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         char *cli_name = lustre_cfg_buf(lcfg, 0);
 271         ENTRY;
 272
 273         /* In a more perfect world, we would hang a ptlrpc_client off of
 274          * obd_type and just use the values from there. */
 275         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 276                 rq_portal = OST_REQUEST_PORTAL;
 277                 rp_portal = OSC_REPLY_PORTAL;
 278                 connect_op = OST_CONNECT;
 279                 cli->cl_sp_me = LUSTRE_SP_CLI;
 280                 cli->cl_sp_to = LUSTRE_SP_OST;
 281                 ns_type = LDLM_NS_TYPE_OSC;
 282         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 283                    !strcmp(name, LUSTRE_LWP_NAME)) {
 284                 rq_portal = MDS_REQUEST_PORTAL;
 285                 rp_portal = MDC_REPLY_PORTAL;
 286                 connect_op = MDS_CONNECT;
 287                 if (is_lwp_on_ost(cli_name))
 288                         cli->cl_sp_me = LUSTRE_SP_OST;
 289                 else if (is_lwp_on_mdt(cli_name))
 290                         cli->cl_sp_me = LUSTRE_SP_MDT;
 291                 else
 292                         cli->cl_sp_me = LUSTRE_SP_CLI;
 293                 cli->cl_sp_to = LUSTRE_SP_MDT;
 294                 ns_type = LDLM_NS_TYPE_MDC;
 295         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 296                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 297                         /* OSP_on_MDT for other MDTs */
 298                         connect_op = MDS_CONNECT;
 299                         cli->cl_sp_to = LUSTRE_SP_MDT;
 300                         ns_type = LDLM_NS_TYPE_MDC;
 301                         rq_portal = OUT_PORTAL;
 302                 } else {
 303                         /* OSP on MDT for OST */
 304                         connect_op = OST_CONNECT;
 305                         cli->cl_sp_to = LUSTRE_SP_OST;
 306                         ns_type = LDLM_NS_TYPE_OSC;
 307                         rq_portal = OST_REQUEST_PORTAL;
 308                 }
 309                 rp_portal = OSC_REPLY_PORTAL;
 310                 cli->cl_sp_me = LUSTRE_SP_MDT;
 311         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 312                 rq_portal = MGS_REQUEST_PORTAL;
 313                 rp_portal = MGC_REPLY_PORTAL;
 314                 connect_op = MGS_CONNECT;
 315                 cli->cl_sp_me = LUSTRE_SP_MGC;
 316                 cli->cl_sp_to = LUSTRE_SP_MGS;
 317                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 318                 ns_type = LDLM_NS_TYPE_MGC;
 319         } else {
 320                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 321                        name);
 322                 RETURN(-EINVAL);
 323         }
 324
 325         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 326                 CERROR("requires a TARGET UUID\n");
 327                 RETURN(-EINVAL);
 328         }
 329
 330         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 331                 CERROR("client UUID must be less than 38 characters\n");
 332                 RETURN(-EINVAL);
 333         }
 334
 335         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 336                 CERROR("setup requires a SERVER UUID\n");
 337                 RETURN(-EINVAL);
 338         }
 339
 340         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 341                 CERROR("target UUID must be less than 38 characters\n");
 342                 RETURN(-EINVAL);
 343         }
 344
 345         init_rwsem(&cli->cl_sem);
 346         mutex_init(&cli->cl_mgc_mutex);
 347         cli->cl_conn_count = 0;
 348         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 349                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 350                      sizeof(server_uuid)));
 351
 352         cli->cl_dirty_pages = 0;
 353         cli->cl_avail_grant = 0;
 354         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 355         /* cl_dirty_max_pages may be changed at connect time in
 356          * ptlrpc_connect_interpret(). */
 357         client_adjust_max_dirty(cli);
 358         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 359         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 360         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 361         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 362         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 363         spin_lock_init(&cli->cl_loi_list_lock);
 364         atomic_set(&cli->cl_pending_w_pages, 0);
 365         atomic_set(&cli->cl_pending_r_pages, 0);
 366         cli->cl_r_in_flight = 0;
 367         cli->cl_w_in_flight = 0;
 368
 369         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 370         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 371         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 372         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 373         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 374         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 375
 376         /* lru for osc. */
 377         INIT_LIST_HEAD(&cli->cl_lru_osc);
 378         atomic_set(&cli->cl_lru_shrinkers, 0);
 379         atomic_long_set(&cli->cl_lru_busy, 0);
 380         atomic_long_set(&cli->cl_lru_in_list, 0);
 381         INIT_LIST_HEAD(&cli->cl_lru_list);
 382         spin_lock_init(&cli->cl_lru_list_lock);
 383         atomic_long_set(&cli->cl_unstable_count, 0);
 384
 385         init_waitqueue_head(&cli->cl_destroy_waitq);
 386         atomic_set(&cli->cl_destroy_in_flight, 0);
 387 #ifdef ENABLE_CHECKSUM
 388         /* Turn on checksumming by default. */
 389         cli->cl_checksum = 1;
 390         /*
 391          * The supported checksum types will be worked out at connect time
 392          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 393          * through procfs.
 394          */
 395         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 396 #endif
 397         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 398
 399         /* This value may be reduced at connect time in
 400          * ptlrpc_connect_interpret() . We initialize it to only
 401          * 1MB until we know what the performance looks like.
 402          * In the future this should likely be increased. LU-1431 */
 403         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
 404                                           LNET_MTU >> PAGE_CACHE_SHIFT);
 405
 406         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 407          * it will be updated at OSC connection time. */
 408         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 409
 410         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 411                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 412         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 413                 cli->cl_max_rpcs_in_flight = 2;
 414         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 415                 cli->cl_max_rpcs_in_flight = 3;
 416         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 417                 cli->cl_max_rpcs_in_flight = 4;
 418         } else {
 419                 if (osc_on_mdt(obddev->obd_name))
 420                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 421                 else
 422                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 423         }
 424
 425         spin_lock_init(&cli->cl_mod_rpcs_lock);
 426         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 427         cli->cl_max_mod_rpcs_in_flight = 0;
 428         cli->cl_mod_rpcs_in_flight = 0;
 429         cli->cl_close_rpcs_in_flight = 0;
 430         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 431         cli->cl_mod_tag_bitmap = NULL;
 432
 433         if (connect_op == MDS_CONNECT) {
 434                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 435                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 436                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 437                 if (cli->cl_mod_tag_bitmap == NULL)
 438                         GOTO(err, rc = -ENOMEM);
 439         }
 440
 441         rc = ldlm_get_ref();
 442         if (rc) {
 443                 CERROR("ldlm_get_ref failed: %d\n", rc);
 444                 GOTO(err, rc);
 445         }
 446
 447         ptlrpc_init_client(rq_portal, rp_portal, name,
 448                            &obddev->obd_ldlm_client);
 449
 450         imp = class_new_import(obddev);
 451         if (imp == NULL)
 452                 GOTO(err_ldlm, rc = -ENOENT);
 453         imp->imp_client = &obddev->obd_ldlm_client;
 454         imp->imp_connect_op = connect_op;
 455         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 456                LUSTRE_CFG_BUFLEN(lcfg, 1));
 457         class_import_put(imp);
 458
 459         rc = client_import_add_conn(imp, &server_uuid, 1);
 460         if (rc) {
 461                 CERROR("can't add initial connection\n");
 462                 GOTO(err_import, rc);
 463         }
 464
 465         cli->cl_import = imp;
 466         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 467         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 468
 469         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 470                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 471                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 472                                name, obddev->obd_name,
 473                                cli->cl_target_uuid.uuid);
 474                         spin_lock(&imp->imp_lock);
 475                         imp->imp_deactive = 1;
 476                         spin_unlock(&imp->imp_lock);
 477                 }
 478         }
 479
 480         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 481                                                    LDLM_NAMESPACE_CLIENT,
 482                                                    LDLM_NAMESPACE_GREEDY,
 483                                                    ns_type);
 484         if (obddev->obd_namespace == NULL) {
 485                 CERROR("Unable to create client namespace - %s\n",
 486                        obddev->obd_name);
 487                 GOTO(err_import, rc = -ENOMEM);
 488         }
 489
 490         RETURN(rc);
 491
 492 err_import:
 493         class_destroy_import(imp);
 494 err_ldlm:
 495         ldlm_put_ref();
 496 err:
 497         if (cli->cl_mod_tag_bitmap != NULL)
 498                 OBD_FREE(cli->cl_mod_tag_bitmap,
 499                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 500         cli->cl_mod_tag_bitmap = NULL;
 501         RETURN(rc);
 502
 503 }
 504 EXPORT_SYMBOL(client_obd_setup);
 505
 506 int client_obd_cleanup(struct obd_device *obddev)
 507 {
 508         struct client_obd *cli = &obddev->u.cli;
 509         ENTRY;
 510
 511         ldlm_namespace_free_post(obddev->obd_namespace);
 512         obddev->obd_namespace = NULL;
 513
 514         obd_cleanup_client_import(obddev);
 515         LASSERT(obddev->u.cli.cl_import == NULL);
 516
 517         ldlm_put_ref();
 518
 519         if (cli->cl_mod_tag_bitmap != NULL)
 520                 OBD_FREE(cli->cl_mod_tag_bitmap,
 521                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 522         cli->cl_mod_tag_bitmap = NULL;
 523
 524         RETURN(0);
 525 }
 526 EXPORT_SYMBOL(client_obd_cleanup);
 527
 528 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 529 int client_connect_import(const struct lu_env *env,
 530                           struct obd_export **exp,
 531                           struct obd_device *obd, struct obd_uuid *cluuid,
 532                           struct obd_connect_data *data, void *localdata)
 533 {
 534         struct client_obd       *cli    = &obd->u.cli;
 535         struct obd_import       *imp    = cli->cl_import;
 536         struct obd_connect_data *ocd;
 537         struct lustre_handle    conn    = { 0 };
 538         int                     rc;
 539         bool                    is_mdc = false;
 540         ENTRY;
 541
 542         *exp = NULL;
 543         down_write(&cli->cl_sem);
 544         if (cli->cl_conn_count > 0)
 545                 GOTO(out_sem, rc = -EALREADY);
 546
 547         rc = class_connect(&conn, obd, cluuid);
 548         if (rc)
 549                 GOTO(out_sem, rc);
 550
 551         cli->cl_conn_count++;
 552         *exp = class_conn2export(&conn);
 553
 554         LASSERT(obd->obd_namespace);
 555
 556         imp->imp_dlm_handle = conn;
 557         rc = ptlrpc_init_import(imp);
 558         if (rc != 0)
 559                 GOTO(out_ldlm, rc);
 560
 561         ocd = &imp->imp_connect_data;
 562         if (data) {
 563                 *ocd = *data;
 564                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
 565                                  LUSTRE_MDC_NAME, 3) == 0;
 566                 if (is_mdc)
 567                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
 568                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 569         }
 570
 571         rc = ptlrpc_connect_import(imp);
 572         if (rc != 0) {
 573                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 574                 GOTO(out_ldlm, rc);
 575         }
 576         LASSERT(*exp != NULL && (*exp)->exp_connection);
 577
 578         if (data) {
 579                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 580                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 581                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 582                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 583                 /* clear the flag as it was not set and is not known
 584                  * by upper layers */
 585                 if (is_mdc)
 586                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 587         }
 588
 589         ptlrpc_pinger_add_import(imp);
 590
 591         EXIT;
 592
 593         if (rc) {
 594 out_ldlm:
 595                 cli->cl_conn_count--;
 596                 class_disconnect(*exp);
 597                 *exp = NULL;
 598         }
 599 out_sem:
 600         up_write(&cli->cl_sem);
 601
 602         return rc;
 603 }
 604 EXPORT_SYMBOL(client_connect_import);
 605
 606 int client_disconnect_export(struct obd_export *exp)
 607 {
 608         struct obd_device *obd = class_exp2obd(exp);
 609         struct client_obd *cli;
 610         struct obd_import *imp;
 611         int rc = 0, err;
 612         ENTRY;
 613
 614         if (!obd) {
 615                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 616                        exp, exp ? exp->exp_handle.h_cookie : -1);
 617                 RETURN(-EINVAL);
 618         }
 619
 620         cli = &obd->u.cli;
 621         imp = cli->cl_import;
 622
 623         down_write(&cli->cl_sem);
 624         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 625                 cli->cl_conn_count);
 626
 627         if (cli->cl_conn_count == 0) {
 628                 CERROR("disconnecting disconnected device (%s)\n",
 629                        obd->obd_name);
