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, 2017, 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 <cl_object.h>
  43 #include <linux/jiffies.h>
  44 #include <linux/kernel.h>
  45 #include <linux/kthread.h>
  46 #include <libcfs/libcfs.h>
  47 #include <obd.h>
  48 #include <obd_class.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /*
  55  * @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         lnet_nid_t nid4refnet = LNET_NID_ANY;
  64         int rc = 0;
  65
  66         ENTRY;
  67
  68         if (!create && !priority) {
  69                 CDEBUG(D_HA, "Nothing to do\n");
  70                 RETURN(-EINVAL);
  71         }
  72
  73         if (imp->imp_connection &&
  74             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  75                 /* nid4refnet is used to restrict network connections */
  76                 nid4refnet = imp->imp_connection->c_self;
  77         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
  78         if (!ptlrpc_conn) {
  79                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  80                 RETURN(-ENOENT);
  81         }
  82
  83         if (create) {
  84                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  85                 if (!imp_conn)
  86                         GOTO(out_put, rc = -ENOMEM);
  87         }
  88
  89         spin_lock(&imp->imp_lock);
  90         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  91                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  92                         if (priority) {
  93                                 list_move(&item->oic_item,
  94                                           &imp->imp_conn_list);
  95                                 item->oic_last_attempt = 0;
  96                         }
  97                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  98                                imp, imp->imp_obd->obd_name, uuid->uuid,
  99                                (priority ? ", moved to head" : ""));
 100                         spin_unlock(&imp->imp_lock);
 101                         GOTO(out_free, rc = 0);
 102                 }
 103         }
 104         /* No existing import connection found for \a uuid. */
 105         if (create) {
 106                 imp_conn->oic_conn = ptlrpc_conn;
 107                 imp_conn->oic_uuid = *uuid;
 108                 imp_conn->oic_last_attempt = 0;
 109                 if (priority)
 110                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 111                 else
 112                         list_add_tail(&imp_conn->oic_item,
 113                                       &imp->imp_conn_list);
 114                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 115                        imp, imp->imp_obd->obd_name, uuid->uuid,
 116                        (priority ? "head" : "tail"));
 117         } else {
 118                 spin_unlock(&imp->imp_lock);
 119                 GOTO(out_free, rc = -ENOENT);
 120         }
 121
 122         spin_unlock(&imp->imp_lock);
 123         RETURN(0);
 124 out_free:
 125         if (imp_conn)
 126                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 127 out_put:
 128         ptlrpc_connection_put(ptlrpc_conn);
 129         RETURN(rc);
 130 }
 131
 132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 133 {
 134         return import_set_conn(imp, uuid, 1, 0);
 135 }
 136
 137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 138                            int priority)
 139 {
 140         return import_set_conn(imp, uuid, priority, 1);
 141 }
 142 EXPORT_SYMBOL(client_import_add_conn);
 143
 144 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 145 {
 146         struct obd_import_conn *imp_conn;
 147         struct obd_export *dlmexp;
 148         int rc = -ENOENT;
 149
 150         ENTRY;
 151
 152         spin_lock(&imp->imp_lock);
 153         if (list_empty(&imp->imp_conn_list)) {
 154                 LASSERT(!imp->imp_connection);
 155                 GOTO(out, rc);
 156         }
 157
 158         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 159                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 160                         continue;
 161                 LASSERT(imp_conn->oic_conn);
 162
 163                 if (imp_conn == imp->imp_conn_current) {
 164                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 165
 166                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 167                             imp->imp_state != LUSTRE_IMP_DISCON) {
 168                                 CERROR("can't remove current connection\n");
 169                                 GOTO(out, rc = -EBUSY);
 170                         }
 171
 172                         ptlrpc_connection_put(imp->imp_connection);
 173                         imp->imp_connection = NULL;
 174
 175                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 176                         if (dlmexp && dlmexp->exp_connection) {
 177                                 LASSERT(dlmexp->exp_connection ==
 178                                         imp_conn->oic_conn);
 179                                 ptlrpc_connection_put(dlmexp->exp_connection);
 180                                 dlmexp->exp_connection = NULL;
 181                         }
 182
 183                         if (dlmexp != NULL)
 184                                 class_export_put(dlmexp);
 185                 }
 186
 187                 list_del(&imp_conn->oic_item);
 188                 ptlrpc_connection_put(imp_conn->oic_conn);
 189                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 190                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 191                        imp, imp->imp_obd->obd_name, uuid->uuid);
 192                 rc = 0;
 193                 break;
 194         }
 195 out:
 196         spin_unlock(&imp->imp_lock);
 197         if (rc == -ENOENT)
 198                 CERROR("connection %s not found\n", uuid->uuid);
 199         RETURN(rc);
 200 }
 201 EXPORT_SYMBOL(client_import_del_conn);
 202
 203 /**
 204  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 205  * to find a conn uuid of \a imp which can reach \a peer.
 206  */
 207 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 208                             struct obd_uuid *uuid)
 209 {
 210         struct obd_import_conn *conn;
 211         int rc = -ENOENT;
 212
 213         ENTRY;
 214
 215         spin_lock(&imp->imp_lock);
 216         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 217                 /* Check if conn UUID does have this peer NID. */
 218                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 219                         *uuid = conn->oic_uuid;
 220                         rc = 0;
 221                         break;
 222                 }
 223         }
 224         spin_unlock(&imp->imp_lock);
 225         RETURN(rc);
 226 }
 227 EXPORT_SYMBOL(client_import_find_conn);
 228
 229 void client_destroy_import(struct obd_import *imp)
 230 {
 231         /*
 232          * Drop security policy instance after all RPCs have finished/aborted
 233          * to let all busy contexts be released.
 234          */
 235         class_import_get(imp);
 236         class_destroy_import(imp);
 237         sptlrpc_import_sec_put(imp);
 238         class_import_put(imp);
 239 }
 240 EXPORT_SYMBOL(client_destroy_import);
 241
 242 /**
 243  * Check whether or not the OSC is on MDT.
 244  * In the config log,
 245  * osc on MDT
 246  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 247  * osc on client
 248  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 249  *
 250  **/
 251 static int osc_on_mdt(char *obdname)
 252 {
 253         char *ptr;
 254
 255         ptr = strrchr(obdname, '-');
 256         if (ptr == NULL)
 257                 return 0;
 258
 259         if (strncmp(ptr + 1, "MDT", 3) == 0)
 260                 return 1;
 261
 262         return 0;
 263 }
 264
 265 /*
 266  * Configure an RPC client OBD device.
 267  *
 268  * lcfg parameters:
 269  * 1 - client UUID
 270  * 2 - server UUID
 271  * 3 - inactive-on-startup
 272  * 4 - restrictive net
 273  */
 274 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 275 {
 276         struct client_obd *cli = &obd->u.cli;
 277         struct obd_import *imp;
 278         struct obd_uuid server_uuid;
 279         int rq_portal, rp_portal, connect_op;
 280         const char *name = obd->obd_type->typ_name;
 281         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 282         char *cli_name = lustre_cfg_buf(lcfg, 0);
 283         struct ptlrpc_connection fake_conn = { .c_self = 0,
 284                                                .c_remote_uuid.uuid[0] = 0 };
 285         int rc;
 286
 287         ENTRY;
 288
 289         /*
 290          * In a more perfect world, we would hang a ptlrpc_client off of
 291          * obd_type and just use the values from there.
 292          */
 293         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 294                 rq_portal = OST_REQUEST_PORTAL;
 295                 rp_portal = OSC_REPLY_PORTAL;
 296                 connect_op = OST_CONNECT;
 297                 cli->cl_sp_me = LUSTRE_SP_CLI;
 298                 cli->cl_sp_to = LUSTRE_SP_OST;
 299                 ns_type = LDLM_NS_TYPE_OSC;
 300         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 301                    !strcmp(name, LUSTRE_LWP_NAME)) {
 302                 rq_portal = MDS_REQUEST_PORTAL;
 303                 rp_portal = MDC_REPLY_PORTAL;
 304                 connect_op = MDS_CONNECT;
 305                 if (is_lwp_on_ost(cli_name))
 306                         cli->cl_sp_me = LUSTRE_SP_OST;
 307                 else if (is_lwp_on_mdt(cli_name))
 308                         cli->cl_sp_me = LUSTRE_SP_MDT;
 309                 else
 310                         cli->cl_sp_me = LUSTRE_SP_CLI;
 311                 cli->cl_sp_to = LUSTRE_SP_MDT;
 312                 ns_type = LDLM_NS_TYPE_MDC;
 313         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 314                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 315                         /* OSP_on_MDT for other MDTs */
 316                         connect_op = MDS_CONNECT;
 317                         cli->cl_sp_to = LUSTRE_SP_MDT;
 318                         ns_type = LDLM_NS_TYPE_MDC;
 319                         rq_portal = OUT_PORTAL;
 320                 } else {
 321                         /* OSP on MDT for OST */
 322                         connect_op = OST_CONNECT;
 323                         cli->cl_sp_to = LUSTRE_SP_OST;
 324                         ns_type = LDLM_NS_TYPE_OSC;
 325                         rq_portal = OST_REQUEST_PORTAL;
 326                 }
 327                 rp_portal = OSC_REPLY_PORTAL;
 328                 cli->cl_sp_me = LUSTRE_SP_MDT;
 329         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 330                 rq_portal = MGS_REQUEST_PORTAL;
 331                 rp_portal = MGC_REPLY_PORTAL;
 332                 connect_op = MGS_CONNECT;
 333                 cli->cl_sp_me = LUSTRE_SP_MGC;
 334                 cli->cl_sp_to = LUSTRE_SP_MGS;
 335                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 336                 ns_type = LDLM_NS_TYPE_MGC;
 337         } else {
 338                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 339                        name);
 340                 RETURN(-EINVAL);
 341         }
 342
 343         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 344                 CERROR("requires a TARGET UUID\n");
 345                 RETURN(-EINVAL);
 346         }
 347
 348         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 349                 CERROR("client UUID must be less than 38 characters\n");
 350                 RETURN(-EINVAL);
 351         }
 352
 353         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 354                 CERROR("setup requires a SERVER UUID\n");
 355                 RETURN(-EINVAL);
 356         }
 357
 358         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 359                 CERROR("target UUID must be less than 38 characters\n");
 360                 RETURN(-EINVAL);
 361         }
 362
 363         init_rwsem(&cli->cl_sem);
 364         mutex_init(&cli->cl_mgc_mutex);
 365         cli->cl_seq = NULL;
 366         init_rwsem(&cli->cl_seq_rwsem);
 367         cli->cl_conn_count = 0;
 368         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 369                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 370                      sizeof(server_uuid)));
 371
 372         cli->cl_dirty_pages = 0;
 373         cli->cl_dirty_max_pages = 0;
 374         cli->cl_avail_grant = 0;
 375         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 376         /*
 377          * cl_dirty_max_pages may be changed at connect time in
 378          * ptlrpc_connect_interpret().
 379          */
 380         client_adjust_max_dirty(cli);
 381         init_waitqueue_head(&cli->cl_cache_waiters);
 382         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 383         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 384         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 385         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 386         spin_lock_init(&cli->cl_loi_list_lock);
 387         atomic_set(&cli->cl_pending_w_pages, 0);
 388         atomic_set(&cli->cl_pending_r_pages, 0);
 389         cli->cl_r_in_flight = 0;
 390         cli->cl_w_in_flight = 0;
 391
 392         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 393         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 394         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 395         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 396         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 397         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 398
 399         /* lru for osc. */
 400         INIT_LIST_HEAD(&cli->cl_lru_osc);
 401         atomic_set(&cli->cl_lru_shrinkers, 0);
 402         atomic_long_set(&cli->cl_lru_busy, 0);
 403         atomic_long_set(&cli->cl_lru_in_list, 0);
 404         INIT_LIST_HEAD(&cli->cl_lru_list);
 405         spin_lock_init(&cli->cl_lru_list_lock);
 406         atomic_long_set(&cli->cl_unstable_count, 0);
 407         INIT_LIST_HEAD(&cli->cl_shrink_list);
 408         INIT_LIST_HEAD(&cli->cl_grant_chain);
 409
 410         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 411         cli->cl_rpcs_in_flight = 0;
 412
 413         init_waitqueue_head(&cli->cl_destroy_waitq);
 414         atomic_set(&cli->cl_destroy_in_flight, 0);
 415
 416
 417         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 418         cli->cl_preferred_cksum_type = 0;
 419 #ifdef ENABLE_CHECKSUM
 420         /* Turn on checksumming by default. */
 421         cli->cl_checksum = 1;
 422         /*
 423          * The supported checksum types will be worked out at connect time
 424          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 425          * through procfs.
 426          */
 427         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 428 #endif
 429         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 430
 431         /*
 432          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 433          * from OFD after connecting.
 434          */
 435         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 436
 437         cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
 438
 439         /*
 440          * set cl_chunkbits default value to PAGE_SHIFT,
 441          * it will be updated at OSC connection time.
