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_LIST_HEAD(&cli->cl_flight_waiters);
 395         cli->cl_rpcs_in_flight = 0;
 396
 397         init_waitqueue_head(&cli->cl_destroy_waitq);
 398         atomic_set(&cli->cl_destroy_in_flight, 0);
 399 #ifdef ENABLE_CHECKSUM
 400         /* Turn on checksumming by default. */
 401         cli->cl_checksum = 1;
 402         /*
 403          * The supported checksum types will be worked out at connect time
 404          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 405          * through procfs.
 406          */
 407         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 408 #endif
 409         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 410
 411         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
 412          * from OFD after connecting. */
 413         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 414
 415         cli->cl_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
 416
 417         /* set cl_chunkbits default value to PAGE_SHIFT,
 418          * it will be updated at OSC connection time. */
 419         cli->cl_chunkbits = PAGE_SHIFT;
 420
 421         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 422                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 423         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 424                 cli->cl_max_rpcs_in_flight = 2;
 425         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 426                 cli->cl_max_rpcs_in_flight = 3;
 427         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 428                 cli->cl_max_rpcs_in_flight = 4;
 429         } else {
 430                 if (osc_on_mdt(obddev->obd_name))
 431                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 432                 else
 433                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 434         }
 435
 436         spin_lock_init(&cli->cl_mod_rpcs_lock);
 437         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 438         cli->cl_max_mod_rpcs_in_flight = 0;
 439         cli->cl_mod_rpcs_in_flight = 0;
 440         cli->cl_close_rpcs_in_flight = 0;
 441         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 442         cli->cl_mod_tag_bitmap = NULL;
 443
 444         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 445
 446         if (connect_op == MDS_CONNECT) {
 447                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 448                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 449                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 450                 if (cli->cl_mod_tag_bitmap == NULL)
 451                         GOTO(err, rc = -ENOMEM);
 452         }
 453
 454         rc = ldlm_get_ref();
 455         if (rc) {
 456                 CERROR("ldlm_get_ref failed: %d\n", rc);
 457                 GOTO(err, rc);
 458         }
 459
 460         ptlrpc_init_client(rq_portal, rp_portal, name,
 461                            &obddev->obd_ldlm_client);
 462
 463         imp = class_new_import(obddev);
 464         if (imp == NULL)
 465                 GOTO(err_ldlm, rc = -ENOENT);
 466         imp->imp_client = &obddev->obd_ldlm_client;
 467         imp->imp_connect_op = connect_op;
 468         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 469                LUSTRE_CFG_BUFLEN(lcfg, 1));
 470         class_import_put(imp);
 471
 472         if (lustre_cfg_buf(lcfg, 4)) {
 473                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 474
 475                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 476                         rc = -EINVAL;
 477                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 478                                obddev->obd_name, lustre_cfg_string(lcfg, 4),
 479                                rc);
 480                         GOTO(err_import, rc);
 481                 }
 482                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 483                 imp->imp_connection = &fake_conn;
 484         }
 485
 486         rc = client_import_add_conn(imp, &server_uuid, 1);
 487         if (rc) {
 488                 CERROR("can't add initial connection\n");
 489                 GOTO(err_import, rc);
 490         }
 491         imp->imp_connection = NULL;
 492
 493         cli->cl_import = imp;
 494         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 495         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 496
 497         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 498                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 499                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 500                                name, obddev->obd_name,
 501                                cli->cl_target_uuid.uuid);
 502                         spin_lock(&imp->imp_lock);
 503                         imp->imp_deactive = 1;
 504                         spin_unlock(&imp->imp_lock);
 505                 }
 506         }
 507
 508         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 509                                                    LDLM_NAMESPACE_CLIENT,
 510                                                    LDLM_NAMESPACE_GREEDY,
 511                                                    ns_type);
 512         if (obddev->obd_namespace == NULL) {
 513                 CERROR("Unable to create client namespace - %s\n",
 514                        obddev->obd_name);
 515                 GOTO(err_import, rc = -ENOMEM);
 516         }
 517
 518         RETURN(rc);
 519
 520 err_import:
 521         class_destroy_import(imp);
 522 err_ldlm:
 523         ldlm_put_ref();
 524 err:
 525         if (cli->cl_mod_tag_bitmap != NULL)
 526                 OBD_FREE(cli->cl_mod_tag_bitmap,
 527                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 528         cli->cl_mod_tag_bitmap = NULL;
 529         RETURN(rc);
 530
 531 }
 532 EXPORT_SYMBOL(client_obd_setup);
 533
 534 int client_obd_cleanup(struct obd_device *obddev)
 535 {
 536         struct client_obd *cli = &obddev->u.cli;
 537         ENTRY;
 538
 539         ldlm_namespace_free_post(obddev->obd_namespace);
 540         obddev->obd_namespace = NULL;
 541
 542         obd_cleanup_client_import(obddev);
 543         LASSERT(obddev->u.cli.cl_import == NULL);
 544
 545         ldlm_put_ref();
 546
 547         if (cli->cl_mod_tag_bitmap != NULL)
 548                 OBD_FREE(cli->cl_mod_tag_bitmap,
 549                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 550         cli->cl_mod_tag_bitmap = NULL;
 551
 552         RETURN(0);
 553 }
 554 EXPORT_SYMBOL(client_obd_cleanup);
 555
 556 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 557 int client_connect_import(const struct lu_env *env,
 558                           struct obd_export **exp,
 559                           struct obd_device *obd, struct obd_uuid *cluuid,
 560                           struct obd_connect_data *data, void *localdata)
 561 {
 562         struct client_obd       *cli    = &obd->u.cli;
 563         struct obd_import       *imp    = cli->cl_import;
 564         struct obd_connect_data *ocd;
 565         struct lustre_handle    conn    = { 0 };
 566         int                     rc;
 567         ENTRY;
 568
 569         *exp = NULL;
 570         down_write(&cli->cl_sem);
 571         if (cli->cl_conn_count > 0)
 572                 GOTO(out_sem, rc = -EALREADY);
 573
 574         rc = class_connect(&conn, obd, cluuid);
 575         if (rc)
 576                 GOTO(out_sem, rc);
 577
 578         cli->cl_conn_count++;
 579         *exp = class_conn2export(&conn);
 580
 581         LASSERT(obd->obd_namespace);
 582
 583         imp->imp_dlm_handle = conn;
 584         rc = ptlrpc_init_import(imp);
 585         if (rc != 0)
 586                 GOTO(out_ldlm, rc);
 587
 588         ocd = &imp->imp_connect_data;
 589         if (data) {
 590                 *ocd = *data;
 591                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 592                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 593         }
 594
 595         rc = ptlrpc_connect_import(imp);
 596         if (rc != 0) {
 597                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 598                 GOTO(out_ldlm, rc);
 599         }
 600         LASSERT(*exp != NULL && (*exp)->exp_connection);
 601
 602         if (data) {
 603                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 604                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 605                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 606                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 607                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 608         }
 609
 610         ptlrpc_pinger_add_import(imp);
 611
 612         EXIT;
 613
 614         if (rc) {
 615 out_ldlm:
 616                 cli->cl_conn_count--;
 617                 class_disconnect(*exp);
 618                 *exp = NULL;
 619         }
 620 out_sem:
 621         up_write(&cli->cl_sem);
 622
 623         return rc;
 624 }
 625 EXPORT_SYMBOL(client_connect_import);
 626
 627 int client_disconnect_export(struct obd_export *exp)
 628 {
 629         struct obd_device *obd = class_exp2obd(exp);
 630         struct client_obd *cli;
 631         struct obd_import *imp;
 632         int rc = 0, err;
 633         ENTRY;
 634
 635         if (!obd) {
 636                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 637                        exp, exp ? exp->exp_handle.h_cookie : -1);
 638                 RETURN(-EINVAL);
 639         }
 640
 641         cli = &obd->u.cli;
 642         imp = cli->cl_import;
 643
 644         down_write(&cli->cl_sem);
 645         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 646                 cli->cl_conn_count);
 647
 648         if (cli->cl_conn_count == 0) {
 649                 CERROR("disconnecting disconnected device (%s)\n",
 650                        obd->obd_name);
 651                 GOTO(out_disconnect, rc = -EINVAL);
 652         }
 653
 654         cli->cl_conn_count--;
 655         if (cli->cl_conn_count != 0)
 656                 GOTO(out_disconnect, rc = 0);
 657
 658         /* Mark import deactivated now, so we don't try to reconnect if any
 659          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 660          * fully deactivate the import, or that would drop all requests. */
 661         spin_lock(&imp->imp_lock);
 662         imp->imp_deactive = 1;
 663         spin_unlock(&imp->imp_lock);
 664
 665         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 666          * delete it regardless.  (It's safe to delete an import that was
 667          * never added.) */
 668         (void)ptlrpc_pinger_del_import(imp);
 669
 670         if (obd->obd_namespace != NULL) {
 671                 /* obd_force == local only */
 672                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 673                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 674                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 675         }
 676
 677         /* There's no need to hold sem while disconnecting an import,
 678          * and it may actually cause deadlock in GSS. */
 679         up_write(&cli->cl_sem);
 680         rc = ptlrpc_disconnect_import(imp, 0);
 681         down_write(&cli->cl_sem);
 682
 683         ptlrpc_invalidate_import(imp);
 684
 685         EXIT;
 686
 687 out_disconnect:
 688         /* Use server style - class_disconnect should be always called for
 689          * o_disconnect. */
 690         err = class_disconnect(exp);
 691         if (!rc && err)
 692                 rc = err;
 693
 694         up_write(&cli->cl_sem);
 695
 696         RETURN(rc);
 697 }
 698 EXPORT_SYMBOL(client_disconnect_export);
 699
 700 #ifdef HAVE_SERVER_SUPPORT
 701 int server_disconnect_export(struct obd_export *exp)
 702 {
 703         int rc;
 704         ENTRY;
 705
 706         /* Disconnect early so that clients can't keep using export. */
 707         rc = class_disconnect(exp);
 708         /* Close import to avoid sending any requests. */
 709         if (exp->exp_imp_reverse)
 710                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 711
 712         ldlm_bl_thread_wakeup();
 713
 714         /* complete all outstanding replies */
 715         spin_lock(&exp->exp_lock);
 716         while (!list_empty(&exp->exp_outstanding_replies)) {
 717                 struct ptlrpc_reply_state *rs =
 718                         list_entry(exp->exp_outstanding_replies.next,
 719                                        struct ptlrpc_reply_state, rs_exp_list);
 720                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 721
 722                 spin_lock(&svcpt->scp_rep_lock);
 723
 724                 list_del_init(&rs->rs_exp_list);
 725
 726                 spin_lock(&rs->rs_lock);
 727                 /* clear rs_convert_lock to make sure rs is handled and put */
 728                 rs->rs_convert_lock = 0;
 729                 ptlrpc_schedule_difficult_reply(rs);
 730                 spin_unlock(&rs->rs_lock);
 731
 732                 spin_unlock(&svcpt->scp_rep_lock);
 733         }
 734         spin_unlock(&exp->exp_lock);
 735
 736         RETURN(rc);
 737 }
 738 EXPORT_SYMBOL(server_disconnect_export);
 739
 740 /* --------------------------------------------------------------------------
 741  * from old lib/target.c
 742  * -------------------------------------------------------------------------- */
 743
 744 static int target_handle_reconnect(struct lustre_handle *conn,
 745                                    struct obd_export *exp,
 746                                    struct obd_uuid *cluuid)
 747 {
 748         struct obd_device *target;
 749         struct lustre_handle *hdl;
 750         time64_t deadline;
 751         time64_t timeout;
 752         time64_t now;
 753         int rc = 0;
 754
 755         ENTRY;
 756         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 757         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 758                 conn->cookie = exp->exp_handle.h_cookie;
 759                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 760                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
 761                 RETURN(0);
 762         }
 763
 764         target = exp->exp_obd;
 765
 766         /* Might be a re-connect after a partition. */
 767         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 768                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 769                               "with handle %#llx. Rejecting client "
 770                               "with the same UUID trying to reconnect "
 771                               "with handle %#llx\n", target->obd_name,
 772                               obd_uuid2str(&exp->exp_client_uuid),
 773                               obd_export_nid2str(exp),
 774                               hdl->cookie, conn->cookie);
 775                 memset(conn, 0, sizeof *conn);
 776                 /* target_handle_connect() treats EALREADY and
 777                  * -EALREADY differently.  -EALREADY is an error
 778                  * (same UUID, different handle). */
 779                 RETURN(-EALREADY);
 780         }
 781
 782         if (!target->obd_recovering) {
 783                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 784                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 785                         obd_export_nid2str(exp));
 786                 GOTO(out_already, rc);
 787         }
 788
 789         now = ktime_get_seconds();
 790         deadline = cfs_duration_sec(target->obd_recovery_timer.expires);
 791         if (now < deadline) {
 792                 struct target_distribute_txn_data *tdtd;
 793                 int size = 0;
 794                 int count = 0;
 795                 char *buf = NULL;
 796
 797                 timeout = deadline - now;
 798                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 799                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 800                         buf = tdtd->tdtd_show_update_logs_retrievers(
 801                                 tdtd->tdtd_show_retrievers_cbdata,
 802                                 &size, &count);
 803
 804                 if (count > 0)
 805                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 806                                       "%lld:%.02lld. It is due to DNE recovery "
 807                                       "failed/stuck on the %d MDT(s):%s. "
 808                                       "Please wait until all MDTs recovered "
 809                                       "or abort the recovery by force.\n",
 810                                       target->obd_name, timeout / 60,
 811                                       timeout % 60, count,
 812                                       buf ? buf : "unknown (not enough RAM)");
 813                 else
 814                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 815                                       "%lld:%.02lld. If you do not want to wait "
 816                                       "more, please abort the recovery by "
 817                                       "force.\n", target->obd_name,
 818                                       timeout / 60, timeout % 60);
 819
 820                 if (buf != NULL)
 821                         OBD_FREE(buf, size);
 822         } else {
 823                 timeout = now - deadline;
 824                 LCONSOLE_WARN("%s: Recovery already passed deadline"
 825                         " %lld:%.02lld, It is most likely due to DNE"
 826                         " recovery is failed or stuck, please wait a"
 827                         " few more minutes or abort the recovery.\n",
 828                         target->obd_name, timeout / 60, timeout % 60);
 829         }
 830
 831 out_already:
 832         conn->cookie = exp->exp_handle.h_cookie;
 833         /* target_handle_connect() treats EALREADY and
 834          * -EALREADY differently.  EALREADY means we are
 835          * doing a valid reconnect from the same client. */
 836         RETURN(EALREADY);
 837 }
 838
 839 static void
 840 check_and_start_recovery_timer(struct obd_device *obd,
 841                                struct ptlrpc_request *req, int new_client);
 842
 843 /**
 844  * update flags for import during reconnect process
 845  */
 846 static int rev_import_flags_update(struct obd_import *revimp,
 847                                    struct ptlrpc_request *req)
 848 {
 849         int rc;
 850         struct obd_connect_data *data;
 851
 852         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 853
 854         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 855                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 856         else
 857                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 858
 859         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 860
 861         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 862         if (rc) {
 863                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 864                         revimp->imp_client->cli_name,
 865                         libcfs_id2str(req->rq_peer), rc);
 866                 return rc;
 867         }
 868
 869         return 0;
 870 }
 871
 872 /**
 873  * Allocate a new reverse import for an export.
