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) 2010, 2014, Intel Corporation.
  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 /**
  38  * This file deals with various client/target related logic including recovery.
  39  *
  40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  41  * should be moved.
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LDLM
  45
  46 #include <linux/kthread.h>
  47 #include <libcfs/libcfs.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_dlm.h>
  51 #include <lustre_net.h>
  52 #include <lustre_sec.h>
  53 #include "ldlm_internal.h"
  54
  55 /* @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         int rc = 0;
  64         ENTRY;
  65
  66         if (!create && !priority) {
  67                 CDEBUG(D_HA, "Nothing to do\n");
  68                 RETURN(-EINVAL);
  69         }
  70
  71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  72         if (!ptlrpc_conn) {
  73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  74                 RETURN (-ENOENT);
  75         }
  76
  77         if (create) {
  78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  79                 if (!imp_conn) {
  80                         GOTO(out_put, rc = -ENOMEM);
  81                 }
  82         }
  83
  84         spin_lock(&imp->imp_lock);
  85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  87                         if (priority) {
  88                                 list_del(&item->oic_item);
  89                                 list_add(&item->oic_item,
  90                                              &imp->imp_conn_list);
  91                                 item->oic_last_attempt = 0;
  92                         }
  93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  94                                imp, imp->imp_obd->obd_name, uuid->uuid,
  95                                (priority ? ", moved to head" : ""));
  96                         spin_unlock(&imp->imp_lock);
  97                         GOTO(out_free, rc = 0);
  98                 }
  99         }
 100         /* No existing import connection found for \a uuid. */
 101         if (create) {
 102                 imp_conn->oic_conn = ptlrpc_conn;
 103                 imp_conn->oic_uuid = *uuid;
 104                 imp_conn->oic_last_attempt = 0;
 105                 if (priority)
 106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 else
 108                         list_add_tail(&imp_conn->oic_item,
 109                                           &imp->imp_conn_list);
 110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 111                        imp, imp->imp_obd->obd_name, uuid->uuid,
 112                        (priority ? "head" : "tail"));
 113         } else {
 114                 spin_unlock(&imp->imp_lock);
 115                 GOTO(out_free, rc = -ENOENT);
 116         }
 117
 118         spin_unlock(&imp->imp_lock);
 119         RETURN(0);
 120 out_free:
 121         if (imp_conn)
 122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 123 out_put:
 124         ptlrpc_connection_put(ptlrpc_conn);
 125         RETURN(rc);
 126 }
 127
 128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 129 {
 130         return import_set_conn(imp, uuid, 1, 0);
 131 }
 132
 133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 134                            int priority)
 135 {
 136         return import_set_conn(imp, uuid, priority, 1);
 137 }
 138 EXPORT_SYMBOL(client_import_add_conn);
 139
 140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 141 {
 142         struct obd_import_conn *imp_conn;
 143         struct obd_export *dlmexp;
 144         int rc = -ENOENT;
 145         ENTRY;
 146
 147         spin_lock(&imp->imp_lock);
 148         if (list_empty(&imp->imp_conn_list)) {
 149                 LASSERT(!imp->imp_connection);
 150                 GOTO(out, rc);
 151         }
 152
 153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 155                         continue;
 156                 LASSERT(imp_conn->oic_conn);
 157
 158                 if (imp_conn == imp->imp_conn_current) {
 159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 160
 161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 162                             imp->imp_state != LUSTRE_IMP_DISCON) {
 163                                 CERROR("can't remove current connection\n");
 164                                 GOTO(out, rc = -EBUSY);
 165                         }
 166
 167                         ptlrpc_connection_put(imp->imp_connection);
 168                         imp->imp_connection = NULL;
 169
 170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 171                         if (dlmexp && dlmexp->exp_connection) {
 172                                 LASSERT(dlmexp->exp_connection ==
 173                                         imp_conn->oic_conn);
 174                                 ptlrpc_connection_put(dlmexp->exp_connection);
 175                                 dlmexp->exp_connection = NULL;
 176                         }
 177                 }
 178
 179                 list_del(&imp_conn->oic_item);
 180                 ptlrpc_connection_put(imp_conn->oic_conn);
 181                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 182                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 183                        imp, imp->imp_obd->obd_name, uuid->uuid);
 184                 rc = 0;
 185                 break;
 186         }
 187 out:
 188         spin_unlock(&imp->imp_lock);
 189         if (rc == -ENOENT)
 190                 CERROR("connection %s not found\n", uuid->uuid);
 191         RETURN(rc);
 192 }
 193 EXPORT_SYMBOL(client_import_del_conn);
 194
 195 /**
 196  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 197  * to find a conn uuid of \a imp which can reach \a peer.
 198  */
 199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 200                             struct obd_uuid *uuid)
 201 {
 202         struct obd_import_conn *conn;
 203         int rc = -ENOENT;
 204         ENTRY;
 205
 206         spin_lock(&imp->imp_lock);
 207         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 208                 /* Check if conn UUID does have this peer NID. */
 209                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 210                         *uuid = conn->oic_uuid;
 211                         rc = 0;
 212                         break;
 213                 }
 214         }
 215         spin_unlock(&imp->imp_lock);
 216         RETURN(rc);
 217 }
 218 EXPORT_SYMBOL(client_import_find_conn);
 219
 220 void client_destroy_import(struct obd_import *imp)
 221 {
 222         /* Drop security policy instance after all RPCs have finished/aborted
 223          * to let all busy contexts be released. */
 224         class_import_get(imp);
 225         class_destroy_import(imp);
 226         sptlrpc_import_sec_put(imp);
 227         class_import_put(imp);
 228 }
 229 EXPORT_SYMBOL(client_destroy_import);
 230
 231 /**
 232  * Check whether or not the OSC is on MDT.
 233  * In the config log,
 234  * osc on MDT
 235  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 236  * osc on client
 237  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 238  *
 239  **/
 240 static int osc_on_mdt(char *obdname)
 241 {
 242         char *ptr;
 243
 244         ptr = strrchr(obdname, '-');
 245         if (ptr == NULL)
 246                 return 0;
 247
 248         if (strncmp(ptr + 1, "MDT", 3) == 0)
 249                 return 1;
 250
 251         return 0;
 252 }
 253
 254 /* Configure an RPC client OBD device.
 255  *
 256  * lcfg parameters:
 257  * 1 - client UUID
 258  * 2 - server UUID
 259  * 3 - inactive-on-startup
 260  */
 261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 262 {
 263         struct client_obd *cli = &obddev->u.cli;
 264         struct obd_import *imp;
 265         struct obd_uuid server_uuid;
 266         int rq_portal, rp_portal, connect_op;
 267         char *name = obddev->obd_type->typ_name;
 268         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 269         int rc;
 270         ENTRY;
 271
 272         /* In a more perfect world, we would hang a ptlrpc_client off of
 273          * obd_type and just use the values from there. */
 274         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 275                 rq_portal = OST_REQUEST_PORTAL;
 276                 rp_portal = OSC_REPLY_PORTAL;
 277                 connect_op = OST_CONNECT;
 278                 cli->cl_sp_me = LUSTRE_SP_CLI;
 279                 cli->cl_sp_to = LUSTRE_SP_OST;
 280                 ns_type = LDLM_NS_TYPE_OSC;
 281         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 282                    !strcmp(name, LUSTRE_LWP_NAME)) {
 283                 rq_portal = MDS_REQUEST_PORTAL;
 284                 rp_portal = MDC_REPLY_PORTAL;
 285                 connect_op = MDS_CONNECT;
 286                 cli->cl_sp_me = LUSTRE_SP_CLI;
 287                 cli->cl_sp_to = LUSTRE_SP_MDT;
 288                 ns_type = LDLM_NS_TYPE_MDC;
 289         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 290                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 291                         /* OSP_on_MDT for other MDTs */
 292                         connect_op = MDS_CONNECT;
 293                         cli->cl_sp_to = LUSTRE_SP_MDT;
 294                         ns_type = LDLM_NS_TYPE_MDC;
 295                         rq_portal = OUT_PORTAL;
 296                 } else {
 297                         /* OSP on MDT for OST */
 298                         connect_op = OST_CONNECT;
 299                         cli->cl_sp_to = LUSTRE_SP_OST;
 300                         ns_type = LDLM_NS_TYPE_OSC;
 301                         rq_portal = OST_REQUEST_PORTAL;
 302                 }
 303                 rp_portal = OSC_REPLY_PORTAL;
 304                 cli->cl_sp_me = LUSTRE_SP_CLI;
 305         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 306                 rq_portal = MGS_REQUEST_PORTAL;
 307                 rp_portal = MGC_REPLY_PORTAL;
 308                 connect_op = MGS_CONNECT;
 309                 cli->cl_sp_me = LUSTRE_SP_MGC;
 310                 cli->cl_sp_to = LUSTRE_SP_MGS;
 311                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 312                 ns_type = LDLM_NS_TYPE_MGC;
 313         } else {
 314                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 315                        name);
 316                 RETURN(-EINVAL);
 317         }
 318
 319         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 320                 CERROR("requires a TARGET UUID\n");
 321                 RETURN(-EINVAL);
 322         }
 323
 324         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 325                 CERROR("client UUID must be less than 38 characters\n");
 326                 RETURN(-EINVAL);
 327         }
 328
 329         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 330                 CERROR("setup requires a SERVER UUID\n");
 331                 RETURN(-EINVAL);
 332         }
 333
 334         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 335                 CERROR("target UUID must be less than 38 characters\n");
 336                 RETURN(-EINVAL);
 337         }
 338
 339         init_rwsem(&cli->cl_sem);
 340         mutex_init(&cli->cl_mgc_mutex);
 341         cli->cl_conn_count = 0;
 342         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 343                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 344                      sizeof(server_uuid)));
 345
 346         cli->cl_dirty_pages = 0;
 347         cli->cl_avail_grant = 0;
 348         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 349         /* cl_dirty_max_pages may be changed at connect time in
 350          * ptlrpc_connect_interpret(). */
 351         client_adjust_max_dirty(cli);
 352         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 353         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 354         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 355         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 356         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 357         spin_lock_init(&cli->cl_loi_list_lock);
 358         atomic_set(&cli->cl_pending_w_pages, 0);
 359         atomic_set(&cli->cl_pending_r_pages, 0);
 360         cli->cl_r_in_flight = 0;
 361         cli->cl_w_in_flight = 0;
 362
 363         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 364         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 365         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 366         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 367         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 368         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 369
 370         /* lru for osc. */
 371         INIT_LIST_HEAD(&cli->cl_lru_osc);
 372         atomic_set(&cli->cl_lru_shrinkers, 0);
 373         atomic_long_set(&cli->cl_lru_busy, 0);
 374         atomic_long_set(&cli->cl_lru_in_list, 0);
 375         INIT_LIST_HEAD(&cli->cl_lru_list);
 376         spin_lock_init(&cli->cl_lru_list_lock);
 377         atomic_long_set(&cli->cl_unstable_count, 0);
 378
 379         init_waitqueue_head(&cli->cl_destroy_waitq);
 380         atomic_set(&cli->cl_destroy_in_flight, 0);
 381 #ifdef ENABLE_CHECKSUM
 382         /* Turn on checksumming by default. */
 383         cli->cl_checksum = 1;
 384         /*
 385          * The supported checksum types will be worked out at connect time
 386          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 387          * through procfs.
 388          */
 389         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 390 #endif
 391         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 392
 393         /* This value may be reduced at connect time in
 394          * ptlrpc_connect_interpret() . We initialize it to only
 395          * 1MB until we know what the performance looks like.
