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_seq = NULL;
 354         init_rwsem(&cli->cl_seq_rwsem);
 355         cli->cl_conn_count = 0;
 356         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 357                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 358                      sizeof(server_uuid)));
 359
 360         cli->cl_dirty_pages = 0;
 361         cli->cl_avail_grant = 0;
 362         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 363         /* cl_dirty_max_pages may be changed at connect time in
 364          * ptlrpc_connect_interpret(). */
 365         client_adjust_max_dirty(cli);
 366         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 367         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 368         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 369         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 370         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 371         spin_lock_init(&cli->cl_loi_list_lock);
 372         atomic_set(&cli->cl_pending_w_pages, 0);
 373         atomic_set(&cli->cl_pending_r_pages, 0);
 374         cli->cl_r_in_flight = 0;
 375         cli->cl_w_in_flight = 0;
 376
 377         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 378         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 379         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 380         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 381         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 382         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 383
 384         /* lru for osc. */
 385         INIT_LIST_HEAD(&cli->cl_lru_osc);
 386         atomic_set(&cli->cl_lru_shrinkers, 0);
 387         atomic_long_set(&cli->cl_lru_busy, 0);
 388         atomic_long_set(&cli->cl_lru_in_list, 0);
 389         INIT_LIST_HEAD(&cli->cl_lru_list);
 390         spin_lock_init(&cli->cl_lru_list_lock);
 391         atomic_long_set(&cli->cl_unstable_count, 0);
 392         INIT_LIST_HEAD(&cli->cl_shrink_list);
 393
 394         init_waitqueue_head(&cli->cl_destroy_waitq);
 395         atomic_set(&cli->cl_destroy_in_flight, 0);
 396 #ifdef ENABLE_CHECKSUM
 397         /* Turn on checksumming by default. */
 398         cli->cl_checksum = 1;
 399         /*
 400          * The supported checksum types will be worked out at connect time
 401          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 402          * through procfs.
 403          */
 404         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 405 #endif
 406         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 407
 408         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
 409          * from OFD after connecting. */
 410         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 411
 412         /* set cl_chunkbits default value to PAGE_SHIFT,
 413          * it will be updated at OSC connection time. */
 414         cli->cl_chunkbits = PAGE_SHIFT;
 415
 416         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 417                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 418         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 419                 cli->cl_max_rpcs_in_flight = 2;
 420         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 421                 cli->cl_max_rpcs_in_flight = 3;
 422         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 423                 cli->cl_max_rpcs_in_flight = 4;
 424         } else {
 425                 if (osc_on_mdt(obddev->obd_name))
 426                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 427                 else
 428                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 429         }
 430
 431         spin_lock_init(&cli->cl_mod_rpcs_lock);
 432         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 433         cli->cl_max_mod_rpcs_in_flight = 0;
 434         cli->cl_mod_rpcs_in_flight = 0;
 435         cli->cl_close_rpcs_in_flight = 0;
 436         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 437         cli->cl_mod_tag_bitmap = NULL;
 438
 439         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 440
 441         if (connect_op == MDS_CONNECT) {
 442                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 443                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 444                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 445                 if (cli->cl_mod_tag_bitmap == NULL)
 446                         GOTO(err, rc = -ENOMEM);
 447         }
 448
 449         rc = ldlm_get_ref();
 450         if (rc) {
 451                 CERROR("ldlm_get_ref failed: %d\n", rc);
 452                 GOTO(err, rc);
 453         }
 454
 455         ptlrpc_init_client(rq_portal, rp_portal, name,
 456                            &obddev->obd_ldlm_client);
 457
 458         imp = class_new_import(obddev);
 459         if (imp == NULL)
 460                 GOTO(err_ldlm, rc = -ENOENT);
 461         imp->imp_client = &obddev->obd_ldlm_client;
 462         imp->imp_connect_op = connect_op;
 463         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 464                LUSTRE_CFG_BUFLEN(lcfg, 1));
 465         class_import_put(imp);
 466
 467         if (lustre_cfg_buf(lcfg, 4)) {
 468                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 469
 470                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 471                         rc = -EINVAL;
 472                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 473                                obddev->obd_name, lustre_cfg_string(lcfg, 4),
 474                                rc);
 475                         GOTO(err_import, rc);
 476                 }
 477                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 478                 imp->imp_connection = &fake_conn;
 479         }
 480
 481         rc = client_import_add_conn(imp, &server_uuid, 1);
 482         if (rc) {
 483                 CERROR("can't add initial connection\n");
 484                 GOTO(err_import, rc);
 485         }
 486         imp->imp_connection = NULL;
 487
 488         cli->cl_import = imp;
 489         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 490         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 491
 492         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 493                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 494                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 495                                name, obddev->obd_name,
 496                                cli->cl_target_uuid.uuid);
 497                         spin_lock(&imp->imp_lock);
 498                         imp->imp_deactive = 1;
 499                         spin_unlock(&imp->imp_lock);
 500                 }
 501         }
 502
 503         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 504                                                    LDLM_NAMESPACE_CLIENT,
 505                                                    LDLM_NAMESPACE_GREEDY,
 506                                                    ns_type);
 507         if (obddev->obd_namespace == NULL) {
 508                 CERROR("Unable to create client namespace - %s\n",
 509                        obddev->obd_name);
 510                 GOTO(err_import, rc = -ENOMEM);
 511         }
 512
 513         RETURN(rc);
 514
 515 err_import:
 516         class_destroy_import(imp);
 517 err_ldlm:
 518         ldlm_put_ref();
 519 err:
 520         if (cli->cl_mod_tag_bitmap != NULL)
 521                 OBD_FREE(cli->cl_mod_tag_bitmap,
 522                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 523         cli->cl_mod_tag_bitmap = NULL;
 524         RETURN(rc);
 525
 526 }
 527 EXPORT_SYMBOL(client_obd_setup);
 528
 529 int client_obd_cleanup(struct obd_device *obddev)
 530 {
 531         struct client_obd *cli = &obddev->u.cli;
 532         ENTRY;
 533
 534         ldlm_namespace_free_post(obddev->obd_namespace);
 535         obddev->obd_namespace = NULL;
 536
 537         obd_cleanup_client_import(obddev);
 538         LASSERT(obddev->u.cli.cl_import == NULL);
 539
 540         ldlm_put_ref();
 541
 542         if (cli->cl_mod_tag_bitmap != NULL)
 543                 OBD_FREE(cli->cl_mod_tag_bitmap,
 544                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 545         cli->cl_mod_tag_bitmap = NULL;
 546
 547         RETURN(0);
 548 }
 549 EXPORT_SYMBOL(client_obd_cleanup);
 550
 551 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 552 int client_connect_import(const struct lu_env *env,
 553                           struct obd_export **exp,
 554                           struct obd_device *obd, struct obd_uuid *cluuid,
 555                           struct obd_connect_data *data, void *localdata)
 556 {
 557         struct client_obd       *cli    = &obd->u.cli;
 558         struct obd_import       *imp    = cli->cl_import;
 559         struct obd_connect_data *ocd;
 560         struct lustre_handle    conn    = { 0 };
 561         int                     rc;
 562         ENTRY;
 563
 564         *exp = NULL;
 565         down_write(&cli->cl_sem);
 566         if (cli->cl_conn_count > 0)
 567                 GOTO(out_sem, rc = -EALREADY);
 568
 569         rc = class_connect(&conn, obd, cluuid);
 570         if (rc)
 571                 GOTO(out_sem, rc);
 572
 573         cli->cl_conn_count++;
 574         *exp = class_conn2export(&conn);
 575
 576         LASSERT(obd->obd_namespace);
 577
 578         imp->imp_dlm_handle = conn;
 579         rc = ptlrpc_init_import(imp);
 580         if (rc != 0)
 581                 GOTO(out_ldlm, rc);
 582
 583         ocd = &imp->imp_connect_data;
 584         if (data) {
 585                 *ocd = *data;
 586                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 587                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 588         }
 589
 590         rc = ptlrpc_connect_import(imp);
 591         if (rc != 0) {
 592                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 593                 GOTO(out_ldlm, rc);
 594         }
 595         LASSERT(*exp != NULL && (*exp)->exp_connection);
 596
 597         if (data) {
 598                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 599                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 600                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 601                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 602                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 603         }
 604
 605         ptlrpc_pinger_add_import(imp);
 606
 607         EXIT;
 608
 609         if (rc) {
 610 out_ldlm:
 611                 cli->cl_conn_count--;
 612                 class_disconnect(*exp);
 613                 *exp = NULL;
 614         }
 615 out_sem:
 616         up_write(&cli->cl_sem);
 617
 618         return rc;
 619 }
 620 EXPORT_SYMBOL(client_connect_import);
 621
 622 int client_disconnect_export(struct obd_export *exp)
 623 {
 624         struct obd_device *obd = class_exp2obd(exp);
 625         struct client_obd *cli;
 626         struct obd_import *imp;
 627         int rc = 0, err;
 628         ENTRY;
 629
 630         if (!obd) {
 631                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 632                        exp, exp ? exp->exp_handle.h_cookie : -1);
 633                 RETURN(-EINVAL);
 634         }
 635
 636         cli = &obd->u.cli;
 637         imp = cli->cl_import;
 638
 639         down_write(&cli->cl_sem);
 640         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 641                 cli->cl_conn_count);
 642
 643         if (cli->cl_conn_count == 0) {
 644                 CERROR("disconnecting disconnected device (%s)\n",
 645                        obd->obd_name);
 646                 GOTO(out_disconnect, rc = -EINVAL);
 647         }
 648
 649         cli->cl_conn_count--;
 650         if (cli->cl_conn_count != 0)
 651                 GOTO(out_disconnect, rc = 0);
 652
 653         /* Mark import deactivated now, so we don't try to reconnect if any
 654          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 655          * fully deactivate the import, or that would drop all requests. */
 656         spin_lock(&imp->imp_lock);
 657         imp->imp_deactive = 1;
 658         spin_unlock(&imp->imp_lock);
 659
 660         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 661          * delete it regardless.  (It's safe to delete an import that was
 662          * never added.) */
 663         (void)ptlrpc_pinger_del_import(imp);
 664
 665         if (obd->obd_namespace != NULL) {
 666                 /* obd_force == local only */
 667                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 668                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 669                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 670         }
 671
 672         /* There's no need to hold sem while disconnecting an import,
 673          * and it may actually cause deadlock in GSS. */
 674         up_write(&cli->cl_sem);
 675         rc = ptlrpc_disconnect_import(imp, 0);
 676         down_write(&cli->cl_sem);
 677
 678         ptlrpc_invalidate_import(imp);
 679
 680         EXIT;
 681
 682 out_disconnect:
 683         /* Use server style - class_disconnect should be always called for
 684          * o_disconnect. */
 685         err = class_disconnect(exp);
 686         if (!rc && err)
 687                 rc = err;
 688
 689         up_write(&cli->cl_sem);
 690
 691         RETURN(rc);
 692 }
 693 EXPORT_SYMBOL(client_disconnect_export);
 694
 695 #ifdef HAVE_SERVER_SUPPORT
 696 int server_disconnect_export(struct obd_export *exp)
 697 {
 698         int rc;
 699         ENTRY;
 700
 701         /* Disconnect early so that clients can't keep using export. */
 702         rc = class_disconnect(exp);
 703         /* Close import to avoid sending any requests. */
 704         if (exp->exp_imp_reverse)
 705                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 706
 707         ldlm_bl_thread_wakeup();
 708
 709         /* complete all outstanding replies */
 710         spin_lock(&exp->exp_lock);
 711         while (!list_empty(&exp->exp_outstanding_replies)) {
 712                 struct ptlrpc_reply_state *rs =
 713                         list_entry(exp->exp_outstanding_replies.next,
 714                                        struct ptlrpc_reply_state, rs_exp_list);
 715                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 716
 717                 spin_lock(&svcpt->scp_rep_lock);
 718
 719                 list_del_init(&rs->rs_exp_list);
 720
 721                 spin_lock(&rs->rs_lock);
 722                 /* clear rs_convert_lock to make sure rs is handled and put */
 723                 rs->rs_convert_lock = 0;
 724                 ptlrpc_schedule_difficult_reply(rs);
 725                 spin_unlock(&rs->rs_lock);
 726
 727                 spin_unlock(&svcpt->scp_rep_lock);
 728         }
 729         spin_unlock(&exp->exp_lock);
 730
 731         RETURN(rc);
 732 }
 733 EXPORT_SYMBOL(server_disconnect_export);
 734
 735 /* --------------------------------------------------------------------------
 736  * from old lib/target.c
 737  * -------------------------------------------------------------------------- */
 738
