lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LDLM
  38
  39 #ifdef __KERNEL__
  40 # include <libcfs/libcfs.h>
  41 #else
  42 # include <liblustre.h>
  43 #endif
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_net.h>
  48 #include <lustre_sec.h>
  49 #include "ldlm_internal.h"
  50
  51 /* @priority: if non-zero, move the selected to the list head
  52  * @create: if zero, only search in existed connections
  53  */
  54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  55                            int priority, int create)
  56 {
  57         struct ptlrpc_connection *ptlrpc_conn;
  58         struct obd_import_conn *imp_conn = NULL, *item;
  59         int rc = 0;
  60         ENTRY;
  61
  62         if (!create && !priority) {
  63                 CDEBUG(D_HA, "Nothing to do\n");
  64                 RETURN(-EINVAL);
  65         }
  66
  67         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  68         if (!ptlrpc_conn) {
  69                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  70                 RETURN (-ENOENT);
  71         }
  72
  73         if (create) {
  74                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  75                 if (!imp_conn) {
  76                         GOTO(out_put, rc = -ENOMEM);
  77                 }
  78         }
  79
  80         cfs_spin_lock(&imp->imp_lock);
  81         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  82                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  83                         if (priority) {
  84                                 cfs_list_del(&item->oic_item);
  85                                 cfs_list_add(&item->oic_item,
  86                                              &imp->imp_conn_list);
  87                                 item->oic_last_attempt = 0;
  88                         }
  89                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  90                                imp, imp->imp_obd->obd_name, uuid->uuid,
  91                                (priority ? ", moved to head" : ""));
  92                         cfs_spin_unlock(&imp->imp_lock);
  93                         GOTO(out_free, rc = 0);
  94                 }
  95         }
  96         /* not found */
  97         if (create) {
  98                 imp_conn->oic_conn = ptlrpc_conn;
  99                 imp_conn->oic_uuid = *uuid;
 100                 imp_conn->oic_last_attempt = 0;
 101                 if (priority)
 102                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 103                 else
 104                         cfs_list_add_tail(&imp_conn->oic_item,
 105                                           &imp->imp_conn_list);
 106                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 107                        imp, imp->imp_obd->obd_name, uuid->uuid,
 108                        (priority ? "head" : "tail"));
 109         } else {
 110                 cfs_spin_unlock(&imp->imp_lock);
 111                 GOTO(out_free, rc = -ENOENT);
 112         }
 113
 114         cfs_spin_unlock(&imp->imp_lock);
 115         RETURN(0);
 116 out_free:
 117         if (imp_conn)
 118                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 119 out_put:
 120         ptlrpc_connection_put(ptlrpc_conn);
 121         RETURN(rc);
 122 }
 123
 124 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 125 {
 126         return import_set_conn(imp, uuid, 1, 0);
 127 }
 128
 129 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 130                            int priority)
 131 {
 132         return import_set_conn(imp, uuid, priority, 1);
 133 }
 134
 135 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 136 {
 137         struct obd_import_conn *imp_conn;
 138         struct obd_export *dlmexp;
 139         int rc = -ENOENT;
 140         ENTRY;
 141
 142         cfs_spin_lock(&imp->imp_lock);
 143         if (cfs_list_empty(&imp->imp_conn_list)) {
 144                 LASSERT(!imp->imp_connection);
 145                 GOTO(out, rc);
 146         }
 147
 148         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 149                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 150                         continue;
 151                 LASSERT(imp_conn->oic_conn);
 152
 153                 /* is current conn? */
 154                 if (imp_conn == imp->imp_conn_current) {
 155                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 156
 157                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 158                             imp->imp_state != LUSTRE_IMP_DISCON) {
 159                                 CERROR("can't remove current connection\n");
 160                                 GOTO(out, rc = -EBUSY);
 161                         }
 162
 163                         ptlrpc_connection_put(imp->imp_connection);
 164                         imp->imp_connection = NULL;
 165
 166                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 167                         if (dlmexp && dlmexp->exp_connection) {
 168                                 LASSERT(dlmexp->exp_connection ==
 169                                         imp_conn->oic_conn);
 170                                 ptlrpc_connection_put(dlmexp->exp_connection);
 171                                 dlmexp->exp_connection = NULL;
 172                         }
 173                 }
 174
 175                 cfs_list_del(&imp_conn->oic_item);
 176                 ptlrpc_connection_put(imp_conn->oic_conn);
 177                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 178                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 179                        imp, imp->imp_obd->obd_name, uuid->uuid);
 180                 rc = 0;
 181                 break;
 182         }
 183 out:
 184         cfs_spin_unlock(&imp->imp_lock);
 185         if (rc == -ENOENT)
 186                 CERROR("connection %s not found\n", uuid->uuid);
 187         RETURN(rc);
 188 }
 189
 190 /**
 191  * Find conn uuid by peer nid. @peer is a server nid. This function is used
 192  * to find a conn uuid of @imp which can reach @peer.
 193  */
 194 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 195                             struct obd_uuid *uuid)
 196 {
 197         struct obd_import_conn *conn;
 198         int rc = -ENOENT;
 199         ENTRY;
 200
 201         cfs_spin_lock(&imp->imp_lock);
 202         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 203                 /* check if conn uuid does have this peer nid */
 204                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 205                         *uuid = conn->oic_uuid;
 206                         rc = 0;
 207                         break;
 208                 }
 209         }
 210         cfs_spin_unlock(&imp->imp_lock);
 211         RETURN(rc);
 212 }
 213 EXPORT_SYMBOL(client_import_find_conn);
 214
 215 void client_destroy_import(struct obd_import *imp)
 216 {
 217         /* drop security policy instance after all rpc finished/aborted
 218          * to let all busy contexts be released. */
 219         class_import_get(imp);
 220         class_destroy_import(imp);
 221         sptlrpc_import_sec_put(imp);
 222         class_import_put(imp);
 223 }
 224 EXPORT_SYMBOL(client_destroy_import);
 225
 226 /**
 227  * check whether the osc is on MDT or not
 228  * In the config log,
 229  * osc on MDT
 230  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 231  * osc on client
 232  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 233  *
 234  **/
 235 static int osc_on_mdt(char *obdname)
 236 {
 237         char *ptr;
 238
 239         ptr = strrchr(obdname, '-');
 240         if (ptr == NULL)
 241                 return 0;
 242
 243         if (strncmp(ptr + 1, "MDT", 3) == 0)
 244                 return 1;
 245
 246         return 0;
 247 }
 248
 249 /* configure an RPC client OBD device
 250  *
 251  * lcfg parameters:
 252  * 1 - client UUID
 253  * 2 - server UUID
 254  * 3 - inactive-on-startup
 255  */
 256 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 257 {
 258         struct client_obd *cli = &obddev->u.cli;
 259         struct obd_import *imp;
 260         struct obd_uuid server_uuid;
 261         int rq_portal, rp_portal, connect_op;
 262         char *name = obddev->obd_type->typ_name;
 263         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 264         int rc;
 265         ENTRY;
 266
 267         /* In a more perfect world, we would hang a ptlrpc_client off of
 268          * obd_type and just use the values from there. */
 269         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 270                 rq_portal = OST_REQUEST_PORTAL;
 271                 rp_portal = OSC_REPLY_PORTAL;
 272                 connect_op = OST_CONNECT;
 273                 cli->cl_sp_me = LUSTRE_SP_CLI;
 274                 cli->cl_sp_to = LUSTRE_SP_OST;
 275                 ns_type = LDLM_NS_TYPE_OSC;
 276
 277         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 278                 rq_portal = MDS_REQUEST_PORTAL;
 279                 rp_portal = MDC_REPLY_PORTAL;
 280                 connect_op = MDS_CONNECT;
 281                 cli->cl_sp_me = LUSTRE_SP_CLI;
 282                 cli->cl_sp_to = LUSTRE_SP_MDT;
 283                 ns_type = LDLM_NS_TYPE_MDC;
 284
 285         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 286                 rq_portal = MGS_REQUEST_PORTAL;
 287                 rp_portal = MGC_REPLY_PORTAL;
 288                 connect_op = MGS_CONNECT;
 289                 cli->cl_sp_me = LUSTRE_SP_MGC;
 290                 cli->cl_sp_to = LUSTRE_SP_MGS;
 291                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 292                 ns_type = LDLM_NS_TYPE_MGC;
 293
 294         } else {
 295                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 296                        name);
 297                 RETURN(-EINVAL);
 298         }
 299
 300         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 301                 CERROR("requires a TARGET UUID\n");
 302                 RETURN(-EINVAL);
 303         }
 304
 305         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 306                 CERROR("client UUID must be less than 38 characters\n");
 307                 RETURN(-EINVAL);
 308         }
 309
 310         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 311                 CERROR("setup requires a SERVER UUID\n");
 312                 RETURN(-EINVAL);
 313         }
 314
 315         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 316                 CERROR("target UUID must be less than 38 characters\n");
 317                 RETURN(-EINVAL);
 318         }
 319
 320         cfs_init_rwsem(&cli->cl_sem);
 321         cfs_sema_init(&cli->cl_mgc_sem, 1);
 322         cli->cl_conn_count = 0;
 323         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 324                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 325                      sizeof(server_uuid)));
 326
 327         cli->cl_dirty = 0;
 328         cli->cl_avail_grant = 0;
 329         /* FIXME: should limit this for the sum of all cl_dirty_max */
 330         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 331         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
 332                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
 333         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 334         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 335         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 336         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 337         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 338         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 339         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
 340         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
 341         cli->cl_r_in_flight = 0;
 342         cli->cl_w_in_flight = 0;
 343
 344         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 345         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 346         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 347         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 348         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 349         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 350         cfs_waitq_init(&cli->cl_destroy_waitq);
 351         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 352 #ifdef ENABLE_CHECKSUM
 353         /* Turn on checksumming by default. */
 354         cli->cl_checksum = 1;
 355         /*
 356          * The supported checksum types will be worked out at connect time
 357          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 358          * through procfs.
