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_cksum.h>
  38 #include <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, 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                 if (imp->imp_replayable) {
 131                         LCONSOLE_WARN("%s: Connection to service %.*s via nid "
 132                                "%s was lost; in progress operations using this "
 133                                "service will wait for recovery to complete.\n",
 134                                imp->imp_obd->obd_name, target_len, target_start,
 135                                libcfs_nid2str(imp->imp_connection->c_peer.nid));
 136                 } else {
 137                         LCONSOLE_ERROR_MSG(0x166, "%s: Connection to service "
 138                                            "%.*s via nid %s was lost; in progress"
 139                                            "operations using this service will"
 140                                            "fail.\n",
 141                                            imp->imp_obd->obd_name,
 142                                            target_len, target_start,
 143                                  libcfs_nid2str(imp->imp_connection->c_peer.nid));
 144                 }
 145                 ptlrpc_deactivate_timeouts(imp);
 146                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
 147                 spin_unlock(&imp->imp_lock);
 148
 149                 if (obd_dump_on_timeout)
 150                         libcfs_debug_dumplog();
 151
 152                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_DISCON);
 153                 rc = 1;
 154         } else {
 155                 spin_unlock(&imp->imp_lock);
 156                 CDEBUG(D_HA, "%s: import %p already %s (conn %u, was %u): %s\n",
 157                        imp->imp_client->cli_name, imp,
 158                        (imp->imp_state == LUSTRE_IMP_FULL &&
 159                         imp->imp_conn_cnt > conn_cnt) ?
 160                        "reconnected" : "not connected", imp->imp_conn_cnt,
 161                        conn_cnt, ptlrpc_import_state_name(imp->imp_state));
 162         }
 163
 164         return rc;
 165 }
 166
 167 /* Must be called with imp_lock held! */
 168 static void ptlrpc_deactivate_and_unlock_import(struct obd_import *imp)
 169 {
 170         ENTRY;
 171         LASSERT_SPIN_LOCKED(&imp->imp_lock);
 172
 173         CDEBUG(D_HA, "setting import %s INVALID\n", obd2cli_tgt(imp->imp_obd));
 174         imp->imp_invalid = 1;
 175         imp->imp_generation++;
 176         spin_unlock(&imp->imp_lock);
 177
 178         ptlrpc_abort_inflight(imp);
 179         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INACTIVE);
 180
 181         EXIT;
 182 }
 183
 184 /*
 185  * This acts as a barrier; all existing requests are rejected, and
 186  * no new requests will be accepted until the import is valid again.
 187  */
 188 void ptlrpc_deactivate_import(struct obd_import *imp)
 189 {
 190         spin_lock(&imp->imp_lock);
 191         ptlrpc_deactivate_and_unlock_import(imp);
 192 }
 193
 194 /*
 195  * This function will invalidate the import, if necessary, then block
 196  * for all the RPC completions, and finally notify the obd to
 197  * invalidate its state (ie cancel locks, clear pending requests,
 198  * etc).
 199  */
 200 void ptlrpc_invalidate_import(struct obd_import *imp)
 201 {
 202         struct l_wait_info lwi;
 203         int rc;
 204
 205         atomic_inc(&imp->imp_inval_count);
 206
 207         /*
 208          * If this is an invalid MGC connection, then don't bother
 209          * waiting for imp_inflight to drop to 0.
 210          */
 211         if (imp->imp_invalid && imp->imp_recon_bk && !imp->imp_obd->obd_no_recov)
 212                 goto out;
 213
 214         if (!imp->imp_invalid || imp->imp_obd->obd_no_recov)
 215                 ptlrpc_deactivate_import(imp);
 216
 217         LASSERT(imp->imp_invalid);
 218
 219         /* wait for all requests to error out and call completion callbacks */
 220         lwi = LWI_TIMEOUT_INTERVAL(cfs_timeout_cap(cfs_time_seconds(obd_timeout)),
 221                                    HZ, NULL, NULL);
 222         rc = l_wait_event(imp->imp_recovery_waitq,
 223                           (atomic_read(&imp->imp_inflight) == 0), &lwi);
 224
 225         if (rc) {
 226                 struct list_head *tmp, *n;
 227                 struct ptlrpc_request *req;
 228
 229                 CERROR("%s: rc = %d waiting for callback (%d != 0)\n",
 230                        obd2cli_tgt(imp->imp_obd), rc,
 231                        atomic_read(&imp->imp_inflight));
 232                 spin_lock(&imp->imp_lock);
 233                 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
 234                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 235                         DEBUG_REQ(D_ERROR, req, "still on sending list");
 236                 }
 237                 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
 238                         req = list_entry(tmp, struct ptlrpc_request, rq_list);
 239                         DEBUG_REQ(D_ERROR, req, "still on delayed list");
 240                 }
 241                 spin_unlock(&imp->imp_lock);
 242                 LASSERT(atomic_read(&imp->imp_inflight) == 0);
 243         }
 244
 245 out:
 246         obd_import_event(imp->imp_obd, imp, IMP_EVENT_INVALIDATE);
 247         sptlrpc_import_flush_all_ctx(imp);
 248
 249         atomic_dec(&imp->imp_inval_count);
 250         cfs_waitq_signal(&imp->imp_recovery_waitq);
 251 }
 252
 253 /* unset imp_invalid */
 254 void ptlrpc_activate_import(struct obd_import *imp)
 255 {
 256         struct obd_device *obd = imp->imp_obd;
 257
 258         spin_lock(&imp->imp_lock);
 259         imp->imp_invalid = 0;
 260         ptlrpc_activate_timeouts(imp);
 261         spin_unlock(&imp->imp_lock);
 262         obd_import_event(obd, imp, IMP_EVENT_ACTIVE);
 263 }
 264
 265 void ptlrpc_fail_import(struct obd_import *imp, __u32 conn_cnt)
 266 {
 267         ENTRY;
 268
 269         LASSERT(!imp->imp_dlm_fake);
 270
 271         if (ptlrpc_set_import_discon(imp, conn_cnt)) {
 272                 if (!imp->imp_replayable) {
 273                         CDEBUG(D_HA, "import %s@%s for %s not replayable, "
 274                                "auto-deactivating\n",