 442          */
 443         cli->cl_chunkbits = PAGE_SHIFT;
 444
 445         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 446                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 447         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 448                 cli->cl_max_rpcs_in_flight = 2;
 449         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 450                 cli->cl_max_rpcs_in_flight = 3;
 451         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 452                 cli->cl_max_rpcs_in_flight = 4;
 453         } else {
 454                 if (osc_on_mdt(obd->obd_name))
 455                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 456                 else
 457                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 458         }
 459
 460         spin_lock_init(&cli->cl_mod_rpcs_lock);
 461         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 462         cli->cl_max_mod_rpcs_in_flight = 0;
 463         cli->cl_mod_rpcs_in_flight = 0;
 464         cli->cl_close_rpcs_in_flight = 0;
 465         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 466         cli->cl_mod_tag_bitmap = NULL;
 467
 468         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 469
 470         if (connect_op == MDS_CONNECT) {
 471                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 472                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 473                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 474                 if (cli->cl_mod_tag_bitmap == NULL)
 475                         GOTO(err, rc = -ENOMEM);
 476         }
 477
 478         rc = ldlm_get_ref();
 479         if (rc) {
 480                 CERROR("ldlm_get_ref failed: %d\n", rc);
 481                 GOTO(err, rc);
 482         }
 483
 484         ptlrpc_init_client(rq_portal, rp_portal, name,
 485                            &obd->obd_ldlm_client);
 486
 487         imp = class_new_import(obd);
 488         if (imp == NULL)
 489                 GOTO(err_ldlm, rc = -ENOENT);
 490         imp->imp_client = &obd->obd_ldlm_client;
 491         imp->imp_connect_op = connect_op;
 492         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 493                LUSTRE_CFG_BUFLEN(lcfg, 1));
 494         class_import_put(imp);
 495
 496         if (lustre_cfg_buf(lcfg, 4)) {
 497                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 498
 499                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 500                         rc = -EINVAL;
 501                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 502                                obd->obd_name, lustre_cfg_string(lcfg, 4),
 503                                rc);
 504                         GOTO(err_import, rc);
 505                 }
 506                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 507                 imp->imp_connection = &fake_conn;
 508         }
 509
 510         rc = client_import_add_conn(imp, &server_uuid, 1);
 511         if (rc) {
 512                 CERROR("can't add initial connection\n");
 513                 GOTO(err_import, rc);
 514         }
 515         imp->imp_connection = NULL;
 516
 517         cli->cl_import = imp;
 518         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 519         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 520
 521         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 522                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 523                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 524                                name, obd->obd_name,
 525                                cli->cl_target_uuid.uuid);
 526                         spin_lock(&imp->imp_lock);
 527                         imp->imp_deactive = 1;
 528                         spin_unlock(&imp->imp_lock);
 529                 }
 530         }
 531
 532         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 533                                                 LDLM_NAMESPACE_CLIENT,
 534                                                 LDLM_NAMESPACE_GREEDY,
 535                                                 ns_type);
 536         if (obd->obd_namespace == NULL) {
 537                 CERROR("Unable to create client namespace - %s\n",
 538                        obd->obd_name);
 539                 GOTO(err_import, rc = -ENOMEM);
 540         }
 541
 542         RETURN(rc);
 543
 544 err_import:
 545         class_destroy_import(imp);
 546 err_ldlm:
 547         ldlm_put_ref();
 548 err:
 549         if (cli->cl_mod_tag_bitmap != NULL)
 550                 OBD_FREE(cli->cl_mod_tag_bitmap,
 551                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 552         cli->cl_mod_tag_bitmap = NULL;
 553         RETURN(rc);
 554
 555 }
 556 EXPORT_SYMBOL(client_obd_setup);
 557
 558 int client_obd_cleanup(struct obd_device *obd)
 559 {
 560         struct client_obd *cli = &obd->u.cli;
 561
 562         ENTRY;
 563
 564         ldlm_namespace_free_post(obd->obd_namespace);
 565         obd->obd_namespace = NULL;
 566
 567         obd_cleanup_client_import(obd);
 568         LASSERT(obd->u.cli.cl_import == NULL);
 569
 570         ldlm_put_ref();
 571
 572         if (cli->cl_mod_tag_bitmap != NULL)
 573                 OBD_FREE(cli->cl_mod_tag_bitmap,
 574                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 575         cli->cl_mod_tag_bitmap = NULL;
 576
 577         RETURN(0);
 578 }
 579 EXPORT_SYMBOL(client_obd_cleanup);
 580
 581 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 582 int client_connect_import(const struct lu_env *env,
 583                           struct obd_export **exp,
 584                           struct obd_device *obd, struct obd_uuid *cluuid,
 585                           struct obd_connect_data *data, void *localdata)
 586 {
 587         struct client_obd *cli = &obd->u.cli;
 588         struct obd_import *imp = cli->cl_import;
 589         struct obd_connect_data *ocd;
 590         struct lustre_handle conn = { 0 };
 591         int rc;
 592
 593         ENTRY;
 594
 595         *exp = NULL;
 596         down_write(&cli->cl_sem);
 597         if (cli->cl_conn_count > 0)
 598                 GOTO(out_sem, rc = -EALREADY);
 599
 600         rc = class_connect(&conn, obd, cluuid);
 601         if (rc)
 602                 GOTO(out_sem, rc);
 603
 604         cli->cl_conn_count++;
 605         *exp = class_conn2export(&conn);
 606
 607         LASSERT(obd->obd_namespace);
 608
 609         imp->imp_dlm_handle = conn;
 610         rc = ptlrpc_init_import(imp);
 611         if (rc != 0)
 612                 GOTO(out_ldlm, rc);
 613
 614         ocd = &imp->imp_connect_data;
 615         if (data) {
 616                 *ocd = *data;
 617                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 618                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 619         }
 620
 621         rc = ptlrpc_connect_import(imp);
 622         if (rc != 0) {
 623                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 624                 GOTO(out_ldlm, rc);
 625         }
 626         LASSERT(*exp != NULL && (*exp)->exp_connection);
 627
 628         if (data) {
 629                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 630                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 631                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 632                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 633                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 634         }
 635
 636         ptlrpc_pinger_add_import(imp);
 637
 638         EXIT;
 639
 640         if (rc) {
 641 out_ldlm:
 642                 cli->cl_conn_count--;
 643                 class_disconnect(*exp);
 644                 *exp = NULL;
 645         }
 646 out_sem:
 647         up_write(&cli->cl_sem);
 648
 649         if (!rc && localdata) {
 650                 LASSERT(cli->cl_cache == NULL); /* only once */
 651                 cli->cl_cache = (struct cl_client_cache *)localdata;
 652                 cl_cache_incref(cli->cl_cache);
 653                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 654
 655                 /* add this osc into entity list */
 656                 LASSERT(list_empty(&cli->cl_lru_osc));
 657                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 658                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 659                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 660         }
 661
 662         return rc;
 663 }
 664 EXPORT_SYMBOL(client_connect_import);
 665
 666 int client_disconnect_export(struct obd_export *exp)
 667 {
 668         struct obd_device *obd = class_exp2obd(exp);
 669         struct client_obd *cli;
 670         struct obd_import *imp;
 671         int rc = 0, err;
 672
 673         ENTRY;
 674
 675         if (!obd) {
 676                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 677                        exp, exp ? exp->exp_handle.h_cookie : -1);
 678                 RETURN(-EINVAL);
 679         }
 680
 681         cli = &obd->u.cli;
 682         imp = cli->cl_import;
 683
 684         down_write(&cli->cl_sem);
 685         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 686                 cli->cl_conn_count);
 687
 688         if (cli->cl_conn_count == 0) {
 689                 CERROR("disconnecting disconnected device (%s)\n",
 690                        obd->obd_name);
 691                 GOTO(out_disconnect, rc = -EINVAL);
 692         }
 693
 694         cli->cl_conn_count--;
 695         if (cli->cl_conn_count != 0)
 696                 GOTO(out_disconnect, rc = 0);
 697
 698         /*
 699          * Mark import deactivated now, so we don't try to reconnect if any
 700          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 701          * fully deactivate the import, or that would drop all requests.
 702          */
 703         spin_lock(&imp->imp_lock);
 704         imp->imp_deactive = 1;
 705         spin_unlock(&imp->imp_lock);
 706
 707         /*
 708          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 709          * delete it regardless.  (It's safe to delete an import that was
 710          * never added.)
 711          */
 712         (void)ptlrpc_pinger_del_import(imp);
 713
 714         if (obd->obd_namespace != NULL) {
 715                 /* obd_force == local only */
 716                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 717                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 718                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 719                                           obd->obd_force);
 720         }
 721
 722         /*
 723          * There's no need to hold sem while disconnecting an import,
 724          * and it may actually cause deadlock in GSS.
 725          */
 726         up_write(&cli->cl_sem);
 727         rc = ptlrpc_disconnect_import(imp, 0);
 728         down_write(&cli->cl_sem);
 729
 730         ptlrpc_invalidate_import(imp);
 731
 732         EXIT;
 733
 734 out_disconnect:
 735         /*
 736          * Use server style - class_disconnect should be always called for
 737          * o_disconnect.
 738          */
 739         err = class_disconnect(exp);
 740         if (!rc && err)
 741                 rc = err;
 742
 743         up_write(&cli->cl_sem);
 744
 745         RETURN(rc);
 746 }
 747 EXPORT_SYMBOL(client_disconnect_export);
 748
 749 #ifdef HAVE_SERVER_SUPPORT
 750 int server_disconnect_export(struct obd_export *exp)
 751 {
 752         int rc;
 753
 754         ENTRY;
 755
 756         /* Disconnect early so that clients can't keep using export. */
 757         rc = class_disconnect(exp);
 758         /* Close import to avoid sending any requests. */
 759         if (exp->exp_imp_reverse)
 760                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 761
 762         ldlm_bl_thread_wakeup();
 763
 764         /* complete all outstanding replies */
 765         spin_lock(&exp->exp_lock);
 766         while (!list_empty(&exp->exp_outstanding_replies)) {
 767                 struct ptlrpc_reply_state *rs =
 768                         list_entry(exp->exp_outstanding_replies.next,
 769                                        struct ptlrpc_reply_state, rs_exp_list);
 770                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 771
 772                 spin_lock(&svcpt->scp_rep_lock);
 773
 774                 list_del_init(&rs->rs_exp_list);
 775
 776                 spin_lock(&rs->rs_lock);
 777                 /* clear rs_convert_lock to make sure rs is handled and put */
 778                 rs->rs_convert_lock = 0;
 779                 ptlrpc_schedule_difficult_reply(rs);
 780                 spin_unlock(&rs->rs_lock);
 781
 782                 spin_unlock(&svcpt->scp_rep_lock);
 783         }
 784         spin_unlock(&exp->exp_lock);
 785
 786         RETURN(rc);
 787 }
 788 EXPORT_SYMBOL(server_disconnect_export);
 789
 790 static inline int target_check_recovery_timer(struct obd_device *target)
 791 {
 792         ktime_t remaining;
 793         s64 timeout;
 794
 795         if (!target->obd_recovering || target->obd_recovery_start == 0)
 796                 return 0;
 797
 798         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 799         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 800         if (timeout > -30)
 801                 return 0;
 802
 803         /* the recovery timer should expire, but it isn't triggered,
 804          * it's better to abort the recovery of this target to speed up
 805          * the recovery of the whole cluster. */
 806         spin_lock(&target->obd_dev_lock);
 807         if (target->obd_recovering) {
 808                 CERROR("%s: Aborting recovery\n", target->obd_name);
 809                 target->obd_abort_recovery = 1;
 810                 wake_up(&target->obd_next_transno_waitq);
 811         }
 812         spin_unlock(&target->obd_dev_lock);
 813         return 0;
 814 }
 815
 816 /*
 817  * --------------------------------------------------------------------------
 818  * from old lib/target.c
 819  * --------------------------------------------------------------------------
 820  */
 821
 822 static int target_handle_reconnect(struct lustre_handle *conn,
 823                                    struct obd_export *exp,
 824                                    struct obd_uuid *cluuid)
 825 {
 826         struct obd_device *target;
 827         struct lustre_handle *hdl;
 828         ktime_t remaining;
 829         s64 timeout;
 830         int rc = 0;
 831
 832         ENTRY;
 833         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 834         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 835                 conn->cookie = exp->exp_handle.h_cookie;
 836                 CDEBUG(D_HA,
 837                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 838                        cluuid->uuid, exp, conn->cookie);
 839                 RETURN(0);
 840         }
 841
 842         target = exp->exp_obd;
 843
 844         /* Might be a re-connect after a partition. */
 845         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 846                 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
 847                               target->obd_name,
 848                               obd_uuid2str(&exp->exp_client_uuid),
 849                               obd_export_nid2str(exp),
 850                               hdl->cookie, conn->cookie);
 851                 memset(conn, 0, sizeof(*conn));
 852                 /*
 853                  * target_handle_connect() treats EALREADY and
 854                  * -EALREADY differently.  -EALREADY is an error
 855                  * (same UUID, different handle).
 856                  */
 857                 RETURN(-EALREADY);
 858         }
 859
 860         if (!target->obd_recovering) {
 861                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 862                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 863                         obd_export_nid2str(exp));
 864                 GOTO(out_already, rc);
 865         }
 866
 867         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 868         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 869         if (timeout > 0) {
 870                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 871                               target->obd_name,
 872                               obd_uuid2str(&exp->exp_client_uuid),
 873                               obd_export_nid2str(exp),
 874                               atomic_read(&target->obd_max_recoverable_clients),
 875                               timeout / 60, timeout % 60);
 876         } else {
 877                 struct target_distribute_txn_data *tdtd;
 878                 int size = 0;
 879                 int count = 0;
 880                 char *buf = NULL;
 881
 882                 target_check_recovery_timer(target);
 883
 884                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 885                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 886                         buf = tdtd->tdtd_show_update_logs_retrievers(
 887                                 tdtd->tdtd_show_retrievers_cbdata,
 888                                 &size, &count);
 889
 890                 if (count > 0)
 891                         LCONSOLE_WARN("%s: Client %s (at %s) reconnecting, waiting for %d MDTs (%s) in recovery for %lld:%.02lld. Please wait until all MDTs recovered or you may force MDT evicition via 'lctl --device %s abort_recovery.\n",
 892                                       target->obd_name,
 893                                       obd_uuid2str(&exp->exp_client_uuid),
 894                                       obd_export_nid2str(exp), count,
 895                                       buf ? buf : "unknown (not enough RAM)",
 896                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 897                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 898                                       target->obd_name);
 899                 else
 900                         LCONSOLE_WARN("%s: Recovery already passed deadline %lld:%.02lld. If you do not want to wait more, you may force taget eviction via 'lctl --device %s abort_recovery.\n",
 901                                       target->obd_name, abs(timeout) / 60,
 902                                       abs(timeout) % 60, target->obd_name);
 903
 904                 if (buf != NULL)
 905                         OBD_FREE(buf, size);
 906         }
 907
 908 out_already:
 909         conn->cookie = exp->exp_handle.h_cookie;
 910         /*
 911          * target_handle_connect() treats EALREADY and
 912          * -EALREADY differently.  EALREADY means we are
 913          * doing a valid reconnect from the same client.