 396          * In the future this should likely be increased. LU-1431 */
 397         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
 398                                           LNET_MTU >> PAGE_CACHE_SHIFT);
 399
 400         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 401          * it will be updated at OSC connection time. */
 402         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 403
 404         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 405                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 406         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 407                 cli->cl_max_rpcs_in_flight = 2;
 408         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 409                 cli->cl_max_rpcs_in_flight = 3;
 410         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 411                 cli->cl_max_rpcs_in_flight = 4;
 412         } else {
 413                 if (osc_on_mdt(obddev->obd_name))
 414                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 415                 else
 416                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 417         }
 418         rc = ldlm_get_ref();
 419         if (rc) {
 420                 CERROR("ldlm_get_ref failed: %d\n", rc);
 421                 GOTO(err, rc);
 422         }
 423
 424         ptlrpc_init_client(rq_portal, rp_portal, name,
 425                            &obddev->obd_ldlm_client);
 426
 427         imp = class_new_import(obddev);
 428         if (imp == NULL)
 429                 GOTO(err_ldlm, rc = -ENOENT);
 430         imp->imp_client = &obddev->obd_ldlm_client;
 431         imp->imp_connect_op = connect_op;
 432         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 433                LUSTRE_CFG_BUFLEN(lcfg, 1));
 434         class_import_put(imp);
 435
 436         rc = client_import_add_conn(imp, &server_uuid, 1);
 437         if (rc) {
 438                 CERROR("can't add initial connection\n");
 439                 GOTO(err_import, rc);
 440         }
 441
 442         cli->cl_import = imp;
 443         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 444         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 445
 446         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 447                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 448                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 449                                name, obddev->obd_name,
 450                                cli->cl_target_uuid.uuid);
 451                         spin_lock(&imp->imp_lock);
 452                         imp->imp_deactive = 1;
 453                         spin_unlock(&imp->imp_lock);
 454                 }
 455         }
 456
 457         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 458                                                    LDLM_NAMESPACE_CLIENT,
 459                                                    LDLM_NAMESPACE_GREEDY,
 460                                                    ns_type);
 461         if (obddev->obd_namespace == NULL) {
 462                 CERROR("Unable to create client namespace - %s\n",
 463                        obddev->obd_name);
 464                 GOTO(err_import, rc = -ENOMEM);
 465         }
 466
 467         RETURN(rc);
 468
 469 err_import:
 470         class_destroy_import(imp);
 471 err_ldlm:
 472         ldlm_put_ref();
 473 err:
 474         RETURN(rc);
 475
 476 }
 477 EXPORT_SYMBOL(client_obd_setup);
 478
 479 int client_obd_cleanup(struct obd_device *obddev)
 480 {
 481         ENTRY;
 482
 483         ldlm_namespace_free_post(obddev->obd_namespace);
 484         obddev->obd_namespace = NULL;
 485
 486         obd_cleanup_client_import(obddev);
 487         LASSERT(obddev->u.cli.cl_import == NULL);
 488
 489         ldlm_put_ref();
 490         RETURN(0);
 491 }
 492 EXPORT_SYMBOL(client_obd_cleanup);
 493
 494 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 495 int client_connect_import(const struct lu_env *env,
 496                           struct obd_export **exp,
 497                           struct obd_device *obd, struct obd_uuid *cluuid,
 498                           struct obd_connect_data *data, void *localdata)
 499 {
 500         struct client_obd       *cli    = &obd->u.cli;
 501         struct obd_import       *imp    = cli->cl_import;
 502         struct obd_connect_data *ocd;
 503         struct lustre_handle    conn    = { 0 };
 504         int                     rc;
 505         ENTRY;
 506
 507         *exp = NULL;
 508         down_write(&cli->cl_sem);
 509         if (cli->cl_conn_count > 0)
 510                 GOTO(out_sem, rc = -EALREADY);
 511
 512         rc = class_connect(&conn, obd, cluuid);
 513         if (rc)
 514                 GOTO(out_sem, rc);
 515
 516         cli->cl_conn_count++;
 517         *exp = class_conn2export(&conn);
 518
 519         LASSERT(obd->obd_namespace);
 520
 521         imp->imp_dlm_handle = conn;
 522         rc = ptlrpc_init_import(imp);
 523         if (rc != 0)
 524                 GOTO(out_ldlm, rc);
 525
 526         ocd = &imp->imp_connect_data;
 527         if (data) {
 528                 *ocd = *data;
 529                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 530         }
 531
 532         rc = ptlrpc_connect_import(imp);
 533         if (rc != 0) {
 534                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 535                 GOTO(out_ldlm, rc);
 536         }
 537         LASSERT(*exp != NULL && (*exp)->exp_connection);
 538
 539         if (data) {
 540                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 541                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 542                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 543                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 544         }
 545
 546         ptlrpc_pinger_add_import(imp);
 547
 548         EXIT;
 549
 550         if (rc) {
 551 out_ldlm:
 552                 cli->cl_conn_count--;
 553                 class_disconnect(*exp);
 554                 *exp = NULL;
 555         }
 556 out_sem:
 557         up_write(&cli->cl_sem);
 558
 559         return rc;
 560 }
 561 EXPORT_SYMBOL(client_connect_import);
 562
 563 int client_disconnect_export(struct obd_export *exp)
 564 {
 565         struct obd_device *obd = class_exp2obd(exp);
 566         struct client_obd *cli;
 567         struct obd_import *imp;
 568         int rc = 0, err;
 569         ENTRY;
 570
 571         if (!obd) {
 572                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 573                        exp, exp ? exp->exp_handle.h_cookie : -1);
 574                 RETURN(-EINVAL);
 575         }
 576
 577         cli = &obd->u.cli;
 578         imp = cli->cl_import;
 579
 580         down_write(&cli->cl_sem);
 581         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 582                 cli->cl_conn_count);
 583
 584         if (cli->cl_conn_count == 0) {
 585                 CERROR("disconnecting disconnected device (%s)\n",
 586                        obd->obd_name);
 587                 GOTO(out_disconnect, rc = -EINVAL);
 588         }
 589
 590         cli->cl_conn_count--;
 591         if (cli->cl_conn_count != 0)
 592                 GOTO(out_disconnect, rc = 0);
 593
 594         /* Mark import deactivated now, so we don't try to reconnect if any
 595          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 596          * fully deactivate the import, or that would drop all requests. */
 597         spin_lock(&imp->imp_lock);
 598         imp->imp_deactive = 1;
 599         spin_unlock(&imp->imp_lock);
 600
 601         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 602          * delete it regardless.  (It's safe to delete an import that was
 603          * never added.) */
 604         (void)ptlrpc_pinger_del_import(imp);
 605
 606         if (obd->obd_namespace != NULL) {
 607                 /* obd_force == local only */
 608                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 609                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 610                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 611         }
 612
 613         /* There's no need to hold sem while disconnecting an import,
 614          * and it may actually cause deadlock in GSS. */
 615         up_write(&cli->cl_sem);
 616         rc = ptlrpc_disconnect_import(imp, 0);
 617         down_write(&cli->cl_sem);
 618
 619         ptlrpc_invalidate_import(imp);
 620
 621         EXIT;
 622
 623 out_disconnect:
 624         /* Use server style - class_disconnect should be always called for
 625          * o_disconnect. */
 626         err = class_disconnect(exp);
 627         if (!rc && err)
 628                 rc = err;
 629
 630         up_write(&cli->cl_sem);
 631
 632         RETURN(rc);
 633 }
 634 EXPORT_SYMBOL(client_disconnect_export);
 635
 636 #ifdef HAVE_SERVER_SUPPORT
 637 int server_disconnect_export(struct obd_export *exp)
 638 {
 639         int rc;
 640         ENTRY;
 641
 642         /* Disconnect early so that clients can't keep using export. */
 643         rc = class_disconnect(exp);
 644         /* Close import to avoid sending any requests. */
 645         if (exp->exp_imp_reverse)
 646                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 647
 648         if (exp->exp_obd->obd_namespace != NULL)
 649                 ldlm_cancel_locks_for_export(exp);
 650
 651         /* complete all outstanding replies */
 652         spin_lock(&exp->exp_lock);
 653         while (!list_empty(&exp->exp_outstanding_replies)) {
 654                 struct ptlrpc_reply_state *rs =
 655                         list_entry(exp->exp_outstanding_replies.next,
 656                                        struct ptlrpc_reply_state, rs_exp_list);
 657                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 658
 659                 spin_lock(&svcpt->scp_rep_lock);
 660
 661                 list_del_init(&rs->rs_exp_list);
 662                 spin_lock(&rs->rs_lock);
 663                 ptlrpc_schedule_difficult_reply(rs);
 664                 spin_unlock(&rs->rs_lock);
 665
 666                 spin_unlock(&svcpt->scp_rep_lock);
 667         }
 668         spin_unlock(&exp->exp_lock);
 669
 670         RETURN(rc);
 671 }
 672 EXPORT_SYMBOL(server_disconnect_export);
 673
 674 /* --------------------------------------------------------------------------
 675  * from old lib/target.c
 676  * -------------------------------------------------------------------------- */
 677
 678 static int target_handle_reconnect(struct lustre_handle *conn,
 679                                    struct obd_export *exp,
 680                                    struct obd_uuid *cluuid)
 681 {
 682         ENTRY;
 683
 684         if (exp->exp_connection && exp->exp_imp_reverse) {
 685                 struct lustre_handle *hdl;
 686                 struct obd_device *target;
 687
 688                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 689                 target = exp->exp_obd;
 690
 691                 /* Might be a re-connect after a partition. */
 692                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 693                         if (target->obd_recovering) {
 694                                 int timeout = cfs_duration_sec(cfs_time_sub(
 695                                         cfs_timer_deadline(
 696                                         &target->obd_recovery_timer),
 697                                         cfs_time_current()));
 698
 699                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 700                                         "ing, waiting for %d clients in recov"
 701                                         "ery for %d:%.02d\n", target->obd_name,
 702                                         obd_uuid2str(&exp->exp_client_uuid),
 703                                         obd_export_nid2str(exp),
 704                                         target->obd_max_recoverable_clients,
 705                                         timeout / 60, timeout % 60);
 706                         } else {
 707                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 708                                         "reconnecting\n", target->obd_name,
 709                                         obd_uuid2str(&exp->exp_client_uuid),
 710                                         obd_export_nid2str(exp));
 711                         }
 712
 713                         conn->cookie = exp->exp_handle.h_cookie;
 714                         /* target_handle_connect() treats EALREADY and
 715                          * -EALREADY differently.  EALREADY means we are
 716                          * doing a valid reconnect from the same client. */
 717                         RETURN(EALREADY);
 718                 } else {
 719                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 720                                       "with handle "LPX64". Rejecting client "
 721                                       "with the same UUID trying to reconnect "
 722                                       "with handle "LPX64"\n", target->obd_name,
 723                                       obd_uuid2str(&exp->exp_client_uuid),
 724                                       obd_export_nid2str(exp),
 725                                       hdl->cookie, conn->cookie);
 726                         memset(conn, 0, sizeof *conn);
 727                         /* target_handle_connect() treats EALREADY and
 728                          * -EALREADY differently.  -EALREADY is an error
 729                          * (same UUID, different handle). */
 730                         RETURN(-EALREADY);
 731                 }
 732         }
 733
 734         conn->cookie = exp->exp_handle.h_cookie;
 735         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 736                cluuid->uuid, exp, conn->cookie);
 737         RETURN(0);
 738 }
 739
 740 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 741                           int error)
 742 {
 743         struct obd_export *exp = cb_data;
 744
 745         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 746                obd->obd_name, exp->exp_client_uuid.uuid);
 747
 748         spin_lock(&exp->exp_lock);
 749         exp->exp_need_sync = 0;
 750         spin_unlock(&exp->exp_lock);
 751         class_export_cb_put(exp);
 752 }
 753
 754 static void
 755 check_and_start_recovery_timer(struct obd_device *obd,
 756                                struct ptlrpc_request *req, int new_client);
 757
 758 int target_handle_connect(struct ptlrpc_request *req)
 759 {
 760         struct obd_device *target = NULL, *targref = NULL;
 761         struct obd_export *export = NULL;
 762         struct obd_import *revimp;
 763         struct obd_import *tmp_imp = NULL;
 764         struct lustre_handle conn;
 765         struct lustre_handle *tmp;
 766         struct obd_uuid tgtuuid;
 767         struct obd_uuid cluuid;
 768         struct obd_uuid remote_uuid;
 769         char *str;
 770         int rc = 0;
 771         char *target_start;
 772         int target_len;
 773         bool     mds_conn = false, lw_client = false;
 774         bool     mds_mds_conn = false;
 775         bool     new_mds_mds_conn = false;
 776         struct obd_connect_data *data, *tmpdata;
 777         int size, tmpsize;
 778         lnet_nid_t *client_nid = NULL;
 779         ENTRY;
 780
 781         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 782
 783         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 784         if (str == NULL) {
 785                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 786                 GOTO(out, rc = -EINVAL);
 787         }
 788
 789         obd_str2uuid(&tgtuuid, str);
 790         target = class_uuid2obd(&tgtuuid);
 791         if (!target)
 792                 target = class_name2obd(str);
 793
 794         if (!target) {
 795                 deuuidify(str, NULL, &target_start, &target_len);
 796                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 797                                    "from %s (no target). If you are running "
 798                                    "an HA pair check that the target is "
 799                                    "mounted on the other server.\n", str,
 800                                    libcfs_nid2str(req->rq_peer.nid));
 801                 GOTO(out, rc = -ENODEV);
 802         }
 803
 804         spin_lock(&target->obd_dev_lock);
 805         if (target->obd_stopping || !target->obd_set_up) {
 806                 spin_unlock(&target->obd_dev_lock);
 807
 808                 deuuidify(str, NULL, &target_start, &target_len);
 809                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 810                               target_len, target_start,
 811                               libcfs_nid2str(req->rq_peer.nid),
 812                               (target->obd_stopping ?
 813                                "stopping" : "not set up"));
 814                 GOTO(out, rc = -ENODEV);
 815         }
 816
 817         if (target->obd_no_conn) {
 818                 spin_unlock(&target->obd_dev_lock);
 819
 820                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 821                                "from %s\n", target->obd_name,
 822                                libcfs_nid2str(req->rq_peer.nid));
 823                 GOTO(out, rc = -EAGAIN);
 824         }
 825
 826         /* Make sure the target isn't cleaned up while we're here. Yes,
 827          * there's still a race between the above check and our incref here.