 275                                obd2cli_tgt(imp->imp_obd),
 276                                imp->imp_connection->c_remote_uuid.uuid,
 277                                imp->imp_obd->obd_name);
 278                         ptlrpc_deactivate_import(imp);
 279                 }
 280
 281                 CDEBUG(D_HA, "%s: waking up pinger\n",
 282                        obd2cli_tgt(imp->imp_obd));
 283
 284                 spin_lock(&imp->imp_lock);
 285                 imp->imp_force_verify = 1;
 286                 spin_unlock(&imp->imp_lock);
 287
 288                 ptlrpc_pinger_wake_up();
 289         }
 290         EXIT;
 291 }
 292
 293 int ptlrpc_reconnect_import(struct obd_import *imp)
 294 {
 295
 296         ptlrpc_set_import_discon(imp, 0);
 297         /* Force a new connect attempt */
 298         ptlrpc_invalidate_import(imp);
 299         /* Do a fresh connect next time by zeroing the handle */
 300         ptlrpc_disconnect_import(imp, 1);
 301         /* Wait for all invalidate calls to finish */
 302         if (atomic_read(&imp->imp_inval_count) > 0) {
 303                 int rc;
 304                 struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL);
 305                 rc = l_wait_event(imp->imp_recovery_waitq,
 306                                   (atomic_read(&imp->imp_inval_count) == 0),
 307                                   &lwi);
 308                 if (rc)
 309                         CERROR("Interrupted, inval=%d\n",
 310                                atomic_read(&imp->imp_inval_count));
 311         }
 312
 313         /* Allow reconnect attempts */
 314         imp->imp_obd->obd_no_recov = 0;
 315         /* Remove 'invalid' flag */
 316         ptlrpc_activate_import(imp);
 317         /* Attempt a new connect */
 318         ptlrpc_recover_import(imp, NULL);
 319         return 0;
 320 }
 321
 322 EXPORT_SYMBOL(ptlrpc_reconnect_import);
 323
 324 static int import_select_connection(struct obd_import *imp)
 325 {
 326         struct obd_import_conn *imp_conn = NULL, *conn;
 327         struct obd_export *dlmexp;
 328         ENTRY;
 329
 330         spin_lock(&imp->imp_lock);
 331
 332         if (list_empty(&imp->imp_conn_list)) {
 333                 CERROR("%s: no connections available\n",
 334                         imp->imp_obd->obd_name);
 335                 spin_unlock(&imp->imp_lock);
 336                 RETURN(-EINVAL);
 337         }
 338
 339         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 340                 CDEBUG(D_HA, "%s: connect to NID %s last attempt "LPU64"\n",
 341                        imp->imp_obd->obd_name,
 342                        libcfs_nid2str(conn->oic_conn->c_peer.nid),
 343                        conn->oic_last_attempt);
 344                 /* Throttle the reconnect rate to once per RECONNECT_INTERVAL */
 345                 if (cfs_time_before_64(conn->oic_last_attempt +
 346                                        RECONNECT_INTERVAL * HZ,
 347                                        cfs_time_current_64())) {
 348                         /* If we have never tried this connection since the
 349                            the last successful attempt, go with this one */
 350                         if (cfs_time_beforeq_64(conn->oic_last_attempt,
 351                                                imp->imp_last_success_conn)) {
 352                                 imp_conn = conn;
 353                                 break;
 354                         }
 355
 356                         /* Both of these connections have already been tried
 357                            since the last successful connection; just choose the
 358                            least recently used */
 359                         if (!imp_conn)
 360                                 imp_conn = conn;
 361                         else if (cfs_time_before_64(conn->oic_last_attempt,
 362                                                     imp_conn->oic_last_attempt))
 363                                 imp_conn = conn;
 364                 }
 365         }
 366
 367         /* if not found, simply choose the current one */
 368         if (!imp_conn) {
 369                 LASSERT(imp->imp_conn_current);
 370                 imp_conn = imp->imp_conn_current;
 371         }
 372         LASSERT(imp_conn->oic_conn);
 373
 374         imp_conn->oic_last_attempt = cfs_time_current_64();
 375
 376         /* switch connection, don't mind if it's same as the current one */
 377         if (imp->imp_connection)
 378                 ptlrpc_put_connection(imp->imp_connection);
 379         imp->imp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 380
 381         dlmexp =  class_conn2export(&imp->imp_dlm_handle);
 382         LASSERT(dlmexp != NULL);
 383         if (dlmexp->exp_connection)
 384                 ptlrpc_put_connection(dlmexp->exp_connection);
 385         dlmexp->exp_connection = ptlrpc_connection_addref(imp_conn->oic_conn);
 386         class_export_put(dlmexp);
 387
 388         if (imp->imp_conn_current != imp_conn) {
 389                 if (imp->imp_conn_current)
 390                         LCONSOLE_INFO("Changing connection for %s to %s/%s\n",
 391                                       imp->imp_obd->obd_name,
 392                                       imp_conn->oic_uuid.uuid,
 393                                       libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 394                 imp->imp_conn_current = imp_conn;
 395         }
 396
 397         CDEBUG(D_HA, "%s: import %p using connection %s/%s\n",
 398                imp->imp_obd->obd_name, imp, imp_conn->oic_uuid.uuid,
 399                libcfs_nid2str(imp_conn->oic_conn->c_peer.nid));
 400
 401         spin_unlock(&imp->imp_lock);
 402
 403         RETURN(0);
 404 }
 405
 406 /*
 407  * must be called under imp_lock
 408  */
 409 int ptlrpc_first_transno(struct obd_import *imp, __u64 *transno)
 410 {
 411         struct ptlrpc_request *req;
 412         struct list_head *tmp;
 413
 414         if (list_empty(&imp->imp_replay_list))
 415                 return 0;
 416         tmp = imp->imp_replay_list.next;
 417         req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
 418         *transno = req->rq_transno;
 419         if (req->rq_transno == 0) {
 420                 DEBUG_REQ(D_ERROR, req, "zero transno in replay");
 421                 LBUG();
 422         }
 423
 424         return 1;