 914          */
 915         RETURN(EALREADY);
 916 }
 917
 918 static void
 919 check_and_start_recovery_timer(struct obd_device *obd,
 920                                struct ptlrpc_request *req, int new_client);
 921
 922 /**
 923  * update flags for import during reconnect process
 924  */
 925 static int rev_import_flags_update(struct obd_import *revimp,
 926                                    struct ptlrpc_request *req)
 927 {
 928         int rc;
 929         struct obd_connect_data *data;
 930
 931         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 932
 933         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 934                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 935         else
 936                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 937
 938         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 939
 940         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 941         if (rc) {
 942                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 943                         revimp->imp_client->cli_name,
 944                         libcfs_id2str(req->rq_peer), rc);
 945                 return rc;
 946         }
 947
 948         return 0;
 949 }
 950
 951 /**
 952  * Allocate a new reverse import for an export.
 953  *
 954  * \retval -errno in case error hit
 955  * \retval 0 if reverse import correctly init
 956  **/
 957 int rev_import_init(struct obd_export *export)
 958 {
 959         struct obd_device *obd = export->exp_obd;
 960         struct obd_import *revimp;
 961
 962         LASSERT(export->exp_imp_reverse == NULL);
 963
 964         revimp = class_new_import(obd);
 965         if (revimp == NULL)
 966                 return -ENOMEM;
 967
 968         revimp->imp_remote_handle.cookie = 0ULL;
 969         revimp->imp_client = &obd->obd_ldlm_client;
 970         revimp->imp_dlm_fake = 1;
 971
 972         /* it is safe to connect import in new state as no sends possible */
 973         spin_lock(&export->exp_lock);
 974         export->exp_imp_reverse = revimp;
 975         spin_unlock(&export->exp_lock);
 976         class_import_put(revimp);
 977
 978         return 0;
 979 }
 980 EXPORT_SYMBOL(rev_import_init);
 981
 982 /**
 983  * Handle reconnect for an export.
 984  *
 985  * \param exp export to handle reconnect process
 986  * \param req client reconnect request
 987  *
 988  * \retval -rc in case securitfy flavor can't be changed
 989  * \retval 0 in case none problems
 990  */
 991 static int rev_import_reconnect(struct obd_export *exp,
 992                                 struct ptlrpc_request *req)
 993 {
 994         struct obd_import *revimp = exp->exp_imp_reverse;
 995         struct lustre_handle *lh;
 996         int rc;
 997
 998         /* avoid sending a request until import flags are changed */
 999         ptlrpc_import_enter_resend(revimp);
1000
1001         if (revimp->imp_connection != NULL)
1002                 ptlrpc_connection_put(revimp->imp_connection);
1003
1004         /*
1005          * client from recovery don't have a handle so we need to take from
1006          * request. it may produce situation when wrong client connected
1007          * to recovery as we trust a client uuid
1008          */
1009         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1010         revimp->imp_remote_handle = *lh;
1011
1012         /*
1013          * unknown versions will be caught in
1014          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1015          */
1016         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1017
1018         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1019
1020         rc = rev_import_flags_update(revimp, req);
1021         if (rc != 0) {
1022                 /*
1023                  * it is safe to still be in RECOVERY phase as we are not able
1024                  * to setup correct security flavor so requests are not able to
1025                  * be delivered correctly
1026                  */
1027                 return rc;
1028         }
1029
1030         /* resend all rpc's via new connection */
1031         return ptlrpc_import_recovery_state_machine(revimp);
1032 }
1033
1034 int target_handle_connect(struct ptlrpc_request *req)
1035 {
1036         struct obd_device *target = NULL;
1037         struct obd_export *export = NULL;
1038         /*
1039          * connect handle - filled from target_handle_reconnect in
1040          * reconnect case
1041          */
1042         struct lustre_handle conn;
1043         struct lustre_handle *tmp;
1044         struct obd_uuid cluuid;
1045         char *str;
1046         int rc = 0;
1047         char *target_start;
1048         int target_len;
1049         bool mds_conn = false, lw_client = false, initial_conn = false;
1050         bool mds_mds_conn = false;
1051         bool new_mds_mds_conn = false;
1052         struct obd_connect_data *data, *tmpdata;
1053         int size, tmpsize;
1054         lnet_nid_t *client_nid = NULL;
1055
1056         ENTRY;
1057
1058         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1059
1060         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1061         if (str == NULL) {
1062                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1063                 GOTO(out, rc = -EINVAL);
1064         }
1065
1066         target = class_dev_by_str(str);
1067         if (!target) {
1068                 deuuidify(str, NULL, &target_start, &target_len);
1069                 LCONSOLE_ERROR_MSG(0x137,
1070                                    "%s: not available for connect from %s (no target). If you are running an HA pair check that the target is mounted on the other server.\n",
1071                                    str, libcfs_nid2str(req->rq_peer.nid));
1072                 GOTO(out, rc = -ENODEV);
1073         }
1074
1075         spin_lock(&target->obd_dev_lock);
1076
1077         target->obd_conn_inprogress++;
1078
1079         if (target->obd_stopping || !target->obd_set_up) {
1080                 spin_unlock(&target->obd_dev_lock);
1081
1082                 deuuidify(str, NULL, &target_start, &target_len);
1083                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1084                               target_len, target_start,
1085                               libcfs_nid2str(req->rq_peer.nid),
1086                               (target->obd_stopping ?
1087                                "stopping" : "not set up"));
1088                 GOTO(out, rc = -ENODEV);
1089         }
1090
1091         if (target->obd_no_conn) {
1092                 spin_unlock(&target->obd_dev_lock);
1093
1094                 CDEBUG(D_INFO,
1095                        "%s: Temporarily refusing client connection from %s\n",
1096                        target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1097                 GOTO(out, rc = -EAGAIN);
1098         }
1099
1100         spin_unlock(&target->obd_dev_lock);
1101
1102         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1103         if (str == NULL) {
1104                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1105                 GOTO(out, rc = -EINVAL);
1106         }
1107
1108         obd_str2uuid(&cluuid, str);
1109
1110         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1111         if (tmp == NULL)
1112                 GOTO(out, rc = -EPROTO);
1113
1114         conn = *tmp;
1115
1116         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1117                                     RCL_CLIENT);
1118         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1119                 GOTO(out, rc = -EPROTO);
1120         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1121         if (!data)
1122                 GOTO(out, rc = -EPROTO);
1123
1124         rc = req_capsule_server_pack(&req->rq_pill);
1125         if (rc)
1126                 GOTO(out, rc);
1127
1128 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1129         /*
1130          * Don't allow clients to connect that are using old 1.8 format
1131          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1132          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1133          * FULL20 flag should be set on all connections since 2.0, but no
1134          * longer affects behaviour.
1135          *
1136          * Later this check will be disabled and the flag can be retired
1137          * completely once interop with 3.0 is no longer needed.
1138          */
1139         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1140                 GOTO(out, rc = -EPROTO);
1141
1142         /*
1143          * Don't allow liblustre clients to connect.
1144          * - testing was disabled in v2_2_50_0-61-g6a75d65
1145          * - building was disabled in v2_5_58_0-28-g7277179
1146          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1147          * - clients were refused connect for version difference > 0.0.1.32
1148          */
1149         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1150                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1151                 GOTO(out, rc = -EPROTO);
1152         }
1153 #endif
1154
1155         /*
1156          * Note: lw_client is needed in MDS-MDS failover during update log
1157          * processing, so we needs to allow lw_client to be connected at
1158          * anytime, instead of only the initial connection
1159          */
1160         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1161
1162         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1163                 initial_conn = true;
1164                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1165                 mds_mds_conn = (data->ocd_connect_flags &
1166                                 OBD_CONNECT_MDS_MDS) != 0;
1167
1168                 /*
1169                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1170                  * Checking OBD_CONNECT_FID can be removed in the future.
1171                  *
1172                  * Via check OBD_CONNECT_FID, we can distinguish whether
1173                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1174                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1175                  */
1176                 if (!lw_client &&
1177                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1178                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1179                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1180                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1181                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1182                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1183
1184                         /*
1185                          * We do not support the MDT-MDT interoperations with
1186                          * different version MDT because of protocol changes.
1187                          */
1188                         if (unlikely(major != LUSTRE_MAJOR ||
1189                                      minor != LUSTRE_MINOR ||
1190                                      abs(patch - LUSTRE_PATCH) > 3)) {
1191                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1192                                               target->obd_name, LUSTRE_MAJOR,
1193                                               LUSTRE_MINOR, LUSTRE_PATCH,
1194                                               LUSTRE_FIX, major, minor, patch,
1195                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1196                                               libcfs_nid2str(req->rq_peer.nid),
1197                                               str);
1198                                 GOTO(out, rc = -EPROTO);
1199                         }
1200                 }
1201         }
1202
1203         /* lctl gets a backstage, all-access pass. */
1204         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1205                 goto dont_check_exports;
1206
1207         export = obd_uuid_lookup(target, &cluuid);
1208         if (!export)
1209                 goto no_export;
1210
1211         /* We've found an export in the hash. */
1212
1213         spin_lock(&export->exp_lock);
1214
1215         if (export->exp_connecting) { /* b=9635, et. al. */
1216                 spin_unlock(&export->exp_lock);
1217                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1218                               export->exp_obd->obd_name, export,
1219                               libcfs_nid2str(req->rq_peer.nid));
1220                 class_export_put(export);
1221                 export = NULL;
1222                 rc = -EALREADY;
1223         } else if ((mds_conn || (lw_client && initial_conn) ||
1224                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1225                    export->exp_connection != NULL) {
1226                 spin_unlock(&export->exp_lock);
1227                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1228                         /* MDS or LWP reconnected after failover. */
1229                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1230                                       target->obd_name,
1231                                       mds_conn ? "MDS" : "LWP",
1232                                       libcfs_nid2str(req->rq_peer.nid),
1233                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
1234                 } else {
1235                         /* New MDS connection from the same NID. */
1236                         LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1237                                       target->obd_name,
1238                                       mds_conn ? "MDS" : "LWP",
1239                                       libcfs_nid2str(req->rq_peer.nid));
1240                 }
1241
1242                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1243                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1244                         /*
1245                          * Because exports between MDTs will always be
1246                          * kept, let's do not fail such export if they
1247                          * come from the same NID, otherwise it might
1248                          * cause eviction between MDTs, which might
1249                          * cause namespace inconsistency
1250                          */
1251                         spin_lock(&export->exp_lock);
1252                         export->exp_connecting = 1;
1253                         export->exp_conn_cnt = 0;
1254                         spin_unlock(&export->exp_lock);
1255                         conn.cookie = export->exp_handle.h_cookie;
1256                         rc = EALREADY;
1257                 } else {
1258                         class_fail_export(export);
1259                         class_export_put(export);
1260                         export = NULL;
1261                         rc = 0;
1262                 }
1263         } else if (export->exp_connection != NULL && initial_conn &&
1264                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1265                 spin_unlock(&export->exp_lock);
1266                 /* In MDS failover we have static UUID but NID can change. */
1267                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1268                               target->obd_name, cluuid.uuid,
1269                               libcfs_nid2str(req->rq_peer.nid),
1270                               libcfs_nid2str(
1271                                         export->exp_connection->c_peer.nid));
1272                 rc = -EALREADY;
1273                 class_export_put(export);
1274                 export = NULL;
1275         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1276                    !lw_client) {
1277                 spin_unlock(&export->exp_lock);
1278                 rc = -EAGAIN;
1279         } else {
1280                 export->exp_connecting = 1;
1281                 spin_unlock(&export->exp_lock);
1282                 LASSERT(export->exp_obd == target);
1283
1284                 rc = target_handle_reconnect(&conn, export, &cluuid);
1285         }
1286
1287         /* If we found an export, we already unlocked. */
1288         if (!export) {
1289 no_export:
1290                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1291         } else if (req->rq_export == NULL &&
1292                    atomic_read(&export->exp_rpc_count) > 0) {
1293                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1294                               target->obd_name, cluuid.uuid,
1295                               libcfs_nid2str(req->rq_peer.nid),
1296                               refcount_read(&export->exp_handle.h_ref));
1297                         GOTO(out, rc = -EBUSY);
1298         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1299                    rc != EALREADY) {
1300                 if (!strstr(cluuid.uuid, "mdt"))
1301                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1302                                       target->obd_name, cluuid.uuid,
1303                                       libcfs_nid2str(req->rq_peer.nid));
1304                 GOTO(out, rc = -EALREADY);
1305         } else {
1306                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1307         }
1308
1309         if (rc < 0)
1310                 GOTO(out, rc);
1311
1312         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1313                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1314                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1315                export, ktime_get_seconds(),
1316                export ? export->exp_last_request_time : 0);
1317
1318         /*
1319          * If this is the first time a client connects, reset the recovery
1320          * timer. Discard lightweight connections which might be local.
1321          */
1322         if (!lw_client && rc == 0 && target->obd_recovering)
1323                 check_and_start_recovery_timer(target, req, export == NULL);
1324
1325         /*
1326          * We want to handle EALREADY but *not* -EALREADY from
1327          * target_handle_reconnect(), return reconnection state in a flag.
1328          */
1329         if (rc == EALREADY) {
1330                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1331                 rc = 0;
1332         } else {
1333                 LASSERT(rc == 0);
1334         }
1335
1336         /* Tell the client if we support replayable requests. */
1337         if (target->obd_replayable)
1338                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1339         client_nid = &req->rq_peer.nid;
1340
1341         if (export == NULL) {
1342                 /* allow lightweight connections during recovery */
1343                 /*
1344                  * allow "new" MDT to be connected during recovery, since we
1345                  * need retrieve recovery update records from it
1346                  */
1347                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1348                         struct hrtimer *timer = &target->obd_recovery_timer;
1349                         ktime_t remaining;
1350                         s64 timeout, left;
1351                         int in_progress;
1352                         int connected;
1353                         int known;
1354                         int stale;
1355                         char *msg;
1356
1357                         connected = atomic_read(&target->obd_connected_clients);
1358                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1359                         known =
1360                            atomic_read(&target->obd_max_recoverable_clients);
1361                         stale = target->obd_stale_clients;
1362                         remaining = hrtimer_expires_remaining(timer);
1363                         left = ktime_divns(remaining, NSEC_PER_SEC);
1364
1365                         if (ktime_to_ns(remaining) > 0) {
1366                                 msg = "to recover in";
1367                                 timeout = left;
1368                         } else {
1369                                 msg = "already passed deadline";
1370                                 timeout = -left;
1371
1372                                 target_check_recovery_timer(target);
1373                         }
1374
1375                         LCONSOLE_WARN("%s: Denying connection for new client %s (at %s), waiting for %d known clients (%d recovered, %d in progress, and %d evicted) %s %lld:%.02lld\n",
1376                                       target->obd_name, cluuid.uuid,
1377                                       libcfs_nid2str(req->rq_peer.nid), known,
1378                                       connected - in_progress, in_progress,
1379                                       stale, msg, timeout / 60, timeout % 60);
1380                         rc = -EBUSY;
1381                 } else {
1382 dont_check_exports:
1383                         rc = obd_connect(req->rq_svc_thread->t_env,
1384                                          &export, target, &cluuid, data,
1385                                          client_nid);
1386                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1387                                 lustre_msg_add_op_flags(req->rq_repmsg,
1388                                                         MSG_CONNECT_RECOVERING);
1389                         if (rc == 0) {
1390                                 conn.cookie = export->exp_handle.h_cookie;
1391                                 rc = rev_import_init(export);
1392                         }
1393
1394                         if (mds_mds_conn)
1395                                 new_mds_mds_conn = true;
1396                 }
1397         } else {
1398                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1399                                    export, target, &cluuid, data, client_nid);
1400         }
1401         if (rc)
1402                 GOTO(out, rc);
1403
1404         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1405         data->ocd_instance = target->u.obt.obt_instance;
1406
1407         /*
1408          * Return only the parts of obd_connect_data that we understand, so the
1409          * client knows that we don't understand the rest.
