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  */
  31
  32 /**
  33  * This file deals with various client/target related logic including recovery.
  34  *
  35  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  36  * should be moved.
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_LDLM
  40
  41 #include <cl_object.h>
  42 #include <linux/jiffies.h>
  43 #include <linux/kernel.h>
  44 #include <linux/kthread.h>
  45 #include <libcfs/libcfs.h>
  46 #include <obd.h>
  47 #include <obd_class.h>
  48 #include <lustre_dlm.h>
  49 #include <lustre_net.h>
  50 #include <lustre_sec.h>
  51 #include "ldlm_internal.h"
  52
  53 /*
  54  * @priority: If non-zero, move the selected connection to the list head.
  55  * @create: If zero, only search in existing connections.
  56  */
  57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  58                            int priority, int create)
  59 {
  60         struct ptlrpc_connection *ptlrpc_conn;
  61         struct obd_import_conn *imp_conn = NULL, *item;
  62         lnet_nid_t nid4refnet = LNET_NID_ANY;
  63         int rc = 0;
  64
  65         ENTRY;
  66
  67         if (!create && !priority) {
  68                 CDEBUG(D_HA, "Nothing to do\n");
  69                 RETURN(-EINVAL);
  70         }
  71
  72         if (imp->imp_connection &&
  73             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  74                 /* nid4refnet is used to restrict network connections */
  75                 nid4refnet = lnet_nid_to_nid4(&imp->imp_connection->c_self);
  76         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
  77         if (!ptlrpc_conn) {
  78                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  79                 RETURN(-ENOENT);
  80         }
  81
  82         if (create) {
  83                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  84                 if (!imp_conn)
  85                         GOTO(out_put, rc = -ENOMEM);
  86         }
  87
  88         spin_lock(&imp->imp_lock);
  89         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  90                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  91                         if (priority) {
  92                                 list_move(&item->oic_item,
  93                                           &imp->imp_conn_list);
  94                                 item->oic_last_attempt = 0;
  95                         }
  96                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  97                                imp, imp->imp_obd->obd_name, uuid->uuid,
  98                                (priority ? ", moved to head" : ""));
  99                         spin_unlock(&imp->imp_lock);
 100                         GOTO(out_free, rc = 0);
 101                 }
 102         }
 103         /* No existing import connection found for \a uuid. */
 104         if (create) {
 105                 imp_conn->oic_conn = ptlrpc_conn;
 106                 imp_conn->oic_uuid = *uuid;
 107                 imp_conn->oic_last_attempt = 0;
 108                 if (priority)
 109                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 110                 else
 111                         list_add_tail(&imp_conn->oic_item,
 112                                       &imp->imp_conn_list);
 113                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 114                        imp, imp->imp_obd->obd_name, uuid->uuid,
 115                        (priority ? "head" : "tail"));
 116         } else {
 117                 spin_unlock(&imp->imp_lock);
 118                 GOTO(out_free, rc = -ENOENT);
 119         }
 120
 121         spin_unlock(&imp->imp_lock);
 122         RETURN(0);
 123 out_free:
 124         if (imp_conn)
 125                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 126 out_put:
 127         ptlrpc_connection_put(ptlrpc_conn);
 128         RETURN(rc);
 129 }
 130
 131 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 132 {
 133         return import_set_conn(imp, uuid, 1, 0);
 134 }
 135
 136 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 137                            int priority)
 138 {
 139         return import_set_conn(imp, uuid, priority, 1);
 140 }
 141 EXPORT_SYMBOL(client_import_add_conn);
 142
 143 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 144                                lnet_nid_t prim_nid, int priority)
 145 {
 146         struct ptlrpc_connection *ptlrpc_conn;
 147         int rc;
 148
 149         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, prim_nid);
 150         if (!ptlrpc_conn) {
 151                 const char *str_uuid = obd_uuid2str(uuid);
 152
 153                 rc = class_add_uuid(str_uuid, prim_nid);
 154                 if (rc) {
 155                         CERROR("%s: failed to add UUID '%s': rc = %d\n",
 156                                imp->imp_obd->obd_name, str_uuid, rc);
 157                         return rc;
 158                 }
 159         }
 160         return import_set_conn(imp, uuid, priority, 1);
 161 }
 162 EXPORT_SYMBOL(client_import_dyn_add_conn);
 163
 164 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
 165                                    int nid_count, struct obd_uuid *uuid)
 166 {
 167         struct obd_import_conn *conn;
 168         int rc = -ENOENT;
 169
 170         ENTRY;
 171         if (nid_count <= 0 || !nids)
 172                 return rc;
 173
 174         spin_lock(&imp->imp_lock);
 175         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 176                 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
 177                         *uuid = conn->oic_uuid;
 178                         spin_unlock(&imp->imp_lock);
 179                         rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
 180                                                     nid_count);
 181                         RETURN(rc);
 182                 }
 183         }
 184         spin_unlock(&imp->imp_lock);
 185         RETURN(rc);
 186 }
 187 EXPORT_SYMBOL(client_import_add_nids_to_conn);
 188
 189 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 190 {
 191         struct obd_import_conn *imp_conn;
 192         struct obd_export *dlmexp;
 193         int rc = -ENOENT;
 194
 195         ENTRY;
 196
 197         spin_lock(&imp->imp_lock);
 198         if (list_empty(&imp->imp_conn_list)) {
 199                 LASSERT(!imp->imp_connection);
 200                 GOTO(out, rc);
 201         }
 202
 203         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 204                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 205                         continue;
 206                 LASSERT(imp_conn->oic_conn);
 207
 208                 if (imp_conn == imp->imp_conn_current) {
 209                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 210
 211                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 212                             imp->imp_state != LUSTRE_IMP_DISCON) {
 213                                 CERROR("can't remove current connection\n");
 214                                 GOTO(out, rc = -EBUSY);
 215                         }
 216
 217                         ptlrpc_connection_put(imp->imp_connection);
 218                         imp->imp_connection = NULL;
 219
 220                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 221                         if (dlmexp && dlmexp->exp_connection) {
 222                                 LASSERT(dlmexp->exp_connection ==
 223                                         imp_conn->oic_conn);
 224                                 ptlrpc_connection_put(dlmexp->exp_connection);
 225                                 dlmexp->exp_connection = NULL;
 226                         }
 227
 228                         if (dlmexp != NULL)
 229                                 class_export_put(dlmexp);
 230                 }
 231
 232                 list_del(&imp_conn->oic_item);
 233                 ptlrpc_connection_put(imp_conn->oic_conn);
 234                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 235                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 236                        imp, imp->imp_obd->obd_name, uuid->uuid);
 237                 rc = 0;
 238                 break;
 239         }
 240 out:
 241         spin_unlock(&imp->imp_lock);
 242         if (rc == -ENOENT)
 243                 CERROR("connection %s not found\n", uuid->uuid);
 244         RETURN(rc);
 245 }
 246 EXPORT_SYMBOL(client_import_del_conn);
 247
 248 /**
 249  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 250  * to find a conn uuid of \a imp which can reach \a peer.
 251  */
 252 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 253                             struct obd_uuid *uuid)
 254 {
 255         struct obd_import_conn *conn;
 256         int rc = -ENOENT;
 257
 258         ENTRY;
 259
 260         spin_lock(&imp->imp_lock);
 261         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 262                 /* Check if conn UUID does have this peer NID. */
 263                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 264                         *uuid = conn->oic_uuid;
 265                         rc = 0;
 266                         break;
 267                 }
 268         }
 269         spin_unlock(&imp->imp_lock);
 270         RETURN(rc);
 271 }
 272 EXPORT_SYMBOL(client_import_find_conn);
 273
 274 void client_destroy_import(struct obd_import *imp)
 275 {
 276         /*
 277          * Drop security policy instance after all RPCs have finished/aborted
 278          * to let all busy contexts be released.
 279          */
 280         class_import_get(imp);
 281         class_destroy_import(imp);
 282         sptlrpc_import_sec_put(imp);
 283         class_import_put(imp);
 284 }
 285 EXPORT_SYMBOL(client_destroy_import);
 286
 287 /**
 288  * Check whether or not the OSC is on MDT.
 289  * In the config log,
 290  * osc on MDT
 291  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 292  * osc on client
 293  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 294  *
 295  **/
 296 static int osc_on_mdt(char *obdname)
 297 {
 298         char *ptr;
 299
 300         ptr = strrchr(obdname, '-');
 301         if (ptr == NULL)
 302                 return 0;
 303
 304         if (strncmp(ptr + 1, "MDT", 3) == 0)
 305                 return 1;
 306
 307         return 0;
 308 }
 309
 310 /*
 311  * Configure an RPC client OBD device.
 312  *
 313  * lcfg parameters:
 314  * 1 - client UUID
 315  * 2 - server UUID
 316  * 3 - inactive-on-startup
 317  * 4 - restrictive net
 318  */
 319 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 320 {
 321         struct client_obd *cli = &obd->u.cli;
 322         struct obd_import *imp;
 323         struct obd_uuid server_uuid;
 324         int rq_portal, rp_portal, connect_op;
 325         const char *name = obd->obd_type->typ_name;
 326         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 327         char *cli_name = lustre_cfg_buf(lcfg, 0);
 328         struct ptlrpc_connection fake_conn = { .c_self = {},
 329                                                .c_remote_uuid.uuid[0] = 0 };
 330         int rc;
 331
 332         ENTRY;
 333
 334         /*
 335          * In a more perfect world, we would hang a ptlrpc_client off of
 336          * obd_type and just use the values from there.
 337          */
 338         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 339                 rq_portal = OST_REQUEST_PORTAL;
 340                 rp_portal = OSC_REPLY_PORTAL;
 341                 connect_op = OST_CONNECT;
 342                 cli->cl_sp_me = LUSTRE_SP_CLI;
 343                 cli->cl_sp_to = LUSTRE_SP_OST;
 344                 ns_type = LDLM_NS_TYPE_OSC;
 345         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 346                    !strcmp(name, LUSTRE_LWP_NAME)) {
 347                 rq_portal = MDS_REQUEST_PORTAL;
 348                 rp_portal = MDC_REPLY_PORTAL;
 349                 connect_op = MDS_CONNECT;
 350                 if (is_lwp_on_ost(cli_name))
 351                         cli->cl_sp_me = LUSTRE_SP_OST;
 352                 else if (is_lwp_on_mdt(cli_name))
 353                         cli->cl_sp_me = LUSTRE_SP_MDT;
 354                 else
 355                         cli->cl_sp_me = LUSTRE_SP_CLI;
 356                 cli->cl_sp_to = LUSTRE_SP_MDT;
 357                 ns_type = LDLM_NS_TYPE_MDC;
 358         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 359                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 360                         /* OSP_on_MDT for other MDTs */
 361                         connect_op = MDS_CONNECT;
 362                         cli->cl_sp_to = LUSTRE_SP_MDT;
 363                         ns_type = LDLM_NS_TYPE_MDC;
 364                         rq_portal = OUT_PORTAL;
 365                 } else {
 366                         /* OSP on MDT for OST */
 367                         connect_op = OST_CONNECT;
 368                         cli->cl_sp_to = LUSTRE_SP_OST;
 369                         ns_type = LDLM_NS_TYPE_OSC;
 370                         rq_portal = OST_REQUEST_PORTAL;
 371                 }
 372                 rp_portal = OSC_REPLY_PORTAL;
 373                 cli->cl_sp_me = LUSTRE_SP_MDT;
 374         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 375                 rq_portal = MGS_REQUEST_PORTAL;
 376                 rp_portal = MGC_REPLY_PORTAL;
 377                 connect_op = MGS_CONNECT;
 378                 cli->cl_sp_me = LUSTRE_SP_MGC;
 379                 cli->cl_sp_to = LUSTRE_SP_MGS;
 380                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 381                 ns_type = LDLM_NS_TYPE_MGC;
 382         } else {
 383                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 384                        name);
 385                 RETURN(-EINVAL);
 386         }
 387
 388         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 389                 CERROR("requires a TARGET UUID\n");
 390                 RETURN(-EINVAL);
 391         }
 392
 393         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 394                 CERROR("client UUID must be less than 38 characters\n");
 395                 RETURN(-EINVAL);
 396         }
 397
 398         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 399                 CERROR("setup requires a SERVER UUID\n");
 400                 RETURN(-EINVAL);
 401         }
 402
 403         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 404                 CERROR("target UUID must be less than 38 characters\n");
 405                 RETURN(-EINVAL);
 406         }
 407
 408         init_rwsem(&cli->cl_sem);
 409         mutex_init(&cli->cl_mgc_mutex);
 410         cli->cl_seq = NULL;
 411         init_rwsem(&cli->cl_seq_rwsem);
 412         cli->cl_conn_count = 0;
 413         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 414                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 415                      sizeof(server_uuid)));
 416
 417         cli->cl_dirty_pages = 0;
 418         cli->cl_dirty_max_pages = 0;
 419         cli->cl_avail_grant = 0;
 420         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 421         /*
 422          * cl_dirty_max_pages may be changed at connect time in
 423          * ptlrpc_connect_interpret().