 630                 GOTO(out_disconnect, rc = -EINVAL);
 631         }
 632
 633         cli->cl_conn_count--;
 634         if (cli->cl_conn_count != 0)
 635                 GOTO(out_disconnect, rc = 0);
 636
 637         /* Mark import deactivated now, so we don't try to reconnect if any
 638          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 639          * fully deactivate the import, or that would drop all requests. */
 640         spin_lock(&imp->imp_lock);
 641         imp->imp_deactive = 1;
 642         spin_unlock(&imp->imp_lock);
 643
 644         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 645          * delete it regardless.  (It's safe to delete an import that was
 646          * never added.) */
 647         (void)ptlrpc_pinger_del_import(imp);
 648
 649         if (obd->obd_namespace != NULL) {
 650                 /* obd_force == local only */
 651                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 652                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 653                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 654         }
 655
 656         /* There's no need to hold sem while disconnecting an import,
 657          * and it may actually cause deadlock in GSS. */
 658         up_write(&cli->cl_sem);
 659         rc = ptlrpc_disconnect_import(imp, 0);
 660         down_write(&cli->cl_sem);
 661
 662         ptlrpc_invalidate_import(imp);
 663
 664         EXIT;
 665
 666 out_disconnect:
 667         /* Use server style - class_disconnect should be always called for
 668          * o_disconnect. */
 669         err = class_disconnect(exp);
 670         if (!rc && err)
 671                 rc = err;
 672
 673         up_write(&cli->cl_sem);
 674
 675         RETURN(rc);
 676 }
 677 EXPORT_SYMBOL(client_disconnect_export);
 678
 679 #ifdef HAVE_SERVER_SUPPORT
 680 int server_disconnect_export(struct obd_export *exp)
 681 {
 682         int rc;
 683         ENTRY;
 684
 685         /* Disconnect early so that clients can't keep using export. */
 686         rc = class_disconnect(exp);
 687         /* Close import to avoid sending any requests. */
 688         if (exp->exp_imp_reverse)
 689                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 690
 691         if (exp->exp_obd->obd_namespace != NULL)
 692                 ldlm_cancel_locks_for_export(exp);
 693
 694         /* complete all outstanding replies */
 695         spin_lock(&exp->exp_lock);
 696         while (!list_empty(&exp->exp_outstanding_replies)) {
 697                 struct ptlrpc_reply_state *rs =
 698                         list_entry(exp->exp_outstanding_replies.next,
 699                                        struct ptlrpc_reply_state, rs_exp_list);
 700                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 701
 702                 spin_lock(&svcpt->scp_rep_lock);
 703
 704                 list_del_init(&rs->rs_exp_list);
 705                 spin_lock(&rs->rs_lock);
 706                 ptlrpc_schedule_difficult_reply(rs);
 707                 spin_unlock(&rs->rs_lock);
 708
 709                 spin_unlock(&svcpt->scp_rep_lock);
 710         }
 711         spin_unlock(&exp->exp_lock);
 712
 713         RETURN(rc);
 714 }
 715 EXPORT_SYMBOL(server_disconnect_export);
 716
 717 /* --------------------------------------------------------------------------
 718  * from old lib/target.c
 719  * -------------------------------------------------------------------------- */
 720
 721 static int target_handle_reconnect(struct lustre_handle *conn,
 722                                    struct obd_export *exp,
 723                                    struct obd_uuid *cluuid)
 724 {
 725         ENTRY;
 726
 727         if (exp->exp_connection && exp->exp_imp_reverse) {
 728                 struct lustre_handle *hdl;
 729                 struct obd_device *target;
 730
 731                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 732                 target = exp->exp_obd;
 733
 734                 /* Might be a re-connect after a partition. */
 735                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 736                         if (target->obd_recovering) {
 737                                 int timeout = cfs_duration_sec(cfs_time_sub(
 738                                         cfs_timer_deadline(
 739                                         &target->obd_recovery_timer),
 740                                         cfs_time_current()));
 741
 742                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 743                                         "ing, waiting for %d clients in recov"
 744                                         "ery for %d:%.02d\n", target->obd_name,
 745                                         obd_uuid2str(&exp->exp_client_uuid),
 746                                         obd_export_nid2str(exp),
 747                                         target->obd_max_recoverable_clients,
 748                                         timeout / 60, timeout % 60);
 749                         } else {
 750                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 751                                         "reconnecting\n", target->obd_name,
 752                                         obd_uuid2str(&exp->exp_client_uuid),
 753                                         obd_export_nid2str(exp));
 754                         }
 755
 756                         conn->cookie = exp->exp_handle.h_cookie;
 757                         /* target_handle_connect() treats EALREADY and
 758                          * -EALREADY differently.  EALREADY means we are
 759                          * doing a valid reconnect from the same client. */
 760                         RETURN(EALREADY);
 761                 } else {
 762                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 763                                       "with handle "LPX64". Rejecting client "
 764                                       "with the same UUID trying to reconnect "
 765                                       "with handle "LPX64"\n", target->obd_name,
 766                                       obd_uuid2str(&exp->exp_client_uuid),
 767                                       obd_export_nid2str(exp),
 768                                       hdl->cookie, conn->cookie);
 769                         memset(conn, 0, sizeof *conn);
 770                         /* target_handle_connect() treats EALREADY and
 771                          * -EALREADY differently.  -EALREADY is an error
 772                          * (same UUID, different handle). */
 773                         RETURN(-EALREADY);
 774                 }
 775         }
 776
 777         conn->cookie = exp->exp_handle.h_cookie;
 778         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 779                cluuid->uuid, exp, conn->cookie);
 780         RETURN(0);
 781 }
 782
 783 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 784                           int error)
 785 {
 786         struct obd_export *exp = cb_data;
 787
 788         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 789                obd->obd_name, exp->exp_client_uuid.uuid);
 790
 791         spin_lock(&exp->exp_lock);
 792         exp->exp_need_sync = 0;
 793         spin_unlock(&exp->exp_lock);
 794         class_export_cb_put(exp);
 795 }
 796
 797 static void
 798 check_and_start_recovery_timer(struct obd_device *obd,
 799                                struct ptlrpc_request *req, int new_client);
 800
 801 int target_handle_connect(struct ptlrpc_request *req)
 802 {
 803         struct obd_device *target = NULL, *targref = NULL;
 804         struct obd_export *export = NULL;
 805         struct obd_import *revimp;
 806         struct obd_import *tmp_imp = NULL;
 807         struct lustre_handle conn;
 808         struct lustre_handle *tmp;
 809         struct obd_uuid tgtuuid;
 810         struct obd_uuid cluuid;
 811         struct obd_uuid remote_uuid;
 812         char *str;
 813         int rc = 0;
 814         char *target_start;
 815         int target_len;
 816         bool     mds_conn = false, lw_client = false;
 817         bool     mds_mds_conn = false;
 818         bool     new_mds_mds_conn = false;
 819         struct obd_connect_data *data, *tmpdata;
 820         int size, tmpsize;
 821         lnet_nid_t *client_nid = NULL;
 822         ENTRY;
 823
 824         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 825
 826         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 827         if (str == NULL) {
 828                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 829                 GOTO(out, rc = -EINVAL);
 830         }
 831
 832         obd_str2uuid(&tgtuuid, str);
 833         target = class_uuid2obd(&tgtuuid);
 834         if (!target)
 835                 target = class_name2obd(str);
 836
 837         if (!target) {
 838                 deuuidify(str, NULL, &target_start, &target_len);
 839                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 840                                    "from %s (no target). If you are running "
 841                                    "an HA pair check that the target is "
 842                                    "mounted on the other server.\n", str,
 843                                    libcfs_nid2str(req->rq_peer.nid));
 844                 GOTO(out, rc = -ENODEV);
 845         }
 846
 847         spin_lock(&target->obd_dev_lock);
 848         if (target->obd_stopping || !target->obd_set_up) {
 849                 spin_unlock(&target->obd_dev_lock);
 850
 851                 deuuidify(str, NULL, &target_start, &target_len);
 852                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 853                               target_len, target_start,
 854                               libcfs_nid2str(req->rq_peer.nid),
 855                               (target->obd_stopping ?
 856                                "stopping" : "not set up"));
 857                 GOTO(out, rc = -ENODEV);
 858         }
 859
 860         if (target->obd_no_conn) {
 861                 spin_unlock(&target->obd_dev_lock);
 862
 863                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 864                                "from %s\n", target->obd_name,
 865                                libcfs_nid2str(req->rq_peer.nid));
 866                 GOTO(out, rc = -EAGAIN);
 867         }
 868
 869         /* Make sure the target isn't cleaned up while we're here. Yes,
 870          * there's still a race between the above check and our incref here.
 871          * Really, class_uuid2obd should take the ref. */
 872         targref = class_incref(target, __FUNCTION__, current);
 873
 874         target->obd_conn_inprogress++;
 875         spin_unlock(&target->obd_dev_lock);
 876
 877         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 878         if (str == NULL) {
 879                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 880                 GOTO(out, rc = -EINVAL);
 881         }
 882
 883         obd_str2uuid(&cluuid, str);
 884
 885         /* XXX Extract a nettype and format accordingly. */
 886         switch (sizeof(lnet_nid_t)) {
 887         /* NB the casts only avoid compiler warnings. */
 888         case 8:
 889                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 890                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 891                 break;
 892         case 4:
 893                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 894                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 895                 break;
 896         default:
 897                 LBUG();
 898         }
 899
 900         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 901         if (tmp == NULL)
 902                 GOTO(out, rc = -EPROTO);
 903
 904         conn = *tmp;
 905
 906         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 907                                     RCL_CLIENT);
 908         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 909         if (!data)
 910                 GOTO(out, rc = -EPROTO);
 911
 912         rc = req_capsule_server_pack(&req->rq_pill);
 913         if (rc)
 914                 GOTO(out, rc);
 915
 916 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 917         /* Don't allow clients to connect that are using old 1.8 format
 918          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
 919          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
 920          * FULL20 flag should be set on all connections since 2.0, but no
 921          * longer affects behaviour.
 922          *
 923          * Later this check will be disabled and the flag can be retired
 924          * completely once interop with 3.0 is no longer needed.
 925          */
 926         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
 927                 GOTO(out, rc = -EPROTO);
 928 #endif
 929
 930         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 931                 if (data->ocd_version < LUSTRE_VERSION_CODE -
 932                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 933                     data->ocd_version > LUSTRE_VERSION_CODE +
 934                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 935                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 936                                   "libclient connection attempt",
 937                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 938                                   "old" : "new",
 939                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 940                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 941                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 942                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 943                         data = req_capsule_server_sized_get(&req->rq_pill,
 944                                                             &RMF_CONNECT_DATA,
 945                                     offsetof(typeof(*data), ocd_version) +
 946                                              sizeof(data->ocd_version));
 947                         if (data) {
 948                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 949                                 data->ocd_version = LUSTRE_VERSION_CODE;
 950                         }
 951                         GOTO(out, rc = -EPROTO);
 952                 }
 953         }
 954
 955         /* Note: lw_client is needed in MDS-MDS failover during update log
 956          * processing, so we needs to allow lw_client to be connected at
 957          * anytime, instead of only the initial connection */
 958         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
 959
 960         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
 961                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
 962                 mds_mds_conn = (data->ocd_connect_flags &
 963                                 OBD_CONNECT_MDS_MDS) != 0;
 964
 965                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
 966                  * for Imperative Recovery connection from MGC to MGS.