1410          */
1411         if (data) {
1412                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1413                                                RCL_SERVER);
1414                 tmpdata = req_capsule_server_get(&req->rq_pill,
1415                                                  &RMF_CONNECT_DATA);
1416                 /*
1417                  * Don't use struct assignment here, because the client reply
1418                  * buffer may be smaller/larger than the local struct
1419                  * obd_connect_data.
1420                  */
1421                 memcpy(tmpdata, data, min(tmpsize, size));
1422         }
1423
1424         /*
1425          * If the client and the server are the same node, we will already
1426          * have an export that really points to the client's DLM export,
1427          * because we have a shared handles table.
1428          *
1429          * XXX this will go away when shaver stops sending the "connect" handle
1430          * in the real "remote handle" field of the request --phik 24 Apr 2003
1431          */
1432         ptlrpc_request_change_export(req, export);
1433
1434         spin_lock(&export->exp_lock);
1435         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1436                 spin_unlock(&export->exp_lock);
1437                 CDEBUG(D_RPCTRACE,
1438                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1439                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1440                        export->exp_conn_cnt,
1441                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1442
1443                 GOTO(out, rc = -EALREADY);
1444         }
1445         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1446         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1447         spin_unlock(&export->exp_lock);
1448
1449         if (export->exp_connection != NULL) {
1450                 /* Check to see if connection came from another NID. */
1451                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1452                     !hlist_unhashed(&export->exp_nid_hash))
1453                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1454                                      &export->exp_connection->c_peer.nid,
1455                                      &export->exp_nid_hash);
1456
1457                 ptlrpc_connection_put(export->exp_connection);
1458         }
1459
1460         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1461                                                        req->rq_self,
1462                                                        &cluuid);
1463         if (hlist_unhashed(&export->exp_nid_hash))
1464                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1465                              &export->exp_connection->c_peer.nid,
1466                              &export->exp_nid_hash);
1467
1468         lustre_msg_set_handle(req->rq_repmsg, &conn);
1469
1470         rc = rev_import_reconnect(export, req);
1471         if (rc != 0)
1472                 GOTO(out, rc);
1473
1474         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1475                 int has_transno;
1476                 __u64 transno = data->ocd_transno;
1477
1478                 spin_lock(&export->exp_lock);
1479                 /*
1480                  * possible race with class_disconnect_stale_exports,
1481                  * export may be already in the eviction process
1482                  */
1483                 if (export->exp_failed) {
1484                         spin_unlock(&export->exp_lock);
1485                         GOTO(out, rc = -ENODEV);
1486                 }
1487                 export->exp_in_recovery = 1;
1488                 export->exp_req_replay_needed = 1;
1489                 export->exp_lock_replay_needed = 1;
1490                 spin_unlock(&export->exp_lock);
1491
1492                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1493                                  MSG_CONNECT_TRANSNO);
1494                 if (has_transno && transno == 0)
1495                         CWARN("Connect with zero transno!\n");
1496
1497                 if (has_transno && transno > 0 &&
1498                     transno < target->obd_next_recovery_transno &&
1499                     transno > target->obd_last_committed) {
1500                         /* Another way is to use cmpxchg() to be lock-free. */
1501                         spin_lock(&target->obd_recovery_task_lock);
1502                         if (transno < target->obd_next_recovery_transno)
1503                                 target->obd_next_recovery_transno = transno;
1504                         spin_unlock(&target->obd_recovery_task_lock);
1505                 }
1506
1507                 atomic_inc(&target->obd_req_replay_clients);
1508                 atomic_inc(&target->obd_lock_replay_clients);
1509                 /*
1510                  * Note: MDS-MDS connection is allowed to be connected during
1511                  * recovery, no matter if the exports needs to be recoveried.
1512                  * Because we need retrieve updates logs from all other MDTs.
1513                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1514                  * also needs to be increased to match other recovery checking
1515                  * condition.
1516                  */
1517                 if (new_mds_mds_conn)
1518                         atomic_inc(&target->obd_max_recoverable_clients);
1519
1520                 if (atomic_inc_return(&target->obd_connected_clients) ==
1521                     atomic_read(&target->obd_max_recoverable_clients))
1522                         wake_up(&target->obd_next_transno_waitq);
1523         }
1524
1525         /* Tell the client we're in recovery, when client is involved in it. */
1526         if (target->obd_recovering && !lw_client)
1527                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1528
1529 out:
1530         if (export) {
1531                 spin_lock(&export->exp_lock);
1532                 export->exp_connecting = 0;
1533                 spin_unlock(&export->exp_lock);
1534
1535                 class_export_put(export);
1536         }
1537         if (target != NULL) {
1538                 spin_lock(&target->obd_dev_lock);
1539                 target->obd_conn_inprogress--;
1540                 spin_unlock(&target->obd_dev_lock);
1541                 class_decref(target, "find", current);
1542         }
1543         req->rq_status = rc;
1544         RETURN(rc);
1545 }
1546
1547 int target_handle_disconnect(struct ptlrpc_request *req)
1548 {
1549         int rc;
1550
1551         ENTRY;
1552
1553         rc = req_capsule_server_pack(&req->rq_pill);
1554         if (rc)
1555                 RETURN(rc);
1556
1557         /* Keep the rq_export around so we can send the reply. */
1558         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1559
1560         RETURN(0);
1561 }
1562
1563 void target_destroy_export(struct obd_export *exp)
1564 {
1565         struct obd_import *imp = NULL;
1566         /*
1567          * exports created from last_rcvd data, and "fake"
1568          * exports created by lctl don't have an import
1569          */
1570         spin_lock(&exp->exp_lock);
1571         if (exp->exp_imp_reverse != NULL) {
1572                 imp = exp->exp_imp_reverse;
1573                 exp->exp_imp_reverse = NULL;
1574         }
1575         spin_unlock(&exp->exp_lock);
1576         if (imp != NULL)
1577                 client_destroy_import(imp);
1578
1579         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1580         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1581         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1582         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1583 }
1584 EXPORT_SYMBOL(target_destroy_export);
1585
1586 /*
1587  * Recovery functions
1588  */
1589 static void target_request_copy_get(struct ptlrpc_request *req)
1590 {
1591         class_export_rpc_inc(req->rq_export);
1592         LASSERT(list_empty(&req->rq_list));
1593         INIT_LIST_HEAD(&req->rq_replay_list);
1594
1595         /* Increase refcount to keep request in queue. */
1596         atomic_inc(&req->rq_refcount);
1597         /* Let export know it has replays to be handled. */
1598         atomic_inc(&req->rq_export->exp_replay_count);
1599 }
1600
1601 static void target_request_copy_put(struct ptlrpc_request *req)
1602 {
1603         LASSERT(list_empty(&req->rq_replay_list));
1604         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1605
1606         atomic_dec(&req->rq_export->exp_replay_count);
1607         class_export_rpc_dec(req->rq_export);
1608         ptlrpc_server_drop_request(req);
1609 }
1610
1611 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1612 {
1613         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1614         struct obd_export *exp = req->rq_export;
1615         struct ptlrpc_request *reqiter;
1616         struct ptlrpc_request *dup_req = NULL;
1617         int dup = 0;
1618
1619         LASSERT(exp);
1620
1621         spin_lock(&exp->exp_lock);
1622         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1623                             rq_replay_list) {
1624                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1625                         dup_req = reqiter;
1626                         dup = 1;
1627                         break;
1628                 }
1629         }
1630
1631         if (dup) {
1632                 /* We expect it with RESENT and REPLAY flags. */
1633                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1634                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1635                         CERROR("invalid flags %x of resent replay\n",
1636                                lustre_msg_get_flags(req->rq_reqmsg));
1637
1638                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1639                         __u32 new_conn;
1640
1641                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1642                         if (new_conn >
1643                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1644                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1645                                                         new_conn);
1646                 }
1647         } else {
1648                 list_add_tail(&req->rq_replay_list,
1649                               &exp->exp_req_replay_queue);
1650         }
1651
1652         spin_unlock(&exp->exp_lock);
1653         return dup;
1654 }
1655
1656 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1657 {
1658         LASSERT(!list_empty(&req->rq_replay_list));
1659         LASSERT(req->rq_export);
1660
1661         spin_lock(&req->rq_export->exp_lock);
1662         list_del_init(&req->rq_replay_list);
1663         spin_unlock(&req->rq_export->exp_lock);
1664 }
1665
1666 static void target_finish_recovery(struct lu_target *lut)
1667 {
1668         struct obd_device *obd = lut->lut_obd;
1669
1670         ENTRY;
1671
1672         /* Only log a recovery message when recovery has occurred. */
1673         if (obd->obd_recovery_start) {
1674                 time64_t now = ktime_get_seconds();
1675                 time64_t elapsed_time;
1676
1677                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1678                                      1);
1679                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1680                               obd->obd_name, elapsed_time / 60,
1681                               elapsed_time % 60,
1682                               atomic_read(&obd->obd_max_recoverable_clients),
1683                               atomic_read(&obd->obd_connected_clients),
1684                               obd->obd_stale_clients,
1685                               obd->obd_stale_clients == 1 ? "was" : "were");
1686         }
1687
1688         ldlm_reprocess_recovery_done(obd->obd_namespace);
1689         spin_lock(&obd->obd_recovery_task_lock);
1690         if (!list_empty(&obd->obd_req_replay_queue) ||
1691             !list_empty(&obd->obd_lock_replay_queue) ||
1692             !list_empty(&obd->obd_final_req_queue)) {
1693                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1694                        obd->obd_name,
1695                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1696                        list_empty(&obd->obd_lock_replay_queue) ? \
1697                                   "" : "lock ",
1698                        list_empty(&obd->obd_final_req_queue) ? \
1699                                   "" : "final ");
1700                 spin_unlock(&obd->obd_recovery_task_lock);
1701                 LBUG();
1702         }
1703         spin_unlock(&obd->obd_recovery_task_lock);
1704
1705         obd->obd_recovery_end = ktime_get_seconds();
1706
1707         /* When recovery finished, cleanup orphans on MDS and OST. */
1708         if (obd->obd_type && OBP(obd, postrecov)) {
1709                 int rc = OBP(obd, postrecov)(obd);
1710
1711                 if (rc < 0)
1712                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1713                                       obd->obd_name, rc);
1714         }
1715         EXIT;
1716 }
1717
1718 static void abort_req_replay_queue(struct obd_device *obd)
1719 {
1720         struct ptlrpc_request *req, *n;
1721         LIST_HEAD(abort_list);
1722
1723         spin_lock(&obd->obd_recovery_task_lock);
1724         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1725         spin_unlock(&obd->obd_recovery_task_lock);
1726         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1727                 DEBUG_REQ(D_WARNING, req, "aborted:");
1728                 req->rq_status = -ENOTCONN;
1729                 if (ptlrpc_error(req)) {
1730                         DEBUG_REQ(D_ERROR, req,
1731                                   "failed abort_req_reply; skipping");
1732                 }
1733                 target_exp_dequeue_req_replay(req);
1734                 target_request_copy_put(req);
1735         }
1736 }
1737
1738 static void abort_lock_replay_queue(struct obd_device *obd)
1739 {
1740         struct ptlrpc_request *req, *n;
1741         LIST_HEAD(abort_list);
1742
1743         spin_lock(&obd->obd_recovery_task_lock);
1744         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1745         spin_unlock(&obd->obd_recovery_task_lock);
1746         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1747                 DEBUG_REQ(D_ERROR, req, "aborted:");
1748                 req->rq_status = -ENOTCONN;
1749                 if (ptlrpc_error(req)) {
1750                         DEBUG_REQ(D_ERROR, req,
1751                                   "failed abort_lock_reply; skipping");
1752                 }
1753                 target_request_copy_put(req);
1754         }
1755 }
1756
1757 /*
1758  * Called from a cleanup function if the device is being cleaned up
1759  * forcefully.  The exports should all have been disconnected already,
1760  * the only thing left to do is
1761  * - clear the recovery flags
1762  * - cancel the timer
1763  * - free queued requests and replies, but don't send replies
1764  * Because the obd_stopping flag is set, no new requests should be received.