 424          */
 425         client_adjust_max_dirty(cli);
 426         init_waitqueue_head(&cli->cl_cache_waiters);
 427         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 428         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 429         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 430         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 431         spin_lock_init(&cli->cl_loi_list_lock);
 432         atomic_set(&cli->cl_pending_w_pages, 0);
 433         atomic_set(&cli->cl_pending_r_pages, 0);
 434         cli->cl_r_in_flight = 0;
 435         cli->cl_w_in_flight = 0;
 436
 437         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 438         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 439         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 440         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 441         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 442         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 443
 444         /* lru for osc. */
 445         INIT_LIST_HEAD(&cli->cl_lru_osc);
 446         atomic_set(&cli->cl_lru_shrinkers, 0);
 447         atomic_long_set(&cli->cl_lru_busy, 0);
 448         atomic_long_set(&cli->cl_lru_in_list, 0);
 449         INIT_LIST_HEAD(&cli->cl_lru_list);
 450         spin_lock_init(&cli->cl_lru_list_lock);
 451         atomic_long_set(&cli->cl_unstable_count, 0);
 452         INIT_LIST_HEAD(&cli->cl_shrink_list);
 453         INIT_LIST_HEAD(&cli->cl_grant_chain);
 454
 455         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 456         cli->cl_rpcs_in_flight = 0;
 457
 458         init_waitqueue_head(&cli->cl_destroy_waitq);
 459         atomic_set(&cli->cl_destroy_in_flight, 0);
 460
 461
 462         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 463         cli->cl_preferred_cksum_type = 0;
 464 #ifdef ENABLE_CHECKSUM
 465         /* Turn on checksumming by default. */
 466         cli->cl_checksum = 1;
 467         /*
 468          * The supported checksum types will be worked out at connect time
 469          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 470          * through procfs.
 471          */
 472         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 473 #endif
 474         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 475
 476         /*
 477          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 478          * from OFD after connecting.
 479          */
 480         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 481
 482         cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
 483
 484         /*
 485          * set cl_chunkbits default value to PAGE_SHIFT,
 486          * it will be updated at OSC connection time.
 487          */
 488         cli->cl_chunkbits = PAGE_SHIFT;
 489
 490         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 491                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 492         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 493                 cli->cl_max_rpcs_in_flight = 2;
 494         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 495                 cli->cl_max_rpcs_in_flight = 3;
 496         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 497                 cli->cl_max_rpcs_in_flight = 4;
 498         } else {
 499                 if (osc_on_mdt(obd->obd_name))
 500                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 501                 else
 502                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 503         }
 504
 505         spin_lock_init(&cli->cl_mod_rpcs_lock);
 506         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 507         cli->cl_max_mod_rpcs_in_flight = 0;
 508         cli->cl_mod_rpcs_in_flight = 0;
 509         cli->cl_close_rpcs_in_flight = 0;
 510         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 511         cli->cl_mod_tag_bitmap = NULL;
 512
 513         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 514
 515         if (connect_op == MDS_CONNECT) {
 516                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 517                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 518                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 519                 if (cli->cl_mod_tag_bitmap == NULL)
 520                         GOTO(err, rc = -ENOMEM);
 521         }
 522
 523         rc = ldlm_get_ref();
 524         if (rc) {
 525                 CERROR("ldlm_get_ref failed: %d\n", rc);
 526                 GOTO(err, rc);
 527         }
 528
 529         ptlrpc_init_client(rq_portal, rp_portal, name,
 530                            &obd->obd_ldlm_client);
 531
 532         imp = class_new_import(obd);
 533         if (imp == NULL)
 534                 GOTO(err_ldlm, rc = -ENOENT);
 535         imp->imp_client = &obd->obd_ldlm_client;
 536         imp->imp_connect_op = connect_op;
 537         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 538                LUSTRE_CFG_BUFLEN(lcfg, 1));
 539         class_import_put(imp);
 540
 541         if (lustre_cfg_buf(lcfg, 4)) {
 542                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 543
 544                 if (refnet == LNET_NET_ANY) {
 545                         rc = -EINVAL;
 546                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 547                                obd->obd_name, lustre_cfg_string(lcfg, 4),
 548                                rc);
 549                         GOTO(err_import, rc);
 550                 }
 551                 lnet_nid4_to_nid(LNET_MKNID(refnet, 0), &fake_conn.c_self);
 552                 imp->imp_connection = &fake_conn;
 553         }
 554
 555         rc = client_import_add_conn(imp, &server_uuid, 1);
 556         if (rc) {
 557                 CERROR("can't add initial connection\n");
 558                 GOTO(err_import, rc);
 559         }
 560         imp->imp_connection = NULL;
 561
 562         cli->cl_import = imp;
 563         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 564         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 565
 566         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 567                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 568                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 569                                name, obd->obd_name,
 570                                cli->cl_target_uuid.uuid);
 571                         spin_lock(&imp->imp_lock);
 572                         imp->imp_deactive = 1;
 573                         spin_unlock(&imp->imp_lock);
 574                 }
 575         }
 576
 577         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 578                                                 LDLM_NAMESPACE_CLIENT,
 579                                                 LDLM_NAMESPACE_GREEDY,
 580                                                 ns_type);
 581         if (IS_ERR(obd->obd_namespace)) {
 582                 rc = PTR_ERR(obd->obd_namespace);
 583                 CERROR("%s: unable to create client namespace: rc = %d\n",
 584                        obd->obd_name, rc);
 585                 obd->obd_namespace = NULL;
 586                 GOTO(err_import, rc);
 587         }
 588
 589         RETURN(rc);
 590
 591 err_import:
 592         class_destroy_import(imp);
 593 err_ldlm:
 594         ldlm_put_ref();
 595 err:
 596         if (cli->cl_mod_tag_bitmap != NULL)
 597                 OBD_FREE(cli->cl_mod_tag_bitmap,
 598                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 599         cli->cl_mod_tag_bitmap = NULL;
 600
 601         RETURN(rc);
 602 }
 603 EXPORT_SYMBOL(client_obd_setup);
 604
 605 int client_obd_cleanup(struct obd_device *obd)
 606 {
 607         struct client_obd *cli = &obd->u.cli;
 608
 609         ENTRY;
 610
 611         ldlm_namespace_free_post(obd->obd_namespace);
 612         obd->obd_namespace = NULL;
 613
 614         obd_cleanup_client_import(obd);
 615         LASSERT(obd->u.cli.cl_import == NULL);
 616
 617         ldlm_put_ref();
 618
 619         if (cli->cl_mod_tag_bitmap != NULL)
 620                 OBD_FREE(cli->cl_mod_tag_bitmap,
 621                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 622         cli->cl_mod_tag_bitmap = NULL;
 623
 624         RETURN(0);
 625 }
 626 EXPORT_SYMBOL(client_obd_cleanup);
 627
 628 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 629 int client_connect_import(const struct lu_env *env,
 630                           struct obd_export **exp,
 631                           struct obd_device *obd, struct obd_uuid *cluuid,
 632                           struct obd_connect_data *data, void *localdata)
 633 {
 634         struct client_obd *cli = &obd->u.cli;
 635         struct obd_import *imp = cli->cl_import;
 636         struct obd_connect_data *ocd;
 637         struct lustre_handle conn = { 0 };
 638         int rc;
 639
 640         ENTRY;
 641
 642         *exp = NULL;
 643         down_write(&cli->cl_sem);
 644         if (cli->cl_conn_count > 0)
 645                 GOTO(out_sem, rc = -EALREADY);
 646
 647         rc = class_connect(&conn, obd, cluuid);
 648         if (rc)
 649                 GOTO(out_sem, rc);
 650
 651         cli->cl_conn_count++;
 652         *exp = class_conn2export(&conn);
 653
 654         LASSERT(obd->obd_namespace);
 655
 656         imp->imp_dlm_handle = conn;
 657         rc = ptlrpc_init_import(imp);
 658         if (rc != 0)
 659                 GOTO(out_ldlm, rc);
 660
 661         ocd = &imp->imp_connect_data;
 662         if (data) {
 663                 *ocd = *data;
 664                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 665                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 666         }
 667
 668         rc = ptlrpc_connect_import(imp);
 669         if (rc != 0) {
 670                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 671                 GOTO(out_ldlm, rc);
 672         }
 673         LASSERT(*exp != NULL && (*exp)->exp_connection);
 674
 675         if (data) {
 676                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 677                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 678                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 679                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 680                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 681         }
 682
 683         ptlrpc_pinger_add_import(imp);
 684
 685         EXIT;
 686
 687         if (rc) {
 688 out_ldlm:
 689                 cli->cl_conn_count--;
 690                 class_disconnect(*exp);
 691                 *exp = NULL;
 692         }
 693 out_sem:
 694         up_write(&cli->cl_sem);
 695
 696         if (!rc && localdata) {
 697                 LASSERT(cli->cl_cache == NULL); /* only once */
 698                 cli->cl_cache = (struct cl_client_cache *)localdata;
 699                 cl_cache_incref(cli->cl_cache);
 700                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 701
 702                 /* add this osc into entity list */
 703                 LASSERT(list_empty(&cli->cl_lru_osc));
 704                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 705                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 706                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 707         }
 708
 709         return rc;
 710 }
 711 EXPORT_SYMBOL(client_connect_import);
 712
 713 int client_disconnect_export(struct obd_export *exp)
 714 {
 715         struct obd_device *obd = class_exp2obd(exp);
 716         struct client_obd *cli;
 717         struct obd_import *imp;
 718         int rc = 0, err;
 719
 720         ENTRY;
 721
 722         if (!obd) {
 723                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 724                        exp, exp ? exp->exp_handle.h_cookie : -1);
 725                 RETURN(-EINVAL);
 726         }
 727
 728         cli = &obd->u.cli;
 729         imp = cli->cl_import;
 730
 731         down_write(&cli->cl_sem);
 732         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 733                 cli->cl_conn_count);
 734
 735         if (cli->cl_conn_count == 0) {
 736                 CERROR("disconnecting disconnected device (%s)\n",
 737                        obd->obd_name);
 738                 GOTO(out_disconnect, rc = -EINVAL);
 739         }
 740
 741         cli->cl_conn_count--;
 742         if (cli->cl_conn_count != 0)
 743                 GOTO(out_disconnect, rc = 0);
 744
 745         /*
 746          * Mark import deactivated now, so we don't try to reconnect if any
 747          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 748          * fully deactivate the import, or that would drop all requests.
 749          */
 750         spin_lock(&imp->imp_lock);
 751         imp->imp_deactive = 1;
 752         spin_unlock(&imp->imp_lock);
 753
 754         /*
 755          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 756          * delete it regardless.  (It's safe to delete an import that was
 757          * never added.)
 758          */
 759         (void)ptlrpc_pinger_del_import(imp);
 760
 761         if (obd->obd_namespace != NULL) {
 762                 /* obd_force == local only */
 763                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 764                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 765                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 766                                           obd->obd_force);
 767         }
 768
 769         /*
 770          * There's no need to hold sem while disconnecting an import,
 771          * and it may actually cause deadlock in GSS.
 772          */
 773         up_write(&cli->cl_sem);
 774         rc = ptlrpc_disconnect_import(imp, 0);
 775         down_write(&cli->cl_sem);
 776
 777         ptlrpc_invalidate_import(imp);
 778
 779         EXIT;
 780
 781 out_disconnect:
 782         /*
 783          * Use server style - class_disconnect should be always called for
 784          * o_disconnect.