 828          * Really, class_uuid2obd should take the ref. */
 829         targref = class_incref(target, __FUNCTION__, current);
 830
 831         target->obd_conn_inprogress++;
 832         spin_unlock(&target->obd_dev_lock);
 833
 834         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 835         if (str == NULL) {
 836                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 837                 GOTO(out, rc = -EINVAL);
 838         }
 839
 840         obd_str2uuid(&cluuid, str);
 841
 842         /* XXX Extract a nettype and format accordingly. */
 843         switch (sizeof(lnet_nid_t)) {
 844         /* NB the casts only avoid compiler warnings. */
 845         case 8:
 846                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 847                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 848                 break;
 849         case 4:
 850                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 851                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 852                 break;
 853         default:
 854                 LBUG();
 855         }
 856
 857         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 858         if (tmp == NULL)
 859                 GOTO(out, rc = -EPROTO);
 860
 861         conn = *tmp;
 862
 863         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 864                                     RCL_CLIENT);
 865         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 866         if (!data)
 867                 GOTO(out, rc = -EPROTO);
 868
 869         rc = req_capsule_server_pack(&req->rq_pill);
 870         if (rc)
 871                 GOTO(out, rc);
 872
 873 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 874         /* Don't allow clients to connect that are using old 1.8 format
 875          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
 876          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
 877          * FULL20 flag should be set on all connections since 2.0, but no
 878          * longer affects behaviour.
 879          *
 880          * Later this check will be disabled and the flag can be retired
 881          * completely once interop with 3.0 is no longer needed.
 882          */
 883         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
 884                 GOTO(out, rc = -EPROTO);
 885 #endif
 886
 887         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 888                 if (data->ocd_version < LUSTRE_VERSION_CODE -
 889                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 890                     data->ocd_version > LUSTRE_VERSION_CODE +
 891                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 892                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 893                                   "libclient connection attempt",
 894                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 895                                   "old" : "new",
 896                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 897                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 898                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 899                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 900                         data = req_capsule_server_sized_get(&req->rq_pill,
 901                                                             &RMF_CONNECT_DATA,
 902                                     offsetof(typeof(*data), ocd_version) +
 903                                              sizeof(data->ocd_version));
 904                         if (data) {
 905                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 906                                 data->ocd_version = LUSTRE_VERSION_CODE;
 907                         }
 908                         GOTO(out, rc = -EPROTO);
 909                 }
 910         }
 911
 912         /* Note: lw_client is needed in MDS-MDS failover during update log
 913          * processing, so we needs to allow lw_client to be connected at
 914          * anytime, instead of only the initial connection */
 915         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
 916
 917         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
 918                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
 919                 mds_mds_conn = (data->ocd_connect_flags &
 920                                 OBD_CONNECT_MDS_MDS) != 0;
 921
 922                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
 923                  * for Imperative Recovery connection from MGC to MGS.
 924                  *
 925                  * Via check OBD_CONNECT_FID, we can distinguish whether
 926                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
 927                  * MGC or MDT. */
 928                 if (!lw_client &&
 929                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
 930                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
 931                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
 932                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
 933                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
 934                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
 935
 936                         /* We do not support the MDT-MDT interoperations with
 937                          * different version MDT because of protocol changes. */
 938                         if (unlikely(major != LUSTRE_MAJOR ||
 939                                      minor != LUSTRE_MINOR ||
 940                                      abs(patch - LUSTRE_PATCH) > 3)) {
 941                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
 942                                         "connection from different version MDT "
 943                                         "(%d.%d.%d.%d) %s %s\n",
 944                                         target->obd_name, LUSTRE_MAJOR,
 945                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
 946                                         major, minor, patch,
 947                                         OBD_OCD_VERSION_FIX(data->ocd_version),
 948                                         libcfs_nid2str(req->rq_peer.nid), str);
 949
 950                                 GOTO(out, rc = -EPROTO);
 951                         }
 952                 }
 953         }
 954
 955         /* lctl gets a backstage, all-access pass. */
 956         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 957                 goto dont_check_exports;
 958
 959         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 960         if (!export)
 961                 goto no_export;
 962
 963         /* We've found an export in the hash. */
 964
 965         spin_lock(&export->exp_lock);
 966
 967         if (export->exp_connecting) { /* bug 9635, et. al. */
 968                 spin_unlock(&export->exp_lock);
 969                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 970                               export->exp_obd->obd_name, export,
 971                               libcfs_nid2str(req->rq_peer.nid));
 972                 class_export_put(export);
 973                 export = NULL;
 974                 rc = -EALREADY;
 975         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
 976                 spin_unlock(&export->exp_lock);
 977                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 978                         /* MDS or LWP reconnected after failover. */
 979                         LCONSOLE_WARN("%s: Received %s connection from "
 980                             "%s, removing former export from %s\n",
 981                             target->obd_name, mds_conn ? "MDS" : "LWP",
 982                             libcfs_nid2str(req->rq_peer.nid),
 983                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 984                 else
 985                         /* New MDS connection from the same NID. */
 986                         LCONSOLE_WARN("%s: Received new %s connection from "
 987                                 "%s, removing former export from same NID\n",
 988                                 target->obd_name, mds_conn ? "MDS" : "LWP",
 989                                 libcfs_nid2str(req->rq_peer.nid));
 990                 class_fail_export(export);
 991                 class_export_put(export);
 992                 export = NULL;
 993                 rc = 0;
 994         } else if (export->exp_connection != NULL &&
 995                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 996                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 997                     MSG_CONNECT_INITIAL)) {
 998                 spin_unlock(&export->exp_lock);
 999                 /* In MDS failover we have static UUID but NID can change. */
1000                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1001                               "existing nid %s is already connected\n",
1002                               target->obd_name, cluuid.uuid,
1003                               libcfs_nid2str(req->rq_peer.nid),
1004                               libcfs_nid2str(
1005                                       export->exp_connection->c_peer.nid));
1006                 rc = -EALREADY;
1007                 class_export_put(export);
1008                 export = NULL;
1009         } else {
1010                 export->exp_connecting = 1;
1011                 spin_unlock(&export->exp_lock);
1012                 LASSERT(export->exp_obd == target);
1013
1014                 rc = target_handle_reconnect(&conn, export, &cluuid);
1015         }
1016
1017         /* If we found an export, we already unlocked. */
1018         if (!export) {
1019 no_export:
1020                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1021         } else if (req->rq_export == NULL &&
1022                    atomic_read(&export->exp_rpc_count) > 0) {
1023                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1024                               "still busy with %d references\n",
1025                               target->obd_name, cluuid.uuid,
1026                               libcfs_nid2str(req->rq_peer.nid),
1027                               atomic_read(&export->exp_refcount));
1028                 GOTO(out, rc = -EBUSY);
1029         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1030                 if (!strstr(cluuid.uuid, "mdt"))
1031                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1032                                       "known client %s (at %s) because it "
1033                                       "is indicating it is a new client",
1034                                       target->obd_name, cluuid.uuid,
1035                                       libcfs_nid2str(req->rq_peer.nid));
1036                 GOTO(out, rc = -EALREADY);
1037         } else {
1038                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1039         }
1040
1041         if (rc < 0) {
1042                 GOTO(out, rc);
1043         }
1044
1045         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1046                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1047               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1048               export, (long)cfs_time_current_sec(),
1049               export ? (long)export->exp_last_request_time : 0);
1050
1051         /* If this is the first time a client connects, reset the recovery
1052          * timer. Discard lightweight connections which might be local. */
1053         if (!lw_client && rc == 0 && target->obd_recovering)
1054                 check_and_start_recovery_timer(target, req, export == NULL);
1055
1056         /* We want to handle EALREADY but *not* -EALREADY from
1057          * target_handle_reconnect(), return reconnection state in a flag. */
1058         if (rc == EALREADY) {
1059                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1060                 rc = 0;
1061         } else {
1062                 LASSERT(rc == 0);
1063         }
1064
1065         /* Tell the client if we support replayable requests. */
1066         if (target->obd_replayable)
1067                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1068         client_nid = &req->rq_peer.nid;
1069
1070         if (export == NULL) {
1071                 /* allow lightweight connections during recovery */
1072                 /* allow "new" MDT to be connected during recovery, since we
1073                  * need retrieve recovery update records from it */
1074                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1075                         cfs_time_t t;
1076                         int     c; /* connected */
1077                         int     i; /* in progress */
1078                         int     k; /* known */
1079                         int     s; /* stale/evicted */
1080
1081                         c = atomic_read(&target->obd_connected_clients);
1082                         i = atomic_read(&target->obd_lock_replay_clients);
1083                         k = target->obd_max_recoverable_clients;
1084                         s = target->obd_stale_clients;
1085                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1086                         t = cfs_time_sub(t, cfs_time_current());
1087                         t = cfs_duration_sec(t);
1088                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1089                                       "(at %s), waiting for %d known clients "
1090                                       "(%d recovered, %d in progress, and %d "
1091                                       "evicted) to recover in %d:%.02d\n",
1092                                       target->obd_name, cluuid.uuid,
1093                                       libcfs_nid2str(req->rq_peer.nid), k,
1094                                       c - i, i, s, (int)t / 60,
1095                                       (int)t % 60);
1096                         rc = -EBUSY;
1097                 } else {
1098 dont_check_exports:
1099                         rc = obd_connect(req->rq_svc_thread->t_env,
1100                                          &export, target, &cluuid, data,
1101                                          client_nid);
1102                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1103                                 lustre_msg_add_op_flags(req->rq_repmsg,
1104                                                         MSG_CONNECT_RECOVERING);
1105                         if (rc == 0)
1106                                 conn.cookie = export->exp_handle.h_cookie;
1107
1108                         if (mds_mds_conn)
1109                                 new_mds_mds_conn = true;
1110                 }
1111         } else {
1112                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1113                                    export, target, &cluuid, data, client_nid);
1114         }
1115         if (rc)
1116                 GOTO(out, rc);
1117
1118         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1119         data->ocd_instance = target->u.obt.obt_instance;
1120
1121         /* Return only the parts of obd_connect_data that we understand, so the
1122          * client knows that we don't understand the rest. */
1123         if (data) {
1124                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1125                                                RCL_SERVER);
1126                 tmpdata = req_capsule_server_get(&req->rq_pill,
1127                                                  &RMF_CONNECT_DATA);
1128                 /* Don't use struct assignment here, because the client reply
1129                  * buffer may be smaller/larger than the local struct
1130                  * obd_connect_data. */
1131                 memcpy(tmpdata, data, min(tmpsize, size));
1132         }
1133
1134         /* If all else goes well, this is our RPC return code. */
1135         req->rq_status = 0;
1136
1137         lustre_msg_set_handle(req->rq_repmsg, &conn);
1138
1139         /* If the client and the server are the same node, we will already
1140          * have an export that really points to the client's DLM export,
1141          * because we have a shared handles table.
1142          *
1143          * XXX this will go away when shaver stops sending the "connect" handle
1144          * in the real "remote handle" field of the request --phik 24 Apr 2003
1145          */
1146         ptlrpc_request_change_export(req, export);
1147
1148         spin_lock(&export->exp_lock);
1149         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1150                 spin_unlock(&export->exp_lock);
1151                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1152                        "or equal conn_cnt: %d >= %d\n",
1153                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1154                        export->exp_conn_cnt,
1155                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1156
1157                 GOTO(out, rc = -EALREADY);
1158         }
1159         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1160         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1161
1162         /* Don't evict liblustre clients for not pinging. */
1163         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1164                 export->exp_libclient = 1;
1165                 spin_unlock(&export->exp_lock);
1166
1167                 spin_lock(&target->obd_dev_lock);
1168                 list_del_init(&export->exp_obd_chain_timed);
1169                 spin_unlock(&target->obd_dev_lock);
1170         } else {
1171                 spin_unlock(&export->exp_lock);
1172         }
1173
1174         if (export->exp_connection != NULL) {
1175                 /* Check to see if connection came from another NID. */
1176                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1177                     !hlist_unhashed(&export->exp_nid_hash))
1178                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1179                                      &export->exp_connection->c_peer.nid,
1180                                      &export->exp_nid_hash);
1181
1182                 ptlrpc_connection_put(export->exp_connection);
1183         }
1184
1185         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1186                                                        req->rq_self,
1187                                                        &remote_uuid);
1188         if (hlist_unhashed(&export->exp_nid_hash)) {
1189                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1190                              &export->exp_connection->c_peer.nid,
1191                              &export->exp_nid_hash);
1192         }
1193
1194         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1195                 int has_transno;
1196                 __u64 transno = data->ocd_transno;
1197
1198                 spin_lock(&export->exp_lock);
1199                 /* possible race with class_disconnect_stale_exports,
1200                  * export may be already in the eviction process */
1201                 if (export->exp_failed) {
1202                         spin_unlock(&export->exp_lock);
1203                         GOTO(out, rc = -ENODEV);
1204                 }
1205                 export->exp_in_recovery = 1;
1206                 export->exp_req_replay_needed = 1;
1207                 export->exp_lock_replay_needed = 1;
1208                 spin_unlock(&export->exp_lock);
1209
1210                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1211                                  MSG_CONNECT_TRANSNO);
1212                 if (has_transno && transno == 0)
1213                         CWARN("Connect with zero transno!\n");
1214
1215                 if (has_transno && transno > 0 &&
1216                     transno < target->obd_next_recovery_transno &&
1217                     transno > target->obd_last_committed) {
1218                         /* Another way is to use cmpxchg() to be lock-free. */
1219                         spin_lock(&target->obd_recovery_task_lock);
1220                         if (transno < target->obd_next_recovery_transno)
1221                                 target->obd_next_recovery_transno = transno;
1222                         spin_unlock(&target->obd_recovery_task_lock);
1223                 }
1224
1225                 atomic_inc(&target->obd_req_replay_clients);
1226                 atomic_inc(&target->obd_lock_replay_clients);
1227                 /* Note: MDS-MDS connection is allowed to be connected during
1228                  * recovery, no matter if the exports needs to be recoveried.