 739 static int target_handle_reconnect(struct lustre_handle *conn,
 740                                    struct obd_export *exp,
 741                                    struct obd_uuid *cluuid)
 742 {
 743         struct obd_device *target;
 744         struct lustre_handle *hdl;
 745         cfs_time_t now;
 746         cfs_time_t deadline;
 747         int timeout;
 748         int rc = 0;
 749         ENTRY;
 750
 751         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 752         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 753                 conn->cookie = exp->exp_handle.h_cookie;
 754                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 755                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
 756                 RETURN(0);
 757         }
 758
 759         target = exp->exp_obd;
 760
 761         /* Might be a re-connect after a partition. */
 762         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 763                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 764                               "with handle %#llx. Rejecting client "
 765                               "with the same UUID trying to reconnect "
 766                               "with handle %#llx\n", target->obd_name,
 767                               obd_uuid2str(&exp->exp_client_uuid),
 768                               obd_export_nid2str(exp),
 769                               hdl->cookie, conn->cookie);
 770                 memset(conn, 0, sizeof *conn);
 771                 /* target_handle_connect() treats EALREADY and
 772                  * -EALREADY differently.  -EALREADY is an error
 773                  * (same UUID, different handle). */
 774                 RETURN(-EALREADY);
 775         }
 776
 777         if (!target->obd_recovering) {
 778                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 779                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 780                         obd_export_nid2str(exp));
 781                 GOTO(out_already, rc);
 782         }
 783
 784         now = cfs_time_current();
 785         deadline = target->obd_recovery_timer.expires;
 786         if (cfs_time_before(now, deadline)) {
 787                 struct target_distribute_txn_data *tdtd =
 788                                         class_exp2tgt(exp)->lut_tdtd;
 789                 int size = 0;
 790                 int count = 0;
 791                 char *buf = NULL;
 792
 793                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
 794                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 795                         buf = tdtd->tdtd_show_update_logs_retrievers(
 796                                 tdtd->tdtd_show_retrievers_cbdata,
 797                                 &size, &count);
 798
 799                 if (count > 0)
 800                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 801                                       "%d:%.02d. It is due to DNE recovery "
 802                                       "failed/stuck on the %d MDT(s):%s. "
 803                                       "Please wait until all MDTs recovered "
 804                                       "or abort the recovery by force.\n",
 805                                       target->obd_name, timeout / 60,
 806                                       timeout % 60, count,
 807                                       buf ? buf : "unknown (not enough RAM)");
 808                 else
 809                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 810                                       "%d:%.02d. If you do not want to wait "
 811                                       "more, please abort the recovery by "
 812                                       "force.\n", target->obd_name,
 813                                       timeout / 60, timeout % 60);
 814
 815                 if (buf != NULL)
 816                         OBD_FREE(buf, size);
 817         } else {
 818                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
 819                 LCONSOLE_WARN("%s: Recovery already passed deadline"
 820                         " %d:%.02d, It is most likely due to DNE"
 821                         " recovery is failed or stuck, please wait a"
 822                         " few more minutes or abort the recovery.\n",
 823                         target->obd_name, timeout / 60, timeout % 60);
 824         }
 825
 826 out_already:
 827         conn->cookie = exp->exp_handle.h_cookie;
 828         /* target_handle_connect() treats EALREADY and
 829          * -EALREADY differently.  EALREADY means we are
 830          * doing a valid reconnect from the same client. */
 831         RETURN(EALREADY);
 832 }
 833
 834 static void
 835 check_and_start_recovery_timer(struct obd_device *obd,
 836                                struct ptlrpc_request *req, int new_client);
 837
 838 /**
 839  * update flags for import during reconnect process
 840  */
 841 static int rev_import_flags_update(struct obd_import *revimp,
 842                                    struct ptlrpc_request *req)
 843 {
 844         int rc;
 845         struct obd_connect_data *data;
 846
 847         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 848
 849         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 850                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 851         else
 852                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 853
 854         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 855
 856         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 857         if (rc) {
 858                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 859                         revimp->imp_client->cli_name,
 860                         libcfs_id2str(req->rq_peer), rc);
 861                 return rc;
 862         }
 863
 864         return 0;
 865 }
 866
 867 /**
 868  * Allocate a new reverse import for an export.
 869  *
 870  * \retval -errno in case error hit
 871  * \retval 0 if reverse import correctly init
 872  **/
 873 int rev_import_init(struct obd_export *export)
 874 {
 875         struct obd_device *obd = export->exp_obd;
 876         struct obd_import *revimp;
 877
 878         LASSERT(export->exp_imp_reverse == NULL);
 879
 880         revimp = class_new_import(obd);
 881         if (revimp == NULL)
 882                 return -ENOMEM;
 883
 884         revimp->imp_remote_handle.cookie = 0ULL;
 885         revimp->imp_client = &obd->obd_ldlm_client;
 886         revimp->imp_dlm_fake = 1;
 887
 888         /* it is safe to connect import in new state as no sends possible */
 889         spin_lock(&export->exp_lock);
 890         export->exp_imp_reverse = revimp;
 891         spin_unlock(&export->exp_lock);
 892         class_import_put(revimp);
 893
 894         return 0;
 895 }
 896 EXPORT_SYMBOL(rev_import_init);
 897
 898 /**
 899  * Handle reconnect for an export.
 900  *
 901  * \param exp export to handle reconnect process
 902  * \param req client reconnect request
 903  *
 904  * \retval -rc in case securitfy flavor can't be changed
 905  * \retval 0 in case none problems
 906  */
 907 static int rev_import_reconnect(struct obd_export *exp,
 908                                 struct ptlrpc_request *req)
 909 {
 910         struct obd_import *revimp = exp->exp_imp_reverse;
 911         struct lustre_handle *lh;
 912         int rc;
 913
 914         /* avoid sending a request until import flags are changed */
 915         ptlrpc_import_enter_resend(revimp);
 916
 917         if (revimp->imp_connection != NULL)
 918                 ptlrpc_connection_put(revimp->imp_connection);
 919
 920         /*
 921          * client from recovery don't have a handle so we need to take from
 922          * request. it may produce situation when wrong client connected
 923          * to recovery as we trust a client uuid
 924          */
 925         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 926         revimp->imp_remote_handle = *lh;
 927
 928         /* unknown versions will be caught in
 929          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 930         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 931
 932         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 933
 934         rc = rev_import_flags_update(revimp, req);
 935         if (rc != 0) {
 936                 /* it is safe to still be in RECOVERY phase as we are not able
 937                  * to setup correct security flavor so requests are not able to
 938                  * be delivered correctly */
 939                 return rc;
 940         }
 941
 942         /* resend all rpc's via new connection */
 943         return ptlrpc_import_recovery_state_machine(revimp);
 944 }
 945
 946 int target_handle_connect(struct ptlrpc_request *req)
 947 {
 948         struct obd_device *target = NULL;
 949         struct obd_export *export = NULL;
 950         /* connect handle - filled from target_handle_reconnect in
 951          * reconnect case */
 952         struct lustre_handle conn;
 953         struct lustre_handle *tmp;
 954         struct obd_uuid cluuid;
 955         char *str;
 956         int rc = 0;
 957         char *target_start;
 958         int target_len;
 959         bool     mds_conn = false, lw_client = false, initial_conn = false;
 960         bool     mds_mds_conn = false;
 961         bool     new_mds_mds_conn = false;
 962         struct obd_connect_data *data, *tmpdata;
 963         int size, tmpsize;
 964         lnet_nid_t *client_nid = NULL;
 965         ENTRY;
 966
 967         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 968
 969         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 970         if (str == NULL) {
 971                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 972                 GOTO(out, rc = -EINVAL);
 973         }
 974
 975         target = class_dev_by_str(str);
 976         if (!target) {
 977                 deuuidify(str, NULL, &target_start, &target_len);
 978                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 979                                    "from %s (no target). If you are running "
 980                                    "an HA pair check that the target is "
 981                                    "mounted on the other server.\n", str,
 982                                    libcfs_nid2str(req->rq_peer.nid));
 983                 GOTO(out, rc = -ENODEV);
 984         }
 985
 986         spin_lock(&target->obd_dev_lock);
 987
 988         target->obd_conn_inprogress++;
 989
 990         if (target->obd_stopping || !target->obd_set_up) {
 991                 spin_unlock(&target->obd_dev_lock);
 992
 993                 deuuidify(str, NULL, &target_start, &target_len);
 994                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 995                               target_len, target_start,
 996                               libcfs_nid2str(req->rq_peer.nid),
 997                               (target->obd_stopping ?
 998                                "stopping" : "not set up"));
 999                 GOTO(out, rc = -ENODEV);
1000         }
1001
1002         if (target->obd_no_conn) {
1003                 spin_unlock(&target->obd_dev_lock);
1004
1005                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1006                                "from %s\n", target->obd_name,
1007                                libcfs_nid2str(req->rq_peer.nid));
1008                 GOTO(out, rc = -EAGAIN);
1009         }
1010
1011         spin_unlock(&target->obd_dev_lock);
1012
1013         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1014         if (str == NULL) {
1015                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1016                 GOTO(out, rc = -EINVAL);
1017         }
1018
1019         obd_str2uuid(&cluuid, str);
1020
1021         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1022         if (tmp == NULL)
1023                 GOTO(out, rc = -EPROTO);
1024
1025         conn = *tmp;
1026
1027         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1028                                     RCL_CLIENT);
1029         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1030         if (!data)
1031                 GOTO(out, rc = -EPROTO);
1032
1033         rc = req_capsule_server_pack(&req->rq_pill);
1034         if (rc)
1035                 GOTO(out, rc);
1036
1037 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1038         /* Don't allow clients to connect that are using old 1.8 format
1039          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1040          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1041          * FULL20 flag should be set on all connections since 2.0, but no
1042          * longer affects behaviour.
1043          *
1044          * Later this check will be disabled and the flag can be retired
1045          * completely once interop with 3.0 is no longer needed.
1046          */
1047         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1048                 GOTO(out, rc = -EPROTO);
1049 #endif
1050
1051         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1052                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1053                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1054                     data->ocd_version > LUSTRE_VERSION_CODE +
1055                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1056                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1057                                   "libclient connection attempt",
1058                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1059                                   "old" : "new",
1060                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1061                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1062                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1063                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1064                         data = req_capsule_server_sized_get(&req->rq_pill,
1065                                                             &RMF_CONNECT_DATA,
1066                                     offsetof(typeof(*data), ocd_version) +
1067                                              sizeof(data->ocd_version));
1068                         if (data) {
1069                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1070                                 data->ocd_version = LUSTRE_VERSION_CODE;
1071                         }
1072                         GOTO(out, rc = -EPROTO);
1073                 }
1074         }
1075
1076         /* Note: lw_client is needed in MDS-MDS failover during update log
1077          * processing, so we needs to allow lw_client to be connected at
1078          * anytime, instead of only the initial connection */
1079         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1080
1081         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1082                 initial_conn = true;
1083                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1084                 mds_mds_conn = (data->ocd_connect_flags &
1085                                 OBD_CONNECT_MDS_MDS) != 0;
1086
1087                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1088                  * for Imperative Recovery connection from MGC to MGS.
