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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 /**
  34  * This file deals with various client/target related logic including recovery.
  35  *
  36  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  37  * should be moved.
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #include <linux/kthread.h>
  43 #include <libcfs/libcfs.h>
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_net.h>
  48 #include <lustre_sec.h>
  49 #include "ldlm_internal.h"
  50
  51 /* @priority: If non-zero, move the selected connection to the list head.
  52  * @create: If zero, only search in existing connections.
  53  */
  54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  55                            int priority, int create)
  56 {
  57         struct ptlrpc_connection *ptlrpc_conn;
  58         struct obd_import_conn *imp_conn = NULL, *item;
  59         lnet_nid_t nid4refnet = LNET_NID_ANY;
  60         int rc = 0;
  61         ENTRY;
  62
  63         if (!create && !priority) {
  64                 CDEBUG(D_HA, "Nothing to do\n");
  65                 RETURN(-EINVAL);
  66         }
  67
  68         if (imp->imp_connection &&
  69             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  70                 /* nid4refnet is used to restrict network connections */
  71                 nid4refnet = imp->imp_connection->c_self;
  72         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
  73         if (!ptlrpc_conn) {
  74                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  75                 RETURN(-ENOENT);
  76         }
  77
  78         if (create) {
  79                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  80                 if (!imp_conn)
  81                         GOTO(out_put, rc = -ENOMEM);
  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                         if (dlmexp != NULL)
 179                                 class_export_put(dlmexp);
 180                 }
 181
 182                 list_del(&imp_conn->oic_item);
 183                 ptlrpc_connection_put(imp_conn->oic_conn);
 184                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 185                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 186                        imp, imp->imp_obd->obd_name, uuid->uuid);
 187                 rc = 0;
 188                 break;
 189         }
 190 out:
 191         spin_unlock(&imp->imp_lock);
 192         if (rc == -ENOENT)
 193                 CERROR("connection %s not found\n", uuid->uuid);
 194         RETURN(rc);
 195 }
 196 EXPORT_SYMBOL(client_import_del_conn);
 197
 198 /**
 199  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 200  * to find a conn uuid of \a imp which can reach \a peer.
 201  */
 202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 203                             struct obd_uuid *uuid)
 204 {
 205         struct obd_import_conn *conn;
 206         int rc = -ENOENT;
 207         ENTRY;
 208
 209         spin_lock(&imp->imp_lock);
 210         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 211                 /* Check if conn UUID does have this peer NID. */
 212                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 213                         *uuid = conn->oic_uuid;
 214                         rc = 0;
 215                         break;
 216                 }
 217         }
 218         spin_unlock(&imp->imp_lock);
 219         RETURN(rc);
 220 }
 221 EXPORT_SYMBOL(client_import_find_conn);
 222
 223 void client_destroy_import(struct obd_import *imp)
 224 {
 225         /* Drop security policy instance after all RPCs have finished/aborted
 226          * to let all busy contexts be released. */
 227         class_import_get(imp);
 228         class_destroy_import(imp);
 229         sptlrpc_import_sec_put(imp);
 230         class_import_put(imp);
 231 }
 232 EXPORT_SYMBOL(client_destroy_import);
 233
 234 /**
 235  * Check whether or not the OSC is on MDT.
 236  * In the config log,
 237  * osc on MDT
 238  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 239  * osc on client
 240  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 241  *
 242  **/
 243 static int osc_on_mdt(char *obdname)
 244 {
 245         char *ptr;
 246
 247         ptr = strrchr(obdname, '-');
 248         if (ptr == NULL)
 249                 return 0;
 250
 251         if (strncmp(ptr + 1, "MDT", 3) == 0)
 252                 return 1;
 253
 254         return 0;
 255 }
 256
 257 /* Configure an RPC client OBD device.
 258  *
 259  * lcfg parameters:
 260  * 1 - client UUID
 261  * 2 - server UUID
 262  * 3 - inactive-on-startup
 263  * 4 - restrictive net
 264  */
 265 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 266 {
 267         struct client_obd *cli = &obddev->u.cli;
 268         struct obd_import *imp;
 269         struct obd_uuid server_uuid;
 270         int rq_portal, rp_portal, connect_op;
 271         char *name = obddev->obd_type->typ_name;
 272         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 273         char *cli_name = lustre_cfg_buf(lcfg, 0);
 274         struct ptlrpc_connection fake_conn = { .c_self = 0,
 275                                                .c_remote_uuid.uuid[0] = 0 };
 276         int rc;
 277         ENTRY;
 278
 279         /* In a more perfect world, we would hang a ptlrpc_client off of
 280          * obd_type and just use the values from there. */
 281         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 282                 rq_portal = OST_REQUEST_PORTAL;
 283                 rp_portal = OSC_REPLY_PORTAL;
 284                 connect_op = OST_CONNECT;
 285                 cli->cl_sp_me = LUSTRE_SP_CLI;
 286                 cli->cl_sp_to = LUSTRE_SP_OST;
 287                 ns_type = LDLM_NS_TYPE_OSC;
 288         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 289                    !strcmp(name, LUSTRE_LWP_NAME)) {
 290                 rq_portal = MDS_REQUEST_PORTAL;
 291                 rp_portal = MDC_REPLY_PORTAL;
 292                 connect_op = MDS_CONNECT;
 293                 if (is_lwp_on_ost(cli_name))
 294                         cli->cl_sp_me = LUSTRE_SP_OST;
 295                 else if (is_lwp_on_mdt(cli_name))
 296                         cli->cl_sp_me = LUSTRE_SP_MDT;
 297                 else
 298                         cli->cl_sp_me = LUSTRE_SP_CLI;
 299                 cli->cl_sp_to = LUSTRE_SP_MDT;
 300                 ns_type = LDLM_NS_TYPE_MDC;
 301         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 302                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 303                         /* OSP_on_MDT for other MDTs */
 304                         connect_op = MDS_CONNECT;
 305                         cli->cl_sp_to = LUSTRE_SP_MDT;
 306                         ns_type = LDLM_NS_TYPE_MDC;
 307                         rq_portal = OUT_PORTAL;
 308                 } else {
 309                         /* OSP on MDT for OST */
 310                         connect_op = OST_CONNECT;
 311                         cli->cl_sp_to = LUSTRE_SP_OST;
 312                         ns_type = LDLM_NS_TYPE_OSC;
 313                         rq_portal = OST_REQUEST_PORTAL;
 314                 }
 315                 rp_portal = OSC_REPLY_PORTAL;
 316                 cli->cl_sp_me = LUSTRE_SP_MDT;
 317         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 318                 rq_portal = MGS_REQUEST_PORTAL;
 319                 rp_portal = MGC_REPLY_PORTAL;
 320                 connect_op = MGS_CONNECT;
 321                 cli->cl_sp_me = LUSTRE_SP_MGC;
 322                 cli->cl_sp_to = LUSTRE_SP_MGS;
 323                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 324                 ns_type = LDLM_NS_TYPE_MGC;
 325         } else {
 326                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 327                        name);
 328                 RETURN(-EINVAL);
 329         }
 330
 331         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 332                 CERROR("requires a TARGET UUID\n");
 333                 RETURN(-EINVAL);
 334         }
 335
 336         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 337                 CERROR("client UUID must be less than 38 characters\n");
 338                 RETURN(-EINVAL);
 339         }
 340
 341         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 342                 CERROR("setup requires a SERVER UUID\n");
 343                 RETURN(-EINVAL);
 344         }
 345
 346         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 347                 CERROR("target UUID must be less than 38 characters\n");
 348                 RETURN(-EINVAL);
 349         }
 350
 351         init_rwsem(&cli->cl_sem);
 352         mutex_init(&cli->cl_mgc_mutex);
 353         cli->cl_conn_count = 0;
 354         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 355                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 356                      sizeof(server_uuid)));
 357
 358         cli->cl_dirty_pages = 0;
 359         cli->cl_avail_grant = 0;
 360         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 361         /* cl_dirty_max_pages may be changed at connect time in
 362          * ptlrpc_connect_interpret(). */
 363         client_adjust_max_dirty(cli);
 364         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 365         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 366         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 367         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 368         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 369         spin_lock_init(&cli->cl_loi_list_lock);
 370         atomic_set(&cli->cl_pending_w_pages, 0);
 371         atomic_set(&cli->cl_pending_r_pages, 0);
 372         cli->cl_r_in_flight = 0;
 373         cli->cl_w_in_flight = 0;
 374
 375         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 376         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 377         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 378         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 379         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 380         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 381
 382         /* lru for osc. */
 383         INIT_LIST_HEAD(&cli->cl_lru_osc);
 384         atomic_set(&cli->cl_lru_shrinkers, 0);
 385         atomic_long_set(&cli->cl_lru_busy, 0);
 386         atomic_long_set(&cli->cl_lru_in_list, 0);
 387         INIT_LIST_HEAD(&cli->cl_lru_list);
 388         spin_lock_init(&cli->cl_lru_list_lock);
 389         atomic_long_set(&cli->cl_unstable_count, 0);
 390         INIT_LIST_HEAD(&cli->cl_shrink_list);
 391
 392         init_waitqueue_head(&cli->cl_destroy_waitq);
 393         atomic_set(&cli->cl_destroy_in_flight, 0);
 394 #ifdef ENABLE_CHECKSUM
 395         /* Turn on checksumming by default. */
 396         cli->cl_checksum = 1;
 397         /*
 398          * The supported checksum types will be worked out at connect time
 399          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 400          * through procfs.
 401          */
 402         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 403 #endif
 404         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 405
 406         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
 407          * from OFD after connecting. */
 408         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 409
 410         /* set cl_chunkbits default value to PAGE_SHIFT,
 411          * it will be updated at OSC connection time. */
 412         cli->cl_chunkbits = PAGE_SHIFT;
 413
 414         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 415                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 416         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 417                 cli->cl_max_rpcs_in_flight = 2;
 418         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 419                 cli->cl_max_rpcs_in_flight = 3;
 420         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 421                 cli->cl_max_rpcs_in_flight = 4;
 422         } else {
 423                 if (osc_on_mdt(obddev->obd_name))
 424                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 425                 else
 426                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 427         }
 428
 429         spin_lock_init(&cli->cl_mod_rpcs_lock);
 430         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 431         cli->cl_max_mod_rpcs_in_flight = 0;
 432         cli->cl_mod_rpcs_in_flight = 0;
 433         cli->cl_close_rpcs_in_flight = 0;
 434         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 435         cli->cl_mod_tag_bitmap = NULL;
 436
 437         if (connect_op == MDS_CONNECT) {
 438                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 439                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 440                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 441                 if (cli->cl_mod_tag_bitmap == NULL)
 442                         GOTO(err, rc = -ENOMEM);
 443         }
 444
 445         rc = ldlm_get_ref();
 446         if (rc) {
 447                 CERROR("ldlm_get_ref failed: %d\n", rc);
 448                 GOTO(err, rc);
 449         }
 450
 451         ptlrpc_init_client(rq_portal, rp_portal, name,
 452                            &obddev->obd_ldlm_client);
 453
 454         imp = class_new_import(obddev);
 455         if (imp == NULL)
 456                 GOTO(err_ldlm, rc = -ENOENT);
 457         imp->imp_client = &obddev->obd_ldlm_client;
 458         imp->imp_connect_op = connect_op;
 459         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 460                LUSTRE_CFG_BUFLEN(lcfg, 1));
 461         class_import_put(imp);
 462
 463         if (lustre_cfg_buf(lcfg, 4)) {
 464                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 465
 466                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 467                         rc = -EINVAL;
 468                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 469                                obddev->obd_name, lustre_cfg_string(lcfg, 4),
 470                                rc);
 471                         GOTO(err_import, rc);
 472                 }
 473                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 474                 imp->imp_connection = &fake_conn;
 475         }
 476
 477         rc = client_import_add_conn(imp, &server_uuid, 1);
 478         if (rc) {
 479                 CERROR("can't add initial connection\n");
 480                 GOTO(err_import, rc);
 481         }
 482         imp->imp_connection = NULL;
 483
 484         cli->cl_import = imp;
 485         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 486         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 487
 488         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 489                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 490                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 491                                name, obddev->obd_name,
 492                                cli->cl_target_uuid.uuid);
 493                         spin_lock(&imp->imp_lock);
 494                         imp->imp_deactive = 1;
 495                         spin_unlock(&imp->imp_lock);
 496                 }
 497         }
 498
 499         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 500                                                    LDLM_NAMESPACE_CLIENT,
 501                                                    LDLM_NAMESPACE_GREEDY,
 502                                                    ns_type);
 503         if (obddev->obd_namespace == NULL) {
 504                 CERROR("Unable to create client namespace - %s\n",
 505                        obddev->obd_name);