1229                  * Because we need retrieve updates logs from all other MDTs.
1230                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1231                  * also needs to be increased to match other recovery checking
1232                  * condition. */
1233                 if (new_mds_mds_conn)
1234                         target->obd_max_recoverable_clients++;
1235                 if (atomic_inc_return(&target->obd_connected_clients) ==
1236                     target->obd_max_recoverable_clients)
1237                         wake_up(&target->obd_next_transno_waitq);
1238         }
1239
1240         /* Tell the client we're in recovery, when client is involved in it. */
1241         if (target->obd_recovering && !lw_client)
1242                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1243
1244         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1245         conn = *tmp;
1246
1247         /* Return -ENOTCONN in case of errors to let client reconnect. */
1248         revimp = class_new_import(target);
1249         if (revimp == NULL) {
1250                 CERROR("fail to alloc new reverse import.\n");
1251                 GOTO(out, rc = -ENOTCONN);
1252         }
1253
1254         spin_lock(&export->exp_lock);
1255         if (export->exp_imp_reverse != NULL)
1256                 /* destroyed import can be still referenced in ctxt */
1257                 tmp_imp = export->exp_imp_reverse;
1258         export->exp_imp_reverse = revimp;
1259         spin_unlock(&export->exp_lock);
1260
1261         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1262         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1263         revimp->imp_remote_handle = conn;
1264         revimp->imp_dlm_fake = 1;
1265         revimp->imp_state = LUSTRE_IMP_FULL;
1266
1267         /* Unknown versions will be caught in
1268          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1269         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1270
1271         if (data->ocd_connect_flags & OBD_CONNECT_AT)
1272                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1273         else
1274                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1275
1276         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1277
1278         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1279         if (rc) {
1280                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1281                 spin_lock(&export->exp_lock);
1282                 export->exp_imp_reverse = NULL;
1283                 spin_unlock(&export->exp_lock);
1284                 class_destroy_import(revimp);
1285         }
1286
1287         class_import_put(revimp);
1288
1289 out:
1290         if (tmp_imp != NULL)
1291                 client_destroy_import(tmp_imp);
1292         if (export) {
1293                 spin_lock(&export->exp_lock);
1294                 export->exp_connecting = 0;
1295                 spin_unlock(&export->exp_lock);
1296
1297                 class_export_put(export);
1298         }
1299         if (targref) {
1300                 spin_lock(&target->obd_dev_lock);
1301                 target->obd_conn_inprogress--;
1302                 spin_unlock(&target->obd_dev_lock);
1303
1304                 class_decref(targref, __func__, current);
1305         }
1306         if (rc)
1307                 req->rq_status = rc;
1308         RETURN(rc);
1309 }
1310
1311 int target_handle_disconnect(struct ptlrpc_request *req)
1312 {
1313         int rc;
1314         ENTRY;
1315
1316         rc = req_capsule_server_pack(&req->rq_pill);
1317         if (rc)
1318                 RETURN(rc);
1319
1320         /* Keep the rq_export around so we can send the reply. */
1321         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1322
1323         RETURN(0);
1324 }
1325
1326 void target_destroy_export(struct obd_export *exp)
1327 {
1328         struct obd_import       *imp = NULL;
1329         /* exports created from last_rcvd data, and "fake"
1330            exports created by lctl don't have an import */
1331         spin_lock(&exp->exp_lock);
1332         if (exp->exp_imp_reverse != NULL) {
1333                 imp = exp->exp_imp_reverse;
1334                 exp->exp_imp_reverse = NULL;
1335         }
1336         spin_unlock(&exp->exp_lock);
1337         if (imp != NULL)
1338                 client_destroy_import(imp);
1339
1340         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1341         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1342         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1343         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1344 }
1345 EXPORT_SYMBOL(target_destroy_export);
1346
1347 /*
1348  * Recovery functions
1349  */
1350 static void target_request_copy_get(struct ptlrpc_request *req)
1351 {
1352         class_export_rpc_inc(req->rq_export);
1353         LASSERT(list_empty(&req->rq_list));
1354         INIT_LIST_HEAD(&req->rq_replay_list);
1355
1356         /* Increase refcount to keep request in queue. */
1357         atomic_inc(&req->rq_refcount);
1358         /* Let export know it has replays to be handled. */
1359         atomic_inc(&req->rq_export->exp_replay_count);
1360 }
1361
1362 static void target_request_copy_put(struct ptlrpc_request *req)
1363 {
1364         LASSERT(list_empty(&req->rq_replay_list));
1365         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1366
1367         atomic_dec(&req->rq_export->exp_replay_count);
1368         class_export_rpc_dec(req->rq_export);
1369         ptlrpc_server_drop_request(req);
1370 }
1371
1372 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1373 {
1374         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1375         struct obd_export     *exp = req->rq_export;
1376         struct ptlrpc_request *reqiter;
1377         int                    dup = 0;
1378
1379         LASSERT(exp);
1380
1381         spin_lock(&exp->exp_lock);
1382         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1383                                 rq_replay_list) {
1384                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1385                         dup = 1;
1386                         break;
1387                 }
1388         }
1389
1390         if (dup) {
1391                 /* We expect it with RESENT and REPLAY flags. */
1392                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1393                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1394                         CERROR("invalid flags %x of resent replay\n",
1395                                lustre_msg_get_flags(req->rq_reqmsg));
1396         } else {
1397                 list_add_tail(&req->rq_replay_list,
1398                                   &exp->exp_req_replay_queue);
1399         }
1400
1401         spin_unlock(&exp->exp_lock);
1402         return dup;
1403 }
1404
1405 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1406 {
1407         LASSERT(!list_empty(&req->rq_replay_list));
1408         LASSERT(req->rq_export);
1409
1410         spin_lock(&req->rq_export->exp_lock);
1411         list_del_init(&req->rq_replay_list);
1412         spin_unlock(&req->rq_export->exp_lock);
1413 }
1414
1415 static void target_finish_recovery(struct lu_target *lut)
1416 {
1417         struct obd_device *obd = lut->lut_obd;
1418         ENTRY;
1419
1420         /* Only log a recovery message when recovery has occurred. */
1421         if (obd->obd_recovery_start) {
1422                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1423                                         obd->obd_recovery_start);
1424                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1425                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1426                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1427                         obd->obd_max_recoverable_clients,
1428                         atomic_read(&obd->obd_connected_clients),
1429                         obd->obd_stale_clients,
1430                         obd->obd_stale_clients == 1 ? "was" : "were");
1431         }
1432
1433         ldlm_reprocess_all_ns(obd->obd_namespace);
1434         spin_lock(&obd->obd_recovery_task_lock);
1435         if (!list_empty(&obd->obd_req_replay_queue) ||
1436             !list_empty(&obd->obd_lock_replay_queue) ||
1437             !list_empty(&obd->obd_final_req_queue)) {
1438                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1439                        obd->obd_name,
1440                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1441                        list_empty(&obd->obd_lock_replay_queue) ? \
1442                                "" : "lock ",
1443                        list_empty(&obd->obd_final_req_queue) ? \
1444                                "" : "final ");
1445                 spin_unlock(&obd->obd_recovery_task_lock);
1446                 LBUG();
1447         }
1448         spin_unlock(&obd->obd_recovery_task_lock);
1449
1450         if (lut->lut_tdtd != NULL &&
1451             !list_empty(&lut->lut_tdtd->tdtd_replay_list))
1452                 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1453
1454         obd->obd_recovery_end = cfs_time_current_sec();
1455
1456         /* When recovery finished, cleanup orphans on MDS and OST. */
1457         if (OBT(obd) && OBP(obd, postrecov)) {
1458                 int rc = OBP(obd, postrecov)(obd);
1459                 if (rc < 0)
1460                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1461                                       obd->obd_name, rc);
1462         }
1463         EXIT;
1464 }
1465
1466 static void abort_req_replay_queue(struct obd_device *obd)
1467 {
1468         struct ptlrpc_request *req, *n;
1469         struct list_head abort_list;
1470
1471         INIT_LIST_HEAD(&abort_list);
1472         spin_lock(&obd->obd_recovery_task_lock);
1473         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1474         spin_unlock(&obd->obd_recovery_task_lock);
1475         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1476                 DEBUG_REQ(D_WARNING, req, "aborted:");
1477                 req->rq_status = -ENOTCONN;
1478                 if (ptlrpc_error(req)) {
1479                         DEBUG_REQ(D_ERROR, req,
1480                                   "failed abort_req_reply; skipping");
1481                 }
1482                 target_exp_dequeue_req_replay(req);
1483                 target_request_copy_put(req);
1484         }
1485 }
1486
1487 static void abort_lock_replay_queue(struct obd_device *obd)
1488 {
1489         struct ptlrpc_request *req, *n;
1490         struct list_head abort_list;
1491
1492         INIT_LIST_HEAD(&abort_list);
1493         spin_lock(&obd->obd_recovery_task_lock);
1494         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1495         spin_unlock(&obd->obd_recovery_task_lock);
1496         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1497                 DEBUG_REQ(D_ERROR, req, "aborted:");
1498                 req->rq_status = -ENOTCONN;
1499                 if (ptlrpc_error(req)) {
1500                         DEBUG_REQ(D_ERROR, req,
1501                                   "failed abort_lock_reply; skipping");
1502                 }
1503                 target_request_copy_put(req);
1504         }
1505 }
1506
1507 /* Called from a cleanup function if the device is being cleaned up
1508    forcefully.  The exports should all have been disconnected already,
1509    the only thing left to do is
1510      - clear the recovery flags
1511      - cancel the timer
1512      - free queued requests and replies, but don't send replies
1513    Because the obd_stopping flag is set, no new requests should be received.
1514
1515 */
1516 void target_cleanup_recovery(struct obd_device *obd)
1517 {
1518         struct ptlrpc_request *req, *n;
1519         struct list_head clean_list;
1520         ENTRY;
1521
1522         INIT_LIST_HEAD(&clean_list);
1523         spin_lock(&obd->obd_dev_lock);
1524         if (!obd->obd_recovering) {
1525                 spin_unlock(&obd->obd_dev_lock);
1526                 EXIT;
1527                 return;
1528         }
1529         obd->obd_recovering = obd->obd_abort_recovery = 0;
1530         obd->obd_force_abort_recovery = 0;
1531         spin_unlock(&obd->obd_dev_lock);
1532
1533         spin_lock(&obd->obd_recovery_task_lock);
1534         target_cancel_recovery_timer(obd);
1535         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1536         spin_unlock(&obd->obd_recovery_task_lock);
1537
1538         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1539                 LASSERT(req->rq_reply_state == NULL);
1540                 target_exp_dequeue_req_replay(req);
1541                 target_request_copy_put(req);
1542         }
1543
1544         spin_lock(&obd->obd_recovery_task_lock);
1545         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1546         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1547         spin_unlock(&obd->obd_recovery_task_lock);
1548
1549         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1550                 LASSERT(req->rq_reply_state == NULL);
1551                 target_request_copy_put(req);
1552         }
1553
1554         EXIT;
1555 }
1556 EXPORT_SYMBOL(target_cleanup_recovery);
1557
1558 /* obd_recovery_task_lock should be held */
1559 void target_cancel_recovery_timer(struct obd_device *obd)
1560 {
1561         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1562         cfs_timer_disarm(&obd->obd_recovery_timer);
1563 }
1564
1565 static void target_start_recovery_timer(struct obd_device *obd)
1566 {
1567         if (obd->obd_recovery_start != 0)
1568                 return;
1569
1570         spin_lock(&obd->obd_dev_lock);
1571         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1572             obd->obd_force_abort_recovery) {
1573                 spin_unlock(&obd->obd_dev_lock);
1574                 return;
1575         }
1576
1577         LASSERT(obd->obd_recovery_timeout != 0);
1578
1579         if (obd->obd_recovery_start != 0) {
1580                 spin_unlock(&obd->obd_dev_lock);
1581                 return;
1582         }
1583
1584         cfs_timer_arm(&obd->obd_recovery_timer,
1585                       cfs_time_shift(obd->obd_recovery_timeout));
1586         obd->obd_recovery_start = cfs_time_current_sec();
1587         spin_unlock(&obd->obd_dev_lock);
1588
1589         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1590                       "or until %d client%s reconnect%s\n",
1591                       obd->obd_name,
1592                       obd->obd_recovery_timeout / 60,
1593                       obd->obd_recovery_timeout % 60,
1594                       obd->obd_max_recoverable_clients,
1595                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1596                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1597 }
1598
1599 /**
1600  * extend recovery window.
