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, 2016, 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 = target->obd_recovery_timer.expires;
 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 static void
 838 check_and_start_recovery_timer(struct obd_device *obd,
 839                                struct ptlrpc_request *req, int new_client);
 840
 841 /**
 842  * update flags for import during reconnect process
 843  */
 844 static int rev_import_flags_update(struct obd_import *revimp,
 845                                    struct ptlrpc_request *req)
 846 {
 847         int rc;
 848         struct obd_connect_data *data;
 849
 850         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 851
 852         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 853                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 854         else
 855                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 856
 857         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 858
 859         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 860         if (rc) {
 861                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 862                         revimp->imp_client->cli_name,
 863                         libcfs_id2str(req->rq_peer), rc);
 864                 return rc;
 865         }
 866
 867         return 0;
 868 }
 869
 870 /**
 871  * Allocate a new reverse import for an export.
 872  *
 873  * \retval -errno in case error hit
 874  * \retval 0 if reverse import correctly init
 875  **/
 876 int rev_import_init(struct obd_export *export)
 877 {
 878         struct obd_device *obd = export->exp_obd;
 879         struct obd_import *revimp;
 880
 881         LASSERT(export->exp_imp_reverse == NULL);
 882
 883         revimp = class_new_import(obd);
 884         if (revimp == NULL)
 885                 return -ENOMEM;
 886
 887         revimp->imp_remote_handle.cookie = 0ULL;
 888         revimp->imp_client = &obd->obd_ldlm_client;
 889         revimp->imp_dlm_fake = 1;
 890
 891         /* it is safe to connect import in new state as no sends possible */
 892         spin_lock(&export->exp_lock);
 893         export->exp_imp_reverse = revimp;
 894         spin_unlock(&export->exp_lock);
 895         class_import_put(revimp);
 896
 897         return 0;
 898 }
 899 EXPORT_SYMBOL(rev_import_init);
 900
 901 /**
 902  * Handle reconnect for an export.
 903  *
 904  * \param exp export to handle reconnect process
 905  * \param req client reconnect request
 906  *
 907  * \retval -rc in case securitfy flavor can't be changed
 908  * \retval 0 in case none problems
 909  */
 910 static int rev_import_reconnect(struct obd_export *exp,
 911                                 struct ptlrpc_request *req)
 912 {
 913         struct obd_import *revimp = exp->exp_imp_reverse;
 914         struct lustre_handle *lh;
 915         int rc;
 916
 917         /* avoid sending a request until import flags are changed */
 918         ptlrpc_import_enter_resend(revimp);
 919
 920         if (revimp->imp_connection != NULL)
 921                 ptlrpc_connection_put(revimp->imp_connection);
 922
 923         /*
 924          * client from recovery don't have a handle so we need to take from
 925          * request. it may produce situation when wrong client connected
 926          * to recovery as we trust a client uuid
 927          */
 928         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 929         revimp->imp_remote_handle = *lh;
 930
 931         /* unknown versions will be caught in
 932          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 933         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 934
 935         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 936
 937         rc = rev_import_flags_update(revimp, req);
 938         if (rc != 0) {
 939                 /* it is safe to still be in RECOVERY phase as we are not able
 940                  * to setup correct security flavor so requests are not able to
 941                  * be delivered correctly */
 942                 return rc;
 943         }
 944
 945         /* resend all rpc's via new connection */
 946         return ptlrpc_import_recovery_state_machine(revimp);
 947 }
 948
 949 int target_handle_connect(struct ptlrpc_request *req)
 950 {
 951         struct obd_device *target = NULL;
 952         struct obd_export *export = NULL;
 953         /* connect handle - filled from target_handle_reconnect in
 954          * reconnect case */
 955         struct lustre_handle conn;
 956         struct lustre_handle *tmp;
 957         struct obd_uuid tgtuuid;
 958         struct obd_uuid cluuid;
 959         char *str;
 960         int rc = 0;
 961         char *target_start;
 962         int target_len;
 963         bool     mds_conn = false, lw_client = false, initial_conn = false;
 964         bool     mds_mds_conn = false;
 965         bool     new_mds_mds_conn = false;
 966         bool     target_referenced = false;
 967         struct obd_connect_data *data, *tmpdata;
 968         int size, tmpsize;
 969         lnet_nid_t *client_nid = NULL;
 970         ENTRY;
 971
 972         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 973
 974         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 975         if (str == NULL) {
 976                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 977                 GOTO(out, rc = -EINVAL);
 978         }
 979
 980         obd_str2uuid(&tgtuuid, str);
 981         target = class_uuid2obd(&tgtuuid);
 982         if (!target)
 983                 target = class_name2obd(str);
 984
 985         if (!target) {
 986                 deuuidify(str, NULL, &target_start, &target_len);
 987                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 988                                    "from %s (no target). If you are running "
 989                                    "an HA pair check that the target is "
 990                                    "mounted on the other server.\n", str,
 991                                    libcfs_nid2str(req->rq_peer.nid));
 992                 GOTO(out, rc = -ENODEV);
 993         }
 994
 995         spin_lock(&target->obd_dev_lock);
 996         if (target->obd_stopping || !target->obd_set_up) {
 997                 spin_unlock(&target->obd_dev_lock);
 998
 999                 deuuidify(str, NULL, &target_start, &target_len);
1000                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1001                               target_len, target_start,
1002                               libcfs_nid2str(req->rq_peer.nid),
1003                               (target->obd_stopping ?
1004                                "stopping" : "not set up"));
1005                 GOTO(out, rc = -ENODEV);
1006         }
1007
1008         if (target->obd_no_conn) {
1009                 spin_unlock(&target->obd_dev_lock);
1010
1011                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1012                                "from %s\n", target->obd_name,
1013                                libcfs_nid2str(req->rq_peer.nid));
1014                 GOTO(out, rc = -EAGAIN);
1015         }
1016
1017         /* Make sure the target isn't cleaned up while we're here. Yes,
1018          * there's still a race between the above check and our incref here.
1019          * Really, class_uuid2obd should take the ref. */
1020         class_incref(target, __func__, current);
1021         target_referenced = true;
1022
1023         target->obd_conn_inprogress++;
1024         spin_unlock(&target->obd_dev_lock);
1025
1026         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1027         if (str == NULL) {
1028                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1029                 GOTO(out, rc = -EINVAL);
1030         }
1031
1032         obd_str2uuid(&cluuid, str);
1033
1034         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1035         if (tmp == NULL)
1036                 GOTO(out, rc = -EPROTO);
1037
1038         conn = *tmp;
1039
1040         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1041                                     RCL_CLIENT);
1042         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1043         if (!data)
1044                 GOTO(out, rc = -EPROTO);
1045
1046         rc = req_capsule_server_pack(&req->rq_pill);
1047         if (rc)
1048                 GOTO(out, rc);
1049
1050 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1051         /* Don't allow clients to connect that are using old 1.8 format
1052          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1053          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1054          * FULL20 flag should be set on all connections since 2.0, but no
1055          * longer affects behaviour.
1056          *
1057          * Later this check will be disabled and the flag can be retired
1058          * completely once interop with 3.0 is no longer needed.
1059          */
1060         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1061                 GOTO(out, rc = -EPROTO);
1062 #endif
1063
1064         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1065                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1066                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1067                     data->ocd_version > LUSTRE_VERSION_CODE +
1068                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1069                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1070                                   "libclient connection attempt",
1071                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1072                                   "old" : "new",
1073                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1074                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1075                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1076                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1077                         data = req_capsule_server_sized_get(&req->rq_pill,
1078                                                             &RMF_CONNECT_DATA,
1079                                     offsetof(typeof(*data), ocd_version) +
1080                                              sizeof(data->ocd_version));
1081                         if (data) {
1082                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1083                                 data->ocd_version = LUSTRE_VERSION_CODE;
1084                         }
1085                         GOTO(out, rc = -EPROTO);
1086                 }
1087         }
1088
1089         /* Note: lw_client is needed in MDS-MDS failover during update log
1090          * processing, so we needs to allow lw_client to be connected at
1091          * anytime, instead of only the initial connection */
1092         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1093
1094         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1095                 initial_conn = true;
1096                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1097                 mds_mds_conn = (data->ocd_connect_flags &
1098                                 OBD_CONNECT_MDS_MDS) != 0;
1099
1100                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1101                  * for Imperative Recovery connection from MGC to MGS.