 425 }
 426
 427 int ptlrpc_connect_import(struct obd_import *imp, char *new_uuid)
 428 {
 429         struct obd_device *obd = imp->imp_obd;
 430         int initial_connect = 0;
 431         int set_transno = 0;
 432         __u64 committed_before_reconnect = 0;
 433         struct ptlrpc_request *request;
 434         char *bufs[] = { NULL,
 435                          obd2cli_tgt(imp->imp_obd),
 436                          obd->obd_uuid.uuid,
 437                          (char *)&imp->imp_dlm_handle,
 438                          (char *)&imp->imp_connect_data };
 439         struct ptlrpc_connect_async_args *aa;
 440         int rc;
 441         ENTRY;
 442
 443         spin_lock(&imp->imp_lock);
 444         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 445                 spin_unlock(&imp->imp_lock);
 446                 CERROR("can't connect to a closed import\n");
 447                 RETURN(-EINVAL);
 448         } else if (imp->imp_state == LUSTRE_IMP_FULL) {
 449                 spin_unlock(&imp->imp_lock);
 450                 CERROR("already connected\n");
 451                 RETURN(0);
 452         } else if (imp->imp_state == LUSTRE_IMP_CONNECTING) {
 453                 spin_unlock(&imp->imp_lock);
 454                 CERROR("already connecting\n");
 455                 RETURN(-EALREADY);
 456         }
 457
 458         IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CONNECTING);
 459
 460         imp->imp_conn_cnt++;
 461         imp->imp_resend_replay = 0;
 462
 463         if (!lustre_handle_is_used(&imp->imp_remote_handle))
 464                 initial_connect = 1;
 465         else
 466                 committed_before_reconnect = imp->imp_peer_committed_transno;
 467
 468         set_transno = ptlrpc_first_transno(imp, &imp->imp_connect_data.ocd_transno);
 469         spin_unlock(&imp->imp_lock);
 470
 471         if (new_uuid) {
 472                 struct obd_uuid uuid;
 473
 474                 obd_str2uuid(&uuid, new_uuid);
 475                 rc = import_set_conn_priority(imp, &uuid);
 476                 if (rc)
 477                         GOTO(out, rc);
 478         }
 479
 480         rc = import_select_connection(imp);
 481         if (rc)
 482                 GOTO(out, rc);
 483
 484         /* last in connection list */
 485         if (imp->imp_conn_current->oic_item.next == &imp->imp_conn_list) {
 486                 if (imp->imp_initial_recov_bk && initial_connect) {
 487                         CDEBUG(D_HA, "Last connection attempt (%d) for %s\n",
 488                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 489                         /* Don't retry if connect fails */
 490                         rc = 0;
 491                         obd_set_info_async(obd->obd_self_export,
 492                                            strlen(KEY_INIT_RECOV),
 493                                            KEY_INIT_RECOV,
 494                                            sizeof(rc), &rc, NULL);
 495                 }
 496                 if (imp->imp_recon_bk) {
 497                         CDEBUG(D_HA, "Last reconnection attempt (%d) for %s\n",
 498                                imp->imp_conn_cnt, obd2cli_tgt(imp->imp_obd));
 499                         spin_lock(&imp->imp_lock);
 500                         imp->imp_last_recon = 1;
 501                         spin_unlock(&imp->imp_lock);
 502                 }
 503         }
 504
 505         rc = sptlrpc_import_sec_adapt(imp, NULL, 0);
 506         if (rc)
 507                 GOTO(out, rc);
 508
 509         /* Reset connect flags to the originally requested flags, in case
 510          * the server is updated on-the-fly we will get the new features. */
 511         imp->imp_connect_data.ocd_connect_flags = imp->imp_connect_flags_orig;
 512         rc = obd_reconnect(NULL, imp->imp_obd->obd_self_export, obd,
 513                            &obd->obd_uuid, &imp->imp_connect_data);
 514         if (rc)
 515                 GOTO(out, rc);
 516
 517         request = ptlrpc_request_alloc(imp, &RQF_MDS_CONNECT);
 518         if (request == NULL)
 519                 GOTO(out, rc = -ENOMEM);
 520
 521         rc = ptlrpc_request_bufs_pack(request, LUSTRE_OBD_VERSION,
 522                                       imp->imp_connect_op, bufs, NULL);
 523         if (rc) {
 524                 ptlrpc_request_free(request);
 525                 GOTO(out, rc);
 526         }
 527
 528 #ifndef __KERNEL__
 529         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_LIBCLIENT);
 530 #endif
 531         lustre_msg_add_op_flags(request->rq_reqmsg, MSG_CONNECT_NEXT_VER);
 532
 533         request->rq_send_state = LUSTRE_IMP_CONNECTING;
 534         /* Allow a slightly larger reply for future growth compatibility */
 535         req_capsule_set_size(&request->rq_pill, &RMF_CONNECT_DATA, RCL_SERVER,
 536                              sizeof(struct obd_connect_data)+16*sizeof(__u64));
 537         ptlrpc_request_set_replen(request);
 538         request->rq_interpret_reply = ptlrpc_connect_interpret;
 539
 540         CLASSERT(sizeof (*aa) <= sizeof (request->rq_async_args));
 541         aa = (struct ptlrpc_connect_async_args *)&request->rq_async_args;
 542         memset(aa, 0, sizeof *aa);
 543
 544         aa->pcaa_peer_committed = committed_before_reconnect;
 545         aa->pcaa_initial_connect = initial_connect;
 546
 547         if (aa->pcaa_initial_connect) {
 548                 spin_lock(&imp->imp_lock);
 549                 imp->imp_replayable = 1;
 550                 spin_unlock(&imp->imp_lock);
 551                 /* On an initial connect, we don't know which one of a
 552                    failover server pair is up.  Don't wait long. */
 553 #ifdef CRAY_XT3
 554                 request->rq_timeout = max((int)(obd_timeout / 2), 5);
 555 #else
 556                 request->rq_timeout = max((int)(obd_timeout / 20), 5);
 557 #endif
 558                 lustre_msg_add_op_flags(request->rq_reqmsg,
 559                                         MSG_CONNECT_INITIAL);
 560         }
 561
 562         if (set_transno)
 563                 lustre_msg_add_op_flags(request->rq_reqmsg,
 564                                         MSG_CONNECT_TRANSNO);
 565
 566         DEBUG_REQ(D_RPCTRACE, request, "(re)connect request");
 567         ptlrpcd_add_req(request);
 568         rc = 0;
 569 out:
 570         if (rc != 0) {