1765  */
1766 void target_cleanup_recovery(struct obd_device *obd)
1767 {
1768         struct ptlrpc_request *req, *n;
1769         LIST_HEAD(clean_list);
1770
1771         spin_lock(&obd->obd_dev_lock);
1772         if (!obd->obd_recovering) {
1773                 spin_unlock(&obd->obd_dev_lock);
1774                 EXIT;
1775                 return;
1776         }
1777         obd->obd_recovering = obd->obd_abort_recovery = 0;
1778         spin_unlock(&obd->obd_dev_lock);
1779
1780         spin_lock(&obd->obd_recovery_task_lock);
1781         target_cancel_recovery_timer(obd);
1782         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1783         spin_unlock(&obd->obd_recovery_task_lock);
1784
1785         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1786                 LASSERT(req->rq_reply_state == NULL);
1787                 target_exp_dequeue_req_replay(req);
1788                 target_request_copy_put(req);
1789         }
1790
1791         spin_lock(&obd->obd_recovery_task_lock);
1792         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1793         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1794         spin_unlock(&obd->obd_recovery_task_lock);
1795
1796         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1797                 LASSERT(req->rq_reply_state == NULL);
1798                 target_request_copy_put(req);
1799         }
1800
1801         EXIT;
1802 }
1803 EXPORT_SYMBOL(target_cleanup_recovery);
1804
1805 /* obd_recovery_task_lock should be held */
1806 void target_cancel_recovery_timer(struct obd_device *obd)
1807 {
1808         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1809         hrtimer_cancel(&obd->obd_recovery_timer);
1810 }
1811
1812 static void target_start_recovery_timer(struct obd_device *obd)
1813 {
1814         ktime_t delay;
1815
1816         if (obd->obd_recovery_start != 0)
1817                 return;
1818
1819         spin_lock(&obd->obd_dev_lock);
1820         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1821                 spin_unlock(&obd->obd_dev_lock);
1822                 return;
1823         }
1824
1825         LASSERT(obd->obd_recovery_timeout != 0);
1826
1827         if (obd->obd_recovery_start != 0) {
1828                 spin_unlock(&obd->obd_dev_lock);
1829                 return;
1830         }
1831
1832         obd->obd_recovery_start = ktime_get_seconds();
1833         delay = ktime_set(obd->obd_recovery_start +
1834                           obd->obd_recovery_timeout, 0);
1835         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1836         spin_unlock(&obd->obd_dev_lock);
1837
1838         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1839                       obd->obd_name,
1840                       obd->obd_recovery_timeout / 60,
1841                       obd->obd_recovery_timeout % 60,
1842                       atomic_read(&obd->obd_max_recoverable_clients),
1843                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1844                       "" : "s",
1845                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1846                       "s" : "");
1847 }
1848
1849 /**
1850  * extend recovery window.
1851  *
1852  * if @extend is true, extend recovery window to have @dr_timeout remaining
1853  * at least; otherwise, make sure the recovery timeout value is not less
1854  * than @dr_timeout.
1855  */
1856 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1857                                   bool extend)
1858 {
1859         ktime_t left_ns;
1860         timeout_t timeout;
1861         timeout_t left;
1862
1863         spin_lock(&obd->obd_dev_lock);
1864         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1865             obd->obd_stopping) {
1866                 spin_unlock(&obd->obd_dev_lock);
1867                 return;
1868         }
1869         LASSERT(obd->obd_recovery_start != 0);
1870
1871         left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1872         left = ktime_divns(left_ns, NSEC_PER_SEC);
1873
1874         if (extend) {
1875                 timeout = obd->obd_recovery_timeout;
1876                 /* dr_timeout will happen after the hrtimer has expired.
1877                  * Add the excess time to the soft recovery timeout without
1878                  * exceeding the hard recovery timeout.
1879                  */
1880                 if (dr_timeout > left) {
1881                         timeout += dr_timeout - left;
1882                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1883                                         timeout);
1884                 }
1885         } else {
1886                 timeout = clamp_t(timeout_t, dr_timeout,
1887                                   obd->obd_recovery_timeout,
1888                                   obd->obd_recovery_time_hard);
1889         }
1890
1891         if (timeout == obd->obd_recovery_time_hard)
1892                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1893                       obd->obd_name, timeout, extend);
1894
1895         if (obd->obd_recovery_timeout < timeout) {
1896                 ktime_t end, now;
1897
1898                 obd->obd_recovery_timeout = timeout;
1899                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1900                 now = ktime_set(ktime_get_seconds(), 0);
1901                 left_ns = ktime_sub(end, now);
1902                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1903                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1904         }
1905         spin_unlock(&obd->obd_dev_lock);
1906
1907         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1908                 obd->obd_name, left);
1909 }
1910
1911 /* Reset the timer with each new client connection */
1912 /*
1913  * This timer is actually reconnect_timer, which is for making sure
1914  * the total recovery window is at least as big as my reconnect
1915  * attempt timing. So the initial recovery time_out will be set to
1916  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1917  * from client is bigger than this, then the recovery time_out will
1918  * be extended to make sure the client could be reconnected, in the
1919  * process, the timeout from the new client should be ignored.
1920  */
1921 static void
1922 check_and_start_recovery_timer(struct obd_device *obd,
1923                                struct ptlrpc_request *req,
1924                                int new_client)
1925 {
1926         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1927         struct obd_device_target *obt = &obd->u.obt;
1928
1929         if (!new_client && service_timeout)
1930                 /*
1931                  * Teach server about old server's estimates, as first guess
1932                  * at how long new requests will take.
1933                  */
1934                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1935                             service_timeout);
1936
1937         target_start_recovery_timer(obd);
1938
1939         /*
1940          * Convert the service time to RPC timeout,
1941          * and reuse service_timeout to limit stack usage.
1942          */
1943         service_timeout = at_est2timeout(service_timeout);
1944
1945         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1946             service_timeout < at_extra)
1947                 service_timeout = at_extra;
1948
1949         /*
1950          * We expect other clients to timeout within service_timeout, then try
1951          * to reconnect, then try the failover server.  The max delay between
1952          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1953          */
1954         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
1955
1956         LASSERT(obt->obt_magic == OBT_MAGIC);
1957         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1958         if (service_timeout > obd->obd_recovery_timeout && !new_client)
1959                 extend_recovery_timer(obd, service_timeout, false);
1960 }
1961
1962 /** Health checking routines */
1963 static inline int exp_connect_healthy(struct obd_export *exp)
1964 {
1965         return exp->exp_in_recovery;
1966 }
1967
1968 /** if export done req_replay or has replay in queue */
1969 static inline int exp_req_replay_healthy(struct obd_export *exp)
1970 {
1971         return (!exp->exp_req_replay_needed ||
1972                 atomic_read(&exp->exp_replay_count) > 0);
1973 }
1974
1975
1976 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1977 {
1978         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1979                exp_req_replay_healthy(exp);
1980 }
1981
1982 /** if export done lock_replay or has replay in queue */
1983 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1984 {
1985         return (!exp->exp_lock_replay_needed ||
1986                 atomic_read(&exp->exp_replay_count) > 0);
1987 }
1988
1989 static inline int exp_vbr_healthy(struct obd_export *exp)
1990 {
1991         return !exp->exp_vbr_failed;
1992 }
1993
1994 static inline int exp_finished(struct obd_export *exp)
1995 {
1996         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1997 }
1998
1999 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2000 {
2001         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2002                 exp_finished(exp);
2003 }
2004
2005 static int check_for_next_transno(struct lu_target *lut)
2006 {
2007         struct ptlrpc_request *req = NULL;
2008         struct obd_device *obd = lut->lut_obd;
2009         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2010         int wake_up = 0, connected, completed, queue_len;
2011         __u64 req_transno = 0;
2012         __u64 update_transno = 0;
2013         __u64 next_transno = 0;
2014
2015         ENTRY;
2016
2017         spin_lock(&obd->obd_recovery_task_lock);
2018         if (!list_empty(&obd->obd_req_replay_queue)) {
2019                 req = list_entry(obd->obd_req_replay_queue.next,
2020                                      struct ptlrpc_request, rq_list);
2021                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2022         }
2023
2024         if (tdtd != NULL)
2025                 update_transno = distribute_txn_get_next_transno(tdtd);
2026
2027         connected = atomic_read(&obd->obd_connected_clients);
2028         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2029         queue_len = obd->obd_requests_queued_for_recovery;
2030         next_transno = obd->obd_next_recovery_transno;
2031
2032         CDEBUG(D_HA,
2033                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2034                atomic_read(&obd->obd_max_recoverable_clients),
2035                connected, completed,
2036                queue_len, req_transno, next_transno);
2037
2038         if (obd->obd_abort_recovery) {
2039                 CDEBUG(D_HA, "waking for aborted recovery\n");
2040                 wake_up = 1;
2041         } else if (obd->obd_recovery_expired) {
2042                 CDEBUG(D_HA, "waking for expired recovery\n");
2043                 wake_up = 1;
2044         } else if (tdtd != NULL && req != NULL &&
2045                    is_req_replayed_by_update(req)) {
2046                 LASSERTF(req_transno < next_transno,
2047                          "req_transno %llu next_transno%llu\n", req_transno,
2048                          next_transno);
2049                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2050                        req_transno);
2051                 wake_up = 1;
2052         } else if (req_transno == next_transno ||
2053                    (update_transno != 0 && update_transno <= next_transno)) {
2054                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2055                 wake_up = 1;
2056         } else if (queue_len > 0 &&
2057                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2058                 /** handle gaps occured due to lost reply or VBR */
2059                 LASSERTF(req_transno >= next_transno,
2060                          "req_transno: %llu, next_transno: %llu\n",
2061                          req_transno, next_transno);
2062                 CDEBUG(D_HA,
2063                        "%s: waking for gap in transno, VBR is %s (skip: %lld, ql: %d, comp: %d, conn: %d, next: %lld, next_update %lld last_committed: %lld)\n",
2064                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2065                        next_transno, queue_len, completed, connected,
2066                        req_transno, update_transno, obd->obd_last_committed);
2067                 obd->obd_next_recovery_transno = req_transno;
2068                 wake_up = 1;
2069         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2070                 CDEBUG(D_HA, "waking for completed recovery\n");
2071                 wake_up = 1;
2072         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2073                 CDEBUG(D_HA,
2074                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2075                        next_transno);
2076                 obd->obd_next_recovery_transno = req_transno;
2077                 wake_up = 1;
2078         }
2079         spin_unlock(&obd->obd_recovery_task_lock);
2080         return wake_up;
2081 }
2082
2083 static int check_for_next_lock(struct lu_target *lut)
2084 {
2085         struct obd_device *obd = lut->lut_obd;
2086         int wake_up = 0;
2087
2088         spin_lock(&obd->obd_recovery_task_lock);
2089         if (!list_empty(&obd->obd_lock_replay_queue)) {
2090                 CDEBUG(D_HA, "waking for next lock\n");
2091                 wake_up = 1;
2092         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2093                 CDEBUG(D_HA, "waking for completed lock replay\n");
2094                 wake_up = 1;
2095         } else if (obd->obd_abort_recovery) {
2096                 CDEBUG(D_HA, "waking for aborted recovery\n");
2097                 wake_up = 1;
2098         } else if (obd->obd_recovery_expired) {
2099                 CDEBUG(D_HA, "waking for expired recovery\n");
2100                 wake_up = 1;
2101         }
2102         spin_unlock(&obd->obd_recovery_task_lock);
2103
2104         return wake_up;
2105 }
2106
2107 /**
2108  * wait for recovery events,
2109  * check its status with help of check_routine
2110  * evict dead clients via health_check
2111  */
2112 static int target_recovery_overseer(struct lu_target *lut,
2113                                     int (*check_routine)(struct lu_target *),
2114                                     int (*health_check)(struct obd_export *))
2115 {
2116         struct obd_device *obd = lut->lut_obd;
2117         struct target_distribute_txn_data *tdtd;
2118         time64_t last = 0;
2119         time64_t now;
2120 repeat:
2121         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2122                 now = ktime_get_seconds();
2123                 if (now - last > 600) {
2124                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2125                                       obd->obd_name);
2126                         last = now;
2127                 }
2128         }
2129         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2130               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2131                 __u64 next_update_transno = 0;
2132
2133                 /*
2134                  * Only abort the recovery if there are no update recovery
2135                  * left in the queue
2136                  */
2137                 spin_lock(&obd->obd_recovery_task_lock);
2138                 if (lut->lut_tdtd != NULL) {
2139                         next_update_transno =
2140                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2141
2142                         tdtd = lut->lut_tdtd;
2143                         /*
2144                          * If next_update_transno == 0, it probably because
2145                          * updatelog retrieve threads did not get any records
2146                          * yet, let's wait those threads stopped
2147                          */
2148                         if (next_update_transno == 0) {
2149                                 spin_unlock(&obd->obd_recovery_task_lock);
2150                                 wait_event_idle(
2151                                         tdtd->tdtd_recovery_threads_waitq,
2152                                         atomic_read(&tdtd->tdtd_recovery_threads_count)
2153                                         == 0);
2154
2155                                 spin_lock(&obd->obd_recovery_task_lock);
2156                                 next_update_transno =
2157                                         distribute_txn_get_next_transno(
2158                                                                 lut->lut_tdtd);
2159                         }
2160                 }
2161
2162                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2163                         obd->obd_next_recovery_transno = next_update_transno;
2164                         spin_unlock(&obd->obd_recovery_task_lock);
2165                         /*
2166                          * Disconnect unfinished exports from clients, and
2167                          * keep connection from MDT to make sure the update
2168                          * recovery will still keep trying until some one
2169                          * manually abort the recovery
2170                          */
2171                         class_disconnect_stale_exports(obd,
2172                                                 exp_finished_or_from_mdt);
2173                         /* Abort all of replay & replay lock req from clients */
2174                         abort_req_replay_queue(obd);
2175                         abort_lock_replay_queue(obd);
2176                         CDEBUG(D_HA,
2177                                "%s: there are still update replay (%#llx)in the queue.\n",
2178                                obd->obd_name, next_update_transno);
2179                 } else {
2180                         obd->obd_abort_recovery = 1;
2181                         spin_unlock(&obd->obd_recovery_task_lock);