1102                  *
1103                  * Via check OBD_CONNECT_FID, we can distinguish whether
1104                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1105                  * MGC or MDT. */
1106                 if (!lw_client &&
1107                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1108                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1109                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1110                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1111                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1112                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1113
1114                         /* We do not support the MDT-MDT interoperations with
1115                          * different version MDT because of protocol changes. */
1116                         if (unlikely(major != LUSTRE_MAJOR ||
1117                                      minor != LUSTRE_MINOR ||
1118                                      abs(patch - LUSTRE_PATCH) > 3)) {
1119                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1120                                         "connection from different version MDT "
1121                                         "(%d.%d.%d.%d) %s %s\n",
1122                                         target->obd_name, LUSTRE_MAJOR,
1123                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1124                                         major, minor, patch,
1125                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1126                                         libcfs_nid2str(req->rq_peer.nid), str);
1127
1128                                 GOTO(out, rc = -EPROTO);
1129                         }
1130                 }
1131         }
1132
1133         /* lctl gets a backstage, all-access pass. */
1134         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1135                 goto dont_check_exports;
1136
1137         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1138         if (!export)
1139                 goto no_export;
1140
1141         /* We've found an export in the hash. */
1142
1143         spin_lock(&export->exp_lock);
1144
1145         if (export->exp_connecting) { /* bug 9635, et. al. */
1146                 spin_unlock(&export->exp_lock);
1147                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1148                               export->exp_obd->obd_name, export,
1149                               libcfs_nid2str(req->rq_peer.nid));
1150                 class_export_put(export);
1151                 export = NULL;
1152                 rc = -EALREADY;
1153         } else if ((mds_conn || (lw_client && initial_conn) ||
1154                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1155                    export->exp_connection != NULL) {
1156                 spin_unlock(&export->exp_lock);
1157                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1158                         /* MDS or LWP reconnected after failover. */
1159                         LCONSOLE_WARN("%s: Received %s connection from "
1160                             "%s, removing former export from %s\n",
1161                             target->obd_name, mds_conn ? "MDS" : "LWP",
1162                             libcfs_nid2str(req->rq_peer.nid),
1163                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1164                 } else {
1165                         /* New MDS connection from the same NID. */
1166                         LCONSOLE_WARN("%s: Received new %s connection from "
1167                                 "%s, removing former export from same NID\n",
1168                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1169                                 libcfs_nid2str(req->rq_peer.nid));
1170                 }
1171
1172                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1173                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1174                         /* Because exports between MDTs will always be
1175                          * kept, let's do not fail such export if they
1176                          * come from the same NID, otherwise it might
1177                          * cause eviction between MDTs, which might
1178                          * cause namespace inconsistency */
1179                         spin_lock(&export->exp_lock);
1180                         export->exp_connecting = 1;
1181                         spin_unlock(&export->exp_lock);
1182                         conn.cookie = export->exp_handle.h_cookie;
1183                         rc = EALREADY;
1184                 } else {
1185                         class_fail_export(export);
1186                         class_export_put(export);
1187                         export = NULL;
1188                         rc = 0;
1189                 }
1190         } else if (export->exp_connection != NULL && initial_conn &&
1191                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1192                 spin_unlock(&export->exp_lock);
1193                 /* In MDS failover we have static UUID but NID can change. */
1194                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1195                               "existing nid %s is already connected\n",
1196                               target->obd_name, cluuid.uuid,
1197                               libcfs_nid2str(req->rq_peer.nid),
1198                               libcfs_nid2str(
1199                                       export->exp_connection->c_peer.nid));
1200                 rc = -EALREADY;
1201                 class_export_put(export);
1202                 export = NULL;
1203         } else {
1204                 export->exp_connecting = 1;
1205                 spin_unlock(&export->exp_lock);
1206                 LASSERT(export->exp_obd == target);
1207
1208                 rc = target_handle_reconnect(&conn, export, &cluuid);
1209         }
1210
1211         /* If we found an export, we already unlocked. */
1212         if (!export) {
1213 no_export:
1214                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1215         } else if (req->rq_export == NULL &&
1216                    atomic_read(&export->exp_rpc_count) > 0) {
1217                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1218                               "still busy with %d references\n",
1219                               target->obd_name, cluuid.uuid,
1220                               libcfs_nid2str(req->rq_peer.nid),
1221                               atomic_read(&export->exp_refcount));
1222                 GOTO(out, rc = -EBUSY);
1223         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1224                 if (!strstr(cluuid.uuid, "mdt"))
1225                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1226                                       "known client %s (at %s) because it "
1227                                       "is indicating it is a new client",
1228                                       target->obd_name, cluuid.uuid,
1229                                       libcfs_nid2str(req->rq_peer.nid));
1230                 GOTO(out, rc = -EALREADY);
1231         } else {
1232                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1233         }
1234
1235         if (rc < 0) {
1236                 GOTO(out, rc);
1237         }
1238
1239         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1240                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1241               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1242               export, (long)cfs_time_current_sec(),
1243               export ? (long)export->exp_last_request_time : 0);
1244
1245         /* If this is the first time a client connects, reset the recovery
1246          * timer. Discard lightweight connections which might be local. */
1247         if (!lw_client && rc == 0 && target->obd_recovering)
1248                 check_and_start_recovery_timer(target, req, export == NULL);
1249
1250         /* We want to handle EALREADY but *not* -EALREADY from
1251          * target_handle_reconnect(), return reconnection state in a flag. */
1252         if (rc == EALREADY) {
1253                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1254                 rc = 0;
1255         } else {
1256                 LASSERT(rc == 0);
1257         }
1258
1259         /* Tell the client if we support replayable requests. */
1260         if (target->obd_replayable)
1261                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1262         client_nid = &req->rq_peer.nid;
1263
1264         if (export == NULL) {
1265                 /* allow lightweight connections during recovery */
1266                 /* allow "new" MDT to be connected during recovery, since we
1267                  * need retrieve recovery update records from it */
1268                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1269                         cfs_time_t t;
1270                         int     c; /* connected */
1271                         int     i; /* in progress */
1272                         int     k; /* known */
1273                         int     s; /* stale/evicted */
1274
1275                         c = atomic_read(&target->obd_connected_clients);
1276                         i = atomic_read(&target->obd_lock_replay_clients);
1277                         k = target->obd_max_recoverable_clients;
1278                         s = target->obd_stale_clients;
1279                         t = target->obd_recovery_timer.expires;
1280                         t = cfs_time_sub(t, cfs_time_current());
1281                         t = cfs_duration_sec(t);
1282                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1283                                       "(at %s), waiting for %d known clients "
1284                                       "(%d recovered, %d in progress, and %d "
1285                                       "evicted) to recover in %d:%.02d\n",
1286                                       target->obd_name, cluuid.uuid,
1287                                       libcfs_nid2str(req->rq_peer.nid), k,
1288                                       c - i, i, s, (int)t / 60,
1289                                       (int)t % 60);
1290                         rc = -EBUSY;
1291                 } else {
1292 dont_check_exports:
1293                         rc = obd_connect(req->rq_svc_thread->t_env,
1294                                          &export, target, &cluuid, data,
1295                                          client_nid);
1296                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1297                                 lustre_msg_add_op_flags(req->rq_repmsg,
1298                                                         MSG_CONNECT_RECOVERING);
1299                         if (rc == 0) {
1300                                 conn.cookie = export->exp_handle.h_cookie;
1301                                 rc = rev_import_init(export);
1302                         }
1303
1304                         if (mds_mds_conn)
1305                                 new_mds_mds_conn = true;
1306                 }
1307         } else {
1308                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1309                                    export, target, &cluuid, data, client_nid);
1310         }
1311         if (rc)
1312                 GOTO(out, rc);
1313
1314         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1315         data->ocd_instance = target->u.obt.obt_instance;
1316
1317         /* Return only the parts of obd_connect_data that we understand, so the
1318          * client knows that we don't understand the rest. */
1319         if (data) {
1320                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1321                                                RCL_SERVER);
1322                 tmpdata = req_capsule_server_get(&req->rq_pill,
1323                                                  &RMF_CONNECT_DATA);
1324                 /* Don't use struct assignment here, because the client reply
1325                  * buffer may be smaller/larger than the local struct
1326                  * obd_connect_data. */
1327                 memcpy(tmpdata, data, min(tmpsize, size));
1328         }
1329
1330         /* If the client and the server are the same node, we will already
1331          * have an export that really points to the client's DLM export,
1332          * because we have a shared handles table.
1333          *
1334          * XXX this will go away when shaver stops sending the "connect" handle
1335          * in the real "remote handle" field of the request --phik 24 Apr 2003
1336          */
1337         ptlrpc_request_change_export(req, export);
1338
1339         spin_lock(&export->exp_lock);
1340         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1341                 spin_unlock(&export->exp_lock);
1342                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1343                        "or equal conn_cnt: %d >= %d\n",
1344                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1345                        export->exp_conn_cnt,
1346                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1347
1348                 GOTO(out, rc = -EALREADY);
1349         }
1350         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1351         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1352
1353         /* Don't evict liblustre clients for not pinging. */
1354         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1355                 export->exp_libclient = 1;
1356                 spin_unlock(&export->exp_lock);
1357
1358                 spin_lock(&target->obd_dev_lock);
1359                 list_del_init(&export->exp_obd_chain_timed);
1360                 spin_unlock(&target->obd_dev_lock);
1361         } else {
1362                 spin_unlock(&export->exp_lock);
1363         }
1364
1365         if (export->exp_connection != NULL) {
1366                 /* Check to see if connection came from another NID. */
1367                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1368                     !hlist_unhashed(&export->exp_nid_hash))
1369                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1370                                      &export->exp_connection->c_peer.nid,
1371                                      &export->exp_nid_hash);
1372
1373                 ptlrpc_connection_put(export->exp_connection);
1374         }
1375
1376         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1377                                                        req->rq_self,
1378                                                        &cluuid);
1379         if (hlist_unhashed(&export->exp_nid_hash))
1380                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1381                              &export->exp_connection->c_peer.nid,
1382                              &export->exp_nid_hash);
1383
1384         lustre_msg_set_handle(req->rq_repmsg, &conn);
1385
1386         rc = rev_import_reconnect(export, req);
1387         if (rc != 0)
1388                 GOTO(out, rc);
1389
1390         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1391                 int has_transno;
1392                 __u64 transno = data->ocd_transno;
1393
1394                 spin_lock(&export->exp_lock);
1395                 /* possible race with class_disconnect_stale_exports,
1396                  * export may be already in the eviction process */
1397                 if (export->exp_failed) {
1398                         spin_unlock(&export->exp_lock);
1399                         GOTO(out, rc = -ENODEV);
1400                 }
1401                 export->exp_in_recovery = 1;
1402                 export->exp_req_replay_needed = 1;
1403                 export->exp_lock_replay_needed = 1;
1404                 spin_unlock(&export->exp_lock);
1405
1406                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1407                                  MSG_CONNECT_TRANSNO);
1408                 if (has_transno && transno == 0)
1409                         CWARN("Connect with zero transno!\n");
1410
1411                 if (has_transno && transno > 0 &&
1412                     transno < target->obd_next_recovery_transno &&
1413                     transno > target->obd_last_committed) {
1414                         /* Another way is to use cmpxchg() to be lock-free. */
1415                         spin_lock(&target->obd_recovery_task_lock);
1416                         if (transno < target->obd_next_recovery_transno)
1417                                 target->obd_next_recovery_transno = transno;
1418                         spin_unlock(&target->obd_recovery_task_lock);
1419                 }
1420
1421                 atomic_inc(&target->obd_req_replay_clients);
1422                 atomic_inc(&target->obd_lock_replay_clients);
1423                 /* Note: MDS-MDS connection is allowed to be connected during
1424                  * recovery, no matter if the exports needs to be recoveried.
1425                  * Because we need retrieve updates logs from all other MDTs.