 785          */
 786         err = class_disconnect(exp);
 787         if (!rc && err)
 788                 rc = err;
 789
 790         up_write(&cli->cl_sem);
 791
 792         RETURN(rc);
 793 }
 794 EXPORT_SYMBOL(client_disconnect_export);
 795
 796 #ifdef HAVE_SERVER_SUPPORT
 797 int server_disconnect_export(struct obd_export *exp)
 798 {
 799         int rc;
 800
 801         ENTRY;
 802
 803         /* Disconnect early so that clients can't keep using export. */
 804         rc = class_disconnect(exp);
 805         /* Close import to avoid sending any requests. */
 806         if (exp->exp_imp_reverse)
 807                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 808
 809         ldlm_bl_thread_wakeup();
 810
 811         /* complete all outstanding replies */
 812         spin_lock(&exp->exp_lock);
 813         while (!list_empty(&exp->exp_outstanding_replies)) {
 814                 struct ptlrpc_reply_state *rs =
 815                         list_entry(exp->exp_outstanding_replies.next,
 816                                        struct ptlrpc_reply_state, rs_exp_list);
 817                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 818
 819                 spin_lock(&svcpt->scp_rep_lock);
 820
 821                 list_del_init(&rs->rs_exp_list);
 822
 823                 spin_lock(&rs->rs_lock);
 824                 /* clear rs_convert_lock to make sure rs is handled and put */
 825                 rs->rs_convert_lock = 0;
 826                 ptlrpc_schedule_difficult_reply(rs);
 827                 spin_unlock(&rs->rs_lock);
 828
 829                 spin_unlock(&svcpt->scp_rep_lock);
 830         }
 831         spin_unlock(&exp->exp_lock);
 832
 833         RETURN(rc);
 834 }
 835 EXPORT_SYMBOL(server_disconnect_export);
 836
 837 static inline int target_check_recovery_timer(struct obd_device *target)
 838 {
 839         ktime_t remaining;
 840         s64 timeout;
 841
 842         if (!target->obd_recovering || target->obd_recovery_start == 0)
 843                 return 0;
 844
 845         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
 846         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 847         if (timeout > -30)
 848                 return 0;
 849
 850         /* the recovery timer should expire, but it isn't triggered,
 851          * it's better to abort the recovery of this target to speed up
 852          * the recovery of the whole cluster. */
 853         spin_lock(&target->obd_dev_lock);
 854         if (target->obd_recovering) {
 855                 CERROR("%s: Aborting recovery\n", target->obd_name);
 856                 target->obd_abort_recovery = 1;
 857                 wake_up(&target->obd_next_transno_waitq);
 858         }
 859         spin_unlock(&target->obd_dev_lock);
 860         return 0;
 861 }
 862
 863 /*
 864  * --------------------------------------------------------------------------
 865  * from old lib/target.c
 866  * --------------------------------------------------------------------------
 867  */
 868 static int target_handle_reconnect(struct lustre_handle *conn,
 869                                    struct obd_export *exp,
 870                                    struct obd_uuid *cluuid)
 871 {
 872         struct obd_device *target;
 873         struct lustre_handle *hdl;
 874         ktime_t remaining;
 875         s64 timeout;
 876         int rc = 0;
 877
 878         ENTRY;
 879         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 880         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 881                 conn->cookie = exp->exp_handle.h_cookie;
 882                 CDEBUG(D_HA,
 883                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 884                        cluuid->uuid, exp, conn->cookie);
 885                 RETURN(0);
 886         }
 887
 888         target = exp->exp_obd;
 889
 890         /* Might be a re-connect after a partition. */
 891         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 892                 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",
 893                               target->obd_name,
 894                               obd_uuid2str(&exp->exp_client_uuid),
 895                               obd_export_nid2str(exp),
 896                               hdl->cookie, conn->cookie);
 897                 memset(conn, 0, sizeof(*conn));
 898                 /*
 899                  * target_handle_connect() treats EALREADY and
 900                  * -EALREADY differently.  -EALREADY is an error
 901                  * (same UUID, different handle).
 902                  */
 903                 RETURN(-EALREADY);
 904         }
 905
 906         if (!target->obd_recovering) {
 907                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 908                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 909                         obd_export_nid2str(exp));
 910                 GOTO(out_already, rc);
 911         }
 912
 913         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
 914         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 915         if (timeout > 0) {
 916                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 917                               target->obd_name,
 918                               obd_uuid2str(&exp->exp_client_uuid),
 919                               obd_export_nid2str(exp),
 920                               atomic_read(&target->obd_max_recoverable_clients),
 921                               timeout / 60, timeout % 60);
 922         } else {
 923                 struct target_distribute_txn_data *tdtd;
 924                 int size = 0;
 925                 int count = 0;
 926                 char *buf = NULL;
 927
 928                 target_check_recovery_timer(target);
 929
 930                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 931                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 932                         buf = tdtd->tdtd_show_update_logs_retrievers(
 933                                 tdtd->tdtd_show_retrievers_cbdata,
 934                                 &size, &count);
 935
 936                 if (count > 0)
 937                         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",
 938                                       target->obd_name,
 939                                       obd_uuid2str(&exp->exp_client_uuid),
 940                                       obd_export_nid2str(exp), count,
 941                                       buf ? buf : "unknown (not enough RAM)",
 942                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 943                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 944                                       target->obd_name);
 945                 else
 946                         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",
 947                                       target->obd_name, abs(timeout) / 60,
 948                                       abs(timeout) % 60, target->obd_name);
 949
 950                 if (buf != NULL)
 951                         OBD_FREE(buf, size);
 952         }
 953
 954 out_already:
 955         conn->cookie = exp->exp_handle.h_cookie;
 956         /*
 957          * target_handle_connect() treats EALREADY and
 958          * -EALREADY differently.  EALREADY means we are
 959          * doing a valid reconnect from the same client.
 960          */
 961         RETURN(EALREADY);
 962 }
 963
 964 static void
 965 check_and_start_recovery_timer(struct obd_device *obd,
 966                                struct ptlrpc_request *req, int new_client);
 967
 968 /**
 969  * update flags for import during reconnect process
 970  */
 971 static int rev_import_flags_update(struct obd_import *revimp,
 972                                    struct ptlrpc_request *req)
 973 {
 974         int rc;
 975         struct obd_connect_data *data;
 976
 977         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 978
 979         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 980                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 981         else
 982                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 983
 984         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 985
 986         revimp->imp_connect_data = *data;
 987         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 988         if (rc) {
 989                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 990                         revimp->imp_client->cli_name,
 991                         libcfs_id2str(req->rq_peer), rc);
 992                 return rc;
 993         }
 994
 995         return 0;
 996 }
 997
 998 /**
 999  * Allocate a new reverse import for an export.
1000  *
1001  * \retval -errno in case error hit
1002  * \retval 0 if reverse import correctly init
1003  **/
1004 int rev_import_init(struct obd_export *export)
1005 {
1006         struct obd_device *obd = export->exp_obd;
1007         struct obd_import *revimp;
1008
1009         LASSERT(export->exp_imp_reverse == NULL);
1010
1011         revimp = class_new_import(obd);
1012         if (revimp == NULL)
1013                 return -ENOMEM;
1014
1015         revimp->imp_remote_handle.cookie = 0ULL;
1016         revimp->imp_client = &obd->obd_ldlm_client;
1017         revimp->imp_dlm_fake = 1;
1018
1019         /* it is safe to connect import in new state as no sends possible */
1020         spin_lock(&export->exp_lock);
1021         export->exp_imp_reverse = revimp;
1022         spin_unlock(&export->exp_lock);
1023         class_import_put(revimp);
1024
1025         return 0;
1026 }
1027 EXPORT_SYMBOL(rev_import_init);
1028
1029 /**
1030  * Handle reconnect for an export.
1031  *
1032  * \param exp export to handle reconnect process
1033  * \param req client reconnect request
1034  *
1035  * \retval -rc in case securitfy flavor can't be changed
1036  * \retval 0 in case none problems
1037  */
1038 static int rev_import_reconnect(struct obd_export *exp,
1039                                 struct ptlrpc_request *req)
1040 {
1041         struct obd_import *revimp = exp->exp_imp_reverse;
1042         struct lustre_handle *lh;
1043         int rc;
1044
1045         /* avoid sending a request until import flags are changed */
1046         ptlrpc_import_enter_resend(revimp);
1047
1048         ptlrpc_connection_put(revimp->imp_connection);
1049
1050         /*
1051          * client from recovery don't have a handle so we need to take from
1052          * request. it may produce situation when wrong client connected
1053          * to recovery as we trust a client uuid
1054          */
1055         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1056         revimp->imp_remote_handle = *lh;
1057
1058         /*
1059          * unknown versions will be caught in
1060          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1061          */
1062         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1063
1064         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1065
1066         rc = rev_import_flags_update(revimp, req);
1067         if (rc != 0) {
1068                 /*
1069                  * it is safe to still be in RECOVERY phase as we are not able
1070                  * to setup correct security flavor so requests are not able to
1071                  * be delivered correctly
1072                  */
1073                 return rc;
1074         }
1075
1076         /* resend all rpc's via new connection */
1077         return ptlrpc_import_recovery_state_machine(revimp);
1078 }
1079
1080 int target_handle_connect(struct ptlrpc_request *req)
1081 {
1082         struct obd_device *target = NULL;
1083         struct obd_export *export = NULL;
1084         /*
1085          * connect handle - filled from target_handle_reconnect in
1086          * reconnect case
1087          */
1088         struct lustre_handle conn;
1089         struct lustre_handle *tmp;
1090         struct obd_uuid cluuid;
1091         char *str;
1092         int rc = 0;
1093         char *target_start;
1094         int target_len;
1095         bool mds_conn = false, lw_client = false, initial_conn = false;
1096         bool mds_mds_conn = false;
1097         bool new_mds_mds_conn = false;
1098         struct obd_connect_data *data, *tmpdata;
1099         int size, tmpsize;
1100         lnet_nid_t *client_nid = NULL;
1101         struct ptlrpc_connection *pcon = NULL;
1102
1103         ENTRY;
1104
1105         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1106
1107         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1108         if (str == NULL) {
1109                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1110                 GOTO(out, rc = -EINVAL);
1111         }
1112
1113         target = class_dev_by_str(str);
1114         if (!target) {
1115                 deuuidify(str, NULL, &target_start, &target_len);
1116                 LCONSOLE_ERROR_MSG(0x137,
1117                                    "%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",
1118                                    str, libcfs_nid2str(req->rq_peer.nid));
1119                 GOTO(out, rc = -ENODEV);
1120         }
1121
1122         spin_lock(&target->obd_dev_lock);
1123
1124         target->obd_conn_inprogress++;
1125
1126         if (target->obd_stopping || !target->obd_set_up) {
1127                 spin_unlock(&target->obd_dev_lock);
1128
1129                 deuuidify(str, NULL, &target_start, &target_len);
1130                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1131                               target_len, target_start,
1132                               libcfs_nid2str(req->rq_peer.nid),
1133                               (target->obd_stopping ?
1134                                "stopping" : "not set up"));
1135                 GOTO(out, rc = -ENODEV);
1136         }
1137
1138         if (target->obd_no_conn) {
1139                 spin_unlock(&target->obd_dev_lock);
1140
1141                 CDEBUG(D_INFO,
1142                        "%s: Temporarily refusing client connection from %s\n",
1143                        target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1144                 GOTO(out, rc = -EAGAIN);
1145         }
1146
1147         spin_unlock(&target->obd_dev_lock);
1148
1149         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1150         if (str == NULL) {
1151                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1152                 GOTO(out, rc = -EINVAL);
1153         }
1154
1155         obd_str2uuid(&cluuid, str);
1156
1157         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1158         if (tmp == NULL)
1159                 GOTO(out, rc = -EPROTO);
1160
1161         conn = *tmp;
1162
1163         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1164                                     RCL_CLIENT);
1165         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1166                 GOTO(out, rc = -EPROTO);
1167         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1168         if (!data)
1169                 GOTO(out, rc = -EPROTO);
1170
1171         rc = req_capsule_server_pack(&req->rq_pill);
1172         if (rc)
1173                 GOTO(out, rc);
1174
1175 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1176         /*
1177          * Don't allow clients to connect that are using old 1.8 format
1178          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1179          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1180          * FULL20 flag should be set on all connections since 2.0, but no
1181          * longer affects behaviour.
1182          *
1183          * Later this check will be disabled and the flag can be retired
1184          * completely once interop with 3.0 is no longer needed.
1185          */
1186         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1187                 GOTO(out, rc = -EPROTO);
1188
1189         /*
1190          * Don't allow liblustre clients to connect.