 506                 GOTO(err_import, rc = -ENOMEM);
 507         }
 508
 509         RETURN(rc);
 510
 511 err_import:
 512         class_destroy_import(imp);
 513 err_ldlm:
 514         ldlm_put_ref();
 515 err:
 516         if (cli->cl_mod_tag_bitmap != NULL)
 517                 OBD_FREE(cli->cl_mod_tag_bitmap,
 518                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 519         cli->cl_mod_tag_bitmap = NULL;
 520         RETURN(rc);
 521
 522 }
 523 EXPORT_SYMBOL(client_obd_setup);
 524
 525 int client_obd_cleanup(struct obd_device *obddev)
 526 {
 527         struct client_obd *cli = &obddev->u.cli;
 528         ENTRY;
 529
 530         ldlm_namespace_free_post(obddev->obd_namespace);
 531         obddev->obd_namespace = NULL;
 532
 533         obd_cleanup_client_import(obddev);
 534         LASSERT(obddev->u.cli.cl_import == NULL);
 535
 536         ldlm_put_ref();
 537
 538         if (cli->cl_mod_tag_bitmap != NULL)
 539                 OBD_FREE(cli->cl_mod_tag_bitmap,
 540                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 541         cli->cl_mod_tag_bitmap = NULL;
 542
 543         RETURN(0);
 544 }
 545 EXPORT_SYMBOL(client_obd_cleanup);
 546
 547 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 548 int client_connect_import(const struct lu_env *env,
 549                           struct obd_export **exp,
 550                           struct obd_device *obd, struct obd_uuid *cluuid,
 551                           struct obd_connect_data *data, void *localdata)
 552 {
 553         struct client_obd       *cli    = &obd->u.cli;
 554         struct obd_import       *imp    = cli->cl_import;
 555         struct obd_connect_data *ocd;
 556         struct lustre_handle    conn    = { 0 };
 557         int                     rc;
 558         bool                    is_mdc = false;
 559         ENTRY;
 560
 561         *exp = NULL;
 562         down_write(&cli->cl_sem);
 563         if (cli->cl_conn_count > 0)
 564                 GOTO(out_sem, rc = -EALREADY);
 565
 566         rc = class_connect(&conn, obd, cluuid);
 567         if (rc)
 568                 GOTO(out_sem, rc);
 569
 570         cli->cl_conn_count++;
 571         *exp = class_conn2export(&conn);
 572
 573         LASSERT(obd->obd_namespace);
 574
 575         imp->imp_dlm_handle = conn;
 576         rc = ptlrpc_init_import(imp);
 577         if (rc != 0)
 578                 GOTO(out_ldlm, rc);
 579
 580         ocd = &imp->imp_connect_data;
 581         if (data) {
 582                 *ocd = *data;
 583                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
 584                                  LUSTRE_MDC_NAME, 3) == 0;
 585                 if (is_mdc)
 586                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
 587                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 588                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 589         }
 590
 591         rc = ptlrpc_connect_import(imp);
 592         if (rc != 0) {
 593                 if (data && is_mdc)
 594                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 595                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 596                 GOTO(out_ldlm, rc);
 597         }
 598         LASSERT(*exp != NULL && (*exp)->exp_connection);
 599
 600         if (data) {
 601                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 602                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 603                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 604                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 605                 /* clear the flag as it was not set and is not known
 606                  * by upper layers */
 607                 if (is_mdc)
 608                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 609         }
 610
 611         ptlrpc_pinger_add_import(imp);
 612
 613         EXIT;
 614
 615         if (rc) {
 616 out_ldlm:
 617                 cli->cl_conn_count--;
 618                 class_disconnect(*exp);
 619                 *exp = NULL;
 620         }
 621 out_sem:
 622         up_write(&cli->cl_sem);
 623
 624         return rc;
 625 }
 626 EXPORT_SYMBOL(client_connect_import);
 627
 628 int client_disconnect_export(struct obd_export *exp)
 629 {
 630         struct obd_device *obd = class_exp2obd(exp);
 631         struct client_obd *cli;
 632         struct obd_import *imp;
 633         int rc = 0, err;
 634         ENTRY;
 635
 636         if (!obd) {
 637                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 638                        exp, exp ? exp->exp_handle.h_cookie : -1);
 639                 RETURN(-EINVAL);
 640         }
 641
 642         cli = &obd->u.cli;
 643         imp = cli->cl_import;
 644
 645         down_write(&cli->cl_sem);
 646         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 647                 cli->cl_conn_count);
 648
 649         if (cli->cl_conn_count == 0) {
 650                 CERROR("disconnecting disconnected device (%s)\n",
 651                        obd->obd_name);
 652                 GOTO(out_disconnect, rc = -EINVAL);
 653         }
 654
 655         cli->cl_conn_count--;
 656         if (cli->cl_conn_count != 0)
 657                 GOTO(out_disconnect, rc = 0);
 658
 659         /* Mark import deactivated now, so we don't try to reconnect if any
 660          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 661          * fully deactivate the import, or that would drop all requests. */
 662         spin_lock(&imp->imp_lock);
 663         imp->imp_deactive = 1;
 664         spin_unlock(&imp->imp_lock);
 665
 666         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 667          * delete it regardless.  (It's safe to delete an import that was
 668          * never added.) */
 669         (void)ptlrpc_pinger_del_import(imp);
 670
 671         if (obd->obd_namespace != NULL) {
 672                 /* obd_force == local only */
 673                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 674                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 675                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 676         }
 677
 678         /* There's no need to hold sem while disconnecting an import,
 679          * and it may actually cause deadlock in GSS. */
 680         up_write(&cli->cl_sem);
 681         rc = ptlrpc_disconnect_import(imp, 0);
 682         down_write(&cli->cl_sem);
 683
 684         ptlrpc_invalidate_import(imp);
 685
 686         EXIT;
 687
 688 out_disconnect:
 689         /* Use server style - class_disconnect should be always called for
 690          * o_disconnect. */
 691         err = class_disconnect(exp);
 692         if (!rc && err)
 693                 rc = err;
 694
 695         up_write(&cli->cl_sem);
 696
 697         RETURN(rc);
 698 }
 699 EXPORT_SYMBOL(client_disconnect_export);
 700
 701 #ifdef HAVE_SERVER_SUPPORT
 702 int server_disconnect_export(struct obd_export *exp)
 703 {
 704         int rc;
 705         ENTRY;
 706
 707         /* Disconnect early so that clients can't keep using export. */
 708         rc = class_disconnect(exp);
 709         /* Close import to avoid sending any requests. */
 710         if (exp->exp_imp_reverse)
 711                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 712
 713         ldlm_bl_thread_wakeup();
 714
 715         /* complete all outstanding replies */
 716         spin_lock(&exp->exp_lock);
 717         while (!list_empty(&exp->exp_outstanding_replies)) {
 718                 struct ptlrpc_reply_state *rs =
 719                         list_entry(exp->exp_outstanding_replies.next,
 720                                        struct ptlrpc_reply_state, rs_exp_list);
 721                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 722
 723                 spin_lock(&svcpt->scp_rep_lock);
 724
 725                 list_del_init(&rs->rs_exp_list);
 726                 spin_lock(&rs->rs_lock);
 727                 ptlrpc_schedule_difficult_reply(rs);
 728                 spin_unlock(&rs->rs_lock);
 729
 730                 spin_unlock(&svcpt->scp_rep_lock);
 731         }
 732         spin_unlock(&exp->exp_lock);
 733
 734         RETURN(rc);
 735 }
 736 EXPORT_SYMBOL(server_disconnect_export);
 737
 738 /* --------------------------------------------------------------------------
 739  * from old lib/target.c
 740  * -------------------------------------------------------------------------- */
 741
 742 static int target_handle_reconnect(struct lustre_handle *conn,
 743                                    struct obd_export *exp,
 744                                    struct obd_uuid *cluuid)
 745 {
 746         struct obd_device *target;
 747         struct lustre_handle *hdl;
 748         cfs_time_t now;
 749         cfs_time_t deadline;
 750         int timeout;
 751         int rc = 0;
 752         ENTRY;
 753
 754         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 755         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 756                 conn->cookie = exp->exp_handle.h_cookie;
 757                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 758                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
 759                 RETURN(0);
 760         }
 761
 762         target = exp->exp_obd;
 763
 764         /* Might be a re-connect after a partition. */
 765         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 766                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 767                               "with handle %#llx. Rejecting client "
 768                               "with the same UUID trying to reconnect "
 769                               "with handle %#llx\n", target->obd_name,
 770                               obd_uuid2str(&exp->exp_client_uuid),
 771                               obd_export_nid2str(exp),
 772                               hdl->cookie, conn->cookie);
 773                 memset(conn, 0, sizeof *conn);
 774                 /* target_handle_connect() treats EALREADY and
 775                  * -EALREADY differently.  -EALREADY is an error
 776                  * (same UUID, different handle). */
 777                 RETURN(-EALREADY);
 778         }
 779
 780         if (!target->obd_recovering) {
 781                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 782                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 783                         obd_export_nid2str(exp));
 784                 GOTO(out_already, rc);
 785         }
 786
 787         now = cfs_time_current();
 788         deadline = cfs_timer_deadline(&target->obd_recovery_timer);
 789         if (cfs_time_before(now, deadline)) {
 790                 struct target_distribute_txn_data *tdtd =
 791                                         class_exp2tgt(exp)->lut_tdtd;
 792                 int size = 0;
 793                 int count = 0;
 794                 char *buf = NULL;
 795
 796                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
 797                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 798                         buf = tdtd->tdtd_show_update_logs_retrievers(
 799                                 tdtd->tdtd_show_retrievers_cbdata,
 800                                 &size, &count);
 801
 802                 if (count > 0)
 803                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 804                                       "%d:%.02d. It is due to DNE recovery "
 805                                       "failed/stuck on the %d MDT(s):%s. "
 806                                       "Please wait until all MDTs recovered "
 807                                       "or abort the recovery by force.\n",
 808                                       target->obd_name, timeout / 60,
 809                                       timeout % 60, count,
 810                                       buf ? buf : "unknown (not enough RAM)");
 811                 else
 812                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 813                                       "%d:%.02d. If you do not want to wait "
 814                                       "more, please abort the recovery by "
 815                                       "force.\n", target->obd_name,
 816                                       timeout / 60, timeout % 60);
 817
 818                 if (buf != NULL)
 819                         OBD_FREE(buf, size);
 820         } else {
 821                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
 822                 LCONSOLE_WARN("%s: Recovery already passed deadline"
 823                         " %d:%.02d, It is most likely due to DNE"
 824                         " recovery is failed or stuck, please wait a"
 825                         " few more minutes or abort the recovery.\n",
 826                         target->obd_name, timeout / 60, timeout % 60);
 827         }
 828
 829 out_already:
 830         conn->cookie = exp->exp_handle.h_cookie;
 831         /* target_handle_connect() treats EALREADY and
 832          * -EALREADY differently.  EALREADY means we are
 833          * doing a valid reconnect from the same client. */
 834         RETURN(EALREADY);
 835 }
 836
 837 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 838                           int error)
 839 {
 840         struct obd_export *exp = cb_data;
 841
 842         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 843                obd->obd_name, exp->exp_client_uuid.uuid);
 844
 845         spin_lock(&exp->exp_lock);
 846         exp->exp_need_sync = 0;
 847         spin_unlock(&exp->exp_lock);
 848         class_export_cb_put(exp);
 849 }
 850
 851 static void
 852 check_and_start_recovery_timer(struct obd_device *obd,
 853                                struct ptlrpc_request *req, int new_client);
 854
 855 /**
 856  * update flags for import during reconnect process
 857  */
 858 static int rev_import_flags_update(struct obd_import *revimp,
 859                                    struct ptlrpc_request *req)
 860 {
 861         int rc;
 862         struct obd_connect_data *data;
 863
 864         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 865
 866         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 867                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 868         else
 869                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 870
 871         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 872
 873         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 874         if (rc) {
 875                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 876                         revimp->imp_client->cli_name,
 877                         libcfs_id2str(req->rq_peer), rc);
 878                 return rc;
 879         }
 880
 881         return 0;
 882 }
 883
 884 /**
 885  * Allocate a new reverse import for an export.
 886  *
 887  * \retval -errno in case error hit
 888  * \retval 0 if reverse import correctly init
 889  **/
 890 int rev_import_init(struct obd_export *export)
 891 {
 892         struct obd_device *obd = export->exp_obd;
 893         struct obd_import *revimp;
 894
 895         LASSERT(export->exp_imp_reverse == NULL);
 896
 897         revimp = class_new_import(obd);
 898         if (revimp == NULL)
 899                 return -ENOMEM;
 900
 901         revimp->imp_remote_handle.cookie = 0ULL;
 902         revimp->imp_client = &obd->obd_ldlm_client;
 903         revimp->imp_dlm_fake = 1;
 904
 905         /* it is safe to connect import in new state as no sends possible */
 906         spin_lock(&export->exp_lock);
 907         export->exp_imp_reverse = revimp;
 908         spin_unlock(&export->exp_lock);
 909         class_import_put(revimp);
 910
 911         return 0;
 912 }
 913 EXPORT_SYMBOL(rev_import_init);
 914
 915 /**
 916  * Handle reconnect for an export.