 967                  *
 968                  * Via check OBD_CONNECT_FID, we can distinguish whether
 969                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
 970                  * MGC or MDT. */
 971                 if (!lw_client &&
 972                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
 973                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
 974                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
 975                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
 976                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
 977                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
 978
 979                         /* We do not support the MDT-MDT interoperations with
 980                          * different version MDT because of protocol changes. */
 981                         if (unlikely(major != LUSTRE_MAJOR ||
 982                                      minor != LUSTRE_MINOR ||
 983                                      abs(patch - LUSTRE_PATCH) > 3)) {
 984                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
 985                                         "connection from different version MDT "
 986                                         "(%d.%d.%d.%d) %s %s\n",
 987                                         target->obd_name, LUSTRE_MAJOR,
 988                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
 989                                         major, minor, patch,
 990                                         OBD_OCD_VERSION_FIX(data->ocd_version),
 991                                         libcfs_nid2str(req->rq_peer.nid), str);
 992
 993                                 GOTO(out, rc = -EPROTO);
 994                         }
 995                 }
 996         }
 997
 998         /* lctl gets a backstage, all-access pass. */
 999         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1000                 goto dont_check_exports;
1001
1002         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1003         if (!export)
1004                 goto no_export;
1005
1006         /* We've found an export in the hash. */
1007
1008         spin_lock(&export->exp_lock);
1009
1010         if (export->exp_connecting) { /* bug 9635, et. al. */
1011                 spin_unlock(&export->exp_lock);
1012                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1013                               export->exp_obd->obd_name, export,
1014                               libcfs_nid2str(req->rq_peer.nid));
1015                 class_export_put(export);
1016                 export = NULL;
1017                 rc = -EALREADY;
1018         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1019                 spin_unlock(&export->exp_lock);
1020                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1021                         /* MDS or LWP reconnected after failover. */
1022                         LCONSOLE_WARN("%s: Received %s connection from "
1023                             "%s, removing former export from %s\n",
1024                             target->obd_name, mds_conn ? "MDS" : "LWP",
1025                             libcfs_nid2str(req->rq_peer.nid),
1026                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1027                 else
1028                         /* New MDS connection from the same NID. */
1029                         LCONSOLE_WARN("%s: Received new %s connection from "
1030                                 "%s, removing former export from same NID\n",
1031                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1032                                 libcfs_nid2str(req->rq_peer.nid));
1033                 class_fail_export(export);
1034                 class_export_put(export);
1035                 export = NULL;
1036                 rc = 0;
1037         } else if (export->exp_connection != NULL &&
1038                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1039                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1040                     MSG_CONNECT_INITIAL)) {
1041                 spin_unlock(&export->exp_lock);
1042                 /* In MDS failover we have static UUID but NID can change. */
1043                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1044                               "existing nid %s is already connected\n",
1045                               target->obd_name, cluuid.uuid,
1046                               libcfs_nid2str(req->rq_peer.nid),
1047                               libcfs_nid2str(
1048                                       export->exp_connection->c_peer.nid));
1049                 rc = -EALREADY;
1050                 class_export_put(export);
1051                 export = NULL;
1052         } else {
1053                 export->exp_connecting = 1;
1054                 spin_unlock(&export->exp_lock);
1055                 LASSERT(export->exp_obd == target);
1056
1057                 rc = target_handle_reconnect(&conn, export, &cluuid);
1058         }
1059
1060         /* If we found an export, we already unlocked. */
1061         if (!export) {
1062 no_export:
1063                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1064         } else if (req->rq_export == NULL &&
1065                    atomic_read(&export->exp_rpc_count) > 0) {
1066                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1067                               "still busy with %d references\n",
1068                               target->obd_name, cluuid.uuid,
1069                               libcfs_nid2str(req->rq_peer.nid),
1070                               atomic_read(&export->exp_refcount));
1071                 GOTO(out, rc = -EBUSY);
1072         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1073                 if (!strstr(cluuid.uuid, "mdt"))
1074                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1075                                       "known client %s (at %s) because it "
1076                                       "is indicating it is a new client",
1077                                       target->obd_name, cluuid.uuid,
1078                                       libcfs_nid2str(req->rq_peer.nid));
1079                 GOTO(out, rc = -EALREADY);
1080         } else {
1081                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1082         }
1083
1084         if (rc < 0) {
1085                 GOTO(out, rc);
1086         }
1087
1088         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1089                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1090               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1091               export, (long)cfs_time_current_sec(),
1092               export ? (long)export->exp_last_request_time : 0);
1093
1094         /* If this is the first time a client connects, reset the recovery
1095          * timer. Discard lightweight connections which might be local. */
1096         if (!lw_client && rc == 0 && target->obd_recovering)
1097                 check_and_start_recovery_timer(target, req, export == NULL);
1098
1099         /* We want to handle EALREADY but *not* -EALREADY from
1100          * target_handle_reconnect(), return reconnection state in a flag. */
1101         if (rc == EALREADY) {
1102                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1103                 rc = 0;
1104         } else {
1105                 LASSERT(rc == 0);
1106         }
1107
1108         /* Tell the client if we support replayable requests. */
1109         if (target->obd_replayable)
1110                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1111         client_nid = &req->rq_peer.nid;
1112
1113         if (export == NULL) {
1114                 /* allow lightweight connections during recovery */
1115                 /* allow "new" MDT to be connected during recovery, since we
1116                  * need retrieve recovery update records from it */
1117                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1118                         cfs_time_t t;
1119                         int     c; /* connected */
1120                         int     i; /* in progress */
1121                         int     k; /* known */
1122                         int     s; /* stale/evicted */
1123
1124                         c = atomic_read(&target->obd_connected_clients);
1125                         i = atomic_read(&target->obd_lock_replay_clients);
1126                         k = target->obd_max_recoverable_clients;
1127                         s = target->obd_stale_clients;
1128                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1129                         t = cfs_time_sub(t, cfs_time_current());
1130                         t = cfs_duration_sec(t);
1131                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1132                                       "(at %s), waiting for %d known clients "
1133                                       "(%d recovered, %d in progress, and %d "
1134                                       "evicted) to recover in %d:%.02d\n",
1135                                       target->obd_name, cluuid.uuid,
1136                                       libcfs_nid2str(req->rq_peer.nid), k,
1137                                       c - i, i, s, (int)t / 60,
1138                                       (int)t % 60);
1139                         rc = -EBUSY;
1140                 } else {
1141 dont_check_exports:
1142                         rc = obd_connect(req->rq_svc_thread->t_env,
1143                                          &export, target, &cluuid, data,
1144                                          client_nid);
1145                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1146                                 lustre_msg_add_op_flags(req->rq_repmsg,
1147                                                         MSG_CONNECT_RECOVERING);
1148                         if (rc == 0)
1149                                 conn.cookie = export->exp_handle.h_cookie;
1150
1151                         if (mds_mds_conn)
1152                                 new_mds_mds_conn = true;
1153                 }
1154         } else {
1155                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1156                                    export, target, &cluuid, data, client_nid);
1157         }
1158         if (rc)
1159                 GOTO(out, rc);
1160
1161         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1162         data->ocd_instance = target->u.obt.obt_instance;
1163
1164         /* Return only the parts of obd_connect_data that we understand, so the
1165          * client knows that we don't understand the rest. */
1166         if (data) {
1167                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1168                                                RCL_SERVER);
1169                 tmpdata = req_capsule_server_get(&req->rq_pill,
1170                                                  &RMF_CONNECT_DATA);
1171                 /* Don't use struct assignment here, because the client reply
1172                  * buffer may be smaller/larger than the local struct
1173                  * obd_connect_data. */
1174                 memcpy(tmpdata, data, min(tmpsize, size));
1175         }
1176
1177         /* If all else goes well, this is our RPC return code. */
1178         req->rq_status = 0;
1179
1180         lustre_msg_set_handle(req->rq_repmsg, &conn);
1181
1182         /* If the client and the server are the same node, we will already
1183          * have an export that really points to the client's DLM export,
1184          * because we have a shared handles table.
1185          *
1186          * XXX this will go away when shaver stops sending the "connect" handle
1187          * in the real "remote handle" field of the request --phik 24 Apr 2003
1188          */
1189         ptlrpc_request_change_export(req, export);
1190
1191         spin_lock(&export->exp_lock);
1192         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1193                 spin_unlock(&export->exp_lock);
1194                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1195                        "or equal conn_cnt: %d >= %d\n",
1196                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1197                        export->exp_conn_cnt,
1198                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1199
1200                 GOTO(out, rc = -EALREADY);
1201         }
1202         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1203         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1204
1205         /* Don't evict liblustre clients for not pinging. */
1206         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1207                 export->exp_libclient = 1;
1208                 spin_unlock(&export->exp_lock);
1209
1210                 spin_lock(&target->obd_dev_lock);
1211                 list_del_init(&export->exp_obd_chain_timed);
1212                 spin_unlock(&target->obd_dev_lock);
1213         } else {
1214                 spin_unlock(&export->exp_lock);
1215         }
1216
1217         if (export->exp_connection != NULL) {
1218                 /* Check to see if connection came from another NID. */
1219                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1220                     !hlist_unhashed(&export->exp_nid_hash))
1221                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1222                                      &export->exp_connection->c_peer.nid,
1223                                      &export->exp_nid_hash);
1224
1225                 ptlrpc_connection_put(export->exp_connection);
1226         }
1227
1228         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1229                                                        req->rq_self,
1230                                                        &remote_uuid);
1231         if (hlist_unhashed(&export->exp_nid_hash)) {
1232                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1233                              &export->exp_connection->c_peer.nid,
1234                              &export->exp_nid_hash);
1235         }
1236
1237         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1238                 int has_transno;
1239                 __u64 transno = data->ocd_transno;
1240
1241                 spin_lock(&export->exp_lock);
1242                 /* possible race with class_disconnect_stale_exports,
1243                  * export may be already in the eviction process */
1244                 if (export->exp_failed) {
1245                         spin_unlock(&export->exp_lock);
1246                         GOTO(out, rc = -ENODEV);
1247                 }
1248                 export->exp_in_recovery = 1;
1249                 export->exp_req_replay_needed = 1;
1250                 export->exp_lock_replay_needed = 1;
1251                 spin_unlock(&export->exp_lock);
1252
1253                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1254                                  MSG_CONNECT_TRANSNO);
1255                 if (has_transno && transno == 0)
1256                         CWARN("Connect with zero transno!\n");
1257
1258                 if (has_transno && transno > 0 &&
1259                     transno < target->obd_next_recovery_transno &&
1260                     transno > target->obd_last_committed) {
1261                         /* Another way is to use cmpxchg() to be lock-free. */
1262                         spin_lock(&target->obd_recovery_task_lock);
1263                         if (transno < target->obd_next_recovery_transno)
1264                                 target->obd_next_recovery_transno = transno;
1265                         spin_unlock(&target->obd_recovery_task_lock);
1266                 }
1267
1268                 atomic_inc(&target->obd_req_replay_clients);
1269                 atomic_inc(&target->obd_lock_replay_clients);
1270                 /* Note: MDS-MDS connection is allowed to be connected during
1271                  * recovery, no matter if the exports needs to be recoveried.
1272                  * Because we need retrieve updates logs from all other MDTs.