 359          */
 360         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 361 #endif
 362         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 363
 364         /* This value may be changed at connect time in
 365            ptlrpc_connect_interpret. */
 366         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 367                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 368
 369         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 370                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 371         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 372                 cli->cl_max_rpcs_in_flight = 2;
 373         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 374                 cli->cl_max_rpcs_in_flight = 3;
 375         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 376                 cli->cl_max_rpcs_in_flight = 4;
 377         } else {
 378                 if (osc_on_mdt(obddev->obd_name))
 379                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
 380                 else
 381                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 382         }
 383         rc = ldlm_get_ref();
 384         if (rc) {
 385                 CERROR("ldlm_get_ref failed: %d\n", rc);
 386                 GOTO(err, rc);
 387         }
 388
 389         ptlrpc_init_client(rq_portal, rp_portal, name,
 390                            &obddev->obd_ldlm_client);
 391
 392         imp = class_new_import(obddev);
 393         if (imp == NULL)
 394                 GOTO(err_ldlm, rc = -ENOENT);
 395         imp->imp_client = &obddev->obd_ldlm_client;
 396         imp->imp_connect_op = connect_op;
 397         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 398         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 399                LUSTRE_CFG_BUFLEN(lcfg, 1));
 400         class_import_put(imp);
 401
 402         rc = client_import_add_conn(imp, &server_uuid, 1);
 403         if (rc) {
 404                 CERROR("can't add initial connection\n");
 405                 GOTO(err_import, rc);
 406         }
 407
 408         cli->cl_import = imp;
 409         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 410         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 411         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 412
 413         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 414                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 415                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 416                                name, obddev->obd_name,
 417                                cli->cl_target_uuid.uuid);
 418                         cfs_spin_lock(&imp->imp_lock);
 419                         imp->imp_deactive = 1;
 420                         cfs_spin_unlock(&imp->imp_lock);
 421                 }
 422         }
 423
 424         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 425                                                    LDLM_NAMESPACE_CLIENT,
 426                                                    LDLM_NAMESPACE_GREEDY,
 427                                                    ns_type);
 428         if (obddev->obd_namespace == NULL) {
 429                 CERROR("Unable to create client namespace - %s\n",
 430                        obddev->obd_name);
 431                 GOTO(err_import, rc = -ENOMEM);
 432         }
 433
 434         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 435
 436         RETURN(rc);
 437
 438 err_import:
 439         class_destroy_import(imp);
 440 err_ldlm:
 441         ldlm_put_ref();
 442 err:
 443         RETURN(rc);
 444
 445 }
 446
 447 int client_obd_cleanup(struct obd_device *obddev)
 448 {
 449         ENTRY;
 450
 451         ldlm_namespace_free_post(obddev->obd_namespace);
 452         obddev->obd_namespace = NULL;
 453
 454         LASSERT(obddev->u.cli.cl_import == NULL);
 455
 456         ldlm_put_ref();
 457         RETURN(0);
 458 }
 459
 460 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 461 int client_connect_import(const struct lu_env *env,
 462                           struct obd_export **exp,
 463                           struct obd_device *obd, struct obd_uuid *cluuid,
 464                           struct obd_connect_data *data, void *localdata)
 465 {
 466         struct client_obd *cli = &obd->u.cli;
 467         struct obd_import *imp = cli->cl_import;
 468         struct obd_connect_data *ocd;
 469         struct lustre_handle conn = { 0 };
 470         int rc;
 471         ENTRY;
 472
 473         *exp = NULL;
 474         cfs_down_write(&cli->cl_sem);
 475         if (cli->cl_conn_count > 0 )
 476                 GOTO(out_sem, rc = -EALREADY);
 477
 478         rc = class_connect(&conn, obd, cluuid);
 479         if (rc)
 480                 GOTO(out_sem, rc);
 481
 482         cli->cl_conn_count++;
 483         *exp = class_conn2export(&conn);
 484
 485         LASSERT(obd->obd_namespace);
 486
 487         imp->imp_dlm_handle = conn;
 488         rc = ptlrpc_init_import(imp);
 489         if (rc != 0)
 490                 GOTO(out_ldlm, rc);
 491
 492         ocd = &imp->imp_connect_data;
 493         if (data) {
 494                 *ocd = *data;
 495                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 496         }
 497
 498         rc = ptlrpc_connect_import(imp);
 499         if (rc != 0) {
 500                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 501                 GOTO(out_ldlm, rc);
 502         }
 503         LASSERT((*exp)->exp_connection);
 504
 505         if (data) {
 506                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 507                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 508                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 509                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 510         }
 511
 512         ptlrpc_pinger_add_import(imp);
 513
 514         EXIT;
 515
 516         if (rc) {
 517 out_ldlm:
 518                 cli->cl_conn_count--;
 519                 class_disconnect(*exp);
 520                 *exp = NULL;
 521         }
 522 out_sem:
 523         cfs_up_write(&cli->cl_sem);
 524
 525         return rc;
 526 }
 527
 528 int client_disconnect_export(struct obd_export *exp)
 529 {
 530         struct obd_device *obd = class_exp2obd(exp);
 531         struct client_obd *cli;
 532         struct obd_import *imp;
 533         int rc = 0, err;
 534         ENTRY;
 535
 536         if (!obd) {
 537                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 538                        exp, exp ? exp->exp_handle.h_cookie : -1);
 539                 RETURN(-EINVAL);
 540         }
 541
 542         cli = &obd->u.cli;
 543         imp = cli->cl_import;
 544
 545         cfs_down_write(&cli->cl_sem);
 546         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 547                cli->cl_conn_count);
 548
 549         if (!cli->cl_conn_count) {
 550                 CERROR("disconnecting disconnected device (%s)\n",
 551                        obd->obd_name);
 552                 GOTO(out_disconnect, rc = -EINVAL);
 553         }
 554
 555         cli->cl_conn_count--;
 556         if (cli->cl_conn_count)
 557                 GOTO(out_disconnect, rc = 0);
 558
 559         /* Mark import deactivated now, so we don't try to reconnect if any
 560          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 561          * fully deactivate the import, or that would drop all requests. */
 562         cfs_spin_lock(&imp->imp_lock);
 563         imp->imp_deactive = 1;
 564         cfs_spin_unlock(&imp->imp_lock);
 565
 566         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 567          * delete it regardless.  (It's safe to delete an import that was
 568          * never added.) */
 569         (void)ptlrpc_pinger_del_import(imp);
 570
 571         if (obd->obd_namespace != NULL) {
 572                 /* obd_force == local only */
 573                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 574                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 575                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 576         }
 577
 578         /*
 579          * there's no need to hold sem during disconnecting an import,
 580          * and actually it may cause deadlock in gss.
 581          */
 582         cfs_up_write(&cli->cl_sem);
 583         rc = ptlrpc_disconnect_import(imp, 0);
 584         cfs_down_write(&cli->cl_sem);
 585
 586         ptlrpc_invalidate_import(imp);
 587
 588         EXIT;
 589
 590  out_disconnect:
 591         /* use server style - class_disconnect should be always called for
 592          * o_disconnect */
 593         err = class_disconnect(exp);
 594         if (!rc && err)
 595                 rc = err;
 596
 597         cfs_up_write(&cli->cl_sem);
 598
 599         RETURN(rc);
 600 }
 601
 602 #ifdef HAVE_SERVER_SUPPORT
 603 int server_disconnect_export(struct obd_export *exp)
 604 {
 605         int rc;
 606         ENTRY;
 607
 608         /* Disconnect early so that clients can't keep using export */
 609         rc = class_disconnect(exp);
 610         /* close import for avoid sending any requests */
 611         if (exp->exp_imp_reverse)
 612                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 613
 614         if (exp->exp_obd->obd_namespace != NULL)
 615                 ldlm_cancel_locks_for_export(exp);
 616
 617         /* complete all outstanding replies */
 618         cfs_spin_lock(&exp->exp_lock);
 619         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 620                 struct ptlrpc_reply_state *rs =
 621                         cfs_list_entry(exp->exp_outstanding_replies.next,
 622                                        struct ptlrpc_reply_state, rs_exp_list);
 623                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 624
 625                 cfs_spin_lock(&svcpt->scp_rep_lock);
 626
 627                 cfs_list_del_init(&rs->rs_exp_list);
 628                 cfs_spin_lock(&rs->rs_lock);
 629                 ptlrpc_schedule_difficult_reply(rs);
 630                 cfs_spin_unlock(&rs->rs_lock);
 631
 632                 cfs_spin_unlock(&svcpt->scp_rep_lock);
 633         }
 634         cfs_spin_unlock(&exp->exp_lock);
 635
 636         RETURN(rc);
 637 }
 638
 639 /* --------------------------------------------------------------------------
 640  * from old lib/target.c
 641  * -------------------------------------------------------------------------- */
 642
 643 static int target_handle_reconnect(struct lustre_handle *conn,
 644                                    struct obd_export *exp,
 645                                    struct obd_uuid *cluuid)
 646 {
 647         ENTRY;
 648
 649         if (exp->exp_connection && exp->exp_imp_reverse) {
 650                 struct lustre_handle *hdl;
 651                 struct obd_device *target;
 652
 653                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 654                 target = exp->exp_obd;
 655
 656                 /* Might be a re-connect after a partition. */
 657                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 658                         if (target->obd_recovering) {
 659                                 int timeout = cfs_duration_sec(cfs_time_sub(
 660                                         cfs_timer_deadline(
 661                                         &target->obd_recovery_timer),
 662                                         cfs_time_current()));
 663
 664                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 665                                         "ing, waiting for %d clients in recov"
 666                                         "ery for %d:%.02d\n", target->obd_name,
 667                                         obd_uuid2str(&exp->exp_client_uuid),
 668                                         obd_export_nid2str(exp),
 669                                         target->obd_max_recoverable_clients,
 670                                         timeout / 60, timeout % 60);
 671                         } else {
 672                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 673                                         "reconnecting\n", target->obd_name,
 674                                         obd_uuid2str(&exp->exp_client_uuid),
 675                                         obd_export_nid2str(exp));
 676                         }
 677
 678                         conn->cookie = exp->exp_handle.h_cookie;
 679                         /* target_handle_connect() treats EALREADY and
 680                          * -EALREADY differently.  EALREADY means we are
 681                          * doing a valid reconnect from the same client. */
 682                         RETURN(EALREADY);
 683                 } else {
 684                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 685                                       "with handle "LPX64". Rejecting client "
 686                                       "with the same UUID trying to reconnect "
 687                                       "with handle "LPX64"\n", target->obd_name,
 688                                       obd_uuid2str(&exp->exp_client_uuid),
 689                                       obd_export_nid2str(exp),
 690                                       hdl->cookie, conn->cookie);
 691                         memset(conn, 0, sizeof *conn);
 692                         /* target_handle_connect() treats EALREADY and
 693                          * -EALREADY differently.  -EALREADY is an error
 694                          * (same UUID, different handle). */
 695                         RETURN(-EALREADY);
 696                 }
 697         }
 698
 699         conn->cookie = exp->exp_handle.h_cookie;
 700         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 701                cluuid->uuid, exp, conn->cookie);
 702         RETURN(0);
 703 }
 704
 705 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 706                           int error)
 707 {
 708         struct obd_export *exp = cb_data;
 709
 710         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 711                obd->obd_name, exp->exp_client_uuid.uuid);
 712
 713         cfs_spin_lock(&exp->exp_lock);
 714         exp->exp_need_sync = 0;
 715         cfs_spin_unlock(&exp->exp_lock);
 716         class_export_cb_put(exp);
 717 }
 718 EXPORT_SYMBOL(target_client_add_cb);
 719
 720 #ifdef __KERNEL__
 721 static void
 722 check_and_start_recovery_timer(struct obd_device *obd,
 723                                struct ptlrpc_request *req, int new_client);
 724 #else
 725 static inline void
 726 check_and_start_recovery_timer(struct obd_device *obd,
 727                                struct ptlrpc_request *req, int new_client)
 728 {
 729 }
 730 #endif
 731
 732 int target_handle_connect(struct ptlrpc_request *req)
 733 {
 734         struct obd_device *target = NULL, *targref = NULL;
 735         struct obd_export *export = NULL;
 736         struct obd_import *revimp;
 737         struct obd_import *tmp_imp = NULL;
 738         struct lustre_handle conn;
 739         struct lustre_handle *tmp;
 740         struct obd_uuid tgtuuid;
 741         struct obd_uuid cluuid;
 742         struct obd_uuid remote_uuid;
 743         char *str;
 744         int rc = 0;
 745         char *target_start;
 746         int target_len;
 747         int mds_conn = 0;
 748         struct obd_connect_data *data, *tmpdata;
 749         int size, tmpsize;
 750         lnet_nid_t *client_nid = NULL;
 751         ENTRY;
 752
 753         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 754
 755         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 756         if (str == NULL) {
 757                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 758                 GOTO(out, rc = -EINVAL);
 759         }
 760
 761         obd_str2uuid(&tgtuuid, str);
 762         target = class_uuid2obd(&tgtuuid);
 763         if (!target)
 764                 target = class_name2obd(str);
 765
 766         if (!target) {
 767                 deuuidify(str, NULL, &target_start, &target_len);
 768                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
 769                                    "connect (no target)\n", str);
 770                 GOTO(out, rc = -ENODEV);
 771         }
 772
 773         cfs_spin_lock(&target->obd_dev_lock);
 774         if (target->obd_stopping || !target->obd_set_up) {
 775                 cfs_spin_unlock(&target->obd_dev_lock);
 776
 777                 deuuidify(str, NULL, &target_start, &target_len);
 778                 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
 779                                    "from %s (%s)\n", target_len, target_start,
 780                                    libcfs_nid2str(req->rq_peer.nid),
 781                                    (target->obd_stopping ?