 917  *
 918  * \param exp export to handle reconnect process
 919  * \param req client reconnect request
 920  *
 921  * \retval -rc in case securitfy flavor can't be changed
 922  * \retval 0 in case none problems
 923  */
 924 static int rev_import_reconnect(struct obd_export *exp,
 925                                 struct ptlrpc_request *req)
 926 {
 927         struct obd_import *revimp = exp->exp_imp_reverse;
 928         struct lustre_handle *lh;
 929         int rc;
 930
 931         /* avoid sending a request until import flags are changed */
 932         ptlrpc_import_enter_resend(revimp);
 933
 934         if (revimp->imp_connection != NULL)
 935                 ptlrpc_connection_put(revimp->imp_connection);
 936
 937         /*
 938          * client from recovery don't have a handle so we need to take from
 939          * request. it may produce situation when wrong client connected
 940          * to recovery as we trust a client uuid
 941          */
 942         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 943         revimp->imp_remote_handle = *lh;
 944
 945         /* unknown versions will be caught in
 946          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 947         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 948
 949         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 950
 951         rc = rev_import_flags_update(revimp, req);
 952         if (rc != 0) {
 953                 /* it is safe to still be in RECOVERY phase as we are not able
 954                  * to setup correct security flavor so requests are not able to
 955                  * be delivered correctly */
 956                 return rc;
 957         }
 958
 959         /* resend all rpc's via new connection */
 960         return ptlrpc_import_recovery_state_machine(revimp);
 961 }
 962
 963 int target_handle_connect(struct ptlrpc_request *req)
 964 {
 965         struct obd_device *target = NULL;
 966         struct obd_export *export = NULL;
 967         /* connect handle - filled from target_handle_reconnect in
 968          * reconnect case */
 969         struct lustre_handle conn;
 970         struct lustre_handle *tmp;
 971         struct obd_uuid tgtuuid;
 972         struct obd_uuid cluuid;
 973         char *str;
 974         int rc = 0;
 975         char *target_start;
 976         int target_len;
 977         bool     mds_conn = false, lw_client = false, initial_conn = false;
 978         bool     mds_mds_conn = false;
 979         bool     new_mds_mds_conn = false;
 980         bool     target_referenced = false;
 981         struct obd_connect_data *data, *tmpdata;
 982         int size, tmpsize;
 983         lnet_nid_t *client_nid = NULL;
 984         ENTRY;
 985
 986         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 987
 988         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 989         if (str == NULL) {
 990                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 991                 GOTO(out, rc = -EINVAL);
 992         }
 993
 994         obd_str2uuid(&tgtuuid, str);
 995         target = class_uuid2obd(&tgtuuid);
 996         if (!target)
 997                 target = class_name2obd(str);
 998
 999         if (!target) {
1000                 deuuidify(str, NULL, &target_start, &target_len);
1001                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
1002                                    "from %s (no target). If you are running "
1003                                    "an HA pair check that the target is "
1004                                    "mounted on the other server.\n", str,
1005                                    libcfs_nid2str(req->rq_peer.nid));
1006                 GOTO(out, rc = -ENODEV);
1007         }
1008
1009         spin_lock(&target->obd_dev_lock);
1010         if (target->obd_stopping || !target->obd_set_up) {
1011                 spin_unlock(&target->obd_dev_lock);
1012
1013                 deuuidify(str, NULL, &target_start, &target_len);
1014                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1015                               target_len, target_start,
1016                               libcfs_nid2str(req->rq_peer.nid),
1017                               (target->obd_stopping ?
1018                                "stopping" : "not set up"));
1019                 GOTO(out, rc = -ENODEV);
1020         }
1021
1022         if (target->obd_no_conn) {
1023                 spin_unlock(&target->obd_dev_lock);
1024
1025                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1026                                "from %s\n", target->obd_name,
1027                                libcfs_nid2str(req->rq_peer.nid));
1028                 GOTO(out, rc = -EAGAIN);
1029         }
1030
1031         /* Make sure the target isn't cleaned up while we're here. Yes,
1032          * there's still a race between the above check and our incref here.
1033          * Really, class_uuid2obd should take the ref. */
1034         class_incref(target, __func__, current);
1035         target_referenced = true;
1036
1037         target->obd_conn_inprogress++;
1038         spin_unlock(&target->obd_dev_lock);
1039
1040         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1041         if (str == NULL) {
1042                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1043                 GOTO(out, rc = -EINVAL);
1044         }
1045
1046         obd_str2uuid(&cluuid, str);
1047
1048         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1049         if (tmp == NULL)
1050                 GOTO(out, rc = -EPROTO);
1051
1052         conn = *tmp;
1053
1054         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1055                                     RCL_CLIENT);
1056         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1057         if (!data)
1058                 GOTO(out, rc = -EPROTO);
1059
1060         rc = req_capsule_server_pack(&req->rq_pill);
1061         if (rc)
1062                 GOTO(out, rc);
1063
1064 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1065         /* Don't allow clients to connect that are using old 1.8 format
1066          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1067          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1068          * FULL20 flag should be set on all connections since 2.0, but no
1069          * longer affects behaviour.
1070          *
1071          * Later this check will be disabled and the flag can be retired
1072          * completely once interop with 3.0 is no longer needed.
1073          */
1074         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1075                 GOTO(out, rc = -EPROTO);
1076 #endif
1077
1078         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1079                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1080                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1081                     data->ocd_version > LUSTRE_VERSION_CODE +
1082                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1083                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1084                                   "libclient connection attempt",
1085                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1086                                   "old" : "new",
1087                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1088                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1089                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1090                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1091                         data = req_capsule_server_sized_get(&req->rq_pill,
1092                                                             &RMF_CONNECT_DATA,
1093                                     offsetof(typeof(*data), ocd_version) +
1094                                              sizeof(data->ocd_version));
1095                         if (data) {
1096                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1097                                 data->ocd_version = LUSTRE_VERSION_CODE;
1098                         }
1099                         GOTO(out, rc = -EPROTO);
1100                 }
1101         }
1102
1103         /* Note: lw_client is needed in MDS-MDS failover during update log
1104          * processing, so we needs to allow lw_client to be connected at
1105          * anytime, instead of only the initial connection */
1106         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1107
1108         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1109                 initial_conn = true;
1110                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1111                 mds_mds_conn = (data->ocd_connect_flags &
1112                                 OBD_CONNECT_MDS_MDS) != 0;
1113
1114                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1115                  * for Imperative Recovery connection from MGC to MGS.
1116                  *
1117                  * Via check OBD_CONNECT_FID, we can distinguish whether
1118                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1119                  * MGC or MDT. */
1120                 if (!lw_client &&
1121                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1122                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1123                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1124                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1125                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1126                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1127
1128                         /* We do not support the MDT-MDT interoperations with
1129                          * different version MDT because of protocol changes. */
1130                         if (unlikely(major != LUSTRE_MAJOR ||
1131                                      minor != LUSTRE_MINOR ||
1132                                      abs(patch - LUSTRE_PATCH) > 3)) {
1133                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1134                                         "connection from different version MDT "
1135                                         "(%d.%d.%d.%d) %s %s\n",
1136                                         target->obd_name, LUSTRE_MAJOR,
1137                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1138                                         major, minor, patch,
1139                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1140                                         libcfs_nid2str(req->rq_peer.nid), str);
1141
1142                                 GOTO(out, rc = -EPROTO);
1143                         }
1144                 }
1145         }
1146
1147         /* lctl gets a backstage, all-access pass. */
1148         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1149                 goto dont_check_exports;
1150
1151         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1152         if (!export)
1153                 goto no_export;
1154
1155         /* We've found an export in the hash. */
1156
1157         spin_lock(&export->exp_lock);
1158
1159         if (export->exp_connecting) { /* bug 9635, et. al. */
1160                 spin_unlock(&export->exp_lock);
1161                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1162                               export->exp_obd->obd_name, export,
1163                               libcfs_nid2str(req->rq_peer.nid));
1164                 class_export_put(export);
1165                 export = NULL;
1166                 rc = -EALREADY;
1167         } else if ((mds_conn || (lw_client && initial_conn) ||
1168                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1169                    export->exp_connection != NULL) {
1170                 spin_unlock(&export->exp_lock);
1171                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1172                         /* MDS or LWP reconnected after failover. */
1173                         LCONSOLE_WARN("%s: Received %s connection from "
1174                             "%s, removing former export from %s\n",
1175                             target->obd_name, mds_conn ? "MDS" : "LWP",
1176                             libcfs_nid2str(req->rq_peer.nid),
1177                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1178                 } else {
1179                         /* New MDS connection from the same NID. */
1180                         LCONSOLE_WARN("%s: Received new %s connection from "
1181                                 "%s, removing former export from same NID\n",
1182                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1183                                 libcfs_nid2str(req->rq_peer.nid));
1184                 }
1185
1186                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1187                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1188                         /* Because exports between MDTs will always be
1189                          * kept, let's do not fail such export if they
1190                          * come from the same NID, otherwise it might
1191                          * cause eviction between MDTs, which might
1192                          * cause namespace inconsistency */
1193                         spin_lock(&export->exp_lock);
1194                         export->exp_connecting = 1;
1195                         spin_unlock(&export->exp_lock);
1196                         conn.cookie = export->exp_handle.h_cookie;
1197                         rc = EALREADY;
1198                 } else {
1199                         class_fail_export(export);
1200                         class_export_put(export);
1201                         export = NULL;
1202                         rc = 0;
1203                 }
1204         } else if (export->exp_connection != NULL && initial_conn &&
1205                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1206                 spin_unlock(&export->exp_lock);
1207                 /* In MDS failover we have static UUID but NID can change. */
1208                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1209                               "existing nid %s is already connected\n",
1210                               target->obd_name, cluuid.uuid,
1211                               libcfs_nid2str(req->rq_peer.nid),
1212                               libcfs_nid2str(
1213                                       export->exp_connection->c_peer.nid));
1214                 rc = -EALREADY;
1215                 class_export_put(export);
1216                 export = NULL;
1217         } else {
1218                 export->exp_connecting = 1;
1219                 spin_unlock(&export->exp_lock);
1220                 LASSERT(export->exp_obd == target);
1221
1222                 rc = target_handle_reconnect(&conn, export, &cluuid);
1223         }
1224
1225         /* If we found an export, we already unlocked. */
1226         if (!export) {
1227 no_export:
1228                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1229         } else if (req->rq_export == NULL &&
1230                    atomic_read(&export->exp_rpc_count) > 0) {
1231                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1232                               "still busy with %d references\n",
1233                               target->obd_name, cluuid.uuid,
1234                               libcfs_nid2str(req->rq_peer.nid),
1235                               atomic_read(&export->exp_refcount));
1236                 GOTO(out, rc = -EBUSY);
1237         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1238                 if (!strstr(cluuid.uuid, "mdt"))
1239                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1240                                       "known client %s (at %s) because it "
1241                                       "is indicating it is a new client",
1242                                       target->obd_name, cluuid.uuid,
1243                                       libcfs_nid2str(req->rq_peer.nid));
1244                 GOTO(out, rc = -EALREADY);
1245         } else {
1246                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1247         }
1248
1249         if (rc < 0) {
1250                 GOTO(out, rc);
1251         }
1252
1253         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1254                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1255               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1256               export, (long)cfs_time_current_sec(),
1257               export ? (long)export->exp_last_request_time : 0);
1258
1259         /* If this is the first time a client connects, reset the recovery
1260          * timer. Discard lightweight connections which might be local. */
1261         if (!lw_client && rc == 0 && target->obd_recovering)
1262                 check_and_start_recovery_timer(target, req, export == NULL);
1263
1264         /* We want to handle EALREADY but *not* -EALREADY from
1265          * target_handle_reconnect(), return reconnection state in a flag. */
1266         if (rc == EALREADY) {
1267                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1268                 rc = 0;
1269         } else {
1270                 LASSERT(rc == 0);
1271         }
1272
1273         /* Tell the client if we support replayable requests. */
1274         if (target->obd_replayable)
1275                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1276         client_nid = &req->rq_peer.nid;
1277
1278         if (export == NULL) {
1279                 /* allow lightweight connections during recovery */
1280                 /* allow "new" MDT to be connected during recovery, since we
1281                  * need retrieve recovery update records from it */
1282                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1283                         cfs_time_t t;
1284                         int     c; /* connected */
1285                         int     i; /* in progress */
1286                         int     k; /* known */
1287                         int     s; /* stale/evicted */
1288
1289                         c = atomic_read(&target->obd_connected_clients);
1290                         i = atomic_read(&target->obd_lock_replay_clients);
1291                         k = target->obd_max_recoverable_clients;
1292                         s = target->obd_stale_clients;
1293                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1294                         t = cfs_time_sub(t, cfs_time_current());
1295                         t = cfs_duration_sec(t);
1296                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1297                                       "(at %s), waiting for %d known clients "
1298                                       "(%d recovered, %d in progress, and %d "
1299                                       "evicted) to recover in %d:%.02d\n",
1300                                       target->obd_name, cluuid.uuid,
1301                                       libcfs_nid2str(req->rq_peer.nid), k,
1302                                       c - i, i, s, (int)t / 60,
1303                                       (int)t % 60);
1304                         rc = -EBUSY;
1305                 } else {
1306 dont_check_exports:
1307                         rc = obd_connect(req->rq_svc_thread->t_env,
1308                                          &export, target, &cluuid, data,
1309                                          client_nid);
1310                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1311                                 lustre_msg_add_op_flags(req->rq_repmsg,
1312                                                         MSG_CONNECT_RECOVERING);
1313                         if (rc == 0) {
1314                                 conn.cookie = export->exp_handle.h_cookie;
1315                                 rc = rev_import_init(export);
1316                         }
1317
1318                         if (mds_mds_conn)
1319                                 new_mds_mds_conn = true;
1320                 }
1321         } else {
1322                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1323                                    export, target, &cluuid, data, client_nid);
1324         }
1325         if (rc)
1326                 GOTO(out, rc);
1327
1328         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1329         data->ocd_instance = target->u.obt.obt_instance;
1330
1331         /* Return only the parts of obd_connect_data that we understand, so the
1332          * client knows that we don't understand the rest. */
1333         if (data) {
1334                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1335                                                RCL_SERVER);
1336                 tmpdata = req_capsule_server_get(&req->rq_pill,
1337                                                  &RMF_CONNECT_DATA);
1338                 /* Don't use struct assignment here, because the client reply
1339                  * buffer may be smaller/larger than the local struct
1340                  * obd_connect_data. */
1341                 memcpy(tmpdata, data, min(tmpsize, size));
1342         }
1343
1344         /* If the client and the server are the same node, we will already
1345          * have an export that really points to the client's DLM export,
1346          * because we have a shared handles table.