1089                  *
1090                  * Via check OBD_CONNECT_FID, we can distinguish whether
1091                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1092                  * MGC or MDT. */
1093                 if (!lw_client &&
1094                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1095                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1096                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1097                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1098                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1099                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1100
1101                         /* We do not support the MDT-MDT interoperations with
1102                          * different version MDT because of protocol changes. */
1103                         if (unlikely(major != LUSTRE_MAJOR ||
1104                                      minor != LUSTRE_MINOR ||
1105                                      abs(patch - LUSTRE_PATCH) > 3)) {
1106                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1107                                         "connection from different version MDT "
1108                                         "(%d.%d.%d.%d) %s %s\n",
1109                                         target->obd_name, LUSTRE_MAJOR,
1110                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1111                                         major, minor, patch,
1112                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1113                                         libcfs_nid2str(req->rq_peer.nid), str);
1114
1115                                 GOTO(out, rc = -EPROTO);
1116                         }
1117                 }
1118         }
1119
1120         /* lctl gets a backstage, all-access pass. */
1121         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1122                 goto dont_check_exports;
1123
1124         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1125         if (!export)
1126                 goto no_export;
1127
1128         /* We've found an export in the hash. */
1129
1130         spin_lock(&export->exp_lock);
1131
1132         if (export->exp_connecting) { /* bug 9635, et. al. */
1133                 spin_unlock(&export->exp_lock);
1134                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1135                               export->exp_obd->obd_name, export,
1136                               libcfs_nid2str(req->rq_peer.nid));
1137                 class_export_put(export);
1138                 export = NULL;
1139                 rc = -EALREADY;
1140         } else if ((mds_conn || (lw_client && initial_conn) ||
1141                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1142                    export->exp_connection != NULL) {
1143                 spin_unlock(&export->exp_lock);
1144                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1145                         /* MDS or LWP reconnected after failover. */
1146                         LCONSOLE_WARN("%s: Received %s connection from "
1147                             "%s, removing former export from %s\n",
1148                             target->obd_name, mds_conn ? "MDS" : "LWP",
1149                             libcfs_nid2str(req->rq_peer.nid),
1150                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1151                 } else {
1152                         /* New MDS connection from the same NID. */
1153                         LCONSOLE_WARN("%s: Received new %s connection from "
1154                                 "%s, removing former export from same NID\n",
1155                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1156                                 libcfs_nid2str(req->rq_peer.nid));
1157                 }
1158
1159                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1160                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1161                         /* Because exports between MDTs will always be
1162                          * kept, let's do not fail such export if they
1163                          * come from the same NID, otherwise it might
1164                          * cause eviction between MDTs, which might
1165                          * cause namespace inconsistency */
1166                         spin_lock(&export->exp_lock);
1167                         export->exp_connecting = 1;
1168                         spin_unlock(&export->exp_lock);
1169                         conn.cookie = export->exp_handle.h_cookie;
1170                         rc = EALREADY;
1171                 } else {
1172                         class_fail_export(export);
1173                         class_export_put(export);
1174                         export = NULL;
1175                         rc = 0;
1176                 }
1177         } else if (export->exp_connection != NULL && initial_conn &&
1178                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1179                 spin_unlock(&export->exp_lock);
1180                 /* In MDS failover we have static UUID but NID can change. */
1181                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1182                               "existing nid %s is already connected\n",
1183                               target->obd_name, cluuid.uuid,
1184                               libcfs_nid2str(req->rq_peer.nid),
1185                               libcfs_nid2str(
1186                                       export->exp_connection->c_peer.nid));
1187                 rc = -EALREADY;
1188                 class_export_put(export);
1189                 export = NULL;
1190         } else {
1191                 export->exp_connecting = 1;
1192                 spin_unlock(&export->exp_lock);
1193                 LASSERT(export->exp_obd == target);
1194
1195                 rc = target_handle_reconnect(&conn, export, &cluuid);
1196         }
1197
1198         /* If we found an export, we already unlocked. */
1199         if (!export) {
1200 no_export:
1201                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1202         } else if (req->rq_export == NULL &&
1203                    atomic_read(&export->exp_rpc_count) > 0) {
1204                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1205                               "still busy with %d references\n",
1206                               target->obd_name, cluuid.uuid,
1207                               libcfs_nid2str(req->rq_peer.nid),
1208                               atomic_read(&export->exp_refcount));
1209                 GOTO(out, rc = -EBUSY);
1210         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1211                 if (!strstr(cluuid.uuid, "mdt"))
1212                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1213                                       "known client %s (at %s) because it "
1214                                       "is indicating it is a new client",
1215                                       target->obd_name, cluuid.uuid,
1216                                       libcfs_nid2str(req->rq_peer.nid));
1217                 GOTO(out, rc = -EALREADY);
1218         } else {
1219                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1220         }
1221
1222         if (rc < 0) {
1223                 GOTO(out, rc);
1224         }
1225
1226         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1227                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1228                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1229                export, ktime_get_real_seconds(),
1230                export ? export->exp_last_request_time : 0);
1231
1232         /* If this is the first time a client connects, reset the recovery
1233          * timer. Discard lightweight connections which might be local. */
1234         if (!lw_client && rc == 0 && target->obd_recovering)
1235                 check_and_start_recovery_timer(target, req, export == NULL);
1236
1237         /* We want to handle EALREADY but *not* -EALREADY from
1238          * target_handle_reconnect(), return reconnection state in a flag. */
1239         if (rc == EALREADY) {
1240                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1241                 rc = 0;
1242         } else {
1243                 LASSERT(rc == 0);
1244         }
1245
1246         /* Tell the client if we support replayable requests. */
1247         if (target->obd_replayable)
1248                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1249         client_nid = &req->rq_peer.nid;
1250
1251         if (export == NULL) {
1252                 /* allow lightweight connections during recovery */
1253                 /* allow "new" MDT to be connected during recovery, since we
1254                  * need retrieve recovery update records from it */
1255                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1256                         cfs_time_t t;
1257                         int     c; /* connected */
1258                         int     i; /* in progress */
1259                         int     k; /* known */
1260                         int     s; /* stale/evicted */
1261
1262                         c = atomic_read(&target->obd_connected_clients);
1263                         i = atomic_read(&target->obd_lock_replay_clients);
1264                         k = target->obd_max_recoverable_clients;
1265                         s = target->obd_stale_clients;
1266                         t = target->obd_recovery_timer.expires;
1267                         t = cfs_time_sub(t, cfs_time_current());
1268                         t = cfs_duration_sec(t);
1269                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1270                                       "(at %s), waiting for %d known clients "
1271                                       "(%d recovered, %d in progress, and %d "
1272                                       "evicted) to recover in %d:%.02d\n",
1273                                       target->obd_name, cluuid.uuid,
1274                                       libcfs_nid2str(req->rq_peer.nid), k,
1275                                       c - i, i, s, (int)t / 60,
1276                                       (int)t % 60);
1277                         rc = -EBUSY;
1278                 } else {
1279 dont_check_exports:
1280                         rc = obd_connect(req->rq_svc_thread->t_env,
1281                                          &export, target, &cluuid, data,
1282                                          client_nid);
1283                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1284                                 lustre_msg_add_op_flags(req->rq_repmsg,
1285                                                         MSG_CONNECT_RECOVERING);
1286                         if (rc == 0) {
1287                                 conn.cookie = export->exp_handle.h_cookie;
1288                                 rc = rev_import_init(export);
1289                         }
1290
1291                         if (mds_mds_conn)
1292                                 new_mds_mds_conn = true;
1293                 }
1294         } else {
1295                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1296                                    export, target, &cluuid, data, client_nid);
1297         }
1298         if (rc)
1299                 GOTO(out, rc);
1300
1301         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1302         data->ocd_instance = target->u.obt.obt_instance;
1303
1304         /* Return only the parts of obd_connect_data that we understand, so the
1305          * client knows that we don't understand the rest. */
1306         if (data) {
1307                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1308                                                RCL_SERVER);
1309                 tmpdata = req_capsule_server_get(&req->rq_pill,
1310                                                  &RMF_CONNECT_DATA);
1311                 /* Don't use struct assignment here, because the client reply
1312                  * buffer may be smaller/larger than the local struct
1313                  * obd_connect_data. */
1314                 memcpy(tmpdata, data, min(tmpsize, size));
1315         }
1316
1317         /* If the client and the server are the same node, we will already
1318          * have an export that really points to the client's DLM export,
1319          * because we have a shared handles table.
1320          *
1321          * XXX this will go away when shaver stops sending the "connect" handle
1322          * in the real "remote handle" field of the request --phik 24 Apr 2003
1323          */
1324         ptlrpc_request_change_export(req, export);
1325
1326         spin_lock(&export->exp_lock);
1327         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1328                 spin_unlock(&export->exp_lock);
1329                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1330                        "or equal conn_cnt: %d >= %d\n",
1331                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1332                        export->exp_conn_cnt,
1333                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1334
1335                 GOTO(out, rc = -EALREADY);
1336         }
1337         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1338         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1339
1340         /* Don't evict liblustre clients for not pinging. */
1341         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1342                 export->exp_libclient = 1;
1343                 spin_unlock(&export->exp_lock);
1344
1345                 spin_lock(&target->obd_dev_lock);
1346                 list_del_init(&export->exp_obd_chain_timed);
1347                 spin_unlock(&target->obd_dev_lock);
1348         } else {
1349                 spin_unlock(&export->exp_lock);
1350         }
1351
1352         if (export->exp_connection != NULL) {
1353                 /* Check to see if connection came from another NID. */
1354                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1355                     !hlist_unhashed(&export->exp_nid_hash))
1356                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1357                                      &export->exp_connection->c_peer.nid,
1358                                      &export->exp_nid_hash);
1359
1360                 ptlrpc_connection_put(export->exp_connection);
1361         }
1362
1363         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1364                                                        req->rq_self,
1365                                                        &cluuid);
1366         if (hlist_unhashed(&export->exp_nid_hash))
1367                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1368                              &export->exp_connection->c_peer.nid,
1369                              &export->exp_nid_hash);
1370
1371         lustre_msg_set_handle(req->rq_repmsg, &conn);
1372
1373         rc = rev_import_reconnect(export, req);
1374         if (rc != 0)
1375                 GOTO(out, rc);
1376
1377         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1378                 int has_transno;
1379                 __u64 transno = data->ocd_transno;
1380
1381                 spin_lock(&export->exp_lock);
1382                 /* possible race with class_disconnect_stale_exports,
1383                  * export may be already in the eviction process */
1384                 if (export->exp_failed) {
1385                         spin_unlock(&export->exp_lock);
1386                         GOTO(out, rc = -ENODEV);
1387                 }
1388                 export->exp_in_recovery = 1;
1389                 export->exp_req_replay_needed = 1;
1390                 export->exp_lock_replay_needed = 1;
1391                 spin_unlock(&export->exp_lock);
1392
1393                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1394                                  MSG_CONNECT_TRANSNO);
1395                 if (has_transno && transno == 0)
1396                         CWARN("Connect with zero transno!\n");
1397
1398                 if (has_transno && transno > 0 &&
1399                     transno < target->obd_next_recovery_transno &&
1400                     transno > target->obd_last_committed) {
1401                         /* Another way is to use cmpxchg() to be lock-free. */
1402                         spin_lock(&target->obd_recovery_task_lock);
1403                         if (transno < target->obd_next_recovery_transno)
1404                                 target->obd_next_recovery_transno = transno;
1405                         spin_unlock(&target->obd_recovery_task_lock);
1406                 }
1407
1408                 atomic_inc(&target->obd_req_replay_clients);
1409                 atomic_inc(&target->obd_lock_replay_clients);
1410                 /* Note: MDS-MDS connection is allowed to be connected during
1411                  * recovery, no matter if the exports needs to be recoveried.
1412                  * Because we need retrieve updates logs from all other MDTs.