1191          * - testing was disabled in v2_2_50_0-61-g6a75d65
1192          * - building was disabled in v2_5_58_0-28-g7277179
1193          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1194          * - clients were refused connect for version difference > 0.0.1.32
1195          */
1196         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1197                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1198                 GOTO(out, rc = -EPROTO);
1199         }
1200 #endif
1201
1202         /*
1203          * Note: lw_client is needed in MDS-MDS failover during update log
1204          * processing, so we needs to allow lw_client to be connected at
1205          * anytime, instead of only the initial connection
1206          */
1207         lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1208
1209         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1210                 initial_conn = true;
1211                 mds_conn = OCD_HAS_FLAG(data, MDS);
1212                 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1213
1214                 /*
1215                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1216                  * Checking OBD_CONNECT_FID can be removed in the future.
1217                  *
1218                  * Via check OBD_CONNECT_FID, we can distinguish whether
1219                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1220                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1221                  */
1222                 if (!lw_client &&
1223                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1224                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1225                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1226                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1227                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1228                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1229
1230                         /*
1231                          * We do not support the MDT-MDT interoperations with
1232                          * different version MDT because of protocol changes.
1233                          */
1234                         if (unlikely(major != LUSTRE_MAJOR ||
1235                                      minor != LUSTRE_MINOR ||
1236                                      abs(patch - LUSTRE_PATCH) > 3)) {
1237                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1238                                               target->obd_name, LUSTRE_MAJOR,
1239                                               LUSTRE_MINOR, LUSTRE_PATCH,
1240                                               LUSTRE_FIX, major, minor, patch,
1241                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1242                                               libcfs_nid2str(req->rq_peer.nid),
1243                                               str);
1244                                 GOTO(out, rc = -EPROTO);
1245                         }
1246                 }
1247         }
1248
1249         /* lctl gets a backstage, all-access pass. */
1250         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1251                 goto dont_check_exports;
1252
1253         export = obd_uuid_lookup(target, &cluuid);
1254         if (!export)
1255                 goto no_export;
1256
1257         /* We've found an export in the hash. */
1258
1259         spin_lock(&export->exp_lock);
1260
1261         if (export->exp_connecting) { /* b=9635, et. al. */
1262                 spin_unlock(&export->exp_lock);
1263                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1264                               export->exp_obd->obd_name, export,
1265                               libcfs_nid2str(req->rq_peer.nid));
1266                 class_export_put(export);
1267                 export = NULL;
1268                 rc = -EALREADY;
1269         } else if ((mds_conn || (lw_client && initial_conn) ||
1270                    OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1271                 spin_unlock(&export->exp_lock);
1272                 if (req->rq_peer.nid !=
1273                     lnet_nid_to_nid4(&export->exp_connection->c_peer.nid)) {
1274                         /* MDS or LWP reconnected after failover. */
1275                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1276                                       target->obd_name,
1277                                       lw_client ? "LWP" : "MDS",
1278                                       libcfs_nid2str(req->rq_peer.nid),
1279                                       libcfs_nidstr(&export->exp_connection->c_peer.nid));
1280                 } else {
1281                         /* New connection from the same NID. */
1282                         LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1283                                       target->obd_name,
1284                                       lw_client ? "LWP" : "MDS",
1285                                       libcfs_nid2str(req->rq_peer.nid),
1286                                       OCD_HAS_FLAG(data, MDS_MDS) ?
1287                                       "keep" : "remove");
1288                 }
1289
1290                 if (req->rq_peer.nid ==
1291                     lnet_nid_to_nid4(&export->exp_connection->c_peer.nid) &&
1292                     OCD_HAS_FLAG(data, MDS_MDS)) {
1293                         /*
1294                          * Because exports between MDTs will always be
1295                          * kept, let's do not fail such export if they
1296                          * come from the same NID, otherwise it might
1297                          * cause eviction between MDTs, which might
1298                          * cause namespace inconsistency
1299                          */
1300                         spin_lock(&export->exp_lock);
1301                         export->exp_connecting = 1;
1302                         export->exp_conn_cnt = 0;
1303                         spin_unlock(&export->exp_lock);
1304                         conn.cookie = export->exp_handle.h_cookie;
1305                         rc = EALREADY;
1306                 } else {
1307                         class_fail_export(export);
1308                         class_export_put(export);
1309                         export = NULL;
1310                         rc = 0;
1311                 }
1312         } else if (export->exp_connection != NULL && initial_conn &&
1313                    req->rq_peer.nid != lnet_nid_to_nid4(&export->exp_connection->c_peer.nid)) {
1314                 spin_unlock(&export->exp_lock);
1315                 /* In MDS failover we have static UUID but NID can change. */
1316                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1317                               target->obd_name, cluuid.uuid,
1318                               libcfs_nid2str(req->rq_peer.nid),
1319                               libcfs_nidstr(
1320                                       &export->exp_connection->c_peer.nid));
1321                 rc = -EALREADY;
1322                 class_export_put(export);
1323                 export = NULL;
1324         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1325                    !lw_client) {
1326                 spin_unlock(&export->exp_lock);
1327                 rc = -EAGAIN;
1328         } else {
1329                 export->exp_connecting = 1;
1330                 spin_unlock(&export->exp_lock);
1331                 LASSERT(export->exp_obd == target);
1332
1333                 rc = target_handle_reconnect(&conn, export, &cluuid);
1334         }
1335
1336         /* If we found an export, we already unlocked. */
1337         if (!export) {
1338 no_export:
1339                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1340         } else if (req->rq_export == NULL &&
1341                    atomic_read(&export->exp_rpc_count) > 0) {
1342                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1343                               target->obd_name, cluuid.uuid,
1344                               libcfs_nid2str(req->rq_peer.nid),
1345                               refcount_read(&export->exp_handle.h_ref));
1346                         GOTO(out, rc = -EBUSY);
1347         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1348                    rc != EALREADY) {
1349                 if (!strstr(cluuid.uuid, "mdt"))
1350                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1351                                       target->obd_name, cluuid.uuid,
1352                                       libcfs_nid2str(req->rq_peer.nid));
1353                 GOTO(out, rc = -EALREADY);
1354         } else {
1355                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1356         }
1357
1358         if (rc < 0)
1359                 GOTO(out, rc);
1360
1361         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1362                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1363                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1364                export, ktime_get_seconds(),
1365                export ? export->exp_last_request_time : 0);
1366
1367         /*
1368          * If this is the first time a client connects, reset the recovery
1369          * timer. Discard lightweight connections which might be local.
1370          */
1371         if (!lw_client && rc == 0 && target->obd_recovering)
1372                 check_and_start_recovery_timer(target, req, export == NULL);
1373
1374         /*
1375          * We want to handle EALREADY but *not* -EALREADY from
1376          * target_handle_reconnect(), return reconnection state in a flag.
1377          */
1378         if (rc == EALREADY) {
1379                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1380                 rc = 0;
1381         } else {
1382                 LASSERT(rc == 0);
1383         }
1384
1385         /* Tell the client if we support replayable requests. */
1386         if (target->obd_replayable)
1387                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1388         client_nid = &req->rq_peer.nid;
1389
1390         if (export == NULL) {
1391                 /* allow lightweight connections during recovery */
1392                 /*
1393                  * allow "new" MDT to be connected during recovery, since we
1394                  * need retrieve recovery update records from it
1395                  */
1396                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1397                         struct hrtimer *timer = &target->obd_recovery_timer;
1398                         ktime_t remaining;
1399                         s64 timeout, left;
1400                         int in_progress;
1401                         int connected;
1402                         int known;
1403                         int stale;
1404                         char *msg;
1405
1406                         connected = atomic_read(&target->obd_connected_clients);
1407                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1408                         known =
1409                            atomic_read(&target->obd_max_recoverable_clients);
1410                         stale = target->obd_stale_clients;
1411                         remaining = hrtimer_get_remaining(timer);
1412                         left = ktime_divns(remaining, NSEC_PER_SEC);
1413
1414                         if (ktime_to_ns(remaining) > 0) {
1415                                 msg = "to recover in";
1416                                 timeout = left;
1417                         } else {
1418                                 msg = "already passed deadline";
1419                                 timeout = -left;
1420
1421                                 target_check_recovery_timer(target);
1422                         }
1423
1424                         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",
1425                                       target->obd_name, cluuid.uuid,
1426                                       libcfs_nid2str(req->rq_peer.nid), known,
1427                                       connected - in_progress, in_progress,
1428                                       stale, msg, timeout / 60, timeout % 60);
1429                         rc = -EBUSY;
1430                 } else {
1431 dont_check_exports:
1432                         rc = obd_connect(req->rq_svc_thread->t_env,
1433                                          &export, target, &cluuid, data,
1434                                          client_nid);
1435                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1436                                 lustre_msg_add_op_flags(req->rq_repmsg,
1437                                                         MSG_CONNECT_RECOVERING);
1438                         if (rc == 0) {
1439                                 conn.cookie = export->exp_handle.h_cookie;
1440                                 rc = rev_import_init(export);
1441                         }
1442
1443                         if (mds_mds_conn)
1444                                 new_mds_mds_conn = true;
1445                 }
1446         } else {
1447                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1448                                    export, target, &cluuid, data, client_nid);
1449         }
1450         if (rc)
1451                 GOTO(out, rc);
1452
1453         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1454         data->ocd_instance = target->u.obt.obt_instance;
1455
1456         /*
1457          * Return only the parts of obd_connect_data that we understand, so the
1458          * client knows that we don't understand the rest.
1459          */
1460         if (data) {
1461                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1462                                                RCL_SERVER);
1463                 tmpdata = req_capsule_server_get(&req->rq_pill,
1464                                                  &RMF_CONNECT_DATA);
1465                 /*
1466                  * Don't use struct assignment here, because the client reply
1467                  * buffer may be smaller/larger than the local struct
1468                  * obd_connect_data.
1469                  */
1470                 memcpy(tmpdata, data, min(tmpsize, size));
1471         }
1472
1473         /*
1474          * If the client and the server are the same node, we will already
1475          * have an export that really points to the client's DLM export,
1476          * because we have a shared handles table.
1477          *
1478          * XXX this will go away when shaver stops sending the "connect" handle
1479          * in the real "remote handle" field of the request --phik 24 Apr 2003
1480          */
1481         ptlrpc_request_change_export(req, export);
1482
1483         pcon = ptlrpc_connection_get(req->rq_peer, req->rq_self, &cluuid);
1484         if (pcon == NULL)
1485                 GOTO(out, rc = -ENOTCONN);
1486
1487         spin_lock(&export->exp_lock);
1488
1489         if (export->exp_disconnected) {
1490                 spin_unlock(&export->exp_lock);
1491                 GOTO(out, rc = -ENODEV);
1492         }
1493         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1494                 spin_unlock(&export->exp_lock);
1495                 CDEBUG(D_RPCTRACE,
1496                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1497                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1498                        export->exp_conn_cnt,
1499                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1500
1501                 GOTO(out, rc = -EALREADY);
1502         }
1503         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1504         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1505
1506         /* Check to see if connection came from another NID. */
1507         if (export->exp_connection != NULL &&
1508             lnet_nid_to_nid4(&export->exp_connection->c_peer.nid) !=
1509             req->rq_peer.nid) {
1510                 obd_nid_del(export->exp_obd, export);
1511                 ptlrpc_connection_put(export->exp_connection);
1512                 export->exp_connection = NULL;
1513         }
1514
1515         if (export->exp_connection == NULL) {
1516                 export->exp_connection = pcon;
1517                 pcon = NULL;
1518         }
1519         obd_nid_add(export->exp_obd, export);
1520
1521         spin_unlock(&export->exp_lock);
1522
1523         lustre_msg_set_handle(req->rq_repmsg, &conn);
1524
1525         rc = rev_import_reconnect(export, req);
1526         if (rc != 0)
1527                 GOTO(out, rc);
1528
1529         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1530                 int has_transno;
1531                 __u64 transno = data->ocd_transno;
1532
1533                 spin_lock(&export->exp_lock);
1534                 /*
1535                  * possible race with class_disconnect_stale_exports,
1536                  * export may be already in the eviction process
1537                  */
1538                 if (export->exp_failed) {
1539                         spin_unlock(&export->exp_lock);
1540                         GOTO(out, rc = -ENODEV);
1541                 }
1542                 export->exp_in_recovery = 1;
1543                 export->exp_req_replay_needed = 1;
1544                 export->exp_lock_replay_needed = 1;
1545                 spin_unlock(&export->exp_lock);
1546
1547                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1548                                  MSG_CONNECT_TRANSNO);
1549                 if (has_transno && transno == 0)
1550                         CWARN("Connect with zero transno!\n");
1551
1552                 if (has_transno && transno > 0 &&
1553                     transno < target->obd_next_recovery_transno &&
1554                     transno > target->obd_last_committed) {
1555                         /* Another way is to use cmpxchg() to be lock-free. */
1556                         spin_lock(&target->obd_recovery_task_lock);
1557                         if (transno < target->obd_next_recovery_transno)
1558                                 target->obd_next_recovery_transno = transno;
1559                         spin_unlock(&target->obd_recovery_task_lock);
1560                 }
1561
1562                 atomic_inc(&target->obd_req_replay_clients);
1563                 atomic_inc(&target->obd_lock_replay_clients);
1564                 /*
1565                  * Note: MDS-MDS connection is allowed to be connected during
1566                  * recovery, no matter if the exports needs to be recoveried.