 782                                    "stopping" : "not set up"));
 783                 GOTO(out, rc = -ENODEV);
 784         }
 785
 786         if (target->obd_no_conn) {
 787                 cfs_spin_unlock(&target->obd_dev_lock);
 788
 789                 LCONSOLE_WARN("%s: Temporarily refusing client connection "
 790                               "from %s\n", target->obd_name,
 791                               libcfs_nid2str(req->rq_peer.nid));
 792                 GOTO(out, rc = -EAGAIN);
 793         }
 794
 795         /* Make sure the target isn't cleaned up while we're here. Yes,
 796            there's still a race between the above check and our incref here.
 797            Really, class_uuid2obd should take the ref. */
 798         targref = class_incref(target, __FUNCTION__, cfs_current());
 799
 800         target->obd_conn_inprogress++;
 801         cfs_spin_unlock(&target->obd_dev_lock);
 802
 803         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 804         if (str == NULL) {
 805                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 806                 GOTO(out, rc = -EINVAL);
 807         }
 808
 809         obd_str2uuid(&cluuid, str);
 810
 811         /* XXX extract a nettype and format accordingly */
 812         switch (sizeof(lnet_nid_t)) {
 813                 /* NB the casts only avoid compiler warnings */
 814         case 8:
 815                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 816                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 817                 break;
 818         case 4:
 819                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 820                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 821                 break;
 822         default:
 823                 LBUG();
 824         }
 825
 826         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 827         if (tmp == NULL)
 828                 GOTO(out, rc = -EPROTO);
 829
 830         conn = *tmp;
 831
 832         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 833                                     RCL_CLIENT);
 834         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 835         if (!data)
 836                 GOTO(out, rc = -EPROTO);
 837
 838         rc = req_capsule_server_pack(&req->rq_pill);
 839         if (rc)
 840                 GOTO(out, rc);
 841
 842         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 843                 if (!data) {
 844                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 845                                   "libclient connection attempt");
 846                         GOTO(out, rc = -EPROTO);
 847                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 848                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 849                            data->ocd_version > LUSTRE_VERSION_CODE +
 850                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 851                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 852                                   "libclient connection attempt",
 853                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 854                                   "old" : "new",
 855                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 856                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 857                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 858                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 859                         data = req_capsule_server_sized_get(&req->rq_pill,
 860                                                         &RMF_CONNECT_DATA,
 861                                     offsetof(typeof(*data), ocd_version) +
 862                                              sizeof(data->ocd_version));
 863                         if (data) {
 864                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 865                                 data->ocd_version = LUSTRE_VERSION_CODE;
 866                         }
 867                         GOTO(out, rc = -EPROTO);
 868                 }
 869         }
 870
 871         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 872             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 873                 mds_conn = 1;
 874
 875         /* lctl gets a backstage, all-access pass. */
 876         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 877                 goto dont_check_exports;
 878
 879         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 880         if (!export)
 881                 goto no_export;
 882
 883         /* we've found an export in the hash */
 884         if (export->exp_connecting) { /* bug 9635, et. al. */
 885                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 886                               export->exp_obd->obd_name, export,
 887                               libcfs_nid2str(req->rq_peer.nid));
 888                 class_export_put(export);
 889                 export = NULL;
 890                 rc = -EALREADY;
 891         } else if (mds_conn && export->exp_connection) {
 892                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 893                         /* mds reconnected after failover */
 894                         LCONSOLE_WARN("%s: Received MDS connection from "
 895                             "%s, removing former export from %s\n",
 896                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 897                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 898                 else
 899                         /* new mds connection from the same nid */
 900                         LCONSOLE_WARN("%s: Received new MDS connection from "
 901                             "%s, removing former export from same NID\n",
 902                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 903                 class_fail_export(export);
 904                 class_export_put(export);
 905                 export = NULL;
 906                 rc = 0;
 907         } else if (export->exp_connection != NULL &&
 908                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 909                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 910                     MSG_CONNECT_INITIAL)) {
 911                 /* in mds failover we have static uuid but nid can be
 912                  * changed*/
 913                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 914                               "existing nid %s is already connected\n",
 915                               target->obd_name, cluuid.uuid,
 916                               libcfs_nid2str(req->rq_peer.nid),
 917                               libcfs_nid2str(
 918                                       export->exp_connection->c_peer.nid));
 919                 rc = -EALREADY;
 920                 class_export_put(export);
 921                 export = NULL;
 922         } else {
 923                 cfs_spin_lock(&export->exp_lock);
 924                 export->exp_connecting = 1;
 925                 cfs_spin_unlock(&export->exp_lock);
 926                 LASSERT(export->exp_obd == target);
 927
 928                 rc = target_handle_reconnect(&conn, export, &cluuid);
 929         }
 930
 931         /* If we found an export, we already unlocked. */
 932         if (!export) {
 933 no_export:
 934                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 935         } else if (req->rq_export == NULL &&
 936                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 937                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 938                               "still busy with %d references\n",
 939                               target->obd_name, cluuid.uuid,
 940                               libcfs_nid2str(req->rq_peer.nid),
 941                               cfs_atomic_read(&export->exp_refcount));
 942                 GOTO(out, rc = -EBUSY);
 943         } else if (req->rq_export != NULL &&
 944                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 945                 /* the current connect rpc has increased exp_rpc_count */
 946                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
 947                               "still busy with %d active RPCs\n",
 948                               target->obd_name, cluuid.uuid,
 949                               libcfs_nid2str(req->rq_peer.nid),
 950                               cfs_atomic_read(&export->exp_rpc_count) - 1);
 951                 cfs_spin_lock(&export->exp_lock);
 952                 if (req->rq_export->exp_conn_cnt <
 953                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 954                         /* try to abort active requests */
 955                         req->rq_export->exp_abort_active_req = 1;
 956                 cfs_spin_unlock(&export->exp_lock);
 957                 GOTO(out, rc = -EBUSY);
 958         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 959                 if (!strstr(cluuid.uuid, "mdt"))
 960                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 961                                       "known client %s (at %s) because it "
 962                                       "is indicating it is a new client",
 963                                       target->obd_name, cluuid.uuid,
 964                                       libcfs_nid2str(req->rq_peer.nid));
 965                 GOTO(out, rc = -EALREADY);
 966         } else {
 967                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 968         }
 969
 970         if (rc < 0) {
 971                 GOTO(out, rc);
 972         }
 973
 974         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 975                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 976               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 977               export, (long)cfs_time_current_sec(),
 978               export ? (long)export->exp_last_request_time : 0);
 979
 980         /* If this is the first time a client connects, reset the recovery
 981          * timer */
 982         if (rc == 0 && target->obd_recovering)
 983                 check_and_start_recovery_timer(target, req, export == NULL);
 984
 985         /* We want to handle EALREADY but *not* -EALREADY from
 986          * target_handle_reconnect(), return reconnection state in a flag */
 987         if (rc == EALREADY) {
 988                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 989                 rc = 0;
 990         } else {
 991                 LASSERT(rc == 0);
 992         }
 993
 994         /* Tell the client if we support replayable requests */
 995         if (target->obd_replayable)
 996                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 997         client_nid = &req->rq_peer.nid;
 998
 999         if (export == NULL) {
1000                 if (target->obd_recovering) {
1001                         cfs_time_t t;
1002                         int        c; /* connected */
1003                         int        i; /* in progress */
1004                         int        k; /* known */
1005
1006                         c = cfs_atomic_read(&target->obd_connected_clients);
1007                         i = cfs_atomic_read(&target->obd_lock_replay_clients);
1008                         k = target->obd_max_recoverable_clients;
1009                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1010                         t = cfs_time_sub(t, cfs_time_current());
1011                         t = cfs_duration_sec(t);
1012                         LCONSOLE_WARN("%s: Denying connection for new client "
1013                                       "%s (at %s), waiting for all %d known "
1014                                       "clients (%d recovered, %d in progress, "
1015                                       "and %d unseen) to recover in %d:%.02d\n",
1016                                       target->obd_name, cluuid.uuid,
1017                                       libcfs_nid2str(req->rq_peer.nid), k,
1018                                       c - i, i, k - c, (int)t / 60,
1019                                       (int)t % 60);
1020                         rc = -EBUSY;
1021                 } else {
1022 dont_check_exports:
1023                         rc = obd_connect(req->rq_svc_thread->t_env,
1024                                          &export, target, &cluuid, data,
1025                                          client_nid);
1026                         if (rc == 0)
1027                                 conn.cookie = export->exp_handle.h_cookie;
1028                 }
1029         } else {
1030                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1031                                    export, target, &cluuid, data, client_nid);
1032         }
1033         if (rc)
1034                 GOTO(out, rc);
1035
1036 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1037         /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1038          * will do the swabbing for if the client is using the same endianness.
1039          *
1040          * This fixup is version-limited, because we don't want to carry the
1041          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1042          * interop with unpatched 2.2 clients.  For newer clients, servers
1043          * will never do MNE swabbing, let the client handle that.  LU-1644 */
1044         export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1045                         !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1046 #else
1047 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1048 #endif
1049
1050         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1051         data->ocd_instance = target->u.obt.obt_instance;
1052
1053         /* Return only the parts of obd_connect_data that we understand, so the
1054          * client knows that we don't understand the rest. */
1055         if (data) {
1056                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1057                                                RCL_SERVER);
1058                 tmpdata = req_capsule_server_get(&req->rq_pill,
1059                                                  &RMF_CONNECT_DATA);
1060                 /* Don't use struct assignment here, because the client reply
1061                  * buffer may be smaller/larger than the local struct
1062                  * obd_connect_data. */
1063                 memcpy(tmpdata, data, min(tmpsize, size));
1064         }
1065
1066         /* If all else goes well, this is our RPC return code. */
1067         req->rq_status = 0;
1068
1069         lustre_msg_set_handle(req->rq_repmsg, &conn);
1070
1071         /* If the client and the server are the same node, we will already
1072          * have an export that really points to the client's DLM export,
1073          * because we have a shared handles table.