1601  *
1602  * if @extend is true, extend recovery window to have @drt remaining at least;
1603  * otherwise, make sure the recovery timeout value is not less than @drt.
1604  */
1605 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1606 {
1607         cfs_time_t now;
1608         cfs_time_t end;
1609         cfs_duration_t left;
1610         int to;
1611
1612         spin_lock(&obd->obd_dev_lock);
1613         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1614             obd->obd_force_abort_recovery) {
1615                 spin_unlock(&obd->obd_dev_lock);
1616                 return;
1617         }
1618         LASSERT(obd->obd_recovery_start != 0);
1619
1620         now  = cfs_time_current_sec();
1621         to   = obd->obd_recovery_timeout;
1622         end  = obd->obd_recovery_start + to;
1623         left = cfs_time_sub(end, now);
1624
1625         if (extend && (drt > left)) {
1626                 to += drt - left;
1627         } else if (!extend && (drt > to)) {
1628                 to = drt;
1629         }
1630
1631         if (to > obd->obd_recovery_time_hard)
1632                 to = obd->obd_recovery_time_hard;
1633         if (obd->obd_recovery_timeout < to) {
1634                 obd->obd_recovery_timeout = to;
1635                 end = obd->obd_recovery_start + to;
1636                 cfs_timer_arm(&obd->obd_recovery_timer,
1637                                 cfs_time_shift(end - now));
1638         }
1639         spin_unlock(&obd->obd_dev_lock);
1640
1641         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1642                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1643 }
1644
1645 /* Reset the timer with each new client connection */
1646 /*
1647  * This timer is actually reconnect_timer, which is for making sure
1648  * the total recovery window is at least as big as my reconnect
1649  * attempt timing. So the initial recovery time_out will be set to
1650  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1651  * from client is bigger than this, then the recovery time_out will
1652  * be extended to make sure the client could be reconnected, in the
1653  * process, the timeout from the new client should be ignored.
1654  */
1655
1656 static void
1657 check_and_start_recovery_timer(struct obd_device *obd,
1658                                struct ptlrpc_request *req,
1659                                int new_client)
1660 {
1661         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1662         struct obd_device_target *obt = &obd->u.obt;
1663
1664         if (!new_client && service_time)
1665                 /* Teach server about old server's estimates, as first guess
1666                  * at how long new requests will take. */
1667                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1668                             service_time);
1669
1670         target_start_recovery_timer(obd);
1671
1672         /* Convert the service time to RPC timeout,
1673          * and reuse service_time to limit stack usage. */
1674         service_time = at_est2timeout(service_time);
1675
1676         /* We expect other clients to timeout within service_time, then try
1677          * to reconnect, then try the failover server.  The max delay between
1678          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1679         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1680
1681         LASSERT(obt->obt_magic == OBT_MAGIC);
1682         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1683         if (service_time > obd->obd_recovery_timeout && !new_client)
1684                 extend_recovery_timer(obd, service_time, false);
1685 }
1686
1687 /** Health checking routines */
1688 static inline int exp_connect_healthy(struct obd_export *exp)
1689 {
1690         return (exp->exp_in_recovery);
1691 }
1692
1693 /** if export done req_replay or has replay in queue */
1694 static inline int exp_req_replay_healthy(struct obd_export *exp)
1695 {
1696         return (!exp->exp_req_replay_needed ||
1697                 atomic_read(&exp->exp_replay_count) > 0);
1698 }
1699 /** if export done lock_replay or has replay in queue */
1700 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1701 {
1702         return (!exp->exp_lock_replay_needed ||
1703                 atomic_read(&exp->exp_replay_count) > 0);
1704 }
1705
1706 static inline int exp_vbr_healthy(struct obd_export *exp)
1707 {
1708         return (!exp->exp_vbr_failed);
1709 }
1710
1711 static inline int exp_finished(struct obd_export *exp)
1712 {
1713         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1714 }
1715
1716 static int check_for_next_transno(struct lu_target *lut)
1717 {
1718         struct ptlrpc_request *req = NULL;
1719         struct obd_device *obd = lut->lut_obd;
1720         int wake_up = 0, connected, completed, queue_len;
1721         __u64 req_transno = 0;
1722         __u64 update_transno = 0;
1723         __u64 next_transno = 0;
1724         ENTRY;
1725
1726         spin_lock(&obd->obd_recovery_task_lock);
1727         if (!list_empty(&obd->obd_req_replay_queue)) {
1728                 req = list_entry(obd->obd_req_replay_queue.next,
1729                                      struct ptlrpc_request, rq_list);
1730                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1731         }
1732
1733         if (lut->lut_tdtd != NULL) {
1734                 struct target_distribute_txn_data *tdtd;
1735
1736                 tdtd = lut->lut_tdtd;
1737                 update_transno = distribute_txn_get_next_transno(lut->lut_tdtd);
1738         }
1739
1740         connected = atomic_read(&obd->obd_connected_clients);
1741         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1742         queue_len = obd->obd_requests_queued_for_recovery;
1743         next_transno = obd->obd_next_recovery_transno;
1744
1745         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1746                "req_transno: "LPU64", next_transno: "LPU64"\n",
1747                obd->obd_max_recoverable_clients, connected, completed,
1748                queue_len, req_transno, next_transno);
1749
1750         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1751                 CDEBUG(D_HA, "waking for aborted recovery\n");
1752                 wake_up = 1;
1753         } else if (obd->obd_recovery_expired) {
1754                 CDEBUG(D_HA, "waking for expired recovery\n");
1755                 wake_up = 1;
1756         } else if (req_transno == next_transno ||
1757                    (update_transno != 0 && update_transno <= next_transno)) {
1758                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1759                 wake_up = 1;
1760         } else if (queue_len > 0 &&
1761                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1762                 int d_lvl = D_HA;
1763                 /** handle gaps occured due to lost reply or VBR */
1764                 LASSERTF(req_transno >= next_transno,
1765                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1766                          req_transno, next_transno);
1767                 if (req_transno > obd->obd_last_committed &&
1768                     !obd->obd_version_recov)
1769                         d_lvl = D_ERROR;
1770                 CDEBUG(d_lvl,
1771                        "%s: waking for gap in transno, VBR is %s (skip: "
1772                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1773                        ", next_update "LPD64" last_committed: "LPD64")\n",
1774                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1775                        next_transno, queue_len, completed, connected,
1776                        req_transno, update_transno, obd->obd_last_committed);
1777                 obd->obd_next_recovery_transno = req_transno;
1778                 wake_up = 1;
1779         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1780                 CDEBUG(D_HA, "waking for completed recovery\n");
1781                 wake_up = 1;
1782         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1783                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1784                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1785                 obd->obd_next_recovery_transno = req_transno;
1786                 wake_up = 1;
1787         }
1788         spin_unlock(&obd->obd_recovery_task_lock);
1789         return wake_up;
1790 }
1791
1792 static int check_for_next_lock(struct lu_target *lut)
1793 {
1794         struct obd_device *obd = lut->lut_obd;
1795         int wake_up = 0;
1796
1797         spin_lock(&obd->obd_recovery_task_lock);
1798         if (!list_empty(&obd->obd_lock_replay_queue)) {
1799                 CDEBUG(D_HA, "waking for next lock\n");
1800                 wake_up = 1;
1801         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1802                 CDEBUG(D_HA, "waking for completed lock replay\n");
1803                 wake_up = 1;
1804         } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1805                 CDEBUG(D_HA, "waking for aborted recovery\n");
1806                 wake_up = 1;
1807         } else if (obd->obd_recovery_expired) {
1808                 CDEBUG(D_HA, "waking for expired recovery\n");
1809                 wake_up = 1;
1810         }
1811         spin_unlock(&obd->obd_recovery_task_lock);
1812
1813         return wake_up;
1814 }
1815
1816 /**
1817  * wait for recovery events,
1818  * check its status with help of check_routine
1819  * evict dead clients via health_check
1820  */
1821 static int target_recovery_overseer(struct lu_target *lut,
1822                                     int (*check_routine)(struct lu_target *),
1823                                     int (*health_check)(struct obd_export *))
1824 {
1825         struct obd_device       *obd = lut->lut_obd;
1826 repeat:
1827         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1828               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1829                 CWARN("recovery is aborted by hard timeout\n");
1830                 obd->obd_abort_recovery = 1;
1831         }
1832
1833         while (wait_event_timeout(obd->obd_next_transno_waitq,
1834                                   check_routine(lut),
1835                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1836                 /* wait indefinitely for event, but don't trigger watchdog */;
1837
1838         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1839                 CWARN("recovery is aborted, evict exports in recovery\n");
1840                 /** evict exports which didn't finish recovery yet */
1841                 class_disconnect_stale_exports(obd, exp_finished);
1842                 return 1;
1843         } else if (obd->obd_recovery_expired) {
1844                 obd->obd_recovery_expired = 0;
1845                 /** If some clients died being recovered, evict them */
1846                 LCONSOLE_WARN("%s: recovery is timed out, "
1847                               "evict stale exports\n", obd->obd_name);
1848                 /** evict cexports with no replay in queue, they are stalled */
1849                 class_disconnect_stale_exports(obd, health_check);
1850                 /** continue with VBR */
1851                 spin_lock(&obd->obd_dev_lock);
1852                 obd->obd_version_recov = 1;
1853                 spin_unlock(&obd->obd_dev_lock);
1854                 /**
1855                  * reset timer, recovery will proceed with versions now,
1856                  * timeout is set just to handle reconnection delays
1857                  */
1858                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1859                 /** Wait for recovery events again, after evicting bad clients */
1860                 goto repeat;
1861         }
1862         return 0;
1863 }
1864
1865 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1866 {
1867         struct obd_device       *obd = lut->lut_obd;
1868         struct ptlrpc_request *req = NULL;
1869
1870         CDEBUG(D_HA, "Waiting for lock\n");
1871         if (target_recovery_overseer(lut, check_for_next_lock,
1872                                      exp_lock_replay_healthy))
1873                 abort_lock_replay_queue(obd);
1874
1875         spin_lock(&obd->obd_recovery_task_lock);
1876         if (!list_empty(&obd->obd_lock_replay_queue)) {
1877                 req = list_entry(obd->obd_lock_replay_queue.next,
1878                                      struct ptlrpc_request, rq_list);
1879                 list_del_init(&req->rq_list);
1880                 spin_unlock(&obd->obd_recovery_task_lock);
1881         } else {
1882                 spin_unlock(&obd->obd_recovery_task_lock);
1883                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1884                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1885                 /** evict exports failed VBR */
1886                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1887         }
1888         return req;
1889 }
1890
1891 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1892 {
1893         struct ptlrpc_request *req = NULL;
1894
1895         spin_lock(&obd->obd_recovery_task_lock);
1896         if (!list_empty(&obd->obd_final_req_queue)) {
1897                 req = list_entry(obd->obd_final_req_queue.next,
1898                                      struct ptlrpc_request, rq_list);
1899                 list_del_init(&req->rq_list);
1900                 spin_unlock(&obd->obd_recovery_task_lock);
1901                 if (req->rq_export->exp_in_recovery) {
1902                         spin_lock(&req->rq_export->exp_lock);
1903                         req->rq_export->exp_in_recovery = 0;
1904                         spin_unlock(&req->rq_export->exp_lock);
1905                 }
1906         } else {
1907                 spin_unlock(&obd->obd_recovery_task_lock);
1908         }
1909         return req;
1910 }
1911
1912 static void handle_recovery_req(struct ptlrpc_thread *thread,
1913                                 struct ptlrpc_request *req,
1914                                 svc_handler_t handler)
1915 {
1916         ENTRY;
1917
1918         /**
1919          * export can be evicted during recovery, no need to handle replays for
1920          * it after that, discard such request silently
1921          */
1922         if (req->rq_export->exp_disconnected)
1923                 RETURN_EXIT;
1924
1925         req->rq_session.lc_thread = thread;
1926         req->rq_svc_thread = thread;
1927         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1928
1929         /* thread context */
1930         lu_context_enter(&thread->t_env->le_ctx);
1931         (void)handler(req);
1932         lu_context_exit(&thread->t_env->le_ctx);
1933
1934         /* don't reset timer for final stage */
1935         if (!exp_finished(req->rq_export)) {
1936                 int to = obd_timeout;
1937
1938                 /**
1939                  * Add request timeout to the recovery time so next request from
1940                  * this client may come in recovery time
1941                  */
1942                 if (!AT_OFF) {
1943                         struct ptlrpc_service_part *svcpt;
1944
1945                         svcpt = req->rq_rqbd->rqbd_svcpt;
1946                         /* If the server sent early reply for this request,
1947                          * the client will recalculate the timeout according to
1948                          * current server estimate service time, so we will
1949                          * use the maxium timeout here for waiting the client
1950                          * sending the next req */
1951                         to = max((int)at_est2timeout(
1952                                  at_get(&svcpt->scp_at_estimate)),
1953                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1954                         /* Add 2 net_latency, one for balance rq_deadline
1955                          * (see ptl_send_rpc), one for resend the req to server,
1956                          * Note: client will pack net_latency in replay req
1957                          * (see ptlrpc_replay_req) */
1958                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1959                 }
1960                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1961         }
1962         EXIT;
1963 }
1964
1965 /** Checking routines for recovery */
1966 static int check_for_recovery_ready(struct lu_target *lut)
1967 {
1968         struct obd_device *obd = lut->lut_obd;
1969         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
1970
1971         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
1972                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
1973                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
1974                obd->obd_recovery_expired);
1975
1976         if (obd->obd_force_abort_recovery)
1977                 return 1;
1978
1979         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
1980                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
1981                 if (clnts + obd->obd_stale_clients <
1982                     obd->obd_max_recoverable_clients)
1983                         return 0;
1984         }
1985
1986         if (lut->lut_tdtd != NULL) {
1987                 if (!lut->lut_tdtd->tdtd_replay_ready) {
1988                         /* Let's extend recovery timer, in case the recovery
1989                          * timer expired, and some clients got evicted */
1990                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
1991                                               true);
1992                         return 0;
1993                 } else {
1994                         dtrq_list_dump(lut->lut_tdtd, D_HA);
1995                 }
1996         }
1997
1998         return 1;
1999 }
2000
2001 enum {
2002         REQUEST_RECOVERY = 1,
2003         UPDATE_RECOVERY = 2,
2004 };
2005
2006 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2007 {
2008         __u64 transno = 0;
2009
2010         if (!list_empty(&obd->obd_req_replay_queue)) {
2011                 struct ptlrpc_request *req;
2012
2013                 req = list_entry(obd->obd_req_replay_queue.next,
2014                                  struct ptlrpc_request, rq_list);
2015                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2016         }
2017
2018         return transno;
2019 }
2020 __u64 get_next_transno(struct lu_target *lut, int *type)
2021 {
2022         struct obd_device *obd = lut->lut_obd;
2023         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2024         __u64 transno = 0;
2025         __u64 update_transno;
2026         ENTRY;
2027
2028         transno = get_next_replay_req_transno(obd);
2029         if (type != NULL)
2030                 *type = REQUEST_RECOVERY;
2031
2032         if (tdtd == NULL)
2033                 RETURN(transno);
2034
2035         update_transno = distribute_txn_get_next_transno(tdtd);
2036         if (transno == 0 || (transno >= update_transno &&
2037                              update_transno != 0)) {
2038                 transno = update_transno;
2039                 if (type != NULL)
2040                         *type = UPDATE_RECOVERY;
2041         }
2042
2043         RETURN(transno);
2044 }
2045
2046 /**
2047  * drop duplicate replay request
2048  *
2049  * Because the operation has been replayed by update recovery, the request
2050  * with the same transno will be dropped and also notify the client to send
2051  * next replay request.