1347          *
1348          * XXX this will go away when shaver stops sending the "connect" handle
1349          * in the real "remote handle" field of the request --phik 24 Apr 2003
1350          */
1351         ptlrpc_request_change_export(req, export);
1352
1353         spin_lock(&export->exp_lock);
1354         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1355                 spin_unlock(&export->exp_lock);
1356                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1357                        "or equal conn_cnt: %d >= %d\n",
1358                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1359                        export->exp_conn_cnt,
1360                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1361
1362                 GOTO(out, rc = -EALREADY);
1363         }
1364         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1365         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1366
1367         /* Don't evict liblustre clients for not pinging. */
1368         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1369                 export->exp_libclient = 1;
1370                 spin_unlock(&export->exp_lock);
1371
1372                 spin_lock(&target->obd_dev_lock);
1373                 list_del_init(&export->exp_obd_chain_timed);
1374                 spin_unlock(&target->obd_dev_lock);
1375         } else {
1376                 spin_unlock(&export->exp_lock);
1377         }
1378
1379         if (export->exp_connection != NULL) {
1380                 /* Check to see if connection came from another NID. */
1381                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1382                     !hlist_unhashed(&export->exp_nid_hash))
1383                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1384                                      &export->exp_connection->c_peer.nid,
1385                                      &export->exp_nid_hash);
1386
1387                 ptlrpc_connection_put(export->exp_connection);
1388         }
1389
1390         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1391                                                        req->rq_self,
1392                                                        &cluuid);
1393         if (hlist_unhashed(&export->exp_nid_hash))
1394                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1395                              &export->exp_connection->c_peer.nid,
1396                              &export->exp_nid_hash);
1397
1398         lustre_msg_set_handle(req->rq_repmsg, &conn);
1399
1400         rc = rev_import_reconnect(export, req);
1401         if (rc != 0)
1402                 GOTO(out, rc);
1403
1404         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1405                 int has_transno;
1406                 __u64 transno = data->ocd_transno;
1407
1408                 spin_lock(&export->exp_lock);
1409                 /* possible race with class_disconnect_stale_exports,
1410                  * export may be already in the eviction process */
1411                 if (export->exp_failed) {
1412                         spin_unlock(&export->exp_lock);
1413                         GOTO(out, rc = -ENODEV);
1414                 }
1415                 export->exp_in_recovery = 1;
1416                 export->exp_req_replay_needed = 1;
1417                 export->exp_lock_replay_needed = 1;
1418                 spin_unlock(&export->exp_lock);
1419
1420                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1421                                  MSG_CONNECT_TRANSNO);
1422                 if (has_transno && transno == 0)
1423                         CWARN("Connect with zero transno!\n");
1424
1425                 if (has_transno && transno > 0 &&
1426                     transno < target->obd_next_recovery_transno &&
1427                     transno > target->obd_last_committed) {
1428                         /* Another way is to use cmpxchg() to be lock-free. */
1429                         spin_lock(&target->obd_recovery_task_lock);
1430                         if (transno < target->obd_next_recovery_transno)
1431                                 target->obd_next_recovery_transno = transno;
1432                         spin_unlock(&target->obd_recovery_task_lock);
1433                 }
1434
1435                 atomic_inc(&target->obd_req_replay_clients);
1436                 atomic_inc(&target->obd_lock_replay_clients);
1437                 /* Note: MDS-MDS connection is allowed to be connected during
1438                  * recovery, no matter if the exports needs to be recoveried.
1439                  * Because we need retrieve updates logs from all other MDTs.
1440                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1441                  * also needs to be increased to match other recovery checking
1442                  * condition. */
1443                 if (new_mds_mds_conn)
1444                         target->obd_max_recoverable_clients++;
1445                 if (atomic_inc_return(&target->obd_connected_clients) ==
1446                     target->obd_max_recoverable_clients)
1447                         wake_up(&target->obd_next_transno_waitq);
1448         }
1449
1450         /* Tell the client we're in recovery, when client is involved in it. */
1451         if (target->obd_recovering && !lw_client)
1452                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1453
1454 out:
1455         if (export) {
1456                 spin_lock(&export->exp_lock);
1457                 export->exp_connecting = 0;
1458                 spin_unlock(&export->exp_lock);
1459
1460                 class_export_put(export);
1461         }
1462         if (target_referenced == true && target != NULL) {
1463                 spin_lock(&target->obd_dev_lock);
1464                 target->obd_conn_inprogress--;
1465                 spin_unlock(&target->obd_dev_lock);
1466
1467                 class_decref(target, __func__, current);
1468         }
1469         req->rq_status = rc;
1470         RETURN(rc);
1471 }
1472
1473 int target_handle_disconnect(struct ptlrpc_request *req)
1474 {
1475         int rc;
1476         ENTRY;
1477
1478         rc = req_capsule_server_pack(&req->rq_pill);
1479         if (rc)
1480                 RETURN(rc);
1481
1482         /* Keep the rq_export around so we can send the reply. */
1483         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1484
1485         RETURN(0);
1486 }
1487
1488 void target_destroy_export(struct obd_export *exp)
1489 {
1490         struct obd_import       *imp = NULL;
1491         /* exports created from last_rcvd data, and "fake"
1492            exports created by lctl don't have an import */
1493         spin_lock(&exp->exp_lock);
1494         if (exp->exp_imp_reverse != NULL) {
1495                 imp = exp->exp_imp_reverse;
1496                 exp->exp_imp_reverse = NULL;
1497         }
1498         spin_unlock(&exp->exp_lock);
1499         if (imp != NULL)
1500                 client_destroy_import(imp);
1501
1502         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1503         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1504         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1505         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1506 }
1507 EXPORT_SYMBOL(target_destroy_export);
1508
1509 /*
1510  * Recovery functions
1511  */
1512 static void target_request_copy_get(struct ptlrpc_request *req)
1513 {
1514         class_export_rpc_inc(req->rq_export);
1515         LASSERT(list_empty(&req->rq_list));
1516         INIT_LIST_HEAD(&req->rq_replay_list);
1517
1518         /* Increase refcount to keep request in queue. */
1519         atomic_inc(&req->rq_refcount);
1520         /* Let export know it has replays to be handled. */
1521         atomic_inc(&req->rq_export->exp_replay_count);
1522 }
1523
1524 static void target_request_copy_put(struct ptlrpc_request *req)
1525 {
1526         LASSERT(list_empty(&req->rq_replay_list));
1527         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1528
1529         atomic_dec(&req->rq_export->exp_replay_count);
1530         class_export_rpc_dec(req->rq_export);
1531         ptlrpc_server_drop_request(req);
1532 }
1533
1534 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1535 {
1536         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1537         struct obd_export     *exp = req->rq_export;
1538         struct ptlrpc_request *reqiter;
1539         struct ptlrpc_request *dup_req = NULL;
1540         int                    dup = 0;
1541
1542         LASSERT(exp);
1543
1544         spin_lock(&exp->exp_lock);
1545         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1546                                 rq_replay_list) {
1547                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1548                         dup_req = reqiter;
1549                         dup = 1;
1550                         break;
1551                 }
1552         }
1553
1554         if (dup) {
1555                 /* We expect it with RESENT and REPLAY flags. */
1556                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1557                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1558                         CERROR("invalid flags %x of resent replay\n",
1559                                lustre_msg_get_flags(req->rq_reqmsg));
1560
1561                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1562                         __u32 new_conn;
1563
1564                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1565                         if (new_conn >
1566                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1567                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1568                                                         new_conn);
1569                 }
1570         } else {
1571                 list_add_tail(&req->rq_replay_list,
1572                                   &exp->exp_req_replay_queue);
1573         }
1574
1575         spin_unlock(&exp->exp_lock);
1576         return dup;
1577 }
1578
1579 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1580 {
1581         LASSERT(!list_empty(&req->rq_replay_list));
1582         LASSERT(req->rq_export);
1583
1584         spin_lock(&req->rq_export->exp_lock);
1585         list_del_init(&req->rq_replay_list);
1586         spin_unlock(&req->rq_export->exp_lock);
1587 }
1588
1589 static void target_finish_recovery(struct lu_target *lut)
1590 {
1591         struct obd_device *obd = lut->lut_obd;
1592         ENTRY;
1593
1594         /* Only log a recovery message when recovery has occurred. */
1595         if (obd->obd_recovery_start) {
1596                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1597                                         obd->obd_recovery_start);
1598                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1599                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1600                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1601                         obd->obd_max_recoverable_clients,
1602                         atomic_read(&obd->obd_connected_clients),
1603                         obd->obd_stale_clients,
1604                         obd->obd_stale_clients == 1 ? "was" : "were");
1605         }
1606
1607         ldlm_reprocess_all_ns(obd->obd_namespace);
1608         spin_lock(&obd->obd_recovery_task_lock);
1609         if (!list_empty(&obd->obd_req_replay_queue) ||
1610             !list_empty(&obd->obd_lock_replay_queue) ||
1611             !list_empty(&obd->obd_final_req_queue)) {
1612                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1613                        obd->obd_name,
1614                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1615                        list_empty(&obd->obd_lock_replay_queue) ? \
1616                                "" : "lock ",
1617                        list_empty(&obd->obd_final_req_queue) ? \
1618                                "" : "final ");
1619                 spin_unlock(&obd->obd_recovery_task_lock);
1620                 LBUG();
1621         }
1622         spin_unlock(&obd->obd_recovery_task_lock);
1623
1624         obd->obd_recovery_end = cfs_time_current_sec();
1625
1626         /* When recovery finished, cleanup orphans on MDS and OST. */
1627         if (OBT(obd) && OBP(obd, postrecov)) {
1628                 int rc = OBP(obd, postrecov)(obd);
1629                 if (rc < 0)
1630                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1631                                       obd->obd_name, rc);
1632         }
1633         EXIT;
1634 }
1635
1636 static void abort_req_replay_queue(struct obd_device *obd)
1637 {
1638         struct ptlrpc_request *req, *n;
1639         struct list_head abort_list;
1640
1641         INIT_LIST_HEAD(&abort_list);
1642         spin_lock(&obd->obd_recovery_task_lock);
1643         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1644         spin_unlock(&obd->obd_recovery_task_lock);
1645         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1646                 DEBUG_REQ(D_WARNING, req, "aborted:");
1647                 req->rq_status = -ENOTCONN;
1648                 if (ptlrpc_error(req)) {
1649                         DEBUG_REQ(D_ERROR, req,
1650                                   "failed abort_req_reply; skipping");
1651                 }
1652                 target_exp_dequeue_req_replay(req);
1653                 target_request_copy_put(req);
1654         }
1655 }
1656
1657 static void abort_lock_replay_queue(struct obd_device *obd)
1658 {
1659         struct ptlrpc_request *req, *n;
1660         struct list_head abort_list;
1661
1662         INIT_LIST_HEAD(&abort_list);
1663         spin_lock(&obd->obd_recovery_task_lock);
1664         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1665         spin_unlock(&obd->obd_recovery_task_lock);
1666         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1667                 DEBUG_REQ(D_ERROR, req, "aborted:");
1668                 req->rq_status = -ENOTCONN;
1669                 if (ptlrpc_error(req)) {
1670                         DEBUG_REQ(D_ERROR, req,
1671                                   "failed abort_lock_reply; skipping");
1672                 }
1673                 target_request_copy_put(req);
1674         }
1675 }
1676
1677 /* Called from a cleanup function if the device is being cleaned up
1678    forcefully.  The exports should all have been disconnected already,
1679    the only thing left to do is
1680      - clear the recovery flags
1681      - cancel the timer
1682      - free queued requests and replies, but don't send replies
1683    Because the obd_stopping flag is set, no new requests should be received.
1684
1685 */
1686 void target_cleanup_recovery(struct obd_device *obd)
1687 {
1688         struct ptlrpc_request *req, *n;
1689         struct list_head clean_list;
1690
1691         INIT_LIST_HEAD(&clean_list);
1692         spin_lock(&obd->obd_dev_lock);
1693         if (!obd->obd_recovering) {
1694                 spin_unlock(&obd->obd_dev_lock);
1695                 EXIT;
1696                 return;
1697         }
1698         obd->obd_recovering = obd->obd_abort_recovery = 0;
1699         spin_unlock(&obd->obd_dev_lock);
1700
1701         spin_lock(&obd->obd_recovery_task_lock);
1702         target_cancel_recovery_timer(obd);
1703         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1704         spin_unlock(&obd->obd_recovery_task_lock);
1705
1706         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1707                 LASSERT(req->rq_reply_state == NULL);
1708                 target_exp_dequeue_req_replay(req);
1709                 target_request_copy_put(req);
1710         }
1711
1712         spin_lock(&obd->obd_recovery_task_lock);
1713         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1714         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1715         spin_unlock(&obd->obd_recovery_task_lock);
1716
1717         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1718                 LASSERT(req->rq_reply_state == NULL);
1719                 target_request_copy_put(req);
1720         }
1721
1722         EXIT;
1723 }
1724 EXPORT_SYMBOL(target_cleanup_recovery);
1725
1726 /* obd_recovery_task_lock should be held */
1727 void target_cancel_recovery_timer(struct obd_device *obd)
1728 {
1729         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1730         cfs_timer_disarm(&obd->obd_recovery_timer);
1731 }
1732
1733 static void target_start_recovery_timer(struct obd_device *obd)
1734 {
1735         if (obd->obd_recovery_start != 0)
1736                 return;
1737
1738         spin_lock(&obd->obd_dev_lock);
1739         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1740                 spin_unlock(&obd->obd_dev_lock);
1741                 return;
1742         }
1743
1744         LASSERT(obd->obd_recovery_timeout != 0);
1745
1746         if (obd->obd_recovery_start != 0) {
1747                 spin_unlock(&obd->obd_dev_lock);
1748                 return;
1749         }
1750
1751         cfs_timer_arm(&obd->obd_recovery_timer,
1752                       cfs_time_shift(obd->obd_recovery_timeout));
1753         obd->obd_recovery_start = cfs_time_current_sec();
1754         spin_unlock(&obd->obd_dev_lock);
1755
1756         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1757                       "or until %d client%s reconnect%s\n",
1758                       obd->obd_name,
1759                       obd->obd_recovery_timeout / 60,
1760                       obd->obd_recovery_timeout % 60,
1761                       obd->obd_max_recoverable_clients,
1762                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1763                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1764 }
1765
1766 /**
1767  * extend recovery window.