1426                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1427                  * also needs to be increased to match other recovery checking
1428                  * condition. */
1429                 if (new_mds_mds_conn)
1430                         target->obd_max_recoverable_clients++;
1431                 if (atomic_inc_return(&target->obd_connected_clients) ==
1432                     target->obd_max_recoverable_clients)
1433                         wake_up(&target->obd_next_transno_waitq);
1434         }
1435
1436         /* Tell the client we're in recovery, when client is involved in it. */
1437         if (target->obd_recovering && !lw_client)
1438                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1439
1440 out:
1441         if (export) {
1442                 spin_lock(&export->exp_lock);
1443                 export->exp_connecting = 0;
1444                 spin_unlock(&export->exp_lock);
1445
1446                 class_export_put(export);
1447         }
1448         if (target_referenced == true && target != NULL) {
1449                 spin_lock(&target->obd_dev_lock);
1450                 target->obd_conn_inprogress--;
1451                 spin_unlock(&target->obd_dev_lock);
1452
1453                 class_decref(target, __func__, current);
1454         }
1455         req->rq_status = rc;
1456         RETURN(rc);
1457 }
1458
1459 int target_handle_disconnect(struct ptlrpc_request *req)
1460 {
1461         int rc;
1462         ENTRY;
1463
1464         rc = req_capsule_server_pack(&req->rq_pill);
1465         if (rc)
1466                 RETURN(rc);
1467
1468         /* Keep the rq_export around so we can send the reply. */
1469         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1470
1471         RETURN(0);
1472 }
1473
1474 void target_destroy_export(struct obd_export *exp)
1475 {
1476         struct obd_import       *imp = NULL;
1477         /* exports created from last_rcvd data, and "fake"
1478            exports created by lctl don't have an import */
1479         spin_lock(&exp->exp_lock);
1480         if (exp->exp_imp_reverse != NULL) {
1481                 imp = exp->exp_imp_reverse;
1482                 exp->exp_imp_reverse = NULL;
1483         }
1484         spin_unlock(&exp->exp_lock);
1485         if (imp != NULL)
1486                 client_destroy_import(imp);
1487
1488         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1489         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1490         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1491         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1492 }
1493 EXPORT_SYMBOL(target_destroy_export);
1494
1495 /*
1496  * Recovery functions
1497  */
1498 static void target_request_copy_get(struct ptlrpc_request *req)
1499 {
1500         class_export_rpc_inc(req->rq_export);
1501         LASSERT(list_empty(&req->rq_list));
1502         INIT_LIST_HEAD(&req->rq_replay_list);
1503
1504         /* Increase refcount to keep request in queue. */
1505         atomic_inc(&req->rq_refcount);
1506         /* Let export know it has replays to be handled. */
1507         atomic_inc(&req->rq_export->exp_replay_count);
1508 }
1509
1510 static void target_request_copy_put(struct ptlrpc_request *req)
1511 {
1512         LASSERT(list_empty(&req->rq_replay_list));
1513         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1514
1515         atomic_dec(&req->rq_export->exp_replay_count);
1516         class_export_rpc_dec(req->rq_export);
1517         ptlrpc_server_drop_request(req);
1518 }
1519
1520 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1521 {
1522         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1523         struct obd_export     *exp = req->rq_export;
1524         struct ptlrpc_request *reqiter;
1525         struct ptlrpc_request *dup_req = NULL;
1526         int                    dup = 0;
1527
1528         LASSERT(exp);
1529
1530         spin_lock(&exp->exp_lock);
1531         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1532                                 rq_replay_list) {
1533                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1534                         dup_req = reqiter;
1535                         dup = 1;
1536                         break;
1537                 }
1538         }
1539
1540         if (dup) {
1541                 /* We expect it with RESENT and REPLAY flags. */
1542                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1543                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1544                         CERROR("invalid flags %x of resent replay\n",
1545                                lustre_msg_get_flags(req->rq_reqmsg));
1546
1547                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1548                         __u32 new_conn;
1549
1550                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1551                         if (new_conn >
1552                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1553                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1554                                                         new_conn);
1555                 }
1556         } else {
1557                 list_add_tail(&req->rq_replay_list,
1558                                   &exp->exp_req_replay_queue);
1559         }
1560
1561         spin_unlock(&exp->exp_lock);
1562         return dup;
1563 }
1564
1565 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1566 {
1567         LASSERT(!list_empty(&req->rq_replay_list));
1568         LASSERT(req->rq_export);
1569
1570         spin_lock(&req->rq_export->exp_lock);
1571         list_del_init(&req->rq_replay_list);
1572         spin_unlock(&req->rq_export->exp_lock);
1573 }
1574
1575 static void target_finish_recovery(struct lu_target *lut)
1576 {
1577         struct obd_device *obd = lut->lut_obd;
1578         ENTRY;
1579
1580         /* Only log a recovery message when recovery has occurred. */
1581         if (obd->obd_recovery_start) {
1582                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1583                                         obd->obd_recovery_start);
1584                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1585                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1586                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1587                         obd->obd_max_recoverable_clients,
1588                         atomic_read(&obd->obd_connected_clients),
1589                         obd->obd_stale_clients,
1590                         obd->obd_stale_clients == 1 ? "was" : "were");
1591         }
1592
1593         ldlm_reprocess_all_ns(obd->obd_namespace);
1594         spin_lock(&obd->obd_recovery_task_lock);
1595         if (!list_empty(&obd->obd_req_replay_queue) ||
1596             !list_empty(&obd->obd_lock_replay_queue) ||
1597             !list_empty(&obd->obd_final_req_queue)) {
1598                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1599                        obd->obd_name,
1600                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1601                        list_empty(&obd->obd_lock_replay_queue) ? \
1602                                "" : "lock ",
1603                        list_empty(&obd->obd_final_req_queue) ? \
1604                                "" : "final ");
1605                 spin_unlock(&obd->obd_recovery_task_lock);
1606                 LBUG();
1607         }
1608         spin_unlock(&obd->obd_recovery_task_lock);
1609
1610         obd->obd_recovery_end = cfs_time_current_sec();
1611
1612         /* When recovery finished, cleanup orphans on MDS and OST. */
1613         if (OBT(obd) && OBP(obd, postrecov)) {
1614                 int rc = OBP(obd, postrecov)(obd);
1615                 if (rc < 0)
1616                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1617                                       obd->obd_name, rc);
1618         }
1619         EXIT;
1620 }
1621
1622 static void abort_req_replay_queue(struct obd_device *obd)
1623 {
1624         struct ptlrpc_request *req, *n;
1625         struct list_head abort_list;
1626
1627         INIT_LIST_HEAD(&abort_list);
1628         spin_lock(&obd->obd_recovery_task_lock);
1629         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1630         spin_unlock(&obd->obd_recovery_task_lock);
1631         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1632                 DEBUG_REQ(D_WARNING, req, "aborted:");
1633                 req->rq_status = -ENOTCONN;
1634                 if (ptlrpc_error(req)) {
1635                         DEBUG_REQ(D_ERROR, req,
1636                                   "failed abort_req_reply; skipping");
1637                 }
1638                 target_exp_dequeue_req_replay(req);
1639                 target_request_copy_put(req);
1640         }
1641 }
1642
1643 static void abort_lock_replay_queue(struct obd_device *obd)
1644 {
1645         struct ptlrpc_request *req, *n;
1646         struct list_head abort_list;
1647
1648         INIT_LIST_HEAD(&abort_list);
1649         spin_lock(&obd->obd_recovery_task_lock);
1650         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1651         spin_unlock(&obd->obd_recovery_task_lock);
1652         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1653                 DEBUG_REQ(D_ERROR, req, "aborted:");
1654                 req->rq_status = -ENOTCONN;
1655                 if (ptlrpc_error(req)) {
1656                         DEBUG_REQ(D_ERROR, req,
1657                                   "failed abort_lock_reply; skipping");
1658                 }
1659                 target_request_copy_put(req);
1660         }
1661 }
1662
1663 /* Called from a cleanup function if the device is being cleaned up
1664    forcefully.  The exports should all have been disconnected already,
1665    the only thing left to do is
1666      - clear the recovery flags
1667      - cancel the timer
1668      - free queued requests and replies, but don't send replies
1669    Because the obd_stopping flag is set, no new requests should be received.
1670
1671 */
1672 void target_cleanup_recovery(struct obd_device *obd)
1673 {
1674         struct ptlrpc_request *req, *n;
1675         struct list_head clean_list;
1676
1677         INIT_LIST_HEAD(&clean_list);
1678         spin_lock(&obd->obd_dev_lock);
1679         if (!obd->obd_recovering) {
1680                 spin_unlock(&obd->obd_dev_lock);
1681                 EXIT;
1682                 return;
1683         }
1684         obd->obd_recovering = obd->obd_abort_recovery = 0;
1685         spin_unlock(&obd->obd_dev_lock);
1686
1687         spin_lock(&obd->obd_recovery_task_lock);
1688         target_cancel_recovery_timer(obd);
1689         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1690         spin_unlock(&obd->obd_recovery_task_lock);
1691
1692         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1693                 LASSERT(req->rq_reply_state == NULL);
1694                 target_exp_dequeue_req_replay(req);
1695                 target_request_copy_put(req);
1696         }
1697
1698         spin_lock(&obd->obd_recovery_task_lock);
1699         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1700         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1701         spin_unlock(&obd->obd_recovery_task_lock);
1702
1703         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1704                 LASSERT(req->rq_reply_state == NULL);
1705                 target_request_copy_put(req);
1706         }
1707
1708         EXIT;
1709 }
1710 EXPORT_SYMBOL(target_cleanup_recovery);
1711
1712 /* obd_recovery_task_lock should be held */
1713 void target_cancel_recovery_timer(struct obd_device *obd)
1714 {
1715         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1716         del_timer(&obd->obd_recovery_timer);
1717 }
1718
1719 static void target_start_recovery_timer(struct obd_device *obd)
1720 {
1721         if (obd->obd_recovery_start != 0)
1722                 return;
1723
1724         spin_lock(&obd->obd_dev_lock);
1725         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1726                 spin_unlock(&obd->obd_dev_lock);
1727                 return;
1728         }
1729
1730         LASSERT(obd->obd_recovery_timeout != 0);
1731
1732         if (obd->obd_recovery_start != 0) {
1733                 spin_unlock(&obd->obd_dev_lock);
1734                 return;
1735         }
1736
1737         mod_timer(&obd->obd_recovery_timer,
1738                   cfs_time_shift(obd->obd_recovery_timeout));
1739         obd->obd_recovery_start = cfs_time_current_sec();
1740         spin_unlock(&obd->obd_dev_lock);
1741
1742         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1743                       "or until %d client%s reconnect%s\n",
1744                       obd->obd_name,
1745                       obd->obd_recovery_timeout / 60,
1746                       obd->obd_recovery_timeout % 60,
1747                       obd->obd_max_recoverable_clients,
1748                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1749                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1750 }
1751
1752 /**
1753  * extend recovery window.
1754  *
1755  * if @extend is true, extend recovery window to have @drt remaining at least;
1756  * otherwise, make sure the recovery timeout value is not less than @drt.