2052  *
2053  * \param[in] env       execution environment
2054  * \param[in] obd       failover obd device
2055  * \param[in] req       request to be dropped
2056  */
2057 static void drop_duplicate_replay_req(struct lu_env *env,
2058                                       struct obd_device *obd,
2059                                       struct ptlrpc_request *req)
2060 {
2061         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2062                   " update records are found.\n",
2063                   lustre_msg_get_transno(req->rq_reqmsg),
2064                   libcfs_nid2str(req->rq_peer.nid));
2065
2066         /* Right now, only for MDS reint operation update replay and
2067          * normal request replay can have the same transno */
2068         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2069                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2070                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2071                 if (likely(req->rq_export))
2072                         target_committed_to_req(req);
2073                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2074                 target_send_reply(req, req->rq_status, 0);
2075         } else {
2076                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2077                 libcfs_nid2str(req->rq_peer.nid));
2078         }
2079         target_exp_dequeue_req_replay(req);
2080         target_request_copy_put(req);
2081         obd->obd_replayed_requests++;
2082 }
2083
2084 /**
2085  * Update last_rcvd of the update
2086  *
2087  * Because update recovery might update the last_rcvd by updates, i.e.
2088  * it will not update the last_rcvd information in memory, so we need
2089  * refresh these information in memory after update recovery.
2090  *
2091  * \param[in] obd       obd_device under recoverying.
2092  * \param[in] dtrq      the update replay requests being replayed.
2093  */
2094 static void target_update_lcd(struct lu_env *env, struct lu_target *lut,
2095                               struct distribute_txn_replay_req *dtrq)
2096 {
2097         struct obd_device       *obd = lut->lut_obd;
2098         struct obd_export       *export;
2099         struct tg_export_data   *ted;
2100         struct distribute_txn_replay_req_sub    *dtrqs;
2101         struct seq_server_site *site;
2102         struct update_records   *ur;
2103         const struct lu_fid     *fid;
2104         struct update_ops       *ops;
2105         struct update_params    *params;
2106         struct update_op        *op;
2107         __u32                   mdt_index;
2108         unsigned int            i;
2109         struct lsd_client_data  *lcd = NULL;
2110
2111         /* if Updates has been executed(committed) on the recovery target,
2112          * i.e. the updates is not being executed on the target, so we do
2113          * not need update it in memory */
2114         site = lu_site2seq(obd->obd_lu_dev->ld_site);
2115         mdt_index = site->ss_node_id;
2116         dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
2117         if (dtrqs != NULL)
2118                 return;
2119
2120         if (dtrq->dtrq_lur == NULL)
2121                 return;
2122
2123         /* Find the update last_rcvd record */
2124         fid = lu_object_fid(&lut->lut_last_rcvd->do_lu);
2125         ur = &dtrq->dtrq_lur->lur_update_rec;
2126         ops = &ur->ur_ops;
2127         params = update_records_get_params(ur);
2128         for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count;
2129              i++, op = update_op_next_op(op)) {
2130                 __u64 pos;
2131                 __u16 size;
2132                 void *buf;
2133
2134                 if (!lu_fid_eq(&op->uop_fid, fid))
2135                         continue;
2136
2137                 if (op->uop_type != OUT_WRITE)
2138                         continue;
2139
2140                 buf = update_params_get_param_buf(params, op->uop_params_off[1],
2141                                                   ur->ur_param_count, NULL);
2142                 if (buf == NULL)
2143                         continue;
2144
2145                 pos = le64_to_cpu(*(__u64 *)buf);
2146                 if (pos == 0)
2147                         continue;
2148
2149                 buf = update_params_get_param_buf(params, op->uop_params_off[0],
2150                                                   ur->ur_param_count, &size);
2151                 if (buf == NULL)
2152                         continue;
2153
2154                 if (size != sizeof(*lcd))
2155                         continue;
2156                 lcd = buf;
2157         }
2158
2159         if (lcd == NULL || lcd->lcd_uuid[0] == '\0')
2160                 return;
2161
2162         /* locate the export then update the exp_target_data if needed */
2163         export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid);
2164         if (export == NULL)
2165                 return;
2166
2167         ted = &export->exp_target_data;
2168         if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) {
2169                 CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n",
2170                        lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid,
2171                        lcd->lcd_last_xid);
2172                 ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid;
2173                 ted->ted_lcd->lcd_last_result = lcd->lcd_last_result;
2174         }
2175         class_export_put(export);
2176 }
2177
2178 static void replay_request_or_update(struct lu_env *env,
2179                                      struct lu_target *lut,
2180                                      struct target_recovery_data *trd,
2181                                      struct ptlrpc_thread *thread)
2182 {
2183         struct obd_device *obd = lut->lut_obd;
2184         struct ptlrpc_request *req = NULL;
2185         int                     type;
2186         __u64                   transno;
2187         ENTRY;
2188
2189         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2190                obd->obd_next_recovery_transno);
2191
2192         /* Replay all of request and update by transno */
2193         do {
2194                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2195
2196                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2197
2198                 /** It is needed to extend recovery window above
2199                  *  recovery_time_soft. Extending is possible only in the
2200                  *  end of recovery window (see more details in
2201                  *  handle_recovery_req()).
2202                  */
2203                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2204
2205                 if (target_recovery_overseer(lut, check_for_next_transno,
2206                                              exp_req_replay_healthy)) {
2207                         abort_req_replay_queue(obd);
2208                         abort_lock_replay_queue(obd);
2209                 }
2210
2211                 spin_lock(&obd->obd_recovery_task_lock);
2212                 transno = get_next_transno(lut, &type);
2213                 if (type == REQUEST_RECOVERY && tdtd != NULL &&
2214                     transno == tdtd->tdtd_last_update_transno) {
2215                         /* Drop replay request from client side, if the
2216                          * replay has been executed by update with the
2217                          * same transno */
2218                         req = list_entry(obd->obd_req_replay_queue.next,
2219                                         struct ptlrpc_request, rq_list);
2220                         list_del_init(&req->rq_list);
2221                         obd->obd_requests_queued_for_recovery--;
2222                         spin_unlock(&obd->obd_recovery_task_lock);
2223                         drop_duplicate_replay_req(env, obd, req);
2224                 } else if (type == REQUEST_RECOVERY && transno != 0) {
2225                         req = list_entry(obd->obd_req_replay_queue.next,
2226                                              struct ptlrpc_request, rq_list);
2227                         list_del_init(&req->rq_list);
2228                         obd->obd_requests_queued_for_recovery--;
2229                         spin_unlock(&obd->obd_recovery_task_lock);
2230                         LASSERT(trd->trd_processing_task == current_pid());
2231                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2232                                   lustre_msg_get_transno(req->rq_reqmsg),
2233                                   libcfs_nid2str(req->rq_peer.nid));
2234
2235                         handle_recovery_req(thread, req,
2236                                             trd->trd_recovery_handler);
2237                         /**
2238                          * bz18031: increase next_recovery_transno before
2239                          * target_request_copy_put() will drop exp_rpc reference
2240                          */
2241                         spin_lock(&obd->obd_recovery_task_lock);
2242                         obd->obd_next_recovery_transno++;
2243                         spin_unlock(&obd->obd_recovery_task_lock);
2244                         target_exp_dequeue_req_replay(req);
2245                         target_request_copy_put(req);
2246                         obd->obd_replayed_requests++;
2247                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2248                         struct distribute_txn_replay_req *dtrq;
2249
2250                         spin_unlock(&obd->obd_recovery_task_lock);
2251
2252                         LASSERT(tdtd != NULL);
2253                         dtrq = distribute_txn_get_next_req(tdtd);
2254                         lu_context_enter(&thread->t_env->le_ctx);
2255                         tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2256                         lu_context_exit(&thread->t_env->le_ctx);
2257                         extend_recovery_timer(obd, obd_timeout, true);
2258                         LASSERT(tdtd->tdtd_last_update_transno <= transno);
2259                         tdtd->tdtd_last_update_transno = transno;
2260                         spin_lock(&obd->obd_recovery_task_lock);
2261                         if (transno > obd->obd_next_recovery_transno)
2262                                 obd->obd_next_recovery_transno = transno;
2263                         spin_unlock(&obd->obd_recovery_task_lock);
2264                         target_update_lcd(env, lut, dtrq);
2265                         dtrq_destroy(dtrq);
2266                 } else {
2267                         spin_unlock(&obd->obd_recovery_task_lock);
2268                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2269                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2270                         /** evict exports failed VBR */
2271                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2272                         break;
2273                 }
2274         } while (1);
2275 }
2276
2277 static int target_recovery_thread(void *arg)
2278 {
2279         struct lu_target *lut = arg;
2280         struct obd_device *obd = lut->lut_obd;
2281         struct ptlrpc_request *req;
2282         struct target_recovery_data *trd = &obd->obd_recovery_data;
2283         unsigned long delta;
2284         struct lu_env *env;
2285         struct ptlrpc_thread *thread = NULL;
2286         int rc = 0;
2287         ENTRY;
2288
2289         unshare_fs_struct();
2290         OBD_ALLOC_PTR(thread);
2291         if (thread == NULL)
2292                 RETURN(-ENOMEM);
2293
2294         OBD_ALLOC_PTR(env);
2295         if (env == NULL) {
2296                 OBD_FREE_PTR(thread);
2297                 RETURN(-ENOMEM);
2298         }
2299
2300         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2301         if (rc) {
2302                 OBD_FREE_PTR(thread);
2303                 OBD_FREE_PTR(env);
2304                 RETURN(rc);
2305         }
2306
2307         thread->t_env = env;
2308         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2309         env->le_ctx.lc_thread = thread;
2310         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2311         thread->t_watchdog = NULL;
2312
2313         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2314                current_pid());
2315         trd->trd_processing_task = current_pid();
2316
2317         spin_lock(&obd->obd_dev_lock);
2318         obd->obd_recovering = 1;
2319         spin_unlock(&obd->obd_dev_lock);
2320         complete(&trd->trd_starting);
2321
2322         /* first of all, we have to know the first transno to replay */
2323         if (target_recovery_overseer(lut, check_for_recovery_ready,
2324                                      exp_connect_healthy)) {
2325                 abort_req_replay_queue(obd);
2326                 abort_lock_replay_queue(obd);
2327                 if (lut->lut_tdtd != NULL)
2328                         dtrq_list_destroy(lut->lut_tdtd);
2329         }
2330
2331         /* next stage: replay requests or update */
2332         delta = jiffies;
2333         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2334                atomic_read(&obd->obd_req_replay_clients),
2335                obd->obd_next_recovery_transno);
2336         replay_request_or_update(env, lut, trd, thread);
2337
2338         /**
2339          * The second stage: replay locks
2340          */
2341         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2342                atomic_read(&obd->obd_lock_replay_clients));
2343         while ((req = target_next_replay_lock(lut))) {
2344                 LASSERT(trd->trd_processing_task == current_pid());
2345                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2346                           libcfs_nid2str(req->rq_peer.nid));
2347                 handle_recovery_req(thread, req,
2348                                     trd->trd_recovery_handler);
2349                 target_request_copy_put(req);
2350                 obd->obd_replayed_locks++;
2351         }
2352
2353         /**