1413                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1414                  * also needs to be increased to match other recovery checking
1415                  * condition. */
1416                 if (new_mds_mds_conn)
1417                         target->obd_max_recoverable_clients++;
1418                 if (atomic_inc_return(&target->obd_connected_clients) ==
1419                     target->obd_max_recoverable_clients)
1420                         wake_up(&target->obd_next_transno_waitq);
1421         }
1422
1423         /* Tell the client we're in recovery, when client is involved in it. */
1424         if (target->obd_recovering && !lw_client)
1425                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1426
1427 out:
1428         if (export) {
1429                 spin_lock(&export->exp_lock);
1430                 export->exp_connecting = 0;
1431                 spin_unlock(&export->exp_lock);
1432
1433                 class_export_put(export);
1434         }
1435         if (target != NULL) {
1436                 spin_lock(&target->obd_dev_lock);
1437                 target->obd_conn_inprogress--;
1438                 spin_unlock(&target->obd_dev_lock);
1439                 class_decref(target, "find", current);
1440         }
1441         req->rq_status = rc;
1442         RETURN(rc);
1443 }
1444
1445 int target_handle_disconnect(struct ptlrpc_request *req)
1446 {
1447         int rc;
1448         ENTRY;
1449
1450         rc = req_capsule_server_pack(&req->rq_pill);
1451         if (rc)
1452                 RETURN(rc);
1453
1454         /* Keep the rq_export around so we can send the reply. */
1455         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1456
1457         RETURN(0);
1458 }
1459
1460 void target_destroy_export(struct obd_export *exp)
1461 {
1462         struct obd_import       *imp = NULL;
1463         /* exports created from last_rcvd data, and "fake"
1464            exports created by lctl don't have an import */
1465         spin_lock(&exp->exp_lock);
1466         if (exp->exp_imp_reverse != NULL) {
1467                 imp = exp->exp_imp_reverse;
1468                 exp->exp_imp_reverse = NULL;
1469         }
1470         spin_unlock(&exp->exp_lock);
1471         if (imp != NULL)
1472                 client_destroy_import(imp);
1473
1474         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1475         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1476         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1477         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1478 }
1479 EXPORT_SYMBOL(target_destroy_export);
1480
1481 /*
1482  * Recovery functions
1483  */
1484 static void target_request_copy_get(struct ptlrpc_request *req)
1485 {
1486         class_export_rpc_inc(req->rq_export);
1487         LASSERT(list_empty(&req->rq_list));
1488         INIT_LIST_HEAD(&req->rq_replay_list);
1489
1490         /* Increase refcount to keep request in queue. */
1491         atomic_inc(&req->rq_refcount);
1492         /* Let export know it has replays to be handled. */
1493         atomic_inc(&req->rq_export->exp_replay_count);
1494 }
1495
1496 static void target_request_copy_put(struct ptlrpc_request *req)
1497 {
1498         LASSERT(list_empty(&req->rq_replay_list));
1499         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1500
1501         atomic_dec(&req->rq_export->exp_replay_count);
1502         class_export_rpc_dec(req->rq_export);
1503         ptlrpc_server_drop_request(req);
1504 }
1505
1506 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1507 {
1508         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1509         struct obd_export     *exp = req->rq_export;
1510         struct ptlrpc_request *reqiter;
1511         struct ptlrpc_request *dup_req = NULL;
1512         int                    dup = 0;
1513
1514         LASSERT(exp);
1515
1516         spin_lock(&exp->exp_lock);
1517         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1518                                 rq_replay_list) {
1519                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1520                         dup_req = reqiter;
1521                         dup = 1;
1522                         break;
1523                 }
1524         }
1525
1526         if (dup) {
1527                 /* We expect it with RESENT and REPLAY flags. */
1528                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1529                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1530                         CERROR("invalid flags %x of resent replay\n",
1531                                lustre_msg_get_flags(req->rq_reqmsg));
1532
1533                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1534                         __u32 new_conn;
1535
1536                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1537                         if (new_conn >
1538                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1539                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1540                                                         new_conn);
1541                 }
1542         } else {
1543                 list_add_tail(&req->rq_replay_list,
1544                                   &exp->exp_req_replay_queue);
1545         }
1546
1547         spin_unlock(&exp->exp_lock);
1548         return dup;
1549 }
1550
1551 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1552 {
1553         LASSERT(!list_empty(&req->rq_replay_list));
1554         LASSERT(req->rq_export);
1555
1556         spin_lock(&req->rq_export->exp_lock);
1557         list_del_init(&req->rq_replay_list);
1558         spin_unlock(&req->rq_export->exp_lock);
1559 }
1560
1561 static void target_finish_recovery(struct lu_target *lut)
1562 {
1563         struct obd_device *obd = lut->lut_obd;
1564         ENTRY;
1565
1566         /* Only log a recovery message when recovery has occurred. */
1567         if (obd->obd_recovery_start) {
1568                 time64_t now = ktime_get_real_seconds();
1569                 time64_t elapsed_time;
1570
1571                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1572                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1573                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1574                         (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1575                         obd->obd_max_recoverable_clients,
1576                         atomic_read(&obd->obd_connected_clients),
1577                         obd->obd_stale_clients,
1578                         obd->obd_stale_clients == 1 ? "was" : "were");
1579         }
1580
1581         ldlm_reprocess_recovery_done(obd->obd_namespace);
1582         spin_lock(&obd->obd_recovery_task_lock);
1583         if (!list_empty(&obd->obd_req_replay_queue) ||
1584             !list_empty(&obd->obd_lock_replay_queue) ||
1585             !list_empty(&obd->obd_final_req_queue)) {
1586                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1587                        obd->obd_name,
1588                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1589                        list_empty(&obd->obd_lock_replay_queue) ? \
1590                                "" : "lock ",
1591                        list_empty(&obd->obd_final_req_queue) ? \
1592                                "" : "final ");
1593                 spin_unlock(&obd->obd_recovery_task_lock);
1594                 LBUG();
1595         }
1596         spin_unlock(&obd->obd_recovery_task_lock);
1597
1598         obd->obd_recovery_end = ktime_get_real_seconds();
1599
1600         /* When recovery finished, cleanup orphans on MDS and OST. */
1601         if (OBT(obd) && OBP(obd, postrecov)) {
1602                 int rc = OBP(obd, postrecov)(obd);
1603                 if (rc < 0)
1604                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1605                                       obd->obd_name, rc);
1606         }
1607         EXIT;
1608 }
1609
1610 static void abort_req_replay_queue(struct obd_device *obd)
1611 {
1612         struct ptlrpc_request *req, *n;
1613         struct list_head abort_list;
1614
1615         INIT_LIST_HEAD(&abort_list);
1616         spin_lock(&obd->obd_recovery_task_lock);
1617         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1618         spin_unlock(&obd->obd_recovery_task_lock);
1619         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1620                 DEBUG_REQ(D_WARNING, req, "aborted:");
1621                 req->rq_status = -ENOTCONN;
1622                 if (ptlrpc_error(req)) {
1623                         DEBUG_REQ(D_ERROR, req,
1624                                   "failed abort_req_reply; skipping");
1625                 }
1626                 target_exp_dequeue_req_replay(req);
1627                 target_request_copy_put(req);
1628         }
1629 }
1630
1631 static void abort_lock_replay_queue(struct obd_device *obd)
1632 {
1633         struct ptlrpc_request *req, *n;
1634         struct list_head abort_list;
1635
1636         INIT_LIST_HEAD(&abort_list);
1637         spin_lock(&obd->obd_recovery_task_lock);
1638         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1639         spin_unlock(&obd->obd_recovery_task_lock);
1640         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1641                 DEBUG_REQ(D_ERROR, req, "aborted:");
1642                 req->rq_status = -ENOTCONN;
1643                 if (ptlrpc_error(req)) {
1644                         DEBUG_REQ(D_ERROR, req,
1645                                   "failed abort_lock_reply; skipping");
1646                 }
1647                 target_request_copy_put(req);
1648         }
1649 }
1650
1651 /* Called from a cleanup function if the device is being cleaned up
1652    forcefully.  The exports should all have been disconnected already,
1653    the only thing left to do is
1654      - clear the recovery flags
1655      - cancel the timer
1656      - free queued requests and replies, but don't send replies
1657    Because the obd_stopping flag is set, no new requests should be received.
1658
1659 */
1660 void target_cleanup_recovery(struct obd_device *obd)
1661 {
1662         struct ptlrpc_request *req, *n;
1663         struct list_head clean_list;
1664
1665         INIT_LIST_HEAD(&clean_list);
1666         spin_lock(&obd->obd_dev_lock);
1667         if (!obd->obd_recovering) {
1668                 spin_unlock(&obd->obd_dev_lock);
1669                 EXIT;
1670                 return;
1671         }
1672         obd->obd_recovering = obd->obd_abort_recovery = 0;
1673         spin_unlock(&obd->obd_dev_lock);
1674
1675         spin_lock(&obd->obd_recovery_task_lock);
1676         target_cancel_recovery_timer(obd);
1677         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1678         spin_unlock(&obd->obd_recovery_task_lock);
1679
1680         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1681                 LASSERT(req->rq_reply_state == NULL);
1682                 target_exp_dequeue_req_replay(req);
1683                 target_request_copy_put(req);
1684         }
1685
1686         spin_lock(&obd->obd_recovery_task_lock);
1687         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1688         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1689         spin_unlock(&obd->obd_recovery_task_lock);
1690
1691         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1692                 LASSERT(req->rq_reply_state == NULL);
1693                 target_request_copy_put(req);
1694         }
1695
1696         EXIT;
1697 }
1698 EXPORT_SYMBOL(target_cleanup_recovery);
1699
1700 /* obd_recovery_task_lock should be held */
1701 void target_cancel_recovery_timer(struct obd_device *obd)
1702 {
1703         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1704         del_timer(&obd->obd_recovery_timer);
1705 }
1706
1707 static void target_start_recovery_timer(struct obd_device *obd)
1708 {
1709         if (obd->obd_recovery_start != 0)
1710                 return;
1711
1712         spin_lock(&obd->obd_dev_lock);
1713         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1714                 spin_unlock(&obd->obd_dev_lock);
1715                 return;
1716         }
1717
1718         LASSERT(obd->obd_recovery_timeout != 0);
1719
1720         if (obd->obd_recovery_start != 0) {
1721                 spin_unlock(&obd->obd_dev_lock);
1722                 return;
1723         }
1724
1725         mod_timer(&obd->obd_recovery_timer,
1726                   cfs_time_shift(obd->obd_recovery_timeout));
1727         obd->obd_recovery_start = ktime_get_real_seconds();
1728         spin_unlock(&obd->obd_dev_lock);
1729
1730         LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1731                       obd->obd_name,
1732                       obd->obd_recovery_timeout / 60,
1733                       obd->obd_recovery_timeout % 60,
1734                       obd->obd_max_recoverable_clients,
1735                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1736                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1737 }
1738
1739 /**
1740  * extend recovery window.
1741  *
1742  * if @extend is true, extend recovery window to have @drt remaining at least;
1743  * otherwise, make sure the recovery timeout value is not less than @drt.
1744  */
1745 static void extend_recovery_timer(struct obd_device *obd, int drt,
1746                                   bool extend)
1747 {
1748         time64_t now;
1749         time64_t end;
1750         time64_t left;
1751         time64_t to;
1752
1753         spin_lock(&obd->obd_dev_lock);
1754         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1755                 spin_unlock(&obd->obd_dev_lock);
1756                 return;
1757         }
1758         LASSERT(obd->obd_recovery_start != 0);
1759
1760         now = ktime_get_real_seconds();
1761         to = obd->obd_recovery_timeout;
1762         end = obd->obd_recovery_start + to;
1763         left = end - now;
1764
1765         if (extend && (drt > left)) {
1766                 to += drt - left;
1767         } else if (!extend && (drt > to)) {
1768                 to = drt;
1769         }
1770
1771         if (to > obd->obd_recovery_time_hard) {
1772                 to = obd->obd_recovery_time_hard;
1773                 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1774                       obd->obd_name, to, extend);
1775         }
1776
1777         if (obd->obd_recovery_timeout < to) {
1778                 obd->obd_recovery_timeout = to;
1779                 end = obd->obd_recovery_start + to;
1780                 mod_timer(&obd->obd_recovery_timer,
1781                           cfs_time_shift(end - now));
1782         }
1783         spin_unlock(&obd->obd_dev_lock);
1784
1785         CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1786                 obd->obd_name, (s64)(end - now));
1787 }
1788
1789 /* Reset the timer with each new client connection */
1790 /*
1791  * This timer is actually reconnect_timer, which is for making sure
1792  * the total recovery window is at least as big as my reconnect
1793  * attempt timing. So the initial recovery time_out will be set to
1794  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1795  * from client is bigger than this, then the recovery time_out will
1796  * be extended to make sure the client could be reconnected, in the
1797  * process, the timeout from the new client should be ignored.