1768  *
1769  * if @extend is true, extend recovery window to have @drt remaining at least;
1770  * otherwise, make sure the recovery timeout value is not less than @drt.
1771  */
1772 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1773 {
1774         cfs_time_t now;
1775         cfs_time_t end;
1776         cfs_duration_t left;
1777         int to;
1778
1779         spin_lock(&obd->obd_dev_lock);
1780         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1781                 spin_unlock(&obd->obd_dev_lock);
1782                 return;
1783         }
1784         LASSERT(obd->obd_recovery_start != 0);
1785
1786         now  = cfs_time_current_sec();
1787         to   = obd->obd_recovery_timeout;
1788         end  = obd->obd_recovery_start + to;
1789         left = cfs_time_sub(end, now);
1790
1791         if (extend && (drt > left)) {
1792                 to += drt - left;
1793         } else if (!extend && (drt > to)) {
1794                 to = drt;
1795         }
1796
1797         if (to > obd->obd_recovery_time_hard)
1798                 to = obd->obd_recovery_time_hard;
1799         if (obd->obd_recovery_timeout < to) {
1800                 obd->obd_recovery_timeout = to;
1801                 end = obd->obd_recovery_start + to;
1802                 cfs_timer_arm(&obd->obd_recovery_timer,
1803                                 cfs_time_shift(end - now));
1804         }
1805         spin_unlock(&obd->obd_dev_lock);
1806
1807         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1808                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1809 }
1810
1811 /* Reset the timer with each new client connection */
1812 /*
1813  * This timer is actually reconnect_timer, which is for making sure
1814  * the total recovery window is at least as big as my reconnect
1815  * attempt timing. So the initial recovery time_out will be set to
1816  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1817  * from client is bigger than this, then the recovery time_out will
1818  * be extended to make sure the client could be reconnected, in the
1819  * process, the timeout from the new client should be ignored.
1820  */
1821
1822 static void
1823 check_and_start_recovery_timer(struct obd_device *obd,
1824                                struct ptlrpc_request *req,
1825                                int new_client)
1826 {
1827         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1828         struct obd_device_target *obt = &obd->u.obt;
1829
1830         if (!new_client && service_time)
1831                 /* Teach server about old server's estimates, as first guess
1832                  * at how long new requests will take. */
1833                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1834                             service_time);
1835
1836         target_start_recovery_timer(obd);
1837
1838         /* Convert the service time to RPC timeout,
1839          * and reuse service_time to limit stack usage. */
1840         service_time = at_est2timeout(service_time);
1841
1842         /* We expect other clients to timeout within service_time, then try
1843          * to reconnect, then try the failover server.  The max delay between
1844          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1845         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1846
1847         LASSERT(obt->obt_magic == OBT_MAGIC);
1848         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1849         if (service_time > obd->obd_recovery_timeout && !new_client)
1850                 extend_recovery_timer(obd, service_time, false);
1851 }
1852
1853 /** Health checking routines */
1854 static inline int exp_connect_healthy(struct obd_export *exp)
1855 {
1856         return (exp->exp_in_recovery);
1857 }
1858
1859 /** if export done req_replay or has replay in queue */
1860 static inline int exp_req_replay_healthy(struct obd_export *exp)
1861 {
1862         return (!exp->exp_req_replay_needed ||
1863                 atomic_read(&exp->exp_replay_count) > 0);
1864 }
1865
1866
1867 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1868 {
1869         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1870                exp_req_replay_healthy(exp);
1871 }
1872
1873 /** if export done lock_replay or has replay in queue */
1874 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1875 {
1876         return (!exp->exp_lock_replay_needed ||
1877                 atomic_read(&exp->exp_replay_count) > 0);
1878 }
1879
1880 static inline int exp_vbr_healthy(struct obd_export *exp)
1881 {
1882         return (!exp->exp_vbr_failed);
1883 }
1884
1885 static inline int exp_finished(struct obd_export *exp)
1886 {
1887         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1888 }
1889
1890 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1891 {
1892         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1893                 exp_finished(exp);
1894 }
1895
1896 static int check_for_next_transno(struct lu_target *lut)
1897 {
1898         struct ptlrpc_request *req = NULL;
1899         struct obd_device *obd = lut->lut_obd;
1900         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1901         int wake_up = 0, connected, completed, queue_len;
1902         __u64 req_transno = 0;
1903         __u64 update_transno = 0;
1904         __u64 next_transno = 0;
1905         ENTRY;
1906
1907         spin_lock(&obd->obd_recovery_task_lock);
1908         if (!list_empty(&obd->obd_req_replay_queue)) {
1909                 req = list_entry(obd->obd_req_replay_queue.next,
1910                                      struct ptlrpc_request, rq_list);
1911                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1912         }
1913
1914         if (tdtd != NULL)
1915                 update_transno = distribute_txn_get_next_transno(tdtd);
1916
1917         connected = atomic_read(&obd->obd_connected_clients);
1918         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1919         queue_len = obd->obd_requests_queued_for_recovery;
1920         next_transno = obd->obd_next_recovery_transno;
1921
1922         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1923                "req_transno: %llu, next_transno: %llu\n",
1924                obd->obd_max_recoverable_clients, connected, completed,
1925                queue_len, req_transno, next_transno);
1926
1927         if (obd->obd_abort_recovery) {
1928                 CDEBUG(D_HA, "waking for aborted recovery\n");
1929                 wake_up = 1;
1930         } else if (obd->obd_recovery_expired) {
1931                 CDEBUG(D_HA, "waking for expired recovery\n");
1932                 wake_up = 1;
1933         } else if (tdtd != NULL && req != NULL &&
1934                    is_req_replayed_by_update(req)) {
1935                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1936                          "next_transno%llu\n", req_transno, next_transno);
1937                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1938                        req_transno);
1939                 wake_up = 1;
1940         } else if (req_transno == next_transno ||
1941                    (update_transno != 0 && update_transno <= next_transno)) {
1942                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1943                 wake_up = 1;
1944         } else if (queue_len > 0 &&
1945                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1946                 /** handle gaps occured due to lost reply or VBR */
1947                 LASSERTF(req_transno >= next_transno,
1948                          "req_transno: %llu, next_transno: %llu\n",
1949                          req_transno, next_transno);
1950                 CDEBUG(D_HA,
1951                        "%s: waking for gap in transno, VBR is %s (skip: "
1952                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1953                        ", next_update %lld last_committed: %lld)\n",
1954                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1955                        next_transno, queue_len, completed, connected,
1956                        req_transno, update_transno, obd->obd_last_committed);
1957                 obd->obd_next_recovery_transno = req_transno;
1958                 wake_up = 1;
1959         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1960                 CDEBUG(D_HA, "waking for completed recovery\n");
1961                 wake_up = 1;
1962         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1963                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1964                        " by fail_lock, waking up (%lld)\n", next_transno);
1965                 obd->obd_next_recovery_transno = req_transno;
1966                 wake_up = 1;
1967         }
1968         spin_unlock(&obd->obd_recovery_task_lock);
1969         return wake_up;
1970 }
1971
1972 static int check_for_next_lock(struct lu_target *lut)
1973 {
1974         struct obd_device *obd = lut->lut_obd;
1975         int wake_up = 0;
1976
1977         spin_lock(&obd->obd_recovery_task_lock);
1978         if (!list_empty(&obd->obd_lock_replay_queue)) {
1979                 CDEBUG(D_HA, "waking for next lock\n");
1980                 wake_up = 1;
1981         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1982                 CDEBUG(D_HA, "waking for completed lock replay\n");
1983                 wake_up = 1;
1984         } else if (obd->obd_abort_recovery) {
1985                 CDEBUG(D_HA, "waking for aborted recovery\n");
1986                 wake_up = 1;
1987         } else if (obd->obd_recovery_expired) {
1988                 CDEBUG(D_HA, "waking for expired recovery\n");
1989                 wake_up = 1;
1990         }
1991         spin_unlock(&obd->obd_recovery_task_lock);
1992
1993         return wake_up;
1994 }
1995
1996 /**
1997  * wait for recovery events,
1998  * check its status with help of check_routine
1999  * evict dead clients via health_check
2000  */
2001 static int target_recovery_overseer(struct lu_target *lut,
2002                                     int (*check_routine)(struct lu_target *),
2003                                     int (*health_check)(struct obd_export *))
2004 {
2005         struct obd_device       *obd = lut->lut_obd;
2006         struct target_distribute_txn_data *tdtd;
2007 repeat:
2008         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
2009               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
2010                 __u64 next_update_transno = 0;
2011
2012                 /* Only abort the recovery if there are no update recovery
2013                  * left in the queue */
2014                 spin_lock(&obd->obd_recovery_task_lock);
2015                 if (lut->lut_tdtd != NULL) {
2016                         next_update_transno =
2017                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2018
2019                         tdtd = lut->lut_tdtd;
2020                         /* If next_update_transno == 0, it probably because
2021                          * updatelog retrieve threads did not get any records
2022                          * yet, let's wait those threads stopped */
2023                         if (next_update_transno == 0) {
2024                                 struct l_wait_info lwi = { 0 };
2025
2026                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2027                                        atomic_read(
2028                                        &tdtd->tdtd_recovery_threads_count) == 0,
2029                                        &lwi);
2030
2031                                 next_update_transno =
2032                                         distribute_txn_get_next_transno(
2033                                                                 lut->lut_tdtd);
2034                         }
2035                 }
2036
2037                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2038                         obd->obd_next_recovery_transno = next_update_transno;
2039                         spin_unlock(&obd->obd_recovery_task_lock);
2040                         /* Disconnect unfinished exports from clients, and
2041                          * keep connection from MDT to make sure the update
2042                          * recovery will still keep trying until some one
2043                          * manually abort the recovery */
2044                         class_disconnect_stale_exports(obd,
2045                                                 exp_finished_or_from_mdt);
2046                         /* Abort all of replay and replay lock req from
2047                          * clients */
2048                         abort_req_replay_queue(obd);
2049                         abort_lock_replay_queue(obd);
2050                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2051                                ")in the queue.\n", obd->obd_name,
2052                                next_update_transno);
2053                 } else {
2054                         obd->obd_abort_recovery = 1;
2055                         spin_unlock(&obd->obd_recovery_task_lock);
2056                         CWARN("%s recovery is aborted by hard timeout\n",
2057                               obd->obd_name);
2058                 }
2059         }
2060
2061         while (wait_event_timeout(obd->obd_next_transno_waitq,
2062                                   check_routine(lut),
2063                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2064                 /* wait indefinitely for event, but don't trigger watchdog */;
2065
2066         if (obd->obd_abort_recovery) {
2067                 CWARN("recovery is aborted, evict exports in recovery\n");
2068                 if (lut->lut_tdtd != NULL) {
2069                         struct l_wait_info lwi = { 0 };
2070
2071                         tdtd = lut->lut_tdtd;
2072                         /* Let's wait all of the update log recovery thread
2073                          * finished */
2074                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2075                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2076                              &lwi);
2077                         /* Then abort the update recovery list */
2078                         dtrq_list_destroy(lut->lut_tdtd);
2079                 }
2080
2081                 /** evict exports which didn't finish recovery yet */
2082                 class_disconnect_stale_exports(obd, exp_finished);
2083                 return 1;
2084         } else if (obd->obd_recovery_expired) {
2085                 obd->obd_recovery_expired = 0;
2086                 /** If some clients died being recovered, evict them */
2087                 LCONSOLE_WARN("%s: recovery is timed out, "
2088                               "evict stale exports\n", obd->obd_name);
2089                 /** evict cexports with no replay in queue, they are stalled */
2090                 class_disconnect_stale_exports(obd, health_check);
2091
2092                 /** continue with VBR */
2093                 spin_lock(&obd->obd_dev_lock);
2094                 obd->obd_version_recov = 1;
2095                 spin_unlock(&obd->obd_dev_lock);
2096                 /**
2097                  * reset timer, recovery will proceed with versions now,
2098                  * timeout is set just to handle reconnection delays
2099                  */
2100                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2101                 /** Wait for recovery events again, after evicting bad clients */
2102                 goto repeat;
2103         }
2104         return 0;
2105 }
2106
2107 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2108 {
2109         struct obd_device       *obd = lut->lut_obd;
2110         struct ptlrpc_request *req = NULL;
2111
2112         CDEBUG(D_HA, "Waiting for lock\n");
2113         if (target_recovery_overseer(lut, check_for_next_lock,
2114                                      exp_lock_replay_healthy))
2115                 abort_lock_replay_queue(obd);
2116
2117         spin_lock(&obd->obd_recovery_task_lock);
2118         if (!list_empty(&obd->obd_lock_replay_queue)) {
2119                 req = list_entry(obd->obd_lock_replay_queue.next,
2120                                      struct ptlrpc_request, rq_list);
2121                 list_del_init(&req->rq_list);
2122                 spin_unlock(&obd->obd_recovery_task_lock);
2123         } else {
2124                 spin_unlock(&obd->obd_recovery_task_lock);
2125                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2126                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2127                 /** evict exports failed VBR */
2128                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2129         }
2130         return req;
2131 }
2132
2133 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2134 {
2135         struct ptlrpc_request *req = NULL;
2136
2137         spin_lock(&obd->obd_recovery_task_lock);
2138         if (!list_empty(&obd->obd_final_req_queue)) {