2354          * The third stage: reply on final pings, at this moment all clients
2355          * must have request in final queue
2356          */
2357         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2358         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2359         /** Update server last boot epoch */
2360         tgt_boot_epoch_update(lut);
2361         /* We drop recoverying flag to forward all new requests
2362          * to regular mds_handle() since now */
2363         spin_lock(&obd->obd_dev_lock);
2364         obd->obd_recovering = obd->obd_abort_recovery = 0;
2365         spin_unlock(&obd->obd_dev_lock);
2366         spin_lock(&obd->obd_recovery_task_lock);
2367         target_cancel_recovery_timer(obd);
2368         spin_unlock(&obd->obd_recovery_task_lock);
2369         while ((req = target_next_final_ping(obd))) {
2370                 LASSERT(trd->trd_processing_task == current_pid());
2371                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2372                           libcfs_nid2str(req->rq_peer.nid));
2373                 handle_recovery_req(thread, req,
2374                                     trd->trd_recovery_handler);
2375                 /* Because the waiting client can not send ping to server,
2376                  * so we need refresh the last_request_time, to avoid the
2377                  * export is being evicted */
2378                 ptlrpc_update_export_timer(req->rq_export, 0);
2379                 target_request_copy_put(req);
2380         }
2381
2382         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2383         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2384               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2385         if (delta > OBD_RECOVERY_TIME_SOFT) {
2386                 CWARN("too long recovery - read logs\n");
2387                 libcfs_debug_dumplog();
2388         }
2389
2390         target_finish_recovery(lut);
2391
2392         lu_context_fini(&env->le_ctx);
2393         trd->trd_processing_task = 0;
2394         complete(&trd->trd_finishing);
2395
2396         tgt_io_thread_done(thread);
2397         OBD_FREE_PTR(thread);
2398         OBD_FREE_PTR(env);
2399         RETURN(rc);
2400 }
2401
2402 static int target_start_recovery_thread(struct lu_target *lut,
2403                                         svc_handler_t handler)
2404 {
2405         struct obd_device *obd = lut->lut_obd;
2406         int rc = 0;
2407         struct target_recovery_data *trd = &obd->obd_recovery_data;
2408
2409         memset(trd, 0, sizeof(*trd));
2410         init_completion(&trd->trd_starting);
2411         init_completion(&trd->trd_finishing);
2412         trd->trd_recovery_handler = handler;
2413
2414         if (!IS_ERR(kthread_run(target_recovery_thread,
2415                                 lut, "tgt_recov"))) {
2416                 wait_for_completion(&trd->trd_starting);
2417                 LASSERT(obd->obd_recovering != 0);
2418         } else {
2419                 rc = -ECHILD;
2420         }
2421
2422         return rc;
2423 }
2424
2425 void target_stop_recovery_thread(struct obd_device *obd)
2426 {
2427         if (obd->obd_recovery_data.trd_processing_task > 0) {
2428                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2429                 /** recovery can be done but postrecovery is not yet */
2430                 spin_lock(&obd->obd_dev_lock);
2431                 if (obd->obd_recovering) {
2432                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2433                         obd->obd_abort_recovery = 1;
2434                         wake_up(&obd->obd_next_transno_waitq);
2435                 }
2436                 spin_unlock(&obd->obd_dev_lock);
2437                 wait_for_completion(&trd->trd_finishing);
2438         }
2439 }
2440 EXPORT_SYMBOL(target_stop_recovery_thread);
2441
2442 void target_recovery_fini(struct obd_device *obd)
2443 {
2444         class_disconnect_exports(obd);
2445         target_stop_recovery_thread(obd);
2446         target_cleanup_recovery(obd);
2447 }
2448 EXPORT_SYMBOL(target_recovery_fini);
2449
2450 static void target_recovery_expired(unsigned long castmeharder)
2451 {
2452         struct obd_device *obd = (struct obd_device *)castmeharder;
2453         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2454                " after %lds (%d clients connected)\n",
2455                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2456                cfs_time_current_sec()- obd->obd_recovery_start,
2457                atomic_read(&obd->obd_connected_clients));
2458
2459         obd->obd_recovery_expired = 1;
2460         wake_up(&obd->obd_next_transno_waitq);
2461 }
2462
2463 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2464 {
2465         struct obd_device *obd = lut->lut_obd;
2466
2467         if (obd->obd_max_recoverable_clients == 0) {
2468                 /** Update server last boot epoch */
2469                 tgt_boot_epoch_update(lut);
2470                 return;
2471         }
2472
2473         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2474                "last_transno "LPU64"\n", obd->obd_name,
2475                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2476         LASSERT(obd->obd_stopping == 0);
2477         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2478         obd->obd_recovery_start = 0;
2479         obd->obd_recovery_end = 0;
2480
2481         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2482         target_start_recovery_thread(lut, handler);
2483 }
2484 EXPORT_SYMBOL(target_recovery_init);
2485
2486 static int target_process_req_flags(struct obd_device *obd,
2487                                     struct ptlrpc_request *req)
2488 {
2489         struct obd_export *exp = req->rq_export;
2490         LASSERT(exp != NULL);
2491         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2492                 /* client declares he's ready to replay locks */
2493                 spin_lock(&exp->exp_lock);
2494                 if (exp->exp_req_replay_needed) {
2495                         exp->exp_req_replay_needed = 0;
2496                         spin_unlock(&exp->exp_lock);
2497
2498                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2499                         atomic_dec(&obd->obd_req_replay_clients);
2500                 } else {
2501                         spin_unlock(&exp->exp_lock);
2502                 }
2503         }
2504         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2505                 /* client declares he's ready to complete recovery
2506                  * so, we put the request on th final queue */
2507                 spin_lock(&exp->exp_lock);
2508                 if (exp->exp_lock_replay_needed) {
2509                         exp->exp_lock_replay_needed = 0;
2510                         spin_unlock(&exp->exp_lock);
2511
2512                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2513                         atomic_dec(&obd->obd_lock_replay_clients);
2514                 } else {
2515                         spin_unlock(&exp->exp_lock);
2516                 }
2517         }
2518         return 0;
2519 }
2520
2521 int target_queue_recovery_request(struct ptlrpc_request *req,
2522                                   struct obd_device *obd)
2523 {
2524         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2525         struct ptlrpc_request *reqiter;
2526         int inserted = 0;
2527         ENTRY;
2528
2529         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2530                 /* Processing the queue right now, don't re-add. */
2531                 RETURN(1);
2532         }
2533
2534         target_process_req_flags(obd, req);
2535
2536         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2537                 /* client declares he's ready to complete recovery
2538                  * so, we put the request on th final queue */
2539                 target_request_copy_get(req);
2540                 DEBUG_REQ(D_HA, req, "queue final req");
2541                 wake_up(&obd->obd_next_transno_waitq);
2542                 spin_lock(&obd->obd_recovery_task_lock);
2543                 if (obd->obd_recovering) {
2544                         list_add_tail(&req->rq_list,
2545                                           &obd->obd_final_req_queue);
2546                 } else {
2547                         spin_unlock(&obd->obd_recovery_task_lock);
2548                         target_request_copy_put(req);
2549                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2550                 }
2551                 spin_unlock(&obd->obd_recovery_task_lock);
2552                 RETURN(0);
2553         }
2554         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2555                 /* client declares he's ready to replay locks */
2556                 target_request_copy_get(req);
2557                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2558                 wake_up(&obd->obd_next_transno_waitq);
2559                 spin_lock(&obd->obd_recovery_task_lock);
2560                 LASSERT(obd->obd_recovering);
2561                 /* usually due to recovery abort */
2562                 if (!req->rq_export->exp_in_recovery) {
2563                         spin_unlock(&obd->obd_recovery_task_lock);
2564                         target_request_copy_put(req);
2565                         RETURN(-ENOTCONN);
2566                 }
2567                 LASSERT(req->rq_export->exp_lock_replay_needed);
2568                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2569                 spin_unlock(&obd->obd_recovery_task_lock);
2570                 RETURN(0);
2571         }
2572
2573         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2574          * (i.e. buflens etc are in my own byte order), but type-dependent
2575          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2576
2577         if (!transno) {
2578                 INIT_LIST_HEAD(&req->rq_list);
2579                 DEBUG_REQ(D_HA, req, "not queueing");
2580                 RETURN(1);
2581         }
2582
2583         /* If we're processing the queue, we want don't want to queue this
2584          * message.
2585          *
2586          * Also, if this request has a transno less than the one we're waiting
2587          * for, we should process it now.  It could (and currently always will)
2588          * be an open request for a descriptor that was opened some time ago.
2589          *
2590          * Also, a resent, replayed request that has already been
2591          * handled will pass through here and be processed immediately.
2592          */
2593         CDEBUG(D_HA, "Next recovery transno: "LPU64
2594                ", current: "LPU64", replaying\n",
2595                obd->obd_next_recovery_transno, transno);
2596         spin_lock(&obd->obd_recovery_task_lock);
2597         if (transno < obd->obd_next_recovery_transno) {
2598                 /* Processing the queue right now, don't re-add. */
2599                 LASSERT(list_empty(&req->rq_list));
2600                 spin_unlock(&obd->obd_recovery_task_lock);
2601                 RETURN(1);
2602         }
2603         spin_unlock(&obd->obd_recovery_task_lock);
2604
2605         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2606                 RETURN(0);
2607
2608         target_request_copy_get(req);
2609         if (!req->rq_export->exp_in_recovery) {
2610                 target_request_copy_put(req);
2611                 RETURN(-ENOTCONN);
2612         }
2613         LASSERT(req->rq_export->exp_req_replay_needed);
2614
2615         if (target_exp_enqueue_req_replay(req)) {
2616                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2617                 target_request_copy_put(req);
2618                 RETURN(0);
2619         }
2620
2621         /* XXX O(n^2) */
2622         spin_lock(&obd->obd_recovery_task_lock);
2623         LASSERT(obd->obd_recovering);
2624         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2625                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2626                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2627                         inserted = 1;
2628                         goto added;
2629                 }
2630
2631                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2632                              transno)) {
2633                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2634                                   "has been claimed by another client");
2635                         spin_unlock(&obd->obd_recovery_task_lock);
2636                         target_exp_dequeue_req_replay(req);
2637                         target_request_copy_put(req);
2638                         RETURN(0);
2639                 }
2640         }
2641 added:
2642         if (!inserted)
2643                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2644
2645         obd->obd_requests_queued_for_recovery++;
2646         spin_unlock(&obd->obd_recovery_task_lock);
2647         wake_up(&obd->obd_next_transno_waitq);
2648         RETURN(0);
2649 }
2650
2651 int target_handle_ping(struct ptlrpc_request *req)
2652 {
2653         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2654         return req_capsule_server_pack(&req->rq_pill);
2655 }
2656
2657 void target_committed_to_req(struct ptlrpc_request *req)
2658 {
2659         struct obd_export *exp = req->rq_export;
2660
2661         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2662                 lustre_msg_set_last_committed(req->rq_repmsg,
2663                                               exp->exp_last_committed);
2664         else
2665                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2666                           "%d)", exp->exp_obd->obd_no_transno,
2667                           req->rq_repmsg == NULL);
2668
2669         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2670                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2671 }
2672
2673 #endif /* HAVE_SERVER_SUPPORT */
2674
2675 /**
2676  * Packs current SLV and Limit into \a req.