1798  */
1799
1800 static void
1801 check_and_start_recovery_timer(struct obd_device *obd,
1802                                struct ptlrpc_request *req,
1803                                int new_client)
1804 {
1805         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1806         struct obd_device_target *obt = &obd->u.obt;
1807
1808         if (!new_client && service_time)
1809                 /* Teach server about old server's estimates, as first guess
1810                  * at how long new requests will take. */
1811                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1812                             service_time);
1813
1814         target_start_recovery_timer(obd);
1815
1816         /* Convert the service time to RPC timeout,
1817          * and reuse service_time to limit stack usage. */
1818         service_time = at_est2timeout(service_time);
1819
1820         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1821             service_time < at_extra)
1822                 service_time = at_extra;
1823
1824         /* We expect other clients to timeout within service_time, then try
1825          * to reconnect, then try the failover server.  The max delay between
1826          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1827         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1828
1829         LASSERT(obt->obt_magic == OBT_MAGIC);
1830         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1831         if (service_time > obd->obd_recovery_timeout && !new_client)
1832                 extend_recovery_timer(obd, service_time, false);
1833 }
1834
1835 /** Health checking routines */
1836 static inline int exp_connect_healthy(struct obd_export *exp)
1837 {
1838         return (exp->exp_in_recovery);
1839 }
1840
1841 /** if export done req_replay or has replay in queue */
1842 static inline int exp_req_replay_healthy(struct obd_export *exp)
1843 {
1844         return (!exp->exp_req_replay_needed ||
1845                 atomic_read(&exp->exp_replay_count) > 0);
1846 }
1847
1848
1849 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1850 {
1851         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1852                exp_req_replay_healthy(exp);
1853 }
1854
1855 /** if export done lock_replay or has replay in queue */
1856 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1857 {
1858         return (!exp->exp_lock_replay_needed ||
1859                 atomic_read(&exp->exp_replay_count) > 0);
1860 }
1861
1862 static inline int exp_vbr_healthy(struct obd_export *exp)
1863 {
1864         return (!exp->exp_vbr_failed);
1865 }
1866
1867 static inline int exp_finished(struct obd_export *exp)
1868 {
1869         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1870 }
1871
1872 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1873 {
1874         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1875                 exp_finished(exp);
1876 }
1877
1878 static int check_for_next_transno(struct lu_target *lut)
1879 {
1880         struct ptlrpc_request *req = NULL;
1881         struct obd_device *obd = lut->lut_obd;
1882         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1883         int wake_up = 0, connected, completed, queue_len;
1884         __u64 req_transno = 0;
1885         __u64 update_transno = 0;
1886         __u64 next_transno = 0;
1887         ENTRY;
1888
1889         spin_lock(&obd->obd_recovery_task_lock);
1890         if (!list_empty(&obd->obd_req_replay_queue)) {
1891                 req = list_entry(obd->obd_req_replay_queue.next,
1892                                      struct ptlrpc_request, rq_list);
1893                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1894         }
1895
1896         if (tdtd != NULL)
1897                 update_transno = distribute_txn_get_next_transno(tdtd);
1898
1899         connected = atomic_read(&obd->obd_connected_clients);
1900         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1901         queue_len = obd->obd_requests_queued_for_recovery;
1902         next_transno = obd->obd_next_recovery_transno;
1903
1904         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1905                "req_transno: %llu, next_transno: %llu\n",
1906                obd->obd_max_recoverable_clients, connected, completed,
1907                queue_len, req_transno, next_transno);
1908
1909         if (obd->obd_abort_recovery) {
1910                 CDEBUG(D_HA, "waking for aborted recovery\n");
1911                 wake_up = 1;
1912         } else if (obd->obd_recovery_expired) {
1913                 CDEBUG(D_HA, "waking for expired recovery\n");
1914                 wake_up = 1;
1915         } else if (tdtd != NULL && req != NULL &&
1916                    is_req_replayed_by_update(req)) {
1917                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1918                          "next_transno%llu\n", req_transno, next_transno);
1919                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1920                        req_transno);
1921                 wake_up = 1;
1922         } else if (req_transno == next_transno ||
1923                    (update_transno != 0 && update_transno <= next_transno)) {
1924                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1925                 wake_up = 1;
1926         } else if (queue_len > 0 &&
1927                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1928                 /** handle gaps occured due to lost reply or VBR */
1929                 LASSERTF(req_transno >= next_transno,
1930                          "req_transno: %llu, next_transno: %llu\n",
1931                          req_transno, next_transno);
1932                 CDEBUG(D_HA,
1933                        "%s: waking for gap in transno, VBR is %s (skip: "
1934                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1935                        ", next_update %lld last_committed: %lld)\n",
1936                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1937                        next_transno, queue_len, completed, connected,
1938                        req_transno, update_transno, obd->obd_last_committed);
1939                 obd->obd_next_recovery_transno = req_transno;
1940                 wake_up = 1;
1941         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1942                 CDEBUG(D_HA, "waking for completed recovery\n");
1943                 wake_up = 1;
1944         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1945                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1946                        " by fail_lock, waking up (%lld)\n", next_transno);
1947                 obd->obd_next_recovery_transno = req_transno;
1948                 wake_up = 1;
1949         }
1950         spin_unlock(&obd->obd_recovery_task_lock);
1951         return wake_up;
1952 }
1953
1954 static int check_for_next_lock(struct lu_target *lut)
1955 {
1956         struct obd_device *obd = lut->lut_obd;
1957         int wake_up = 0;
1958
1959         spin_lock(&obd->obd_recovery_task_lock);
1960         if (!list_empty(&obd->obd_lock_replay_queue)) {
1961                 CDEBUG(D_HA, "waking for next lock\n");
1962                 wake_up = 1;
1963         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1964                 CDEBUG(D_HA, "waking for completed lock replay\n");
1965                 wake_up = 1;
1966         } else if (obd->obd_abort_recovery) {
1967                 CDEBUG(D_HA, "waking for aborted recovery\n");
1968                 wake_up = 1;
1969         } else if (obd->obd_recovery_expired) {
1970                 CDEBUG(D_HA, "waking for expired recovery\n");
1971                 wake_up = 1;
1972         }
1973         spin_unlock(&obd->obd_recovery_task_lock);
1974
1975         return wake_up;
1976 }
1977
1978 /**
1979  * wait for recovery events,
1980  * check its status with help of check_routine
1981  * evict dead clients via health_check
1982  */
1983 static int target_recovery_overseer(struct lu_target *lut,
1984                                     int (*check_routine)(struct lu_target *),
1985                                     int (*health_check)(struct obd_export *))
1986 {
1987         struct obd_device       *obd = lut->lut_obd;
1988         struct target_distribute_txn_data *tdtd;
1989 repeat:
1990         if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1991               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1992                 __u64 next_update_transno = 0;
1993
1994                 /* Only abort the recovery if there are no update recovery
1995                  * left in the queue */
1996                 spin_lock(&obd->obd_recovery_task_lock);
1997                 if (lut->lut_tdtd != NULL) {
1998                         next_update_transno =
1999                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2000
2001                         tdtd = lut->lut_tdtd;
2002                         /* If next_update_transno == 0, it probably because
2003                          * updatelog retrieve threads did not get any records
2004                          * yet, let's wait those threads stopped */
2005                         if (next_update_transno == 0) {
2006                                 struct l_wait_info lwi = { 0 };
2007
2008                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2009                                        atomic_read(
2010                                        &tdtd->tdtd_recovery_threads_count) == 0,
2011                                        &lwi);
2012
2013                                 next_update_transno =
2014                                         distribute_txn_get_next_transno(
2015                                                                 lut->lut_tdtd);
2016                         }
2017                 }
2018
2019                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2020                         obd->obd_next_recovery_transno = next_update_transno;
2021                         spin_unlock(&obd->obd_recovery_task_lock);
2022                         /* Disconnect unfinished exports from clients, and
2023                          * keep connection from MDT to make sure the update
2024                          * recovery will still keep trying until some one
2025                          * manually abort the recovery */
2026                         class_disconnect_stale_exports(obd,
2027                                                 exp_finished_or_from_mdt);
2028                         /* Abort all of replay and replay lock req from
2029                          * clients */
2030                         abort_req_replay_queue(obd);
2031                         abort_lock_replay_queue(obd);
2032                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2033                                ")in the queue.\n", obd->obd_name,
2034                                next_update_transno);
2035                 } else {
2036                         obd->obd_abort_recovery = 1;
2037                         spin_unlock(&obd->obd_recovery_task_lock);
2038                         CWARN("%s recovery is aborted by hard timeout\n",
2039                               obd->obd_name);
2040                 }
2041         }
2042
2043         while (wait_event_timeout(obd->obd_next_transno_waitq,
2044                                   check_routine(lut),
2045                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2046                 /* wait indefinitely for event, but don't trigger watchdog */;
2047
2048         if (obd->obd_abort_recovery) {
2049                 CWARN("recovery is aborted, evict exports in recovery\n");
2050                 if (lut->lut_tdtd != NULL) {
2051                         struct l_wait_info lwi = { 0 };
2052
2053                         tdtd = lut->lut_tdtd;
2054                         /* Let's wait all of the update log recovery thread
2055                          * finished */
2056                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2057                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2058                              &lwi);
2059                         /* Then abort the update recovery list */
2060                         dtrq_list_destroy(lut->lut_tdtd);
2061                 }
2062
2063                 /** evict exports which didn't finish recovery yet */
2064                 class_disconnect_stale_exports(obd, exp_finished);
2065                 return 1;
2066         } else if (obd->obd_recovery_expired) {
2067                 obd->obd_recovery_expired = 0;
2068                 /** If some clients died being recovered, evict them */
2069                 LCONSOLE_WARN("%s: recovery is timed out, "
2070                               "evict stale exports\n", obd->obd_name);
2071                 /** evict cexports with no replay in queue, they are stalled */
2072                 class_disconnect_stale_exports(obd, health_check);
2073
2074                 /** continue with VBR */
2075                 spin_lock(&obd->obd_dev_lock);
2076                 obd->obd_version_recov = 1;
2077                 spin_unlock(&obd->obd_dev_lock);
2078                 /**
2079                  * reset timer, recovery will proceed with versions now,
2080                  * timeout is set just to handle reconnection delays
2081                  */
2082                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2083                 /** Wait for recovery events again, after evicting bad clients */
2084                 goto repeat;
2085         }
2086         return 0;
2087 }
2088
2089 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2090 {
2091         struct obd_device       *obd = lut->lut_obd;
2092         struct ptlrpc_request *req = NULL;
2093
2094         CDEBUG(D_HA, "Waiting for lock\n");
2095         if (target_recovery_overseer(lut, check_for_next_lock,
2096                                      exp_lock_replay_healthy))
2097                 abort_lock_replay_queue(obd);
2098
2099         spin_lock(&obd->obd_recovery_task_lock);
2100         if (!list_empty(&obd->obd_lock_replay_queue)) {
2101                 req = list_entry(obd->obd_lock_replay_queue.next,
2102                                      struct ptlrpc_request, rq_list);
2103                 list_del_init(&req->rq_list);
2104                 spin_unlock(&obd->obd_recovery_task_lock);
2105         } else {
2106                 spin_unlock(&obd->obd_recovery_task_lock);
2107                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2108                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2109                 /** evict exports failed VBR */
2110                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2111         }
2112         return req;
2113 }
2114
2115 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2116 {
2117         struct ptlrpc_request *req = NULL;
2118
2119         spin_lock(&obd->obd_recovery_task_lock);
2120         if (!list_empty(&obd->obd_final_req_queue)) {
2121                 req = list_entry(obd->obd_final_req_queue.next,
2122                                      struct ptlrpc_request, rq_list);
2123                 list_del_init(&req->rq_list);
2124                 spin_unlock(&obd->obd_recovery_task_lock);
2125                 if (req->rq_export->exp_in_recovery) {
2126                         spin_lock(&req->rq_export->exp_lock);
2127                         req->rq_export->exp_in_recovery = 0;
2128                         spin_unlock(&req->rq_export->exp_lock);
2129                 }
2130         } else {
2131                 spin_unlock(&obd->obd_recovery_task_lock);
2132         }
2133         return req;
2134 }
2135
2136 static void handle_recovery_req(struct ptlrpc_thread *thread,
2137                                 struct ptlrpc_request *req,
2138                                 svc_handler_t handler)
2139 {
2140         ENTRY;
2141
2142         /**
2143          * export can be evicted during recovery, no need to handle replays for
2144          * it after that, discard such request silently
2145          */
2146         if (req->rq_export->exp_disconnected)
2147                 RETURN_EXIT;
2148
2149         req->rq_session.lc_thread = thread;
2150         req->rq_svc_thread = thread;
2151         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2152
2153         /* thread context */
2154         lu_context_enter(&thread->t_env->le_ctx);
2155         (void)handler(req);
2156         lu_context_exit(&thread->t_env->le_ctx);
2157
2158         /* don't reset timer for final stage */
2159         if (!exp_finished(req->rq_export)) {
2160                 int to = obd_timeout;
2161
2162                 /**
2163                  * Add request timeout to the recovery time so next request from
2164                  * this client may come in recovery time
2165                  */
2166                 if (!AT_OFF) {
2167                         struct ptlrpc_service_part *svcpt;
2168
2169                         svcpt = req->rq_rqbd->rqbd_svcpt;
2170                         /* If the server sent early reply for this request,
2171                          * the client will recalculate the timeout according to
2172                          * current server estimate service time, so we will
2173                          * use the maxium timeout here for waiting the client
2174                          * sending the next req */
2175                         to = max((int)at_est2timeout(
2176                                  at_get(&svcpt->scp_at_estimate)),
2177                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2178                         /* Add 2 net_latency, one for balance rq_deadline
2179                          * (see ptl_send_rpc), one for resend the req to server,
2180                          * Note: client will pack net_latency in replay req
2181                          * (see ptlrpc_replay_req) */
2182                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2183                 }
2184                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2185         }
2186         EXIT;
2187 }
2188
2189 /** Checking routines for recovery */
2190 static int check_for_recovery_ready(struct lu_target *lut)
2191 {
2192         struct obd_device *obd = lut->lut_obd;
2193         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2194
2195         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2196                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2197                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2198                obd->obd_recovery_expired);
2199
2200         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2201                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2202                 if (clnts + obd->obd_stale_clients <
2203                     obd->obd_max_recoverable_clients)
2204                         return 0;
2205         }
2206
2207         if (lut->lut_tdtd != NULL) {
2208                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2209                     !obd->obd_abort_recovery) {
2210                         /* Let's extend recovery timer, in case the recovery
2211                          * timer expired, and some clients got evicted */
2212                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2213                                               true);
2214                         CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2215                                obd->obd_name, obd->obd_recovery_timeout);
2216                         return 0;
2217                 }
2218         }
2219
2220         return 1;
2221 }
2222
2223 enum {
2224         REQUEST_RECOVERY = 1,
2225         UPDATE_RECOVERY = 2,
2226 };
2227
2228 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2229 {
2230         __u64 transno = 0;
2231
2232         if (!list_empty(&obd->obd_req_replay_queue)) {
2233                 struct ptlrpc_request *req;
2234
2235                 req = list_entry(obd->obd_req_replay_queue.next,
2236                                  struct ptlrpc_request, rq_list);
2237                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2238         }
2239
2240         return transno;
2241 }
2242
2243 static __u64 get_next_transno(struct lu_target *lut, int *type)
2244 {
2245         struct obd_device *obd = lut->lut_obd;
2246         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2247         __u64 transno = 0;
2248         __u64 update_transno;
2249         ENTRY;
2250
2251         transno = get_next_replay_req_transno(obd);
2252         if (type != NULL)
2253                 *type = REQUEST_RECOVERY;
2254
2255         if (tdtd == NULL)
2256                 RETURN(transno);
2257
2258         update_transno = distribute_txn_get_next_transno(tdtd);
2259         if (transno == 0 || (transno >= update_transno &&
2260                              update_transno != 0)) {
2261                 transno = update_transno;
2262                 if (type != NULL)
2263                         *type = UPDATE_RECOVERY;
2264         }
2265
2266         RETURN(transno);
2267 }
2268
2269 /**
2270  * drop duplicate replay request
2271  *
2272  * Because the operation has been replayed by update recovery, the request
2273  * with the same transno will be dropped and also notify the client to send
2274  * next replay request.