2139                 req = list_entry(obd->obd_final_req_queue.next,
2140                                      struct ptlrpc_request, rq_list);
2141                 list_del_init(&req->rq_list);
2142                 spin_unlock(&obd->obd_recovery_task_lock);
2143                 if (req->rq_export->exp_in_recovery) {
2144                         spin_lock(&req->rq_export->exp_lock);
2145                         req->rq_export->exp_in_recovery = 0;
2146                         spin_unlock(&req->rq_export->exp_lock);
2147                 }
2148         } else {
2149                 spin_unlock(&obd->obd_recovery_task_lock);
2150         }
2151         return req;
2152 }
2153
2154 static void handle_recovery_req(struct ptlrpc_thread *thread,
2155                                 struct ptlrpc_request *req,
2156                                 svc_handler_t handler)
2157 {
2158         ENTRY;
2159
2160         /**
2161          * export can be evicted during recovery, no need to handle replays for
2162          * it after that, discard such request silently
2163          */
2164         if (req->rq_export->exp_disconnected)
2165                 RETURN_EXIT;
2166
2167         req->rq_session.lc_thread = thread;
2168         req->rq_svc_thread = thread;
2169         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2170
2171         /* thread context */
2172         lu_context_enter(&thread->t_env->le_ctx);
2173         (void)handler(req);
2174         lu_context_exit(&thread->t_env->le_ctx);
2175
2176         /* don't reset timer for final stage */
2177         if (!exp_finished(req->rq_export)) {
2178                 int to = obd_timeout;
2179
2180                 /**
2181                  * Add request timeout to the recovery time so next request from
2182                  * this client may come in recovery time
2183                  */
2184                 if (!AT_OFF) {
2185                         struct ptlrpc_service_part *svcpt;
2186
2187                         svcpt = req->rq_rqbd->rqbd_svcpt;
2188                         /* If the server sent early reply for this request,
2189                          * the client will recalculate the timeout according to
2190                          * current server estimate service time, so we will
2191                          * use the maxium timeout here for waiting the client
2192                          * sending the next req */
2193                         to = max((int)at_est2timeout(
2194                                  at_get(&svcpt->scp_at_estimate)),
2195                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2196                         /* Add 2 net_latency, one for balance rq_deadline
2197                          * (see ptl_send_rpc), one for resend the req to server,
2198                          * Note: client will pack net_latency in replay req
2199                          * (see ptlrpc_replay_req) */
2200                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2201                 }
2202                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2203         }
2204         EXIT;
2205 }
2206
2207 /** Checking routines for recovery */
2208 static int check_for_recovery_ready(struct lu_target *lut)
2209 {
2210         struct obd_device *obd = lut->lut_obd;
2211         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2212
2213         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2214                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2215                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2216                obd->obd_recovery_expired);
2217
2218         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2219                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2220                 if (clnts + obd->obd_stale_clients <
2221                     obd->obd_max_recoverable_clients)
2222                         return 0;
2223         }
2224
2225         if (lut->lut_tdtd != NULL) {
2226                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2227                     !obd->obd_abort_recovery) {
2228                         /* Let's extend recovery timer, in case the recovery
2229                          * timer expired, and some clients got evicted */
2230                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2231                                               true);
2232                         CDEBUG(D_HA, "%s update recovery is not ready,"
2233                                " extend recovery %d\n", obd->obd_name,
2234                                obd->obd_recovery_timeout);
2235                         return 0;
2236                 }
2237         }
2238
2239         return 1;
2240 }
2241
2242 enum {
2243         REQUEST_RECOVERY = 1,
2244         UPDATE_RECOVERY = 2,
2245 };
2246
2247 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2248 {
2249         __u64 transno = 0;
2250
2251         if (!list_empty(&obd->obd_req_replay_queue)) {
2252                 struct ptlrpc_request *req;
2253
2254                 req = list_entry(obd->obd_req_replay_queue.next,
2255                                  struct ptlrpc_request, rq_list);
2256                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2257         }
2258
2259         return transno;
2260 }
2261
2262 static __u64 get_next_transno(struct lu_target *lut, int *type)
2263 {
2264         struct obd_device *obd = lut->lut_obd;
2265         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2266         __u64 transno = 0;
2267         __u64 update_transno;
2268         ENTRY;
2269
2270         transno = get_next_replay_req_transno(obd);
2271         if (type != NULL)
2272                 *type = REQUEST_RECOVERY;
2273
2274         if (tdtd == NULL)
2275                 RETURN(transno);
2276
2277         update_transno = distribute_txn_get_next_transno(tdtd);
2278         if (transno == 0 || (transno >= update_transno &&
2279                              update_transno != 0)) {
2280                 transno = update_transno;
2281                 if (type != NULL)
2282                         *type = UPDATE_RECOVERY;
2283         }
2284
2285         RETURN(transno);
2286 }
2287
2288 /**
2289  * drop duplicate replay request
2290  *
2291  * Because the operation has been replayed by update recovery, the request
2292  * with the same transno will be dropped and also notify the client to send
2293  * next replay request.
2294  *
2295  * \param[in] env       execution environment
2296  * \param[in] obd       failover obd device
2297  * \param[in] req       request to be dropped
2298  */
2299 static void drop_duplicate_replay_req(struct lu_env *env,
2300                                       struct obd_device *obd,
2301                                       struct ptlrpc_request *req)
2302 {
2303         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2304                   " update records are found.\n",
2305                   lustre_msg_get_transno(req->rq_reqmsg),
2306                   libcfs_nid2str(req->rq_peer.nid));
2307
2308         /* Right now, only for MDS reint operation update replay and
2309          * normal request replay can have the same transno */
2310         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2311                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2312                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2313                 if (likely(req->rq_export))
2314                         target_committed_to_req(req);
2315                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2316                 target_send_reply(req, req->rq_status, 0);
2317         } else {
2318                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2319                 libcfs_nid2str(req->rq_peer.nid));
2320         }
2321         target_exp_dequeue_req_replay(req);
2322         target_request_copy_put(req);
2323         obd->obd_replayed_requests++;
2324 }
2325
2326 static void replay_request_or_update(struct lu_env *env,
2327                                      struct lu_target *lut,
2328                                      struct target_recovery_data *trd,
2329                                      struct ptlrpc_thread *thread)
2330 {
2331         struct obd_device *obd = lut->lut_obd;
2332         struct ptlrpc_request *req = NULL;
2333         int                     type;
2334         __u64                   transno;
2335         ENTRY;
2336
2337         CDEBUG(D_HA, "Waiting for transno %lld\n",
2338                obd->obd_next_recovery_transno);
2339
2340         /* Replay all of request and update by transno */
2341         do {
2342                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2343
2344                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2345
2346                 /** It is needed to extend recovery window above
2347                  *  recovery_time_soft. Extending is possible only in the
2348                  *  end of recovery window (see more details in
2349                  *  handle_recovery_req()).
2350                  */
2351                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2352
2353                 if (target_recovery_overseer(lut, check_for_next_transno,
2354                                         exp_req_replay_healthy_or_from_mdt)) {
2355                         abort_req_replay_queue(obd);
2356                         abort_lock_replay_queue(obd);
2357                         goto abort;
2358                 }
2359
2360                 spin_lock(&obd->obd_recovery_task_lock);
2361                 transno = get_next_transno(lut, &type);
2362                 if (type == REQUEST_RECOVERY && transno != 0) {
2363                         /* Drop replay request from client side, if the
2364                          * replay has been executed by update with the
2365                          * same transno */
2366                         req = list_entry(obd->obd_req_replay_queue.next,
2367                                         struct ptlrpc_request, rq_list);
2368
2369                         list_del_init(&req->rq_list);
2370                         obd->obd_requests_queued_for_recovery--;
2371                         spin_unlock(&obd->obd_recovery_task_lock);
2372
2373                         /* Let's check if the request has been redone by
2374                          * update replay */
2375                         if (is_req_replayed_by_update(req)) {
2376                                 struct distribute_txn_replay_req *dtrq;
2377
2378                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2379                                                                    req->rq_xid);
2380                                 LASSERT(dtrq != NULL);
2381                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2382                                 list_del_init(&dtrq->dtrq_list);
2383                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2384                                 dtrq_destroy(dtrq);
2385
2386                                 drop_duplicate_replay_req(env, obd, req);
2387
2388                                 continue;
2389                         }
2390
2391                         LASSERT(trd->trd_processing_task == current_pid());
2392                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2393                                   lustre_msg_get_transno(req->rq_reqmsg),
2394                                   libcfs_nid2str(req->rq_peer.nid));
2395
2396                         handle_recovery_req(thread, req,
2397                                             trd->trd_recovery_handler);
2398                         /**
2399                          * bz18031: increase next_recovery_transno before
2400                          * target_request_copy_put() will drop exp_rpc reference
2401                          */
2402                         spin_lock(&obd->obd_recovery_task_lock);
2403                         obd->obd_next_recovery_transno++;
2404                         spin_unlock(&obd->obd_recovery_task_lock);
2405                         target_exp_dequeue_req_replay(req);
2406                         target_request_copy_put(req);
2407                         obd->obd_replayed_requests++;
2408                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2409                         struct distribute_txn_replay_req *dtrq;
2410                         int rc;
2411
2412                         spin_unlock(&obd->obd_recovery_task_lock);
2413
2414                         LASSERT(tdtd != NULL);
2415                         dtrq = distribute_txn_get_next_req(tdtd);
2416                         lu_context_enter(&thread->t_env->le_ctx);
2417                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2418                         lu_context_exit(&thread->t_env->le_ctx);
2419                         extend_recovery_timer(obd, obd_timeout, true);
2420
2421                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2422                                 CDEBUG(D_HA, "Move x%llu t%llu"
2423                                        " to finish list\n", dtrq->dtrq_xid,
2424                                        dtrq->dtrq_master_transno);
2425
2426                                 /* Add it to the replay finish list */
2427                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2428                                 list_add(&dtrq->dtrq_list,
2429                                          &tdtd->tdtd_replay_finish_list);
2430                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2431
2432                                 spin_lock(&obd->obd_recovery_task_lock);
2433                                 if (transno == obd->obd_next_recovery_transno)
2434                                         obd->obd_next_recovery_transno++;
2435                                 else if (transno >
2436                                          obd->obd_next_recovery_transno)
2437                                         obd->obd_next_recovery_transno =
2438                                                                 transno + 1;
2439                                 spin_unlock(&obd->obd_recovery_task_lock);
2440                         } else {
2441                                 dtrq_destroy(dtrq);
2442                         }
2443                 } else {
2444                         spin_unlock(&obd->obd_recovery_task_lock);
2445 abort:
2446                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2447                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2448                         /** evict exports failed VBR */
2449                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2450                         break;
2451                 }
2452         } while (1);
2453 }
2454
2455 static int target_recovery_thread(void *arg)
2456 {
2457         struct lu_target *lut = arg;
2458         struct obd_device *obd = lut->lut_obd;
2459         struct ptlrpc_request *req;
2460         struct target_recovery_data *trd = &obd->obd_recovery_data;
2461         unsigned long delta;
2462         struct lu_env *env;
2463         struct ptlrpc_thread *thread = NULL;
2464         int rc = 0;
2465         ENTRY;
2466
2467         unshare_fs_struct();
2468         OBD_ALLOC_PTR(thread);
2469         if (thread == NULL)
2470                 RETURN(-ENOMEM);
2471
2472         OBD_ALLOC_PTR(env);
2473         if (env == NULL) {
2474                 OBD_FREE_PTR(thread);
2475                 RETURN(-ENOMEM);
2476         }
2477
2478         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2479         if (rc) {
2480                 OBD_FREE_PTR(thread);
2481                 OBD_FREE_PTR(env);
2482                 RETURN(rc);
2483         }
2484
2485         thread->t_env = env;
2486         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2487         env->le_ctx.lc_thread = thread;
2488         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2489         thread->t_watchdog = NULL;
2490
2491         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2492                current_pid());
2493         trd->trd_processing_task = current_pid();
2494
2495         spin_lock(&obd->obd_dev_lock);
2496         obd->obd_recovering = 1;
2497         spin_unlock(&obd->obd_dev_lock);
2498         complete(&trd->trd_starting);
2499
2500         /* first of all, we have to know the first transno to replay */
2501         if (target_recovery_overseer(lut, check_for_recovery_ready,
2502                                      exp_connect_healthy)) {
2503                 abort_req_replay_queue(obd);
2504                 abort_lock_replay_queue(obd);
2505                 if (lut->lut_tdtd != NULL)
2506                         dtrq_list_destroy(lut->lut_tdtd);
2507         }
2508
2509         /* next stage: replay requests or update */
2510         delta = jiffies;
2511         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2512                atomic_read(&obd->obd_req_replay_clients),
2513                obd->obd_next_recovery_transno);
2514         replay_request_or_update(env, lut, trd, thread);