 874  *
 875  * \retval -errno in case error hit
 876  * \retval 0 if reverse import correctly init
 877  **/
 878 int rev_import_init(struct obd_export *export)
 879 {
 880         struct obd_device *obd = export->exp_obd;
 881         struct obd_import *revimp;
 882
 883         LASSERT(export->exp_imp_reverse == NULL);
 884
 885         revimp = class_new_import(obd);
 886         if (revimp == NULL)
 887                 return -ENOMEM;
 888
 889         revimp->imp_remote_handle.cookie = 0ULL;
 890         revimp->imp_client = &obd->obd_ldlm_client;
 891         revimp->imp_dlm_fake = 1;
 892
 893         /* it is safe to connect import in new state as no sends possible */
 894         spin_lock(&export->exp_lock);
 895         export->exp_imp_reverse = revimp;
 896         spin_unlock(&export->exp_lock);
 897         class_import_put(revimp);
 898
 899         return 0;
 900 }
 901 EXPORT_SYMBOL(rev_import_init);
 902
 903 /**
 904  * Handle reconnect for an export.
 905  *
 906  * \param exp export to handle reconnect process
 907  * \param req client reconnect request
 908  *
 909  * \retval -rc in case securitfy flavor can't be changed
 910  * \retval 0 in case none problems
 911  */
 912 static int rev_import_reconnect(struct obd_export *exp,
 913                                 struct ptlrpc_request *req)
 914 {
 915         struct obd_import *revimp = exp->exp_imp_reverse;
 916         struct lustre_handle *lh;
 917         int rc;
 918
 919         /* avoid sending a request until import flags are changed */
 920         ptlrpc_import_enter_resend(revimp);
 921
 922         if (revimp->imp_connection != NULL)
 923                 ptlrpc_connection_put(revimp->imp_connection);
 924
 925         /*
 926          * client from recovery don't have a handle so we need to take from
 927          * request. it may produce situation when wrong client connected
 928          * to recovery as we trust a client uuid
 929          */
 930         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 931         revimp->imp_remote_handle = *lh;
 932
 933         /* unknown versions will be caught in
 934          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 935         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 936
 937         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 938
 939         rc = rev_import_flags_update(revimp, req);
 940         if (rc != 0) {
 941                 /* it is safe to still be in RECOVERY phase as we are not able
 942                  * to setup correct security flavor so requests are not able to
 943                  * be delivered correctly */
 944                 return rc;
 945         }
 946
 947         /* resend all rpc's via new connection */
 948         return ptlrpc_import_recovery_state_machine(revimp);
 949 }
 950
 951 int target_handle_connect(struct ptlrpc_request *req)
 952 {
 953         struct obd_device *target = NULL;
 954         struct obd_export *export = NULL;
 955         /* connect handle - filled from target_handle_reconnect in
 956          * reconnect case */
 957         struct lustre_handle conn;
 958         struct lustre_handle *tmp;
 959         struct obd_uuid cluuid;
 960         char *str;
 961         int rc = 0;
 962         char *target_start;
 963         int target_len;
 964         bool     mds_conn = false, lw_client = false, initial_conn = false;
 965         bool     mds_mds_conn = false;
 966         bool     new_mds_mds_conn = false;
 967         struct obd_connect_data *data, *tmpdata;
 968         int size, tmpsize;
 969         lnet_nid_t *client_nid = NULL;
 970         ENTRY;
 971
 972         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 973
 974         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 975         if (str == NULL) {
 976                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 977                 GOTO(out, rc = -EINVAL);
 978         }
 979
 980         target = class_dev_by_str(str);
 981         if (!target) {
 982                 deuuidify(str, NULL, &target_start, &target_len);
 983                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 984                                    "from %s (no target). If you are running "
 985                                    "an HA pair check that the target is "
 986                                    "mounted on the other server.\n", str,
 987                                    libcfs_nid2str(req->rq_peer.nid));
 988                 GOTO(out, rc = -ENODEV);
 989         }
 990
 991         spin_lock(&target->obd_dev_lock);
 992
 993         target->obd_conn_inprogress++;
 994
 995         if (target->obd_stopping || !target->obd_set_up) {
 996                 spin_unlock(&target->obd_dev_lock);
 997
 998                 deuuidify(str, NULL, &target_start, &target_len);
 999                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1000                               target_len, target_start,
1001                               libcfs_nid2str(req->rq_peer.nid),
1002                               (target->obd_stopping ?
1003                                "stopping" : "not set up"));
1004                 GOTO(out, rc = -ENODEV);
1005         }
1006
1007         if (target->obd_no_conn) {
1008                 spin_unlock(&target->obd_dev_lock);
1009
1010                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1011                                "from %s\n", target->obd_name,
1012                                libcfs_nid2str(req->rq_peer.nid));
1013                 GOTO(out, rc = -EAGAIN);
1014         }
1015
1016         spin_unlock(&target->obd_dev_lock);
1017
1018         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1019         if (str == NULL) {
1020                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1021                 GOTO(out, rc = -EINVAL);
1022         }
1023
1024         obd_str2uuid(&cluuid, str);
1025
1026         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1027         if (tmp == NULL)
1028                 GOTO(out, rc = -EPROTO);
1029
1030         conn = *tmp;
1031
1032         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1033                                     RCL_CLIENT);
1034         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1035         if (!data)
1036                 GOTO(out, rc = -EPROTO);
1037
1038         rc = req_capsule_server_pack(&req->rq_pill);
1039         if (rc)
1040                 GOTO(out, rc);
1041
1042 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1043         /* Don't allow clients to connect that are using old 1.8 format
1044          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1045          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1046          * FULL20 flag should be set on all connections since 2.0, but no
1047          * longer affects behaviour.
1048          *
1049          * Later this check will be disabled and the flag can be retired
1050          * completely once interop with 3.0 is no longer needed.
1051          */
1052         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1053                 GOTO(out, rc = -EPROTO);
1054
1055         /* Don't allow liblustre clients to connect.
1056          * - testing was disabled in v2_2_50_0-61-g6a75d65
1057          * - building was disabled in v2_5_58_0-28-g7277179
1058          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1059          * - clients were refused connect for version difference > 0.0.1.32  */
1060         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1061                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1062                 GOTO(out, rc = -EPROTO);
1063         }
1064 #endif
1065
1066         /* Note: lw_client is needed in MDS-MDS failover during update log
1067          * processing, so we needs to allow lw_client to be connected at
1068          * anytime, instead of only the initial connection */
1069         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1070
1071         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1072                 initial_conn = true;
1073                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1074                 mds_mds_conn = (data->ocd_connect_flags &
1075                                 OBD_CONNECT_MDS_MDS) != 0;
1076
1077                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1078                  * for Imperative Recovery connection from MGC to MGS.