2677  */
2678 int target_pack_pool_reply(struct ptlrpc_request *req)
2679 {
2680         struct obd_device *obd;
2681         ENTRY;
2682
2683         /* Check that we still have all structures alive as this may
2684          * be some late RPC at shutdown time. */
2685         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2686                      !exp_connect_lru_resize(req->rq_export))) {
2687                 lustre_msg_set_slv(req->rq_repmsg, 0);
2688                 lustre_msg_set_limit(req->rq_repmsg, 0);
2689                 RETURN(0);
2690         }
2691
2692         /* OBD is alive here as export is alive, which we checked above. */
2693         obd = req->rq_export->exp_obd;
2694
2695         read_lock(&obd->obd_pool_lock);
2696         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2697         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2698         read_unlock(&obd->obd_pool_lock);
2699
2700         RETURN(0);
2701 }
2702
2703 static int target_send_reply_msg(struct ptlrpc_request *req,
2704                                  int rc, int fail_id)
2705 {
2706         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2707                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2708                 return (-ECOMM);
2709         }
2710
2711         if (unlikely(rc)) {
2712                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2713                 req->rq_status = rc;
2714                 return (ptlrpc_send_error(req, 1));
2715         } else {
2716                 DEBUG_REQ(D_NET, req, "sending reply");
2717         }
2718
2719         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2720 }
2721
2722 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2723 {
2724         struct ptlrpc_service_part *svcpt;
2725         int                        netrc;
2726         struct ptlrpc_reply_state *rs;
2727         struct obd_export         *exp;
2728         ENTRY;
2729
2730         if (req->rq_no_reply) {
2731                 EXIT;
2732                 return;
2733         }
2734
2735         svcpt = req->rq_rqbd->rqbd_svcpt;
2736         rs = req->rq_reply_state;
2737         if (rs == NULL || !rs->rs_difficult) {
2738                 /* no notifiers */
2739                 target_send_reply_msg (req, rc, fail_id);
2740                 EXIT;
2741                 return;
2742         }
2743
2744         /* must be an export if locks saved */
2745         LASSERT(req->rq_export != NULL);
2746         /* req/reply consistent */
2747         LASSERT(rs->rs_svcpt == svcpt);
2748
2749         /* "fresh" reply */
2750         LASSERT(!rs->rs_scheduled);
2751         LASSERT(!rs->rs_scheduled_ever);
2752         LASSERT(!rs->rs_handled);
2753         LASSERT(!rs->rs_on_net);
2754         LASSERT(rs->rs_export == NULL);
2755         LASSERT(list_empty(&rs->rs_obd_list));
2756         LASSERT(list_empty(&rs->rs_exp_list));
2757
2758         exp = class_export_get(req->rq_export);
2759
2760         /* disable reply scheduling while I'm setting up */
2761         rs->rs_scheduled = 1;
2762         rs->rs_on_net    = 1;
2763         rs->rs_xid       = req->rq_xid;
2764         rs->rs_transno   = req->rq_transno;
2765         rs->rs_export    = exp;
2766         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2767
2768         spin_lock(&exp->exp_uncommitted_replies_lock);
2769         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2770                rs->rs_transno, exp->exp_last_committed);
2771         if (rs->rs_transno > exp->exp_last_committed) {
2772                 /* not committed already */
2773                 list_add_tail(&rs->rs_obd_list,
2774                                   &exp->exp_uncommitted_replies);
2775         }
2776         spin_unlock(&exp->exp_uncommitted_replies_lock);
2777
2778         spin_lock(&exp->exp_lock);
2779         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2780         spin_unlock(&exp->exp_lock);
2781
2782         netrc = target_send_reply_msg(req, rc, fail_id);
2783
2784         spin_lock(&svcpt->scp_rep_lock);
2785
2786         atomic_inc(&svcpt->scp_nreps_difficult);
2787
2788         if (netrc != 0) {
2789                 /* error sending: reply is off the net.  Also we need +1
2790                  * reply ref until ptlrpc_handle_rs() is done
2791                  * with the reply state (if the send was successful, there
2792                  * would have been +1 ref for the net, which
2793                  * reply_out_callback leaves alone) */
2794                 rs->rs_on_net = 0;
2795                 ptlrpc_rs_addref(rs);
2796         }
2797
2798         spin_lock(&rs->rs_lock);
2799         if (rs->rs_transno <= exp->exp_last_committed ||
2800             (!rs->rs_on_net && !rs->rs_no_ack) ||
2801             list_empty(&rs->rs_exp_list) ||     /* completed already */
2802             list_empty(&rs->rs_obd_list)) {
2803                 CDEBUG(D_HA, "Schedule reply immediately\n");
2804                 ptlrpc_dispatch_difficult_reply(rs);
2805         } else {
2806                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2807                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2808         }
2809         spin_unlock(&rs->rs_lock);
2810         spin_unlock(&svcpt->scp_rep_lock);
2811         EXIT;
2812 }
2813
2814 ldlm_mode_t lck_compat_array[] = {
2815         [LCK_EX]    = LCK_COMPAT_EX,
2816         [LCK_PW]    = LCK_COMPAT_PW,
2817         [LCK_PR]    = LCK_COMPAT_PR,
2818         [LCK_CW]    = LCK_COMPAT_CW,
2819         [LCK_CR]    = LCK_COMPAT_CR,
2820         [LCK_NL]    = LCK_COMPAT_NL,
2821         [LCK_GROUP] = LCK_COMPAT_GROUP,
2822         [LCK_COS]   = LCK_COMPAT_COS,
2823 };
2824
2825 /**
2826  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2827  * not escape to the user level.
2828  */
2829 int ldlm_error2errno(ldlm_error_t error)
2830 {
2831         int result;
2832
2833         switch (error) {
2834         case ELDLM_OK:
2835         case ELDLM_LOCK_MATCHED:
2836                 result = 0;
2837                 break;
2838         case ELDLM_LOCK_CHANGED:
2839                 result = -ESTALE;
2840                 break;
2841         case ELDLM_LOCK_ABORTED:
2842                 result = -ENAVAIL;
2843                 break;
2844         case ELDLM_LOCK_REPLACED:
2845                 result = -ESRCH;
2846                 break;
2847         case ELDLM_NO_LOCK_DATA:
2848                 result = -ENOENT;
2849                 break;
2850         case ELDLM_NAMESPACE_EXISTS:
2851                 result = -EEXIST;
2852                 break;
2853         case ELDLM_BAD_NAMESPACE:
2854                 result = -EBADF;
2855                 break;
2856         default:
2857                 if (((int)error) < 0)  /* cast to signed type */
2858                         result = error; /* as ldlm_error_t can be unsigned */
2859                 else {
2860                         CERROR("Invalid DLM result code: %d\n", error);
2861                         result = -EPROTO;
2862                 }
2863         }
2864         return result;
2865 }
2866 EXPORT_SYMBOL(ldlm_error2errno);
2867
2868 /**
2869  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2870  */
2871 ldlm_error_t ldlm_errno2error(int err_no)
2872 {
2873         int error;
2874
2875         switch (err_no) {
2876         case 0:
2877                 error = ELDLM_OK;
2878                 break;
2879         case -ESTALE:
2880                 error = ELDLM_LOCK_CHANGED;
2881                 break;
2882         case -ENAVAIL:
2883                 error = ELDLM_LOCK_ABORTED;
2884                 break;
2885         case -ESRCH:
2886                 error = ELDLM_LOCK_REPLACED;
2887                 break;
2888         case -ENOENT:
2889                 error = ELDLM_NO_LOCK_DATA;
2890                 break;
2891         case -EEXIST:
2892                 error = ELDLM_NAMESPACE_EXISTS;
2893                 break;
2894         case -EBADF:
2895                 error = ELDLM_BAD_NAMESPACE;
2896                 break;
2897         default:
2898                 error = err_no;
2899         }
2900         return error;
2901 }
2902
2903 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2904 void ldlm_dump_export_locks(struct obd_export *exp)
2905 {
2906         spin_lock(&exp->exp_locks_list_guard);
2907         if (!list_empty(&exp->exp_locks_list)) {
2908                 struct ldlm_lock *lock;
2909
2910                 CERROR("dumping locks for export %p,"
2911                        "ignore if the unmount doesn't hang\n", exp);
2912                 list_for_each_entry(lock, &exp->exp_locks_list,
2913                                         l_exp_refs_link)
2914                         LDLM_ERROR(lock, "lock:");
2915         }
2916         spin_unlock(&exp->exp_locks_list_guard);
2917 }
2918 #endif
2919
2920 #ifdef HAVE_SERVER_SUPPORT
2921 static int target_bulk_timeout(void *data)
2922 {
2923         ENTRY;
2924         /* We don't fail the connection here, because having the export
2925          * killed makes the (vital) call to commitrw very sad.
2926          */
2927         RETURN(1);
2928 }
2929
2930 static inline const char *bulk2type(struct ptlrpc_request *req)
2931 {
2932         if (req->rq_bulk_read)
2933                 return "READ";
2934         if (req->rq_bulk_write)
2935                 return "WRITE";
2936         return "UNKNOWN";
2937 }
2938
2939 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2940                    struct l_wait_info *lwi)
2941 {
2942         struct ptlrpc_request   *req = desc->bd_req;
2943         time_t                   start = cfs_time_current_sec();
2944         time_t                   deadline;
2945         int                      rc = 0;
2946
2947         ENTRY;
2948
2949         /* If there is eviction in progress, wait for it to finish. */
2950         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2951                 *lwi = LWI_INTR(NULL, NULL);
2952                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2953                                   !atomic_read(&exp->exp_obd->
2954                                                    obd_evict_inprogress),
2955                                   lwi);
2956         }
2957
2958         /* Check if client was evicted or reconnected already. */
2959         if (exp->exp_failed ||
2960             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2961                 rc = -ENOTCONN;
2962         } else {
2963                 if (req->rq_bulk_read)
2964                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2965
2966                 if ((exp->exp_connect_data.ocd_connect_flags &
2967                      OBD_CONNECT_BULK_MBITS) != 0)
2968                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2969                 else /* old version, bulk matchbits is rq_xid */
2970                         req->rq_mbits = req->rq_xid;
2971
2972                 if (rc == 0)
2973                         rc = ptlrpc_start_bulk_transfer(desc);
2974         }
2975
2976         if (rc < 0) {
2977                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2978                           bulk2type(req), rc);
2979                 RETURN(rc);
2980         }
2981
2982         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2983                 ptlrpc_abort_bulk(desc);
2984                 RETURN(0);
2985         }
2986
2987         /* limit actual bulk transfer to bulk_timeout seconds */
2988         deadline = start + bulk_timeout;
2989         if (deadline > req->rq_deadline)
2990                 deadline = req->rq_deadline;
2991
2992         do {
2993                 long timeoutl = deadline - cfs_time_current_sec();
2994                 cfs_duration_t timeout = timeoutl <= 0 ?
2995                                          CFS_TICK : cfs_time_seconds(timeoutl);
2996                 time_t  rq_deadline;
2997
2998                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2999                                             target_bulk_timeout, desc);
3000                 rc = l_wait_event(desc->bd_waitq,
3001                                   !ptlrpc_server_bulk_active(desc) ||
3002                                   exp->exp_failed ||
3003                                   exp->exp_conn_cnt >
3004                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3005                                   lwi);
3006                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3007                 /* Wait again if we changed rq_deadline. */
3008                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3009                 deadline = start + bulk_timeout;
3010                 if (deadline > rq_deadline)
3011                         deadline = rq_deadline;
3012         } while ((rc == -ETIMEDOUT) &&
3013                  (deadline > cfs_time_current_sec()));
3014
3015         if (rc == -ETIMEDOUT) {
3016                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3017                           bulk2type(req), deadline - start,
3018                           cfs_time_current_sec() - deadline);
3019                 ptlrpc_abort_bulk(desc);
3020         } else if (exp->exp_failed) {
3021                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3022                           bulk2type(req));
3023                 rc = -ENOTCONN;
3024                 ptlrpc_abort_bulk(desc);
3025         } else if (exp->exp_conn_cnt >
3026                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3027                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3028                           bulk2type(req));
3029                 /* We don't reply anyway. */
3030                 rc = -ETIMEDOUT;
3031                 ptlrpc_abort_bulk(desc);
3032         } else if (desc->bd_failure) {
3033                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3034                           bulk2type(req));
3035                 /* XXX should this be a different errno? */
3036                 rc = -ETIMEDOUT;
3037         } else {
3038                 if (req->rq_bulk_write)
3039                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3040                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3041                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3042                                   bulk2type(req), desc->bd_nob_transferred,
3043                                   desc->bd_nob);
3044                         /* XXX should this be a different errno? */
3045                         rc = -ETIMEDOUT;
3046                 }
3047         }
3048
3049         RETURN(rc);
3050 }
3051 EXPORT_SYMBOL(target_bulk_io);
3052
3053 #endif /* HAVE_SERVER_SUPPORT */