1567                  * Because we need retrieve updates logs from all other MDTs.
1568                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1569                  * also needs to be increased to match other recovery checking
1570                  * condition.
1571                  */
1572                 if (new_mds_mds_conn)
1573                         atomic_inc(&target->obd_max_recoverable_clients);
1574
1575                 if (atomic_inc_return(&target->obd_connected_clients) ==
1576                     atomic_read(&target->obd_max_recoverable_clients))
1577                         wake_up(&target->obd_next_transno_waitq);
1578         }
1579
1580         /* Tell the client we're in recovery, when client is involved in it. */
1581         if (target->obd_recovering && !lw_client)
1582                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1583
1584 out:
1585         if (export) {
1586                 spin_lock(&export->exp_lock);
1587                 export->exp_connecting = 0;
1588                 spin_unlock(&export->exp_lock);
1589
1590                 class_export_put(export);
1591         }
1592         if (target != NULL) {
1593                 spin_lock(&target->obd_dev_lock);
1594                 target->obd_conn_inprogress--;
1595                 spin_unlock(&target->obd_dev_lock);
1596                 class_decref(target, "find", current);
1597         }
1598         if (pcon)
1599                 ptlrpc_connection_put(pcon);
1600         req->rq_status = rc;
1601         RETURN(rc);
1602 }
1603
1604 int target_handle_disconnect(struct ptlrpc_request *req)
1605 {
1606         int rc;
1607
1608         ENTRY;
1609
1610         rc = req_capsule_server_pack(&req->rq_pill);
1611         if (rc)
1612                 RETURN(rc);
1613
1614         /* In case of target disconnect, updating sec ctx immediately is
1615          * required in order to record latest sequence number used.
1616          * Sequence is normally updated on export destroy, but this event
1617          * can occur too late, ie after a new target connect request has
1618          * been processed.
1619          * Maintaining correct sequence when client connection becomes idle
1620          * ensures that GSS does not erroneously consider requests as replays.
1621          */
1622         rc = sptlrpc_export_update_ctx(req->rq_export);
1623         if (rc)
1624                 RETURN(rc);
1625
1626         /* Keep the rq_export around so we can send the reply. */
1627         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1628
1629         RETURN(0);
1630 }
1631
1632 void target_destroy_export(struct obd_export *exp)
1633 {
1634         struct obd_import *imp = NULL;
1635         /*
1636          * exports created from last_rcvd data, and "fake"
1637          * exports created by lctl don't have an import
1638          */
1639         spin_lock(&exp->exp_lock);
1640         if (exp->exp_imp_reverse != NULL) {
1641                 imp = exp->exp_imp_reverse;
1642                 exp->exp_imp_reverse = NULL;
1643         }
1644         spin_unlock(&exp->exp_lock);
1645         if (imp != NULL)
1646                 client_destroy_import(imp);
1647
1648         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1649         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1650         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1651         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1652 }
1653 EXPORT_SYMBOL(target_destroy_export);
1654
1655 /*
1656  * Recovery functions
1657  */
1658 static void target_request_copy_get(struct ptlrpc_request *req)
1659 {
1660         class_export_rpc_inc(req->rq_export);
1661         LASSERT(list_empty(&req->rq_list));
1662         INIT_LIST_HEAD(&req->rq_replay_list);
1663
1664         /* Increase refcount to keep request in queue. */
1665         atomic_inc(&req->rq_refcount);
1666         /* Let export know it has replays to be handled. */
1667         atomic_inc(&req->rq_export->exp_replay_count);
1668 }
1669
1670 static void target_request_copy_put(struct ptlrpc_request *req)
1671 {
1672         LASSERT(list_empty(&req->rq_replay_list));
1673         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1674
1675         atomic_dec(&req->rq_export->exp_replay_count);
1676         class_export_rpc_dec(req->rq_export);
1677         ptlrpc_server_drop_request(req);
1678 }
1679
1680 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1681 {
1682         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1683         struct obd_export *exp = req->rq_export;
1684         struct ptlrpc_request *reqiter;
1685         struct ptlrpc_request *dup_req = NULL;
1686         int dup = 0;
1687
1688         LASSERT(exp);
1689
1690         spin_lock(&exp->exp_lock);
1691         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1692                             rq_replay_list) {
1693                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1694                         dup_req = reqiter;
1695                         dup = 1;
1696                         break;
1697                 }
1698         }
1699
1700         if (dup) {
1701                 /* We expect it with RESENT and REPLAY flags. */
1702                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1703                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1704                         CERROR("invalid flags %x of resent replay\n",
1705                                lustre_msg_get_flags(req->rq_reqmsg));
1706
1707                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1708                         __u32 new_conn;
1709
1710                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1711                         if (new_conn >
1712                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1713                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1714                                                         new_conn);
1715                 }
1716         } else {
1717                 list_add_tail(&req->rq_replay_list,
1718                               &exp->exp_req_replay_queue);
1719         }
1720
1721         spin_unlock(&exp->exp_lock);
1722         return dup;
1723 }
1724
1725 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1726 {
1727         LASSERT(!list_empty(&req->rq_replay_list));
1728         LASSERT(req->rq_export);
1729
1730         spin_lock(&req->rq_export->exp_lock);
1731         list_del_init(&req->rq_replay_list);
1732         spin_unlock(&req->rq_export->exp_lock);
1733 }
1734
1735 static void target_finish_recovery(struct lu_target *lut)
1736 {
1737         struct obd_device *obd = lut->lut_obd;
1738
1739         ENTRY;
1740
1741         /* Only log a recovery message when recovery has occurred. */
1742         if (obd->obd_recovery_start) {
1743                 time64_t now = ktime_get_seconds();
1744                 time64_t elapsed_time;
1745
1746                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1747                                      1);
1748                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1749                               obd->obd_name, elapsed_time / 60,
1750                               elapsed_time % 60,
1751                               atomic_read(&obd->obd_max_recoverable_clients),
1752                               atomic_read(&obd->obd_connected_clients),
1753                               obd->obd_stale_clients,
1754                               obd->obd_stale_clients == 1 ? "was" : "were");
1755                 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1756                         libcfs_debug_dumplog();
1757         }
1758
1759         ldlm_reprocess_recovery_done(obd->obd_namespace);
1760         spin_lock(&obd->obd_recovery_task_lock);
1761         if (!list_empty(&obd->obd_req_replay_queue) ||
1762             !list_empty(&obd->obd_lock_replay_queue) ||
1763             !list_empty(&obd->obd_final_req_queue)) {
1764                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1765                        obd->obd_name,
1766                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1767                        list_empty(&obd->obd_lock_replay_queue) ? \
1768                                   "" : "lock ",
1769                        list_empty(&obd->obd_final_req_queue) ? \
1770                                   "" : "final ");
1771                 spin_unlock(&obd->obd_recovery_task_lock);
1772                 LBUG();
1773         }
1774         spin_unlock(&obd->obd_recovery_task_lock);
1775
1776         obd->obd_recovery_end = ktime_get_seconds();
1777
1778         /* When recovery finished, cleanup orphans on MDS and OST. */
1779         if (obd->obd_type && OBP(obd, postrecov)) {
1780                 int rc = OBP(obd, postrecov)(obd);
1781
1782                 if (rc < 0)
1783                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1784                                       obd->obd_name, rc);
1785         }
1786         EXIT;
1787 }
1788
1789 static void abort_req_replay_queue(struct obd_device *obd)
1790 {
1791         struct ptlrpc_request *req, *n;
1792         LIST_HEAD(abort_list);
1793
1794         spin_lock(&obd->obd_recovery_task_lock);
1795         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1796         spin_unlock(&obd->obd_recovery_task_lock);
1797         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1798                 DEBUG_REQ(D_WARNING, req, "aborted:");
1799                 req->rq_status = -ENOTCONN;
1800                 if (ptlrpc_error(req)) {
1801                         DEBUG_REQ(D_ERROR, req,
1802                                   "failed abort_req_reply; skipping");
1803                 }
1804                 target_exp_dequeue_req_replay(req);
1805                 target_request_copy_put(req);
1806         }
1807 }
1808
1809 static void abort_lock_replay_queue(struct obd_device *obd)
1810 {
1811         struct ptlrpc_request *req, *n;
1812         LIST_HEAD(abort_list);
1813
1814         spin_lock(&obd->obd_recovery_task_lock);
1815         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1816         spin_unlock(&obd->obd_recovery_task_lock);
1817         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1818                 DEBUG_REQ(D_ERROR, req, "aborted:");
1819                 req->rq_status = -ENOTCONN;
1820                 if (ptlrpc_error(req)) {
1821                         DEBUG_REQ(D_ERROR, req,
1822                                   "failed abort_lock_reply; skipping");
1823                 }
1824                 target_request_copy_put(req);
1825         }
1826 }
1827
1828 /*
1829  * Called from a cleanup function if the device is being cleaned up
1830  * forcefully.  The exports should all have been disconnected already,
1831  * the only thing left to do is
1832  * - clear the recovery flags
1833  * - cancel the timer
1834  * - free queued requests and replies, but don't send replies
1835  * Because the obd_stopping flag is set, no new requests should be received.
1836  */
1837 void target_cleanup_recovery(struct obd_device *obd)
1838 {
1839         struct ptlrpc_request *req, *n;
1840         LIST_HEAD(clean_list);
1841
1842         spin_lock(&obd->obd_dev_lock);
1843         if (!obd->obd_recovering) {
1844                 spin_unlock(&obd->obd_dev_lock);
1845                 EXIT;
1846                 return;
1847         }
1848         obd->obd_recovering = obd->obd_abort_recovery = 0;
1849         obd->obd_abort_recov_mdt = 0;
1850         spin_unlock(&obd->obd_dev_lock);
1851
1852         spin_lock(&obd->obd_recovery_task_lock);
1853         target_cancel_recovery_timer(obd);
1854         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1855         spin_unlock(&obd->obd_recovery_task_lock);
1856
1857         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1858                 LASSERT(req->rq_reply_state == NULL);
1859                 target_exp_dequeue_req_replay(req);
1860                 target_request_copy_put(req);
1861         }
1862
1863         spin_lock(&obd->obd_recovery_task_lock);
1864         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1865         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1866         spin_unlock(&obd->obd_recovery_task_lock);
1867
1868         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1869                 LASSERT(req->rq_reply_state == NULL);
1870                 target_request_copy_put(req);
1871         }
1872
1873         EXIT;
1874 }
1875 EXPORT_SYMBOL(target_cleanup_recovery);
1876
1877 /* obd_recovery_task_lock should be held */
1878 void target_cancel_recovery_timer(struct obd_device *obd)
1879 {
1880         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1881         hrtimer_cancel(&obd->obd_recovery_timer);
1882 }
1883
1884 static void target_start_recovery_timer(struct obd_device *obd)
1885 {
1886         ktime_t delay;
1887
1888         if (obd->obd_recovery_start != 0)
1889                 return;
1890
1891         spin_lock(&obd->obd_dev_lock);
1892         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1893                 spin_unlock(&obd->obd_dev_lock);
1894                 return;
1895         }
1896
1897         LASSERT(obd->obd_recovery_timeout != 0);
1898
1899         if (obd->obd_recovery_start != 0) {
1900                 spin_unlock(&obd->obd_dev_lock);
1901                 return;
1902         }
1903
1904         obd->obd_recovery_start = ktime_get_seconds();
1905         delay = ktime_set(obd->obd_recovery_start +
1906                           obd->obd_recovery_timeout, 0);
1907         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1908         spin_unlock(&obd->obd_dev_lock);
1909
1910         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1911                       obd->obd_name,
1912                       obd->obd_recovery_timeout / 60,
1913                       obd->obd_recovery_timeout % 60,
1914                       atomic_read(&obd->obd_max_recoverable_clients),
1915                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1916                       "" : "s",
1917                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1918                       "s" : "");
1919 }
1920
1921 /**
1922  * extend recovery window.
1923  *
1924  * if @extend is true, extend recovery window to have @dr_timeout remaining
1925  * at least; otherwise, make sure the recovery timeout value is not less
1926  * than @dr_timeout.