1273                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1274                  * also needs to be increased to match other recovery checking
1275                  * condition. */
1276                 if (new_mds_mds_conn)
1277                         target->obd_max_recoverable_clients++;
1278                 if (atomic_inc_return(&target->obd_connected_clients) ==
1279                     target->obd_max_recoverable_clients)
1280                         wake_up(&target->obd_next_transno_waitq);
1281         }
1282
1283         /* Tell the client we're in recovery, when client is involved in it. */
1284         if (target->obd_recovering && !lw_client)
1285                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1286
1287         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1288         conn = *tmp;
1289
1290         /* Return -ENOTCONN in case of errors to let client reconnect. */
1291         revimp = class_new_import(target);
1292         if (revimp == NULL) {
1293                 CERROR("fail to alloc new reverse import.\n");
1294                 GOTO(out, rc = -ENOTCONN);
1295         }
1296
1297         spin_lock(&export->exp_lock);
1298         if (export->exp_imp_reverse != NULL)
1299                 /* destroyed import can be still referenced in ctxt */
1300                 tmp_imp = export->exp_imp_reverse;
1301         export->exp_imp_reverse = revimp;
1302         spin_unlock(&export->exp_lock);
1303
1304         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1305         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1306         revimp->imp_remote_handle = conn;
1307         revimp->imp_dlm_fake = 1;
1308         revimp->imp_state = LUSTRE_IMP_FULL;
1309
1310         /* Unknown versions will be caught in
1311          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1312         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1313
1314         if (data->ocd_connect_flags & OBD_CONNECT_AT)
1315                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1316         else
1317                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1318
1319         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1320
1321         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1322         if (rc) {
1323                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1324                 spin_lock(&export->exp_lock);
1325                 export->exp_imp_reverse = NULL;
1326                 spin_unlock(&export->exp_lock);
1327                 class_destroy_import(revimp);
1328         }
1329
1330         class_import_put(revimp);
1331
1332 out:
1333         if (tmp_imp != NULL)
1334                 client_destroy_import(tmp_imp);
1335         if (export) {
1336                 spin_lock(&export->exp_lock);
1337                 export->exp_connecting = 0;
1338                 spin_unlock(&export->exp_lock);
1339
1340                 class_export_put(export);
1341         }
1342         if (targref) {
1343                 spin_lock(&target->obd_dev_lock);
1344                 target->obd_conn_inprogress--;
1345                 spin_unlock(&target->obd_dev_lock);
1346
1347                 class_decref(targref, __func__, current);
1348         }
1349         if (rc)
1350                 req->rq_status = rc;
1351         RETURN(rc);
1352 }
1353
1354 int target_handle_disconnect(struct ptlrpc_request *req)
1355 {
1356         int rc;
1357         ENTRY;
1358
1359         rc = req_capsule_server_pack(&req->rq_pill);
1360         if (rc)
1361                 RETURN(rc);
1362
1363         /* Keep the rq_export around so we can send the reply. */
1364         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1365
1366         RETURN(0);
1367 }
1368
1369 void target_destroy_export(struct obd_export *exp)
1370 {
1371         struct obd_import       *imp = NULL;
1372         /* exports created from last_rcvd data, and "fake"
1373            exports created by lctl don't have an import */
1374         spin_lock(&exp->exp_lock);
1375         if (exp->exp_imp_reverse != NULL) {
1376                 imp = exp->exp_imp_reverse;
1377                 exp->exp_imp_reverse = NULL;
1378         }
1379         spin_unlock(&exp->exp_lock);
1380         if (imp != NULL)
1381                 client_destroy_import(imp);
1382
1383         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1384         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1385         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1386         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1387 }
1388 EXPORT_SYMBOL(target_destroy_export);
1389
1390 /*
1391  * Recovery functions
1392  */
1393 static void target_request_copy_get(struct ptlrpc_request *req)
1394 {
1395         class_export_rpc_inc(req->rq_export);
1396         LASSERT(list_empty(&req->rq_list));
1397         INIT_LIST_HEAD(&req->rq_replay_list);
1398
1399         /* Increase refcount to keep request in queue. */
1400         atomic_inc(&req->rq_refcount);
1401         /* Let export know it has replays to be handled. */
1402         atomic_inc(&req->rq_export->exp_replay_count);
1403 }
1404
1405 static void target_request_copy_put(struct ptlrpc_request *req)
1406 {
1407         LASSERT(list_empty(&req->rq_replay_list));
1408         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1409
1410         atomic_dec(&req->rq_export->exp_replay_count);
1411         class_export_rpc_dec(req->rq_export);
1412         ptlrpc_server_drop_request(req);
1413 }
1414
1415 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1416 {
1417         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1418         struct obd_export     *exp = req->rq_export;
1419         struct ptlrpc_request *reqiter;
1420         struct ptlrpc_request *dup_req = NULL;
1421         int                    dup = 0;
1422
1423         LASSERT(exp);
1424
1425         spin_lock(&exp->exp_lock);
1426         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1427                                 rq_replay_list) {
1428                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1429                         dup_req = reqiter;
1430                         dup = 1;
1431                         break;
1432                 }
1433         }
1434
1435         if (dup) {
1436                 /* We expect it with RESENT and REPLAY flags. */
1437                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1438                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1439                         CERROR("invalid flags %x of resent replay\n",
1440                                lustre_msg_get_flags(req->rq_reqmsg));
1441
1442                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1443                         __u32 new_conn;
1444
1445                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1446                         if (new_conn >
1447                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1448                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1449                                                         new_conn);
1450                 }
1451         } else {
1452                 list_add_tail(&req->rq_replay_list,
1453                                   &exp->exp_req_replay_queue);
1454         }
1455
1456         spin_unlock(&exp->exp_lock);
1457         return dup;
1458 }
1459
1460 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1461 {
1462         LASSERT(!list_empty(&req->rq_replay_list));
1463         LASSERT(req->rq_export);
1464
1465         spin_lock(&req->rq_export->exp_lock);
1466         list_del_init(&req->rq_replay_list);
1467         spin_unlock(&req->rq_export->exp_lock);
1468 }
1469
1470 static void target_finish_recovery(struct lu_target *lut)
1471 {
1472         struct obd_device *obd = lut->lut_obd;
1473         ENTRY;
1474
1475         /* Only log a recovery message when recovery has occurred. */
1476         if (obd->obd_recovery_start) {
1477                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1478                                         obd->obd_recovery_start);
1479                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1480                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1481                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1482                         obd->obd_max_recoverable_clients,
1483                         atomic_read(&obd->obd_connected_clients),
1484                         obd->obd_stale_clients,
1485                         obd->obd_stale_clients == 1 ? "was" : "were");
1486         }
1487
1488         ldlm_reprocess_all_ns(obd->obd_namespace);
1489         spin_lock(&obd->obd_recovery_task_lock);
1490         if (!list_empty(&obd->obd_req_replay_queue) ||
1491             !list_empty(&obd->obd_lock_replay_queue) ||
1492             !list_empty(&obd->obd_final_req_queue)) {
1493                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1494                        obd->obd_name,
1495                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1496                        list_empty(&obd->obd_lock_replay_queue) ? \
1497                                "" : "lock ",
1498                        list_empty(&obd->obd_final_req_queue) ? \
1499                                "" : "final ");
1500                 spin_unlock(&obd->obd_recovery_task_lock);
1501                 LBUG();
1502         }
1503         spin_unlock(&obd->obd_recovery_task_lock);
1504
1505         if (lut->lut_tdtd != NULL &&
1506             !list_empty(&lut->lut_tdtd->tdtd_replay_list))
1507                 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1508
1509         obd->obd_recovery_end = cfs_time_current_sec();
1510
1511         /* When recovery finished, cleanup orphans on MDS and OST. */
1512         if (OBT(obd) && OBP(obd, postrecov)) {
1513                 int rc = OBP(obd, postrecov)(obd);
1514                 if (rc < 0)
1515                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1516                                       obd->obd_name, rc);
1517         }
1518         EXIT;
1519 }
1520
1521 static void abort_req_replay_queue(struct obd_device *obd)
1522 {
1523         struct ptlrpc_request *req, *n;
1524         struct list_head abort_list;
1525
1526         INIT_LIST_HEAD(&abort_list);
1527         spin_lock(&obd->obd_recovery_task_lock);
1528         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1529         spin_unlock(&obd->obd_recovery_task_lock);
1530         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1531                 DEBUG_REQ(D_WARNING, req, "aborted:");
1532                 req->rq_status = -ENOTCONN;
1533                 if (ptlrpc_error(req)) {
1534                         DEBUG_REQ(D_ERROR, req,
1535                                   "failed abort_req_reply; skipping");
1536                 }
1537                 target_exp_dequeue_req_replay(req);
1538                 target_request_copy_put(req);
1539         }
1540 }
1541
1542 static void abort_lock_replay_queue(struct obd_device *obd)
1543 {
1544         struct ptlrpc_request *req, *n;
1545         struct list_head abort_list;
1546
1547         INIT_LIST_HEAD(&abort_list);
1548         spin_lock(&obd->obd_recovery_task_lock);
1549         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1550         spin_unlock(&obd->obd_recovery_task_lock);
1551         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1552                 DEBUG_REQ(D_ERROR, req, "aborted:");
1553                 req->rq_status = -ENOTCONN;
1554                 if (ptlrpc_error(req)) {
1555                         DEBUG_REQ(D_ERROR, req,
1556                                   "failed abort_lock_reply; skipping");
1557                 }
1558                 target_request_copy_put(req);
1559         }
1560 }
1561
1562 /* Called from a cleanup function if the device is being cleaned up
1563    forcefully.  The exports should all have been disconnected already,
1564    the only thing left to do is
1565      - clear the recovery flags
1566      - cancel the timer
1567      - free queued requests and replies, but don't send replies
1568    Because the obd_stopping flag is set, no new requests should be received.
1569
1570 */
1571 void target_cleanup_recovery(struct obd_device *obd)
1572 {
1573         struct ptlrpc_request *req, *n;
1574         struct list_head clean_list;
1575         ENTRY;
1576
1577         INIT_LIST_HEAD(&clean_list);
1578         spin_lock(&obd->obd_dev_lock);
1579         if (!obd->obd_recovering) {
1580                 spin_unlock(&obd->obd_dev_lock);
1581                 EXIT;
1582                 return;
1583         }
1584         obd->obd_recovering = obd->obd_abort_recovery = 0;
1585         obd->obd_force_abort_recovery = 0;
1586         spin_unlock(&obd->obd_dev_lock);
1587
1588         spin_lock(&obd->obd_recovery_task_lock);
1589         target_cancel_recovery_timer(obd);
1590         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1591         spin_unlock(&obd->obd_recovery_task_lock);
1592
1593         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1594                 LASSERT(req->rq_reply_state == NULL);
1595                 target_exp_dequeue_req_replay(req);
1596                 target_request_copy_put(req);
1597         }
1598
1599         spin_lock(&obd->obd_recovery_task_lock);
1600         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1601         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1602         spin_unlock(&obd->obd_recovery_task_lock);
1603
1604         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1605                 LASSERT(req->rq_reply_state == NULL);
1606                 target_request_copy_put(req);
1607         }
1608
1609         EXIT;
1610 }
1611 EXPORT_SYMBOL(target_cleanup_recovery);
1612
1613 /* obd_recovery_task_lock should be held */
1614 void target_cancel_recovery_timer(struct obd_device *obd)
1615 {
1616         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1617         cfs_timer_disarm(&obd->obd_recovery_timer);
1618 }
1619
1620 static void target_start_recovery_timer(struct obd_device *obd)
1621 {
1622         if (obd->obd_recovery_start != 0)
1623                 return;
1624
1625         spin_lock(&obd->obd_dev_lock);
1626         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1627             obd->obd_force_abort_recovery) {
1628                 spin_unlock(&obd->obd_dev_lock);
1629                 return;
1630         }
1631
1632         LASSERT(obd->obd_recovery_timeout != 0);
1633
1634         if (obd->obd_recovery_start != 0) {
1635                 spin_unlock(&obd->obd_dev_lock);
1636                 return;
1637         }
1638
1639         cfs_timer_arm(&obd->obd_recovery_timer,
1640                       cfs_time_shift(obd->obd_recovery_timeout));
1641         obd->obd_recovery_start = cfs_time_current_sec();
1642         spin_unlock(&obd->obd_dev_lock);
1643
1644         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1645                       "or until %d client%s reconnect%s\n",
1646                       obd->obd_name,
1647                       obd->obd_recovery_timeout / 60,
1648                       obd->obd_recovery_timeout % 60,
1649                       obd->obd_max_recoverable_clients,
1650                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1651                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1652 }
1653
1654 /**
1655  * extend recovery window.
1656  *
1657  * if @extend is true, extend recovery window to have @drt remaining at least;
1658  * otherwise, make sure the recovery timeout value is not less than @drt.
1659  */
1660 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1661 {
1662         cfs_time_t now;
1663         cfs_time_t end;
1664         cfs_duration_t left;
1665         int to;
1666
1667         spin_lock(&obd->obd_dev_lock);
1668         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1669             obd->obd_force_abort_recovery) {
1670                 spin_unlock(&obd->obd_dev_lock);
1671                 return;
1672         }
1673         LASSERT(obd->obd_recovery_start != 0);
1674
1675         now  = cfs_time_current_sec();
1676         to   = obd->obd_recovery_timeout;
1677         end  = obd->obd_recovery_start + to;
1678         left = cfs_time_sub(end, now);
1679
1680         if (extend && (drt > left)) {
1681                 to += drt - left;
1682         } else if (!extend && (drt > to)) {
1683                 to = drt;
1684         }
1685
1686         if (to > obd->obd_recovery_time_hard)
1687                 to = obd->obd_recovery_time_hard;
1688         if (obd->obd_recovery_timeout < to) {
1689                 obd->obd_recovery_timeout = to;
1690                 end = obd->obd_recovery_start + to;
1691                 cfs_timer_arm(&obd->obd_recovery_timer,
1692                                 cfs_time_shift(end - now));
1693         }
1694         spin_unlock(&obd->obd_dev_lock);
1695
1696         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1697                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1698 }
1699
1700 /* Reset the timer with each new client connection */
1701 /*
1702  * This timer is actually reconnect_timer, which is for making sure
1703  * the total recovery window is at least as big as my reconnect
1704  * attempt timing. So the initial recovery time_out will be set to
1705  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1706  * from client is bigger than this, then the recovery time_out will
1707  * be extended to make sure the client could be reconnected, in the
1708  * process, the timeout from the new client should be ignored.