1079                  *
1080                  * Via check OBD_CONNECT_FID, we can distinguish whether
1081                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1082                  * MGC or MDT. */
1083                 if (!lw_client &&
1084                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1085                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1086                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1087                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1088                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1089                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1090
1091                         /* We do not support the MDT-MDT interoperations with
1092                          * different version MDT because of protocol changes. */
1093                         if (unlikely(major != LUSTRE_MAJOR ||
1094                                      minor != LUSTRE_MINOR ||
1095                                      abs(patch - LUSTRE_PATCH) > 3)) {
1096                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1097                                         "connection from different version MDT "
1098                                         "(%d.%d.%d.%d) %s %s\n",
1099                                         target->obd_name, LUSTRE_MAJOR,
1100                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1101                                         major, minor, patch,
1102                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1103                                         libcfs_nid2str(req->rq_peer.nid), str);
1104
1105                                 GOTO(out, rc = -EPROTO);
1106                         }
1107                 }
1108         }
1109
1110         /* lctl gets a backstage, all-access pass. */
1111         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1112                 goto dont_check_exports;
1113
1114         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1115         if (!export)
1116                 goto no_export;
1117
1118         /* We've found an export in the hash. */
1119
1120         spin_lock(&export->exp_lock);
1121
1122         if (export->exp_connecting) { /* bug 9635, et. al. */
1123                 spin_unlock(&export->exp_lock);
1124                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1125                               export->exp_obd->obd_name, export,
1126                               libcfs_nid2str(req->rq_peer.nid));
1127                 class_export_put(export);
1128                 export = NULL;
1129                 rc = -EALREADY;
1130         } else if ((mds_conn || (lw_client && initial_conn) ||
1131                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1132                    export->exp_connection != NULL) {
1133                 spin_unlock(&export->exp_lock);
1134                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1135                         /* MDS or LWP reconnected after failover. */
1136                         LCONSOLE_WARN("%s: Received %s connection from "
1137                             "%s, removing former export from %s\n",
1138                             target->obd_name, mds_conn ? "MDS" : "LWP",
1139                             libcfs_nid2str(req->rq_peer.nid),
1140                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1141                 } else {
1142                         /* New MDS connection from the same NID. */
1143                         LCONSOLE_WARN("%s: Received new %s connection from "
1144                                 "%s, removing former export from same NID\n",
1145                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1146                                 libcfs_nid2str(req->rq_peer.nid));
1147                 }
1148
1149                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1150                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1151                         /* Because exports between MDTs will always be
1152                          * kept, let's do not fail such export if they
1153                          * come from the same NID, otherwise it might
1154                          * cause eviction between MDTs, which might
1155                          * cause namespace inconsistency */
1156                         spin_lock(&export->exp_lock);
1157                         export->exp_connecting = 1;
1158                         spin_unlock(&export->exp_lock);
1159                         conn.cookie = export->exp_handle.h_cookie;
1160                         rc = EALREADY;
1161                 } else {
1162                         class_fail_export(export);
1163                         class_export_put(export);
1164                         export = NULL;
1165                         rc = 0;
1166                 }
1167         } else if (export->exp_connection != NULL && initial_conn &&
1168                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1169                 spin_unlock(&export->exp_lock);
1170                 /* In MDS failover we have static UUID but NID can change. */
1171                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1172                               "existing nid %s is already connected\n",
1173                               target->obd_name, cluuid.uuid,
1174                               libcfs_nid2str(req->rq_peer.nid),
1175                               libcfs_nid2str(
1176                                       export->exp_connection->c_peer.nid));
1177                 rc = -EALREADY;
1178                 class_export_put(export);
1179                 export = NULL;
1180         } else {
1181                 export->exp_connecting = 1;
1182                 spin_unlock(&export->exp_lock);
1183                 LASSERT(export->exp_obd == target);
1184
1185                 rc = target_handle_reconnect(&conn, export, &cluuid);
1186         }
1187
1188         /* If we found an export, we already unlocked. */
1189         if (!export) {
1190 no_export:
1191                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1192         } else if (req->rq_export == NULL &&
1193                    atomic_read(&export->exp_rpc_count) > 0) {
1194                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1195                               "still busy with %d references\n",
1196                               target->obd_name, cluuid.uuid,
1197                               libcfs_nid2str(req->rq_peer.nid),
1198                               atomic_read(&export->exp_refcount));
1199                 GOTO(out, rc = -EBUSY);
1200         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1201                 if (!strstr(cluuid.uuid, "mdt"))
1202                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1203                                       "known client %s (at %s) because it "
1204                                       "is indicating it is a new client",
1205                                       target->obd_name, cluuid.uuid,
1206                                       libcfs_nid2str(req->rq_peer.nid));
1207                 GOTO(out, rc = -EALREADY);
1208         } else {
1209                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1210         }
1211
1212         if (rc < 0) {
1213                 GOTO(out, rc);
1214         }
1215
1216         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1217                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1218                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1219                export, ktime_get_real_seconds(),
1220                export ? export->exp_last_request_time : 0);
1221
1222         /* If this is the first time a client connects, reset the recovery
1223          * timer. Discard lightweight connections which might be local. */
1224         if (!lw_client && rc == 0 && target->obd_recovering)
1225                 check_and_start_recovery_timer(target, req, export == NULL);
1226
1227         /* We want to handle EALREADY but *not* -EALREADY from
1228          * target_handle_reconnect(), return reconnection state in a flag. */
1229         if (rc == EALREADY) {
1230                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1231                 rc = 0;
1232         } else {
1233                 LASSERT(rc == 0);
1234         }
1235
1236         /* Tell the client if we support replayable requests. */
1237         if (target->obd_replayable)
1238                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1239         client_nid = &req->rq_peer.nid;
1240
1241         if (export == NULL) {
1242                 /* allow lightweight connections during recovery */
1243                 /* allow "new" MDT to be connected during recovery, since we
1244                  * need retrieve recovery update records from it */
1245                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1246                         time64_t t;
1247                         int c; /* connected */
1248                         int i; /* in progress */
1249                         int k; /* known */
1250                         int s; /* stale/evicted */
1251
1252                         c = atomic_read(&target->obd_connected_clients);
1253                         i = atomic_read(&target->obd_lock_replay_clients);
1254                         k = target->obd_max_recoverable_clients;
1255                         s = target->obd_stale_clients;
1256                         t = target->obd_recovery_timer.expires;
1257                         t = cfs_duration_sec(target->obd_recovery_timer.expires);
1258                         t -= ktime_get_seconds();
1259                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1260                                       "(at %s), waiting for %d known clients "
1261                                       "(%d recovered, %d in progress, and %d "
1262                                       "evicted) to recover in %lld:%.02lld\n",
1263                                       target->obd_name, cluuid.uuid,
1264                                       libcfs_nid2str(req->rq_peer.nid), k,
1265                                       c - i, i, s, t / 60, t % 60);
1266                         rc = -EBUSY;
1267                 } else {
1268 dont_check_exports:
1269                         rc = obd_connect(req->rq_svc_thread->t_env,
1270                                          &export, target, &cluuid, data,
1271                                          client_nid);
1272                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1273                                 lustre_msg_add_op_flags(req->rq_repmsg,
1274                                                         MSG_CONNECT_RECOVERING);
1275                         if (rc == 0) {
1276                                 conn.cookie = export->exp_handle.h_cookie;
1277                                 rc = rev_import_init(export);
1278                         }
1279
1280                         if (mds_mds_conn)
1281                                 new_mds_mds_conn = true;
1282                 }
1283         } else {
1284                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1285                                    export, target, &cluuid, data, client_nid);
1286         }
1287         if (rc)
1288                 GOTO(out, rc);
1289
1290         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1291         data->ocd_instance = target->u.obt.obt_instance;
1292
1293         /* Return only the parts of obd_connect_data that we understand, so the
1294          * client knows that we don't understand the rest. */
1295         if (data) {
1296                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1297                                                RCL_SERVER);
1298                 tmpdata = req_capsule_server_get(&req->rq_pill,
1299                                                  &RMF_CONNECT_DATA);
1300                 /* Don't use struct assignment here, because the client reply
1301                  * buffer may be smaller/larger than the local struct
1302                  * obd_connect_data. */
1303                 memcpy(tmpdata, data, min(tmpsize, size));
1304         }
1305
1306         /* If the client and the server are the same node, we will already
1307          * have an export that really points to the client's DLM export,
1308          * because we have a shared handles table.
1309          *
1310          * XXX this will go away when shaver stops sending the "connect" handle
1311          * in the real "remote handle" field of the request --phik 24 Apr 2003
1312          */
1313         ptlrpc_request_change_export(req, export);
1314
1315         spin_lock(&export->exp_lock);
1316         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1317                 spin_unlock(&export->exp_lock);
1318                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1319                        "or equal conn_cnt: %d >= %d\n",
1320                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1321                        export->exp_conn_cnt,
1322                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1323
1324                 GOTO(out, rc = -EALREADY);
1325         }
1326         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1327         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1328         spin_unlock(&export->exp_lock);
1329
1330         if (export->exp_connection != NULL) {
1331                 /* Check to see if connection came from another NID. */
1332                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1333                     !hlist_unhashed(&export->exp_nid_hash))
1334                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1335                                      &export->exp_connection->c_peer.nid,
1336                                      &export->exp_nid_hash);
1337
1338                 ptlrpc_connection_put(export->exp_connection);
1339         }
1340
1341         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1342                                                        req->rq_self,
1343                                                        &cluuid);
1344         if (hlist_unhashed(&export->exp_nid_hash))
1345                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1346                              &export->exp_connection->c_peer.nid,
1347                              &export->exp_nid_hash);
1348
1349         lustre_msg_set_handle(req->rq_repmsg, &conn);
1350
1351         rc = rev_import_reconnect(export, req);
1352         if (rc != 0)
1353                 GOTO(out, rc);
1354
1355         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1356                 int has_transno;
1357                 __u64 transno = data->ocd_transno;
1358
1359                 spin_lock(&export->exp_lock);
1360                 /* possible race with class_disconnect_stale_exports,
1361                  * export may be already in the eviction process */
1362                 if (export->exp_failed) {
1363                         spin_unlock(&export->exp_lock);
1364                         GOTO(out, rc = -ENODEV);
1365                 }
1366                 export->exp_in_recovery = 1;
1367                 export->exp_req_replay_needed = 1;
1368                 export->exp_lock_replay_needed = 1;
1369                 spin_unlock(&export->exp_lock);
1370
1371                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1372                                  MSG_CONNECT_TRANSNO);
1373                 if (has_transno && transno == 0)
1374                         CWARN("Connect with zero transno!\n");
1375
1376                 if (has_transno && transno > 0 &&
1377                     transno < target->obd_next_recovery_transno &&
1378                     transno > target->obd_last_committed) {
1379                         /* Another way is to use cmpxchg() to be lock-free. */
1380                         spin_lock(&target->obd_recovery_task_lock);
1381                         if (transno < target->obd_next_recovery_transno)
1382                                 target->obd_next_recovery_transno = transno;
1383                         spin_unlock(&target->obd_recovery_task_lock);
1384                 }
1385
1386                 atomic_inc(&target->obd_req_replay_clients);
1387                 atomic_inc(&target->obd_lock_replay_clients);
1388                 /* Note: MDS-MDS connection is allowed to be connected during
1389                  * recovery, no matter if the exports needs to be recoveried.
1390                  * Because we need retrieve updates logs from all other MDTs.