1927  */
1928 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1929                                   bool extend)
1930 {
1931         ktime_t left_ns;
1932         timeout_t timeout;
1933         timeout_t left;
1934
1935         spin_lock(&obd->obd_dev_lock);
1936         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1937             obd->obd_stopping) {
1938                 spin_unlock(&obd->obd_dev_lock);
1939                 return;
1940         }
1941         LASSERT(obd->obd_recovery_start != 0);
1942
1943         left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1944         left = ktime_divns(left_ns, NSEC_PER_SEC);
1945
1946         if (extend) {
1947                 timeout = obd->obd_recovery_timeout;
1948                 /* dr_timeout will happen after the hrtimer has expired.
1949                  * Add the excess time to the soft recovery timeout without
1950                  * exceeding the hard recovery timeout.
1951                  */
1952                 if (dr_timeout > left) {
1953                         timeout += dr_timeout - left;
1954                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1955                                         timeout);
1956                 }
1957         } else {
1958                 timeout = clamp_t(timeout_t, dr_timeout,
1959                                   obd->obd_recovery_timeout,
1960                                   obd->obd_recovery_time_hard);
1961         }
1962
1963         if (timeout == obd->obd_recovery_time_hard)
1964                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1965                       obd->obd_name, timeout, extend);
1966
1967         if (obd->obd_recovery_timeout < timeout) {
1968                 ktime_t end, now;
1969
1970                 obd->obd_recovery_timeout = timeout;
1971                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1972                 now = ktime_set(ktime_get_seconds(), 0);
1973                 left_ns = ktime_sub(end, now);
1974                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1975                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1976         }
1977         spin_unlock(&obd->obd_dev_lock);
1978
1979         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1980                 obd->obd_name, left);
1981 }
1982
1983 /* Reset the timer with each new client connection */
1984 /*
1985  * This timer is actually reconnect_timer, which is for making sure
1986  * the total recovery window is at least as big as my reconnect
1987  * attempt timing. So the initial recovery time_out will be set to
1988  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1989  * from client is bigger than this, then the recovery time_out will
1990  * be extended to make sure the client could be reconnected, in the
1991  * process, the timeout from the new client should be ignored.
1992  */
1993 static void
1994 check_and_start_recovery_timer(struct obd_device *obd,
1995                                struct ptlrpc_request *req,
1996                                int new_client)
1997 {
1998         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1999         struct obd_device_target *obt = &obd->u.obt;
2000
2001         if (!new_client && service_timeout)
2002                 /*
2003                  * Teach server about old server's estimates, as first guess
2004                  * at how long new requests will take.
2005                  */
2006                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2007                             service_timeout);
2008
2009         target_start_recovery_timer(obd);
2010
2011         /*
2012          * Convert the service time to RPC timeout,
2013          * and reuse service_timeout to limit stack usage.
2014          */
2015         service_timeout = at_est2timeout(service_timeout);
2016
2017         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2018             service_timeout < at_extra)
2019                 service_timeout = at_extra;
2020
2021         /*
2022          * We expect other clients to timeout within service_timeout, then try
2023          * to reconnect, then try the failover server.  The max delay between
2024          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2025          */
2026         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2027
2028         LASSERT(obt->obt_magic == OBT_MAGIC);
2029         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2030         if (service_timeout > obd->obd_recovery_timeout && !new_client)
2031                 extend_recovery_timer(obd, service_timeout, false);
2032 }
2033
2034 /** Health checking routines */
2035 static inline int exp_connect_healthy(struct obd_export *exp)
2036 {
2037         return exp->exp_in_recovery;
2038 }
2039
2040 /** if export done req_replay or has replay in queue */
2041 static inline int exp_req_replay_healthy(struct obd_export *exp)
2042 {
2043         return (!exp->exp_req_replay_needed ||
2044                 atomic_read(&exp->exp_replay_count) > 0);
2045 }
2046
2047
2048 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2049 {
2050         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2051                exp_req_replay_healthy(exp);
2052 }
2053
2054 /** if export done lock_replay or has replay in queue */
2055 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2056 {
2057         return (!exp->exp_lock_replay_needed ||
2058                 atomic_read(&exp->exp_replay_count) > 0);
2059 }
2060
2061 static inline int exp_vbr_healthy(struct obd_export *exp)
2062 {
2063         return !exp->exp_vbr_failed;
2064 }
2065
2066 static inline int exp_finished(struct obd_export *exp)
2067 {
2068         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2069 }
2070
2071 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2072 {
2073         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2074                 exp_finished(exp);
2075 }
2076
2077 static int check_for_next_transno(struct lu_target *lut)
2078 {
2079         struct ptlrpc_request *req = NULL;
2080         struct obd_device *obd = lut->lut_obd;
2081         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2082         int wake_up = 0, connected, completed, queue_len;
2083         __u64 req_transno = 0;
2084         __u64 update_transno = 0;
2085         __u64 next_transno = 0;
2086
2087         ENTRY;
2088
2089         spin_lock(&obd->obd_recovery_task_lock);
2090         if (!list_empty(&obd->obd_req_replay_queue)) {
2091                 req = list_entry(obd->obd_req_replay_queue.next,
2092                                      struct ptlrpc_request, rq_list);
2093                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2094         }
2095
2096         if (!obd->obd_abort_recov_mdt && tdtd)
2097                 update_transno = distribute_txn_get_next_transno(tdtd);
2098
2099         connected = atomic_read(&obd->obd_connected_clients);
2100         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2101         queue_len = obd->obd_requests_queued_for_recovery;
2102         next_transno = obd->obd_next_recovery_transno;
2103
2104         CDEBUG(D_HA,
2105                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2106                atomic_read(&obd->obd_max_recoverable_clients),
2107                connected, completed,
2108                queue_len, req_transno, next_transno);
2109
2110         if (obd->obd_abort_recovery) {
2111                 CDEBUG(D_HA, "waking for aborted recovery\n");
2112                 wake_up = 1;
2113         } else if (obd->obd_recovery_expired) {
2114                 CDEBUG(D_HA, "waking for expired recovery\n");
2115                 wake_up = 1;
2116         } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2117                    is_req_replayed_by_update(req)) {
2118                 LASSERTF(req_transno < next_transno,
2119                          "req_transno %llu next_transno%llu\n", req_transno,
2120                          next_transno);
2121                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2122                        req_transno);
2123                 wake_up = 1;
2124         } else if (req_transno == next_transno ||
2125                    (update_transno != 0 && update_transno <= next_transno)) {
2126                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2127                 wake_up = 1;
2128         } else if (queue_len > 0 &&
2129                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2130                 /** handle gaps occured due to lost reply or VBR */
2131                 LASSERTF(req_transno >= next_transno,
2132                          "req_transno: %llu, next_transno: %llu\n",
2133                          req_transno, next_transno);
2134                 CDEBUG(D_HA,
2135                        "%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",
2136                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2137                        next_transno, queue_len, completed, connected,
2138                        req_transno, update_transno, obd->obd_last_committed);
2139                 obd->obd_next_recovery_transno = req_transno;
2140                 wake_up = 1;
2141         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2142                 CDEBUG(D_HA, "waking for completed recovery\n");
2143                 wake_up = 1;
2144         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2145                 CDEBUG(D_HA,
2146                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2147                        next_transno);
2148                 obd->obd_next_recovery_transno = req_transno;
2149                 wake_up = 1;
2150         }
2151         spin_unlock(&obd->obd_recovery_task_lock);
2152         return wake_up;
2153 }
2154
2155 static int check_for_next_lock(struct lu_target *lut)
2156 {
2157         struct obd_device *obd = lut->lut_obd;
2158         int wake_up = 0;
2159
2160         spin_lock(&obd->obd_recovery_task_lock);
2161         if (!list_empty(&obd->obd_lock_replay_queue)) {
2162                 CDEBUG(D_HA, "waking for next lock\n");
2163                 wake_up = 1;
2164         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2165                 CDEBUG(D_HA, "waking for completed lock replay\n");
2166                 wake_up = 1;
2167         } else if (obd->obd_abort_recovery) {
2168                 CDEBUG(D_HA, "waking for aborted recovery\n");
2169                 wake_up = 1;
2170         } else if (obd->obd_recovery_expired) {
2171                 CDEBUG(D_HA, "waking for expired recovery\n");
2172                 wake_up = 1;
2173         }
2174         spin_unlock(&obd->obd_recovery_task_lock);
2175
2176         return wake_up;
2177 }
2178
2179 static int check_update_llog(struct lu_target *lut)
2180 {
2181         struct obd_device *obd = lut->lut_obd;
2182         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2183
2184         if (obd->obd_abort_recovery) {
2185                 CDEBUG(D_HA, "waking for aborted recovery\n");
2186                 return 1;
2187         }
2188
2189         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2190                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2191                 return 1;
2192         }
2193
2194         return 0;
2195 }
2196
2197 /**
2198  * wait for recovery events,
2199  * check its status with help of check_routine
2200  * evict dead clients via health_check
2201  */
2202 static int target_recovery_overseer(struct lu_target *lut,
2203                                     int (*check_routine)(struct lu_target *),
2204                                     int (*health_check)(struct obd_export *))
2205 {
2206         struct obd_device *obd = lut->lut_obd;
2207         struct target_distribute_txn_data *tdtd;
2208         time64_t last = 0;
2209         time64_t now;
2210 repeat:
2211         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2212                 now = ktime_get_seconds();
2213                 if (now - last > 600) {
2214                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2215                                       obd->obd_name);
2216                         last = now;
2217                 }
2218         }
2219         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2220               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2221                 __u64 next_update_transno = 0;
2222
2223                 /*
2224                  * Only abort the recovery if there are no update recovery
2225                  * left in the queue
2226                  */
2227                 spin_lock(&obd->obd_recovery_task_lock);
2228                 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2229                         next_update_transno =
2230                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2231
2232                         tdtd = lut->lut_tdtd;
2233                         /*
2234                          * If next_update_transno == 0, it probably because
2235                          * updatelog retrieve threads did not get any records
2236                          * yet, let's wait those threads stopped
2237                          */
2238                         if (next_update_transno == 0) {
2239                                 spin_unlock(&obd->obd_recovery_task_lock);
2240
2241                                 while (wait_event_timeout(
2242                                         tdtd->tdtd_recovery_threads_waitq,
2243                                         check_update_llog(lut),
2244                                         cfs_time_seconds(60)) == 0);
2245
2246                                 spin_lock(&obd->obd_recovery_task_lock);
2247                                 next_update_transno =
2248                                         distribute_txn_get_next_transno(tdtd);
2249                         }
2250                 }
2251
2252                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2253                         obd->obd_next_recovery_transno = next_update_transno;
2254                         spin_unlock(&obd->obd_recovery_task_lock);
2255                         /*
2256                          * Disconnect unfinished exports from clients, and
2257                          * keep connection from MDT to make sure the update
2258                          * recovery will still keep trying until some one