1074          *
1075          * XXX this will go away when shaver stops sending the "connect" handle
1076          * in the real "remote handle" field of the request --phik 24 Apr 2003
1077          */
1078         if (req->rq_export != NULL)
1079                 class_export_put(req->rq_export);
1080
1081         /* request takes one export refcount */
1082         req->rq_export = class_export_get(export);
1083
1084         cfs_spin_lock(&export->exp_lock);
1085         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1086                 cfs_spin_unlock(&export->exp_lock);
1087                 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1088                        "conn_cnt: %d > %d\n",
1089                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1090                        export->exp_conn_cnt,
1091                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1092
1093                 GOTO(out, rc = -EALREADY);
1094         }
1095         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1096         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1097         export->exp_abort_active_req = 0;
1098
1099         /* request from liblustre?  Don't evict it for not pinging. */
1100         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1101                 export->exp_libclient = 1;
1102                 cfs_spin_unlock(&export->exp_lock);
1103
1104                 cfs_spin_lock(&target->obd_dev_lock);
1105                 cfs_list_del_init(&export->exp_obd_chain_timed);
1106                 cfs_spin_unlock(&target->obd_dev_lock);
1107         } else {
1108                 cfs_spin_unlock(&export->exp_lock);
1109         }
1110
1111         if (export->exp_connection != NULL) {
1112                 /* Check to see if connection came from another NID */
1113                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1114                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1115                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1116                                      &export->exp_connection->c_peer.nid,
1117                                      &export->exp_nid_hash);
1118
1119                 ptlrpc_connection_put(export->exp_connection);
1120         }
1121
1122         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1123                                                        req->rq_self,
1124                                                        &remote_uuid);
1125         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1126                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1127                              &export->exp_connection->c_peer.nid,
1128                              &export->exp_nid_hash);
1129         }
1130
1131         if (target->obd_recovering && !export->exp_in_recovery) {
1132                 int has_transno;
1133                 __u64 transno = data->ocd_transno;
1134
1135                 cfs_spin_lock(&export->exp_lock);
1136                 /* possible race with class_disconnect_stale_exports,
1137                  * export may be already in the eviction process */
1138                 if (export->exp_failed) {
1139                         cfs_spin_unlock(&export->exp_lock);
1140                         GOTO(out, rc = -ENODEV);
1141                 }
1142                 export->exp_in_recovery = 1;
1143                 export->exp_req_replay_needed = 1;
1144                 export->exp_lock_replay_needed = 1;
1145                 cfs_spin_unlock(&export->exp_lock);
1146
1147                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1148                                  MSG_CONNECT_TRANSNO);
1149                 if (has_transno && transno == 0)
1150                         CWARN("Connect with zero transno!\n");
1151
1152                 if (has_transno && transno > 0 &&
1153                     transno < target->obd_next_recovery_transno &&
1154                     transno > target->obd_last_committed) {
1155                         /* another way is to use cmpxchg() so it will be
1156                          * lock free */
1157                         cfs_spin_lock(&target->obd_recovery_task_lock);
1158                         if (transno < target->obd_next_recovery_transno)
1159                                 target->obd_next_recovery_transno = transno;
1160                         cfs_spin_unlock(&target->obd_recovery_task_lock);
1161                 }
1162
1163                 cfs_atomic_inc(&target->obd_req_replay_clients);
1164                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1165                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1166                     target->obd_max_recoverable_clients)
1167                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1168         }
1169
1170         /* Tell the client we're in recovery, when client is involved in it. */
1171         if (target->obd_recovering)
1172                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1173
1174         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1175         conn = *tmp;
1176
1177         /* for the rest part, we return -ENOTCONN in case of errors
1178          * in order to let client initialize connection again.
1179          */
1180         revimp = class_new_import(target);
1181         if (revimp == NULL) {
1182                 CERROR("fail to alloc new reverse import.\n");
1183                 GOTO(out, rc = -ENOTCONN);
1184         }
1185
1186         cfs_spin_lock(&export->exp_lock);
1187         if (export->exp_imp_reverse != NULL) {
1188                 /* destroyed import can be still referenced in ctxt */
1189                 obd_set_info_async(req->rq_svc_thread->t_env, export,
1190                                    sizeof(KEY_REVIMP_UPD), KEY_REVIMP_UPD,
1191                                    0, NULL, NULL);
1192                 tmp_imp = export->exp_imp_reverse;
1193         }
1194         export->exp_imp_reverse = revimp;
1195         cfs_spin_unlock(&export->exp_lock);
1196
1197         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1198         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1199         revimp->imp_remote_handle = conn;
1200         revimp->imp_dlm_fake = 1;
1201         revimp->imp_state = LUSTRE_IMP_FULL;
1202
1203         /* unknown versions will be caught in
1204          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1205         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1206
1207         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1208             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1209                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1210         else
1211                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1212
1213         if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1214             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1215                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1216         else
1217                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1218
1219         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1220         if (rc) {
1221                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1222                 cfs_spin_lock(&export->exp_lock);
1223                 export->exp_imp_reverse = NULL;
1224                 cfs_spin_unlock(&export->exp_lock);
1225                 class_destroy_import(revimp);
1226         }
1227
1228         class_import_put(revimp);
1229
1230 out:
1231         if (tmp_imp != NULL)
1232                 client_destroy_import(tmp_imp);
1233         if (export) {
1234                 cfs_spin_lock(&export->exp_lock);
1235                 export->exp_connecting = 0;
1236                 cfs_spin_unlock(&export->exp_lock);
1237
1238                 class_export_put(export);
1239         }
1240         if (targref) {
1241                 cfs_spin_lock(&target->obd_dev_lock);
1242                 target->obd_conn_inprogress--;
1243                 cfs_spin_unlock(&target->obd_dev_lock);
1244
1245                 class_decref(targref, __FUNCTION__, cfs_current());
1246         }
1247         if (rc)
1248                 req->rq_status = rc;
1249         RETURN(rc);
1250 }
1251
1252 int target_handle_disconnect(struct ptlrpc_request *req)
1253 {
1254         int rc;
1255         ENTRY;
1256
1257         rc = req_capsule_server_pack(&req->rq_pill);
1258         if (rc)
1259                 RETURN(rc);
1260
1261         /* keep the rq_export around so we can send the reply */
1262         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1263
1264         RETURN(0);
1265 }
1266
1267 void target_destroy_export(struct obd_export *exp)
1268 {
1269         struct obd_import       *imp = NULL;
1270         /* exports created from last_rcvd data, and "fake"
1271            exports created by lctl don't have an import */
1272         cfs_spin_lock(&exp->exp_lock);
1273         if (exp->exp_imp_reverse != NULL) {
1274                 imp = exp->exp_imp_reverse;
1275                 exp->exp_imp_reverse = NULL;
1276         }
1277         cfs_spin_unlock(&exp->exp_lock);
1278         if (imp != NULL)
1279                 client_destroy_import(imp);
1280
1281         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1282         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1283         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1284         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1285 }
1286
1287 /*
1288  * Recovery functions
1289  */
1290 static void target_request_copy_get(struct ptlrpc_request *req)
1291 {
1292         class_export_rpc_get(req->rq_export);
1293         LASSERT(cfs_list_empty(&req->rq_list));
1294         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1295
1296         /* increase refcount to keep request in queue */
1297         cfs_atomic_inc(&req->rq_refcount);
1298         /** let export know it has replays to be handled */
1299         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1300 }
1301
1302 static void target_request_copy_put(struct ptlrpc_request *req)
1303 {
1304         LASSERT(cfs_list_empty(&req->rq_replay_list));
1305         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1306
1307         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1308         class_export_rpc_put(req->rq_export);
1309         ptlrpc_server_drop_request(req);
1310 }
1311
1312 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1313 {
1314         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1315         struct obd_export     *exp = req->rq_export;
1316         struct ptlrpc_request *reqiter;
1317         int                    dup = 0;
1318
1319         LASSERT(exp);
1320
1321         cfs_spin_lock(&exp->exp_lock);
1322         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1323                                 rq_replay_list) {
1324                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1325                         dup = 1;
1326                         break;
1327                 }
1328         }
1329
1330         if (dup) {
1331                 /* we expect it with RESENT and REPLAY flags */
1332                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1333                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1334                         CERROR("invalid flags %x of resent replay\n",
1335                                lustre_msg_get_flags(req->rq_reqmsg));
1336         } else {
1337                 cfs_list_add_tail(&req->rq_replay_list,
1338                                   &exp->exp_req_replay_queue);
1339         }
1340
1341         cfs_spin_unlock(&exp->exp_lock);
1342         return dup;
1343 }
1344
1345 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1346 {
1347         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1348         LASSERT(req->rq_export);
1349
1350         cfs_spin_lock(&req->rq_export->exp_lock);
1351         cfs_list_del_init(&req->rq_replay_list);
1352         cfs_spin_unlock(&req->rq_export->exp_lock);
1353 }
1354
1355 #ifdef __KERNEL__
1356 static void target_finish_recovery(struct obd_device *obd)
1357 {
1358         time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1359                                     obd->obd_recovery_start);
1360         ENTRY;
1361
1362         LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1363                       "%d recovered and %d %s evicted.\n", obd->obd_name,
1364                       (int)elapsed_time / 60, (int)elapsed_time % 60,
1365                       obd->obd_max_recoverable_clients,
1366                       cfs_atomic_read(&obd->obd_connected_clients),
1367                       obd->obd_stale_clients,
1368                       obd->obd_stale_clients == 1 ? "was" : "were");
1369
1370         ldlm_reprocess_all_ns(obd->obd_namespace);
1371         cfs_spin_lock(&obd->obd_recovery_task_lock);
1372         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1373             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1374             !cfs_list_empty(&obd->obd_final_req_queue)) {
1375                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1376                        obd->obd_name,
1377                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1378                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1379                                "" : "lock ",
1380                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1381                                "" : "final ");
1382                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1383                 LBUG();
1384         }
1385         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1386
1387         obd->obd_recovery_end = cfs_time_current_sec();
1388
1389         /* when recovery finished, cleanup orphans on mds and ost */
1390         if (OBT(obd) && OBP(obd, postrecov)) {
1391                 int rc = OBP(obd, postrecov)(obd);
1392                 if (rc < 0)
1393                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1394                                       obd->obd_name, rc);
1395         }
1396         EXIT;
1397 }
1398
1399 static void abort_req_replay_queue(struct obd_device *obd)
1400 {
1401         struct ptlrpc_request *req, *n;
1402         cfs_list_t abort_list;
1403
1404         CFS_INIT_LIST_HEAD(&abort_list);
1405         cfs_spin_lock(&obd->obd_recovery_task_lock);
1406         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1407         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1408         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1409                 DEBUG_REQ(D_WARNING, req, "aborted:");
1410                 req->rq_status = -ENOTCONN;
1411                 if (ptlrpc_error(req)) {
1412                         DEBUG_REQ(D_ERROR, req,
1413                                   "failed abort_req_reply; skipping");
1414                 }
1415                 target_exp_dequeue_req_replay(req);
1416                 target_request_copy_put(req);
1417         }
1418 }
1419
1420 static void abort_lock_replay_queue(struct obd_device *obd)
1421 {
1422         struct ptlrpc_request *req, *n;
1423         cfs_list_t abort_list;
1424
1425         CFS_INIT_LIST_HEAD(&abort_list);
1426         cfs_spin_lock(&obd->obd_recovery_task_lock);
1427         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1428         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1429         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1430                 DEBUG_REQ(D_ERROR, req, "aborted:");
1431                 req->rq_status = -ENOTCONN;
1432                 if (ptlrpc_error(req)) {
1433                         DEBUG_REQ(D_ERROR, req,
1434                                   "failed abort_lock_reply; skipping");
1435                 }
1436                 target_request_copy_put(req);