1391                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1392                  * also needs to be increased to match other recovery checking
1393                  * condition. */
1394                 if (new_mds_mds_conn)
1395                         target->obd_max_recoverable_clients++;
1396                 if (atomic_inc_return(&target->obd_connected_clients) ==
1397                     target->obd_max_recoverable_clients)
1398                         wake_up(&target->obd_next_transno_waitq);
1399         }
1400
1401         /* Tell the client we're in recovery, when client is involved in it. */
1402         if (target->obd_recovering && !lw_client)
1403                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1404
1405 out:
1406         if (export) {
1407                 spin_lock(&export->exp_lock);
1408                 export->exp_connecting = 0;
1409                 spin_unlock(&export->exp_lock);
1410
1411                 class_export_put(export);
1412         }
1413         if (target != NULL) {
1414                 spin_lock(&target->obd_dev_lock);
1415                 target->obd_conn_inprogress--;
1416                 spin_unlock(&target->obd_dev_lock);
1417                 class_decref(target, "find", current);
1418         }
1419         req->rq_status = rc;
1420         RETURN(rc);
1421 }
1422
1423 int target_handle_disconnect(struct ptlrpc_request *req)
1424 {
1425         int rc;
1426         ENTRY;
1427
1428         rc = req_capsule_server_pack(&req->rq_pill);
1429         if (rc)
1430                 RETURN(rc);
1431
1432         /* Keep the rq_export around so we can send the reply. */
1433         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1434
1435         RETURN(0);
1436 }
1437
1438 void target_destroy_export(struct obd_export *exp)
1439 {
1440         struct obd_import       *imp = NULL;
1441         /* exports created from last_rcvd data, and "fake"
1442            exports created by lctl don't have an import */
1443         spin_lock(&exp->exp_lock);
1444         if (exp->exp_imp_reverse != NULL) {
1445                 imp = exp->exp_imp_reverse;
1446                 exp->exp_imp_reverse = NULL;
1447         }
1448         spin_unlock(&exp->exp_lock);
1449         if (imp != NULL)
1450                 client_destroy_import(imp);
1451
1452         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1453         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1454         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1455         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1456 }
1457 EXPORT_SYMBOL(target_destroy_export);
1458
1459 /*
1460  * Recovery functions
1461  */
1462 static void target_request_copy_get(struct ptlrpc_request *req)
1463 {
1464         class_export_rpc_inc(req->rq_export);
1465         LASSERT(list_empty(&req->rq_list));
1466         INIT_LIST_HEAD(&req->rq_replay_list);
1467
1468         /* Increase refcount to keep request in queue. */
1469         atomic_inc(&req->rq_refcount);
1470         /* Let export know it has replays to be handled. */
1471         atomic_inc(&req->rq_export->exp_replay_count);
1472 }
1473
1474 static void target_request_copy_put(struct ptlrpc_request *req)
1475 {
1476         LASSERT(list_empty(&req->rq_replay_list));
1477         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1478
1479         atomic_dec(&req->rq_export->exp_replay_count);
1480         class_export_rpc_dec(req->rq_export);
1481         ptlrpc_server_drop_request(req);
1482 }
1483
1484 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1485 {
1486         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1487         struct obd_export     *exp = req->rq_export;
1488         struct ptlrpc_request *reqiter;
1489         struct ptlrpc_request *dup_req = NULL;
1490         int                    dup = 0;
1491
1492         LASSERT(exp);
1493
1494         spin_lock(&exp->exp_lock);
1495         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1496                                 rq_replay_list) {
1497                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1498                         dup_req = reqiter;
1499                         dup = 1;
1500                         break;
1501                 }
1502         }
1503
1504         if (dup) {
1505                 /* We expect it with RESENT and REPLAY flags. */
1506                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1507                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1508                         CERROR("invalid flags %x of resent replay\n",
1509                                lustre_msg_get_flags(req->rq_reqmsg));
1510
1511                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1512                         __u32 new_conn;
1513
1514                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1515                         if (new_conn >
1516                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1517                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1518                                                         new_conn);
1519                 }
1520         } else {
1521                 list_add_tail(&req->rq_replay_list,
1522                                   &exp->exp_req_replay_queue);
1523         }
1524
1525         spin_unlock(&exp->exp_lock);
1526         return dup;
1527 }
1528
1529 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1530 {
1531         LASSERT(!list_empty(&req->rq_replay_list));
1532         LASSERT(req->rq_export);
1533
1534         spin_lock(&req->rq_export->exp_lock);
1535         list_del_init(&req->rq_replay_list);
1536         spin_unlock(&req->rq_export->exp_lock);
1537 }
1538
1539 static void target_finish_recovery(struct lu_target *lut)
1540 {
1541         struct obd_device *obd = lut->lut_obd;
1542         ENTRY;
1543
1544         /* Only log a recovery message when recovery has occurred. */
1545         if (obd->obd_recovery_start) {
1546                 time64_t now = ktime_get_real_seconds();
1547                 time64_t elapsed_time;
1548
1549                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start, 1);
1550                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients "
1551                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1552                         (s64)elapsed_time / 60, (s64)elapsed_time % 60,
1553                         obd->obd_max_recoverable_clients,
1554                         atomic_read(&obd->obd_connected_clients),
1555                         obd->obd_stale_clients,
1556                         obd->obd_stale_clients == 1 ? "was" : "were");
1557         }
1558
1559         ldlm_reprocess_recovery_done(obd->obd_namespace);
1560         spin_lock(&obd->obd_recovery_task_lock);
1561         if (!list_empty(&obd->obd_req_replay_queue) ||
1562             !list_empty(&obd->obd_lock_replay_queue) ||
1563             !list_empty(&obd->obd_final_req_queue)) {
1564                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1565                        obd->obd_name,
1566                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1567                        list_empty(&obd->obd_lock_replay_queue) ? \
1568                                "" : "lock ",
1569                        list_empty(&obd->obd_final_req_queue) ? \
1570                                "" : "final ");
1571                 spin_unlock(&obd->obd_recovery_task_lock);
1572                 LBUG();
1573         }
1574         spin_unlock(&obd->obd_recovery_task_lock);
1575
1576         obd->obd_recovery_end = ktime_get_real_seconds();
1577
1578         /* When recovery finished, cleanup orphans on MDS and OST. */
1579         if (obd->obd_type && OBP(obd, postrecov)) {
1580                 int rc = OBP(obd, postrecov)(obd);
1581
1582                 if (rc < 0)
1583                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1584                                       obd->obd_name, rc);
1585         }
1586         EXIT;
1587 }
1588
1589 static void abort_req_replay_queue(struct obd_device *obd)
1590 {
1591         struct ptlrpc_request *req, *n;
1592         struct list_head abort_list;
1593
1594         INIT_LIST_HEAD(&abort_list);
1595         spin_lock(&obd->obd_recovery_task_lock);
1596         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1597         spin_unlock(&obd->obd_recovery_task_lock);
1598         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1599                 DEBUG_REQ(D_WARNING, req, "aborted:");
1600                 req->rq_status = -ENOTCONN;
1601                 if (ptlrpc_error(req)) {
1602                         DEBUG_REQ(D_ERROR, req,
1603                                   "failed abort_req_reply; skipping");
1604                 }
1605                 target_exp_dequeue_req_replay(req);
1606                 target_request_copy_put(req);
1607         }
1608 }
1609
1610 static void abort_lock_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_lock_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_ERROR, req, "aborted:");
1621                 req->rq_status = -ENOTCONN;
1622                 if (ptlrpc_error(req)) {
1623                         DEBUG_REQ(D_ERROR, req,
1624                                   "failed abort_lock_reply; skipping");
1625                 }
1626                 target_request_copy_put(req);
1627         }
1628 }
1629
1630 /* Called from a cleanup function if the device is being cleaned up
1631    forcefully.  The exports should all have been disconnected already,
1632    the only thing left to do is
1633      - clear the recovery flags
1634      - cancel the timer
1635      - free queued requests and replies, but don't send replies
1636    Because the obd_stopping flag is set, no new requests should be received.
1637
1638 */
1639 void target_cleanup_recovery(struct obd_device *obd)
1640 {
1641         struct ptlrpc_request *req, *n;
1642         struct list_head clean_list;
1643
1644         INIT_LIST_HEAD(&clean_list);
1645         spin_lock(&obd->obd_dev_lock);
1646         if (!obd->obd_recovering) {
1647                 spin_unlock(&obd->obd_dev_lock);
1648                 EXIT;
1649                 return;
1650         }
1651         obd->obd_recovering = obd->obd_abort_recovery = 0;
1652         spin_unlock(&obd->obd_dev_lock);
1653
1654         spin_lock(&obd->obd_recovery_task_lock);
1655         target_cancel_recovery_timer(obd);
1656         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1657         spin_unlock(&obd->obd_recovery_task_lock);
1658
1659         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1660                 LASSERT(req->rq_reply_state == NULL);
1661                 target_exp_dequeue_req_replay(req);
1662                 target_request_copy_put(req);
1663         }
1664
1665         spin_lock(&obd->obd_recovery_task_lock);
1666         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1667         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1668         spin_unlock(&obd->obd_recovery_task_lock);
1669
1670         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1671                 LASSERT(req->rq_reply_state == NULL);
1672                 target_request_copy_put(req);
1673         }
1674
1675         EXIT;
1676 }
1677 EXPORT_SYMBOL(target_cleanup_recovery);
1678
1679 /* obd_recovery_task_lock should be held */
1680 void target_cancel_recovery_timer(struct obd_device *obd)
1681 {
1682         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1683         del_timer(&obd->obd_recovery_timer);
1684 }
1685
1686 static void target_start_recovery_timer(struct obd_device *obd)
1687 {
1688         if (obd->obd_recovery_start != 0)
1689                 return;
1690
1691         spin_lock(&obd->obd_dev_lock);
1692         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1693                 spin_unlock(&obd->obd_dev_lock);
1694                 return;
1695         }
1696
1697         LASSERT(obd->obd_recovery_timeout != 0);
1698
1699         if (obd->obd_recovery_start != 0) {
1700                 spin_unlock(&obd->obd_dev_lock);
1701                 return;
1702         }
1703
1704         mod_timer(&obd->obd_recovery_timer,
1705                   jiffies + cfs_time_seconds(obd->obd_recovery_timeout));
1706         obd->obd_recovery_start = ktime_get_real_seconds();
1707         spin_unlock(&obd->obd_dev_lock);
1708
1709         LCONSOLE_WARN("%s: Will be in recovery for at least %llu:%02llu, or until %d client%s reconnect%s\n",
1710                       obd->obd_name,
1711                       obd->obd_recovery_timeout / 60,
1712                       obd->obd_recovery_timeout % 60,
1713                       obd->obd_max_recoverable_clients,
1714                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1715                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1716 }
1717
1718 /**
1719  * extend recovery window.
1720  *
1721  * if @extend is true, extend recovery window to have @drt remaining at least;
1722  * otherwise, make sure the recovery timeout value is not less than @drt.
1723  */
1724 static void extend_recovery_timer(struct obd_device *obd, time64_t drt,
1725                                   bool extend)
1726 {
1727         time64_t now;
1728         time64_t end;
1729         time64_t left;
1730         time64_t to;
1731
1732         spin_lock(&obd->obd_dev_lock);
1733         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1734                 spin_unlock(&obd->obd_dev_lock);
1735                 return;
1736         }
1737         LASSERT(obd->obd_recovery_start != 0);
1738
1739         now = ktime_get_real_seconds();
1740         to = obd->obd_recovery_timeout;
1741         end = obd->obd_recovery_start + to;
1742         left = end - now;
1743
1744         if (extend && (drt > left)) {
1745                 to += drt - left;
1746         } else if (!extend && (drt > to)) {
1747                 to = drt;
1748         }
1749
1750         if (to > obd->obd_recovery_time_hard) {
1751                 to = obd->obd_recovery_time_hard;
1752                 CWARN("%s: extended recovery timer reaching hard limit: %lld, extend: %d\n",
1753                       obd->obd_name, to, extend);
1754         }
1755
1756         if (obd->obd_recovery_timeout < to) {
1757                 obd->obd_recovery_timeout = to;
1758                 end = obd->obd_recovery_start + to;
1759                 mod_timer(&obd->obd_recovery_timer,
1760                           jiffies + cfs_time_seconds(end - now));
1761         }
1762         spin_unlock(&obd->obd_dev_lock);
1763
1764         CDEBUG(D_HA, "%s: recovery timer will expire in %lld seconds\n",
1765                 obd->obd_name, (s64)(end - now));
1766 }
1767
1768 /* Reset the timer with each new client connection */
1769 /*
1770  * This timer is actually reconnect_timer, which is for making sure
1771  * the total recovery window is at least as big as my reconnect
1772  * attempt timing. So the initial recovery time_out will be set to
1773  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1774  * from client is bigger than this, then the recovery time_out will
1775  * be extended to make sure the client could be reconnected, in the
1776  * process, the timeout from the new client should be ignored.
1777  */
1778
1779 static void
1780 check_and_start_recovery_timer(struct obd_device *obd,
1781                                struct ptlrpc_request *req,
1782                                int new_client)
1783 {
1784         time64_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1785         struct obd_device_target *obt = &obd->u.obt;
1786
1787         if (!new_client && service_time)
1788                 /* Teach server about old server's estimates, as first guess
1789                  * at how long new requests will take. */
1790                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1791                             service_time);
1792
1793         target_start_recovery_timer(obd);
1794
1795         /* Convert the service time to RPC timeout,
1796          * and reuse service_time to limit stack usage.