1709  */
1710
1711 static void
1712 check_and_start_recovery_timer(struct obd_device *obd,
1713                                struct ptlrpc_request *req,
1714                                int new_client)
1715 {
1716         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1717         struct obd_device_target *obt = &obd->u.obt;
1718
1719         if (!new_client && service_time)
1720                 /* Teach server about old server's estimates, as first guess
1721                  * at how long new requests will take. */
1722                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1723                             service_time);
1724
1725         target_start_recovery_timer(obd);
1726
1727         /* Convert the service time to RPC timeout,
1728          * and reuse service_time to limit stack usage. */
1729         service_time = at_est2timeout(service_time);
1730
1731         /* We expect other clients to timeout within service_time, then try
1732          * to reconnect, then try the failover server.  The max delay between
1733          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1734         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1735
1736         LASSERT(obt->obt_magic == OBT_MAGIC);
1737         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1738         if (service_time > obd->obd_recovery_timeout && !new_client)
1739                 extend_recovery_timer(obd, service_time, false);
1740 }
1741
1742 /** Health checking routines */
1743 static inline int exp_connect_healthy(struct obd_export *exp)
1744 {
1745         return (exp->exp_in_recovery);
1746 }
1747
1748 /** if export done req_replay or has replay in queue */
1749 static inline int exp_req_replay_healthy(struct obd_export *exp)
1750 {
1751         return (!exp->exp_req_replay_needed ||
1752                 atomic_read(&exp->exp_replay_count) > 0);
1753 }
1754 /** if export done lock_replay or has replay in queue */
1755 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1756 {
1757         return (!exp->exp_lock_replay_needed ||
1758                 atomic_read(&exp->exp_replay_count) > 0);
1759 }
1760
1761 static inline int exp_vbr_healthy(struct obd_export *exp)
1762 {
1763         return (!exp->exp_vbr_failed);
1764 }
1765
1766 static inline int exp_finished(struct obd_export *exp)
1767 {
1768         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1769 }
1770
1771 static int check_for_next_transno(struct lu_target *lut)
1772 {
1773         struct ptlrpc_request *req = NULL;
1774         struct obd_device *obd = lut->lut_obd;
1775         int wake_up = 0, connected, completed, queue_len;
1776         __u64 req_transno = 0;
1777         __u64 update_transno = 0;
1778         __u64 next_transno = 0;
1779         ENTRY;
1780
1781         spin_lock(&obd->obd_recovery_task_lock);
1782         if (!list_empty(&obd->obd_req_replay_queue)) {
1783                 req = list_entry(obd->obd_req_replay_queue.next,
1784                                      struct ptlrpc_request, rq_list);
1785                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1786         }
1787
1788         if (lut->lut_tdtd != NULL) {
1789                 struct target_distribute_txn_data *tdtd;
1790
1791                 tdtd = lut->lut_tdtd;
1792                 update_transno = distribute_txn_get_next_transno(lut->lut_tdtd);
1793         }
1794
1795         connected = atomic_read(&obd->obd_connected_clients);
1796         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1797         queue_len = obd->obd_requests_queued_for_recovery;
1798         next_transno = obd->obd_next_recovery_transno;
1799
1800         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1801                "req_transno: "LPU64", next_transno: "LPU64"\n",
1802                obd->obd_max_recoverable_clients, connected, completed,
1803                queue_len, req_transno, next_transno);
1804
1805         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1806                 CDEBUG(D_HA, "waking for aborted recovery\n");
1807                 wake_up = 1;
1808         } else if (obd->obd_recovery_expired) {
1809                 CDEBUG(D_HA, "waking for expired recovery\n");
1810                 wake_up = 1;
1811         } else if (req_transno == next_transno ||
1812                    (update_transno != 0 && update_transno <= next_transno)) {
1813                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1814                 wake_up = 1;
1815         } else if (queue_len > 0 &&
1816                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1817                 int d_lvl = D_HA;
1818                 /** handle gaps occured due to lost reply or VBR */
1819                 LASSERTF(req_transno >= next_transno,
1820                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1821                          req_transno, next_transno);
1822                 if (req_transno > obd->obd_last_committed &&
1823                     !obd->obd_version_recov)
1824                         d_lvl = D_ERROR;
1825                 CDEBUG(d_lvl,
1826                        "%s: waking for gap in transno, VBR is %s (skip: "
1827                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1828                        ", next_update "LPD64" last_committed: "LPD64")\n",
1829                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1830                        next_transno, queue_len, completed, connected,
1831                        req_transno, update_transno, obd->obd_last_committed);
1832                 obd->obd_next_recovery_transno = req_transno;
1833                 wake_up = 1;
1834         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1835                 CDEBUG(D_HA, "waking for completed recovery\n");
1836                 wake_up = 1;
1837         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1838                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1839                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1840                 obd->obd_next_recovery_transno = req_transno;
1841                 wake_up = 1;
1842         }
1843         spin_unlock(&obd->obd_recovery_task_lock);
1844         return wake_up;
1845 }
1846
1847 static int check_for_next_lock(struct lu_target *lut)
1848 {
1849         struct obd_device *obd = lut->lut_obd;
1850         int wake_up = 0;
1851
1852         spin_lock(&obd->obd_recovery_task_lock);
1853         if (!list_empty(&obd->obd_lock_replay_queue)) {
1854                 CDEBUG(D_HA, "waking for next lock\n");
1855                 wake_up = 1;
1856         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1857                 CDEBUG(D_HA, "waking for completed lock replay\n");
1858                 wake_up = 1;
1859         } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1860                 CDEBUG(D_HA, "waking for aborted recovery\n");
1861                 wake_up = 1;
1862         } else if (obd->obd_recovery_expired) {
1863                 CDEBUG(D_HA, "waking for expired recovery\n");
1864                 wake_up = 1;
1865         }
1866         spin_unlock(&obd->obd_recovery_task_lock);
1867
1868         return wake_up;
1869 }
1870
1871 /**
1872  * wait for recovery events,
1873  * check its status with help of check_routine
1874  * evict dead clients via health_check
1875  */
1876 static int target_recovery_overseer(struct lu_target *lut,
1877                                     int (*check_routine)(struct lu_target *),
1878                                     int (*health_check)(struct obd_export *))
1879 {
1880         struct obd_device       *obd = lut->lut_obd;
1881 repeat:
1882         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1883               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1884                 CWARN("recovery is aborted by hard timeout\n");
1885                 obd->obd_abort_recovery = 1;
1886         }
1887
1888         while (wait_event_timeout(obd->obd_next_transno_waitq,
1889                                   check_routine(lut),
1890                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1891                 /* wait indefinitely for event, but don't trigger watchdog */;
1892
1893         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1894                 CWARN("recovery is aborted, evict exports in recovery\n");
1895                 /** evict exports which didn't finish recovery yet */
1896                 class_disconnect_stale_exports(obd, exp_finished);
1897                 return 1;
1898         } else if (obd->obd_recovery_expired) {
1899                 obd->obd_recovery_expired = 0;
1900                 /** If some clients died being recovered, evict them */
1901                 LCONSOLE_WARN("%s: recovery is timed out, "
1902                               "evict stale exports\n", obd->obd_name);
1903                 /** evict cexports with no replay in queue, they are stalled */
1904                 class_disconnect_stale_exports(obd, health_check);
1905                 /** continue with VBR */
1906                 spin_lock(&obd->obd_dev_lock);
1907                 obd->obd_version_recov = 1;
1908                 spin_unlock(&obd->obd_dev_lock);
1909                 /**
1910                  * reset timer, recovery will proceed with versions now,
1911                  * timeout is set just to handle reconnection delays
1912                  */
1913                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1914                 /** Wait for recovery events again, after evicting bad clients */
1915                 goto repeat;
1916         }
1917         return 0;
1918 }
1919
1920 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1921 {
1922         struct obd_device       *obd = lut->lut_obd;
1923         struct ptlrpc_request *req = NULL;
1924
1925         CDEBUG(D_HA, "Waiting for lock\n");
1926         if (target_recovery_overseer(lut, check_for_next_lock,
1927                                      exp_lock_replay_healthy))
1928                 abort_lock_replay_queue(obd);
1929
1930         spin_lock(&obd->obd_recovery_task_lock);
1931         if (!list_empty(&obd->obd_lock_replay_queue)) {
1932                 req = list_entry(obd->obd_lock_replay_queue.next,
1933                                      struct ptlrpc_request, rq_list);
1934                 list_del_init(&req->rq_list);
1935                 spin_unlock(&obd->obd_recovery_task_lock);
1936         } else {
1937                 spin_unlock(&obd->obd_recovery_task_lock);
1938                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1939                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1940                 /** evict exports failed VBR */
1941                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1942         }
1943         return req;
1944 }
1945
1946 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1947 {
1948         struct ptlrpc_request *req = NULL;
1949
1950         spin_lock(&obd->obd_recovery_task_lock);
1951         if (!list_empty(&obd->obd_final_req_queue)) {
1952                 req = list_entry(obd->obd_final_req_queue.next,
1953                                      struct ptlrpc_request, rq_list);
1954                 list_del_init(&req->rq_list);
1955                 spin_unlock(&obd->obd_recovery_task_lock);
1956                 if (req->rq_export->exp_in_recovery) {
1957                         spin_lock(&req->rq_export->exp_lock);
1958                         req->rq_export->exp_in_recovery = 0;
1959                         spin_unlock(&req->rq_export->exp_lock);
1960                 }
1961         } else {
1962                 spin_unlock(&obd->obd_recovery_task_lock);
1963         }
1964         return req;
1965 }
1966
1967 static void handle_recovery_req(struct ptlrpc_thread *thread,
1968                                 struct ptlrpc_request *req,
1969                                 svc_handler_t handler)
1970 {
1971         ENTRY;
1972
1973         /**
1974          * export can be evicted during recovery, no need to handle replays for
1975          * it after that, discard such request silently
1976          */
1977         if (req->rq_export->exp_disconnected)
1978                 RETURN_EXIT;
1979
1980         req->rq_session.lc_thread = thread;
1981         req->rq_svc_thread = thread;
1982         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1983
1984         /* thread context */
1985         lu_context_enter(&thread->t_env->le_ctx);
1986         (void)handler(req);
1987         lu_context_exit(&thread->t_env->le_ctx);
1988
1989         /* don't reset timer for final stage */
1990         if (!exp_finished(req->rq_export)) {
1991                 int to = obd_timeout;
1992
1993                 /**
1994                  * Add request timeout to the recovery time so next request from
1995                  * this client may come in recovery time
1996                  */
1997                 if (!AT_OFF) {
1998                         struct ptlrpc_service_part *svcpt;
1999
2000                         svcpt = req->rq_rqbd->rqbd_svcpt;
2001                         /* If the server sent early reply for this request,
2002                          * the client will recalculate the timeout according to
2003                          * current server estimate service time, so we will
2004                          * use the maxium timeout here for waiting the client
2005                          * sending the next req */
2006                         to = max((int)at_est2timeout(
2007                                  at_get(&svcpt->scp_at_estimate)),
2008                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2009                         /* Add 2 net_latency, one for balance rq_deadline
2010                          * (see ptl_send_rpc), one for resend the req to server,
2011                          * Note: client will pack net_latency in replay req
2012                          * (see ptlrpc_replay_req) */
2013                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2014                 }
2015                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2016         }
2017         EXIT;
2018 }
2019
2020 /** Checking routines for recovery */
2021 static int check_for_recovery_ready(struct lu_target *lut)
2022 {
2023         struct obd_device *obd = lut->lut_obd;
2024         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2025
2026         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2027                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2028                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2029                obd->obd_recovery_expired);
2030
2031         if (obd->obd_force_abort_recovery)
2032                 return 1;
2033
2034         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2035                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2036                 if (clnts + obd->obd_stale_clients <
2037                     obd->obd_max_recoverable_clients)
2038                         return 0;
2039         }
2040
2041         if (lut->lut_tdtd != NULL) {
2042                 if (!lut->lut_tdtd->tdtd_replay_ready) {
2043                         /* Let's extend recovery timer, in case the recovery
2044                          * timer expired, and some clients got evicted */
2045                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2046                                               true);
2047                         return 0;
2048                 } else {
2049                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2050                 }
2051         }
2052
2053         return 1;
2054 }
2055
2056 enum {
2057         REQUEST_RECOVERY = 1,
2058         UPDATE_RECOVERY = 2,
2059 };
2060
2061 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2062 {
2063         __u64 transno = 0;
2064
2065         if (!list_empty(&obd->obd_req_replay_queue)) {
2066                 struct ptlrpc_request *req;
2067
2068                 req = list_entry(obd->obd_req_replay_queue.next,
2069                                  struct ptlrpc_request, rq_list);
2070                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2071         }
2072
2073         return transno;
2074 }
2075 __u64 get_next_transno(struct lu_target *lut, int *type)
2076 {
2077         struct obd_device *obd = lut->lut_obd;
2078         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2079         __u64 transno = 0;
2080         __u64 update_transno;
2081         ENTRY;
2082
2083         transno = get_next_replay_req_transno(obd);
2084         if (type != NULL)
2085                 *type = REQUEST_RECOVERY;
2086
2087         if (tdtd == NULL)
2088                 RETURN(transno);
2089
2090         update_transno = distribute_txn_get_next_transno(tdtd);
2091         if (transno == 0 || (transno >= update_transno &&
2092                              update_transno != 0)) {
2093                 transno = update_transno;
2094                 if (type != NULL)
2095                         *type = UPDATE_RECOVERY;
2096         }
2097
2098         RETURN(transno);
2099 }
2100
2101 /**
2102  * drop duplicate replay request
2103  *
2104  * Because the operation has been replayed by update recovery, the request
2105  * with the same transno will be dropped and also notify the client to send
2106  * next replay request.