2259                          * manually abort the recovery
2260                          */
2261                         class_disconnect_stale_exports(obd,
2262                                                 exp_finished_or_from_mdt);
2263                         /* Abort all of replay & replay lock req from clients */
2264                         abort_req_replay_queue(obd);
2265                         abort_lock_replay_queue(obd);
2266                         CDEBUG(D_HA,
2267                                "%s: there are still update replay (%#llx)in the queue.\n",
2268                                obd->obd_name, next_update_transno);
2269                 } else {
2270                         obd->obd_abort_recovery = 1;
2271                         spin_unlock(&obd->obd_recovery_task_lock);
2272                         CWARN("%s recovery is aborted by hard timeout\n",
2273                               obd->obd_name);
2274                 }
2275         }
2276
2277         while (wait_event_timeout(obd->obd_next_transno_waitq,
2278                                   check_routine(lut),
2279                                   cfs_time_seconds(60)) == 0)
2280                 ; /* wait indefinitely for event, but don't trigger watchdog */
2281
2282         if (obd->obd_abort_recovery) {
2283                 CWARN("recovery is aborted, evict exports in recovery\n");
2284                 if (lut->lut_tdtd != NULL) {
2285                         tdtd = lut->lut_tdtd;
2286                         /*
2287                          * Let's wait all of the update log recovery thread
2288                          * finished
2289                          */
2290                         wait_event_idle(
2291                                 tdtd->tdtd_recovery_threads_waitq,
2292                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2293                                 == 0);
2294                         /* Then abort the update recovery list */
2295                         dtrq_list_destroy(lut->lut_tdtd);
2296                 }
2297
2298                 /** evict exports which didn't finish recovery yet */
2299                 class_disconnect_stale_exports(obd, exp_finished);
2300                 return 1;
2301         } else if (obd->obd_recovery_expired) {
2302                 obd->obd_recovery_expired = 0;
2303
2304                 /** If some clients died being recovered, evict them */
2305                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2306                               obd->obd_name);
2307                 /** evict cexports with no replay in queue, they are stalled */
2308                 class_disconnect_stale_exports(obd, health_check);
2309
2310                 /** continue with VBR */
2311                 spin_lock(&obd->obd_dev_lock);
2312                 obd->obd_version_recov = 1;
2313                 spin_unlock(&obd->obd_dev_lock);
2314                 /**
2315                  * reset timer, recovery will proceed with versions now,
2316                  * timeout is set just to handle reconnection delays
2317                  */
2318                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2319                 /**
2320                  * Wait for recovery events again, after evicting bad clients
2321                  */
2322                 goto repeat;
2323         }
2324         return 0;
2325 }
2326
2327 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2328 {
2329         struct obd_device *obd = lut->lut_obd;
2330         struct ptlrpc_request *req = NULL;
2331
2332         CDEBUG(D_HA, "Waiting for lock\n");
2333         if (target_recovery_overseer(lut, check_for_next_lock,
2334                                      exp_lock_replay_healthy))
2335                 abort_lock_replay_queue(obd);
2336
2337         spin_lock(&obd->obd_recovery_task_lock);
2338         if (!list_empty(&obd->obd_lock_replay_queue)) {
2339                 req = list_entry(obd->obd_lock_replay_queue.next,
2340                                      struct ptlrpc_request, rq_list);
2341                 list_del_init(&req->rq_list);
2342                 spin_unlock(&obd->obd_recovery_task_lock);
2343         } else {
2344                 spin_unlock(&obd->obd_recovery_task_lock);
2345                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2346                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2347                 /** evict exports failed VBR */
2348                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2349         }
2350         return req;
2351 }
2352
2353 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2354 {
2355         struct ptlrpc_request *req = NULL;
2356
2357         spin_lock(&obd->obd_recovery_task_lock);
2358         if (!list_empty(&obd->obd_final_req_queue)) {
2359                 req = list_entry(obd->obd_final_req_queue.next,
2360                                      struct ptlrpc_request, rq_list);
2361                 list_del_init(&req->rq_list);
2362                 spin_unlock(&obd->obd_recovery_task_lock);
2363                 if (req->rq_export->exp_in_recovery) {
2364                         spin_lock(&req->rq_export->exp_lock);
2365                         req->rq_export->exp_in_recovery = 0;
2366                         spin_unlock(&req->rq_export->exp_lock);
2367                 }
2368         } else {
2369                 spin_unlock(&obd->obd_recovery_task_lock);
2370         }
2371         return req;
2372 }
2373
2374 static void handle_recovery_req(struct ptlrpc_thread *thread,
2375                                 struct ptlrpc_request *req,
2376                                 svc_handler_t handler)
2377 {
2378         ENTRY;
2379
2380         /**
2381          * export can be evicted during recovery, no need to handle replays for
2382          * it after that, discard such request silently
2383          */
2384         if (req->rq_export->exp_disconnected)
2385                 RETURN_EXIT;
2386
2387         req->rq_session.lc_thread = thread;
2388         req->rq_svc_thread = thread;
2389         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2390
2391         /* thread context */
2392         lu_context_enter(&thread->t_env->le_ctx);
2393         (void)handler(req);
2394         lu_context_exit(&thread->t_env->le_ctx);
2395
2396         req->rq_svc_thread->t_env->le_ses = NULL;
2397
2398         /* don't reset timer for final stage */
2399         if (!exp_finished(req->rq_export)) {
2400                 timeout_t timeout = obd_timeout;
2401
2402                 /**
2403                  * Add request @timeout to the recovery time so next request from
2404                  * this client may come in recovery time
2405                  */
2406                 if (!AT_OFF) {
2407                         struct ptlrpc_service_part *svcpt;
2408                         timeout_t est_timeout;
2409
2410                         svcpt = req->rq_rqbd->rqbd_svcpt;
2411                         /*
2412                          * If the server sent early reply for this request,
2413                          * the client will recalculate the timeout according to
2414                          * current server estimate service time, so we will
2415                          * use the maxium timeout here for waiting the client
2416                          * sending the next req
2417                          */
2418                         est_timeout = at_get(&svcpt->scp_at_estimate);
2419                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2420                                         lustre_msg_get_timeout(req->rq_reqmsg));
2421                         /*
2422                          * Add 2 net_latency, one for balance rq_deadline
2423                          * (see ptl_send_rpc), one for resend the req to server,
2424                          * Note: client will pack net_latency in replay req
2425                          * (see ptlrpc_replay_req)
2426                          */
2427                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2428                 }
2429                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2430                                       true);
2431         }
2432         EXIT;
2433 }
2434
2435 /** Checking routines for recovery */
2436 static int check_for_recovery_ready(struct lu_target *lut)
2437 {
2438         struct obd_device *obd = lut->lut_obd;
2439         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2440
2441         CDEBUG(D_HA,
2442                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2443                clnts, obd->obd_stale_clients,
2444                atomic_read(&obd->obd_max_recoverable_clients),
2445                obd->obd_abort_recovery, obd->obd_recovery_expired);
2446
2447         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2448                 LASSERT(clnts <=
2449                         atomic_read(&obd->obd_max_recoverable_clients));
2450                 if (clnts + obd->obd_stale_clients <
2451                     atomic_read(&obd->obd_max_recoverable_clients))
2452                         return 0;
2453         }
2454
2455         if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2456                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2457                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2458                         /*
2459                          * Let's extend recovery timer, in case the recovery
2460                          * timer expired, and some clients got evicted
2461                          */
2462                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2463                                               true);
2464                         CDEBUG(D_HA,
2465                                "%s update recovery is not ready, extend recovery %d\n",
2466                                obd->obd_name, obd->obd_recovery_timeout);
2467                         return 0;
2468                 }
2469         }
2470
2471         return 1;
2472 }
2473
2474 enum {
2475         REQUEST_RECOVERY = 1,
2476         UPDATE_RECOVERY = 2,
2477 };
2478
2479 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2480 {
2481         __u64 transno = 0;
2482
2483         if (!list_empty(&obd->obd_req_replay_queue)) {
2484                 struct ptlrpc_request *req;
2485
2486                 req = list_entry(obd->obd_req_replay_queue.next,
2487                                  struct ptlrpc_request, rq_list);
2488                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2489         }
2490
2491         return transno;
2492 }
2493
2494 static __u64 get_next_transno(struct lu_target *lut, int *type)
2495 {
2496         struct obd_device *obd = lut->lut_obd;
2497         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2498         __u64 transno = 0;
2499         __u64 update_transno;
2500
2501         ENTRY;
2502
2503         transno = get_next_replay_req_transno(obd);
2504         if (type != NULL)
2505                 *type = REQUEST_RECOVERY;
2506
2507         if (!tdtd || obd->obd_abort_recov_mdt)
2508                 RETURN(transno);
2509
2510         update_transno = distribute_txn_get_next_transno(tdtd);
2511         if (transno == 0 || (transno >= update_transno &&
2512                              update_transno != 0)) {
2513                 transno = update_transno;
2514                 if (type != NULL)
2515                         *type = UPDATE_RECOVERY;
2516         }
2517
2518         RETURN(transno);
2519 }
2520
2521 /**
2522  * drop duplicate replay request
2523  *
2524  * Because the operation has been replayed by update recovery, the request
2525  * with the same transno will be dropped and also notify the client to send
2526  * next replay request.
2527  *
2528  * \param[in] env       execution environment
2529  * \param[in] obd       failover obd device
2530  * \param[in] req       request to be dropped
2531  */
2532 static void drop_duplicate_replay_req(struct lu_env *env,
2533                                       struct obd_device *obd,
2534                                       struct ptlrpc_request *req)
2535 {
2536         DEBUG_REQ(D_HA, req,
2537                   "remove t%lld from %s because duplicate update records found",
2538                   lustre_msg_get_transno(req->rq_reqmsg),
2539                   libcfs_nid2str(req->rq_peer.nid));
2540
2541         /*
2542          * Right now, only for MDS reint operation update replay and
2543          * normal request replay can have the same transno
2544          */
2545         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2546                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2547                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2548                 if (likely(req->rq_export))
2549                         target_committed_to_req(req);
2550                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2551                 target_send_reply(req, req->rq_status, 0);
2552         } else {
2553                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2554                 libcfs_nid2str(req->rq_peer.nid));
2555         }
2556         target_exp_dequeue_req_replay(req);
2557         target_request_copy_put(req);
2558         obd->obd_replayed_requests++;
2559 }
2560
2561 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2562
2563 static void replay_request_or_update(struct lu_env *env,
2564                                      struct lu_target *lut,
2565                                      struct target_recovery_data *trd,
2566                                      struct ptlrpc_thread *thread)
2567 {
2568         struct obd_device *obd = lut->lut_obd;
2569         struct ptlrpc_request *req = NULL;
2570         int type;
2571         __u64 transno;
2572
2573         ENTRY;
2574
2575         CDEBUG(D_HA, "Waiting for transno %lld\n",
2576                obd->obd_next_recovery_transno);
2577
2578         /* Replay all of request and update by transno */
2579         do {
2580                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2581
2582                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2583
2584                 /**
2585                  * It is needed to extend recovery window above
2586                  *  recovery_time_soft. Extending is possible only in the
2587                  *  end of recovery window (see more details in
2588                  *  handle_recovery_req()).