 571                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 572         }
 573
 574         RETURN(rc);
 575 }
 576 EXPORT_SYMBOL(ptlrpc_connect_import);
 577
 578 static void ptlrpc_maybe_ping_import_soon(struct obd_import *imp)
 579 {
 580 #ifdef __KERNEL__
 581         struct obd_import_conn *imp_conn;
 582 #endif
 583         int wake_pinger = 0;
 584
 585         ENTRY;
 586
 587         spin_lock(&imp->imp_lock);
 588         if (list_empty(&imp->imp_conn_list))
 589                 GOTO(unlock, 0);
 590
 591 #ifdef __KERNEL__
 592         imp_conn = list_entry(imp->imp_conn_list.prev,
 593                               struct obd_import_conn,
 594                               oic_item);
 595
 596         /* XXX: When the failover node is the primary node, it is possible
 597          * to have two identical connections in imp_conn_list. We must
 598          * compare not conn's pointers but NIDs, otherwise we can defeat
 599          * connection throttling. (See bug 14774.) */
 600         if (imp->imp_conn_current->oic_conn->c_self !=
 601                                 imp_conn->oic_conn->c_self) {
 602                 ptlrpc_ping_import_soon(imp);
 603                 wake_pinger = 1;
 604         }
 605 #else
 606         /* liblustre has no pinger thead, so we wakup pinger anyway */
 607         wake_pinger = 1;
 608 #endif
 609
 610  unlock:
 611         spin_unlock(&imp->imp_lock);
 612
 613         if (wake_pinger)
 614                 ptlrpc_pinger_wake_up();
 615
 616         EXIT;
 617 }
 618
 619 static int ptlrpc_connect_interpret(struct ptlrpc_request *request,
 620                                     void * data, int rc)
 621 {
 622         struct ptlrpc_connect_async_args *aa = data;
 623         struct obd_import *imp = request->rq_import;
 624         struct client_obd *cli = &imp->imp_obd->u.cli;
 625         struct lustre_handle old_hdl;
 626         int msg_flags;
 627         ENTRY;
 628
 629         spin_lock(&imp->imp_lock);
 630         if (imp->imp_state == LUSTRE_IMP_CLOSED) {
 631                 spin_unlock(&imp->imp_lock);
 632                 RETURN(0);
 633         }
 634         spin_unlock(&imp->imp_lock);
 635
 636         if (rc)
 637                 GOTO(out, rc);
 638
 639         LASSERT(imp->imp_conn_current);
 640
 641         msg_flags = lustre_msg_get_op_flags(request->rq_repmsg);
 642
 643         /* All imports are pingable */
 644         spin_lock(&imp->imp_lock);
 645         imp->imp_pingable = 1;
 646
 647         if (aa->pcaa_initial_connect) {
 648                 if (msg_flags & MSG_CONNECT_REPLAYABLE) {
 649                         imp->imp_replayable = 1;
 650                         spin_unlock(&imp->imp_lock);
 651                         CDEBUG(D_HA, "connected to replayable target: %s\n",
 652                                obd2cli_tgt(imp->imp_obd));
 653                 } else {
 654                         imp->imp_replayable = 0;
 655                         spin_unlock(&imp->imp_lock);
 656                 }
 657
 658                 /* if applies, adjust the imp->imp_msg_magic here
 659                  * according to reply flags */
 660
 661                 imp->imp_remote_handle =
 662                                 *lustre_msg_get_handle(request->rq_repmsg);
 663
 664                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
 665                 spin_lock(&imp->imp_lock);
 666                 if (imp->imp_invalid) {
 667                         spin_unlock(&imp->imp_lock);
 668                         ptlrpc_activate_import(imp);
 669                 } else {
 670                         spin_unlock(&imp->imp_lock);
 671                 }
 672
 673                 GOTO(finish, rc = 0);
 674         } else {
 675                 spin_unlock(&imp->imp_lock);
 676         }
 677
 678         /* Determine what recovery state to move the import to. */
 679         if (MSG_CONNECT_RECONNECT & msg_flags) {
 680                 memset(&old_hdl, 0, sizeof(old_hdl));
 681                 if (!memcmp(&old_hdl, lustre_msg_get_handle(request->rq_repmsg),
 682                             sizeof (old_hdl))) {
 683                         CERROR("%s@%s didn't like our handle "LPX64
 684                                ", failed\n", obd2cli_tgt(imp->imp_obd),
 685                                imp->imp_connection->c_remote_uuid.uuid,
 686                                imp->imp_dlm_handle.cookie);
 687                         GOTO(out, rc = -ENOTCONN);
 688                 }
 689
 690                 if (memcmp(&imp->imp_remote_handle,
 691                            lustre_msg_get_handle(request->rq_repmsg),
 692                            sizeof(imp->imp_remote_handle))) {
 693
 694                         CWARN("%s@%s changed server handle from "
 695                                LPX64" to "LPX64" - evicting.\n",
 696                                obd2cli_tgt(imp->imp_obd),
 697                                imp->imp_connection->c_remote_uuid.uuid,
 698                                imp->imp_remote_handle.cookie,
 699                                lustre_msg_get_handle(request->rq_repmsg)->
 700                                          cookie);
 701                         imp->imp_remote_handle =
 702                                      *lustre_msg_get_handle(request->rq_repmsg);
 703
 704                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 705                         GOTO(finish, rc = 0);
 706                 } else {
 707                         CDEBUG(D_HA, "reconnected to %s@%s after partition\n",
 708                                obd2cli_tgt(imp->imp_obd),
 709                                imp->imp_connection->c_remote_uuid.uuid);
 710                 }
 711
 712                 if (imp->imp_invalid) {
 713                         IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 714                 } else if (MSG_CONNECT_RECOVERING & msg_flags) {
 715                         CDEBUG(D_HA, "%s: reconnected to %s during replay\n",
 716                                imp->imp_obd->obd_name,
 717                                obd2cli_tgt(imp->imp_obd));
 718
 719                         spin_lock(&imp->imp_lock);
 720                         imp->imp_resend_replay = 1;
 721                         spin_unlock(&imp->imp_lock);
 722
 723                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 724                 } else {