1797          */
1798         service_time = at_est2timeout(service_time);
1799
1800         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1801             service_time < at_extra)
1802                 service_time = at_extra;
1803
1804         /* We expect other clients to timeout within service_time, then try
1805          * to reconnect, then try the failover server.  The max delay between
1806          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1807         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1808
1809         LASSERT(obt->obt_magic == OBT_MAGIC);
1810         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1811         if (service_time > obd->obd_recovery_timeout && !new_client)
1812                 extend_recovery_timer(obd, service_time, false);
1813 }
1814
1815 /** Health checking routines */
1816 static inline int exp_connect_healthy(struct obd_export *exp)
1817 {
1818         return (exp->exp_in_recovery);
1819 }
1820
1821 /** if export done req_replay or has replay in queue */
1822 static inline int exp_req_replay_healthy(struct obd_export *exp)
1823 {
1824         return (!exp->exp_req_replay_needed ||
1825                 atomic_read(&exp->exp_replay_count) > 0);
1826 }
1827
1828
1829 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1830 {
1831         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1832                exp_req_replay_healthy(exp);
1833 }
1834
1835 /** if export done lock_replay or has replay in queue */
1836 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1837 {
1838         return (!exp->exp_lock_replay_needed ||
1839                 atomic_read(&exp->exp_replay_count) > 0);
1840 }
1841
1842 static inline int exp_vbr_healthy(struct obd_export *exp)
1843 {
1844         return (!exp->exp_vbr_failed);
1845 }
1846
1847 static inline int exp_finished(struct obd_export *exp)
1848 {
1849         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1850 }
1851
1852 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1853 {
1854         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1855                 exp_finished(exp);
1856 }
1857
1858 static int check_for_next_transno(struct lu_target *lut)
1859 {
1860         struct ptlrpc_request *req = NULL;
1861         struct obd_device *obd = lut->lut_obd;
1862         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1863         int wake_up = 0, connected, completed, queue_len;
1864         __u64 req_transno = 0;
1865         __u64 update_transno = 0;
1866         __u64 next_transno = 0;
1867         ENTRY;
1868
1869         spin_lock(&obd->obd_recovery_task_lock);
1870         if (!list_empty(&obd->obd_req_replay_queue)) {
1871                 req = list_entry(obd->obd_req_replay_queue.next,
1872                                      struct ptlrpc_request, rq_list);
1873                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1874         }
1875
1876         if (tdtd != NULL)
1877                 update_transno = distribute_txn_get_next_transno(tdtd);
1878
1879         connected = atomic_read(&obd->obd_connected_clients);
1880         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1881         queue_len = obd->obd_requests_queued_for_recovery;
1882         next_transno = obd->obd_next_recovery_transno;
1883
1884         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1885                "req_transno: %llu, next_transno: %llu\n",
1886                obd->obd_max_recoverable_clients, connected, completed,
1887                queue_len, req_transno, next_transno);
1888
1889         if (obd->obd_abort_recovery) {
1890                 CDEBUG(D_HA, "waking for aborted recovery\n");
1891                 wake_up = 1;
1892         } else if (obd->obd_recovery_expired) {
1893                 CDEBUG(D_HA, "waking for expired recovery\n");
1894                 wake_up = 1;
1895         } else if (tdtd != NULL && req != NULL &&
1896                    is_req_replayed_by_update(req)) {
1897                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1898                          "next_transno%llu\n", req_transno, next_transno);
1899                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1900                        req_transno);
1901                 wake_up = 1;
1902         } else if (req_transno == next_transno ||
1903                    (update_transno != 0 && update_transno <= next_transno)) {
1904                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1905                 wake_up = 1;
1906         } else if (queue_len > 0 &&
1907                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1908                 /** handle gaps occured due to lost reply or VBR */
1909                 LASSERTF(req_transno >= next_transno,
1910                          "req_transno: %llu, next_transno: %llu\n",
1911                          req_transno, next_transno);
1912                 CDEBUG(D_HA,
1913                        "%s: waking for gap in transno, VBR is %s (skip: "
1914                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1915                        ", next_update %lld last_committed: %lld)\n",
1916                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1917                        next_transno, queue_len, completed, connected,
1918                        req_transno, update_transno, obd->obd_last_committed);
1919                 obd->obd_next_recovery_transno = req_transno;
1920                 wake_up = 1;
1921         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1922                 CDEBUG(D_HA, "waking for completed recovery\n");
1923                 wake_up = 1;
1924         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1925                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1926                        " by fail_lock, waking up (%lld)\n", next_transno);
1927                 obd->obd_next_recovery_transno = req_transno;
1928                 wake_up = 1;
1929         }
1930         spin_unlock(&obd->obd_recovery_task_lock);
1931         return wake_up;
1932 }
1933
1934 static int check_for_next_lock(struct lu_target *lut)
1935 {
1936         struct obd_device *obd = lut->lut_obd;
1937         int wake_up = 0;
1938
1939         spin_lock(&obd->obd_recovery_task_lock);
1940         if (!list_empty(&obd->obd_lock_replay_queue)) {
1941                 CDEBUG(D_HA, "waking for next lock\n");
1942                 wake_up = 1;
1943         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1944                 CDEBUG(D_HA, "waking for completed lock replay\n");
1945                 wake_up = 1;
1946         } else if (obd->obd_abort_recovery) {
1947                 CDEBUG(D_HA, "waking for aborted recovery\n");
1948                 wake_up = 1;
1949         } else if (obd->obd_recovery_expired) {
1950                 CDEBUG(D_HA, "waking for expired recovery\n");
1951                 wake_up = 1;
1952         }
1953         spin_unlock(&obd->obd_recovery_task_lock);
1954
1955         return wake_up;
1956 }
1957
1958 /**
1959  * wait for recovery events,
1960  * check its status with help of check_routine
1961  * evict dead clients via health_check
1962  */
1963 static int target_recovery_overseer(struct lu_target *lut,
1964                                     int (*check_routine)(struct lu_target *),
1965                                     int (*health_check)(struct obd_export *))
1966 {
1967         struct obd_device       *obd = lut->lut_obd;
1968         struct target_distribute_txn_data *tdtd;
1969 repeat:
1970         if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
1971               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
1972                 __u64 next_update_transno = 0;
1973
1974                 /* Only abort the recovery if there are no update recovery
1975                  * left in the queue */
1976                 spin_lock(&obd->obd_recovery_task_lock);
1977                 if (lut->lut_tdtd != NULL) {
1978                         next_update_transno =
1979                                 distribute_txn_get_next_transno(lut->lut_tdtd);
1980
1981                         tdtd = lut->lut_tdtd;
1982                         /* If next_update_transno == 0, it probably because
1983                          * updatelog retrieve threads did not get any records
1984                          * yet, let's wait those threads stopped */
1985                         if (next_update_transno == 0) {
1986                                 struct l_wait_info lwi = { 0 };
1987
1988                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1989                                        atomic_read(
1990                                        &tdtd->tdtd_recovery_threads_count) == 0,
1991                                        &lwi);
1992
1993                                 next_update_transno =
1994                                         distribute_txn_get_next_transno(
1995                                                                 lut->lut_tdtd);
1996                         }
1997                 }
1998
1999                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2000                         obd->obd_next_recovery_transno = next_update_transno;
2001                         spin_unlock(&obd->obd_recovery_task_lock);
2002                         /* Disconnect unfinished exports from clients, and
2003                          * keep connection from MDT to make sure the update
2004                          * recovery will still keep trying until some one
2005                          * manually abort the recovery */
2006                         class_disconnect_stale_exports(obd,
2007                                                 exp_finished_or_from_mdt);
2008                         /* Abort all of replay and replay lock req from
2009                          * clients */
2010                         abort_req_replay_queue(obd);
2011                         abort_lock_replay_queue(obd);
2012                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2013                                ")in the queue.\n", obd->obd_name,
2014                                next_update_transno);
2015                 } else {
2016                         obd->obd_abort_recovery = 1;
2017                         spin_unlock(&obd->obd_recovery_task_lock);
2018                         CWARN("%s recovery is aborted by hard timeout\n",
2019                               obd->obd_name);
2020                 }
2021         }
2022
2023         while (wait_event_timeout(obd->obd_next_transno_waitq,
2024                                   check_routine(lut),
2025                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2026                 /* wait indefinitely for event, but don't trigger watchdog */;
2027
2028         if (obd->obd_abort_recovery) {
2029                 CWARN("recovery is aborted, evict exports in recovery\n");
2030                 if (lut->lut_tdtd != NULL) {
2031                         struct l_wait_info lwi = { 0 };
2032
2033                         tdtd = lut->lut_tdtd;
2034                         /* Let's wait all of the update log recovery thread
2035                          * finished */
2036                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2037                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2038                              &lwi);
2039                         /* Then abort the update recovery list */
2040                         dtrq_list_destroy(lut->lut_tdtd);
2041                 }
2042
2043                 /** evict exports which didn't finish recovery yet */
2044                 class_disconnect_stale_exports(obd, exp_finished);
2045                 return 1;
2046         } else if (obd->obd_recovery_expired) {
2047                 obd->obd_recovery_expired = 0;
2048                 /** If some clients died being recovered, evict them */
2049                 LCONSOLE_WARN("%s: recovery is timed out, "
2050                               "evict stale exports\n", obd->obd_name);
2051                 /** evict cexports with no replay in queue, they are stalled */
2052                 class_disconnect_stale_exports(obd, health_check);
2053
2054                 /** continue with VBR */
2055                 spin_lock(&obd->obd_dev_lock);
2056                 obd->obd_version_recov = 1;
2057                 spin_unlock(&obd->obd_dev_lock);
2058                 /**
2059                  * reset timer, recovery will proceed with versions now,
2060                  * timeout is set just to handle reconnection delays
2061                  */
2062                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2063                 /** Wait for recovery events again, after evicting bad clients */
2064                 goto repeat;
2065         }
2066         return 0;
2067 }
2068
2069 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2070 {
2071         struct obd_device       *obd = lut->lut_obd;
2072         struct ptlrpc_request *req = NULL;
2073
2074         CDEBUG(D_HA, "Waiting for lock\n");
2075         if (target_recovery_overseer(lut, check_for_next_lock,
2076                                      exp_lock_replay_healthy))
2077                 abort_lock_replay_queue(obd);
2078
2079         spin_lock(&obd->obd_recovery_task_lock);
2080         if (!list_empty(&obd->obd_lock_replay_queue)) {
2081                 req = list_entry(obd->obd_lock_replay_queue.next,
2082                                      struct ptlrpc_request, rq_list);
2083                 list_del_init(&req->rq_list);
2084                 spin_unlock(&obd->obd_recovery_task_lock);
2085         } else {
2086                 spin_unlock(&obd->obd_recovery_task_lock);
2087                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2088                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2089                 /** evict exports failed VBR */
2090                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2091         }
2092         return req;
2093 }
2094
2095 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2096 {
2097         struct ptlrpc_request *req = NULL;
2098
2099         spin_lock(&obd->obd_recovery_task_lock);
2100         if (!list_empty(&obd->obd_final_req_queue)) {
2101                 req = list_entry(obd->obd_final_req_queue.next,
2102                                      struct ptlrpc_request, rq_list);
2103                 list_del_init(&req->rq_list);
2104                 spin_unlock(&obd->obd_recovery_task_lock);
2105                 if (req->rq_export->exp_in_recovery) {
2106                         spin_lock(&req->rq_export->exp_lock);
2107                         req->rq_export->exp_in_recovery = 0;
2108                         spin_unlock(&req->rq_export->exp_lock);
2109                 }
2110         } else {
2111                 spin_unlock(&obd->obd_recovery_task_lock);
2112         }
2113         return req;
2114 }
2115
2116 static void handle_recovery_req(struct ptlrpc_thread *thread,
2117                                 struct ptlrpc_request *req,
2118                                 svc_handler_t handler)
2119 {
2120         ENTRY;
2121
2122         /**
2123          * export can be evicted during recovery, no need to handle replays for
2124          * it after that, discard such request silently
2125          */
2126         if (req->rq_export->exp_disconnected)
2127                 RETURN_EXIT;
2128
2129         req->rq_session.lc_thread = thread;
2130         req->rq_svc_thread = thread;
2131         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2132
2133         /* thread context */
2134         lu_context_enter(&thread->t_env->le_ctx);
2135         (void)handler(req);
2136         lu_context_exit(&thread->t_env->le_ctx);
2137
2138         /* don't reset timer for final stage */
2139         if (!exp_finished(req->rq_export)) {
2140                 time64_t to = obd_timeout;
2141
2142                 /**
2143                  * Add request timeout to the recovery time so next request from
2144                  * this client may come in recovery time
2145                  */
2146                 if (!AT_OFF) {
2147                         struct ptlrpc_service_part *svcpt;
2148
2149                         svcpt = req->rq_rqbd->rqbd_svcpt;
2150                         /* If the server sent early reply for this request,
2151                          * the client will recalculate the timeout according to
2152                          * current server estimate service time, so we will
2153                          * use the maxium timeout here for waiting the client
2154                          * sending the next req */
2155                         to = max((int)at_est2timeout(
2156                                  at_get(&svcpt->scp_at_estimate)),
2157                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2158                         /* Add 2 net_latency, one for balance rq_deadline
2159                          * (see ptl_send_rpc), one for resend the req to server,
2160                          * Note: client will pack net_latency in replay req
2161                          * (see ptlrpc_replay_req) */
2162                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2163                 }
2164                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2165         }
2166         EXIT;
2167 }
2168
2169 /** Checking routines for recovery */
2170 static int check_for_recovery_ready(struct lu_target *lut)
2171 {
2172         struct obd_device *obd = lut->lut_obd;
2173         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2174
2175         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2176                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2177                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2178                obd->obd_recovery_expired);
2179
2180         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2181                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2182                 if (clnts + obd->obd_stale_clients <
2183                     obd->obd_max_recoverable_clients)
2184                         return 0;
2185         }
2186
2187         if (lut->lut_tdtd != NULL) {
2188                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2189                     !obd->obd_abort_recovery) {
2190                         /* Let's extend recovery timer, in case the recovery
2191                          * timer expired, and some clients got evicted */
2192                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2193                                               true);
2194                         CDEBUG(D_HA, "%s update recovery is not ready, extend recovery %llu\n",
2195                                obd->obd_name, obd->obd_recovery_timeout);
2196                         return 0;
2197                 }
2198         }
2199
2200         return 1;
2201 }
2202
2203 enum {
2204         REQUEST_RECOVERY = 1,
2205         UPDATE_RECOVERY = 2,
2206 };
2207
2208 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2209 {
2210         __u64 transno = 0;
2211
2212         if (!list_empty(&obd->obd_req_replay_queue)) {
2213                 struct ptlrpc_request *req;
2214
2215                 req = list_entry(obd->obd_req_replay_queue.next,
2216                                  struct ptlrpc_request, rq_list);
2217                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2218         }
2219
2220         return transno;
2221 }
2222
2223 static __u64 get_next_transno(struct lu_target *lut, int *type)
2224 {
2225         struct obd_device *obd = lut->lut_obd;
2226         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2227         __u64 transno = 0;
2228         __u64 update_transno;
2229         ENTRY;
2230
2231         transno = get_next_replay_req_transno(obd);
2232         if (type != NULL)
2233                 *type = REQUEST_RECOVERY;
2234
2235         if (tdtd == NULL)
2236                 RETURN(transno);
2237
2238         update_transno = distribute_txn_get_next_transno(tdtd);
2239         if (transno == 0 || (transno >= update_transno &&
2240                              update_transno != 0)) {
2241                 transno = update_transno;
2242                 if (type != NULL)
2243                         *type = UPDATE_RECOVERY;
2244         }
2245
2246         RETURN(transno);
2247 }
2248
2249 /**
2250  * drop duplicate replay request
2251  *
2252  * Because the operation has been replayed by update recovery, the request
2253  * with the same transno will be dropped and also notify the client to send
2254  * next replay request.