1757  */
1758 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1759 {
1760         cfs_time_t now;
1761         cfs_time_t end;
1762         cfs_duration_t left;
1763         int to;
1764
1765         spin_lock(&obd->obd_dev_lock);
1766         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1767                 spin_unlock(&obd->obd_dev_lock);
1768                 return;
1769         }
1770         LASSERT(obd->obd_recovery_start != 0);
1771
1772         now  = cfs_time_current_sec();
1773         to   = obd->obd_recovery_timeout;
1774         end  = obd->obd_recovery_start + to;
1775         left = cfs_time_sub(end, now);
1776
1777         if (extend && (drt > left)) {
1778                 to += drt - left;
1779         } else if (!extend && (drt > to)) {
1780                 to = drt;
1781         }
1782
1783         if (to > obd->obd_recovery_time_hard) {
1784                 to = obd->obd_recovery_time_hard;
1785                 CWARN("%s: extended recovery timer reaching hard "
1786                       "limit: %d, extend: %d\n",
1787                       obd->obd_name, to, extend);
1788         }
1789
1790         if (obd->obd_recovery_timeout < to) {
1791                 obd->obd_recovery_timeout = to;
1792                 end = obd->obd_recovery_start + to;
1793                 mod_timer(&obd->obd_recovery_timer,
1794                           cfs_time_shift(end - now));
1795         }
1796         spin_unlock(&obd->obd_dev_lock);
1797
1798         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1799                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1800 }
1801
1802 /* Reset the timer with each new client connection */
1803 /*
1804  * This timer is actually reconnect_timer, which is for making sure
1805  * the total recovery window is at least as big as my reconnect
1806  * attempt timing. So the initial recovery time_out will be set to
1807  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1808  * from client is bigger than this, then the recovery time_out will
1809  * be extended to make sure the client could be reconnected, in the
1810  * process, the timeout from the new client should be ignored.
1811  */
1812
1813 static void
1814 check_and_start_recovery_timer(struct obd_device *obd,
1815                                struct ptlrpc_request *req,
1816                                int new_client)
1817 {
1818         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1819         struct obd_device_target *obt = &obd->u.obt;
1820
1821         if (!new_client && service_time)
1822                 /* Teach server about old server's estimates, as first guess
1823                  * at how long new requests will take. */
1824                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1825                             service_time);
1826
1827         target_start_recovery_timer(obd);
1828
1829         /* Convert the service time to RPC timeout,
1830          * and reuse service_time to limit stack usage. */
1831         service_time = at_est2timeout(service_time);
1832
1833         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1834             service_time < at_extra)
1835                 service_time = at_extra;
1836
1837         /* We expect other clients to timeout within service_time, then try
1838          * to reconnect, then try the failover server.  The max delay between
1839          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1840         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1841
1842         LASSERT(obt->obt_magic == OBT_MAGIC);
1843         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1844         if (service_time > obd->obd_recovery_timeout && !new_client)
1845                 extend_recovery_timer(obd, service_time, false);
1846 }
1847
1848 /** Health checking routines */
1849 static inline int exp_connect_healthy(struct obd_export *exp)
1850 {
1851         return (exp->exp_in_recovery);
1852 }
1853
1854 /** if export done req_replay or has replay in queue */
1855 static inline int exp_req_replay_healthy(struct obd_export *exp)
1856 {
1857         return (!exp->exp_req_replay_needed ||
1858                 atomic_read(&exp->exp_replay_count) > 0);
1859 }
1860
1861
1862 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1863 {
1864         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1865                exp_req_replay_healthy(exp);
1866 }
1867
1868 /** if export done lock_replay or has replay in queue */
1869 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1870 {
1871         return (!exp->exp_lock_replay_needed ||
1872                 atomic_read(&exp->exp_replay_count) > 0);
1873 }
1874
1875 static inline int exp_vbr_healthy(struct obd_export *exp)
1876 {
1877         return (!exp->exp_vbr_failed);
1878 }
1879
1880 static inline int exp_finished(struct obd_export *exp)
1881 {
1882         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1883 }
1884
1885 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1886 {
1887         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1888                 exp_finished(exp);
1889 }
1890
1891 static int check_for_next_transno(struct lu_target *lut)
1892 {
1893         struct ptlrpc_request *req = NULL;
1894         struct obd_device *obd = lut->lut_obd;
1895         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1896         int wake_up = 0, connected, completed, queue_len;
1897         __u64 req_transno = 0;
1898         __u64 update_transno = 0;
1899         __u64 next_transno = 0;
1900         ENTRY;
1901
1902         spin_lock(&obd->obd_recovery_task_lock);
1903         if (!list_empty(&obd->obd_req_replay_queue)) {
1904                 req = list_entry(obd->obd_req_replay_queue.next,
1905                                      struct ptlrpc_request, rq_list);
1906                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1907         }
1908
1909         if (tdtd != NULL)
1910                 update_transno = distribute_txn_get_next_transno(tdtd);
1911
1912         connected = atomic_read(&obd->obd_connected_clients);
1913         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1914         queue_len = obd->obd_requests_queued_for_recovery;
1915         next_transno = obd->obd_next_recovery_transno;
1916
1917         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1918                "req_transno: %llu, next_transno: %llu\n",
1919                obd->obd_max_recoverable_clients, connected, completed,
1920                queue_len, req_transno, next_transno);
1921
1922         if (obd->obd_abort_recovery) {
1923                 CDEBUG(D_HA, "waking for aborted recovery\n");
1924                 wake_up = 1;
1925         } else if (obd->obd_recovery_expired) {
1926                 CDEBUG(D_HA, "waking for expired recovery\n");
1927                 wake_up = 1;
1928         } else if (tdtd != NULL && req != NULL &&
1929                    is_req_replayed_by_update(req)) {
1930                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1931                          "next_transno%llu\n", req_transno, next_transno);
1932                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1933                        req_transno);
1934                 wake_up = 1;
1935         } else if (req_transno == next_transno ||
1936                    (update_transno != 0 && update_transno <= next_transno)) {
1937                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1938                 wake_up = 1;
1939         } else if (queue_len > 0 &&
1940                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1941                 /** handle gaps occured due to lost reply or VBR */
1942                 LASSERTF(req_transno >= next_transno,
1943                          "req_transno: %llu, next_transno: %llu\n",
1944                          req_transno, next_transno);
1945                 CDEBUG(D_HA,
1946                        "%s: waking for gap in transno, VBR is %s (skip: "
1947                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1948                        ", next_update %lld last_committed: %lld)\n",
1949                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1950                        next_transno, queue_len, completed, connected,
1951                        req_transno, update_transno, obd->obd_last_committed);
1952                 obd->obd_next_recovery_transno = req_transno;
1953                 wake_up = 1;
1954         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1955                 CDEBUG(D_HA, "waking for completed recovery\n");
1956                 wake_up = 1;
1957         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1958                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1959                        " by fail_lock, waking up (%lld)\n", next_transno);
1960                 obd->obd_next_recovery_transno = req_transno;
1961                 wake_up = 1;
1962         }
1963         spin_unlock(&obd->obd_recovery_task_lock);
1964         return wake_up;
1965 }
1966
1967 static int check_for_next_lock(struct lu_target *lut)
1968 {
1969         struct obd_device *obd = lut->lut_obd;
1970         int wake_up = 0;
1971
1972         spin_lock(&obd->obd_recovery_task_lock);
1973         if (!list_empty(&obd->obd_lock_replay_queue)) {
1974                 CDEBUG(D_HA, "waking for next lock\n");
1975                 wake_up = 1;
1976         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1977                 CDEBUG(D_HA, "waking for completed lock replay\n");
1978                 wake_up = 1;
1979         } else if (obd->obd_abort_recovery) {
1980                 CDEBUG(D_HA, "waking for aborted recovery\n");
1981                 wake_up = 1;
1982         } else if (obd->obd_recovery_expired) {
1983                 CDEBUG(D_HA, "waking for expired recovery\n");
1984                 wake_up = 1;
1985         }
1986         spin_unlock(&obd->obd_recovery_task_lock);
1987
1988         return wake_up;
1989 }
1990
1991 /**
1992  * wait for recovery events,
1993  * check its status with help of check_routine
1994  * evict dead clients via health_check
1995  */
1996 static int target_recovery_overseer(struct lu_target *lut,
1997                                     int (*check_routine)(struct lu_target *),
1998                                     int (*health_check)(struct obd_export *))
1999 {
2000         struct obd_device       *obd = lut->lut_obd;
2001         struct target_distribute_txn_data *tdtd;
2002 repeat:
2003         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
2004               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
2005                 __u64 next_update_transno = 0;
2006
2007                 /* Only abort the recovery if there are no update recovery
2008                  * left in the queue */
2009                 spin_lock(&obd->obd_recovery_task_lock);
2010                 if (lut->lut_tdtd != NULL) {
2011                         next_update_transno =
2012                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2013
2014                         tdtd = lut->lut_tdtd;
2015                         /* If next_update_transno == 0, it probably because
2016                          * updatelog retrieve threads did not get any records
2017                          * yet, let's wait those threads stopped */
2018                         if (next_update_transno == 0) {
2019                                 struct l_wait_info lwi = { 0 };
2020
2021                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2022                                        atomic_read(
2023                                        &tdtd->tdtd_recovery_threads_count) == 0,
2024                                        &lwi);
2025
2026                                 next_update_transno =
2027                                         distribute_txn_get_next_transno(
2028                                                                 lut->lut_tdtd);
2029                         }
2030                 }
2031
2032                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2033                         obd->obd_next_recovery_transno = next_update_transno;
2034                         spin_unlock(&obd->obd_recovery_task_lock);
2035                         /* Disconnect unfinished exports from clients, and
2036                          * keep connection from MDT to make sure the update
2037                          * recovery will still keep trying until some one
2038                          * manually abort the recovery */
2039                         class_disconnect_stale_exports(obd,
2040                                                 exp_finished_or_from_mdt);
2041                         /* Abort all of replay and replay lock req from
2042                          * clients */
2043                         abort_req_replay_queue(obd);
2044                         abort_lock_replay_queue(obd);
2045                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2046                                ")in the queue.\n", obd->obd_name,
2047                                next_update_transno);