2275  *
2276  * \param[in] env       execution environment
2277  * \param[in] obd       failover obd device
2278  * \param[in] req       request to be dropped
2279  */
2280 static void drop_duplicate_replay_req(struct lu_env *env,
2281                                       struct obd_device *obd,
2282                                       struct ptlrpc_request *req)
2283 {
2284         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2285                   " update records are found.\n",
2286                   lustre_msg_get_transno(req->rq_reqmsg),
2287                   libcfs_nid2str(req->rq_peer.nid));
2288
2289         /* Right now, only for MDS reint operation update replay and
2290          * normal request replay can have the same transno */
2291         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2292                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2293                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2294                 if (likely(req->rq_export))
2295                         target_committed_to_req(req);
2296                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2297                 target_send_reply(req, req->rq_status, 0);
2298         } else {
2299                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2300                 libcfs_nid2str(req->rq_peer.nid));
2301         }
2302         target_exp_dequeue_req_replay(req);
2303         target_request_copy_put(req);
2304         obd->obd_replayed_requests++;
2305 }
2306
2307 static void replay_request_or_update(struct lu_env *env,
2308                                      struct lu_target *lut,
2309                                      struct target_recovery_data *trd,
2310                                      struct ptlrpc_thread *thread)
2311 {
2312         struct obd_device *obd = lut->lut_obd;
2313         struct ptlrpc_request *req = NULL;
2314         int                     type;
2315         __u64                   transno;
2316         ENTRY;
2317
2318         CDEBUG(D_HA, "Waiting for transno %lld\n",
2319                obd->obd_next_recovery_transno);
2320
2321         /* Replay all of request and update by transno */
2322         do {
2323                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2324
2325                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2326
2327                 /** It is needed to extend recovery window above
2328                  *  recovery_time_soft. Extending is possible only in the
2329                  *  end of recovery window (see more details in
2330                  *  handle_recovery_req()).
2331                  */
2332                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2333
2334                 if (target_recovery_overseer(lut, check_for_next_transno,
2335                                         exp_req_replay_healthy_or_from_mdt)) {
2336                         abort_req_replay_queue(obd);
2337                         abort_lock_replay_queue(obd);
2338                         goto abort;
2339                 }
2340
2341                 spin_lock(&obd->obd_recovery_task_lock);
2342                 transno = get_next_transno(lut, &type);
2343                 if (type == REQUEST_RECOVERY && transno != 0) {
2344                         /* Drop replay request from client side, if the
2345                          * replay has been executed by update with the
2346                          * same transno */
2347                         req = list_entry(obd->obd_req_replay_queue.next,
2348                                         struct ptlrpc_request, rq_list);
2349
2350                         list_del_init(&req->rq_list);
2351                         obd->obd_requests_queued_for_recovery--;
2352                         spin_unlock(&obd->obd_recovery_task_lock);
2353
2354                         /* Let's check if the request has been redone by
2355                          * update replay */
2356                         if (is_req_replayed_by_update(req)) {
2357                                 struct distribute_txn_replay_req *dtrq;
2358
2359                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2360                                                                    req->rq_xid);
2361                                 LASSERT(dtrq != NULL);
2362                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2363                                 list_del_init(&dtrq->dtrq_list);
2364                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2365                                 dtrq_destroy(dtrq);
2366
2367                                 drop_duplicate_replay_req(env, obd, req);
2368
2369                                 continue;
2370                         }
2371
2372                         LASSERT(trd->trd_processing_task == current_pid());
2373                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2374                                   lustre_msg_get_transno(req->rq_reqmsg),
2375                                   libcfs_nid2str(req->rq_peer.nid));
2376
2377                         handle_recovery_req(thread, req,
2378                                             trd->trd_recovery_handler);
2379                         /**
2380                          * bz18031: increase next_recovery_transno before
2381                          * target_request_copy_put() will drop exp_rpc reference
2382                          */
2383                         spin_lock(&obd->obd_recovery_task_lock);
2384                         obd->obd_next_recovery_transno++;
2385                         spin_unlock(&obd->obd_recovery_task_lock);
2386                         target_exp_dequeue_req_replay(req);
2387                         target_request_copy_put(req);
2388                         obd->obd_replayed_requests++;
2389                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2390                         struct distribute_txn_replay_req *dtrq;
2391                         int rc;
2392
2393                         spin_unlock(&obd->obd_recovery_task_lock);
2394
2395                         LASSERT(tdtd != NULL);
2396                         dtrq = distribute_txn_get_next_req(tdtd);
2397                         lu_context_enter(&thread->t_env->le_ctx);
2398                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2399                         lu_context_exit(&thread->t_env->le_ctx);
2400                         extend_recovery_timer(obd, obd_timeout, true);
2401
2402                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2403                                 CDEBUG(D_HA, "Move x%llu t%llu"
2404                                        " to finish list\n", dtrq->dtrq_xid,
2405                                        dtrq->dtrq_master_transno);
2406
2407                                 /* Add it to the replay finish list */
2408                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2409                                 list_add(&dtrq->dtrq_list,
2410                                          &tdtd->tdtd_replay_finish_list);
2411                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2412
2413                                 spin_lock(&obd->obd_recovery_task_lock);
2414                                 if (transno == obd->obd_next_recovery_transno)
2415                                         obd->obd_next_recovery_transno++;
2416                                 else if (transno >
2417                                          obd->obd_next_recovery_transno)
2418                                         obd->obd_next_recovery_transno =
2419                                                                 transno + 1;
2420                                 spin_unlock(&obd->obd_recovery_task_lock);
2421                         } else {
2422                                 dtrq_destroy(dtrq);
2423                         }
2424                 } else {
2425                         spin_unlock(&obd->obd_recovery_task_lock);
2426 abort:
2427                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2428                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2429                         /** evict exports failed VBR */
2430                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2431                         break;
2432                 }
2433         } while (1);
2434 }
2435
2436 static int target_recovery_thread(void *arg)
2437 {
2438         struct lu_target *lut = arg;
2439         struct obd_device *obd = lut->lut_obd;
2440         struct ptlrpc_request *req;
2441         struct target_recovery_data *trd = &obd->obd_recovery_data;
2442         unsigned long delta;
2443         struct lu_env *env;
2444         struct ptlrpc_thread *thread = NULL;
2445         int rc = 0;
2446         ENTRY;
2447
2448         unshare_fs_struct();
2449         OBD_ALLOC_PTR(thread);
2450         if (thread == NULL)
2451                 RETURN(-ENOMEM);
2452
2453         OBD_ALLOC_PTR(env);
2454         if (env == NULL) {
2455                 OBD_FREE_PTR(thread);
2456                 RETURN(-ENOMEM);
2457         }
2458
2459         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2460         if (rc) {
2461                 OBD_FREE_PTR(thread);
2462                 OBD_FREE_PTR(env);
2463                 RETURN(rc);
2464         }
2465
2466         thread->t_env = env;
2467         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2468         env->le_ctx.lc_thread = thread;
2469         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2470         thread->t_watchdog = NULL;
2471
2472         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2473                current_pid());
2474         trd->trd_processing_task = current_pid();
2475
2476         spin_lock(&obd->obd_dev_lock);
2477         obd->obd_recovering = 1;
2478         spin_unlock(&obd->obd_dev_lock);
2479         complete(&trd->trd_starting);
2480
2481         /* first of all, we have to know the first transno to replay */
2482         if (target_recovery_overseer(lut, check_for_recovery_ready,
2483                                      exp_connect_healthy)) {
2484                 abort_req_replay_queue(obd);
2485                 abort_lock_replay_queue(obd);
2486                 if (lut->lut_tdtd != NULL)
2487                         dtrq_list_destroy(lut->lut_tdtd);
2488         }
2489
2490         /* next stage: replay requests or update */
2491         delta = jiffies;
2492         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2493                atomic_read(&obd->obd_req_replay_clients),
2494                obd->obd_next_recovery_transno);
2495         replay_request_or_update(env, lut, trd, thread);
2496
2497         /**
2498          * The second stage: replay locks
2499          */
2500         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2501                atomic_read(&obd->obd_lock_replay_clients));
2502         while ((req = target_next_replay_lock(lut))) {
2503                 LASSERT(trd->trd_processing_task == current_pid());
2504                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2505                           libcfs_nid2str(req->rq_peer.nid));
2506                 handle_recovery_req(thread, req,
2507                                     trd->trd_recovery_handler);
2508                 target_request_copy_put(req);
2509                 obd->obd_replayed_locks++;
2510         }
2511
2512         /**
2513          * The third stage: reply on final pings, at this moment all clients
2514          * must have request in final queue
2515          */
2516         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2517         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2518         /** Update server last boot epoch */
2519         tgt_boot_epoch_update(lut);
2520         /* We drop recoverying flag to forward all new requests
2521          * to regular mds_handle() since now */
2522         spin_lock(&obd->obd_dev_lock);
2523         obd->obd_recovering = obd->obd_abort_recovery = 0;
2524         spin_unlock(&obd->obd_dev_lock);
2525         spin_lock(&obd->obd_recovery_task_lock);
2526         target_cancel_recovery_timer(obd);
2527         spin_unlock(&obd->obd_recovery_task_lock);
2528         while ((req = target_next_final_ping(obd))) {
2529                 LASSERT(trd->trd_processing_task == current_pid());
2530                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2531                           libcfs_nid2str(req->rq_peer.nid));
2532                 handle_recovery_req(thread, req,
2533                                     trd->trd_recovery_handler);
2534                 /* Because the waiting client can not send ping to server,
2535                  * so we need refresh the last_request_time, to avoid the
2536                  * export is being evicted */
2537                 ptlrpc_update_export_timer(req->rq_export, 0);
2538                 target_request_copy_put(req);
2539         }
2540
2541         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2542         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2543               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2544         if (delta > OBD_RECOVERY_TIME_SOFT) {
2545                 CWARN("too long recovery - read logs\n");
2546                 libcfs_debug_dumplog();
2547         }
2548
2549         target_finish_recovery(lut);
2550
2551         lu_context_fini(&env->le_ctx);
2552         trd->trd_processing_task = 0;
2553         complete(&trd->trd_finishing);
2554
2555         tgt_io_thread_done(thread);
2556         OBD_FREE_PTR(thread);
2557         OBD_FREE_PTR(env);
2558         RETURN(rc);
2559 }
2560
2561 static int target_start_recovery_thread(struct lu_target *lut,
2562                                         svc_handler_t handler)
2563 {
2564         struct obd_device *obd = lut->lut_obd;
2565         int rc = 0;
2566         struct target_recovery_data *trd = &obd->obd_recovery_data;