2255  *
2256  * \param[in] env       execution environment
2257  * \param[in] obd       failover obd device
2258  * \param[in] req       request to be dropped
2259  */
2260 static void drop_duplicate_replay_req(struct lu_env *env,
2261                                       struct obd_device *obd,
2262                                       struct ptlrpc_request *req)
2263 {
2264         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2265                   " update records are found.\n",
2266                   lustre_msg_get_transno(req->rq_reqmsg),
2267                   libcfs_nid2str(req->rq_peer.nid));
2268
2269         /* Right now, only for MDS reint operation update replay and
2270          * normal request replay can have the same transno */
2271         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2272                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2273                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2274                 if (likely(req->rq_export))
2275                         target_committed_to_req(req);
2276                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2277                 target_send_reply(req, req->rq_status, 0);
2278         } else {
2279                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2280                 libcfs_nid2str(req->rq_peer.nid));
2281         }
2282         target_exp_dequeue_req_replay(req);
2283         target_request_copy_put(req);
2284         obd->obd_replayed_requests++;
2285 }
2286
2287 static void replay_request_or_update(struct lu_env *env,
2288                                      struct lu_target *lut,
2289                                      struct target_recovery_data *trd,
2290                                      struct ptlrpc_thread *thread)
2291 {
2292         struct obd_device *obd = lut->lut_obd;
2293         struct ptlrpc_request *req = NULL;
2294         int                     type;
2295         __u64                   transno;
2296         ENTRY;
2297
2298         CDEBUG(D_HA, "Waiting for transno %lld\n",
2299                obd->obd_next_recovery_transno);
2300
2301         /* Replay all of request and update by transno */
2302         do {
2303                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2304
2305                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2306
2307                 /** It is needed to extend recovery window above
2308                  *  recovery_time_soft. Extending is possible only in the
2309                  *  end of recovery window (see more details in
2310                  *  handle_recovery_req()).
2311                  */
2312                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2313
2314                 if (target_recovery_overseer(lut, check_for_next_transno,
2315                                         exp_req_replay_healthy_or_from_mdt)) {
2316                         abort_req_replay_queue(obd);
2317                         abort_lock_replay_queue(obd);
2318                         goto abort;
2319                 }
2320
2321                 spin_lock(&obd->obd_recovery_task_lock);
2322                 transno = get_next_transno(lut, &type);
2323                 if (type == REQUEST_RECOVERY && transno != 0) {
2324                         /* Drop replay request from client side, if the
2325                          * replay has been executed by update with the
2326                          * same transno */
2327                         req = list_entry(obd->obd_req_replay_queue.next,
2328                                         struct ptlrpc_request, rq_list);
2329
2330                         list_del_init(&req->rq_list);
2331                         obd->obd_requests_queued_for_recovery--;
2332                         spin_unlock(&obd->obd_recovery_task_lock);
2333
2334                         /* Let's check if the request has been redone by
2335                          * update replay */
2336                         if (is_req_replayed_by_update(req)) {
2337                                 struct distribute_txn_replay_req *dtrq;
2338
2339                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2340                                                                    req->rq_xid);
2341                                 LASSERT(dtrq != NULL);
2342                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2343                                 list_del_init(&dtrq->dtrq_list);
2344                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2345                                 dtrq_destroy(dtrq);
2346
2347                                 drop_duplicate_replay_req(env, obd, req);
2348
2349                                 continue;
2350                         }
2351
2352                         LASSERT(trd->trd_processing_task == current_pid());
2353                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2354                                   lustre_msg_get_transno(req->rq_reqmsg),
2355                                   libcfs_nid2str(req->rq_peer.nid));
2356
2357                         handle_recovery_req(thread, req,
2358                                             trd->trd_recovery_handler);
2359                         /**
2360                          * bz18031: increase next_recovery_transno before
2361                          * target_request_copy_put() will drop exp_rpc reference
2362                          */
2363                         spin_lock(&obd->obd_recovery_task_lock);
2364                         obd->obd_next_recovery_transno++;
2365                         spin_unlock(&obd->obd_recovery_task_lock);
2366                         target_exp_dequeue_req_replay(req);
2367                         target_request_copy_put(req);
2368                         obd->obd_replayed_requests++;
2369                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2370                         struct distribute_txn_replay_req *dtrq;
2371                         int rc;
2372
2373                         spin_unlock(&obd->obd_recovery_task_lock);
2374
2375                         LASSERT(tdtd != NULL);
2376                         dtrq = distribute_txn_get_next_req(tdtd);
2377                         lu_context_enter(&thread->t_env->le_ctx);
2378                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2379                         lu_context_exit(&thread->t_env->le_ctx);
2380                         extend_recovery_timer(obd, obd_timeout, true);
2381
2382                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2383                                 CDEBUG(D_HA, "Move x%llu t%llu"
2384                                        " to finish list\n", dtrq->dtrq_xid,
2385                                        dtrq->dtrq_master_transno);
2386
2387                                 /* Add it to the replay finish list */
2388                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2389                                 list_add(&dtrq->dtrq_list,
2390                                          &tdtd->tdtd_replay_finish_list);
2391                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2392
2393                                 spin_lock(&obd->obd_recovery_task_lock);
2394                                 if (transno == obd->obd_next_recovery_transno)
2395                                         obd->obd_next_recovery_transno++;
2396                                 else if (transno >
2397                                          obd->obd_next_recovery_transno)
2398                                         obd->obd_next_recovery_transno =
2399                                                                 transno + 1;
2400                                 spin_unlock(&obd->obd_recovery_task_lock);
2401                         } else {
2402                                 dtrq_destroy(dtrq);
2403                         }
2404                 } else {
2405                         spin_unlock(&obd->obd_recovery_task_lock);
2406 abort:
2407                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2408                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2409                         /** evict exports failed VBR */
2410                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2411                         break;
2412                 }
2413         } while (1);
2414 }
2415
2416 static int target_recovery_thread(void *arg)
2417 {
2418         struct lu_target *lut = arg;
2419         struct obd_device *obd = lut->lut_obd;
2420         struct ptlrpc_request *req;
2421         struct target_recovery_data *trd = &obd->obd_recovery_data;
2422         unsigned long delta;
2423         struct lu_env *env;
2424         struct ptlrpc_thread *thread = NULL;
2425         int rc = 0;
2426         ENTRY;
2427
2428         unshare_fs_struct();
2429         OBD_ALLOC_PTR(thread);
2430         if (thread == NULL)
2431                 RETURN(-ENOMEM);
2432
2433         OBD_ALLOC_PTR(env);
2434         if (env == NULL) {
2435                 OBD_FREE_PTR(thread);
2436                 RETURN(-ENOMEM);
2437         }
2438
2439         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2440         if (rc) {
2441                 OBD_FREE_PTR(thread);
2442                 OBD_FREE_PTR(env);
2443                 RETURN(rc);
2444         }
2445
2446         thread->t_env = env;
2447         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2448         env->le_ctx.lc_thread = thread;
2449         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2450         thread->t_watchdog = NULL;
2451
2452         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2453                current_pid());
2454         trd->trd_processing_task = current_pid();
2455
2456         spin_lock(&obd->obd_dev_lock);
2457         obd->obd_recovering = 1;
2458         spin_unlock(&obd->obd_dev_lock);
2459         complete(&trd->trd_starting);
2460
2461         /* first of all, we have to know the first transno to replay */
2462         if (target_recovery_overseer(lut, check_for_recovery_ready,
2463                                      exp_connect_healthy)) {
2464                 abort_req_replay_queue(obd);
2465                 abort_lock_replay_queue(obd);
2466                 if (lut->lut_tdtd != NULL)
2467                         dtrq_list_destroy(lut->lut_tdtd);
2468         }
2469
2470         /* next stage: replay requests or update */
2471         delta = jiffies;
2472         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2473                atomic_read(&obd->obd_req_replay_clients),
2474                obd->obd_next_recovery_transno);
2475         replay_request_or_update(env, lut, trd, thread);
2476
2477         /**
2478          * The second stage: replay locks
2479          */
2480         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2481                atomic_read(&obd->obd_lock_replay_clients));
2482         while ((req = target_next_replay_lock(lut))) {
2483                 LASSERT(trd->trd_processing_task == current_pid());
2484                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2485                           libcfs_nid2str(req->rq_peer.nid));
2486                 handle_recovery_req(thread, req,
2487                                     trd->trd_recovery_handler);
2488                 target_request_copy_put(req);
2489                 obd->obd_replayed_locks++;
2490         }
2491
2492         /**
2493          * The third stage: reply on final pings, at this moment all clients
2494          * must have request in final queue
2495          */
2496         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2497         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2498         /** Update server last boot epoch */
2499         tgt_boot_epoch_update(lut);
2500         /* We drop recoverying flag to forward all new requests
2501          * to regular mds_handle() since now */
2502         spin_lock(&obd->obd_dev_lock);
2503         obd->obd_recovering = obd->obd_abort_recovery = 0;
2504         spin_unlock(&obd->obd_dev_lock);
2505         spin_lock(&obd->obd_recovery_task_lock);
2506         target_cancel_recovery_timer(obd);
2507         spin_unlock(&obd->obd_recovery_task_lock);
2508         while ((req = target_next_final_ping(obd))) {
2509                 LASSERT(trd->trd_processing_task == current_pid());
2510                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2511                           libcfs_nid2str(req->rq_peer.nid));
2512                 handle_recovery_req(thread, req,
2513                                     trd->trd_recovery_handler);
2514                 /* Because the waiting client can not send ping to server,
2515                  * so we need refresh the last_request_time, to avoid the
2516                  * export is being evicted */
2517                 ptlrpc_update_export_timer(req->rq_export, 0);
2518                 target_request_copy_put(req);
2519         }
2520
2521         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2522         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2523               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2524         if (delta > OBD_RECOVERY_TIME_SOFT) {
2525                 CWARN("too long recovery - read logs\n");
2526                 libcfs_debug_dumplog();
2527         }
2528
2529         target_finish_recovery(lut);
2530
2531         lu_context_fini(&env->le_ctx);
2532         trd->trd_processing_task = 0;
2533         complete(&trd->trd_finishing);
2534
2535         tgt_io_thread_done(thread);
2536         OBD_FREE_PTR(thread);
2537         OBD_FREE_PTR(env);
2538         RETURN(rc);
2539 }
2540
2541 static int target_start_recovery_thread(struct lu_target *lut,
2542                                         svc_handler_t handler)
2543 {
2544         struct obd_device *obd = lut->lut_obd;
2545         int rc = 0;
2546         struct target_recovery_data *trd = &obd->obd_recovery_data;