2048                 } else {
2049                         obd->obd_abort_recovery = 1;
2050                         spin_unlock(&obd->obd_recovery_task_lock);
2051                         CWARN("%s recovery is aborted by hard timeout\n",
2052                               obd->obd_name);
2053                 }
2054         }
2055
2056         while (wait_event_timeout(obd->obd_next_transno_waitq,
2057                                   check_routine(lut),
2058                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2059                 /* wait indefinitely for event, but don't trigger watchdog */;
2060
2061         if (obd->obd_abort_recovery) {
2062                 CWARN("recovery is aborted, evict exports in recovery\n");
2063                 if (lut->lut_tdtd != NULL) {
2064                         struct l_wait_info lwi = { 0 };
2065
2066                         tdtd = lut->lut_tdtd;
2067                         /* Let's wait all of the update log recovery thread
2068                          * finished */
2069                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2070                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2071                              &lwi);
2072                         /* Then abort the update recovery list */
2073                         dtrq_list_destroy(lut->lut_tdtd);
2074                 }
2075
2076                 /** evict exports which didn't finish recovery yet */
2077                 class_disconnect_stale_exports(obd, exp_finished);
2078                 return 1;
2079         } else if (obd->obd_recovery_expired) {
2080                 obd->obd_recovery_expired = 0;
2081                 /** If some clients died being recovered, evict them */
2082                 LCONSOLE_WARN("%s: recovery is timed out, "
2083                               "evict stale exports\n", obd->obd_name);
2084                 /** evict cexports with no replay in queue, they are stalled */
2085                 class_disconnect_stale_exports(obd, health_check);
2086
2087                 /** continue with VBR */
2088                 spin_lock(&obd->obd_dev_lock);
2089                 obd->obd_version_recov = 1;
2090                 spin_unlock(&obd->obd_dev_lock);
2091                 /**
2092                  * reset timer, recovery will proceed with versions now,
2093                  * timeout is set just to handle reconnection delays
2094                  */
2095                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2096                 /** Wait for recovery events again, after evicting bad clients */
2097                 goto repeat;
2098         }
2099         return 0;
2100 }
2101
2102 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2103 {
2104         struct obd_device       *obd = lut->lut_obd;
2105         struct ptlrpc_request *req = NULL;
2106
2107         CDEBUG(D_HA, "Waiting for lock\n");
2108         if (target_recovery_overseer(lut, check_for_next_lock,
2109                                      exp_lock_replay_healthy))
2110                 abort_lock_replay_queue(obd);
2111
2112         spin_lock(&obd->obd_recovery_task_lock);
2113         if (!list_empty(&obd->obd_lock_replay_queue)) {
2114                 req = list_entry(obd->obd_lock_replay_queue.next,
2115                                      struct ptlrpc_request, rq_list);
2116                 list_del_init(&req->rq_list);
2117                 spin_unlock(&obd->obd_recovery_task_lock);
2118         } else {
2119                 spin_unlock(&obd->obd_recovery_task_lock);
2120                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2121                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2122                 /** evict exports failed VBR */
2123                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2124         }
2125         return req;
2126 }
2127
2128 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2129 {
2130         struct ptlrpc_request *req = NULL;
2131
2132         spin_lock(&obd->obd_recovery_task_lock);
2133         if (!list_empty(&obd->obd_final_req_queue)) {
2134                 req = list_entry(obd->obd_final_req_queue.next,
2135                                      struct ptlrpc_request, rq_list);
2136                 list_del_init(&req->rq_list);
2137                 spin_unlock(&obd->obd_recovery_task_lock);
2138                 if (req->rq_export->exp_in_recovery) {
2139                         spin_lock(&req->rq_export->exp_lock);
2140                         req->rq_export->exp_in_recovery = 0;
2141                         spin_unlock(&req->rq_export->exp_lock);
2142                 }
2143         } else {
2144                 spin_unlock(&obd->obd_recovery_task_lock);
2145         }
2146         return req;
2147 }
2148
2149 static void handle_recovery_req(struct ptlrpc_thread *thread,
2150                                 struct ptlrpc_request *req,
2151                                 svc_handler_t handler)
2152 {
2153         ENTRY;
2154
2155         /**
2156          * export can be evicted during recovery, no need to handle replays for
2157          * it after that, discard such request silently
2158          */
2159         if (req->rq_export->exp_disconnected)
2160                 RETURN_EXIT;
2161
2162         req->rq_session.lc_thread = thread;
2163         req->rq_svc_thread = thread;
2164         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2165
2166         /* thread context */
2167         lu_context_enter(&thread->t_env->le_ctx);
2168         (void)handler(req);
2169         lu_context_exit(&thread->t_env->le_ctx);
2170
2171         /* don't reset timer for final stage */
2172         if (!exp_finished(req->rq_export)) {
2173                 int to = obd_timeout;
2174
2175                 /**
2176                  * Add request timeout to the recovery time so next request from
2177                  * this client may come in recovery time
2178                  */
2179                 if (!AT_OFF) {
2180                         struct ptlrpc_service_part *svcpt;
2181
2182                         svcpt = req->rq_rqbd->rqbd_svcpt;
2183                         /* If the server sent early reply for this request,
2184                          * the client will recalculate the timeout according to
2185                          * current server estimate service time, so we will
2186                          * use the maxium timeout here for waiting the client
2187                          * sending the next req */
2188                         to = max((int)at_est2timeout(
2189                                  at_get(&svcpt->scp_at_estimate)),
2190                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2191                         /* Add 2 net_latency, one for balance rq_deadline
2192                          * (see ptl_send_rpc), one for resend the req to server,
2193                          * Note: client will pack net_latency in replay req
2194                          * (see ptlrpc_replay_req) */
2195                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2196                 }
2197                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2198         }
2199         EXIT;
2200 }
2201
2202 /** Checking routines for recovery */
2203 static int check_for_recovery_ready(struct lu_target *lut)
2204 {
2205         struct obd_device *obd = lut->lut_obd;
2206         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2207
2208         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2209                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2210                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2211                obd->obd_recovery_expired);
2212
2213         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2214                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2215                 if (clnts + obd->obd_stale_clients <
2216                     obd->obd_max_recoverable_clients)
2217                         return 0;
2218         }
2219
2220         if (lut->lut_tdtd != NULL) {
2221                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2222                     !obd->obd_abort_recovery) {
2223                         /* Let's extend recovery timer, in case the recovery
2224                          * timer expired, and some clients got evicted */
2225                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2226                                               true);
2227                         CDEBUG(D_HA, "%s update recovery is not ready,"
2228                                " extend recovery %d\n", obd->obd_name,
2229                                obd->obd_recovery_timeout);
2230                         return 0;
2231                 }
2232         }
2233
2234         return 1;
2235 }
2236
2237 enum {
2238         REQUEST_RECOVERY = 1,
2239         UPDATE_RECOVERY = 2,
2240 };
2241
2242 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2243 {
2244         __u64 transno = 0;
2245
2246         if (!list_empty(&obd->obd_req_replay_queue)) {
2247                 struct ptlrpc_request *req;
2248
2249                 req = list_entry(obd->obd_req_replay_queue.next,
2250                                  struct ptlrpc_request, rq_list);
2251                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2252         }
2253
2254         return transno;
2255 }
2256
2257 static __u64 get_next_transno(struct lu_target *lut, int *type)
2258 {
2259         struct obd_device *obd = lut->lut_obd;
2260         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2261         __u64 transno = 0;
2262         __u64 update_transno;
2263         ENTRY;
2264
2265         transno = get_next_replay_req_transno(obd);
2266         if (type != NULL)
2267                 *type = REQUEST_RECOVERY;
2268
2269         if (tdtd == NULL)
2270                 RETURN(transno);
2271
2272         update_transno = distribute_txn_get_next_transno(tdtd);
2273         if (transno == 0 || (transno >= update_transno &&
2274                              update_transno != 0)) {
2275                 transno = update_transno;
2276                 if (type != NULL)
2277                         *type = UPDATE_RECOVERY;
2278         }
2279
2280         RETURN(transno);
2281 }
2282
2283 /**
2284  * drop duplicate replay request
2285  *
2286  * Because the operation has been replayed by update recovery, the request
2287  * with the same transno will be dropped and also notify the client to send
2288  * next replay request.
2289  *
2290  * \param[in] env       execution environment
2291  * \param[in] obd       failover obd device
2292  * \param[in] req       request to be dropped
2293  */
2294 static void drop_duplicate_replay_req(struct lu_env *env,
2295                                       struct obd_device *obd,
2296                                       struct ptlrpc_request *req)
2297 {
2298         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2299                   " update records are found.\n",
2300                   lustre_msg_get_transno(req->rq_reqmsg),
2301                   libcfs_nid2str(req->rq_peer.nid));
2302
2303         /* Right now, only for MDS reint operation update replay and
2304          * normal request replay can have the same transno */
2305         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2306                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2307                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2308                 if (likely(req->rq_export))
2309                         target_committed_to_req(req);
2310                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2311                 target_send_reply(req, req->rq_status, 0);
2312         } else {
2313                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2314                 libcfs_nid2str(req->rq_peer.nid));
2315         }
2316         target_exp_dequeue_req_replay(req);
2317         target_request_copy_put(req);
2318         obd->obd_replayed_requests++;
2319 }
2320
2321 static void replay_request_or_update(struct lu_env *env,
2322                                      struct lu_target *lut,
2323                                      struct target_recovery_data *trd,
2324                                      struct ptlrpc_thread *thread)
2325 {
2326         struct obd_device *obd = lut->lut_obd;
2327         struct ptlrpc_request *req = NULL;
2328         int                     type;
2329         __u64                   transno;
2330         ENTRY;
2331
2332         CDEBUG(D_HA, "Waiting for transno %lld\n",
2333                obd->obd_next_recovery_transno);
2334
2335         /* Replay all of request and update by transno */
2336         do {
2337                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2338
2339                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2340
2341                 /** It is needed to extend recovery window above
2342                  *  recovery_time_soft. Extending is possible only in the
2343                  *  end of recovery window (see more details in
2344                  *  handle_recovery_req()).