1437         }
1438 }
1439
1440 /* Called from a cleanup function if the device is being cleaned up
1441    forcefully.  The exports should all have been disconnected already,
1442    the only thing left to do is
1443      - clear the recovery flags
1444      - cancel the timer
1445      - free queued requests and replies, but don't send replies
1446    Because the obd_stopping flag is set, no new requests should be received.
1447
1448 */
1449 void target_cleanup_recovery(struct obd_device *obd)
1450 {
1451         struct ptlrpc_request *req, *n;
1452         cfs_list_t clean_list;
1453         ENTRY;
1454
1455         CFS_INIT_LIST_HEAD(&clean_list);
1456         cfs_spin_lock(&obd->obd_dev_lock);
1457         if (!obd->obd_recovering) {
1458                 cfs_spin_unlock(&obd->obd_dev_lock);
1459                 EXIT;
1460                 return;
1461         }
1462         obd->obd_recovering = obd->obd_abort_recovery = 0;
1463         cfs_spin_unlock(&obd->obd_dev_lock);
1464
1465         cfs_spin_lock(&obd->obd_recovery_task_lock);
1466         target_cancel_recovery_timer(obd);
1467         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1468         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1469
1470         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1471                 LASSERT(req->rq_reply_state == 0);
1472                 target_exp_dequeue_req_replay(req);
1473                 target_request_copy_put(req);
1474         }
1475
1476         cfs_spin_lock(&obd->obd_recovery_task_lock);
1477         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1478         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1479         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1480
1481         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1482                 LASSERT(req->rq_reply_state == 0);
1483                 target_request_copy_put(req);
1484         }
1485
1486         EXIT;
1487 }
1488
1489 /* obd_recovery_task_lock should be held */
1490 void target_cancel_recovery_timer(struct obd_device *obd)
1491 {
1492         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1493         cfs_timer_disarm(&obd->obd_recovery_timer);
1494 }
1495
1496 static void target_start_recovery_timer(struct obd_device *obd)
1497 {
1498         if (obd->obd_recovery_start != 0)
1499                 return;
1500
1501         cfs_spin_lock(&obd->obd_dev_lock);
1502         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1503                 cfs_spin_unlock(&obd->obd_dev_lock);
1504                 return;
1505         }
1506
1507         LASSERT(obd->obd_recovery_timeout != 0);
1508
1509         if (obd->obd_recovery_start != 0) {
1510                 cfs_spin_unlock(&obd->obd_dev_lock);
1511                 return;
1512         }
1513
1514         cfs_timer_arm(&obd->obd_recovery_timer,
1515                       cfs_time_shift(obd->obd_recovery_timeout));
1516         obd->obd_recovery_start = cfs_time_current_sec();
1517         cfs_spin_unlock(&obd->obd_dev_lock);
1518
1519         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1520                       "or until %d client%s reconnect%s\n",
1521                       obd->obd_name,
1522                       obd->obd_recovery_timeout / 60,
1523                       obd->obd_recovery_timeout % 60,
1524                       obd->obd_max_recoverable_clients,
1525                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1526                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1527 }
1528
1529 /**
1530  * extend recovery window.
1531  *
1532  * if @extend is true, extend recovery window to have @drt remaining at least;
1533  * otherwise, make sure the recovery timeout value is not less than @drt.
1534  */
1535 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1536 {
1537         cfs_time_t now;
1538         cfs_time_t end;
1539         cfs_duration_t left;
1540         int to;
1541
1542         cfs_spin_lock(&obd->obd_dev_lock);
1543         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1544                 cfs_spin_unlock(&obd->obd_dev_lock);
1545                 return;
1546         }
1547         LASSERT(obd->obd_recovery_start != 0);
1548
1549         now  = cfs_time_current_sec();
1550         to   = obd->obd_recovery_timeout;
1551         end  = obd->obd_recovery_start + to;
1552         left = cfs_time_sub(end, now);
1553
1554         if (extend && (drt > left)) {
1555                 to += drt - left;
1556         } else if (!extend && (drt > to)) {
1557                 to = drt;
1558                 /* reduce drt by already passed time */
1559                 drt -= obd->obd_recovery_timeout - left;
1560         }
1561
1562         if (to > obd->obd_recovery_time_hard)
1563                 to = obd->obd_recovery_time_hard;
1564         if (obd->obd_recovery_timeout < to) {
1565                 obd->obd_recovery_timeout = to;
1566                 cfs_timer_arm(&obd->obd_recovery_timer,
1567                               cfs_time_shift(drt));
1568         }
1569         cfs_spin_unlock(&obd->obd_dev_lock);
1570
1571         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1572                obd->obd_name, (unsigned)drt);
1573 }
1574
1575 /* Reset the timer with each new client connection */
1576 /*
1577  * This timer is actually reconnect_timer, which is for making sure
1578  * the total recovery window is at least as big as my reconnect
1579  * attempt timing. So the initial recovery time_out will be set to
1580  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1581  * from client is bigger than this, then the recovery time_out will
1582  * be extended to make sure the client could be reconnected, in the
1583  * process, the timeout from the new client should be ignored.
1584  */
1585
1586 static void
1587 check_and_start_recovery_timer(struct obd_device *obd,
1588                                struct ptlrpc_request *req,
1589                                int new_client)
1590 {
1591         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1592         struct obd_device_target *obt = &obd->u.obt;
1593         struct lustre_sb_info *lsi;
1594
1595         if (!new_client && service_time)
1596                 /* Teach server about old server's estimates, as first guess
1597                  * at how long new requests will take. */
1598                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1599                             service_time);
1600
1601         target_start_recovery_timer(obd);
1602
1603         /* convert the service time to rpc timeout,
1604          * reuse service_time to limit stack usage */
1605         service_time = at_est2timeout(service_time);
1606
1607         /* We expect other clients to timeout within service_time, then try
1608          * to reconnect, then try the failover server.  The max delay between
1609          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1610         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1611
1612         LASSERT(obt->obt_magic == OBT_MAGIC);
1613         lsi = s2lsi(obt->obt_sb);
1614         if (!(lsi->lsi_flags | LDD_F_IR_CAPABLE))
1615                 service_time += 2 * (CONNECTION_SWITCH_MAX +
1616                                      CONNECTION_SWITCH_INC);
1617         if (service_time > obd->obd_recovery_timeout && !new_client)
1618                 extend_recovery_timer(obd, service_time, false);
1619 }
1620
1621 /** Health checking routines */
1622 static inline int exp_connect_healthy(struct obd_export *exp)
1623 {
1624         return (exp->exp_in_recovery);
1625 }
1626
1627 /** if export done req_replay or has replay in queue */
1628 static inline int exp_req_replay_healthy(struct obd_export *exp)
1629 {
1630         return (!exp->exp_req_replay_needed ||
1631                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1632 }
1633 /** if export done lock_replay or has replay in queue */
1634 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1635 {
1636         return (!exp->exp_lock_replay_needed ||
1637                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1638 }
1639
1640 static inline int exp_vbr_healthy(struct obd_export *exp)
1641 {
1642         return (!exp->exp_vbr_failed);
1643 }
1644
1645 static inline int exp_finished(struct obd_export *exp)
1646 {
1647         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1648 }
1649
1650 /** Checking routines for recovery */
1651 static int check_for_clients(struct obd_device *obd)
1652 {
1653         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1654
1655         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1656                 return 1;
1657         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1658         if (obd->obd_no_conn == 0 &&
1659             clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1660                 return 1;
1661         return 0;
1662 }
1663
1664 static int check_for_next_transno(struct obd_device *obd)
1665 {
1666         struct ptlrpc_request *req = NULL;
1667         int wake_up = 0, connected, completed, queue_len;
1668         __u64 next_transno, req_transno;
1669         ENTRY;
1670
1671         cfs_spin_lock(&obd->obd_recovery_task_lock);
1672         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1673                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1674                                      struct ptlrpc_request, rq_list);
1675                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1676         } else {
1677                 req_transno = 0;
1678         }
1679
1680         connected = cfs_atomic_read(&obd->obd_connected_clients);
1681         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1682         queue_len = obd->obd_requests_queued_for_recovery;
1683         next_transno = obd->obd_next_recovery_transno;
1684
1685         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1686                "req_transno: "LPU64", next_transno: "LPU64"\n",
1687                obd->obd_max_recoverable_clients, connected, completed,
1688                queue_len, req_transno, next_transno);
1689
1690         if (obd->obd_abort_recovery) {
1691                 CDEBUG(D_HA, "waking for aborted recovery\n");
1692                 wake_up = 1;
1693         } else if (obd->obd_recovery_expired) {
1694                 CDEBUG(D_HA, "waking for expired recovery\n");
1695                 wake_up = 1;
1696         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1697                 CDEBUG(D_HA, "waking for completed recovery\n");
1698                 wake_up = 1;
1699         } else if (req_transno == next_transno) {
1700                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1701                 wake_up = 1;
1702         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1703                 int d_lvl = D_HA;
1704                 /** handle gaps occured due to lost reply or VBR */
1705                 LASSERTF(req_transno >= next_transno,
1706                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1707                          req_transno, next_transno);
1708                 if (req_transno > obd->obd_last_committed &&
1709                     !obd->obd_version_recov)
1710                         d_lvl = D_ERROR;
1711                 CDEBUG(d_lvl,
1712                        "%s: waking for gap in transno, VBR is %s (skip: "
1713                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1714                        ", last_committed: "LPD64")\n",
1715                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1716                        next_transno, queue_len, completed, connected,
1717                        req_transno, obd->obd_last_committed);
1718                 obd->obd_next_recovery_transno = req_transno;
1719                 wake_up = 1;
1720         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1721                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1722                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1723                 obd->obd_next_recovery_transno = req_transno;
1724                 wake_up = 1;
1725         }
1726         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1727         return wake_up;
1728 }
1729
1730 static int check_for_next_lock(struct obd_device *obd)
1731 {
1732         int wake_up = 0;
1733
1734         cfs_spin_lock(&obd->obd_recovery_task_lock);
1735         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1736                 CDEBUG(D_HA, "waking for next lock\n");
1737                 wake_up = 1;
1738         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1739                 CDEBUG(D_HA, "waking for completed lock replay\n");
1740                 wake_up = 1;
1741         } else if (obd->obd_abort_recovery) {
1742                 CDEBUG(D_HA, "waking for aborted recovery\n");
1743                 wake_up = 1;
1744         } else if (obd->obd_recovery_expired) {
1745                 CDEBUG(D_HA, "waking for expired recovery\n");
1746                 wake_up = 1;
1747         }
1748         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1749
1750         return wake_up;
1751 }
1752
1753 /**
1754  * wait for recovery events,
1755  * check its status with help of check_routine
1756  * evict dead clients via health_check
1757  */
1758 static int target_recovery_overseer(struct obd_device *obd,