2107  *
2108  * \param[in] env       execution environment
2109  * \param[in] obd       failover obd device
2110  * \param[in] req       request to be dropped
2111  */
2112 static void drop_duplicate_replay_req(struct lu_env *env,
2113                                       struct obd_device *obd,
2114                                       struct ptlrpc_request *req)
2115 {
2116         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2117                   " update records are found.\n",
2118                   lustre_msg_get_transno(req->rq_reqmsg),
2119                   libcfs_nid2str(req->rq_peer.nid));
2120
2121         /* Right now, only for MDS reint operation update replay and
2122          * normal request replay can have the same transno */
2123         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2124                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2125                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2126                 if (likely(req->rq_export))
2127                         target_committed_to_req(req);
2128                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2129                 target_send_reply(req, req->rq_status, 0);
2130         } else {
2131                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2132                 libcfs_nid2str(req->rq_peer.nid));
2133         }
2134         target_exp_dequeue_req_replay(req);
2135         target_request_copy_put(req);
2136         obd->obd_replayed_requests++;
2137 }
2138
2139 /**
2140  * Update last_rcvd of the update
2141  *
2142  * Because update recovery might update the last_rcvd by updates, i.e.
2143  * it will not update the last_rcvd information in memory, so we need
2144  * refresh these information in memory after update recovery.
2145  *
2146  * \param[in] obd       obd_device under recoverying.
2147  * \param[in] dtrq      the update replay requests being replayed.
2148  */
2149 static void target_update_lcd(struct lu_env *env, struct lu_target *lut,
2150                               struct distribute_txn_replay_req *dtrq)
2151 {
2152         struct obd_device       *obd = lut->lut_obd;
2153         struct obd_export       *export;
2154         struct tg_export_data   *ted;
2155         struct distribute_txn_replay_req_sub    *dtrqs;
2156         struct seq_server_site *site;
2157         struct update_records   *ur;
2158         const struct lu_fid     *fid;
2159         struct update_ops       *ops;
2160         struct update_params    *params;
2161         struct update_op        *op;
2162         __u32                   mdt_index;
2163         unsigned int            i;
2164         struct lsd_client_data  *lcd = NULL;
2165
2166         /* if Updates has been executed(committed) on the recovery target,
2167          * i.e. the updates is not being executed on the target, so we do
2168          * not need update it in memory */
2169         site = lu_site2seq(obd->obd_lu_dev->ld_site);
2170         mdt_index = site->ss_node_id;
2171         dtrqs = dtrq_sub_lookup(dtrq, mdt_index);
2172         if (dtrqs != NULL)
2173                 return;
2174
2175         if (dtrq->dtrq_lur == NULL)
2176                 return;
2177
2178         /* Find the update last_rcvd record */
2179         fid = lu_object_fid(&lut->lut_last_rcvd->do_lu);
2180         ur = &dtrq->dtrq_lur->lur_update_rec;
2181         ops = &ur->ur_ops;
2182         params = update_records_get_params(ur);
2183         for (i = 0, op = &ops->uops_op[0]; i < ur->ur_update_count;
2184              i++, op = update_op_next_op(op)) {
2185                 __u64 pos;
2186                 __u16 size;
2187                 void *buf;
2188
2189                 if (!lu_fid_eq(&op->uop_fid, fid))
2190                         continue;
2191
2192                 if (op->uop_type != OUT_WRITE)
2193                         continue;
2194
2195                 buf = update_params_get_param_buf(params, op->uop_params_off[1],
2196                                                   ur->ur_param_count, NULL);
2197                 if (buf == NULL)
2198                         continue;
2199
2200                 pos = le64_to_cpu(*(__u64 *)buf);
2201                 if (pos == 0)
2202                         continue;
2203
2204                 buf = update_params_get_param_buf(params, op->uop_params_off[0],
2205                                                   ur->ur_param_count, &size);
2206                 if (buf == NULL)
2207                         continue;
2208
2209                 if (size != sizeof(*lcd))
2210                         continue;
2211                 lcd = buf;
2212         }
2213
2214         if (lcd == NULL || lcd->lcd_uuid[0] == '\0')
2215                 return;
2216
2217         /* locate the export then update the exp_target_data if needed */
2218         export = cfs_hash_lookup(obd->obd_uuid_hash, lcd->lcd_uuid);
2219         if (export == NULL)
2220                 return;
2221
2222         ted = &export->exp_target_data;
2223         if (lcd->lcd_last_xid > ted->ted_lcd->lcd_last_xid) {
2224                 CDEBUG(D_HA, "%s update xid from "LPU64" to "LPU64"\n",
2225                        lut->lut_obd->obd_name, ted->ted_lcd->lcd_last_xid,
2226                        lcd->lcd_last_xid);
2227                 ted->ted_lcd->lcd_last_xid = lcd->lcd_last_xid;
2228                 ted->ted_lcd->lcd_last_result = lcd->lcd_last_result;
2229         }
2230         class_export_put(export);
2231 }
2232
2233 static void replay_request_or_update(struct lu_env *env,
2234                                      struct lu_target *lut,
2235                                      struct target_recovery_data *trd,
2236                                      struct ptlrpc_thread *thread)
2237 {
2238         struct obd_device *obd = lut->lut_obd;
2239         struct ptlrpc_request *req = NULL;
2240         int                     type;
2241         __u64                   transno;
2242         ENTRY;
2243
2244         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2245                obd->obd_next_recovery_transno);
2246
2247         /* Replay all of request and update by transno */
2248         do {
2249                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2250
2251                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2252
2253                 /** It is needed to extend recovery window above
2254                  *  recovery_time_soft. Extending is possible only in the
2255                  *  end of recovery window (see more details in
2256                  *  handle_recovery_req()).
2257                  */
2258                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2259
2260                 if (target_recovery_overseer(lut, check_for_next_transno,
2261                                              exp_req_replay_healthy)) {
2262                         abort_req_replay_queue(obd);
2263                         abort_lock_replay_queue(obd);
2264                 }
2265
2266                 spin_lock(&obd->obd_recovery_task_lock);
2267                 transno = get_next_transno(lut, &type);
2268                 if (type == REQUEST_RECOVERY && tdtd != NULL &&
2269                     transno == tdtd->tdtd_last_update_transno) {
2270                         /* Drop replay request from client side, if the
2271                          * replay has been executed by update with the
2272                          * same transno */
2273                         req = list_entry(obd->obd_req_replay_queue.next,
2274                                         struct ptlrpc_request, rq_list);
2275                         list_del_init(&req->rq_list);
2276                         obd->obd_requests_queued_for_recovery--;
2277                         spin_unlock(&obd->obd_recovery_task_lock);
2278                         drop_duplicate_replay_req(env, obd, req);
2279                 } else if (type == REQUEST_RECOVERY && transno != 0) {
2280                         req = list_entry(obd->obd_req_replay_queue.next,
2281                                              struct ptlrpc_request, rq_list);
2282                         list_del_init(&req->rq_list);
2283                         obd->obd_requests_queued_for_recovery--;
2284                         spin_unlock(&obd->obd_recovery_task_lock);
2285                         LASSERT(trd->trd_processing_task == current_pid());
2286                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2287                                   lustre_msg_get_transno(req->rq_reqmsg),
2288                                   libcfs_nid2str(req->rq_peer.nid));
2289
2290                         handle_recovery_req(thread, req,
2291                                             trd->trd_recovery_handler);
2292                         /**
2293                          * bz18031: increase next_recovery_transno before
2294                          * target_request_copy_put() will drop exp_rpc reference
2295                          */
2296                         spin_lock(&obd->obd_recovery_task_lock);
2297                         obd->obd_next_recovery_transno++;
2298                         spin_unlock(&obd->obd_recovery_task_lock);
2299                         target_exp_dequeue_req_replay(req);
2300                         target_request_copy_put(req);
2301                         obd->obd_replayed_requests++;
2302                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2303                         struct distribute_txn_replay_req *dtrq;
2304
2305                         spin_unlock(&obd->obd_recovery_task_lock);
2306
2307                         LASSERT(tdtd != NULL);
2308                         dtrq = distribute_txn_get_next_req(tdtd);
2309                         lu_context_enter(&thread->t_env->le_ctx);
2310                         tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2311                         lu_context_exit(&thread->t_env->le_ctx);
2312                         extend_recovery_timer(obd, obd_timeout, true);
2313                         LASSERT(tdtd->tdtd_last_update_transno <= transno);
2314                         tdtd->tdtd_last_update_transno = transno;
2315                         spin_lock(&obd->obd_recovery_task_lock);
2316                         if (transno > obd->obd_next_recovery_transno)
2317                                 obd->obd_next_recovery_transno = transno;
2318                         spin_unlock(&obd->obd_recovery_task_lock);
2319                         target_update_lcd(env, lut, dtrq);
2320                         dtrq_destroy(dtrq);
2321                 } else {
2322                         spin_unlock(&obd->obd_recovery_task_lock);
2323                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2324                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2325                         /** evict exports failed VBR */
2326                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2327                         break;
2328                 }
2329         } while (1);
2330 }
2331
2332 static int target_recovery_thread(void *arg)
2333 {
2334         struct lu_target *lut = arg;
2335         struct obd_device *obd = lut->lut_obd;
2336         struct ptlrpc_request *req;
2337         struct target_recovery_data *trd = &obd->obd_recovery_data;
2338         unsigned long delta;
2339         struct lu_env *env;
2340         struct ptlrpc_thread *thread = NULL;
2341         int rc = 0;
2342         ENTRY;
2343
2344         unshare_fs_struct();
2345         OBD_ALLOC_PTR(thread);
2346         if (thread == NULL)
2347                 RETURN(-ENOMEM);
2348
2349         OBD_ALLOC_PTR(env);
2350         if (env == NULL) {
2351                 OBD_FREE_PTR(thread);
2352                 RETURN(-ENOMEM);
2353         }
2354
2355         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2356         if (rc) {
2357                 OBD_FREE_PTR(thread);
2358                 OBD_FREE_PTR(env);
2359                 RETURN(rc);
2360         }
2361
2362         thread->t_env = env;
2363         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2364         env->le_ctx.lc_thread = thread;
2365         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2366         thread->t_watchdog = NULL;
2367
2368         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2369                current_pid());
2370         trd->trd_processing_task = current_pid();
2371
2372         spin_lock(&obd->obd_dev_lock);
2373         obd->obd_recovering = 1;
2374         spin_unlock(&obd->obd_dev_lock);
2375         complete(&trd->trd_starting);
2376
2377         /* first of all, we have to know the first transno to replay */
2378         if (target_recovery_overseer(lut, check_for_recovery_ready,
2379                                      exp_connect_healthy)) {
2380                 abort_req_replay_queue(obd);
2381                 abort_lock_replay_queue(obd);
2382                 if (lut->lut_tdtd != NULL)
2383                         dtrq_list_destroy(lut->lut_tdtd);
2384         }
2385
2386         /* next stage: replay requests or update */
2387         delta = jiffies;
2388         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2389                atomic_read(&obd->obd_req_replay_clients),
2390                obd->obd_next_recovery_transno);
2391         replay_request_or_update(env, lut, trd, thread);
2392
2393         /**
2394          * The second stage: replay locks
2395          */
2396         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2397                atomic_read(&obd->obd_lock_replay_clients));
2398         while ((req = target_next_replay_lock(lut))) {
2399                 LASSERT(trd->trd_processing_task == current_pid());
2400                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2401                           libcfs_nid2str(req->rq_peer.nid));
2402                 handle_recovery_req(thread, req,
2403                                     trd->trd_recovery_handler);
2404                 target_request_copy_put(req);
2405                 obd->obd_replayed_locks++;
2406         }
2407
2408         /**