2547         int index;
2548
2549         memset(trd, 0, sizeof(*trd));
2550         init_completion(&trd->trd_starting);
2551         init_completion(&trd->trd_finishing);
2552         trd->trd_recovery_handler = handler;
2553
2554         rc = server_name2index(obd->obd_name, &index, NULL);
2555         if (rc < 0)
2556                 return rc;
2557
2558         if (!IS_ERR(kthread_run(target_recovery_thread,
2559                                 lut, "tgt_recover_%d", index))) {
2560                 wait_for_completion(&trd->trd_starting);
2561                 LASSERT(obd->obd_recovering != 0);
2562         } else {
2563                 rc = -ECHILD;
2564         }
2565
2566         return rc;
2567 }
2568
2569 void target_stop_recovery_thread(struct obd_device *obd)
2570 {
2571         if (obd->obd_recovery_data.trd_processing_task > 0) {
2572                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2573                 /** recovery can be done but postrecovery is not yet */
2574                 spin_lock(&obd->obd_dev_lock);
2575                 if (obd->obd_recovering) {
2576                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2577                         obd->obd_abort_recovery = 1;
2578                         wake_up(&obd->obd_next_transno_waitq);
2579                 }
2580                 spin_unlock(&obd->obd_dev_lock);
2581                 wait_for_completion(&trd->trd_finishing);
2582         }
2583 }
2584 EXPORT_SYMBOL(target_stop_recovery_thread);
2585
2586 void target_recovery_fini(struct obd_device *obd)
2587 {
2588         class_disconnect_exports(obd);
2589         target_stop_recovery_thread(obd);
2590         target_cleanup_recovery(obd);
2591 }
2592 EXPORT_SYMBOL(target_recovery_fini);
2593
2594 static void target_recovery_expired(unsigned long castmeharder)
2595 {
2596         struct obd_device *obd = (struct obd_device *)castmeharder;
2597         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2598                " after %llus (%d clients connected)\n",
2599                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2600                (s64)(ktime_get_real_seconds() - obd->obd_recovery_start),
2601                atomic_read(&obd->obd_connected_clients));
2602
2603         obd->obd_recovery_expired = 1;
2604         wake_up(&obd->obd_next_transno_waitq);
2605 }
2606
2607 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2608 {
2609         struct obd_device *obd = lut->lut_obd;
2610
2611         if (lut->lut_bottom->dd_rdonly)
2612                 return;
2613
2614         if (obd->obd_max_recoverable_clients == 0) {
2615                 /** Update server last boot epoch */
2616                 tgt_boot_epoch_update(lut);
2617                 return;
2618         }
2619
2620         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2621                "last_transno %llu\n", obd->obd_name,
2622                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2623         LASSERT(obd->obd_stopping == 0);
2624         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2625         obd->obd_recovery_start = 0;
2626         obd->obd_recovery_end = 0;
2627
2628         setup_timer(&obd->obd_recovery_timer, target_recovery_expired,
2629                     (unsigned long)obd);
2630         target_start_recovery_thread(lut, handler);
2631 }
2632 EXPORT_SYMBOL(target_recovery_init);
2633
2634 static int target_process_req_flags(struct obd_device *obd,
2635                                     struct ptlrpc_request *req)
2636 {
2637         struct obd_export *exp = req->rq_export;
2638         LASSERT(exp != NULL);
2639         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2640                 /* client declares he's ready to replay locks */
2641                 spin_lock(&exp->exp_lock);
2642                 if (exp->exp_req_replay_needed) {
2643                         exp->exp_req_replay_needed = 0;
2644                         spin_unlock(&exp->exp_lock);
2645
2646                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2647                         atomic_dec(&obd->obd_req_replay_clients);
2648                 } else {
2649                         spin_unlock(&exp->exp_lock);
2650                 }
2651         }
2652         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2653                 /* client declares he's ready to complete recovery
2654                  * so, we put the request on th final queue */
2655                 spin_lock(&exp->exp_lock);
2656                 if (exp->exp_lock_replay_needed) {
2657                         exp->exp_lock_replay_needed = 0;
2658                         spin_unlock(&exp->exp_lock);
2659
2660                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2661                         atomic_dec(&obd->obd_lock_replay_clients);
2662                 } else {
2663                         spin_unlock(&exp->exp_lock);
2664                 }
2665         }
2666         return 0;
2667 }
2668
2669 int target_queue_recovery_request(struct ptlrpc_request *req,
2670                                   struct obd_device *obd)
2671 {
2672         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2673         struct ptlrpc_request *reqiter;
2674         int inserted = 0;
2675         ENTRY;
2676
2677         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2678                 /* Processing the queue right now, don't re-add. */
2679                 RETURN(1);
2680         }
2681
2682         target_process_req_flags(obd, req);
2683
2684         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2685                 /* client declares he's ready to complete recovery
2686                  * so, we put the request on th final queue */
2687                 target_request_copy_get(req);
2688                 DEBUG_REQ(D_HA, req, "queue final req");
2689                 wake_up(&obd->obd_next_transno_waitq);
2690                 spin_lock(&obd->obd_recovery_task_lock);
2691                 if (obd->obd_recovering) {
2692                         struct ptlrpc_request *tmp;
2693                         struct ptlrpc_request *duplicate = NULL;
2694
2695                         if (likely(!req->rq_export->exp_replay_done)) {
2696                                 req->rq_export->exp_replay_done = 1;
2697                                 list_add_tail(&req->rq_list,
2698                                               &obd->obd_final_req_queue);
2699                                 spin_unlock(&obd->obd_recovery_task_lock);
2700                                 RETURN(0);
2701                         }
2702
2703                         /* XXX O(n), but only happens if final ping is
2704                          * timed out, probably reorganize the list as
2705                          * a hash list later */
2706                         list_for_each_entry_safe(reqiter, tmp,
2707                                                  &obd->obd_final_req_queue,
2708                                                  rq_list) {
2709                                 if (reqiter->rq_export == req->rq_export) {
2710                                         list_del_init(&reqiter->rq_list);
2711                                         duplicate = reqiter;
2712                                         break;
2713                                 }
2714                         }
2715
2716                         list_add_tail(&req->rq_list,
2717                                       &obd->obd_final_req_queue);
2718                         req->rq_export->exp_replay_done = 1;
2719                         spin_unlock(&obd->obd_recovery_task_lock);
2720
2721                         if (duplicate != NULL) {
2722                                 DEBUG_REQ(D_HA, duplicate,
2723                                           "put prev final req\n");
2724                                 target_request_copy_put(duplicate);
2725                         }
2726                         RETURN(0);
2727                 } else {
2728                         spin_unlock(&obd->obd_recovery_task_lock);
2729                         target_request_copy_put(req);
2730                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2731                 }
2732         }
2733         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2734                 /* client declares he's ready to replay locks */
2735                 target_request_copy_get(req);
2736                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2737                 wake_up(&obd->obd_next_transno_waitq);
2738                 spin_lock(&obd->obd_recovery_task_lock);
2739                 LASSERT(obd->obd_recovering);
2740                 /* usually due to recovery abort */
2741                 if (!req->rq_export->exp_in_recovery) {
2742                         spin_unlock(&obd->obd_recovery_task_lock);
2743                         target_request_copy_put(req);
2744                         RETURN(-ENOTCONN);
2745                 }
2746                 LASSERT(req->rq_export->exp_lock_replay_needed);
2747                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2748                 spin_unlock(&obd->obd_recovery_task_lock);
2749                 RETURN(0);
2750         }
2751
2752         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2753          * (i.e. buflens etc are in my own byte order), but type-dependent
2754          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2755
2756         if (!transno) {
2757                 INIT_LIST_HEAD(&req->rq_list);
2758                 DEBUG_REQ(D_HA, req, "not queueing");
2759                 RETURN(1);
2760         }
2761
2762         /* If we're processing the queue, we want don't want to queue this
2763          * message.
2764          *
2765          * Also, if this request has a transno less than the one we're waiting
2766          * for, we should process it now.  It could (and currently always will)
2767          * be an open request for a descriptor that was opened some time ago.
2768          *
2769          * Also, a resent, replayed request that has already been
2770          * handled will pass through here and be processed immediately.
2771          */
2772         CDEBUG(D_HA, "Next recovery transno: %llu"
2773                ", current: %llu, replaying\n",
2774                obd->obd_next_recovery_transno, transno);
2775
2776         /* If the request has been replayed by update replay, then sends this
2777          * request to the recovery thread (replay_request_or_update()), where
2778          * it will be handled */
2779         spin_lock(&obd->obd_recovery_task_lock);
2780         if (transno < obd->obd_next_recovery_transno &&
2781             !is_req_replayed_by_update(req)) {
2782                 /* Processing the queue right now, don't re-add. */
2783                 LASSERT(list_empty(&req->rq_list));
2784                 spin_unlock(&obd->obd_recovery_task_lock);
2785                 RETURN(1);
2786         }
2787         spin_unlock(&obd->obd_recovery_task_lock);
2788
2789         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2790                 RETURN(0);
2791
2792         target_request_copy_get(req);
2793         if (!req->rq_export->exp_in_recovery) {
2794                 target_request_copy_put(req);
2795                 RETURN(-ENOTCONN);
2796         }
2797         LASSERT(req->rq_export->exp_req_replay_needed);
2798
2799         if (target_exp_enqueue_req_replay(req)) {
2800                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2801                 target_request_copy_put(req);
2802                 RETURN(0);
2803         }
2804
2805         /* XXX O(n^2) */
2806         spin_lock(&obd->obd_recovery_task_lock);
2807         LASSERT(obd->obd_recovering);
2808         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2809                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2810                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2811                         inserted = 1;
2812                         goto added;
2813                 }
2814
2815                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2816                              transno)) {
2817                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2818                                   "has been claimed by another client");
2819                         spin_unlock(&obd->obd_recovery_task_lock);
2820                         target_exp_dequeue_req_replay(req);
2821                         target_request_copy_put(req);
2822                         RETURN(0);
2823                 }
2824         }
2825 added:
2826         if (!inserted)
2827                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2828
2829         obd->obd_requests_queued_for_recovery++;
2830         spin_unlock(&obd->obd_recovery_task_lock);
2831         wake_up(&obd->obd_next_transno_waitq);
2832         RETURN(0);
2833 }
2834
2835 int target_handle_ping(struct ptlrpc_request *req)
2836 {
2837         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2838         return req_capsule_server_pack(&req->rq_pill);
2839 }
2840
2841 void target_committed_to_req(struct ptlrpc_request *req)
2842 {
2843         struct obd_export *exp = req->rq_export;
2844
2845         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2846                 lustre_msg_set_last_committed(req->rq_repmsg,
2847                                               exp->exp_last_committed);
2848         else
2849                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2850                           "%d)", exp->exp_obd->obd_no_transno,
2851                           req->rq_repmsg == NULL);
2852
2853         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2854                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2855 }
2856
2857 #endif /* HAVE_SERVER_SUPPORT */
2858
2859 /**
2860  * Packs current SLV and Limit into \a req.
