(out, rc);
 369         }
 370
 371         rc = import_select_connection(imp);
 372         if (rc)
 373                 GOTO(out, rc);
 374
 375         request = ptlrpc_prep_req(imp, imp->imp_connect_op, 4, size, tmp);
 376         if (!request)
 377                 GOTO(out, rc = -ENOMEM);
 378
 379 #ifndef __KERNEL__
 380         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 381 #endif
 382
 383         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 384         size[0] = sizeof(struct obd_connect_data);
 385         request->rq_replen = lustre_msg_size(1, size);
 386         request->rq_interpret_reply = ptlrpc_connect_interpret;
 387
 388         LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
 389         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 390         memset(aa, 0, sizeof *aa);
 391
 392         aa->pcaa_peer_committed = committed_before_reconnect;
 393         aa->pcaa_initial_connect = initial_connect;
 394
 395         if (aa->pcaa_initial_connect)
 396                 imp->imp_replayable = 1;
 397
 398         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 399         ptlrpcd_add_req(request);
 400         rc = 0;
 401 out:
 402         if (rc != 0) {
 403                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 404         }
 405
 406         RETURN(rc);
 407 }
 408
 409 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 410                                     void * data, int rc)
 411 {
 412         struct ptlrpc_connect_async_args *aa = data;
 413         struct obd_import *imp = request->rq_import;
 414         struct lustre_handle old_hdl;
 415         unsigned long flags;
 416         int msg_flags;
 417         ENTRY;
 418
 419         spin_lock_irqsave(&imp->imp_lock, flags);
 420         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 421                 spin_unlock_irqrestore(&imp->imp_lock, flags);
 422                 RETURN(0);
 423         }
 424         spin_unlock_irqrestore(&imp->imp_lock, flags);
 425
 426         if (rc)
 427                 GOTO(out, rc);
 428
 429         LASSERT(imp->imp_conn_current);
 430         imp->imp_conn_current->oic_last_attempt = 0;
 431
 432         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 433
 434         /* All imports are pingable */
 435         imp->imp_pingable = 1;
 436
 437         if (aa->pcaa_initial_connect) {
 438                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 439                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 440                                imp->imp_target_uuid.uuid);
 441                         imp->imp_replayable = 1;
 442                 } else {
 443                         imp->imp_replayable = 0;
 444                 }
 445                 imp->imp_remote_handle = request->rq_repmsg->handle;
 446                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 447                 GOTO(finish, rc = 0);
 448         }
 449
 450         /* Determine what recovery state to move the import to. */
 451         if (MSG_CONNECT_RECONNECT & msg_flags) {
 452                 memset(&old_hdl, 0, sizeof(old_hdl));
 453                 if (!memcmp(&old_hdl, &request->rq_repmsg->handle,
 454                             sizeof (old_hdl))) {
 455                         CERROR("%s@%s didn't like our handle "LPX64
 456                                ", failed\n", imp->imp_target_uuid.uuid,
 457                                imp->imp_connection->c_remote_uuid.uuid,
 458                                imp->imp_dlm_handle.cookie);
 459                         GOTO(out, rc = -ENOTCONN);
 460                 }
 461
 462                 if (memcmp(&imp->imp_remote_handle, &request->rq_repmsg->handle,
 463                            sizeof(imp->imp_remote_handle))) {
 464                         CERROR("%s@%s changed handle from "LPX64" to "LPX64
 465                                "; copying, but this may foreshadow disaster\n",
 466                                imp->imp_target_uuid.uuid,
 467                                imp->imp_connection->c_remote_uuid.uuid,
 468                                imp->imp_remote_handle.cookie,
 469                                request->rq_repmsg->handle.cookie);
 470                         imp->imp_remote_handle = request->rq_repmsg->handle;
 471                 } else {
 472                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 473                                imp->imp_target_uuid.uuid,
 474                                imp->imp_connection->c_remote_uuid.uuid);
 475                 }
 476
 477                 if (imp->imp_invalid) {
 478                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 479                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 480                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 481                                imp->imp_obd->obd_name,
 482                                imp->imp_target_uuid.uuid);
 483                         imp->imp_resend_replay = 1;
 484                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 485                 } else {
 486                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 487                 }
 488         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 489                 LASSERT(imp->imp_replayable);
 490                 imp->imp_remote_handle = request->rq_repmsg->handle;
 491                 imp->imp_last_replay_transno = 0;