1759                                     int (*check_routine)(struct obd_device *),
1760                                     int (*health_check)(struct obd_export *))
1761 {
1762 repeat:
1763         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1764         if (obd->obd_abort_recovery) {
1765                 CWARN("recovery is aborted, evict exports in recovery\n");
1766                 /** evict exports which didn't finish recovery yet */
1767                 class_disconnect_stale_exports(obd, exp_finished);
1768                 return 1;
1769         } else if (obd->obd_recovery_expired) {
1770                 obd->obd_recovery_expired = 0;
1771                 /** If some clients died being recovered, evict them */
1772                 LCONSOLE_WARN("%s: recovery is timed out, "
1773                               "evict stale exports\n", obd->obd_name);
1774                 /** evict cexports with no replay in queue, they are stalled */
1775                 class_disconnect_stale_exports(obd, health_check);
1776                 /** continue with VBR */
1777                 cfs_spin_lock(&obd->obd_dev_lock);
1778                 obd->obd_version_recov = 1;
1779                 cfs_spin_unlock(&obd->obd_dev_lock);
1780                 /**
1781                  * reset timer, recovery will proceed with versions now,
1782                  * timeout is set just to handle reconnection delays
1783                  */
1784                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1785                 /** Wait for recovery events again, after evicting bad clients */
1786                 goto repeat;
1787         }
1788         return 0;
1789 }
1790
1791 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1792 {
1793         struct ptlrpc_request *req = NULL;
1794         ENTRY;
1795
1796         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1797                obd->obd_next_recovery_transno);
1798
1799         if (target_recovery_overseer(obd, check_for_next_transno,
1800                                      exp_req_replay_healthy)) {
1801                 abort_req_replay_queue(obd);
1802                 abort_lock_replay_queue(obd);
1803         }
1804
1805         cfs_spin_lock(&obd->obd_recovery_task_lock);
1806         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1807                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1808                                      struct ptlrpc_request, rq_list);
1809                 cfs_list_del_init(&req->rq_list);
1810                 obd->obd_requests_queued_for_recovery--;
1811                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1812         } else {
1813                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1814                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1815                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1816                 /** evict exports failed VBR */
1817                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1818         }
1819         RETURN(req);
1820 }
1821
1822 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1823 {
1824         struct ptlrpc_request *req = NULL;
1825
1826         CDEBUG(D_HA, "Waiting for lock\n");
1827         if (target_recovery_overseer(obd, check_for_next_lock,
1828                                      exp_lock_replay_healthy))
1829                 abort_lock_replay_queue(obd);
1830
1831         cfs_spin_lock(&obd->obd_recovery_task_lock);
1832         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1833                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1834                                      struct ptlrpc_request, rq_list);
1835                 cfs_list_del_init(&req->rq_list);
1836                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1837         } else {
1838                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1839                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1840                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1841                 /** evict exports failed VBR */
1842                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1843         }
1844         return req;
1845 }
1846
1847 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1848 {
1849         struct ptlrpc_request *req = NULL;
1850
1851         cfs_spin_lock(&obd->obd_recovery_task_lock);
1852         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1853                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1854                                      struct ptlrpc_request, rq_list);
1855                 cfs_list_del_init(&req->rq_list);
1856                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1857                 if (req->rq_export->exp_in_recovery) {
1858                         cfs_spin_lock(&req->rq_export->exp_lock);
1859                         req->rq_export->exp_in_recovery = 0;
1860                         cfs_spin_unlock(&req->rq_export->exp_lock);
1861                 }
1862         } else {
1863                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1864         }
1865         return req;
1866 }
1867
1868 static int handle_recovery_req(struct ptlrpc_thread *thread,
1869                                struct ptlrpc_request *req,
1870                                svc_handler_t handler)
1871 {
1872         int rc;
1873         ENTRY;
1874
1875         /**
1876          * export can be evicted during recovery, no need to handle replays for
1877          * it after that, discard such request silently
1878          */
1879         if (req->rq_export->exp_disconnected)
1880                 GOTO(reqcopy_put, rc = 0);
1881
1882         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1883         if (rc) {
1884                 CERROR("Failure to initialize session: %d\n", rc);
1885                 GOTO(reqcopy_put, rc);
1886         }
1887
1888         req->rq_recov_session.lc_thread = thread;
1889         lu_context_enter(&req->rq_recov_session);
1890         req->rq_svc_thread = thread;
1891         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1892
1893         /* thread context */
1894         lu_context_enter(&thread->t_env->le_ctx);
1895         (void)handler(req);
1896         lu_context_exit(&thread->t_env->le_ctx);
1897
1898         lu_context_exit(&req->rq_recov_session);
1899         lu_context_fini(&req->rq_recov_session);
1900         /* don't reset timer for final stage */
1901         if (!exp_finished(req->rq_export)) {
1902                 int to = obd_timeout;
1903
1904                 /**
1905                  * Add request timeout to the recovery time so next request from
1906                  * this client may come in recovery time
1907                  */
1908                 if (!AT_OFF) {
1909                         struct ptlrpc_service_part *svcpt;
1910
1911                         svcpt = req->rq_rqbd->rqbd_svcpt;
1912                         /* If the server sent early reply for this request,
1913                          * the client will recalculate the timeout according to
1914                          * current server estimate service time, so we will
1915                          * use the maxium timeout here for waiting the client
1916                          * sending the next req */
1917                         to = max((int)at_est2timeout(
1918                                  at_get(&svcpt->scp_at_estimate)),
1919                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1920                         /* Add net_latency (see ptlrpc_replay_req) */
1921                         to += lustre_msg_get_service_time(req->rq_reqmsg);
1922                 }
1923                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1924         }
1925 reqcopy_put:
1926         RETURN(rc);
1927 }
1928
1929 static int target_recovery_thread(void *arg)
1930 {
1931         struct lu_target *lut = arg;
1932         struct obd_device *obd = lut->lut_obd;
1933         struct ptlrpc_request *req;
1934         struct target_recovery_data *trd = &obd->obd_recovery_data;
1935         unsigned long delta;
1936         unsigned long flags;
1937         struct lu_env *env;
1938         struct ptlrpc_thread *thread = NULL;
1939         int rc = 0;
1940         ENTRY;
1941
1942         cfs_daemonize_ctxt("tgt_recov");
1943
1944         SIGNAL_MASK_LOCK(current, flags);
1945         sigfillset(&current->blocked);
1946         RECALC_SIGPENDING;
1947         SIGNAL_MASK_UNLOCK(current, flags);
1948
1949         OBD_ALLOC_PTR(thread);
1950         if (thread == NULL)
1951                 RETURN(-ENOMEM);
1952
1953         OBD_ALLOC_PTR(env);
1954         if (env == NULL) {
1955                 OBD_FREE_PTR(thread);
1956                 RETURN(-ENOMEM);
1957         }
1958
1959         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1960         if (rc) {
1961                 OBD_FREE_PTR(thread);
1962                 OBD_FREE_PTR(env);
1963                 RETURN(rc);
1964         }
1965
1966         thread->t_env = env;
1967         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1968         env->le_ctx.lc_thread = thread;
1969         thread->t_data = NULL;
1970         thread->t_watchdog = NULL;
1971
1972         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1973                cfs_curproc_pid());
1974         trd->trd_processing_task = cfs_curproc_pid();
1975
1976         cfs_spin_lock(&obd->obd_dev_lock);
1977         obd->obd_recovering = 1;
1978         cfs_spin_unlock(&obd->obd_dev_lock);
1979         cfs_complete(&trd->trd_starting);
1980
1981         /* first of all, we have to know the first transno to replay */
1982         if (target_recovery_overseer(obd, check_for_clients,
1983                                      exp_connect_healthy)) {
1984                 abort_req_replay_queue(obd);
1985                 abort_lock_replay_queue(obd);
1986         }
1987
1988         /* next stage: replay requests */
1989         delta = jiffies;
1990         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1991                cfs_atomic_read(&obd->obd_req_replay_clients),
1992                obd->obd_next_recovery_transno);
1993         while ((req = target_next_replay_req(obd))) {
1994                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1995                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1996                           lustre_msg_get_transno(req->rq_reqmsg),
1997                           libcfs_nid2str(req->rq_peer.nid));
1998                 handle_recovery_req(thread, req,
1999                                     trd->trd_recovery_handler);
2000                 /**
2001                  * bz18031: increase next_recovery_transno before
2002                  * target_request_copy_put() will drop exp_rpc reference
2003                  */
2004                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2005                 obd->obd_next_recovery_transno++;
2006                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2007                 target_exp_dequeue_req_replay(req);
2008                 target_request_copy_put(req);
2009                 obd->obd_replayed_requests++;
2010         }
2011
2012         /**
2013          * The second stage: replay locks
2014          */
2015         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2016                cfs_atomic_read(&obd->obd_lock_replay_clients));
2017         while ((req = target_next_replay_lock(obd))) {
2018                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2019                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2020                           libcfs_nid2str(req->rq_peer.nid));
2021                 handle_recovery_req(thread, req,
2022                                     trd->trd_recovery_handler);
2023                 target_request_copy_put(req);
2024                 obd->obd_replayed_locks++;
2025         }
2026
2027         /**
2028          * The third stage: reply on final pings, at this moment all clients
2029          * must have request in final queue
2030          */
2031         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2032         /** Update server last boot epoch */
2033         lut_boot_epoch_update(lut);
2034         /* We drop recoverying flag to forward all new requests
2035          * to regular mds_handle() since now */
2036         cfs_spin_lock(&obd->obd_dev_lock);
2037         obd->obd_recovering = obd->obd_abort_recovery = 0;
2038         cfs_spin_unlock(&obd->obd_dev_lock);
2039         cfs_spin_lock(&obd->obd_recovery_task_lock);
2040         target_cancel_recovery_timer(obd);
2041         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2042         while ((req = target_next_final_ping(obd))) {
2043                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2044                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2045                           libcfs_nid2str(req->rq_peer.nid));
2046                 handle_recovery_req(thread, req,
2047                                     trd->trd_recovery_handler);
2048                 target_request_copy_put(req);
2049         }
2050
2051         delta = (jiffies - delta) / CFS_HZ;
2052         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2053               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2054         if (delta > OBD_RECOVERY_TIME_SOFT) {