2515
2516         /**
2517          * The second stage: replay locks
2518          */
2519         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2520                atomic_read(&obd->obd_lock_replay_clients));
2521         while ((req = target_next_replay_lock(lut))) {
2522                 LASSERT(trd->trd_processing_task == current_pid());
2523                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2524                           libcfs_nid2str(req->rq_peer.nid));
2525                 handle_recovery_req(thread, req,
2526                                     trd->trd_recovery_handler);
2527                 target_request_copy_put(req);
2528                 obd->obd_replayed_locks++;
2529         }
2530
2531         /**
2532          * The third stage: reply on final pings, at this moment all clients
2533          * must have request in final queue
2534          */
2535         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2536         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2537         /** Update server last boot epoch */
2538         tgt_boot_epoch_update(lut);
2539         /* We drop recoverying flag to forward all new requests
2540          * to regular mds_handle() since now */
2541         spin_lock(&obd->obd_dev_lock);
2542         obd->obd_recovering = obd->obd_abort_recovery = 0;
2543         spin_unlock(&obd->obd_dev_lock);
2544         spin_lock(&obd->obd_recovery_task_lock);
2545         target_cancel_recovery_timer(obd);
2546         spin_unlock(&obd->obd_recovery_task_lock);
2547         while ((req = target_next_final_ping(obd))) {
2548                 LASSERT(trd->trd_processing_task == current_pid());
2549                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2550                           libcfs_nid2str(req->rq_peer.nid));
2551                 handle_recovery_req(thread, req,
2552                                     trd->trd_recovery_handler);
2553                 /* Because the waiting client can not send ping to server,
2554                  * so we need refresh the last_request_time, to avoid the
2555                  * export is being evicted */
2556                 ptlrpc_update_export_timer(req->rq_export, 0);
2557                 target_request_copy_put(req);
2558         }
2559
2560         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2561         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2562               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2563         if (delta > OBD_RECOVERY_TIME_SOFT) {
2564                 CWARN("too long recovery - read logs\n");
2565                 libcfs_debug_dumplog();
2566         }
2567
2568         target_finish_recovery(lut);
2569
2570         lu_context_fini(&env->le_ctx);
2571         trd->trd_processing_task = 0;
2572         complete(&trd->trd_finishing);
2573
2574         tgt_io_thread_done(thread);
2575         OBD_FREE_PTR(thread);
2576         OBD_FREE_PTR(env);
2577         RETURN(rc);
2578 }
2579
2580 static int target_start_recovery_thread(struct lu_target *lut,
2581                                         svc_handler_t handler)
2582 {
2583         struct obd_device *obd = lut->lut_obd;
2584         int rc = 0;
2585         struct target_recovery_data *trd = &obd->obd_recovery_data;
2586         int index;
2587
2588         memset(trd, 0, sizeof(*trd));
2589         init_completion(&trd->trd_starting);
2590         init_completion(&trd->trd_finishing);
2591         trd->trd_recovery_handler = handler;
2592
2593         rc = server_name2index(obd->obd_name, &index, NULL);
2594         if (rc < 0)
2595                 return rc;
2596
2597         if (!IS_ERR(kthread_run(target_recovery_thread,
2598                                 lut, "tgt_recover_%d", index))) {
2599                 wait_for_completion(&trd->trd_starting);
2600                 LASSERT(obd->obd_recovering != 0);
2601         } else {
2602                 rc = -ECHILD;
2603         }
2604
2605         return rc;
2606 }
2607
2608 void target_stop_recovery_thread(struct obd_device *obd)
2609 {
2610         if (obd->obd_recovery_data.trd_processing_task > 0) {
2611                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2612                 /** recovery can be done but postrecovery is not yet */
2613                 spin_lock(&obd->obd_dev_lock);
2614                 if (obd->obd_recovering) {
2615                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2616                         obd->obd_abort_recovery = 1;
2617                         wake_up(&obd->obd_next_transno_waitq);
2618                 }
2619                 spin_unlock(&obd->obd_dev_lock);
2620                 wait_for_completion(&trd->trd_finishing);
2621         }
2622 }
2623 EXPORT_SYMBOL(target_stop_recovery_thread);
2624
2625 void target_recovery_fini(struct obd_device *obd)
2626 {
2627         class_disconnect_exports(obd);
2628         target_stop_recovery_thread(obd);
2629         target_cleanup_recovery(obd);
2630 }
2631 EXPORT_SYMBOL(target_recovery_fini);
2632
2633 static void target_recovery_expired(unsigned long castmeharder)
2634 {
2635         struct obd_device *obd = (struct obd_device *)castmeharder;
2636         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2637                " after %lds (%d clients connected)\n",
2638                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2639                cfs_time_current_sec()- obd->obd_recovery_start,
2640                atomic_read(&obd->obd_connected_clients));
2641
2642         obd->obd_recovery_expired = 1;
2643         wake_up(&obd->obd_next_transno_waitq);
2644 }
2645
2646 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2647 {
2648         struct obd_device *obd = lut->lut_obd;
2649
2650         if (obd->obd_max_recoverable_clients == 0) {
2651                 /** Update server last boot epoch */
2652                 tgt_boot_epoch_update(lut);
2653                 return;
2654         }
2655
2656         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2657                "last_transno %llu\n", obd->obd_name,
2658                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2659         LASSERT(obd->obd_stopping == 0);
2660         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2661         obd->obd_recovery_start = 0;
2662         obd->obd_recovery_end = 0;
2663
2664         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2665         target_start_recovery_thread(lut, handler);
2666 }
2667 EXPORT_SYMBOL(target_recovery_init);
2668
2669 static int target_process_req_flags(struct obd_device *obd,
2670                                     struct ptlrpc_request *req)
2671 {
2672         struct obd_export *exp = req->rq_export;
2673         LASSERT(exp != NULL);
2674         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2675                 /* client declares he's ready to replay locks */
2676                 spin_lock(&exp->exp_lock);
2677                 if (exp->exp_req_replay_needed) {
2678                         exp->exp_req_replay_needed = 0;
2679                         spin_unlock(&exp->exp_lock);
2680
2681                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2682                         atomic_dec(&obd->obd_req_replay_clients);
2683                 } else {
2684                         spin_unlock(&exp->exp_lock);
2685                 }
2686         }
2687         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2688                 /* client declares he's ready to complete recovery
2689                  * so, we put the request on th final queue */
2690                 spin_lock(&exp->exp_lock);
2691                 if (exp->exp_lock_replay_needed) {
2692                         exp->exp_lock_replay_needed = 0;
2693                         spin_unlock(&exp->exp_lock);
2694
2695                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2696                         atomic_dec(&obd->obd_lock_replay_clients);
2697                 } else {
2698                         spin_unlock(&exp->exp_lock);
2699                 }
2700         }
2701         return 0;
2702 }
2703
2704 int target_queue_recovery_request(struct ptlrpc_request *req,
2705                                   struct obd_device *obd)
2706 {
2707         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2708         struct ptlrpc_request *reqiter;
2709         int inserted = 0;
2710         ENTRY;
2711
2712         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2713                 /* Processing the queue right now, don't re-add. */
2714                 RETURN(1);
2715         }
2716
2717         target_process_req_flags(obd, req);
2718
2719         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2720                 /* client declares he's ready to complete recovery
2721                  * so, we put the request on th final queue */
2722                 target_request_copy_get(req);
2723                 DEBUG_REQ(D_HA, req, "queue final req");
2724                 wake_up(&obd->obd_next_transno_waitq);
2725                 spin_lock(&obd->obd_recovery_task_lock);
2726                 if (obd->obd_recovering) {
2727                         struct ptlrpc_request *tmp;
2728                         struct ptlrpc_request *duplicate = NULL;
2729
2730                         if (likely(!req->rq_export->exp_replay_done)) {
2731                                 req->rq_export->exp_replay_done = 1;
2732                                 list_add_tail(&req->rq_list,
2733                                               &obd->obd_final_req_queue);
2734                                 spin_unlock(&obd->obd_recovery_task_lock);
2735                                 RETURN(0);
2736                         }
2737
2738                         /* XXX O(n), but only happens if final ping is
2739                          * timed out, probably reorganize the list as
2740                          * a hash list later */
2741                         list_for_each_entry_safe(reqiter, tmp,
2742                                                  &obd->obd_final_req_queue,
2743                                                  rq_list) {
2744                                 if (reqiter->rq_export == req->rq_export) {
2745                                         list_del_init(&reqiter->rq_list);
2746                                         duplicate = reqiter;
2747                                         break;
2748                                 }
2749                         }
2750
2751                         list_add_tail(&req->rq_list,
2752                                       &obd->obd_final_req_queue);
2753                         req->rq_export->exp_replay_done = 1;
2754                         spin_unlock(&obd->obd_recovery_task_lock);
2755
2756                         if (duplicate != NULL) {
2757                                 DEBUG_REQ(D_HA, duplicate,
2758                                           "put prev final req\n");
2759                                 target_request_copy_put(duplicate);
2760                         }
2761                         RETURN(0);
2762                 } else {
2763                         spin_unlock(&obd->obd_recovery_task_lock);
2764                         target_request_copy_put(req);
2765                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2766                 }
2767         }
2768         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2769                 /* client declares he's ready to replay locks */
2770                 target_request_copy_get(req);
2771                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2772                 wake_up(&obd->obd_next_transno_waitq);
2773                 spin_lock(&obd->obd_recovery_task_lock);
2774                 LASSERT(obd->obd_recovering);
2775                 /* usually due to recovery abort */
2776                 if (!req->rq_export->exp_in_recovery) {
2777                         spin_unlock(&obd->obd_recovery_task_lock);
2778                         target_request_copy_put(req);
2779                         RETURN(-ENOTCONN);
2780                 }
2781                 LASSERT(req->rq_export->exp_lock_replay_needed);
2782                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2783                 spin_unlock(&obd->obd_recovery_task_lock);
2784                 RETURN(0);
2785         }
2786
2787         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2788          * (i.e. buflens etc are in my own byte order), but type-dependent
2789          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2790
2791         if (!transno) {
2792                 INIT_LIST_HEAD(&req->rq_list);
2793                 DEBUG_REQ(D_HA, req, "not queueing");
2794                 RETURN(1);
2795         }
2796
2797         /* If we're processing the queue, we want don't want to queue this
2798          * message.
2799          *
2800          * Also, if this request has a transno less than the one we're waiting
2801          * for, we should process it now.  It could (and currently always will)
2802          * be an open request for a descriptor that was opened some time ago.
2803          *
2804          * Also, a resent, replayed request that has already been
2805          * handled will pass through here and be processed immediately.
2806          */
2807         CDEBUG(D_HA, "Next recovery transno: %llu"
2808                ", current: %llu, replaying\n",
2809                obd->obd_next_recovery_transno, transno);
2810
2811         /* If the request has been replayed by update replay, then sends this
2812          * request to the recovery thread (replay_request_or_update()), where
2813          * it will be handled */
2814         spin_lock(&obd->obd_recovery_task_lock);
2815         if (transno < obd->obd_next_recovery_transno &&
2816             !is_req_replayed_by_update(req)) {
2817                 /* Processing the queue right now, don't re-add. */
2818                 LASSERT(list_empty(&req->rq_list));
2819                 spin_unlock(&obd->obd_recovery_task_lock);
2820                 RETURN(1);
2821         }
2822         spin_unlock(&obd->obd_recovery_task_lock);
2823
2824         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2825                 RETURN(0);
2826
2827         target_request_copy_get(req);
2828         if (!req->rq_export->exp_in_recovery) {
2829                 target_request_copy_put(req);
2830                 RETURN(-ENOTCONN);
2831         }
2832         LASSERT(req->rq_export->exp_req_replay_needed);
2833
2834         if (target_exp_enqueue_req_replay(req)) {
2835                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2836                 target_request_copy_put(req);
2837                 RETURN(0);
2838         }
2839
2840         /* XXX O(n^2) */
2841         spin_lock(&obd->obd_recovery_task_lock);
2842         LASSERT(obd->obd_recovering);
2843         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2844                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2845                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2846                         inserted = 1;
2847                         goto added;
2848                 }
2849
2850                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2851                              transno)) {
2852                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2853                                   "has been claimed by another client");
2854                         spin_unlock(&obd->obd_recovery_task_lock);
2855                         target_exp_dequeue_req_replay(req);
2856                         target_request_copy_put(req);
2857                         RETURN(0);
2858                 }
2859         }
2860 added:
2861         if (!inserted)
2862                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2863
2864         obd->obd_requests_queued_for_recovery++;
2865         spin_unlock(&obd->obd_recovery_task_lock);
2866         wake_up(&obd->obd_next_transno_waitq);
2867         RETURN(0);
2868 }
2869
2870 int target_handle_ping(struct ptlrpc_request *req)
2871 {
2872         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2873         return req_capsule_server_pack(&req->rq_pill);
2874 }
2875
2876 void target_committed_to_req(struct ptlrpc_request *req)
2877 {
2878         struct obd_export *exp = req->rq_export;
2879
2880         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2881                 lustre_msg_set_last_committed(req->rq_repmsg,
2882                                               exp->exp_last_committed);
2883         else
2884                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2885                           "%d)", exp->exp_obd->obd_no_transno,
2886                           req->rq_repmsg == NULL);
2887
2888         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2889                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2890 }
2891
2892 #endif /* HAVE_SERVER_SUPPORT */
2893
2894 /**
2895  * Packs current SLV and Limit into \a req.