2567         int index;
2568
2569         memset(trd, 0, sizeof(*trd));
2570         init_completion(&trd->trd_starting);
2571         init_completion(&trd->trd_finishing);
2572         trd->trd_recovery_handler = handler;
2573
2574         rc = server_name2index(obd->obd_name, &index, NULL);
2575         if (rc < 0)
2576                 return rc;
2577
2578         if (!IS_ERR(kthread_run(target_recovery_thread,
2579                                 lut, "tgt_recover_%d", index))) {
2580                 wait_for_completion(&trd->trd_starting);
2581                 LASSERT(obd->obd_recovering != 0);
2582         } else {
2583                 rc = -ECHILD;
2584         }
2585
2586         return rc;
2587 }
2588
2589 void target_stop_recovery_thread(struct obd_device *obd)
2590 {
2591         if (obd->obd_recovery_data.trd_processing_task > 0) {
2592                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2593                 /** recovery can be done but postrecovery is not yet */
2594                 spin_lock(&obd->obd_dev_lock);
2595                 if (obd->obd_recovering) {
2596                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2597                         obd->obd_abort_recovery = 1;
2598                         wake_up(&obd->obd_next_transno_waitq);
2599                 }
2600                 spin_unlock(&obd->obd_dev_lock);
2601                 wait_for_completion(&trd->trd_finishing);
2602         }
2603 }
2604 EXPORT_SYMBOL(target_stop_recovery_thread);
2605
2606 void target_recovery_fini(struct obd_device *obd)
2607 {
2608         class_disconnect_exports(obd);
2609         target_stop_recovery_thread(obd);
2610         target_cleanup_recovery(obd);
2611 }
2612 EXPORT_SYMBOL(target_recovery_fini);
2613
2614 static void target_recovery_expired(unsigned long castmeharder)
2615 {
2616         struct obd_device *obd = (struct obd_device *)castmeharder;
2617         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2618                " after %llus (%d clients connected)\n",
2619                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2620                (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2621                atomic_read(&obd->obd_connected_clients));
2622
2623         obd->obd_recovery_expired = 1;
2624         wake_up(&obd->obd_next_transno_waitq);
2625 }
2626
2627 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2628 {
2629         struct obd_device *obd = lut->lut_obd;
2630
2631         if (lut->lut_bottom->dd_rdonly)
2632                 return;
2633
2634         if (obd->obd_max_recoverable_clients == 0) {
2635                 /** Update server last boot epoch */
2636                 tgt_boot_epoch_update(lut);
2637                 return;
2638         }
2639
2640         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2641                "last_transno %llu\n", obd->obd_name,
2642                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2643         LASSERT(obd->obd_stopping == 0);
2644         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2645         obd->obd_recovery_start = 0;
2646         obd->obd_recovery_end = 0;
2647
2648         setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2649                     (unsigned long)obd);
2650         target_start_recovery_thread(lut, handler);
2651 }
2652 EXPORT_SYMBOL(target_recovery_init);
2653
2654 static int target_process_req_flags(struct obd_device *obd,
2655                                     struct ptlrpc_request *req)
2656 {
2657         struct obd_export *exp = req->rq_export;
2658         LASSERT(exp != NULL);
2659         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2660                 /* client declares he's ready to replay locks */
2661                 spin_lock(&exp->exp_lock);
2662                 if (exp->exp_req_replay_needed) {
2663                         exp->exp_req_replay_needed = 0;
2664                         spin_unlock(&exp->exp_lock);
2665
2666                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2667                         atomic_dec(&obd->obd_req_replay_clients);
2668                 } else {
2669                         spin_unlock(&exp->exp_lock);
2670                 }
2671         }
2672         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2673                 /* client declares he's ready to complete recovery
2674                  * so, we put the request on th final queue */
2675                 spin_lock(&exp->exp_lock);
2676                 if (exp->exp_lock_replay_needed) {
2677                         exp->exp_lock_replay_needed = 0;
2678                         spin_unlock(&exp->exp_lock);
2679
2680                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2681                         atomic_dec(&obd->obd_lock_replay_clients);
2682                 } else {
2683                         spin_unlock(&exp->exp_lock);
2684                 }
2685         }
2686         return 0;
2687 }
2688
2689 int target_queue_recovery_request(struct ptlrpc_request *req,
2690                                   struct obd_device *obd)
2691 {
2692         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2693         struct ptlrpc_request *reqiter;
2694         int inserted = 0;
2695         ENTRY;
2696
2697         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2698                 /* Processing the queue right now, don't re-add. */
2699                 RETURN(1);
2700         }
2701
2702         target_process_req_flags(obd, req);
2703
2704         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2705                 /* client declares he's ready to complete recovery
2706                  * so, we put the request on th final queue */
2707                 target_request_copy_get(req);
2708                 DEBUG_REQ(D_HA, req, "queue final req");
2709                 wake_up(&obd->obd_next_transno_waitq);
2710                 spin_lock(&obd->obd_recovery_task_lock);
2711                 if (obd->obd_recovering) {
2712                         struct ptlrpc_request *tmp;
2713                         struct ptlrpc_request *duplicate = NULL;
2714
2715                         if (likely(!req->rq_export->exp_replay_done)) {
2716                                 req->rq_export->exp_replay_done = 1;
2717                                 list_add_tail(&req->rq_list,
2718                                               &obd->obd_final_req_queue);
2719                                 spin_unlock(&obd->obd_recovery_task_lock);
2720                                 RETURN(0);
2721                         }
2722
2723                         /* XXX O(n), but only happens if final ping is
2724                          * timed out, probably reorganize the list as
2725                          * a hash list later */
2726                         list_for_each_entry_safe(reqiter, tmp,
2727                                                  &obd->obd_final_req_queue,
2728                                                  rq_list) {
2729                                 if (reqiter->rq_export == req->rq_export) {
2730                                         list_del_init(&reqiter->rq_list);
2731                                         duplicate = reqiter;
2732                                         break;
2733                                 }
2734                         }
2735
2736                         list_add_tail(&req->rq_list,
2737                                       &obd->obd_final_req_queue);
2738                         req->rq_export->exp_replay_done = 1;
2739                         spin_unlock(&obd->obd_recovery_task_lock);
2740
2741                         if (duplicate != NULL) {
2742                                 DEBUG_REQ(D_HA, duplicate,
2743                                           "put prev final req\n");
2744                                 target_request_copy_put(duplicate);
2745                         }
2746                         RETURN(0);
2747                 } else {
2748                         spin_unlock(&obd->obd_recovery_task_lock);
2749                         target_request_copy_put(req);
2750                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2751                 }
2752         }
2753         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2754                 /* client declares he's ready to replay locks */
2755                 target_request_copy_get(req);
2756                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2757                 wake_up(&obd->obd_next_transno_waitq);
2758                 spin_lock(&obd->obd_recovery_task_lock);
2759                 LASSERT(obd->obd_recovering);
2760                 /* usually due to recovery abort */
2761                 if (!req->rq_export->exp_in_recovery) {
2762                         spin_unlock(&obd->obd_recovery_task_lock);
2763                         target_request_copy_put(req);
2764                         RETURN(-ENOTCONN);
2765                 }
2766                 LASSERT(req->rq_export->exp_lock_replay_needed);
2767                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2768                 spin_unlock(&obd->obd_recovery_task_lock);
2769                 RETURN(0);
2770         }
2771
2772         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2773          * (i.e. buflens etc are in my own byte order), but type-dependent
2774          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2775
2776         if (!transno) {
2777                 INIT_LIST_HEAD(&req->rq_list);
2778                 DEBUG_REQ(D_HA, req, "not queueing");
2779                 RETURN(1);
2780         }
2781
2782         /* If we're processing the queue, we want don't want to queue this
2783          * message.
2784          *
2785          * Also, if this request has a transno less than the one we're waiting
2786          * for, we should process it now.  It could (and currently always will)
2787          * be an open request for a descriptor that was opened some time ago.
2788          *
2789          * Also, a resent, replayed request that has already been
2790          * handled will pass through here and be processed immediately.
2791          */
2792         CDEBUG(D_HA, "Next recovery transno: %llu"
2793                ", current: %llu, replaying\n",
2794                obd->obd_next_recovery_transno, transno);
2795
2796         /* If the request has been replayed by update replay, then sends this
2797          * request to the recovery thread (replay_request_or_update()), where
2798          * it will be handled */
2799         spin_lock(&obd->obd_recovery_task_lock);
2800         if (transno < obd->obd_next_recovery_transno &&
2801             !is_req_replayed_by_update(req)) {
2802                 /* Processing the queue right now, don't re-add. */
2803                 LASSERT(list_empty(&req->rq_list));
2804                 spin_unlock(&obd->obd_recovery_task_lock);
2805                 RETURN(1);
2806         }
2807         spin_unlock(&obd->obd_recovery_task_lock);
2808
2809         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2810                 RETURN(0);
2811
2812         target_request_copy_get(req);
2813         if (!req->rq_export->exp_in_recovery) {
2814                 target_request_copy_put(req);
2815                 RETURN(-ENOTCONN);
2816         }
2817         LASSERT(req->rq_export->exp_req_replay_needed);
2818
2819         if (target_exp_enqueue_req_replay(req)) {
2820                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2821                 target_request_copy_put(req);
2822                 RETURN(0);
2823         }
2824
2825         /* XXX O(n^2) */
2826         spin_lock(&obd->obd_recovery_task_lock);
2827         LASSERT(obd->obd_recovering);
2828         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2829                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2830                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2831                         inserted = 1;
2832                         goto added;
2833                 }
2834
2835                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2836                              transno)) {
2837                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2838                                   "has been claimed by another client");
2839                         spin_unlock(&obd->obd_recovery_task_lock);
2840                         target_exp_dequeue_req_replay(req);
2841                         target_request_copy_put(req);
2842                         RETURN(0);
2843                 }
2844         }
2845 added:
2846         if (!inserted)
2847                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2848
2849         obd->obd_requests_queued_for_recovery++;
2850         spin_unlock(&obd->obd_recovery_task_lock);
2851         wake_up(&obd->obd_next_transno_waitq);
2852         RETURN(0);
2853 }
2854
2855 int target_handle_ping(struct ptlrpc_request *req)
2856 {
2857         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2858         return req_capsule_server_pack(&req->rq_pill);
2859 }
2860
2861 void target_committed_to_req(struct ptlrpc_request *req)
2862 {
2863         struct obd_export *exp = req->rq_export;
2864
2865         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2866                 lustre_msg_set_last_committed(req->rq_repmsg,
2867                                               exp->exp_last_committed);
2868         else
2869                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2870                           "%d)", exp->exp_obd->obd_no_transno,
2871                           req->rq_repmsg == NULL);
2872
2873         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2874                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2875 }
2876
2877 #endif /* HAVE_SERVER_SUPPORT */
2878
2879 /**
2880  * Packs current SLV and Limit into \a req.