2055                 CWARN("too long recovery - read logs\n");
2056                 libcfs_debug_dumplog();
2057         }
2058
2059         target_finish_recovery(obd);
2060
2061         lu_context_fini(&env->le_ctx);
2062         trd->trd_processing_task = 0;
2063         cfs_complete(&trd->trd_finishing);
2064
2065         OBD_FREE_PTR(thread);
2066         OBD_FREE_PTR(env);
2067         RETURN(rc);
2068 }
2069
2070 static int target_start_recovery_thread(struct lu_target *lut,
2071                                         svc_handler_t handler)
2072 {
2073         struct obd_device *obd = lut->lut_obd;
2074         int rc = 0;
2075         struct target_recovery_data *trd = &obd->obd_recovery_data;
2076
2077         memset(trd, 0, sizeof(*trd));
2078         cfs_init_completion(&trd->trd_starting);
2079         cfs_init_completion(&trd->trd_finishing);
2080         trd->trd_recovery_handler = handler;
2081
2082         if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2083                 cfs_wait_for_completion(&trd->trd_starting);
2084                 LASSERT(obd->obd_recovering != 0);
2085         } else
2086                 rc = -ECHILD;
2087
2088         return rc;
2089 }
2090
2091 void target_stop_recovery_thread(struct obd_device *obd)
2092 {
2093         if (obd->obd_recovery_data.trd_processing_task > 0) {
2094                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2095                 /** recovery can be done but postrecovery is not yet */
2096                 cfs_spin_lock(&obd->obd_dev_lock);
2097                 if (obd->obd_recovering) {
2098                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2099                         obd->obd_abort_recovery = 1;
2100                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2101                 }
2102                 cfs_spin_unlock(&obd->obd_dev_lock);
2103                 cfs_wait_for_completion(&trd->trd_finishing);
2104         }
2105 }
2106
2107 void target_recovery_fini(struct obd_device *obd)
2108 {
2109         class_disconnect_exports(obd);
2110         target_stop_recovery_thread(obd);
2111         target_cleanup_recovery(obd);
2112 }
2113 EXPORT_SYMBOL(target_recovery_fini);
2114
2115 static void target_recovery_expired(unsigned long castmeharder)
2116 {
2117         struct obd_device *obd = (struct obd_device *)castmeharder;
2118         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2119                " after %lds (%d clients connected)\n",
2120                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2121                cfs_time_current_sec()- obd->obd_recovery_start,
2122                cfs_atomic_read(&obd->obd_connected_clients));
2123
2124         obd->obd_recovery_expired = 1;
2125         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2126 }
2127
2128 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2129 {
2130         struct obd_device *obd = lut->lut_obd;
2131         if (obd->obd_max_recoverable_clients == 0) {
2132                 /** Update server last boot epoch */
2133                 lut_boot_epoch_update(lut);
2134                 return;
2135         }
2136
2137         CWARN("RECOVERY: service %s, %d recoverable clients, "
2138               "last_transno "LPU64"\n", obd->obd_name,
2139               obd->obd_max_recoverable_clients, obd->obd_last_committed);
2140         LASSERT(obd->obd_stopping == 0);
2141         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2142         obd->obd_recovery_start = 0;
2143         obd->obd_recovery_end = 0;
2144
2145         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2146         target_start_recovery_thread(lut, handler);
2147 }
2148 EXPORT_SYMBOL(target_recovery_init);
2149
2150 #endif /* __KERNEL__ */
2151
2152 static int target_process_req_flags(struct obd_device *obd,
2153                                     struct ptlrpc_request *req)
2154 {
2155         struct obd_export *exp = req->rq_export;
2156         LASSERT(exp != NULL);
2157         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2158                 /* client declares he's ready to replay locks */
2159                 cfs_spin_lock(&exp->exp_lock);
2160                 if (exp->exp_req_replay_needed) {
2161                         exp->exp_req_replay_needed = 0;
2162                         cfs_spin_unlock(&exp->exp_lock);
2163
2164                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2165                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2166                 } else {
2167                         cfs_spin_unlock(&exp->exp_lock);
2168                 }
2169         }
2170         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2171                 /* client declares he's ready to complete recovery
2172                  * so, we put the request on th final queue */
2173                 cfs_spin_lock(&exp->exp_lock);
2174                 if (exp->exp_lock_replay_needed) {
2175                         exp->exp_lock_replay_needed = 0;
2176                         cfs_spin_unlock(&exp->exp_lock);
2177
2178                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2179                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2180                 } else {
2181                         cfs_spin_unlock(&exp->exp_lock);
2182                 }
2183         }
2184         return 0;
2185 }
2186
2187 int target_queue_recovery_request(struct ptlrpc_request *req,
2188                                   struct obd_device *obd)
2189 {
2190         cfs_list_t *tmp;
2191         int inserted = 0;
2192         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2193         ENTRY;
2194
2195         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2196                 /* Processing the queue right now, don't re-add. */
2197                 RETURN(1);
2198         }
2199
2200         target_process_req_flags(obd, req);
2201
2202         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2203                 /* client declares he's ready to complete recovery
2204                  * so, we put the request on th final queue */
2205                 target_request_copy_get(req);
2206                 DEBUG_REQ(D_HA, req, "queue final req");
2207                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2208                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2209                 if (obd->obd_recovering) {
2210                         cfs_list_add_tail(&req->rq_list,
2211                                           &obd->obd_final_req_queue);
2212                 } else {
2213                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2214                         target_request_copy_put(req);
2215                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2216                 }
2217                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2218                 RETURN(0);
2219         }
2220         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2221                 /* client declares he's ready to replay locks */
2222                 target_request_copy_get(req);
2223                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2224                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2225                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2226                 LASSERT(obd->obd_recovering);
2227                 /* usually due to recovery abort */
2228                 if (!req->rq_export->exp_in_recovery) {
2229                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2230                         target_request_copy_put(req);
2231                         RETURN(-ENOTCONN);
2232                 }
2233                 LASSERT(req->rq_export->exp_lock_replay_needed);
2234                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2235                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2236                 RETURN(0);
2237         }
2238
2239         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2240          * (i.e. buflens etc are in my own byte order), but type-dependent
2241          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2242
2243         if (!transno) {
2244                 CFS_INIT_LIST_HEAD(&req->rq_list);
2245                 DEBUG_REQ(D_HA, req, "not queueing");
2246                 RETURN(1);
2247         }
2248
2249         /* If we're processing the queue, we want don't want to queue this
2250          * message.
2251          *
2252          * Also, if this request has a transno less than the one we're waiting
2253          * for, we should process it now.  It could (and currently always will)
2254          * be an open request for a descriptor that was opened some time ago.
2255          *
2256          * Also, a resent, replayed request that has already been
2257          * handled will pass through here and be processed immediately.
2258          */
2259         CDEBUG(D_HA, "Next recovery transno: "LPU64
2260                ", current: "LPU64", replaying\n",
2261                obd->obd_next_recovery_transno, transno);
2262         cfs_spin_lock(&obd->obd_recovery_task_lock);
2263         if (transno < obd->obd_next_recovery_transno) {
2264                 /* Processing the queue right now, don't re-add. */
2265                 LASSERT(cfs_list_empty(&req->rq_list));
2266                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2267                 RETURN(1);
2268         }
2269         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2270
2271         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2272                 RETURN(0);
2273
2274         target_request_copy_get(req);
2275         if (!req->rq_export->exp_in_recovery) {
2276                 target_request_copy_put(req);
2277                 RETURN(-ENOTCONN);
2278         }
2279         LASSERT(req->rq_export->exp_req_replay_needed);
2280
2281         if (target_exp_enqueue_req_replay(req)) {
2282                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2283                 target_request_copy_put(req);
2284                 RETURN(0);
2285         }
2286
2287         /* XXX O(n^2) */
2288         cfs_spin_lock(&obd->obd_recovery_task_lock);
2289         LASSERT(obd->obd_recovering);
2290         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2291                 struct ptlrpc_request *reqiter =
2292                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2293
2294                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2295                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2296                         inserted = 1;
2297                         break;
2298                 }
2299
2300                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2301                              transno)) {
2302                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2303                                   "has been claimed by another client");
2304                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2305                         target_exp_dequeue_req_replay(req);
2306                         target_request_copy_put(req);
2307                         RETURN(0);
2308                 }
2309         }
2310
2311         if (!inserted)
2312                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2313
2314         obd->obd_requests_queued_for_recovery++;
2315         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2316         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2317         RETURN(0);
2318 }
2319
2320 int target_handle_ping(struct ptlrpc_request *req)
2321 {
2322         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2323         return req_capsule_server_pack(&req->rq_pill);
2324 }
2325
2326 void target_committed_to_req(struct ptlrpc_request *req)
2327 {
2328         struct obd_export *exp = req->rq_export;
2329
2330         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2331                 lustre_msg_set_last_committed(req->rq_repmsg,
2332                                               exp->exp_last_committed);
2333         else
2334                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2335                           "%d)", exp->exp_obd->obd_no_transno,
2336                           req->rq_repmsg == NULL);
2337
2338         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2339                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2340 }
2341 EXPORT_SYMBOL(target_committed_to_req);
2342
2343 #endif /* HAVE_SERVER_SUPPORT */
2344
2345 /**
2346  * Packs current SLV and Limit into \a req.
2347  */
2348 int target_pack_pool_reply(struct ptlrpc_request *req)
2349 {
2350         struct obd_device *obd;
2351         ENTRY;
2352
2353         /*
2354          * Check that we still have all structures alive as this may
2355          * be some late rpc in shutdown time.
2356          */
2357         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2358                      !exp_connect_lru_resize(req->rq_export))) {
2359                 lustre_msg_set_slv(req->rq_repmsg, 0);
2360                 lustre_msg_set_limit(req->rq_repmsg, 0);
2361                 RETURN(0);
2362         }
2363
2364         /*
2365          * OBD is alive here as export is alive, which we checked above.