2409          * The third stage: reply on final pings, at this moment all clients
2410          * must have request in final queue
2411          */
2412         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2413         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2414         /** Update server last boot epoch */
2415         tgt_boot_epoch_update(lut);
2416         /* We drop recoverying flag to forward all new requests
2417          * to regular mds_handle() since now */
2418         spin_lock(&obd->obd_dev_lock);
2419         obd->obd_recovering = obd->obd_abort_recovery = 0;
2420         spin_unlock(&obd->obd_dev_lock);
2421         spin_lock(&obd->obd_recovery_task_lock);
2422         target_cancel_recovery_timer(obd);
2423         spin_unlock(&obd->obd_recovery_task_lock);
2424         while ((req = target_next_final_ping(obd))) {
2425                 LASSERT(trd->trd_processing_task == current_pid());
2426                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2427                           libcfs_nid2str(req->rq_peer.nid));
2428                 handle_recovery_req(thread, req,
2429                                     trd->trd_recovery_handler);
2430                 /* Because the waiting client can not send ping to server,
2431                  * so we need refresh the last_request_time, to avoid the
2432                  * export is being evicted */
2433                 ptlrpc_update_export_timer(req->rq_export, 0);
2434                 target_request_copy_put(req);
2435         }
2436
2437         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2438         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2439               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2440         if (delta > OBD_RECOVERY_TIME_SOFT) {
2441                 CWARN("too long recovery - read logs\n");
2442                 libcfs_debug_dumplog();
2443         }
2444
2445         target_finish_recovery(lut);
2446
2447         lu_context_fini(&env->le_ctx);
2448         trd->trd_processing_task = 0;
2449         complete(&trd->trd_finishing);
2450
2451         tgt_io_thread_done(thread);
2452         OBD_FREE_PTR(thread);
2453         OBD_FREE_PTR(env);
2454         RETURN(rc);
2455 }
2456
2457 static int target_start_recovery_thread(struct lu_target *lut,
2458                                         svc_handler_t handler)
2459 {
2460         struct obd_device *obd = lut->lut_obd;
2461         int rc = 0;
2462         struct target_recovery_data *trd = &obd->obd_recovery_data;
2463
2464         memset(trd, 0, sizeof(*trd));
2465         init_completion(&trd->trd_starting);
2466         init_completion(&trd->trd_finishing);
2467         trd->trd_recovery_handler = handler;
2468
2469         if (!IS_ERR(kthread_run(target_recovery_thread,
2470                                 lut, "tgt_recov"))) {
2471                 wait_for_completion(&trd->trd_starting);
2472                 LASSERT(obd->obd_recovering != 0);
2473         } else {
2474                 rc = -ECHILD;
2475         }
2476
2477         return rc;
2478 }
2479
2480 void target_stop_recovery_thread(struct obd_device *obd)
2481 {
2482         if (obd->obd_recovery_data.trd_processing_task > 0) {
2483                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2484                 /** recovery can be done but postrecovery is not yet */
2485                 spin_lock(&obd->obd_dev_lock);
2486                 if (obd->obd_recovering) {
2487                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2488                         obd->obd_abort_recovery = 1;
2489                         wake_up(&obd->obd_next_transno_waitq);
2490                 }
2491                 spin_unlock(&obd->obd_dev_lock);
2492                 wait_for_completion(&trd->trd_finishing);
2493         }
2494 }
2495 EXPORT_SYMBOL(target_stop_recovery_thread);
2496
2497 void target_recovery_fini(struct obd_device *obd)
2498 {
2499         class_disconnect_exports(obd);
2500         target_stop_recovery_thread(obd);
2501         target_cleanup_recovery(obd);
2502 }
2503 EXPORT_SYMBOL(target_recovery_fini);
2504
2505 static void target_recovery_expired(unsigned long castmeharder)
2506 {
2507         struct obd_device *obd = (struct obd_device *)castmeharder;
2508         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2509                " after %lds (%d clients connected)\n",
2510                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2511                cfs_time_current_sec()- obd->obd_recovery_start,
2512                atomic_read(&obd->obd_connected_clients));
2513
2514         obd->obd_recovery_expired = 1;
2515         wake_up(&obd->obd_next_transno_waitq);
2516 }
2517
2518 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2519 {
2520         struct obd_device *obd = lut->lut_obd;
2521
2522         if (obd->obd_max_recoverable_clients == 0) {
2523                 /** Update server last boot epoch */
2524                 tgt_boot_epoch_update(lut);
2525                 return;
2526         }
2527
2528         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2529                "last_transno "LPU64"\n", obd->obd_name,
2530                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2531         LASSERT(obd->obd_stopping == 0);
2532         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2533         obd->obd_recovery_start = 0;
2534         obd->obd_recovery_end = 0;
2535
2536         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2537         target_start_recovery_thread(lut, handler);
2538 }
2539 EXPORT_SYMBOL(target_recovery_init);
2540
2541 static int target_process_req_flags(struct obd_device *obd,
2542                                     struct ptlrpc_request *req)
2543 {
2544         struct obd_export *exp = req->rq_export;
2545         LASSERT(exp != NULL);
2546         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2547                 /* client declares he's ready to replay locks */
2548                 spin_lock(&exp->exp_lock);
2549                 if (exp->exp_req_replay_needed) {
2550                         exp->exp_req_replay_needed = 0;
2551                         spin_unlock(&exp->exp_lock);
2552
2553                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2554                         atomic_dec(&obd->obd_req_replay_clients);
2555                 } else {
2556                         spin_unlock(&exp->exp_lock);
2557                 }
2558         }
2559         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2560                 /* client declares he's ready to complete recovery
2561                  * so, we put the request on th final queue */
2562                 spin_lock(&exp->exp_lock);
2563                 if (exp->exp_lock_replay_needed) {
2564                         exp->exp_lock_replay_needed = 0;
2565                         spin_unlock(&exp->exp_lock);
2566
2567                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2568                         atomic_dec(&obd->obd_lock_replay_clients);
2569                 } else {
2570                         spin_unlock(&exp->exp_lock);
2571                 }
2572         }
2573         return 0;
2574 }
2575
2576 int target_queue_recovery_request(struct ptlrpc_request *req,
2577                                   struct obd_device *obd)
2578 {
2579         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2580         struct ptlrpc_request *reqiter;
2581         int inserted = 0;
2582         ENTRY;
2583
2584         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2585                 /* Processing the queue right now, don't re-add. */
2586                 RETURN(1);
2587         }
2588
2589         target_process_req_flags(obd, req);
2590
2591         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2592                 /* client declares he's ready to complete recovery
2593                  * so, we put the request on th final queue */
2594                 target_request_copy_get(req);
2595                 DEBUG_REQ(D_HA, req, "queue final req");
2596                 wake_up(&obd->obd_next_transno_waitq);
2597                 spin_lock(&obd->obd_recovery_task_lock);
2598                 if (obd->obd_recovering) {
2599                         list_add_tail(&req->rq_list,
2600                                           &obd->obd_final_req_queue);
2601                 } else {
2602                         spin_unlock(&obd->obd_recovery_task_lock);
2603                         target_request_copy_put(req);
2604                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2605                 }
2606                 spin_unlock(&obd->obd_recovery_task_lock);
2607                 RETURN(0);
2608         }
2609         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2610                 /* client declares he's ready to replay locks */
2611                 target_request_copy_get(req);
2612                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2613                 wake_up(&obd->obd_next_transno_waitq);
2614                 spin_lock(&obd->obd_recovery_task_lock);
2615                 LASSERT(obd->obd_recovering);
2616                 /* usually due to recovery abort */
2617                 if (!req->rq_export->exp_in_recovery) {
2618                         spin_unlock(&obd->obd_recovery_task_lock);
2619                         target_request_copy_put(req);
2620                         RETURN(-ENOTCONN);
2621                 }
2622                 LASSERT(req->rq_export->exp_lock_replay_needed);
2623                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2624                 spin_unlock(&obd->obd_recovery_task_lock);
2625                 RETURN(0);
2626         }
2627
2628         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2629          * (i.e. buflens etc are in my own byte order), but type-dependent
2630          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2631
2632         if (!transno) {
2633                 INIT_LIST_HEAD(&req->rq_list);
2634                 DEBUG_REQ(D_HA, req, "not queueing");
2635                 RETURN(1);
2636         }
2637
2638         /* If we're processing the queue, we want don't want to queue this
2639          * message.
2640          *
2641          * Also, if this request has a transno less than the one we're waiting
2642          * for, we should process it now.  It could (and currently always will)
2643          * be an open request for a descriptor that was opened some time ago.
2644          *
2645          * Also, a resent, replayed request that has already been
2646          * handled will pass through here and be processed immediately.
2647          */
2648         CDEBUG(D_HA, "Next recovery transno: "LPU64
2649                ", current: "LPU64", replaying\n",
2650                obd->obd_next_recovery_transno, transno);
2651         spin_lock(&obd->obd_recovery_task_lock);
2652         if (transno < obd->obd_next_recovery_transno) {
2653                 /* Processing the queue right now, don't re-add. */
2654                 LASSERT(list_empty(&req->rq_list));
2655                 spin_unlock(&obd->obd_recovery_task_lock);
2656                 RETURN(1);
2657         }
2658         spin_unlock(&obd->obd_recovery_task_lock);
2659
2660         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2661                 RETURN(0);
2662
2663         target_request_copy_get(req);
2664         if (!req->rq_export->exp_in_recovery) {
2665                 target_request_copy_put(req);
2666                 RETURN(-ENOTCONN);
2667         }
2668         LASSERT(req->rq_export->exp_req_replay_needed);
2669
2670         if (target_exp_enqueue_req_replay(req)) {
2671                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2672                 target_request_copy_put(req);
2673                 RETURN(0);
2674         }
2675
2676         /* XXX O(n^2) */
2677         spin_lock(&obd->obd_recovery_task_lock);
2678         LASSERT(obd->obd_recovering);
2679         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2680                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2681                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2682                         inserted = 1;
2683                         goto added;
2684                 }
2685
2686                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2687                              transno)) {
2688                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2689                                   "has been claimed by another client");
2690                         spin_unlock(&obd->obd_recovery_task_lock);
2691                         target_exp_dequeue_req_replay(req);
2692                         target_request_copy_put(req);
2693                         RETURN(0);
2694                 }
2695         }
2696 added:
2697         if (!inserted)
2698                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2699
2700         obd->obd_requests_queued_for_recovery++;
2701         spin_unlock(&obd->obd_recovery_task_lock);
2702         wake_up(&obd->obd_next_transno_waitq);
2703         RETURN(0);
2704 }
2705
2706 int target_handle_ping(struct ptlrpc_request *req)
2707 {
2708         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2709         return req_capsule_server_pack(&req->rq_pill);
2710 }
2711
2712 void target_committed_to_req(struct ptlrpc_request *req)
2713 {
2714         struct obd_export *exp = req->rq_export;
2715
2716         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2717                 lustre_msg_set_last_committed(req->rq_repmsg,
2718                                               exp->exp_last_committed);
2719         else
2720                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2721                           "%d)", exp->exp_obd->obd_no_transno,
2722                           req->rq_repmsg == NULL);
2723
2724         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2725                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2726 }
2727
2728 #endif /* HAVE_SERVER_SUPPORT */
2729
2730 /**
2731  * Packs current SLV and Limit into \a req.