2182                         CWARN("%s recovery is aborted by hard timeout\n",
2183                               obd->obd_name);
2184                 }
2185         }
2186
2187         while (wait_event_timeout(obd->obd_next_transno_waitq,
2188                                   check_routine(lut),
2189                                   cfs_time_seconds(60)) == 0)
2190                 ; /* wait indefinitely for event, but don't trigger watchdog */
2191
2192         if (obd->obd_abort_recovery) {
2193                 CWARN("recovery is aborted, evict exports in recovery\n");
2194                 if (lut->lut_tdtd != NULL) {
2195                         tdtd = lut->lut_tdtd;
2196                         /*
2197                          * Let's wait all of the update log recovery thread
2198                          * finished
2199                          */
2200                         wait_event_idle(
2201                                 tdtd->tdtd_recovery_threads_waitq,
2202                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2203                                 == 0);
2204                         /* Then abort the update recovery list */
2205                         dtrq_list_destroy(lut->lut_tdtd);
2206                 }
2207
2208                 /** evict exports which didn't finish recovery yet */
2209                 class_disconnect_stale_exports(obd, exp_finished);
2210                 return 1;
2211         } else if (obd->obd_recovery_expired &&
2212                    obd->obd_recovery_timeout < obd->obd_recovery_time_hard) {
2213                 obd->obd_recovery_expired = 0;
2214
2215                 /** If some clients died being recovered, evict them */
2216                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2217                               obd->obd_name);
2218                 /** evict cexports with no replay in queue, they are stalled */
2219                 class_disconnect_stale_exports(obd, health_check);
2220
2221                 /** continue with VBR */
2222                 spin_lock(&obd->obd_dev_lock);
2223                 obd->obd_version_recov = 1;
2224                 spin_unlock(&obd->obd_dev_lock);
2225                 /**
2226                  * reset timer, recovery will proceed with versions now,
2227                  * timeout is set just to handle reconnection delays
2228                  */
2229                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2230                 /**
2231                  * Wait for recovery events again, after evicting bad clients
2232                  */
2233                 goto repeat;
2234         }
2235         return 0;
2236 }
2237
2238 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2239 {
2240         struct obd_device *obd = lut->lut_obd;
2241         struct ptlrpc_request *req = NULL;
2242
2243         CDEBUG(D_HA, "Waiting for lock\n");
2244         if (target_recovery_overseer(lut, check_for_next_lock,
2245                                      exp_lock_replay_healthy))
2246                 abort_lock_replay_queue(obd);
2247
2248         spin_lock(&obd->obd_recovery_task_lock);
2249         if (!list_empty(&obd->obd_lock_replay_queue)) {
2250                 req = list_entry(obd->obd_lock_replay_queue.next,
2251                                      struct ptlrpc_request, rq_list);
2252                 list_del_init(&req->rq_list);
2253                 spin_unlock(&obd->obd_recovery_task_lock);
2254         } else {
2255                 spin_unlock(&obd->obd_recovery_task_lock);
2256                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2257                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2258                 /** evict exports failed VBR */
2259                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2260         }
2261         return req;
2262 }
2263
2264 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2265 {
2266         struct ptlrpc_request *req = NULL;
2267
2268         spin_lock(&obd->obd_recovery_task_lock);
2269         if (!list_empty(&obd->obd_final_req_queue)) {
2270                 req = list_entry(obd->obd_final_req_queue.next,
2271                                      struct ptlrpc_request, rq_list);
2272                 list_del_init(&req->rq_list);
2273                 spin_unlock(&obd->obd_recovery_task_lock);
2274                 if (req->rq_export->exp_in_recovery) {
2275                         spin_lock(&req->rq_export->exp_lock);
2276                         req->rq_export->exp_in_recovery = 0;
2277                         spin_unlock(&req->rq_export->exp_lock);
2278                 }
2279         } else {
2280                 spin_unlock(&obd->obd_recovery_task_lock);
2281         }
2282         return req;
2283 }
2284
2285 static void handle_recovery_req(struct ptlrpc_thread *thread,
2286                                 struct ptlrpc_request *req,
2287                                 svc_handler_t handler)
2288 {
2289         ENTRY;
2290
2291         /**
2292          * export can be evicted during recovery, no need to handle replays for
2293          * it after that, discard such request silently
2294          */
2295         if (req->rq_export->exp_disconnected)
2296                 RETURN_EXIT;
2297
2298         req->rq_session.lc_thread = thread;
2299         req->rq_svc_thread = thread;
2300         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2301
2302         /* thread context */
2303         lu_context_enter(&thread->t_env->le_ctx);
2304         (void)handler(req);
2305         lu_context_exit(&thread->t_env->le_ctx);
2306
2307         req->rq_svc_thread->t_env->le_ses = NULL;
2308
2309         /* don't reset timer for final stage */
2310         if (!exp_finished(req->rq_export)) {
2311                 timeout_t timeout = obd_timeout;
2312
2313                 /**
2314                  * Add request @timeout to the recovery time so next request from
2315                  * this client may come in recovery time
2316                  */
2317                 if (!AT_OFF) {
2318                         struct ptlrpc_service_part *svcpt;
2319                         timeout_t est_timeout;
2320
2321                         svcpt = req->rq_rqbd->rqbd_svcpt;
2322                         /*
2323                          * If the server sent early reply for this request,
2324                          * the client will recalculate the timeout according to
2325                          * current server estimate service time, so we will
2326                          * use the maxium timeout here for waiting the client
2327                          * sending the next req
2328                          */
2329                         est_timeout = at_get(&svcpt->scp_at_estimate);
2330                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2331                                         lustre_msg_get_timeout(req->rq_reqmsg));
2332                         /*
2333                          * Add 2 net_latency, one for balance rq_deadline
2334                          * (see ptl_send_rpc), one for resend the req to server,
2335                          * Note: client will pack net_latency in replay req
2336                          * (see ptlrpc_replay_req)
2337                          */
2338                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2339                 }
2340                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2341                                       true);
2342         }
2343         EXIT;
2344 }
2345
2346 /** Checking routines for recovery */
2347 static int check_for_recovery_ready(struct lu_target *lut)
2348 {
2349         struct obd_device *obd = lut->lut_obd;
2350         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2351
2352         CDEBUG(D_HA,
2353                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2354                clnts, obd->obd_stale_clients,
2355                atomic_read(&obd->obd_max_recoverable_clients),
2356                obd->obd_abort_recovery, obd->obd_recovery_expired);
2357
2358         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2359                 LASSERT(clnts <=
2360                         atomic_read(&obd->obd_max_recoverable_clients));
2361                 if (clnts + obd->obd_stale_clients <
2362                     atomic_read(&obd->obd_max_recoverable_clients))
2363                         return 0;
2364         }
2365
2366         if (lut->lut_tdtd != NULL) {
2367                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2368                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2369                         /*
2370                          * Let's extend recovery timer, in case the recovery
2371                          * timer expired, and some clients got evicted
2372                          */
2373                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2374                                               true);
2375                         CDEBUG(D_HA,
2376                                "%s update recovery is not ready, extend recovery %d\n",
2377                                obd->obd_name, obd->obd_recovery_timeout);
2378                         return 0;
2379                 }
2380         }
2381
2382         return 1;
2383 }
2384
2385 enum {
2386         REQUEST_RECOVERY = 1,
2387         UPDATE_RECOVERY = 2,
2388 };
2389
2390 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2391 {
2392         __u64 transno = 0;
2393
2394         if (!list_empty(&obd->obd_req_replay_queue)) {
2395                 struct ptlrpc_request *req;
2396
2397                 req = list_entry(obd->obd_req_replay_queue.next,
2398                                  struct ptlrpc_request, rq_list);
2399                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2400         }
2401
2402         return transno;
2403 }
2404
2405 static __u64 get_next_transno(struct lu_target *lut, int *type)
2406 {
2407         struct obd_device *obd = lut->lut_obd;
2408         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2409         __u64 transno = 0;
2410         __u64 update_transno;
2411
2412         ENTRY;
2413
2414         transno = get_next_replay_req_transno(obd);
2415         if (type != NULL)
2416                 *type = REQUEST_RECOVERY;
2417
2418         if (tdtd == NULL)
2419                 RETURN(transno);
2420
2421         update_transno = distribute_txn_get_next_transno(tdtd);
2422         if (transno == 0 || (transno >= update_transno &&
2423                              update_transno != 0)) {
2424                 transno = update_transno;
2425                 if (type != NULL)
2426                         *type = UPDATE_RECOVERY;
2427         }
2428
2429         RETURN(transno);
2430 }
2431
2432 /**
2433  * drop duplicate replay request
2434  *
2435  * Because the operation has been replayed by update recovery, the request
2436  * with the same transno will be dropped and also notify the client to send
2437  * next replay request.
2438  *
2439  * \param[in] env       execution environment
2440  * \param[in] obd       failover obd device
2441  * \param[in] req       request to be dropped
2442  */
2443 static void drop_duplicate_replay_req(struct lu_env *env,
2444                                       struct obd_device *obd,
2445                                       struct ptlrpc_request *req)
2446 {
2447         DEBUG_REQ(D_HA, req,
2448                   "remove t%lld from %s because duplicate update records found",
2449                   lustre_msg_get_transno(req->rq_reqmsg),
2450                   libcfs_nid2str(req->rq_peer.nid));
2451
2452         /*
2453          * Right now, only for MDS reint operation update replay and
2454          * normal request replay can have the same transno
2455          */
2456         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2457                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2458                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2459                 if (likely(req->rq_export))
2460                         target_committed_to_req(req);
2461                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2462                 target_send_reply(req, req->rq_status, 0);
2463         } else {
2464                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2465                 libcfs_nid2str(req->rq_peer.nid));
2466         }
2467         target_exp_dequeue_req_replay(req);
2468         target_request_copy_put(req);
2469         obd->obd_replayed_requests++;
2470 }
2471
2472 static void replay_request_or_update(struct lu_env *env,
2473                                      struct lu_target *lut,
2474                                      struct target_recovery_data *trd,
2475                                      struct ptlrpc_thread *thread)
2476 {
2477         struct obd_device *obd = lut->lut_obd;
2478         struct ptlrpc_request *req = NULL;
2479         int type;
2480         __u64 transno;
2481
2482         ENTRY;
2483
2484         CDEBUG(D_HA, "Waiting for transno %lld\n",
2485                obd->obd_next_recovery_transno);
2486
2487         /* Replay all of request and update by transno */
2488         do {
2489                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2490
2491                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2492
2493                 /**
2494                  * It is needed to extend recovery window above
2495                  *  recovery_time_soft. Extending is possible only in the
2496                  *  end of recovery window (see more details in
2497                  *  handle_recovery_req()).
2498                  */
2499                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2500
2501                 if (target_recovery_overseer(lut, check_for_next_transno,
2502                                         exp_req_replay_healthy_or_from_mdt)) {
2503                         abort_req_replay_queue(obd);
2504                         abort_lock_replay_queue(obd);
2505                         goto abort;
2506                 }
2507
2508                 spin_lock(&obd->obd_recovery_task_lock);
2509                 transno = get_next_transno(lut, &type);
2510                 if (type == REQUEST_RECOVERY && transno != 0) {
2511                         /*
2512                          * Drop replay request from client side, if the
2513                          * replay has been executed by update with the
2514                          * same transno
2515                          */
2516                         req = list_entry(obd->obd_req_replay_queue.next,
2517                                         struct ptlrpc_request, rq_list);
2518
2519                         list_del_init(&req->rq_list);
2520                         obd->obd_requests_queued_for_recovery--;
2521                         spin_unlock(&obd->obd_recovery_task_lock);
2522
2523                         /*
2524                          * Let's check if the request has been redone by
2525                          * update replay
2526                          */
2527                         if (is_req_replayed_by_update(req)) {
2528                                 struct distribute_txn_replay_req *dtrq;
2529
2530                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2531                                                                       transno);
2532                                 LASSERT(dtrq != NULL);
2533                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2534                                 list_del_init(&dtrq->dtrq_list);
2535                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2536                                 dtrq_destroy(dtrq);
2537
2538                                 drop_duplicate_replay_req(env, obd, req);
2539
2540                                 continue;
2541                         }
2542
2543                         LASSERT(trd->trd_processing_task == current->pid);
2544                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2545                                   req->rq_xid,
2546                                   lustre_msg_get_transno(req->rq_reqmsg),
2547                                   libcfs_nid2str(req->rq_peer.nid));
2548
2549                         handle_recovery_req(thread, req,
2550                                             trd->trd_recovery_handler);
2551                         /**
2552                          * bz18031: increase next_recovery_transno before
2553                          * target_request_copy_put() will drop exp_rpc reference
2554                          */
2555                         spin_lock(&obd->obd_recovery_task_lock);
2556                         obd->obd_next_recovery_transno++;
2557                         spin_unlock(&obd->obd_recovery_task_lock);
2558                         target_exp_dequeue_req_replay(req);
2559                         target_request_copy_put(req);
2560                         obd->obd_replayed_requests++;
2561                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2562                         struct distribute_txn_replay_req *dtrq;
2563                         int rc;
2564
2565                         spin_unlock(&obd->obd_recovery_task_lock);
2566
2567                         LASSERT(tdtd != NULL);
2568                         dtrq = distribute_txn_get_next_req(tdtd);
2569                         lu_context_enter(&thread->t_env->le_ctx);
2570                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2571                         lu_context_exit(&thread->t_env->le_ctx);
2572                         extend_recovery_timer(obd, obd_timeout, true);
2573
2574                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2575                                 CDEBUG(D_HA,
2576                                        "Move x%llu t%llu to finish list\n",
2577                                        dtrq->dtrq_xid,
2578                                        dtrq->dtrq_master_transno);
2579
2580                                 /* Add it to the replay finish list */
2581                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2582                                 list_add(&dtrq->dtrq_list,