2366          */
2367         obd = req->rq_export->exp_obd;
2368
2369         cfs_read_lock(&obd->obd_pool_lock);
2370         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2371         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2372         cfs_read_unlock(&obd->obd_pool_lock);
2373
2374         RETURN(0);
2375 }
2376
2377 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2378 {
2379         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2380                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2381                 return (-ECOMM);
2382         }
2383
2384         if (unlikely(rc)) {
2385                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2386                 req->rq_status = rc;
2387                 return (ptlrpc_send_error(req, 1));
2388         } else {
2389                 DEBUG_REQ(D_NET, req, "sending reply");
2390         }
2391
2392         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2393 }
2394
2395 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2396 {
2397         struct ptlrpc_service_part *svcpt;
2398         int                        netrc;
2399         struct ptlrpc_reply_state *rs;
2400         struct obd_export         *exp;
2401         ENTRY;
2402
2403         if (req->rq_no_reply) {
2404                 EXIT;
2405                 return;
2406         }
2407
2408         svcpt = req->rq_rqbd->rqbd_svcpt;
2409         rs = req->rq_reply_state;
2410         if (rs == NULL || !rs->rs_difficult) {
2411                 /* no notifiers */
2412                 target_send_reply_msg (req, rc, fail_id);
2413                 EXIT;
2414                 return;
2415         }
2416
2417         /* must be an export if locks saved */
2418         LASSERT (req->rq_export != NULL);
2419         /* req/reply consistent */
2420         LASSERT(rs->rs_svcpt == svcpt);
2421
2422         /* "fresh" reply */
2423         LASSERT (!rs->rs_scheduled);
2424         LASSERT (!rs->rs_scheduled_ever);
2425         LASSERT (!rs->rs_handled);
2426         LASSERT (!rs->rs_on_net);
2427         LASSERT (rs->rs_export == NULL);
2428         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2429         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2430
2431         exp = class_export_get (req->rq_export);
2432
2433         /* disable reply scheduling while I'm setting up */
2434         rs->rs_scheduled = 1;
2435         rs->rs_on_net    = 1;
2436         rs->rs_xid       = req->rq_xid;
2437         rs->rs_transno   = req->rq_transno;
2438         rs->rs_export    = exp;
2439         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2440
2441         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2442         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2443                rs->rs_transno, exp->exp_last_committed);
2444         if (rs->rs_transno > exp->exp_last_committed) {
2445                 /* not committed already */
2446                 cfs_list_add_tail(&rs->rs_obd_list,
2447                                   &exp->exp_uncommitted_replies);
2448         }
2449         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2450
2451         cfs_spin_lock(&exp->exp_lock);
2452         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2453         cfs_spin_unlock(&exp->exp_lock);
2454
2455         netrc = target_send_reply_msg (req, rc, fail_id);
2456
2457         cfs_spin_lock(&svcpt->scp_rep_lock);
2458
2459         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2460
2461         if (netrc != 0) {
2462                 /* error sending: reply is off the net.  Also we need +1
2463                  * reply ref until ptlrpc_handle_rs() is done
2464                  * with the reply state (if the send was successful, there
2465                  * would have been +1 ref for the net, which
2466                  * reply_out_callback leaves alone) */
2467                 rs->rs_on_net = 0;
2468                 ptlrpc_rs_addref(rs);
2469         }
2470
2471         cfs_spin_lock(&rs->rs_lock);
2472         if (rs->rs_transno <= exp->exp_last_committed ||
2473             (!rs->rs_on_net && !rs->rs_no_ack) ||
2474              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2475              cfs_list_empty(&rs->rs_obd_list)) {
2476                 CDEBUG(D_HA, "Schedule reply immediately\n");
2477                 ptlrpc_dispatch_difficult_reply(rs);
2478         } else {
2479                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2480                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2481         }
2482         cfs_spin_unlock(&rs->rs_lock);
2483         cfs_spin_unlock(&svcpt->scp_rep_lock);
2484         EXIT;
2485 }
2486
2487 int target_handle_qc_callback(struct ptlrpc_request *req)
2488 {
2489         struct obd_quotactl *oqctl;
2490         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2491
2492         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2493         if (oqctl == NULL) {
2494                 CERROR("Can't unpack obd_quotactl\n");
2495                 RETURN(-EPROTO);
2496         }
2497
2498         cli->cl_qchk_stat = oqctl->qc_stat;
2499
2500         return 0;
2501 }
2502
2503 #ifdef HAVE_QUOTA_SUPPORT
2504 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2505 {
2506 #ifdef __KERNEL__
2507         struct obd_device *obd = req->rq_export->exp_obd;
2508         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2509         struct obd_device_target *obt;
2510         struct lustre_quota_ctxt *qctxt;
2511         struct qunit_data *qdata = NULL;
2512         int rc = 0;
2513         ENTRY;
2514
2515         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2516                 RETURN(rc);
2517
2518         rc = req_capsule_server_pack(&req->rq_pill);
2519         if (rc) {
2520                 CERROR("packing reply failed!: rc = %d\n", rc);
2521                 RETURN(rc);
2522         }
2523
2524         LASSERT(req->rq_export);
2525
2526         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2527         if (IS_ERR(qdata)) {
2528                 rc = PTR_ERR(qdata);
2529                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2530                 req->rq_status = rc;
2531                 GOTO(out, rc);
2532         }
2533
2534         /* we use the observer */
2535         if (obd_pin_observer(obd, &lov_obd) ||
2536             obd_pin_observer(lov_obd, &master_obd)) {
2537                 CERROR("Can't find the observer, it is recovering\n");
2538                 req->rq_status = -EAGAIN;
2539                 GOTO(out, rc);
2540         }
2541
2542         obt = &master_obd->u.obt;
2543         qctxt = &obt->obt_qctxt;
2544
2545         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2546                 /* quota_type has not been processed yet, return EAGAIN
2547                  * until we know whether or not quotas are supposed to
2548                  * be enabled */
2549                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2550                        "-EAGAIN\n");
2551                 req->rq_status = -EAGAIN;
2552                 GOTO(out, rc);
2553         }
2554
2555         cfs_down_read(&obt->obt_rwsem);
2556         if (qctxt->lqc_lqs_hash == NULL) {
2557                 cfs_up_read(&obt->obt_rwsem);
2558                 /* quota_type has not been processed yet, return EAGAIN
2559                  * until we know whether or not quotas are supposed to
2560                  * be enabled */
2561                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2562                        "-EAGAIN\n");
2563                 req->rq_status = -EAGAIN;
2564                 GOTO(out, rc);
2565         }
2566
2567         LASSERT(qctxt->lqc_handler);
2568         rc = qctxt->lqc_handler(master_obd, qdata,
2569                                 lustre_msg_get_opc(req->rq_reqmsg));
2570         cfs_up_read(&obt->obt_rwsem);
2571         if (rc && rc != -EDQUOT)
2572                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2573                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2574         req->rq_status = rc;
2575
2576         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2577         if (rc < 0) {
2578                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2579                 GOTO(out, rc);
2580         }
2581
2582         /* Block the quota req. b=14840 */
2583         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2584         EXIT;
2585
2586 out:
2587         if (master_obd)
2588                 obd_unpin_observer(lov_obd);
2589         if (lov_obd)
2590                 obd_unpin_observer(obd);
2591
2592         rc = ptlrpc_reply(req);
2593         return rc;
2594 #else
2595         return 0;
2596 #endif /* !__KERNEL__ */
2597 }
2598 #endif /* HAVE_QUOTA_SUPPORT */
2599
2600 ldlm_mode_t lck_compat_array[] = {
2601         [LCK_EX] LCK_COMPAT_EX,
2602         [LCK_PW] LCK_COMPAT_PW,
2603         [LCK_PR] LCK_COMPAT_PR,
2604         [LCK_CW] LCK_COMPAT_CW,
2605         [LCK_CR] LCK_COMPAT_CR,
2606         [LCK_NL] LCK_COMPAT_NL,
2607         [LCK_GROUP] LCK_COMPAT_GROUP,
2608         [LCK_COS] LCK_COMPAT_COS,
2609 };
2610
2611 /**
2612  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2613  * not escape to the user level.
2614  */
2615 int ldlm_error2errno(ldlm_error_t error)
2616 {
2617         int result;
2618
2619         switch (error) {
2620         case ELDLM_OK:
2621                 result = 0;
2622                 break;
2623         case ELDLM_LOCK_CHANGED:
2624                 result = -ESTALE;
2625                 break;
2626         case ELDLM_LOCK_ABORTED:
2627                 result = -ENAVAIL;
2628                 break;
2629         case ELDLM_LOCK_REPLACED:
2630                 result = -ESRCH;
2631                 break;
2632         case ELDLM_NO_LOCK_DATA:
2633                 result = -ENOENT;
2634                 break;
2635         case ELDLM_NAMESPACE_EXISTS:
2636                 result = -EEXIST;
2637                 break;
2638         case ELDLM_BAD_NAMESPACE:
2639                 result = -EBADF;
2640                 break;
2641         default:
2642                 if (((int)error) < 0)  /* cast to signed type */
2643                         result = error; /* as ldlm_error_t can be unsigned */
2644                 else {
2645                         CERROR("Invalid DLM result code: %d\n", error);
2646                         result = -EPROTO;
2647                 }
2648         }
2649         return result;
2650 }
2651 EXPORT_SYMBOL(ldlm_error2errno);
2652
2653 /**
2654  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2655  */
2656 ldlm_error_t ldlm_errno2error(int err_no)
2657 {
2658         int error;
2659
2660         switch (err_no) {
2661         case 0:
2662                 error = ELDLM_OK;
2663                 break;
2664         case -ESTALE:
2665                 error = ELDLM_LOCK_CHANGED;
2666                 break;
2667         case -ENAVAIL:
2668                 error = ELDLM_LOCK_ABORTED;
2669                 break;
2670         case -ESRCH:
2671                 error = ELDLM_LOCK_REPLACED;
2672                 break;
2673         case -ENOENT:
2674                 error = ELDLM_NO_LOCK_DATA;
2675                 break;
2676         case -EEXIST:
2677                 error = ELDLM_NAMESPACE_EXISTS;
2678                 break;
2679         case -EBADF:
2680                 error = ELDLM_BAD_NAMESPACE;
2681                 break;
2682         default:
2683                 error = err_no;
2684         }
2685         return error;
2686 }
2687 EXPORT_SYMBOL(ldlm_errno2error);
2688
2689 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2690 void ldlm_dump_export_locks(struct obd_export *exp)
2691 {
2692         cfs_spin_lock(&exp->exp_locks_list_guard);
2693         if (!cfs_list_empty(&exp->exp_locks_list)) {
2694             struct ldlm_lock *lock;
2695
2696             CERROR("dumping locks for export %p,"
2697                    "ignore if the unmount doesn't hang\n", exp);
2698             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2699                 LDLM_ERROR(lock, "lock:");
2700         }
2701         cfs_spin_unlock(&exp->exp_locks_list_guard);
2702 }
2703 #endif
2704
2705 #ifdef HAVE_SERVER_SUPPORT
2706 static int target_bulk_timeout(void *data)
2707 {
2708         ENTRY;
2709         /* We don't fail the connection here, because having the export
2710          * killed makes the (vital) call to commitrw very sad.
2711          */
2712         RETURN(1);
2713 }
2714
2715 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2716 {
2717         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2718 }
2719
2720 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2721                    struct l_wait_info *lwi)
2722 {
2723         struct ptlrpc_request *req = desc->bd_req;
2724         int rc = 0;
2725         ENTRY;
2726
2727         /* Check if there is eviction in progress, and if so, wait for
2728          * it to finish */
2729         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2730                 *lwi = LWI_INTR(NULL, NULL);
2731                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2732                                   !cfs_atomic_read(&exp->exp_obd->
2733                                                    obd_evict_inprogress),
2734                                   lwi);
2735         }
2736
2737         /* Check if client was evicted or tried to reconnect already */
2738         if (exp->exp_failed || exp->exp_abort_active_req) {
2739                 rc = -ENOTCONN;
2740         } else {
2741                 if (desc->bd_type == BULK_PUT_SINK)
2742                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2743                 if (rc == 0)
2744                         rc = ptlrpc_start_bulk_transfer(desc);
2745         }
2746
2747         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2748                 ptlrpc_abort_bulk(desc);
2749         } else if (rc == 0) {
2750                 time_t start = cfs_time_current_sec();
2751                 do {
2752                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2753                         cfs_duration_t timeout = timeoutl <= 0 ?
2754                                 CFS_TICK : cfs_time_seconds(timeoutl);
2755                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2756                                                     cfs_time_seconds(1),
2757                                                    target_bulk_timeout,
2758                                                    desc);
2759                         rc = l_wait_event(desc->bd_waitq,
2760                                           !ptlrpc_server_bulk_active(desc) ||
2761                                           exp->exp_failed ||
2762                                           exp->exp_abort_active_req,
2763                                           lwi);
2764                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2765                         /* Wait again if we changed deadline */
2766                 } while ((rc == -ETIMEDOUT) &&
2767                          (req->rq_deadline > cfs_time_current_sec()));
2768
2769                 if (rc == -ETIMEDOUT) {
2770                         DEBUG_REQ(D_ERROR, req,
2771                                   "timeout on bulk %s after %ld%+lds",
2772                                   bulk2type(desc),
2773                                   req->rq_deadline - start,
2774                                   cfs_time_current_sec() -
2775                                   req->rq_deadline);
2776                         ptlrpc_abort_bulk(desc);
2777                 } else if (exp->exp_failed) {
2778                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2779                                   bulk2type(desc));
2780                         rc = -ENOTCONN;
2781                         ptlrpc_abort_bulk(desc);
2782                 } else if (exp->exp_abort_active_req) {
2783                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2784                                   bulk2type(desc));
2785                         /* we don't reply anyway */
2786                         rc = -ETIMEDOUT;
2787                         ptlrpc_abort_bulk(desc);
2788                 } else if (!desc->bd_success ||
2789                            desc->bd_nob_transferred != desc->bd_nob) {
2790                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2791                                   desc->bd_success ?
2792                                   "truncated" : "network error on",
2793                                   bulk2type(desc),
2794                                   desc->bd_nob_transferred,
2795                                   desc->bd_nob);
2796                         /* XXX should this be a different errno? */
2797                         rc = -ETIMEDOUT;
2798                 } else if (desc->bd_type == BULK_GET_SINK) {
2799                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2800                 }
2801         } else {
2802                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2803                           bulk2type(desc), rc);
2804         }
2805
2806         RETURN(rc);
2807 }
2808 EXPORT_SYMBOL(target_bulk_io);
2809
2810 #endif /* HAVE_SERVER_SUPPORT */