 492                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 493         } else {
 494                 imp->imp_remote_handle = request->rq_repmsg->handle;
 495                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 496         }
 497
 498         /* Sanity checks for a reconnected import. */
 499         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 500                 CERROR("imp_replayable flag does not match server "
 501                        "after reconnect. We should LBUG right here.\n");
 502         }
 503
 504         if (request->rq_repmsg->last_committed < aa->pcaa_peer_committed) {
 505                 CERROR("%s went back in time (transno "LPD64
 506                        " was previously committed, server now claims "LPD64
 507                        ")! is shared storage not coherent?\n",
 508                        imp->imp_target_uuid.uuid,
 509                        aa->pcaa_peer_committed,
 510                        request->rq_repmsg->last_committed);
 511         }
 512
 513 finish:
 514         rc = ptlrpc_import_recovery_state_machine(imp);
 515         if (rc != 0) {
 516                 if (rc == -ENOTCONN) {
 517                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 518                                "invalidating and reconnecting\n",
 519                                imp->imp_target_uuid.uuid,
 520                                imp->imp_connection->c_remote_uuid.uuid);
 521                         ptlrpc_connect_import(imp, NULL);
 522                         RETURN(0);
 523                 }
 524         }
 525  out:
 526         if (rc != 0) {
 527                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 528                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov) {
 529                         ptlrpc_deactivate_import(imp);
 530                 }
 531                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 532                        imp->imp_target_uuid.uuid,
 533                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 534         }
 535
 536         wake_up(&imp->imp_recovery_waitq);
 537         RETURN(rc);
 538 }
 539
 540 static int completed_replay_interpret(struct ptlrpc_request *req,
 541                                     void * data, int rc)
 542 {
 543         atomic_dec(&req->rq_import->imp_replay_inflight);
 544         if (req->rq_status == 0) {
 545                 ptlrpc_import_recovery_state_machine(req->rq_import);
 546         } else {
 547                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 548                        "reconnecting\n",
 549                        req->rq_import->imp_obd->obd_name, req->rq_status);
 550                 ptlrpc_connect_import(req->rq_import, NULL);
 551         }
 552
 553         RETURN(0);
 554 }
 555
 556 static int signal_completed_replay(struct obd_import *imp)
 557 {
 558         struct ptlrpc_request *req;
 559         ENTRY;
 560
 561         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 562         atomic_inc(&imp->imp_replay_inflight);
 563
 564         req = ptlrpc_prep_req(imp, OBD_PING, 0, NULL, NULL);
 565         if (!req) {
 566                 atomic_dec(&imp->imp_replay_inflight);
 567                 RETURN(-ENOMEM);
 568         }
 569
 570         req->rq_replen = lustre_msg_size(0, NULL);
 571         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 572         req->rq_reqmsg->flags |= MSG_LAST_REPLAY;
 573         req->rq_timeout *= 3;
 574         req->rq_interpret_reply = completed_replay_interpret;
 575
 576         ptlrpcd_add_req(req);
 577         RETURN(0);
 578 }
 579
 580 static int ptlrpc_invalidate_import_thread(void *data)
 581 {
 582         struct obd_import *imp = data;
 583         unsigned long flags;
 584
 585         ENTRY;
 586
 587         lock_kernel();
 588         ptlrpc_daemonize();
 589
 590         SIGNAL_MASK_LOCK(current, flags);
 591         sigfillset(&current->blocked);
 592         RECALC_SIGPENDING;
 593         SIGNAL_MASK_UNLOCK(current, flags);
 594         THREAD_NAME(current->comm, sizeof(current->comm), "ll_imp_inval");
 595         unlock_kernel();
 596
 597         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
 598                imp->imp_obd->obd_name, imp->imp_target_uuid.uuid,
 599                imp->imp_connection->c_remote_uuid.uuid);
 600
 601         ptlrpc_invalidate_import(imp);
 602
 603         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 604         ptlrpc_import_recovery_state_machine(imp);
 605
 606         RETURN(0);
 607 }
 608
 609 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
 610 {
 611         int rc = 0;
 612         int inflight;
 613         char *target_start;
 614         int target_len;
 615
 616         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
 617                 deuuidify(imp->imp_target_uuid.uuid, NULL,
 618                           &target_start, &target_len);
 619                 LCONSOLE_ERROR("This client was evicted by %.*s; in progress "
 620                                "operations using this service will %s.\n",
 621                                target_len, target_start,
 622                                imp->imp_replayable
 623                                ? "be reattempted"
 624                                : "fail");
 625                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
 626                        imp->imp_target_uuid.uuid,
 627                        imp->imp_connection->c_remote_uuid.uuid);