2861  */
2862 int target_pack_pool_reply(struct ptlrpc_request *req)
2863 {
2864         struct obd_device *obd;
2865         ENTRY;
2866
2867         /* Check that we still have all structures alive as this may
2868          * be some late RPC at shutdown time. */
2869         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2870                      !exp_connect_lru_resize(req->rq_export))) {
2871                 lustre_msg_set_slv(req->rq_repmsg, 0);
2872                 lustre_msg_set_limit(req->rq_repmsg, 0);
2873                 RETURN(0);
2874         }
2875
2876         /* OBD is alive here as export is alive, which we checked above. */
2877         obd = req->rq_export->exp_obd;
2878
2879         read_lock(&obd->obd_pool_lock);
2880         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2881         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2882         read_unlock(&obd->obd_pool_lock);
2883
2884         RETURN(0);
2885 }
2886
2887 static int target_send_reply_msg(struct ptlrpc_request *req,
2888                                  int rc, int fail_id)
2889 {
2890         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2891                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2892                 return -ECOMM;
2893         }
2894         /* We can have a null rq_reqmsg in the event of bad signature or
2895          * no context when unwrapping */
2896         if (req->rq_reqmsg &&
2897             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2898             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2899                 return -ECOMM;
2900
2901         if (unlikely(rc)) {
2902                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2903                 req->rq_status = rc;
2904                 return ptlrpc_send_error(req, 1);
2905         } else {
2906                 DEBUG_REQ(D_NET, req, "sending reply");
2907         }
2908
2909         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2910 }
2911
2912 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2913 {
2914         struct ptlrpc_service_part *svcpt;
2915         int                        netrc;
2916         struct ptlrpc_reply_state *rs;
2917         struct obd_export         *exp;
2918         ENTRY;
2919
2920         if (req->rq_no_reply) {
2921                 EXIT;
2922                 return;
2923         }
2924
2925         svcpt = req->rq_rqbd->rqbd_svcpt;
2926         rs = req->rq_reply_state;
2927         if (rs == NULL || !rs->rs_difficult) {
2928                 /* no notifiers */
2929                 target_send_reply_msg (req, rc, fail_id);
2930                 EXIT;
2931                 return;
2932         }
2933
2934         /* must be an export if locks saved */
2935         LASSERT(req->rq_export != NULL);
2936         /* req/reply consistent */
2937         LASSERT(rs->rs_svcpt == svcpt);
2938
2939         /* "fresh" reply */
2940         LASSERT(!rs->rs_scheduled);
2941         LASSERT(!rs->rs_scheduled_ever);
2942         LASSERT(!rs->rs_handled);
2943         LASSERT(!rs->rs_on_net);
2944         LASSERT(rs->rs_export == NULL);
2945         LASSERT(list_empty(&rs->rs_obd_list));
2946         LASSERT(list_empty(&rs->rs_exp_list));
2947
2948         exp = class_export_get(req->rq_export);
2949
2950         /* disable reply scheduling while I'm setting up */
2951         rs->rs_scheduled = 1;
2952         rs->rs_on_net    = 1;
2953         rs->rs_xid       = req->rq_xid;
2954         rs->rs_transno   = req->rq_transno;
2955         rs->rs_export    = exp;
2956         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2957
2958         spin_lock(&exp->exp_uncommitted_replies_lock);
2959         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2960                rs->rs_transno, exp->exp_last_committed);
2961         if (rs->rs_transno > exp->exp_last_committed) {
2962                 /* not committed already */
2963                 list_add_tail(&rs->rs_obd_list,
2964                                   &exp->exp_uncommitted_replies);
2965         }
2966         spin_unlock(&exp->exp_uncommitted_replies_lock);
2967
2968         spin_lock(&exp->exp_lock);
2969         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2970         spin_unlock(&exp->exp_lock);
2971
2972         netrc = target_send_reply_msg(req, rc, fail_id);
2973
2974         spin_lock(&svcpt->scp_rep_lock);
2975
2976         atomic_inc(&svcpt->scp_nreps_difficult);
2977
2978         if (netrc != 0) {
2979                 /* error sending: reply is off the net.  Also we need +1
2980                  * reply ref until ptlrpc_handle_rs() is done
2981                  * with the reply state (if the send was successful, there
2982                  * would have been +1 ref for the net, which
2983                  * reply_out_callback leaves alone) */
2984                 rs->rs_on_net = 0;
2985                 ptlrpc_rs_addref(rs);
2986         }
2987
2988         spin_lock(&rs->rs_lock);
2989         if (rs->rs_transno <= exp->exp_last_committed ||
2990             (!rs->rs_on_net && !rs->rs_no_ack) ||
2991             list_empty(&rs->rs_exp_list) ||     /* completed already */
2992             list_empty(&rs->rs_obd_list)) {
2993                 CDEBUG(D_HA, "Schedule reply immediately\n");
2994                 ptlrpc_dispatch_difficult_reply(rs);
2995         } else {
2996                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2997                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2998         }
2999         spin_unlock(&rs->rs_lock);
3000         spin_unlock(&svcpt->scp_rep_lock);
3001         EXIT;
3002 }
3003
3004 enum ldlm_mode lck_compat_array[] = {
3005         [LCK_EX]    = LCK_COMPAT_EX,
3006         [LCK_PW]    = LCK_COMPAT_PW,
3007         [LCK_PR]    = LCK_COMPAT_PR,
3008         [LCK_CW]    = LCK_COMPAT_CW,
3009         [LCK_CR]    = LCK_COMPAT_CR,
3010         [LCK_NL]    = LCK_COMPAT_NL,
3011         [LCK_GROUP] = LCK_COMPAT_GROUP,
3012         [LCK_COS]   = LCK_COMPAT_COS,
3013 };
3014
3015 /**
3016  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3017  * not escape to the user level.
3018  */
3019 int ldlm_error2errno(enum ldlm_error error)
3020 {
3021         int result;
3022
3023         switch (error) {
3024         case ELDLM_OK:
3025         case ELDLM_LOCK_MATCHED:
3026                 result = 0;
3027                 break;
3028         case ELDLM_LOCK_CHANGED:
3029                 result = -ESTALE;
3030                 break;
3031         case ELDLM_LOCK_ABORTED:
3032                 result = -ENAVAIL;
3033                 break;
3034         case ELDLM_LOCK_REPLACED:
3035                 result = -ESRCH;
3036                 break;
3037         case ELDLM_NO_LOCK_DATA:
3038                 result = -ENOENT;
3039                 break;
3040         case ELDLM_NAMESPACE_EXISTS:
3041                 result = -EEXIST;
3042                 break;
3043         case ELDLM_BAD_NAMESPACE:
3044                 result = -EBADF;
3045                 break;
3046         default:
3047                 if (((int)error) < 0) { /* cast to signed type */
3048                         result = error; /* as ldlm_error can be unsigned */
3049                 } else {
3050                         CERROR("Invalid DLM result code: %d\n", error);
3051                         result = -EPROTO;
3052                 }
3053         }
3054         return result;
3055 }
3056 EXPORT_SYMBOL(ldlm_error2errno);
3057
3058 /**
3059  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3060  */
3061 enum ldlm_error ldlm_errno2error(int err_no)
3062 {
3063         int error;
3064
3065         switch (err_no) {
3066         case 0:
3067                 error = ELDLM_OK;
3068                 break;
3069         case -ESTALE:
3070                 error = ELDLM_LOCK_CHANGED;
3071                 break;
3072         case -ENAVAIL:
3073                 error = ELDLM_LOCK_ABORTED;
3074                 break;
3075         case -ESRCH:
3076                 error = ELDLM_LOCK_REPLACED;
3077                 break;
3078         case -ENOENT:
3079                 error = ELDLM_NO_LOCK_DATA;
3080                 break;
3081         case -EEXIST:
3082                 error = ELDLM_NAMESPACE_EXISTS;
3083                 break;
3084         case -EBADF:
3085                 error = ELDLM_BAD_NAMESPACE;
3086                 break;
3087         default:
3088                 error = err_no;
3089         }
3090         return error;
3091 }
3092
3093 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3094 void ldlm_dump_export_locks(struct obd_export *exp)
3095 {
3096         spin_lock(&exp->exp_locks_list_guard);
3097         if (!list_empty(&exp->exp_locks_list)) {
3098                 struct ldlm_lock *lock;
3099
3100                 CERROR("dumping locks for export %p,"
3101                        "ignore if the unmount doesn't hang\n", exp);
3102                 list_for_each_entry(lock, &exp->exp_locks_list,
3103                                         l_exp_refs_link)
3104                         LDLM_ERROR(lock, "lock:");
3105         }
3106         spin_unlock(&exp->exp_locks_list_guard);
3107 }
3108 #endif
3109
3110 #ifdef HAVE_SERVER_SUPPORT
3111 static int target_bulk_timeout(void *data)
3112 {
3113         ENTRY;
3114         /* We don't fail the connection here, because having the export
3115          * killed makes the (vital) call to commitrw very sad.
3116          */
3117         RETURN(1);
3118 }
3119
3120 static inline const char *bulk2type(struct ptlrpc_request *req)
3121 {
3122         if (req->rq_bulk_read)
3123                 return "READ";
3124         if (req->rq_bulk_write)
3125                 return "WRITE";
3126         return "UNKNOWN";
3127 }
3128
3129 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3130                    struct l_wait_info *lwi)
3131 {
3132         struct ptlrpc_request *req = desc->bd_req;
3133         time64_t start = ktime_get_real_seconds();
3134         time64_t deadline;
3135         int rc = 0;
3136
3137         ENTRY;
3138
3139         /* If there is eviction in progress, wait for it to finish. */
3140         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3141                 *lwi = LWI_INTR(NULL, NULL);
3142                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3143                                   !atomic_read(&exp->exp_obd->
3144                                                    obd_evict_inprogress),
3145                                   lwi);
3146         }
3147
3148         /* Check if client was evicted or reconnected already. */
3149         if (exp->exp_failed ||
3150             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3151                 rc = -ENOTCONN;
3152         } else {
3153                 if (req->rq_bulk_read)
3154                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3155
3156                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3157                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3158                 else /* old version, bulk matchbits is rq_xid */
3159                         req->rq_mbits = req->rq_xid;
3160
3161                 if (rc == 0)
3162                         rc = ptlrpc_start_bulk_transfer(desc);
3163         }
3164
3165         if (rc < 0) {
3166                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3167                           bulk2type(req), rc);
3168                 RETURN(rc);
3169         }
3170
3171         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3172                 ptlrpc_abort_bulk(desc);
3173                 RETURN(0);
3174         }
3175
3176         /* limit actual bulk transfer to bulk_timeout seconds */
3177         deadline = start + bulk_timeout;
3178         if (deadline > req->rq_deadline)
3179                 deadline = req->rq_deadline;
3180
3181         do {
3182                 time64_t timeoutl = deadline - ktime_get_real_seconds();
3183                 long timeout_jiffies = timeoutl <= 0 ?
3184                                        1 : cfs_time_seconds(timeoutl);
3185                 time64_t rq_deadline;
3186
3187                 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3188                                             cfs_time_seconds(1),
3189                                             target_bulk_timeout, desc);
3190                 rc = l_wait_event(desc->bd_waitq,
3191                                   !ptlrpc_server_bulk_active(desc) ||
3192                                   exp->exp_failed ||
3193                                   exp->exp_conn_cnt >
3194                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3195                                   lwi);
3196                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3197                 /* Wait again if we changed rq_deadline. */
3198                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3199                 deadline = start + bulk_timeout;
3200                 if (deadline > rq_deadline)
3201                         deadline = rq_deadline;
3202         } while (rc == -ETIMEDOUT &&
3203                  deadline > ktime_get_real_seconds());
3204
3205         if (rc == -ETIMEDOUT) {
3206                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3207                           bulk2type(req), deadline - start,
3208                           ktime_get_real_seconds() - deadline);
3209                 ptlrpc_abort_bulk(desc);
3210         } else if (exp->exp_failed) {
3211                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3212                           bulk2type(req));
3213                 rc = -ENOTCONN;
3214                 ptlrpc_abort_bulk(desc);
3215         } else if (exp->exp_conn_cnt >
3216                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3217                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3218                           bulk2type(req));
3219                 /* We don't reply anyway. */
3220                 rc = -ETIMEDOUT;
3221                 ptlrpc_abort_bulk(desc);
3222         } else if (desc->bd_failure) {
3223                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3224                           bulk2type(req));
3225                 /* XXX should this be a different errno? */
3226                 rc = -ETIMEDOUT;
3227         } else {
3228                 if (req->rq_bulk_write)
3229                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3230                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3231                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3232                                   bulk2type(req), desc->bd_nob_transferred,
3233                                   desc->bd_nob);
3234                         /* XXX should this be a different errno? */
3235                         rc = -ETIMEDOUT;
3236                 }
3237         }
3238
3239         RETURN(rc);
3240 }
3241 EXPORT_SYMBOL(target_bulk_io);
3242
3243 #endif /* HAVE_SERVER_SUPPORT */