2345                  */
2346                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2347
2348                 if (target_recovery_overseer(lut, check_for_next_transno,
2349                                         exp_req_replay_healthy_or_from_mdt)) {
2350                         abort_req_replay_queue(obd);
2351                         abort_lock_replay_queue(obd);
2352                         goto abort;
2353                 }
2354
2355                 spin_lock(&obd->obd_recovery_task_lock);
2356                 transno = get_next_transno(lut, &type);
2357                 if (type == REQUEST_RECOVERY && transno != 0) {
2358                         /* Drop replay request from client side, if the
2359                          * replay has been executed by update with the
2360                          * same transno */
2361                         req = list_entry(obd->obd_req_replay_queue.next,
2362                                         struct ptlrpc_request, rq_list);
2363
2364                         list_del_init(&req->rq_list);
2365                         obd->obd_requests_queued_for_recovery--;
2366                         spin_unlock(&obd->obd_recovery_task_lock);
2367
2368                         /* Let's check if the request has been redone by
2369                          * update replay */
2370                         if (is_req_replayed_by_update(req)) {
2371                                 struct distribute_txn_replay_req *dtrq;
2372
2373                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2374                                                                    req->rq_xid);
2375                                 LASSERT(dtrq != NULL);
2376                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2377                                 list_del_init(&dtrq->dtrq_list);
2378                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2379                                 dtrq_destroy(dtrq);
2380
2381                                 drop_duplicate_replay_req(env, obd, req);
2382
2383                                 continue;
2384                         }
2385
2386                         LASSERT(trd->trd_processing_task == current_pid());
2387                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2388                                   lustre_msg_get_transno(req->rq_reqmsg),
2389                                   libcfs_nid2str(req->rq_peer.nid));
2390
2391                         handle_recovery_req(thread, req,
2392                                             trd->trd_recovery_handler);
2393                         /**
2394                          * bz18031: increase next_recovery_transno before
2395                          * target_request_copy_put() will drop exp_rpc reference
2396                          */
2397                         spin_lock(&obd->obd_recovery_task_lock);
2398                         obd->obd_next_recovery_transno++;
2399                         spin_unlock(&obd->obd_recovery_task_lock);
2400                         target_exp_dequeue_req_replay(req);
2401                         target_request_copy_put(req);
2402                         obd->obd_replayed_requests++;
2403                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2404                         struct distribute_txn_replay_req *dtrq;
2405                         int rc;
2406
2407                         spin_unlock(&obd->obd_recovery_task_lock);
2408
2409                         LASSERT(tdtd != NULL);
2410                         dtrq = distribute_txn_get_next_req(tdtd);
2411                         lu_context_enter(&thread->t_env->le_ctx);
2412                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2413                         lu_context_exit(&thread->t_env->le_ctx);
2414                         extend_recovery_timer(obd, obd_timeout, true);
2415
2416                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2417                                 CDEBUG(D_HA, "Move x%llu t%llu"
2418                                        " to finish list\n", dtrq->dtrq_xid,
2419                                        dtrq->dtrq_master_transno);
2420
2421                                 /* Add it to the replay finish list */
2422                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2423                                 list_add(&dtrq->dtrq_list,
2424                                          &tdtd->tdtd_replay_finish_list);
2425                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2426
2427                                 spin_lock(&obd->obd_recovery_task_lock);
2428                                 if (transno == obd->obd_next_recovery_transno)
2429                                         obd->obd_next_recovery_transno++;
2430                                 else if (transno >
2431                                          obd->obd_next_recovery_transno)
2432                                         obd->obd_next_recovery_transno =
2433                                                                 transno + 1;
2434                                 spin_unlock(&obd->obd_recovery_task_lock);
2435                         } else {
2436                                 dtrq_destroy(dtrq);
2437                         }
2438                 } else {
2439                         spin_unlock(&obd->obd_recovery_task_lock);
2440 abort:
2441                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2442                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2443                         /** evict exports failed VBR */
2444                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2445                         break;
2446                 }
2447         } while (1);
2448 }
2449
2450 static int target_recovery_thread(void *arg)
2451 {
2452         struct lu_target *lut = arg;
2453         struct obd_device *obd = lut->lut_obd;
2454         struct ptlrpc_request *req;
2455         struct target_recovery_data *trd = &obd->obd_recovery_data;
2456         unsigned long delta;
2457         struct lu_env *env;
2458         struct ptlrpc_thread *thread = NULL;
2459         int rc = 0;
2460         ENTRY;
2461
2462         unshare_fs_struct();
2463         OBD_ALLOC_PTR(thread);
2464         if (thread == NULL)
2465                 RETURN(-ENOMEM);
2466
2467         OBD_ALLOC_PTR(env);
2468         if (env == NULL) {
2469                 OBD_FREE_PTR(thread);
2470                 RETURN(-ENOMEM);
2471         }
2472
2473         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2474         if (rc) {
2475                 OBD_FREE_PTR(thread);
2476                 OBD_FREE_PTR(env);
2477                 RETURN(rc);
2478         }
2479
2480         thread->t_env = env;
2481         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2482         env->le_ctx.lc_thread = thread;
2483         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2484         thread->t_watchdog = NULL;
2485
2486         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2487                current_pid());
2488         trd->trd_processing_task = current_pid();
2489
2490         spin_lock(&obd->obd_dev_lock);
2491         obd->obd_recovering = 1;
2492         spin_unlock(&obd->obd_dev_lock);
2493         complete(&trd->trd_starting);
2494
2495         /* first of all, we have to know the first transno to replay */
2496         if (target_recovery_overseer(lut, check_for_recovery_ready,
2497                                      exp_connect_healthy)) {
2498                 abort_req_replay_queue(obd);
2499                 abort_lock_replay_queue(obd);
2500                 if (lut->lut_tdtd != NULL)
2501                         dtrq_list_destroy(lut->lut_tdtd);
2502         }
2503
2504         /* next stage: replay requests or update */
2505         delta = jiffies;
2506         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2507                atomic_read(&obd->obd_req_replay_clients),
2508                obd->obd_next_recovery_transno);
2509         replay_request_or_update(env, lut, trd, thread);
2510
2511         /**
2512          * The second stage: replay locks
2513          */
2514         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2515                atomic_read(&obd->obd_lock_replay_clients));
2516         while ((req = target_next_replay_lock(lut))) {
2517                 LASSERT(trd->trd_processing_task == current_pid());
2518                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2519                           libcfs_nid2str(req->rq_peer.nid));
2520                 handle_recovery_req(thread, req,
2521                                     trd->trd_recovery_handler);
2522                 target_request_copy_put(req);
2523                 obd->obd_replayed_locks++;
2524         }
2525
2526         /**
2527          * The third stage: reply on final pings, at this moment all clients
2528          * must have request in final queue
2529          */
2530         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2531         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2532         /** Update server last boot epoch */
2533         tgt_boot_epoch_update(lut);
2534         /* We drop recoverying flag to forward all new requests
2535          * to regular mds_handle() since now */
2536         spin_lock(&obd->obd_dev_lock);
2537         obd->obd_recovering = obd->obd_abort_recovery = 0;
2538         spin_unlock(&obd->obd_dev_lock);
2539         spin_lock(&obd->obd_recovery_task_lock);
2540         target_cancel_recovery_timer(obd);
2541         spin_unlock(&obd->obd_recovery_task_lock);
2542         while ((req = target_next_final_ping(obd))) {
2543                 LASSERT(trd->trd_processing_task == current_pid());
2544                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2545                           libcfs_nid2str(req->rq_peer.nid));
2546                 handle_recovery_req(thread, req,
2547                                     trd->trd_recovery_handler);
2548                 /* Because the waiting client can not send ping to server,
2549                  * so we need refresh the last_request_time, to avoid the
2550                  * export is being evicted */
2551                 ptlrpc_update_export_timer(req->rq_export, 0);
2552                 target_request_copy_put(req);
2553         }
2554
2555         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2556         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2557               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2558         if (delta > OBD_RECOVERY_TIME_SOFT) {
2559                 CWARN("too long recovery - read logs\n");
2560                 libcfs_debug_dumplog();
2561         }
2562
2563         target_finish_recovery(lut);
2564
2565         lu_context_fini(&env->le_ctx);
2566         trd->trd_processing_task = 0;
2567         complete(&trd->trd_finishing);
2568
2569         tgt_io_thread_done(thread);
2570         OBD_FREE_PTR(thread);
2571         OBD_FREE_PTR(env);
2572         RETURN(rc);
2573 }
2574
2575 static int target_start_recovery_thread(struct lu_target *lut,
2576                                         svc_handler_t handler)
2577 {
2578         struct obd_device *obd = lut->lut_obd;
2579         int rc = 0;
2580         struct target_recovery_data *trd = &obd->obd_recovery_data;
2581         int index;
2582
2583         memset(trd, 0, sizeof(*trd));
2584         init_completion(&trd->trd_starting);
2585         init_completion(&trd->trd_finishing);
2586         trd->trd_recovery_handler = handler;
2587
2588         rc = server_name2index(obd->obd_name, &index, NULL);
2589         if (rc < 0)
2590                 return rc;
2591
2592         if (!IS_ERR(kthread_run(target_recovery_thread,
2593                                 lut, "tgt_recover_%d", index))) {
2594                 wait_for_completion(&trd->trd_starting);
2595                 LASSERT(obd->obd_recovering != 0);
2596         } else {
2597                 rc = -ECHILD;
2598         }
2599
2600         return rc;
2601 }
2602
2603 void target_stop_recovery_thread(struct obd_device *obd)
2604 {
2605         if (obd->obd_recovery_data.trd_processing_task > 0) {
2606                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2607                 /** recovery can be done but postrecovery is not yet */
2608                 spin_lock(&obd->obd_dev_lock);
2609                 if (obd->obd_recovering) {
2610                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2611                         obd->obd_abort_recovery = 1;