2583                                          &tdtd->tdtd_replay_finish_list);
2584                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2585
2586                                 spin_lock(&obd->obd_recovery_task_lock);
2587                                 if (transno == obd->obd_next_recovery_transno)
2588                                         obd->obd_next_recovery_transno++;
2589                                 else if (transno >
2590                                          obd->obd_next_recovery_transno)
2591                                         obd->obd_next_recovery_transno =
2592                                                                 transno + 1;
2593                                 spin_unlock(&obd->obd_recovery_task_lock);
2594                         } else {
2595                                 dtrq_destroy(dtrq);
2596                         }
2597                 } else {
2598                         spin_unlock(&obd->obd_recovery_task_lock);
2599 abort:
2600                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2601                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2602                         /** evict exports failed VBR */
2603                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2604                         break;
2605                 }
2606         } while (1);
2607 }
2608
2609 static int target_recovery_thread(void *arg)
2610 {
2611         struct lu_target *lut = arg;
2612         struct obd_device *obd = lut->lut_obd;
2613         struct ptlrpc_request *req;
2614         struct target_recovery_data *trd = &obd->obd_recovery_data;
2615         unsigned long delta;
2616         struct lu_env *env;
2617         struct ptlrpc_thread *thread = NULL;
2618         int rc = 0;
2619
2620         ENTRY;
2621
2622         unshare_fs_struct();
2623         OBD_ALLOC_PTR(thread);
2624         if (thread == NULL)
2625                 RETURN(-ENOMEM);
2626
2627         OBD_ALLOC_PTR(env);
2628         if (env == NULL)
2629                 GOTO(out_thread, rc = -ENOMEM);
2630         rc = lu_env_add(env);
2631         if (rc)
2632                 GOTO(out_env, rc);
2633
2634         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2635         if (rc)
2636                 GOTO(out_env_remove, rc);
2637
2638         thread->t_env = env;
2639         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2640         env->le_ctx.lc_thread = thread;
2641         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2642
2643         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2644                current->pid);
2645         trd->trd_processing_task = current->pid;
2646
2647         spin_lock(&obd->obd_dev_lock);
2648         obd->obd_recovering = 1;
2649         spin_unlock(&obd->obd_dev_lock);
2650         complete(&trd->trd_starting);
2651
2652         /* first of all, we have to know the first transno to replay */
2653         if (target_recovery_overseer(lut, check_for_recovery_ready,
2654                                      exp_connect_healthy)) {
2655                 abort_req_replay_queue(obd);
2656                 abort_lock_replay_queue(obd);
2657                 if (lut->lut_tdtd != NULL)
2658                         dtrq_list_destroy(lut->lut_tdtd);
2659         }
2660
2661         /* next stage: replay requests or update */
2662         delta = jiffies;
2663         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2664                atomic_read(&obd->obd_req_replay_clients),
2665                obd->obd_next_recovery_transno);
2666         replay_request_or_update(env, lut, trd, thread);
2667
2668         /**
2669          * The second stage: replay locks
2670          */
2671         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2672                atomic_read(&obd->obd_lock_replay_clients));
2673         while ((req = target_next_replay_lock(lut))) {
2674                 LASSERT(trd->trd_processing_task == current->pid);
2675                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2676                           libcfs_nid2str(req->rq_peer.nid));
2677                 handle_recovery_req(thread, req,
2678                                     trd->trd_recovery_handler);
2679                 target_request_copy_put(req);
2680                 obd->obd_replayed_locks++;
2681         }
2682
2683         /**
2684          * The third stage: reply on final pings, at this moment all clients
2685          * must have request in final queue
2686          */
2687         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2688         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2689         /** Update server last boot epoch */
2690         tgt_boot_epoch_update(lut);
2691         /*
2692          * We drop recoverying flag to forward all new requests
2693          * to regular mds_handle() since now
2694          */
2695         spin_lock(&obd->obd_dev_lock);
2696         obd->obd_recovering = obd->obd_abort_recovery = 0;
2697         spin_unlock(&obd->obd_dev_lock);
2698         spin_lock(&obd->obd_recovery_task_lock);
2699         target_cancel_recovery_timer(obd);
2700         spin_unlock(&obd->obd_recovery_task_lock);
2701         while ((req = target_next_final_ping(obd))) {
2702                 LASSERT(trd->trd_processing_task == current->pid);
2703                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2704                           libcfs_nid2str(req->rq_peer.nid));
2705                 handle_recovery_req(thread, req,
2706                                     trd->trd_recovery_handler);
2707                 /*
2708                  * Because the waiting client can not send ping to server,
2709                  * so we need refresh the last_request_time, to avoid the
2710                  * export is being evicted
2711                  */
2712                 ptlrpc_update_export_timer(req->rq_export, 0);
2713                 target_request_copy_put(req);
2714         }
2715
2716         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2717         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2718                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2719         if (delta > OBD_RECOVERY_TIME_SOFT) {
2720                 CWARN("too long recovery - read logs\n");
2721                 libcfs_debug_dumplog();
2722         }
2723
2724         target_finish_recovery(lut);
2725         lu_context_fini(&env->le_ctx);
2726         trd->trd_processing_task = 0;
2727         complete_all(&trd->trd_finishing);
2728         tgt_io_thread_done(thread);
2729 out_env_remove:
2730         lu_env_remove(env);
2731 out_env:
2732         OBD_FREE_PTR(env);
2733 out_thread:
2734         OBD_FREE_PTR(thread);
2735         RETURN(rc);
2736 }
2737
2738 static int target_start_recovery_thread(struct lu_target *lut,
2739                                         svc_handler_t handler)
2740 {
2741         struct obd_device *obd = lut->lut_obd;
2742         int rc = 0;
2743         struct target_recovery_data *trd = &obd->obd_recovery_data;
2744         int index;
2745
2746         memset(trd, 0, sizeof(*trd));
2747         init_completion(&trd->trd_starting);
2748         init_completion(&trd->trd_finishing);
2749         trd->trd_recovery_handler = handler;
2750
2751         rc = server_name2index(obd->obd_name, &index, NULL);
2752         if (rc < 0)
2753                 return rc;
2754
2755         if (!IS_ERR(kthread_run(target_recovery_thread,
2756                                 lut, "tgt_recover_%d", index))) {
2757                 wait_for_completion(&trd->trd_starting);
2758                 LASSERT(obd->obd_recovering != 0);
2759         } else {
2760                 rc = -ECHILD;
2761         }
2762
2763         return rc;
2764 }
2765
2766 void target_stop_recovery_thread(struct obd_device *obd)
2767 {
2768         if (obd->obd_recovery_data.trd_processing_task > 0) {
2769                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2770                 /** recovery can be done but postrecovery is not yet */
2771                 spin_lock(&obd->obd_dev_lock);
2772                 if (obd->obd_recovering) {
2773                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2774                         obd->obd_abort_recovery = 1;
2775                         wake_up(&obd->obd_next_transno_waitq);
2776                 }
2777                 spin_unlock(&obd->obd_dev_lock);
2778                 wait_for_completion(&trd->trd_finishing);
2779         }
2780 }
2781 EXPORT_SYMBOL(target_stop_recovery_thread);
2782
2783 void target_recovery_fini(struct obd_device *obd)
2784 {
2785         class_disconnect_exports(obd);
2786         target_stop_recovery_thread(obd);
2787         target_cleanup_recovery(obd);
2788 }
2789 EXPORT_SYMBOL(target_recovery_fini);
2790
2791 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2792 {
2793         struct obd_device *obd = container_of(timer, struct obd_device,
2794                                               obd_recovery_timer);
2795
2796         CDEBUG(D_HA,
2797                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2798                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2799                ktime_get_seconds() - obd->obd_recovery_start,
2800                atomic_read(&obd->obd_connected_clients));
2801
2802         obd->obd_recovery_expired = 1;
2803         wake_up(&obd->obd_next_transno_waitq);
2804         return HRTIMER_NORESTART;
2805 }
2806
2807 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2808 {
2809         struct obd_device *obd = lut->lut_obd;
2810
2811         if (lut->lut_bottom->dd_rdonly)
2812                 return;
2813
2814         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2815                 /** Update server last boot epoch */
2816                 tgt_boot_epoch_update(lut);
2817                 return;
2818         }
2819
2820         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2821                "last_transno %llu\n", obd->obd_name,
2822                atomic_read(&obd->obd_max_recoverable_clients),
2823                obd->obd_last_committed);
2824         LASSERT(obd->obd_stopping == 0);
2825         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2826         obd->obd_recovery_start = 0;
2827         obd->obd_recovery_end = 0;
2828
2829         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2830                      HRTIMER_MODE_ABS);
2831         obd->obd_recovery_timer.function = &target_recovery_expired;
2832         target_start_recovery_thread(lut, handler);
2833 }
2834 EXPORT_SYMBOL(target_recovery_init);
2835
2836 static int target_process_req_flags(struct obd_device *obd,
2837                                     struct ptlrpc_request *req)
2838 {
2839         struct obd_export *exp = req->rq_export;
2840
2841         LASSERT(exp != NULL);
2842         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2843                 /* client declares he's ready to replay locks */
2844                 spin_lock(&exp->exp_lock);
2845                 if (exp->exp_req_replay_needed) {
2846                         exp->exp_req_replay_needed = 0;
2847                         spin_unlock(&exp->exp_lock);
2848
2849                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2850                         atomic_dec(&obd->obd_req_replay_clients);
2851                 } else {
2852                         spin_unlock(&exp->exp_lock);
2853                 }
2854         }
2855         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2856                 /*
2857                  * client declares he's ready to complete recovery
2858                  * so, we put the request on th final queue
2859                  */
2860                 spin_lock(&exp->exp_lock);
2861                 if (exp->exp_lock_replay_needed) {
2862                         exp->exp_lock_replay_needed = 0;
2863                         spin_unlock(&exp->exp_lock);
2864
2865                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2866                         atomic_dec(&obd->obd_lock_replay_clients);
2867                 } else {
2868                         spin_unlock(&exp->exp_lock);
2869                 }
2870         }
2871         return 0;
2872 }
2873
2874 int target_queue_recovery_request(struct ptlrpc_request *req,
2875                                   struct obd_device *obd)
2876 {
2877         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2878         struct ptlrpc_request *reqiter;
2879         int inserted = 0;
2880
2881         ENTRY;
2882
2883         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2884                 /* Processing the queue right now, don't re-add. */
2885                 RETURN(1);
2886         }
2887
2888         target_process_req_flags(obd, req);
2889
2890         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2891                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2892                         if (cfs_fail_val == 1) {
2893                                 cfs_race_state = 1;
2894                                 cfs_fail_val = 0;
2895                                 wake_up(&cfs_race_waitq);
2896
2897                                 schedule_timeout_interruptible(
2898                                         cfs_time_seconds(1));
2899                         }
2900                 }
2901
2902                 /*
2903                  * client declares he's ready to complete recovery
2904                  * so, we put the request on th final queue
2905                  */
2906                 target_request_copy_get(req);
2907                 DEBUG_REQ(D_HA, req, "queue final req");
2908                 wake_up(&obd->obd_next_transno_waitq);
2909                 spin_lock(&obd->obd_recovery_task_lock);
2910                 if (obd->obd_recovering) {
2911                         struct ptlrpc_request *tmp;
2912                         struct ptlrpc_request *duplicate = NULL;
2913
2914                         if (likely(!req->rq_export->exp_replay_done)) {
2915                                 req->rq_export->exp_replay_done = 1;
2916                                 list_add_tail(&req->rq_list,
2917                                               &obd->obd_final_req_queue);
2918                                 spin_unlock(&obd->obd_recovery_task_lock);
2919                                 RETURN(0);
2920                         }
2921
2922                         /*
2923                          * XXX O(n), but only happens if final ping is
2924                          * timed out, probably reorganize the list as
2925                          * a hash list later
2926                          */
2927                         list_for_each_entry_safe(reqiter, tmp,
2928                                                  &obd->obd_final_req_queue,
2929                                                  rq_list) {
2930                                 if (reqiter->rq_export == req->rq_export) {
2931                                         list_del_init(&reqiter->rq_list);
2932                                         duplicate = reqiter;
2933                                         break;
2934                                 }
2935                         }
2936
2937                         list_add_tail(&req->rq_list,
2938                                       &obd->obd_final_req_queue);
2939                         req->rq_export->exp_replay_done = 1;
2940                         spin_unlock(&obd->obd_recovery_task_lock);
2941
2942                         if (duplicate != NULL) {
2943                                 DEBUG_REQ(D_HA, duplicate,
2944                                           "put prev final req");
2945                                 target_request_copy_put(duplicate);
2946                         }
2947                         RETURN(0);
2948                 } else {
2949                         spin_unlock(&obd->obd_recovery_task_lock);
2950                         target_request_copy_put(req);
2951                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2952                 }
2953         }
2954         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2955                 /* client declares he's ready to replay locks */
2956                 target_request_copy_get(req);
2957                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2958                 wake_up(&obd->obd_next_transno_waitq);
2959                 spin_lock(&obd->obd_recovery_task_lock);
2960                 LASSERT(obd->obd_recovering);
2961                 /* usually due to recovery abort */
2962                 if (!req->rq_export->exp_in_recovery) {
2963                         spin_unlock(&obd->obd_recovery_task_lock);
2964                         target_request_copy_put(req);
2965                         RETURN(-ENOTCONN);
2966                 }
2967                 LASSERT(req->rq_export->exp_lock_replay_needed);
2968                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2969                 spin_unlock(&obd->obd_recovery_task_lock);
2970                 RETURN(0);
2971         }
2972
2973         /*
2974          * CAVEAT EMPTOR: The incoming request message has been swabbed
2975          * (i.e. buflens etc are in my own byte order), but type-dependent
2976          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2977          */
2978
2979         if (!transno) {
2980                 INIT_LIST_HEAD(&req->rq_list);
2981                 DEBUG_REQ(D_HA, req, "not queueing");
2982                 RETURN(1);
2983         }
2984
2985         /*
2986          * If we're processing the queue, we want don't want to queue this
2987          * message.
2988          *
2989          * Also, if this request has a transno less than the one we're waiting
2990          * for, we should process it now.  It could (and currently always will)
2991          * be an open request for a descriptor that was opened some time ago.
2992          *
2993          * Also, a resent, replayed request that has already been
2994          * handled will pass through here and be processed immediately.