2881  */
2882 int target_pack_pool_reply(struct ptlrpc_request *req)
2883 {
2884         struct obd_device *obd;
2885         ENTRY;
2886
2887         /* Check that we still have all structures alive as this may
2888          * be some late RPC at shutdown time. */
2889         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2890                      !exp_connect_lru_resize(req->rq_export))) {
2891                 lustre_msg_set_slv(req->rq_repmsg, 0);
2892                 lustre_msg_set_limit(req->rq_repmsg, 0);
2893                 RETURN(0);
2894         }
2895
2896         /* OBD is alive here as export is alive, which we checked above. */
2897         obd = req->rq_export->exp_obd;
2898
2899         read_lock(&obd->obd_pool_lock);
2900         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2901         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2902         read_unlock(&obd->obd_pool_lock);
2903
2904         RETURN(0);
2905 }
2906
2907 static int target_send_reply_msg(struct ptlrpc_request *req,
2908                                  int rc, int fail_id)
2909 {
2910         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2911                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2912                 return -ECOMM;
2913         }
2914         /* We can have a null rq_reqmsg in the event of bad signature or
2915          * no context when unwrapping */
2916         if (req->rq_reqmsg &&
2917             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2918             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2919                 return -ECOMM;
2920
2921         if (unlikely(rc)) {
2922                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2923                 req->rq_status = rc;
2924                 return ptlrpc_send_error(req, 1);
2925         } else {
2926                 DEBUG_REQ(D_NET, req, "sending reply");
2927         }
2928
2929         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2930 }
2931
2932 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2933 {
2934         struct ptlrpc_service_part *svcpt;
2935         int                        netrc;
2936         struct ptlrpc_reply_state *rs;
2937         struct obd_export         *exp;
2938         ENTRY;
2939
2940         if (req->rq_no_reply) {
2941                 EXIT;
2942                 return;
2943         }
2944
2945         svcpt = req->rq_rqbd->rqbd_svcpt;
2946         rs = req->rq_reply_state;
2947         if (rs == NULL || !rs->rs_difficult) {
2948                 /* no notifiers */
2949                 target_send_reply_msg (req, rc, fail_id);
2950                 EXIT;
2951                 return;
2952         }
2953
2954         /* must be an export if locks saved */
2955         LASSERT(req->rq_export != NULL);
2956         /* req/reply consistent */
2957         LASSERT(rs->rs_svcpt == svcpt);
2958
2959         /* "fresh" reply */
2960         LASSERT(!rs->rs_scheduled);
2961         LASSERT(!rs->rs_scheduled_ever);
2962         LASSERT(!rs->rs_handled);
2963         LASSERT(!rs->rs_on_net);
2964         LASSERT(rs->rs_export == NULL);
2965         LASSERT(list_empty(&rs->rs_obd_list));
2966         LASSERT(list_empty(&rs->rs_exp_list));
2967
2968         exp = class_export_get(req->rq_export);
2969
2970         /* disable reply scheduling while I'm setting up */
2971         rs->rs_scheduled = 1;
2972         rs->rs_on_net    = 1;
2973         rs->rs_xid       = req->rq_xid;
2974         rs->rs_transno   = req->rq_transno;
2975         rs->rs_export    = exp;
2976         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2977
2978         spin_lock(&exp->exp_uncommitted_replies_lock);
2979         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2980                rs->rs_transno, exp->exp_last_committed);
2981         if (rs->rs_transno > exp->exp_last_committed) {
2982                 /* not committed already */
2983                 list_add_tail(&rs->rs_obd_list,
2984                                   &exp->exp_uncommitted_replies);
2985         }
2986         spin_unlock(&exp->exp_uncommitted_replies_lock);
2987
2988         spin_lock(&exp->exp_lock);
2989         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2990         spin_unlock(&exp->exp_lock);
2991
2992         netrc = target_send_reply_msg(req, rc, fail_id);
2993
2994         spin_lock(&svcpt->scp_rep_lock);
2995
2996         atomic_inc(&svcpt->scp_nreps_difficult);
2997
2998         if (netrc != 0) {
2999                 /* error sending: reply is off the net.  Also we need +1
3000                  * reply ref until ptlrpc_handle_rs() is done
3001                  * with the reply state (if the send was successful, there
3002                  * would have been +1 ref for the net, which
3003                  * reply_out_callback leaves alone) */
3004                 rs->rs_on_net = 0;
3005                 ptlrpc_rs_addref(rs);
3006         }
3007
3008         spin_lock(&rs->rs_lock);
3009         if (rs->rs_transno <= exp->exp_last_committed ||
3010             (!rs->rs_on_net && !rs->rs_no_ack) ||
3011             list_empty(&rs->rs_exp_list) ||     /* completed already */
3012             list_empty(&rs->rs_obd_list)) {
3013                 CDEBUG(D_HA, "Schedule reply immediately\n");
3014                 ptlrpc_dispatch_difficult_reply(rs);
3015         } else {
3016                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3017                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3018         }
3019         spin_unlock(&rs->rs_lock);
3020         spin_unlock(&svcpt->scp_rep_lock);
3021         EXIT;
3022 }
3023
3024 enum ldlm_mode lck_compat_array[] = {
3025         [LCK_EX]    = LCK_COMPAT_EX,
3026         [LCK_PW]    = LCK_COMPAT_PW,
3027         [LCK_PR]    = LCK_COMPAT_PR,
3028         [LCK_CW]    = LCK_COMPAT_CW,
3029         [LCK_CR]    = LCK_COMPAT_CR,
3030         [LCK_NL]    = LCK_COMPAT_NL,
3031         [LCK_GROUP] = LCK_COMPAT_GROUP,
3032         [LCK_COS]   = LCK_COMPAT_COS,
3033 };
3034
3035 /**
3036  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3037  * not escape to the user level.
3038  */
3039 int ldlm_error2errno(enum ldlm_error error)
3040 {
3041         int result;
3042
3043         switch (error) {
3044         case ELDLM_OK:
3045         case ELDLM_LOCK_MATCHED:
3046                 result = 0;
3047                 break;
3048         case ELDLM_LOCK_CHANGED:
3049                 result = -ESTALE;
3050                 break;
3051         case ELDLM_LOCK_ABORTED:
3052                 result = -ENAVAIL;
3053                 break;
3054         case ELDLM_LOCK_REPLACED:
3055                 result = -ESRCH;
3056                 break;
3057         case ELDLM_NO_LOCK_DATA:
3058                 result = -ENOENT;
3059                 break;
3060         case ELDLM_NAMESPACE_EXISTS:
3061                 result = -EEXIST;
3062                 break;
3063         case ELDLM_BAD_NAMESPACE:
3064                 result = -EBADF;
3065                 break;
3066         default:
3067                 if (((int)error) < 0) { /* cast to signed type */
3068                         result = error; /* as ldlm_error can be unsigned */
3069                 } else {
3070                         CERROR("Invalid DLM result code: %d\n", error);
3071                         result = -EPROTO;
3072                 }
3073         }
3074         return result;
3075 }
3076 EXPORT_SYMBOL(ldlm_error2errno);
3077
3078 /**
3079  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3080  */
3081 enum ldlm_error ldlm_errno2error(int err_no)
3082 {
3083         int error;
3084
3085         switch (err_no) {
3086         case 0:
3087                 error = ELDLM_OK;
3088                 break;
3089         case -ESTALE:
3090                 error = ELDLM_LOCK_CHANGED;
3091                 break;
3092         case -ENAVAIL:
3093                 error = ELDLM_LOCK_ABORTED;
3094                 break;
3095         case -ESRCH:
3096                 error = ELDLM_LOCK_REPLACED;
3097                 break;
3098         case -ENOENT:
3099                 error = ELDLM_NO_LOCK_DATA;
3100                 break;
3101         case -EEXIST:
3102                 error = ELDLM_NAMESPACE_EXISTS;
3103                 break;
3104         case -EBADF:
3105                 error = ELDLM_BAD_NAMESPACE;
3106                 break;
3107         default:
3108                 error = err_no;
3109         }
3110         return error;
3111 }
3112
3113 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3114 void ldlm_dump_export_locks(struct obd_export *exp)
3115 {
3116         spin_lock(&exp->exp_locks_list_guard);
3117         if (!list_empty(&exp->exp_locks_list)) {
3118                 struct ldlm_lock *lock;
3119
3120                 CERROR("dumping locks for export %p,"
3121                        "ignore if the unmount doesn't hang\n", exp);
3122                 list_for_each_entry(lock, &exp->exp_locks_list,
3123                                         l_exp_refs_link)
3124                         LDLM_ERROR(lock, "lock:");
3125         }
3126         spin_unlock(&exp->exp_locks_list_guard);
3127 }
3128 #endif
3129
3130 #ifdef HAVE_SERVER_SUPPORT
3131 static int target_bulk_timeout(void *data)
3132 {
3133         ENTRY;
3134         /* We don't fail the connection here, because having the export
3135          * killed makes the (vital) call to commitrw very sad.
3136          */
3137         RETURN(1);
3138 }
3139
3140 static inline const char *bulk2type(struct ptlrpc_request *req)
3141 {
3142         if (req->rq_bulk_read)
3143                 return "READ";
3144         if (req->rq_bulk_write)
3145                 return "WRITE";
3146         return "UNKNOWN";
3147 }
3148
3149 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3150                    struct l_wait_info *lwi)
3151 {
3152         struct ptlrpc_request   *req = desc->bd_req;
3153         time_t                   start = cfs_time_current_sec();
3154         time_t                   deadline;
3155         int                      rc = 0;
3156
3157         ENTRY;
3158
3159         /* If there is eviction in progress, wait for it to finish. */
3160         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3161                 *lwi = LWI_INTR(NULL, NULL);
3162                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3163                                   !atomic_read(&exp->exp_obd->
3164                                                    obd_evict_inprogress),
3165                                   lwi);
3166         }
3167
3168         /* Check if client was evicted or reconnected already. */
3169         if (exp->exp_failed ||
3170             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3171                 rc = -ENOTCONN;
3172         } else {
3173                 if (req->rq_bulk_read)
3174                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3175
3176                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3177                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3178                 else /* old version, bulk matchbits is rq_xid */
3179                         req->rq_mbits = req->rq_xid;
3180
3181                 if (rc == 0)
3182                         rc = ptlrpc_start_bulk_transfer(desc);
3183         }
3184
3185         if (rc < 0) {
3186                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3187                           bulk2type(req), rc);
3188                 RETURN(rc);
3189         }
3190
3191         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3192                 ptlrpc_abort_bulk(desc);
3193                 RETURN(0);
3194         }
3195
3196         /* limit actual bulk transfer to bulk_timeout seconds */
3197         deadline = start + bulk_timeout;
3198         if (deadline > req->rq_deadline)
3199                 deadline = req->rq_deadline;
3200
3201         do {
3202                 long timeoutl = deadline - cfs_time_current_sec();
3203                 cfs_duration_t timeout = timeoutl <= 0 ?
3204                                          CFS_TICK : cfs_time_seconds(timeoutl);
3205                 time_t  rq_deadline;
3206
3207                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3208                                             target_bulk_timeout, desc);
3209                 rc = l_wait_event(desc->bd_waitq,
3210                                   !ptlrpc_server_bulk_active(desc) ||
3211                                   exp->exp_failed ||
3212                                   exp->exp_conn_cnt >
3213                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3214                                   lwi);
3215                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3216                 /* Wait again if we changed rq_deadline. */
3217                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3218                 deadline = start + bulk_timeout;
3219                 if (deadline > rq_deadline)
3220                         deadline = rq_deadline;
3221         } while ((rc == -ETIMEDOUT) &&
3222                  (deadline > cfs_time_current_sec()));
3223
3224         if (rc == -ETIMEDOUT) {
3225                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3226                           bulk2type(req), deadline - start,
3227                           cfs_time_current_sec() - deadline);
3228                 ptlrpc_abort_bulk(desc);
3229         } else if (exp->exp_failed) {
3230                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3231                           bulk2type(req));
3232                 rc = -ENOTCONN;
3233                 ptlrpc_abort_bulk(desc);
3234         } else if (exp->exp_conn_cnt >
3235                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3236                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3237                           bulk2type(req));
3238                 /* We don't reply anyway. */
3239                 rc = -ETIMEDOUT;
3240                 ptlrpc_abort_bulk(desc);
3241         } else if (desc->bd_failure) {
3242                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3243                           bulk2type(req));
3244                 /* XXX should this be a different errno? */
3245                 rc = -ETIMEDOUT;
3246         } else {
3247                 if (req->rq_bulk_write)
3248                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3249                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3250                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3251                                   bulk2type(req), desc->bd_nob_transferred,
3252                                   desc->bd_nob);
3253                         /* XXX should this be a different errno? */
3254                         rc = -ETIMEDOUT;
3255                 }
3256         }
3257
3258         RETURN(rc);
3259 }
3260 EXPORT_SYMBOL(target_bulk_io);
3261
3262 #endif /* HAVE_SERVER_SUPPORT */