 725                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
 726                 }
 727         } else if ((MSG_CONNECT_RECOVERING & msg_flags) && !imp->imp_invalid) {
 728                 LASSERT(imp->imp_replayable);
 729                 imp->imp_remote_handle =
 730                                 *lustre_msg_get_handle(request->rq_repmsg);
 731                 imp->imp_last_replay_transno = 0;
 732                 IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY);
 733         } else {
 734                 imp->imp_remote_handle =
 735                                 *lustre_msg_get_handle(request->rq_repmsg);
 736                 IMPORT_SET_STATE(imp, LUSTRE_IMP_EVICTED);
 737         }
 738
 739         /* Sanity checks for a reconnected import. */
 740         if (!(imp->imp_replayable) != !(msg_flags & MSG_CONNECT_REPLAYABLE)) {
 741                 CERROR("imp_replayable flag does not match server "
 742                        "after reconnect. We should LBUG right here.\n");
 743         }
 744
 745         if (lustre_msg_get_last_committed(request->rq_repmsg) <
 746             aa->pcaa_peer_committed) {
 747                 CERROR("%s went back in time (transno "LPD64
 748                        " was previously committed, server now claims "LPD64
 749                        ")!  See https://bugzilla.clusterfs.com/"
 750                        "long_list.cgi?buglist=9646\n",
 751                        obd2cli_tgt(imp->imp_obd), aa->pcaa_peer_committed,
 752                        lustre_msg_get_last_committed(request->rq_repmsg));
 753         }
 754
 755 finish:
 756         rc = ptlrpc_import_recovery_state_machine(imp);
 757         if (rc != 0) {
 758                 if (rc == -ENOTCONN) {
 759                         CDEBUG(D_HA, "evicted/aborted by %s@%s during recovery;"
 760                                "invalidating and reconnecting\n",
 761                                obd2cli_tgt(imp->imp_obd),
 762                                imp->imp_connection->c_remote_uuid.uuid);
 763                         ptlrpc_connect_import(imp, NULL);
 764                         RETURN(0);
 765                 }
 766         } else {
 767                 struct obd_connect_data *ocd;
 768                 struct obd_export *exp;
 769                 int ret;
 770                 ret = req_capsule_get_size(&request->rq_pill, &RMF_CONNECT_DATA,
 771                                            RCL_SERVER);
 772                 /* server replied obd_connect_data is always bigger */
 773                 ocd = req_capsule_server_sized_get(&request->rq_pill,
 774                                                    &RMF_CONNECT_DATA, ret);
 775
 776                 spin_lock(&imp->imp_lock);
 777                 list_del(&imp->imp_conn_current->oic_item);
 778                 list_add(&imp->imp_conn_current->oic_item, &imp->imp_conn_list);
 779                 imp->imp_last_success_conn =
 780                         imp->imp_conn_current->oic_last_attempt;
 781
 782                 if (ocd == NULL) {
 783                         spin_unlock(&imp->imp_lock);
 784                         CERROR("Wrong connect data from server\n");
 785                         rc = -EPROTO;
 786                         GOTO(out, rc);
 787                 }
 788
 789                 imp->imp_connect_data = *ocd;
 790
 791                 exp = class_conn2export(&imp->imp_dlm_handle);
 792                 spin_unlock(&imp->imp_lock);
 793
 794                 /* check that server granted subset of flags we asked for. */
 795                 LASSERTF((ocd->ocd_connect_flags &
 796                           imp->imp_connect_flags_orig) ==
 797                          ocd->ocd_connect_flags, LPX64" != "LPX64,
 798                          imp->imp_connect_flags_orig, ocd->ocd_connect_flags);
 799
 800                 if (!exp) {
 801                         /* This could happen if export is cleaned during the
 802                            connect attempt */
 803                         CERROR("Missing export for %s\n",
 804                                imp->imp_obd->obd_name);
 805                         GOTO(out, rc = -ENODEV);
 806                 }
 807                 exp->exp_connect_flags = ocd->ocd_connect_flags;
 808                 imp->imp_obd->obd_self_export->exp_connect_flags =
 809                                                         ocd->ocd_connect_flags;
 810                 class_export_put(exp);
 811
 812                 obd_import_event(imp->imp_obd, imp, IMP_EVENT_OCD);
 813
 814                 if (!ocd->ocd_ibits_known &&
 815                     ocd->ocd_connect_flags & OBD_CONNECT_IBITS)
 816                         CERROR("Inodebits aware server returned zero compatible"
 817                                " bits?\n");
 818
 819                 if ((ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 820                     (ocd->ocd_version > LUSTRE_VERSION_CODE +
 821                                         LUSTRE_VERSION_OFFSET_WARN ||
 822                      ocd->ocd_version < LUSTRE_VERSION_CODE -
 823                                         LUSTRE_VERSION_OFFSET_WARN)) {
 824                         /* Sigh, some compilers do not like #ifdef in the middle
 825                            of macro arguments */
 826 #ifdef __KERNEL__
 827                         const char *older =
 828                                 "older. Consider upgrading this client";
 829 #else
 830                         const char *older =
 831                                 "older. Consider recompiling this application";
 832 #endif
 833                         const char *newer = "newer than client version";
 834
 835                         LCONSOLE_WARN("Server %s version (%d.%d.%d.%d) "
 836                                       "is much %s (%s)\n",
 837                                       obd2cli_tgt(imp->imp_obd),
 838                                       OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 839                                       OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 840                                       OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 841                                       OBD_OCD_VERSION_FIX(ocd->ocd_version),
 842                                       ocd->ocd_version > LUSTRE_VERSION_CODE ?