2995          */
2996         CDEBUG(D_HA,
2997                "Next recovery transno: %llu, current: %llu, replaying\n",
2998                obd->obd_next_recovery_transno, transno);
2999
3000         /*
3001          * If the request has been replayed by update replay, then sends this
3002          * request to the recovery thread (replay_request_or_update()), where
3003          * it will be handled
3004          */
3005         spin_lock(&obd->obd_recovery_task_lock);
3006         if (transno < obd->obd_next_recovery_transno &&
3007             !is_req_replayed_by_update(req)) {
3008                 /* Processing the queue right now, don't re-add. */
3009                 LASSERT(list_empty(&req->rq_list));
3010                 spin_unlock(&obd->obd_recovery_task_lock);
3011                 RETURN(1);
3012         }
3013         spin_unlock(&obd->obd_recovery_task_lock);
3014
3015         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3016                 RETURN(0);
3017
3018         target_request_copy_get(req);
3019         if (!req->rq_export->exp_in_recovery) {
3020                 target_request_copy_put(req);
3021                 RETURN(-ENOTCONN);
3022         }
3023         LASSERT(req->rq_export->exp_req_replay_needed);
3024
3025         if (target_exp_enqueue_req_replay(req)) {
3026                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3027                 target_request_copy_put(req);
3028                 RETURN(0);
3029         }
3030
3031         /* XXX O(n^2) */
3032         spin_lock(&obd->obd_recovery_task_lock);
3033         LASSERT(obd->obd_recovering);
3034         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3035                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3036                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3037                         inserted = 1;
3038                         goto added;
3039                 }
3040
3041                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3042                              transno)) {
3043                         DEBUG_REQ(D_ERROR, req,
3044                                   "dropping replay: transno has been claimed by another client");
3045                         spin_unlock(&obd->obd_recovery_task_lock);
3046                         target_exp_dequeue_req_replay(req);
3047                         target_request_copy_put(req);
3048                         RETURN(0);
3049                 }
3050         }
3051 added:
3052         if (!inserted)
3053                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3054
3055         obd->obd_requests_queued_for_recovery++;
3056         spin_unlock(&obd->obd_recovery_task_lock);
3057         wake_up(&obd->obd_next_transno_waitq);
3058         RETURN(0);
3059 }
3060
3061 void target_committed_to_req(struct ptlrpc_request *req)
3062 {
3063         struct obd_export *exp = req->rq_export;
3064
3065         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3066                 lustre_msg_set_last_committed(req->rq_repmsg,
3067                                               exp->exp_last_committed);
3068         else
3069                 DEBUG_REQ(D_IOCTL, req,
3070                           "not sending last_committed update (%d/%d)",
3071                           exp->exp_obd->obd_no_transno,
3072                           req->rq_repmsg == NULL);
3073
3074         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3075                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3076 }
3077
3078 #endif /* HAVE_SERVER_SUPPORT */
3079
3080 /**
3081  * Packs current SLV and Limit into \a req.
3082  */
3083 int target_pack_pool_reply(struct ptlrpc_request *req)
3084 {
3085         struct obd_device *obd;
3086
3087         ENTRY;
3088
3089         /*
3090          * Check that we still have all structures alive as this may
3091          * be some late RPC at shutdown time.
3092          */
3093         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3094                      !exp_connect_lru_resize(req->rq_export))) {
3095                 lustre_msg_set_slv(req->rq_repmsg, 0);
3096                 lustre_msg_set_limit(req->rq_repmsg, 0);
3097                 RETURN(0);
3098         }
3099
3100         /* OBD is alive here as export is alive, which we checked above. */
3101         obd = req->rq_export->exp_obd;
3102
3103         read_lock(&obd->obd_pool_lock);
3104         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3105         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3106         read_unlock(&obd->obd_pool_lock);
3107
3108         RETURN(0);
3109 }
3110
3111 static int target_send_reply_msg(struct ptlrpc_request *req,
3112                                  int rc, int fail_id)
3113 {
3114         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3115                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3116                 return -ECOMM;
3117         }
3118         /*
3119          * We can have a null rq_reqmsg in the event of bad signature or
3120          * no context when unwrapping
3121          */
3122         if (req->rq_reqmsg &&
3123             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3124             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3125                 return -ECOMM;
3126
3127         if (unlikely(rc)) {
3128                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3129                 req->rq_status = rc;
3130                 return ptlrpc_send_error(req, 1);
3131         }
3132         DEBUG_REQ(D_NET, req, "sending reply");
3133
3134         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3135 }
3136
3137 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3138 {
3139         struct ptlrpc_service_part *svcpt;
3140         int netrc;
3141         struct ptlrpc_reply_state *rs;
3142         struct obd_export *exp;
3143
3144         ENTRY;
3145
3146         if (req->rq_no_reply) {
3147                 EXIT;
3148                 return;
3149         }
3150
3151         svcpt = req->rq_rqbd->rqbd_svcpt;
3152         rs = req->rq_reply_state;
3153         if (rs == NULL || !rs->rs_difficult) {
3154                 /* no notifiers */
3155                 target_send_reply_msg(req, rc, fail_id);
3156                 EXIT;
3157                 return;
3158         }
3159
3160         /* must be an export if locks saved */
3161         LASSERT(req->rq_export != NULL);
3162         /* req/reply consistent */
3163         LASSERT(rs->rs_svcpt == svcpt);
3164
3165         /* "fresh" reply */
3166         LASSERT(!rs->rs_scheduled);
3167         LASSERT(!rs->rs_scheduled_ever);
3168         LASSERT(!rs->rs_handled);
3169         LASSERT(!rs->rs_on_net);
3170         LASSERT(rs->rs_export == NULL);
3171         LASSERT(list_empty(&rs->rs_obd_list));
3172         LASSERT(list_empty(&rs->rs_exp_list));
3173
3174         exp = class_export_get(req->rq_export);
3175
3176         /* disable reply scheduling while I'm setting up */
3177         rs->rs_scheduled = 1;
3178         rs->rs_on_net    = 1;
3179         rs->rs_xid       = req->rq_xid;
3180         rs->rs_transno   = req->rq_transno;
3181         rs->rs_export    = exp;
3182         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3183
3184         spin_lock(&exp->exp_uncommitted_replies_lock);
3185         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3186                rs->rs_transno, exp->exp_last_committed);
3187         if (rs->rs_transno > exp->exp_last_committed) {
3188                 /* not committed already */
3189                 list_add_tail(&rs->rs_obd_list,
3190                                   &exp->exp_uncommitted_replies);
3191         }
3192         spin_unlock(&exp->exp_uncommitted_replies_lock);
3193
3194         spin_lock(&exp->exp_lock);
3195         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3196         spin_unlock(&exp->exp_lock);
3197
3198         netrc = target_send_reply_msg(req, rc, fail_id);
3199
3200         spin_lock(&svcpt->scp_rep_lock);
3201
3202         atomic_inc(&svcpt->scp_nreps_difficult);
3203
3204         if (netrc != 0) {
3205                 /*
3206                  * error sending: reply is off the net.  Also we need +1
3207                  * reply ref until ptlrpc_handle_rs() is done
3208                  * with the reply state (if the send was successful, there
3209                  * would have been +1 ref for the net, which
3210                  * reply_out_callback leaves alone)
3211                  */
3212                 rs->rs_on_net = 0;
3213                 ptlrpc_rs_addref(rs);
3214         }
3215
3216         spin_lock(&rs->rs_lock);
3217         if (rs->rs_transno <= exp->exp_last_committed ||
3218             (!rs->rs_on_net && !rs->rs_no_ack) ||
3219             list_empty(&rs->rs_exp_list) ||     /* completed already */
3220             list_empty(&rs->rs_obd_list)) {
3221                 CDEBUG(D_HA, "Schedule reply immediately\n");
3222                 ptlrpc_dispatch_difficult_reply(rs);
3223         } else {
3224                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3225                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3226         }
3227         spin_unlock(&rs->rs_lock);
3228         spin_unlock(&svcpt->scp_rep_lock);
3229         EXIT;
3230 }
3231
3232 enum ldlm_mode lck_compat_array[] = {
3233         [LCK_EX]    = LCK_COMPAT_EX,
3234         [LCK_PW]    = LCK_COMPAT_PW,
3235         [LCK_PR]    = LCK_COMPAT_PR,
3236         [LCK_CW]    = LCK_COMPAT_CW,
3237         [LCK_CR]    = LCK_COMPAT_CR,
3238         [LCK_NL]    = LCK_COMPAT_NL,
3239         [LCK_GROUP] = LCK_COMPAT_GROUP,
3240         [LCK_COS]   = LCK_COMPAT_COS,
3241 };
3242
3243 /**
3244  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3245  * not escape to the user level.
3246  */
3247 int ldlm_error2errno(enum ldlm_error error)
3248 {
3249         int result;
3250
3251         switch (error) {
3252         case ELDLM_OK:
3253         case ELDLM_LOCK_MATCHED:
3254                 result = 0;
3255                 break;
3256         case ELDLM_LOCK_CHANGED:
3257                 result = -ESTALE;
3258                 break;
3259         case ELDLM_LOCK_ABORTED:
3260                 result = -ENAVAIL;
3261                 break;
3262         case ELDLM_LOCK_REPLACED:
3263                 result = -ESRCH;
3264                 break;
3265         case ELDLM_NO_LOCK_DATA:
3266                 result = -ENOENT;
3267                 break;
3268         case ELDLM_NAMESPACE_EXISTS:
3269                 result = -EEXIST;
3270                 break;
3271         case ELDLM_BAD_NAMESPACE:
3272                 result = -EBADF;
3273                 break;
3274         default:
3275                 if (((int)error) < 0) { /* cast to signed type */
3276                         result = error; /* as ldlm_error can be unsigned */
3277                 } else {
3278                         CERROR("Invalid DLM result code: %d\n", error);
3279                         result = -EPROTO;
3280                 }
3281         }
3282         return result;
3283 }
3284 EXPORT_SYMBOL(ldlm_error2errno);
3285
3286 /**
3287  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3288  */
3289 enum ldlm_error ldlm_errno2error(int err_no)
3290 {
3291         int error;
3292
3293         switch (err_no) {
3294         case 0:
3295                 error = ELDLM_OK;
3296                 break;
3297         case -ESTALE:
3298                 error = ELDLM_LOCK_CHANGED;
3299                 break;
3300         case -ENAVAIL:
3301                 error = ELDLM_LOCK_ABORTED;
3302                 break;
3303         case -ESRCH:
3304                 error = ELDLM_LOCK_REPLACED;
3305                 break;
3306         case -ENOENT:
3307                 error = ELDLM_NO_LOCK_DATA;
3308                 break;
3309         case -EEXIST:
3310                 error = ELDLM_NAMESPACE_EXISTS;
3311                 break;
3312         case -EBADF:
3313                 error = ELDLM_BAD_NAMESPACE;
3314                 break;
3315         default:
3316                 error = err_no;
3317         }
3318         return error;
3319 }
3320
3321 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3322 void ldlm_dump_export_locks(struct obd_export *exp)
3323 {
3324         spin_lock(&exp->exp_locks_list_guard);
3325         if (!list_empty(&exp->exp_locks_list)) {
3326                 struct ldlm_lock *lock;
3327
3328                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3329                        exp);
3330                 list_for_each_entry(lock, &exp->exp_locks_list,
3331                                         l_exp_refs_link)
3332                         LDLM_ERROR(lock, "lock:");
3333         }
3334         spin_unlock(&exp->exp_locks_list_guard);
3335 }
3336 #endif
3337
3338 #ifdef HAVE_SERVER_SUPPORT
3339 static inline const char *bulk2type(struct ptlrpc_request *req)
3340 {
3341         if (req->rq_bulk_read)
3342                 return "READ";
3343         if (req->rq_bulk_write)
3344                 return "WRITE";
3345         return "UNKNOWN";
3346 }
3347
3348 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3349 {
3350         struct ptlrpc_request *req = desc->bd_req;
3351         time64_t start = ktime_get_seconds();
3352         time64_t deadline;
3353         int rc = 0;
3354
3355         ENTRY;
3356
3357         /* If there is eviction in progress, wait for it to finish. */
3358         wait_event_idle(
3359                 exp->exp_obd->obd_evict_inprogress_waitq,
3360                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3361
3362         /* Check if client was evicted or reconnected already. */
3363         if (exp->exp_failed ||
3364             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3365                 rc = -ENOTCONN;
3366         } else {
3367                 if (req->rq_bulk_read)
3368                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3369
3370                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3371                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3372                 else /* old version, bulk matchbits is rq_xid */
3373                         req->rq_mbits = req->rq_xid;
3374
3375                 if (rc == 0)
3376                         rc = ptlrpc_start_bulk_transfer(desc);
3377         }
3378
3379         if (rc < 0) {
3380                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3381                           bulk2type(req), rc);
3382                 RETURN(rc);
3383         }
3384
3385         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3386                 ptlrpc_abort_bulk(desc);
3387                 RETURN(0);
3388         }
3389
3390         /* limit actual bulk transfer to bulk_timeout seconds */
3391         deadline = start + bulk_timeout;
3392         if (deadline > req->rq_deadline)
3393                 deadline = req->rq_deadline;
3394
3395         do {
3396                 time64_t timeoutl = deadline - ktime_get_seconds();
3397                 time64_t rq_deadline;
3398
3399                 while (timeoutl >= 0 &&
3400                        wait_event_idle_timeout(
3401                                desc->bd_waitq,
3402                                !ptlrpc_server_bulk_active(desc) ||
3403                                exp->exp_failed ||
3404                                exp->exp_conn_cnt >
3405                                lustre_msg_get_conn_cnt(req->rq_reqmsg),
3406                                timeoutl ? cfs_time_seconds(1) : 1) == 0)
3407                         timeoutl -= 1;
3408                 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3409
3410                 /* Wait again if we changed rq_deadline. */
3411                 rq_deadline = READ_ONCE(req->rq_deadline);
3412                 deadline = start + bulk_timeout;
3413                 if (deadline > rq_deadline)
3414                         deadline = rq_deadline;
3415         } while (rc == -ETIMEDOUT &&
3416                  deadline > ktime_get_seconds());
3417
3418         if (rc == -ETIMEDOUT) {
3419                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3420                           bulk2type(req), deadline - start,
3421                           ktime_get_real_seconds() - deadline);
3422                 ptlrpc_abort_bulk(desc);
3423         } else if (exp->exp_failed) {
3424                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3425                           bulk2type(req));
3426                 rc = -ENOTCONN;
3427                 ptlrpc_abort_bulk(desc);
3428         } else if (exp->exp_conn_cnt >
3429                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3430                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3431                           bulk2type(req));
3432                 /* We don't reply anyway. */
3433                 rc = -ETIMEDOUT;
3434                 ptlrpc_abort_bulk(desc);
3435         } else if (desc->bd_failure) {
3436                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3437                           bulk2type(req));
3438                 /* XXX should this be a different errno? */
3439                 rc = -ETIMEDOUT;
3440         } else {
3441                 if (req->rq_bulk_write)
3442                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3443                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3444                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3445                                   bulk2type(req), desc->bd_nob_transferred,
3446                                   desc->bd_nob);
3447                         /* XXX should this be a different errno? */
3448                         rc = -ETIMEDOUT;
3449                 }
3450         }
3451
3452         RETURN(rc);
3453 }
3454 EXPORT_SYMBOL(target_bulk_io);
3455
3456 #endif /* HAVE_SERVER_SUPPORT */