2896  */
2897 int target_pack_pool_reply(struct ptlrpc_request *req)
2898 {
2899         struct obd_device *obd;
2900         ENTRY;
2901
2902         /* Check that we still have all structures alive as this may
2903          * be some late RPC at shutdown time. */
2904         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2905                      !exp_connect_lru_resize(req->rq_export))) {
2906                 lustre_msg_set_slv(req->rq_repmsg, 0);
2907                 lustre_msg_set_limit(req->rq_repmsg, 0);
2908                 RETURN(0);
2909         }
2910
2911         /* OBD is alive here as export is alive, which we checked above. */
2912         obd = req->rq_export->exp_obd;
2913
2914         read_lock(&obd->obd_pool_lock);
2915         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2916         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2917         read_unlock(&obd->obd_pool_lock);
2918
2919         RETURN(0);
2920 }
2921
2922 static int target_send_reply_msg(struct ptlrpc_request *req,
2923                                  int rc, int fail_id)
2924 {
2925         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2926                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2927                 return -ECOMM;
2928         }
2929         /* We can have a null rq_reqmsg in the event of bad signature or
2930          * no context when unwrapping */
2931         if (req->rq_reqmsg &&
2932             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2933             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2934                 return -ECOMM;
2935
2936         if (unlikely(rc)) {
2937                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2938                 req->rq_status = rc;
2939                 return ptlrpc_send_error(req, 1);
2940         } else {
2941                 DEBUG_REQ(D_NET, req, "sending reply");
2942         }
2943
2944         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2945 }
2946
2947 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2948 {
2949         struct ptlrpc_service_part *svcpt;
2950         int                        netrc;
2951         struct ptlrpc_reply_state *rs;
2952         struct obd_export         *exp;
2953         ENTRY;
2954
2955         if (req->rq_no_reply) {
2956                 EXIT;
2957                 return;
2958         }
2959
2960         svcpt = req->rq_rqbd->rqbd_svcpt;
2961         rs = req->rq_reply_state;
2962         if (rs == NULL || !rs->rs_difficult) {
2963                 /* no notifiers */
2964                 target_send_reply_msg (req, rc, fail_id);
2965                 EXIT;
2966                 return;
2967         }
2968
2969         /* must be an export if locks saved */
2970         LASSERT(req->rq_export != NULL);
2971         /* req/reply consistent */
2972         LASSERT(rs->rs_svcpt == svcpt);
2973
2974         /* "fresh" reply */
2975         LASSERT(!rs->rs_scheduled);
2976         LASSERT(!rs->rs_scheduled_ever);
2977         LASSERT(!rs->rs_handled);
2978         LASSERT(!rs->rs_on_net);
2979         LASSERT(rs->rs_export == NULL);
2980         LASSERT(list_empty(&rs->rs_obd_list));
2981         LASSERT(list_empty(&rs->rs_exp_list));
2982
2983         exp = class_export_get(req->rq_export);
2984
2985         /* disable reply scheduling while I'm setting up */
2986         rs->rs_scheduled = 1;
2987         rs->rs_on_net    = 1;
2988         rs->rs_xid       = req->rq_xid;
2989         rs->rs_transno   = req->rq_transno;
2990         rs->rs_export    = exp;
2991         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2992
2993         spin_lock(&exp->exp_uncommitted_replies_lock);
2994         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2995                rs->rs_transno, exp->exp_last_committed);
2996         if (rs->rs_transno > exp->exp_last_committed) {
2997                 /* not committed already */
2998                 list_add_tail(&rs->rs_obd_list,
2999                                   &exp->exp_uncommitted_replies);
3000         }
3001         spin_unlock(&exp->exp_uncommitted_replies_lock);
3002
3003         spin_lock(&exp->exp_lock);
3004         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3005         spin_unlock(&exp->exp_lock);
3006
3007         netrc = target_send_reply_msg(req, rc, fail_id);
3008
3009         spin_lock(&svcpt->scp_rep_lock);
3010
3011         atomic_inc(&svcpt->scp_nreps_difficult);
3012
3013         if (netrc != 0) {
3014                 /* error sending: reply is off the net.  Also we need +1
3015                  * reply ref until ptlrpc_handle_rs() is done
3016                  * with the reply state (if the send was successful, there
3017                  * would have been +1 ref for the net, which
3018                  * reply_out_callback leaves alone) */
3019                 rs->rs_on_net = 0;
3020                 ptlrpc_rs_addref(rs);
3021         }
3022
3023         spin_lock(&rs->rs_lock);
3024         if (rs->rs_transno <= exp->exp_last_committed ||
3025             (!rs->rs_on_net && !rs->rs_no_ack) ||
3026             list_empty(&rs->rs_exp_list) ||     /* completed already */
3027             list_empty(&rs->rs_obd_list)) {
3028                 CDEBUG(D_HA, "Schedule reply immediately\n");
3029                 ptlrpc_dispatch_difficult_reply(rs);
3030         } else {
3031                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3032                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3033         }
3034         spin_unlock(&rs->rs_lock);
3035         spin_unlock(&svcpt->scp_rep_lock);
3036         EXIT;
3037 }
3038
3039 enum ldlm_mode lck_compat_array[] = {
3040         [LCK_EX]    = LCK_COMPAT_EX,
3041         [LCK_PW]    = LCK_COMPAT_PW,
3042         [LCK_PR]    = LCK_COMPAT_PR,
3043         [LCK_CW]    = LCK_COMPAT_CW,
3044         [LCK_CR]    = LCK_COMPAT_CR,
3045         [LCK_NL]    = LCK_COMPAT_NL,
3046         [LCK_GROUP] = LCK_COMPAT_GROUP,
3047         [LCK_COS]   = LCK_COMPAT_COS,
3048 };
3049
3050 /**
3051  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3052  * not escape to the user level.
3053  */
3054 int ldlm_error2errno(enum ldlm_error error)
3055 {
3056         int result;
3057
3058         switch (error) {
3059         case ELDLM_OK:
3060         case ELDLM_LOCK_MATCHED:
3061                 result = 0;
3062                 break;
3063         case ELDLM_LOCK_CHANGED:
3064                 result = -ESTALE;
3065                 break;
3066         case ELDLM_LOCK_ABORTED:
3067                 result = -ENAVAIL;
3068                 break;
3069         case ELDLM_LOCK_REPLACED:
3070                 result = -ESRCH;
3071                 break;
3072         case ELDLM_NO_LOCK_DATA:
3073                 result = -ENOENT;
3074                 break;
3075         case ELDLM_NAMESPACE_EXISTS:
3076                 result = -EEXIST;
3077                 break;
3078         case ELDLM_BAD_NAMESPACE:
3079                 result = -EBADF;
3080                 break;
3081         default:
3082                 if (((int)error) < 0) { /* cast to signed type */
3083                         result = error; /* as ldlm_error can be unsigned */
3084                 } else {
3085                         CERROR("Invalid DLM result code: %d\n", error);
3086                         result = -EPROTO;
3087                 }
3088         }
3089         return result;
3090 }
3091 EXPORT_SYMBOL(ldlm_error2errno);
3092
3093 /**
3094  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3095  */
3096 enum ldlm_error ldlm_errno2error(int err_no)
3097 {
3098         int error;
3099
3100         switch (err_no) {
3101         case 0:
3102                 error = ELDLM_OK;
3103                 break;
3104         case -ESTALE:
3105                 error = ELDLM_LOCK_CHANGED;
3106                 break;
3107         case -ENAVAIL:
3108                 error = ELDLM_LOCK_ABORTED;
3109                 break;
3110         case -ESRCH:
3111                 error = ELDLM_LOCK_REPLACED;
3112                 break;
3113         case -ENOENT:
3114                 error = ELDLM_NO_LOCK_DATA;
3115                 break;
3116         case -EEXIST:
3117                 error = ELDLM_NAMESPACE_EXISTS;
3118                 break;
3119         case -EBADF:
3120                 error = ELDLM_BAD_NAMESPACE;
3121                 break;
3122         default:
3123                 error = err_no;
3124         }
3125         return error;
3126 }
3127
3128 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3129 void ldlm_dump_export_locks(struct obd_export *exp)
3130 {
3131         spin_lock(&exp->exp_locks_list_guard);
3132         if (!list_empty(&exp->exp_locks_list)) {
3133                 struct ldlm_lock *lock;
3134
3135                 CERROR("dumping locks for export %p,"
3136                        "ignore if the unmount doesn't hang\n", exp);
3137                 list_for_each_entry(lock, &exp->exp_locks_list,
3138                                         l_exp_refs_link)
3139                         LDLM_ERROR(lock, "lock:");
3140         }
3141         spin_unlock(&exp->exp_locks_list_guard);
3142 }
3143 #endif
3144
3145 #ifdef HAVE_SERVER_SUPPORT
3146 static int target_bulk_timeout(void *data)
3147 {
3148         ENTRY;
3149         /* We don't fail the connection here, because having the export
3150          * killed makes the (vital) call to commitrw very sad.
3151          */
3152         RETURN(1);
3153 }
3154
3155 static inline const char *bulk2type(struct ptlrpc_request *req)
3156 {
3157         if (req->rq_bulk_read)
3158                 return "READ";
3159         if (req->rq_bulk_write)
3160                 return "WRITE";
3161         return "UNKNOWN";
3162 }
3163
3164 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3165                    struct l_wait_info *lwi)
3166 {
3167         struct ptlrpc_request   *req = desc->bd_req;
3168         time_t                   start = cfs_time_current_sec();
3169         time_t                   deadline;
3170         int                      rc = 0;
3171
3172         ENTRY;
3173
3174         /* If there is eviction in progress, wait for it to finish. */
3175         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3176                 *lwi = LWI_INTR(NULL, NULL);
3177                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3178                                   !atomic_read(&exp->exp_obd->
3179                                                    obd_evict_inprogress),
3180                                   lwi);
3181         }
3182
3183         /* Check if client was evicted or reconnected already. */
3184         if (exp->exp_failed ||
3185             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3186                 rc = -ENOTCONN;
3187         } else {
3188                 if (req->rq_bulk_read)
3189                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3190
3191                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3192                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3193                 else /* old version, bulk matchbits is rq_xid */
3194                         req->rq_mbits = req->rq_xid;
3195
3196                 if (rc == 0)
3197                         rc = ptlrpc_start_bulk_transfer(desc);
3198         }
3199
3200         if (rc < 0) {
3201                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3202                           bulk2type(req), rc);
3203                 RETURN(rc);
3204         }
3205
3206         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3207                 ptlrpc_abort_bulk(desc);
3208                 RETURN(0);
3209         }
3210
3211         /* limit actual bulk transfer to bulk_timeout seconds */
3212         deadline = start + bulk_timeout;
3213         if (deadline > req->rq_deadline)
3214                 deadline = req->rq_deadline;
3215
3216         do {
3217                 long timeoutl = deadline - cfs_time_current_sec();
3218                 cfs_duration_t timeout = timeoutl <= 0 ?
3219                                          CFS_TICK : cfs_time_seconds(timeoutl);
3220                 time_t  rq_deadline;
3221
3222                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3223                                             target_bulk_timeout, desc);
3224                 rc = l_wait_event(desc->bd_waitq,
3225                                   !ptlrpc_server_bulk_active(desc) ||
3226                                   exp->exp_failed ||
3227                                   exp->exp_conn_cnt >
3228                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3229                                   lwi);
3230                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3231                 /* Wait again if we changed rq_deadline. */
3232                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3233                 deadline = start + bulk_timeout;
3234                 if (deadline > rq_deadline)
3235                         deadline = rq_deadline;
3236         } while ((rc == -ETIMEDOUT) &&
3237                  (deadline > cfs_time_current_sec()));
3238
3239         if (rc == -ETIMEDOUT) {
3240                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3241                           bulk2type(req), deadline - start,
3242                           cfs_time_current_sec() - deadline);
3243                 ptlrpc_abort_bulk(desc);
3244         } else if (exp->exp_failed) {
3245                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3246                           bulk2type(req));
3247                 rc = -ENOTCONN;
3248                 ptlrpc_abort_bulk(desc);
3249         } else if (exp->exp_conn_cnt >
3250                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3251                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3252                           bulk2type(req));
3253                 /* We don't reply anyway. */
3254                 rc = -ETIMEDOUT;
3255                 ptlrpc_abort_bulk(desc);
3256         } else if (desc->bd_failure) {
3257                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3258                           bulk2type(req));
3259                 /* XXX should this be a different errno? */
3260                 rc = -ETIMEDOUT;
3261         } else {
3262                 if (req->rq_bulk_write)
3263                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3264                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3265                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3266                                   bulk2type(req), desc->bd_nob_transferred,
3267                                   desc->bd_nob);
3268                         /* XXX should this be a different errno? */
3269                         rc = -ETIMEDOUT;
3270                 }
3271         }
3272
3273         RETURN(rc);
3274 }
3275 EXPORT_SYMBOL(target_bulk_io);
3276
3277 #endif /* HAVE_SERVER_SUPPORT */