 843                                       newer : older, LUSTRE_VERSION_STRING);
 844                 }
 845
 846                 if (ocd->ocd_connect_flags & OBD_CONNECT_CKSUM) {
 847                         /* We sent to the server ocd_cksum_types with bits set
 848                          * for algorithms we understand. The server masked off
 849                          * the checksum types it doesn't support */
 850                         if ((ocd->ocd_cksum_types & OBD_CKSUM_ALL) == 0) {
 851                                 LCONSOLE_WARN("The negotiation of the checksum "
 852                                               "alogrithm to use with server %s "
 853                                               "failed (%x/%x), disabling "
 854                                               "checksums\n",
 855                                               obd2cli_tgt(imp->imp_obd),
 856                                               ocd->ocd_cksum_types,
 857                                               OBD_CKSUM_ALL);
 858                                 cli->cl_checksum = 0;
 859                                 cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 860                                 cli->cl_cksum_type = OBD_CKSUM_CRC32;
 861                         } else {
 862                                 cli->cl_supp_cksum_types = ocd->ocd_cksum_types;
 863
 864                                 if (ocd->ocd_cksum_types & OSC_DEFAULT_CKSUM)
 865                                         cli->cl_cksum_type = OSC_DEFAULT_CKSUM;
 866                                 else if (ocd->ocd_cksum_types & OBD_CKSUM_ADLER)
 867                                         cli->cl_cksum_type = OBD_CKSUM_ADLER;
 868                                 else
 869                                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 870                         }
 871                 } else {
 872                         /* The server does not support OBD_CONNECT_CKSUM.
 873                          * Enforce CRC32 for backward compatibility*/
 874                         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 875                         cli->cl_cksum_type = OBD_CKSUM_CRC32;
 876                 }
 877
 878                 if (ocd->ocd_connect_flags & OBD_CONNECT_BRW_SIZE) {
 879                         cli->cl_max_pages_per_rpc =
 880                                 ocd->ocd_brw_size >> CFS_PAGE_SHIFT;
 881                 }
 882
 883                 imp->imp_obd->obd_namespace->ns_connect_flags =
 884                                                         ocd->ocd_connect_flags;
 885                 imp->imp_obd->obd_namespace->ns_orig_connect_flags =
 886                                                         ocd->ocd_connect_flags;
 887
 888                 LASSERT((cli->cl_max_pages_per_rpc <= PTLRPC_MAX_BRW_PAGES) &&
 889                         (cli->cl_max_pages_per_rpc > 0));
 890         }
 891
 892 out:
 893         if (rc != 0) {
 894                 IMPORT_SET_STATE(imp, LUSTRE_IMP_DISCON);
 895                 spin_lock(&imp->imp_lock);
 896                 if (aa->pcaa_initial_connect && !imp->imp_initial_recov &&
 897                     (request->rq_import_generation == imp->imp_generation))
 898                         ptlrpc_deactivate_and_unlock_import(imp);
 899                 else
 900                         spin_unlock(&imp->imp_lock);
 901
 902                 if ((imp->imp_recon_bk && imp->imp_last_recon) ||
 903                     (rc == -EACCES)) {
 904                         /*
 905                          * Give up trying to reconnect
 906                          * EACCES means client has no permission for connection
 907                          */
 908                         imp->imp_obd->obd_no_recov = 1;
 909                         ptlrpc_deactivate_import(imp);
 910                 }
 911
 912                 if (rc == -EPROTO) {
 913                         struct obd_connect_data *ocd;
 914
 915                         /* reply message might not be ready */
 916                         if (request->rq_repmsg == NULL)
 917                                 RETURN(-EPROTO);
 918
 919                         ocd = req_capsule_server_get(&request->rq_pill,
 920                                                      &RMF_CONNECT_DATA);
 921                         if (ocd &&
 922                             (ocd->ocd_connect_flags & OBD_CONNECT_VERSION) &&
 923                             (ocd->ocd_version != LUSTRE_VERSION_CODE)) {
 924                            /* Actually servers are only supposed to refuse
 925                               connection from liblustre clients, so we should
 926                               never see this from VFS context */
 927                                 LCONSOLE_ERROR_MSG(0x16a, "Server %s version "
 928                                         "(%d.%d.%d.%d)"
 929                                         " refused connection from this client "
 930                                         "with an incompatible version (%s).  "
 931                                         "Client must be recompiled\n",
 932                                         obd2cli_tgt(imp->imp_obd),
 933                                         OBD_OCD_VERSION_MAJOR(ocd->ocd_version),
 934                                         OBD_OCD_VERSION_MINOR(ocd->ocd_version),