2612                         wake_up(&obd->obd_next_transno_waitq);
2613                 }
2614                 spin_unlock(&obd->obd_dev_lock);
2615                 wait_for_completion(&trd->trd_finishing);
2616         }
2617 }
2618 EXPORT_SYMBOL(target_stop_recovery_thread);
2619
2620 void target_recovery_fini(struct obd_device *obd)
2621 {
2622         class_disconnect_exports(obd);
2623         target_stop_recovery_thread(obd);
2624         target_cleanup_recovery(obd);
2625 }
2626 EXPORT_SYMBOL(target_recovery_fini);
2627
2628 static void target_recovery_expired(unsigned long castmeharder)
2629 {
2630         struct obd_device *obd = (struct obd_device *)castmeharder;
2631         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2632                " after %lds (%d clients connected)\n",
2633                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2634                cfs_time_current_sec()- obd->obd_recovery_start,
2635                atomic_read(&obd->obd_connected_clients));
2636
2637         obd->obd_recovery_expired = 1;
2638         wake_up(&obd->obd_next_transno_waitq);
2639 }
2640
2641 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2642 {
2643         struct obd_device *obd = lut->lut_obd;
2644
2645         if (obd->obd_max_recoverable_clients == 0) {
2646                 /** Update server last boot epoch */
2647                 tgt_boot_epoch_update(lut);
2648                 return;
2649         }
2650
2651         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2652                "last_transno %llu\n", obd->obd_name,
2653                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2654         LASSERT(obd->obd_stopping == 0);
2655         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2656         obd->obd_recovery_start = 0;
2657         obd->obd_recovery_end = 0;
2658
2659         setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2660                     (unsigned long)obd);
2661         target_start_recovery_thread(lut, handler);
2662 }
2663 EXPORT_SYMBOL(target_recovery_init);
2664
2665 static int target_process_req_flags(struct obd_device *obd,
2666                                     struct ptlrpc_request *req)
2667 {
2668         struct obd_export *exp = req->rq_export;
2669         LASSERT(exp != NULL);
2670         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2671                 /* client declares he's ready to replay locks */
2672                 spin_lock(&exp->exp_lock);
2673                 if (exp->exp_req_replay_needed) {
2674                         exp->exp_req_replay_needed = 0;
2675                         spin_unlock(&exp->exp_lock);
2676
2677                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2678                         atomic_dec(&obd->obd_req_replay_clients);
2679                 } else {
2680                         spin_unlock(&exp->exp_lock);
2681                 }
2682         }
2683         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2684                 /* client declares he's ready to complete recovery
2685                  * so, we put the request on th final queue */
2686                 spin_lock(&exp->exp_lock);
2687                 if (exp->exp_lock_replay_needed) {
2688                         exp->exp_lock_replay_needed = 0;
2689                         spin_unlock(&exp->exp_lock);
2690
2691                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2692                         atomic_dec(&obd->obd_lock_replay_clients);
2693                 } else {
2694                         spin_unlock(&exp->exp_lock);
2695                 }
2696         }
2697         return 0;
2698 }
2699
2700 int target_queue_recovery_request(struct ptlrpc_request *req,
2701                                   struct obd_device *obd)
2702 {
2703         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2704         struct ptlrpc_request *reqiter;
2705         int inserted = 0;
2706         ENTRY;
2707
2708         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2709                 /* Processing the queue right now, don't re-add. */
2710                 RETURN(1);
2711         }
2712
2713         target_process_req_flags(obd, req);
2714
2715         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2716                 /* client declares he's ready to complete recovery
2717                  * so, we put the request on th final queue */
2718                 target_request_copy_get(req);
2719                 DEBUG_REQ(D_HA, req, "queue final req");
2720                 wake_up(&obd->obd_next_transno_waitq);
2721                 spin_lock(&obd->obd_recovery_task_lock);
2722                 if (obd->obd_recovering) {
2723                         struct ptlrpc_request *tmp;
2724                         struct ptlrpc_request *duplicate = NULL;
2725
2726                         if (likely(!req->rq_export->exp_replay_done)) {
2727                                 req->rq_export->exp_replay_done = 1;
2728                                 list_add_tail(&req->rq_list,
2729                                               &obd->obd_final_req_queue);
2730                                 spin_unlock(&obd->obd_recovery_task_lock);
2731                                 RETURN(0);
2732                         }
2733
2734                         /* XXX O(n), but only happens if final ping is
2735                          * timed out, probably reorganize the list as
2736                          * a hash list later */
2737                         list_for_each_entry_safe(reqiter, tmp,
2738                                                  &obd->obd_final_req_queue,
2739                                                  rq_list) {
2740                                 if (reqiter->rq_export == req->rq_export) {
2741                                         list_del_init(&reqiter->rq_list);
2742                                         duplicate = reqiter;
2743                                         break;
2744                                 }
2745                         }
2746
2747                         list_add_tail(&req->rq_list,
2748                                       &obd->obd_final_req_queue);
2749                         req->rq_export->exp_replay_done = 1;
2750                         spin_unlock(&obd->obd_recovery_task_lock);
2751
2752                         if (duplicate != NULL) {
2753                                 DEBUG_REQ(D_HA, duplicate,
2754                                           "put prev final req\n");
2755                                 target_request_copy_put(duplicate);
2756                         }
2757                         RETURN(0);
2758                 } else {
2759                         spin_unlock(&obd->obd_recovery_task_lock);
2760                         target_request_copy_put(req);
2761                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2762                 }
2763         }
2764         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2765                 /* client declares he's ready to replay locks */
2766                 target_request_copy_get(req);
2767                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2768                 wake_up(&obd->obd_next_transno_waitq);
2769                 spin_lock(&obd->obd_recovery_task_lock);
2770                 LASSERT(obd->obd_recovering);
2771                 /* usually due to recovery abort */
2772                 if (!req->rq_export->exp_in_recovery) {
2773                         spin_unlock(&obd->obd_recovery_task_lock);
2774                         target_request_copy_put(req);
2775                         RETURN(-ENOTCONN);
2776                 }
2777                 LASSERT(req->rq_export->exp_lock_replay_needed);
2778                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2779                 spin_unlock(&obd->obd_recovery_task_lock);
2780                 RETURN(0);
2781         }
2782
2783         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2784          * (i.e. buflens etc are in my own byte order), but type-dependent
2785          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2786
2787         if (!transno) {
2788                 INIT_LIST_HEAD(&req->rq_list);
2789                 DEBUG_REQ(D_HA, req, "not queueing");
2790                 RETURN(1);
2791         }
2792
2793         /* If we're processing the queue, we want don't want to queue this
2794          * message.
2795          *
2796          * Also, if this request has a transno less than the one we're waiting
2797          * for, we should process it now.  It could (and currently always will)
2798          * be an open request for a descriptor that was opened some time ago.
2799          *
2800          * Also, a resent, replayed request that has already been
2801          * handled will pass through here and be processed immediately.
2802          */
2803         CDEBUG(D_HA, "Next recovery transno: %llu"
2804                ", current: %llu, replaying\n",
2805                obd->obd_next_recovery_transno, transno);
2806
2807         /* If the request has been replayed by update replay, then sends this
2808          * request to the recovery thread (replay_request_or_update()), where
2809          * it will be handled */
2810         spin_lock(&obd->obd_recovery_task_lock);
2811         if (transno < obd->obd_next_recovery_transno &&
2812             !is_req_replayed_by_update(req)) {
2813                 /* Processing the queue right now, don't re-add. */
2814                 LASSERT(list_empty(&req->rq_list));
2815                 spin_unlock(&obd->obd_recovery_task_lock);
2816                 RETURN(1);
2817         }
2818         spin_unlock(&obd->obd_recovery_task_lock);
2819
2820         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2821                 RETURN(0);
2822
2823         target_request_copy_get(req);
2824         if (!req->rq_export->exp_in_recovery) {
2825                 target_request_copy_put(req);
2826                 RETURN(-ENOTCONN);
2827         }
2828         LASSERT(req->rq_export->exp_req_replay_needed);
2829
2830         if (target_exp_enqueue_req_replay(req)) {
2831                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2832                 target_request_copy_put(req);
2833                 RETURN(0);
2834         }
2835
2836         /* XXX O(n^2) */
2837         spin_lock(&obd->obd_recovery_task_lock);
2838         LASSERT(obd->obd_recovering);
2839         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2840                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2841                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2842                         inserted = 1;
2843                         goto added;
2844                 }
2845
2846                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2847                              transno)) {
2848                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2849                                   "has been claimed by another client");
2850                         spin_unlock(&obd->obd_recovery_task_lock);
2851                         target_exp_dequeue_req_replay(req);
2852                         target_request_copy_put(req);
2853                         RETURN(0);
2854                 }
2855         }
2856 added:
2857         if (!inserted)
2858                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2859
2860         obd->obd_requests_queued_for_recovery++;
2861         spin_unlock(&obd->obd_recovery_task_lock);
2862         wake_up(&obd->obd_next_transno_waitq);
2863         RETURN(0);
2864 }
2865
2866 int target_handle_ping(struct ptlrpc_request *req)
2867 {
2868         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2869         return req_capsule_server_pack(&req->rq_pill);
2870 }
2871
2872 void target_committed_to_req(struct ptlrpc_request *req)
2873 {
2874         struct obd_export *exp = req->rq_export;
2875
2876         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2877                 lustre_msg_set_last_committed(req->rq_repmsg,
2878                                               exp->exp_last_committed);
2879         else
2880                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2881                           "%d)", exp->exp_obd->obd_no_transno,
2882                           req->rq_repmsg == NULL);
2883
2884         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2885                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2886 }
2887
2888 #endif /* HAVE_SERVER_SUPPORT */
2889
2890 /**
2891  * Packs current SLV and Limit into \a req.