2732  */
2733 int target_pack_pool_reply(struct ptlrpc_request *req)
2734 {
2735         struct obd_device *obd;
2736         ENTRY;
2737
2738         /* Check that we still have all structures alive as this may
2739          * be some late RPC at shutdown time. */
2740         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2741                      !exp_connect_lru_resize(req->rq_export))) {
2742                 lustre_msg_set_slv(req->rq_repmsg, 0);
2743                 lustre_msg_set_limit(req->rq_repmsg, 0);
2744                 RETURN(0);
2745         }
2746
2747         /* OBD is alive here as export is alive, which we checked above. */
2748         obd = req->rq_export->exp_obd;
2749
2750         read_lock(&obd->obd_pool_lock);
2751         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2752         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2753         read_unlock(&obd->obd_pool_lock);
2754
2755         RETURN(0);
2756 }
2757
2758 static int target_send_reply_msg(struct ptlrpc_request *req,
2759                                  int rc, int fail_id)
2760 {
2761         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2762                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2763                 return (-ECOMM);
2764         }
2765
2766         if (unlikely(rc)) {
2767                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2768                 req->rq_status = rc;
2769                 return (ptlrpc_send_error(req, 1));
2770         } else {
2771                 DEBUG_REQ(D_NET, req, "sending reply");
2772         }
2773
2774         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2775 }
2776
2777 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2778 {
2779         struct ptlrpc_service_part *svcpt;
2780         int                        netrc;
2781         struct ptlrpc_reply_state *rs;
2782         struct obd_export         *exp;
2783         ENTRY;
2784
2785         if (req->rq_no_reply) {
2786                 EXIT;
2787                 return;
2788         }
2789
2790         svcpt = req->rq_rqbd->rqbd_svcpt;
2791         rs = req->rq_reply_state;
2792         if (rs == NULL || !rs->rs_difficult) {
2793                 /* no notifiers */
2794                 target_send_reply_msg (req, rc, fail_id);
2795                 EXIT;
2796                 return;
2797         }
2798
2799         /* must be an export if locks saved */
2800         LASSERT(req->rq_export != NULL);
2801         /* req/reply consistent */
2802         LASSERT(rs->rs_svcpt == svcpt);
2803
2804         /* "fresh" reply */
2805         LASSERT(!rs->rs_scheduled);
2806         LASSERT(!rs->rs_scheduled_ever);
2807         LASSERT(!rs->rs_handled);
2808         LASSERT(!rs->rs_on_net);
2809         LASSERT(rs->rs_export == NULL);
2810         LASSERT(list_empty(&rs->rs_obd_list));
2811         LASSERT(list_empty(&rs->rs_exp_list));
2812
2813         exp = class_export_get(req->rq_export);
2814
2815         /* disable reply scheduling while I'm setting up */
2816         rs->rs_scheduled = 1;
2817         rs->rs_on_net    = 1;
2818         rs->rs_xid       = req->rq_xid;
2819         rs->rs_transno   = req->rq_transno;
2820         rs->rs_export    = exp;
2821         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2822
2823         spin_lock(&exp->exp_uncommitted_replies_lock);
2824         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2825                rs->rs_transno, exp->exp_last_committed);
2826         if (rs->rs_transno > exp->exp_last_committed) {
2827                 /* not committed already */
2828                 list_add_tail(&rs->rs_obd_list,
2829                                   &exp->exp_uncommitted_replies);
2830         }
2831         spin_unlock(&exp->exp_uncommitted_replies_lock);
2832
2833         spin_lock(&exp->exp_lock);
2834         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2835         spin_unlock(&exp->exp_lock);
2836
2837         netrc = target_send_reply_msg(req, rc, fail_id);
2838
2839         spin_lock(&svcpt->scp_rep_lock);
2840
2841         atomic_inc(&svcpt->scp_nreps_difficult);
2842
2843         if (netrc != 0) {
2844                 /* error sending: reply is off the net.  Also we need +1
2845                  * reply ref until ptlrpc_handle_rs() is done
2846                  * with the reply state (if the send was successful, there
2847                  * would have been +1 ref for the net, which
2848                  * reply_out_callback leaves alone) */
2849                 rs->rs_on_net = 0;
2850                 ptlrpc_rs_addref(rs);
2851         }
2852
2853         spin_lock(&rs->rs_lock);
2854         if (rs->rs_transno <= exp->exp_last_committed ||
2855             (!rs->rs_on_net && !rs->rs_no_ack) ||
2856             list_empty(&rs->rs_exp_list) ||     /* completed already */
2857             list_empty(&rs->rs_obd_list)) {
2858                 CDEBUG(D_HA, "Schedule reply immediately\n");
2859                 ptlrpc_dispatch_difficult_reply(rs);
2860         } else {
2861                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2862                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2863         }
2864         spin_unlock(&rs->rs_lock);
2865         spin_unlock(&svcpt->scp_rep_lock);
2866         EXIT;
2867 }
2868
2869 ldlm_mode_t lck_compat_array[] = {
2870         [LCK_EX]    = LCK_COMPAT_EX,
2871         [LCK_PW]    = LCK_COMPAT_PW,
2872         [LCK_PR]    = LCK_COMPAT_PR,
2873         [LCK_CW]    = LCK_COMPAT_CW,
2874         [LCK_CR]    = LCK_COMPAT_CR,
2875         [LCK_NL]    = LCK_COMPAT_NL,
2876         [LCK_GROUP] = LCK_COMPAT_GROUP,
2877         [LCK_COS]   = LCK_COMPAT_COS,
2878 };
2879
2880 /**
2881  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2882  * not escape to the user level.
2883  */
2884 int ldlm_error2errno(ldlm_error_t error)
2885 {
2886         int result;
2887
2888         switch (error) {
2889         case ELDLM_OK:
2890         case ELDLM_LOCK_MATCHED:
2891                 result = 0;
2892                 break;
2893         case ELDLM_LOCK_CHANGED:
2894                 result = -ESTALE;
2895                 break;
2896         case ELDLM_LOCK_ABORTED:
2897                 result = -ENAVAIL;
2898                 break;
2899         case ELDLM_LOCK_REPLACED:
2900                 result = -ESRCH;
2901                 break;
2902         case ELDLM_NO_LOCK_DATA:
2903                 result = -ENOENT;
2904                 break;
2905         case ELDLM_NAMESPACE_EXISTS:
2906                 result = -EEXIST;
2907                 break;
2908         case ELDLM_BAD_NAMESPACE:
2909                 result = -EBADF;
2910                 break;
2911         default:
2912                 if (((int)error) < 0)  /* cast to signed type */
2913                         result = error; /* as ldlm_error_t can be unsigned */
2914                 else {
2915                         CERROR("Invalid DLM result code: %d\n", error);
2916                         result = -EPROTO;
2917                 }
2918         }
2919         return result;
2920 }
2921 EXPORT_SYMBOL(ldlm_error2errno);
2922
2923 /**
2924  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2925  */
2926 ldlm_error_t ldlm_errno2error(int err_no)
2927 {
2928         int error;
2929
2930         switch (err_no) {
2931         case 0:
2932                 error = ELDLM_OK;
2933                 break;
2934         case -ESTALE:
2935                 error = ELDLM_LOCK_CHANGED;
2936                 break;
2937         case -ENAVAIL:
2938                 error = ELDLM_LOCK_ABORTED;
2939                 break;
2940         case -ESRCH:
2941                 error = ELDLM_LOCK_REPLACED;
2942                 break;
2943         case -ENOENT:
2944                 error = ELDLM_NO_LOCK_DATA;
2945                 break;
2946         case -EEXIST:
2947                 error = ELDLM_NAMESPACE_EXISTS;
2948                 break;
2949         case -EBADF:
2950                 error = ELDLM_BAD_NAMESPACE;
2951                 break;
2952         default:
2953                 error = err_no;
2954         }
2955         return error;
2956 }
2957
2958 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2959 void ldlm_dump_export_locks(struct obd_export *exp)
2960 {
2961         spin_lock(&exp->exp_locks_list_guard);
2962         if (!list_empty(&exp->exp_locks_list)) {
2963                 struct ldlm_lock *lock;
2964
2965                 CERROR("dumping locks for export %p,"
2966                        "ignore if the unmount doesn't hang\n", exp);
2967                 list_for_each_entry(lock, &exp->exp_locks_list,
2968                                         l_exp_refs_link)
2969                         LDLM_ERROR(lock, "lock:");
2970         }
2971         spin_unlock(&exp->exp_locks_list_guard);
2972 }
2973 #endif
2974
2975 #ifdef HAVE_SERVER_SUPPORT
2976 static int target_bulk_timeout(void *data)
2977 {
2978         ENTRY;
2979         /* We don't fail the connection here, because having the export
2980          * killed makes the (vital) call to commitrw very sad.
2981          */
2982         RETURN(1);
2983 }
2984
2985 static inline const char *bulk2type(struct ptlrpc_request *req)
2986 {
2987         if (req->rq_bulk_read)
2988                 return "READ";
2989         if (req->rq_bulk_write)
2990                 return "WRITE";
2991         return "UNKNOWN";
2992 }
2993
2994 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2995                    struct l_wait_info *lwi)
2996 {
2997         struct ptlrpc_request   *req = desc->bd_req;
2998         time_t                   start = cfs_time_current_sec();
2999         time_t                   deadline;
3000         int                      rc = 0;
3001
3002         ENTRY;
3003
3004         /* If there is eviction in progress, wait for it to finish. */
3005         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3006                 *lwi = LWI_INTR(NULL, NULL);
3007                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3008                                   !atomic_read(&exp->exp_obd->
3009                                                    obd_evict_inprogress),
3010                                   lwi);
3011         }
3012
3013         /* Check if client was evicted or reconnected already. */
3014         if (exp->exp_failed ||
3015             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3016                 rc = -ENOTCONN;
3017         } else {
3018                 if (req->rq_bulk_read)
3019                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3020
3021                 if ((exp->exp_connect_data.ocd_connect_flags &
3022                      OBD_CONNECT_BULK_MBITS) != 0)
3023                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3024                 else /* old version, bulk matchbits is rq_xid */
3025                         req->rq_mbits = req->rq_xid;
3026
3027                 if (rc == 0)
3028                         rc = ptlrpc_start_bulk_transfer(desc);
3029         }
3030
3031         if (rc < 0) {
3032                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3033                           bulk2type(req), rc);
3034                 RETURN(rc);
3035         }
3036
3037         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3038                 ptlrpc_abort_bulk(desc);
3039                 RETURN(0);
3040         }
3041
3042         /* limit actual bulk transfer to bulk_timeout seconds */
3043         deadline = start + bulk_timeout;
3044         if (deadline > req->rq_deadline)
3045                 deadline = req->rq_deadline;
3046
3047         do {
3048                 long timeoutl = deadline - cfs_time_current_sec();
3049                 cfs_duration_t timeout = timeoutl <= 0 ?
3050                                          CFS_TICK : cfs_time_seconds(timeoutl);
3051                 time_t  rq_deadline;
3052
3053                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3054                                             target_bulk_timeout, desc);
3055                 rc = l_wait_event(desc->bd_waitq,
3056                                   !ptlrpc_server_bulk_active(desc) ||
3057                                   exp->exp_failed ||
3058                                   exp->exp_conn_cnt >
3059                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3060                                   lwi);
3061                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3062                 /* Wait again if we changed rq_deadline. */
3063                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3064                 deadline = start + bulk_timeout;
3065                 if (deadline > rq_deadline)
3066                         deadline = rq_deadline;
3067         } while ((rc == -ETIMEDOUT) &&
3068                  (deadline > cfs_time_current_sec()));
3069
3070         if (rc == -ETIMEDOUT) {
3071                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3072                           bulk2type(req), deadline - start,
3073                           cfs_time_current_sec() - deadline);
3074                 ptlrpc_abort_bulk(desc);
3075         } else if (exp->exp_failed) {
3076                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3077                           bulk2type(req));
3078                 rc = -ENOTCONN;
3079                 ptlrpc_abort_bulk(desc);
3080         } else if (exp->exp_conn_cnt >
3081                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3082                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3083                           bulk2type(req));
3084                 /* We don't reply anyway. */
3085                 rc = -ETIMEDOUT;
3086                 ptlrpc_abort_bulk(desc);
3087         } else if (desc->bd_failure) {
3088                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3089                           bulk2type(req));
3090                 /* XXX should this be a different errno? */
3091                 rc = -ETIMEDOUT;
3092         } else {
3093                 if (req->rq_bulk_write)
3094                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3095                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3096                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3097                                   bulk2type(req), desc->bd_nob_transferred,
3098                                   desc->bd_nob);
3099                         /* XXX should this be a different errno? */
3100                         rc = -ETIMEDOUT;
3101                 }
3102         }
3103
3104         RETURN(rc);
3105 }
3106 EXPORT_SYMBOL(target_bulk_io);
3107
3108 #endif /* HAVE_SERVER_SUPPORT */