 935                                         OBD_OCD_VERSION_PATCH(ocd->ocd_version),
 936                                         OBD_OCD_VERSION_FIX(ocd->ocd_version),
 937                                         LUSTRE_VERSION_STRING);
 938                                 ptlrpc_deactivate_import(imp);
 939                                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CLOSED);
 940                         }
 941                         RETURN(-EPROTO);
 942                 }
 943
 944                 ptlrpc_maybe_ping_import_soon(imp);
 945
 946                 CDEBUG(D_HA, "recovery of %s on %s failed (%d)\n",
 947                        obd2cli_tgt(imp->imp_obd),
 948                        (char *)imp->imp_connection->c_remote_uuid.uuid, rc);
 949         }
 950
 951         spin_lock(&imp->imp_lock);
 952         imp->imp_last_recon = 0;
 953         spin_unlock(&imp->imp_lock);
 954
 955         cfs_waitq_signal(&imp->imp_recovery_waitq);
 956         RETURN(rc);
 957 }
 958
 959 static int completed_replay_interpret(struct ptlrpc_request *req,
 960                                     void * data, int rc)
 961 {
 962         ENTRY;
 963         atomic_dec(&req->rq_import->imp_replay_inflight);
 964         if (req->rq_status == 0) {
 965                 ptlrpc_import_recovery_state_machine(req->rq_import);
 966         } else {
 967                 CDEBUG(D_HA, "%s: LAST_REPLAY message error: %d, "
 968                        "reconnecting\n",
 969                        req->rq_import->imp_obd->obd_name, req->rq_status);
 970                 ptlrpc_connect_import(req->rq_import, NULL);
 971         }
 972
 973         RETURN(0);
 974 }
 975
 976 static int signal_completed_replay(struct obd_import *imp)
 977 {
 978         struct ptlrpc_request *req;
 979         ENTRY;
 980
 981         LASSERT(atomic_read(&imp->imp_replay_inflight) == 0);
 982         atomic_inc(&imp->imp_replay_inflight);
 983
 984         req = ptlrpc_request_alloc_pack(imp, &RQF_OBD_PING, LUSTRE_OBD_VERSION,
 985                                         OBD_PING);
 986         if (req == NULL) {
 987                 atomic_dec(&imp->imp_replay_inflight);
 988                 RETURN(-ENOMEM);
 989         }
 990
 991         ptlrpc_request_set_replen(req);
 992         req->rq_send_state = LUSTRE_IMP_REPLAY_WAIT;
 993         lustre_msg_add_flags(req->rq_reqmsg,
 994                              MSG_LOCK_REPLAY_DONE | MSG_REQ_REPLAY_DONE);
 995         req->rq_timeout *= 3;
 996         req->rq_interpret_reply = completed_replay_interpret;
 997
 998         ptlrpcd_add_req(req);
 999         RETURN(0);
1000 }
1001
1002 #ifdef __KERNEL__
1003 static int ptlrpc_invalidate_import_thread(void *data)
1004 {
1005         struct obd_import *imp = data;
1006
1007         ENTRY;
1008
1009         ptlrpc_daemonize("ll_imp_inval");
1010
1011         CDEBUG(D_HA, "thread invalidate import %s to %s@%s\n",
1012                imp->imp_obd->obd_name, obd2cli_tgt(imp->imp_obd),
1013                imp->imp_connection->c_remote_uuid.uuid);
1014
1015         ptlrpc_invalidate_import(imp);
1016
1017         if (obd_dump_on_eviction) {
1018                 CERROR("dump the log upon eviction\n");
1019                 libcfs_debug_dumplog();
1020         }
1021
1022         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1023         ptlrpc_import_recovery_state_machine(imp);
1024
1025         RETURN(0);
1026 }
1027 #endif
1028
1029 int ptlrpc_import_recovery_state_machine(struct obd_import *imp)
1030 {
1031         int rc = 0;
1032         int inflight;
1033         char *target_start;
1034         int target_len;
1035
1036         ENTRY;
1037         if (imp->imp_state == LUSTRE_IMP_EVICTED) {
1038                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1039                           &target_start, &target_len);
1040                 /* Don't care about MGC eviction */
1041                 if (strcmp(imp->imp_obd->obd_type->typ_name,
1042                            LUSTRE_MGC_NAME) != 0) {
1043                         LCONSOLE_ERROR_MSG(0x167, "This client was evicted by "
1044                                            "%.*s; in progress operations using "
1045                                            "this service will fail.\n",
1046                                            target_len, target_start);
1047                 }
1048                 CDEBUG(D_HA, "evicted from %s@%s; invalidating\n",
1049                        obd2cli_tgt(imp->imp_obd),
1050                        imp->imp_connection->c_remote_uuid.uuid);
1051
1052 #ifdef __KERNEL__
1053                 rc = cfs_kernel_thread(ptlrpc_invalidate_import_thread, imp,
1054                                        CLONE_VM | CLONE_FILES);
1055                 if (rc < 0)
1056                         CERROR("error starting invalidate thread: %d\n", rc);
1057                 else
1058                         rc = 0;
1059                 RETURN(rc);
1060 #else
1061                 ptlrpc_invalidate_import(imp);
1062
1063                 IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1064 #endif
1065         }
1066
1067         if (imp->imp_state == LUSTRE_IMP_REPLAY) {
1068                 CDEBUG(D_HA, "replay requested by %s\n",
1069                        obd2cli_tgt(imp->imp_obd));
1070                 rc = ptlrpc_replay_next(imp, &inflight);