 628
 629                 rc = kernel_thread(ptlrpc_invalidate_import_thread, imp,
 630                                    CLONE_VM | CLONE_FILES);
 631                 if (rc < 0)
 632                         CERROR("error starting invalidate thread: %d\n", rc);
 633                 RETURN(rc);
 634         }
 635
 636         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
 637                 CDEBUG(D_HA, "replay requested by %s\n",
 638                        imp->imp_target_uuid.uuid);
 639                 rc = ptlrpc_replay_next(imp, &inflight);
 640                 if (inflight == 0 &&
 641                     atomic_read(&imp->imp_replay_inflight) == 0) {
 642                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
 643                         rc = ldlm_replay_locks(imp);
 644                         if (rc)
 645                                 GOTO(out, rc);
 646                 }
 647                 rc = 0;
 648         }
 649
 650         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
 651                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 652                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
 653                         rc = signal_completed_replay(imp);
 654                         if (rc)
 655                                 GOTO(out, rc);
 656                 }
 657
 658         }
 659
 660         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
 661                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
 662                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 663                 }
 664         }
 665
 666         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
 667                 char nidbuf[PTL_NALFMT_SIZE];
 668
 669                 CDEBUG(D_HA, "reconnected to %s@%s\n",
 670                        imp->imp_target_uuid.uuid,
 671                        imp->imp_connection->c_remote_uuid.uuid);
 672
 673                 rc = ptlrpc_resend(imp);
 674                 if (rc)
 675                         GOTO(out, rc);
 676                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 677                 ptlrpc_activate_import(imp);
 678
 679                 deuuidify(imp->imp_target_uuid.uuid, NULL,
 680                           &target_start, &target_len);
 681                 ptlrpc_peernid2str(&imp->imp_connection->c_peer,
 682                                    nidbuf);
 683
 684                 LCONSOLE_INFO("Connection restored to service %.*s using nid "
 685                               "%s.\n",
 686                               target_len, target_start, nidbuf);
 687
 688                 CWARN("%s: connection restored to %s@%s\n",
 689                       imp->imp_obd->obd_name,
 690                       imp->imp_target_uuid.uuid,
 691                       imp->imp_connection->c_remote_uuid.uuid);
 692         }
 693
 694         if (imp->imp_state == LUSTRE_IMP_FULL) {
 695                 wake_up(&imp->imp_recovery_waitq);
 696                 ptlrpc_wake_delayed(imp);
 697         }
 698
 699  out:
 700         RETURN(rc);
 701 }
 702
 703 static int back_to_sleep(void *unused)
 704 {
 705         return 0;
 706 }
 707
 708 int ptlrpc_disconnect_import(struct obd_import *imp)
 709 {
 710         struct ptlrpc_request *request;
 711         int rq_opc;
 712         int rc = 0;
 713         unsigned long flags;
 714         ENTRY;
 715
 716         switch (imp->imp_connect_op) {
 717         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
 718         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
 719         case MGMT_CONNECT:rq_opc = MGMT_DISCONNECT;break;
 720         default:
 721                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
 722                        imp->imp_target_uuid.uuid, imp->imp_connect_op);
 723                 RETURN(-EINVAL);
 724         }
 725
 726
 727         if (ptlrpc_import_in_recovery(imp)) {
 728                 struct l_wait_info lwi;
 729                 lwi = LWI_TIMEOUT_INTR(MAX(obd_timeout * HZ, 1), back_to_sleep,
 730                                        NULL, NULL);
 731                 rc = l_wait_event(imp->imp_recovery_waitq,
 732                                   !ptlrpc_import_in_recovery(imp), &lwi);
 733
 734         }
 735
 736         spin_lock_irqsave(&imp->imp_lock, flags);
 737         if (imp->imp_state != LUSTRE_IMP_FULL) {
 738                 GOTO(out, 0);
 739         }
 740         spin_unlock_irqrestore(&imp->imp_lock, flags);
 741
 742         request = ptlrpc_prep_req(imp, rq_opc, 0, NULL, NULL);
 743         if (request) {
 744                 /* For non-replayable connections, don't attempt
 745                    reconnect if this fails */
 746                 if (!imp->imp_replayable) {
 747                         request->rq_no_resend = 1;
 748                         IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
 749                         request->rq_send_state =  LUSTRE_IMP_CONNECTING;
 750                 }
 751                 request->rq_replen = lustre_msg_size(0, NULL);
 752                 rc = ptlrpc_queue_wait(request);
 753                 ptlrpc_req_finished(request);
 754         }
 755
 756         spin_lock_irqsave(&imp->imp_lock, flags);
 757 out:
 758         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
 759         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
 760         spin_unlock_irqrestore(&imp->imp_lock, flags);
 761
 762         RETURN(rc);
 763 }
 764