2892  */
2893 int target_pack_pool_reply(struct ptlrpc_request *req)
2894 {
2895         struct obd_device *obd;
2896         ENTRY;
2897
2898         /* Check that we still have all structures alive as this may
2899          * be some late RPC at shutdown time. */
2900         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2901                      !exp_connect_lru_resize(req->rq_export))) {
2902                 lustre_msg_set_slv(req->rq_repmsg, 0);
2903                 lustre_msg_set_limit(req->rq_repmsg, 0);
2904                 RETURN(0);
2905         }
2906
2907         /* OBD is alive here as export is alive, which we checked above. */
2908         obd = req->rq_export->exp_obd;
2909
2910         read_lock(&obd->obd_pool_lock);
2911         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2912         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2913         read_unlock(&obd->obd_pool_lock);
2914
2915         RETURN(0);
2916 }
2917
2918 static int target_send_reply_msg(struct ptlrpc_request *req,
2919                                  int rc, int fail_id)
2920 {
2921         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2922                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2923                 return -ECOMM;
2924         }
2925         /* We can have a null rq_reqmsg in the event of bad signature or
2926          * no context when unwrapping */
2927         if (req->rq_reqmsg &&
2928             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2929             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2930                 return -ECOMM;
2931
2932         if (unlikely(rc)) {
2933                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2934                 req->rq_status = rc;
2935                 return ptlrpc_send_error(req, 1);
2936         } else {
2937                 DEBUG_REQ(D_NET, req, "sending reply");
2938         }
2939
2940         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2941 }
2942
2943 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2944 {
2945         struct ptlrpc_service_part *svcpt;
2946         int                        netrc;
2947         struct ptlrpc_reply_state *rs;
2948         struct obd_export         *exp;
2949         ENTRY;
2950
2951         if (req->rq_no_reply) {
2952                 EXIT;
2953                 return;
2954         }
2955
2956         svcpt = req->rq_rqbd->rqbd_svcpt;
2957         rs = req->rq_reply_state;
2958         if (rs == NULL || !rs->rs_difficult) {
2959                 /* no notifiers */
2960                 target_send_reply_msg (req, rc, fail_id);
2961                 EXIT;
2962                 return;
2963         }
2964
2965         /* must be an export if locks saved */
2966         LASSERT(req->rq_export != NULL);
2967         /* req/reply consistent */
2968         LASSERT(rs->rs_svcpt == svcpt);
2969
2970         /* "fresh" reply */
2971         LASSERT(!rs->rs_scheduled);
2972         LASSERT(!rs->rs_scheduled_ever);
2973         LASSERT(!rs->rs_handled);
2974         LASSERT(!rs->rs_on_net);
2975         LASSERT(rs->rs_export == NULL);
2976         LASSERT(list_empty(&rs->rs_obd_list));
2977         LASSERT(list_empty(&rs->rs_exp_list));
2978
2979         exp = class_export_get(req->rq_export);
2980
2981         /* disable reply scheduling while I'm setting up */
2982         rs->rs_scheduled = 1;
2983         rs->rs_on_net    = 1;
2984         rs->rs_xid       = req->rq_xid;
2985         rs->rs_transno   = req->rq_transno;
2986         rs->rs_export    = exp;
2987         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2988
2989         spin_lock(&exp->exp_uncommitted_replies_lock);
2990         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2991                rs->rs_transno, exp->exp_last_committed);
2992         if (rs->rs_transno > exp->exp_last_committed) {
2993                 /* not committed already */
2994                 list_add_tail(&rs->rs_obd_list,
2995                                   &exp->exp_uncommitted_replies);
2996         }
2997         spin_unlock(&exp->exp_uncommitted_replies_lock);
2998
2999         spin_lock(&exp->exp_lock);
3000         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3001         spin_unlock(&exp->exp_lock);
3002
3003         netrc = target_send_reply_msg(req, rc, fail_id);
3004
3005         spin_lock(&svcpt->scp_rep_lock);
3006
3007         atomic_inc(&svcpt->scp_nreps_difficult);
3008
3009         if (netrc != 0) {
3010                 /* error sending: reply is off the net.  Also we need +1
3011                  * reply ref until ptlrpc_handle_rs() is done
3012                  * with the reply state (if the send was successful, there
3013                  * would have been +1 ref for the net, which
3014                  * reply_out_callback leaves alone) */
3015                 rs->rs_on_net = 0;
3016                 ptlrpc_rs_addref(rs);
3017         }
3018
3019         spin_lock(&rs->rs_lock);
3020         if (rs->rs_transno <= exp->exp_last_committed ||
3021             (!rs->rs_on_net && !rs->rs_no_ack) ||
3022             list_empty(&rs->rs_exp_list) ||     /* completed already */
3023             list_empty(&rs->rs_obd_list)) {
3024                 CDEBUG(D_HA, "Schedule reply immediately\n");
3025                 ptlrpc_dispatch_difficult_reply(rs);
3026         } else {
3027                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3028                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3029         }
3030         spin_unlock(&rs->rs_lock);
3031         spin_unlock(&svcpt->scp_rep_lock);
3032         EXIT;
3033 }
3034
3035 enum ldlm_mode lck_compat_array[] = {
3036         [LCK_EX]    = LCK_COMPAT_EX,
3037         [LCK_PW]    = LCK_COMPAT_PW,
3038         [LCK_PR]    = LCK_COMPAT_PR,
3039         [LCK_CW]    = LCK_COMPAT_CW,
3040         [LCK_CR]    = LCK_COMPAT_CR,
3041         [LCK_NL]    = LCK_COMPAT_NL,
3042         [LCK_GROUP] = LCK_COMPAT_GROUP,
3043         [LCK_COS]   = LCK_COMPAT_COS,
3044 };
3045
3046 /**
3047  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3048  * not escape to the user level.
3049  */
3050 int ldlm_error2errno(enum ldlm_error error)
3051 {
3052         int result;
3053
3054         switch (error) {
3055         case ELDLM_OK:
3056         case ELDLM_LOCK_MATCHED:
3057                 result = 0;
3058                 break;
3059         case ELDLM_LOCK_CHANGED:
3060                 result = -ESTALE;
3061                 break;
3062         case ELDLM_LOCK_ABORTED:
3063                 result = -ENAVAIL;
3064                 break;
3065         case ELDLM_LOCK_REPLACED:
3066                 result = -ESRCH;
3067                 break;
3068         case ELDLM_NO_LOCK_DATA:
3069                 result = -ENOENT;
3070                 break;
3071         case ELDLM_NAMESPACE_EXISTS:
3072                 result = -EEXIST;
3073                 break;
3074         case ELDLM_BAD_NAMESPACE:
3075                 result = -EBADF;
3076                 break;
3077         default:
3078                 if (((int)error) < 0) { /* cast to signed type */
3079                         result = error; /* as ldlm_error can be unsigned */
3080                 } else {
3081                         CERROR("Invalid DLM result code: %d\n", error);
3082                         result = -EPROTO;
3083                 }
3084         }
3085         return result;
3086 }
3087 EXPORT_SYMBOL(ldlm_error2errno);
3088
3089 /**
3090  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3091  */
3092 enum ldlm_error ldlm_errno2error(int err_no)
3093 {
3094         int error;
3095
3096         switch (err_no) {
3097         case 0:
3098                 error = ELDLM_OK;
3099                 break;
3100         case -ESTALE:
3101                 error = ELDLM_LOCK_CHANGED;
3102                 break;
3103         case -ENAVAIL:
3104                 error = ELDLM_LOCK_ABORTED;
3105                 break;
3106         case -ESRCH:
3107                 error = ELDLM_LOCK_REPLACED;
3108                 break;
3109         case -ENOENT:
3110                 error = ELDLM_NO_LOCK_DATA;
3111                 break;
3112         case -EEXIST:
3113                 error = ELDLM_NAMESPACE_EXISTS;
3114                 break;
3115         case -EBADF:
3116                 error = ELDLM_BAD_NAMESPACE;
3117                 break;
3118         default:
3119                 error = err_no;
3120         }
3121         return error;
3122 }
3123
3124 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3125 void ldlm_dump_export_locks(struct obd_export *exp)
3126 {
3127         spin_lock(&exp->exp_locks_list_guard);
3128         if (!list_empty(&exp->exp_locks_list)) {
3129                 struct ldlm_lock *lock;
3130
3131                 CERROR("dumping locks for export %p,"
3132                        "ignore if the unmount doesn't hang\n", exp);
3133                 list_for_each_entry(lock, &exp->exp_locks_list,
3134                                         l_exp_refs_link)
3135                         LDLM_ERROR(lock, "lock:");
3136         }
3137         spin_unlock(&exp->exp_locks_list_guard);
3138 }
3139 #endif
3140
3141 #ifdef HAVE_SERVER_SUPPORT
3142 static int target_bulk_timeout(void *data)
3143 {
3144         ENTRY;
3145         /* We don't fail the connection here, because having the export
3146          * killed makes the (vital) call to commitrw very sad.
3147          */
3148         RETURN(1);
3149 }
3150
3151 static inline const char *bulk2type(struct ptlrpc_request *req)
3152 {
3153         if (req->rq_bulk_read)
3154                 return "READ";
3155         if (req->rq_bulk_write)
3156                 return "WRITE";
3157         return "UNKNOWN";
3158 }
3159
3160 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3161                    struct l_wait_info *lwi)
3162 {
3163         struct ptlrpc_request   *req = desc->bd_req;
3164         time_t                   start = cfs_time_current_sec();
3165         time_t                   deadline;
3166         int                      rc = 0;
3167
3168         ENTRY;
3169
3170         /* If there is eviction in progress, wait for it to finish. */
3171         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3172                 *lwi = LWI_INTR(NULL, NULL);
3173                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3174                                   !atomic_read(&exp->exp_obd->
3175                                                    obd_evict_inprogress),
3176                                   lwi);
3177         }
3178
3179         /* Check if client was evicted or reconnected already. */
3180         if (exp->exp_failed ||
3181             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3182                 rc = -ENOTCONN;
3183         } else {
3184                 if (req->rq_bulk_read)
3185                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3186
3187                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3188                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3189                 else /* old version, bulk matchbits is rq_xid */
3190                         req->rq_mbits = req->rq_xid;
3191
3192                 if (rc == 0)
3193                         rc = ptlrpc_start_bulk_transfer(desc);
3194         }
3195
3196         if (rc < 0) {
3197                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3198                           bulk2type(req), rc);
3199                 RETURN(rc);
3200         }
3201
3202         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3203                 ptlrpc_abort_bulk(desc);
3204                 RETURN(0);
3205         }
3206
3207         /* limit actual bulk transfer to bulk_timeout seconds */
3208         deadline = start + bulk_timeout;
3209         if (deadline > req->rq_deadline)
3210                 deadline = req->rq_deadline;
3211
3212         do {
3213                 long timeoutl = deadline - cfs_time_current_sec();
3214                 cfs_duration_t timeout = timeoutl <= 0 ?
3215                                          CFS_TICK : cfs_time_seconds(timeoutl);
3216                 time_t  rq_deadline;
3217
3218                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3219                                             target_bulk_timeout, desc);
3220                 rc = l_wait_event(desc->bd_waitq,
3221                                   !ptlrpc_server_bulk_active(desc) ||
3222                                   exp->exp_failed ||
3223                                   exp->exp_conn_cnt >
3224                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3225                                   lwi);
3226                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3227                 /* Wait again if we changed rq_deadline. */
3228                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3229                 deadline = start + bulk_timeout;
3230                 if (deadline > rq_deadline)
3231                         deadline = rq_deadline;
3232         } while ((rc == -ETIMEDOUT) &&
3233                  (deadline > cfs_time_current_sec()));
3234
3235         if (rc == -ETIMEDOUT) {
3236                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3237                           bulk2type(req), deadline - start,
3238                           cfs_time_current_sec() - deadline);
3239                 ptlrpc_abort_bulk(desc);
3240         } else if (exp->exp_failed) {
3241                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3242                           bulk2type(req));
3243                 rc = -ENOTCONN;
3244                 ptlrpc_abort_bulk(desc);
3245         } else if (exp->exp_conn_cnt >
3246                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3247                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3248                           bulk2type(req));
3249                 /* We don't reply anyway. */
3250                 rc = -ETIMEDOUT;
3251                 ptlrpc_abort_bulk(desc);
3252         } else if (desc->bd_failure) {
3253                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3254                           bulk2type(req));
3255                 /* XXX should this be a different errno? */
3256                 rc = -ETIMEDOUT;
3257         } else {
3258                 if (req->rq_bulk_write)
3259                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3260                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3261                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3262                                   bulk2type(req), desc->bd_nob_transferred,
3263                                   desc->bd_nob);
3264                         /* XXX should this be a different errno? */
3265                         rc = -ETIMEDOUT;
3266                 }
3267         }
3268
3269         RETURN(rc);
3270 }
3271 EXPORT_SYMBOL(target_bulk_io);
3272
3273 #endif /* HAVE_SERVER_SUPPORT */