1071                 if (inflight == 0 &&
1072                     atomic_read(&imp->imp_replay_inflight) == 0) {
1073                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_LOCKS);
1074                         rc = ldlm_replay_locks(imp);
1075                         if (rc)
1076                                 GOTO(out, rc);
1077                 }
1078                 rc = 0;
1079         }
1080
1081         if (imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS) {
1082                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1083                         IMPORT_SET_STATE(imp, LUSTRE_IMP_REPLAY_WAIT);
1084                         rc = signal_completed_replay(imp);
1085                         if (rc)
1086                                 GOTO(out, rc);
1087                 }
1088
1089         }
1090
1091         if (imp->imp_state == LUSTRE_IMP_REPLAY_WAIT) {
1092                 if (atomic_read(&imp->imp_replay_inflight) == 0) {
1093                         IMPORT_SET_STATE(imp, LUSTRE_IMP_RECOVER);
1094                 }
1095         }
1096
1097         if (imp->imp_state == LUSTRE_IMP_RECOVER) {
1098                 CDEBUG(D_HA, "reconnected to %s@%s\n",
1099                        obd2cli_tgt(imp->imp_obd),
1100                        imp->imp_connection->c_remote_uuid.uuid);
1101
1102                 rc = ptlrpc_resend(imp);
1103                 if (rc)
1104                         GOTO(out, rc);
1105                 IMPORT_SET_STATE(imp, LUSTRE_IMP_FULL);
1106                 ptlrpc_activate_import(imp);
1107
1108                 deuuidify(obd2cli_tgt(imp->imp_obd), NULL,
1109                           &target_start, &target_len);
1110                 LCONSOLE_INFO("%s: Connection restored to service %.*s "
1111                               "using nid %s.\n", imp->imp_obd->obd_name,
1112                               target_len, target_start,
1113                               libcfs_nid2str(imp->imp_connection->c_peer.nid));
1114         }
1115
1116         if (imp->imp_state == LUSTRE_IMP_FULL) {
1117                 cfs_waitq_signal(&imp->imp_recovery_waitq);
1118                 ptlrpc_wake_delayed(imp);
1119         }
1120
1121 out:
1122         RETURN(rc);
1123 }
1124
1125 static int back_to_sleep(void *unused)
1126 {
1127         return 0;
1128 }
1129
1130 int ptlrpc_disconnect_import(struct obd_import *imp, int noclose)
1131 {
1132         struct ptlrpc_request *req;
1133         int rq_opc, rc = 0;
1134         int nowait = imp->imp_obd->obd_force;
1135         ENTRY;
1136
1137         if (nowait)
1138                 GOTO(set_state, rc);
1139
1140         switch (imp->imp_connect_op) {
1141         case OST_CONNECT: rq_opc = OST_DISCONNECT; break;
1142         case MDS_CONNECT: rq_opc = MDS_DISCONNECT; break;
1143         case MGS_CONNECT: rq_opc = MGS_DISCONNECT; break;
1144         default:
1145                 CERROR("don't know how to disconnect from %s (connect_op %d)\n",
1146                        obd2cli_tgt(imp->imp_obd), imp->imp_connect_op);
1147                 RETURN(-EINVAL);
1148         }
1149
1150         if (ptlrpc_import_in_recovery(imp)) {
1151                 struct l_wait_info lwi;
1152                 cfs_duration_t timeout;
1153                 if (imp->imp_server_timeout)
1154                         timeout = cfs_time_seconds(obd_timeout / 2);
1155                 else
1156                         timeout = cfs_time_seconds(obd_timeout);
1157
1158                 timeout = MAX(timeout * HZ, 1);
1159
1160                 lwi = LWI_TIMEOUT_INTR(cfs_timeout_cap(timeout),
1161                                        back_to_sleep, LWI_ON_SIGNAL_NOOP, NULL);
1162                 rc = l_wait_event(imp->imp_recovery_waitq,
1163                                   !ptlrpc_import_in_recovery(imp), &lwi);
1164
1165         }
1166
1167         spin_lock(&imp->imp_lock);
1168         if (imp->imp_state != LUSTRE_IMP_FULL)
1169                 GOTO(out, 0);
1170
1171         spin_unlock(&imp->imp_lock);
1172
1173         req = ptlrpc_request_alloc_pack(imp, &RQF_MDS_DISCONNECT,
1174                                         LUSTRE_OBD_VERSION, rq_opc);
1175         if (req) {
1176                 /* We are disconnecting, do not retry a failed DISCONNECT rpc if
1177                  * it fails.  We can get through the above with a down server
1178                  * if the client doesn't know the server is gone yet. */
1179                 req->rq_no_resend = 1;
1180 #ifdef CRAY_XT3
1181                 req->rq_timeout = obd_timeout / 3;
1182 #else
1183                 req->rq_timeout = 5;
1184 #endif
1185                 IMPORT_SET_STATE(imp, LUSTRE_IMP_CONNECTING);
1186                 req->rq_send_state =  LUSTRE_IMP_CONNECTING;
1187                 ptlrpc_request_set_replen(req);
1188                 rc = ptlrpc_queue_wait(req);
1189                 ptlrpc_req_finished(req);
1190         }
1191
1192 set_state:
1193         spin_lock(&imp->imp_lock);
1194 out:
1195         if (noclose)
1196                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_DISCON);
1197         else
1198                 IMPORT_SET_STATE_NOLOCK(imp, LUSTRE_IMP_CLOSED);
1199         memset(&imp->imp_remote_handle, 0, sizeof(imp->imp_remote_handle));
1200         imp->imp_conn_cnt = 0;
1201         imp->imp_last_recon = 0;
1202         spin_unlock(&imp->imp_lock);
1203
1204         RETURN(rc);
1205 }
1206