2589                  */
2590                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2591
2592                 if (target_recovery_overseer(lut, check_for_next_transno,
2593                                         exp_req_replay_healthy_or_from_mdt)) {
2594                         abort_req_replay_queue(obd);
2595                         abort_lock_replay_queue(obd);
2596                         goto abort;
2597                 }
2598
2599                 spin_lock(&obd->obd_recovery_task_lock);
2600                 transno = get_next_transno(lut, &type);
2601                 if (type == REQUEST_RECOVERY && transno != 0) {
2602                         /*
2603                          * Drop replay request from client side, if the
2604                          * replay has been executed by update with the
2605                          * same transno
2606                          */
2607                         req = list_entry(obd->obd_req_replay_queue.next,
2608                                         struct ptlrpc_request, rq_list);
2609
2610                         list_del_init(&req->rq_list);
2611                         obd->obd_requests_queued_for_recovery--;
2612                         spin_unlock(&obd->obd_recovery_task_lock);
2613
2614                         /*
2615                          * Let's check if the request has been redone by
2616                          * update replay
2617                          */
2618                         if (is_req_replayed_by_update(req)) {
2619                                 struct distribute_txn_replay_req *dtrq;
2620
2621                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2622                                                                       transno);
2623                                 LASSERT(dtrq != NULL);
2624                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2625                                 list_del_init(&dtrq->dtrq_list);
2626                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2627                                 dtrq_destroy(dtrq);
2628
2629                                 drop_duplicate_replay_req(env, obd, req);
2630
2631                                 continue;
2632                         }
2633
2634                         LASSERT(trd->trd_processing_task == current->pid);
2635                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2636                                   req->rq_xid,
2637                                   lustre_msg_get_transno(req->rq_reqmsg),
2638                                   libcfs_nid2str(req->rq_peer.nid));
2639
2640                         ptlrpc_watchdog_init(&thread->t_watchdog,
2641                                              WATCHDOG_TIMEOUT);
2642                         handle_recovery_req(thread, req,
2643                                             trd->trd_recovery_handler);
2644                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2645
2646                         /**
2647                          * bz18031: increase next_recovery_transno before
2648                          * target_request_copy_put() will drop exp_rpc reference
2649                          */
2650                         spin_lock(&obd->obd_recovery_task_lock);
2651                         obd->obd_next_recovery_transno++;
2652                         spin_unlock(&obd->obd_recovery_task_lock);
2653                         target_exp_dequeue_req_replay(req);
2654                         target_request_copy_put(req);
2655                         obd->obd_replayed_requests++;
2656                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2657                         struct distribute_txn_replay_req *dtrq;
2658                         int rc;
2659
2660                         spin_unlock(&obd->obd_recovery_task_lock);
2661
2662                         LASSERT(tdtd != NULL);
2663                         dtrq = distribute_txn_get_next_req(tdtd);
2664                         lu_context_enter(&thread->t_env->le_ctx);
2665                         ptlrpc_watchdog_init(&thread->t_watchdog,
2666                                              WATCHDOG_TIMEOUT);
2667                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2668                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2669                         lu_context_exit(&thread->t_env->le_ctx);
2670                         extend_recovery_timer(obd, obd_timeout, true);
2671
2672                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2673                                 CDEBUG(D_HA,
2674                                        "Move x%llu t%llu to finish list\n",
2675                                        dtrq->dtrq_xid,
2676                                        dtrq->dtrq_master_transno);
2677
2678                                 /* Add it to the replay finish list */
2679                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2680                                 list_add(&dtrq->dtrq_list,
2681                                          &tdtd->tdtd_replay_finish_list);
2682                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2683
2684                                 spin_lock(&obd->obd_recovery_task_lock);
2685                                 if (transno == obd->obd_next_recovery_transno)
2686                                         obd->obd_next_recovery_transno++;
2687                                 else if (transno >
2688                                          obd->obd_next_recovery_transno)
2689                                         obd->obd_next_recovery_transno =
2690                                                                 transno + 1;
2691                                 spin_unlock(&obd->obd_recovery_task_lock);
2692                         } else {
2693                                 dtrq_destroy(dtrq);
2694                         }
2695                 } else {
2696                         spin_unlock(&obd->obd_recovery_task_lock);
2697 abort:
2698                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2699                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2700                         /** evict exports failed VBR */
2701                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2702                         break;
2703                 }
2704         } while (1);
2705 }
2706
2707 static int target_recovery_thread(void *arg)
2708 {
2709         struct lu_target *lut = arg;
2710         struct obd_device *obd = lut->lut_obd;
2711         struct ptlrpc_request *req;
2712         struct target_recovery_data *trd = &obd->obd_recovery_data;
2713         unsigned long delta;
2714         struct lu_env *env;
2715         struct ptlrpc_thread *thread = NULL;
2716         int rc = 0;
2717
2718         ENTRY;
2719         OBD_ALLOC_PTR(thread);
2720         if (thread == NULL)
2721                 RETURN(-ENOMEM);
2722
2723         OBD_ALLOC_PTR(env);
2724         if (env == NULL)
2725                 GOTO(out_thread, rc = -ENOMEM);
2726         rc = lu_env_add(env);
2727         if (rc)
2728                 GOTO(out_env, rc);
2729
2730         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2731         if (rc)
2732                 GOTO(out_env_remove, rc);
2733
2734         thread->t_env = env;
2735         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2736         thread->t_task = current;
2737         env->le_ctx.lc_thread = thread;
2738         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2739
2740         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2741                current->pid);
2742         trd->trd_processing_task = current->pid;
2743
2744         spin_lock(&obd->obd_dev_lock);
2745         obd->obd_recovering = 1;
2746         spin_unlock(&obd->obd_dev_lock);
2747         complete(&trd->trd_starting);
2748
2749         /* first of all, we have to know the first transno to replay */
2750         if (target_recovery_overseer(lut, check_for_recovery_ready,
2751                                      exp_connect_healthy)) {
2752                 abort_req_replay_queue(obd);
2753                 abort_lock_replay_queue(obd);
2754                 if (lut->lut_tdtd != NULL)
2755                         dtrq_list_destroy(lut->lut_tdtd);
2756         }
2757
2758         /* next stage: replay requests or update */
2759         delta = jiffies;
2760         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2761                atomic_read(&obd->obd_req_replay_clients),
2762                obd->obd_next_recovery_transno);
2763         replay_request_or_update(env, lut, trd, thread);
2764
2765         /**
2766          * The second stage: replay locks
2767          */
2768         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2769                atomic_read(&obd->obd_lock_replay_clients));
2770         while ((req = target_next_replay_lock(lut))) {
2771                 LASSERT(trd->trd_processing_task == current->pid);
2772                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2773                           libcfs_nid2str(req->rq_peer.nid));
2774                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2775                         req->rq_status = -ENODEV;
2776                         target_request_copy_put(req);
2777                         continue;
2778                 }
2779                 handle_recovery_req(thread, req,
2780                                     trd->trd_recovery_handler);
2781                 target_request_copy_put(req);
2782                 obd->obd_replayed_locks++;
2783         }
2784
2785         /**
2786          * The third stage: reply on final pings, at this moment all clients
2787          * must have request in final queue
2788          */
2789         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2790         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2791         /** Update server last boot epoch */
2792         tgt_boot_epoch_update(lut);
2793         /*
2794          * We drop recoverying flag to forward all new requests
2795          * to regular mds_handle() since now
2796          */
2797         spin_lock(&obd->obd_dev_lock);
2798         obd->obd_recovering = obd->obd_abort_recovery = 0;
2799         obd->obd_abort_recov_mdt = 0;
2800         spin_unlock(&obd->obd_dev_lock);
2801         spin_lock(&obd->obd_recovery_task_lock);
2802         target_cancel_recovery_timer(obd);
2803         spin_unlock(&obd->obd_recovery_task_lock);
2804         while ((req = target_next_final_ping(obd))) {
2805                 LASSERT(trd->trd_processing_task == current->pid);
2806                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2807                           libcfs_nid2str(req->rq_peer.nid));
2808                 handle_recovery_req(thread, req,
2809                                     trd->trd_recovery_handler);
2810                 /*
2811                  * Because the waiting client can not send ping to server,
2812                  * so we need refresh the last_request_time, to avoid the
2813                  * export is being evicted
2814                  */
2815                 ptlrpc_update_export_timer(req->rq_export, 0);
2816                 target_request_copy_put(req);
2817         }
2818
2819         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2820         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2821                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2822         if (delta > OBD_RECOVERY_TIME_SOFT) {
2823                 CWARN("too long recovery - read logs\n");
2824                 libcfs_debug_dumplog();
2825         }
2826
2827         target_finish_recovery(lut);
2828         lu_context_fini(&env->le_ctx);
2829         trd->trd_processing_task = 0;
2830         complete_all(&trd->trd_finishing);
2831         tgt_io_thread_done(thread);
2832 out_env_remove:
2833         lu_env_remove(env);
2834 out_env:
2835         OBD_FREE_PTR(env);
2836 out_thread:
2837         OBD_FREE_PTR(thread);
2838         RETURN(rc);
2839 }
2840
2841 static int target_start_recovery_thread(struct lu_target *lut,
2842                                         svc_handler_t handler)
2843 {
2844         struct obd_device *obd = lut->lut_obd;
2845         int rc = 0;
2846         struct target_recovery_data *trd = &obd->obd_recovery_data;
2847         int index;
2848
2849         memset(trd, 0, sizeof(*trd));
2850         init_completion(&trd->trd_starting);
2851         init_completion(&trd->trd_finishing);
2852         trd->trd_recovery_handler = handler;
2853
2854         rc = server_name2index(obd->obd_name, &index, NULL);
2855         if (rc < 0)
2856                 return rc;
2857
2858         if (!IS_ERR(kthread_run(target_recovery_thread,
2859                                 lut, "tgt_recover_%d", index))) {
2860                 wait_for_completion(&trd->trd_starting);
2861                 LASSERT(obd->obd_recovering != 0);
2862         } else {
2863                 rc = -ECHILD;
2864         }
2865
2866         return rc;
2867 }
2868
2869 void target_stop_recovery_thread(struct obd_device *obd)
2870 {
2871         if (obd->obd_recovery_data.trd_processing_task > 0) {
2872                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2873                 /** recovery can be done but postrecovery is not yet */
2874                 spin_lock(&obd->obd_dev_lock);
2875                 if (obd->obd_recovering) {
2876                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2877                         obd->obd_abort_recovery = 1;
2878                         wake_up(&obd->obd_next_transno_waitq);
2879                 }
2880                 spin_unlock(&obd->obd_dev_lock);
2881                 wait_for_completion(&trd->trd_finishing);
2882         }
2883 }
2884 EXPORT_SYMBOL(target_stop_recovery_thread);
2885
2886 void target_recovery_fini(struct obd_device *obd)
2887 {
2888         class_disconnect_exports(obd);
2889         target_stop_recovery_thread(obd);
2890         target_cleanup_recovery(obd);
2891 }
2892 EXPORT_SYMBOL(target_recovery_fini);
2893
2894 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2895 {
2896         struct obd_device *obd = container_of(timer, struct obd_device,
2897                                               obd_recovery_timer);
2898
2899         CDEBUG(D_HA,
2900                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2901                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2902                ktime_get_seconds() - obd->obd_recovery_start,
2903                atomic_read(&obd->obd_connected_clients));
2904
2905         obd->obd_recovery_expired = 1;
2906         wake_up(&obd->obd_next_transno_waitq);
2907         return HRTIMER_NORESTART;
2908 }
2909
2910 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2911 {
2912         struct obd_device *obd = lut->lut_obd;
2913
2914         if (lut->lut_bottom->dd_rdonly)
2915                 return;
2916
2917         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2918                 /** Update server last boot epoch */
2919                 tgt_boot_epoch_update(lut);
2920                 return;
2921         }
2922
2923         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2924                "last_transno %llu\n", obd->obd_name,
2925                atomic_read(&obd->obd_max_recoverable_clients),
2926                obd->obd_last_committed);
2927         LASSERT(obd->obd_stopping == 0);
2928         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2929         obd->obd_recovery_start = 0;
2930         obd->obd_recovery_end = 0;
2931
2932         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2933                      HRTIMER_MODE_ABS);
2934         obd->obd_recovery_timer.function = &target_recovery_expired;
2935         target_start_recovery_thread(lut, handler);
2936 }
2937 EXPORT_SYMBOL(target_recovery_init);
2938
2939 static int target_process_req_flags(struct obd_device *obd,
2940                                     struct ptlrpc_request *req)
2941 {
2942         struct obd_export *exp = req->rq_export;
2943
2944         LASSERT(exp != NULL);
2945         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2946                 /* client declares he's ready to replay locks */
2947                 spin_lock(&exp->exp_lock);
2948                 if (exp->exp_req_replay_needed) {
2949                         exp->exp_req_replay_needed = 0;
2950                         spin_unlock(&exp->exp_lock);
2951
2952                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2953                         atomic_dec(&obd->obd_req_replay_clients);
2954                 } else {
2955                         spin_unlock(&exp->exp_lock);
2956                 }
2957         }
2958         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2959                 /*
2960                  * client declares he's ready to complete recovery
2961                  * so, we put the request on th final queue
2962                  */
2963                 spin_lock(&exp->exp_lock);
2964                 if (exp->exp_lock_replay_needed) {
2965                         exp->exp_lock_replay_needed = 0;
2966                         spin_unlock(&exp->exp_lock);
2967
2968                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2969                         atomic_dec(&obd->obd_lock_replay_clients);
2970                 } else {
2971                         spin_unlock(&exp->exp_lock);
2972                 }
2973         }
2974         return 0;
2975 }
2976
2977 int target_queue_recovery_request(struct ptlrpc_request *req,
2978                                   struct obd_device *obd)
2979 {
2980         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2981         struct ptlrpc_request *reqiter;
2982         int inserted = 0;
2983
2984         ENTRY;
2985
2986         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2987                 /* Processing the queue right now, don't re-add. */
2988                 RETURN(1);
2989         }
2990
2991         target_process_req_flags(obd, req);
2992
2993         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2994                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2995                         if (cfs_fail_val == 1) {
2996                                 cfs_race_state = 1;
2997                                 cfs_fail_val = 0;
2998                                 wake_up(&cfs_race_waitq);
2999
3000                                 schedule_timeout_interruptible(
3001                                         cfs_time_seconds(1));
3002                         }
3003                 }
3004
3005                 /*
3006                  * client declares he's ready to complete recovery
3007                  * so, we put the request on th final queue
3008                  */
3009                 target_request_copy_get(req);
3010                 DEBUG_REQ(D_HA, req, "queue final req");
3011                 wake_up(&obd->obd_next_transno_waitq);
3012                 spin_lock(&obd->obd_recovery_task_lock);
3013                 if (obd->obd_recovering) {
3014                         struct ptlrpc_request *tmp;
3015                         struct ptlrpc_request *duplicate = NULL;
3016
3017                         if (likely(!req->rq_export->exp_replay_done)) {
3018                                 req->rq_export->exp_replay_done = 1;
3019                                 list_add_tail(&req->rq_list,
3020                                               &obd->obd_final_req_queue);
3021                                 spin_unlock(&obd->obd_recovery_task_lock);
3022                                 RETURN(0);
3023                         }
3024
3025                         /*
3026                          * XXX O(n), but only happens if final ping is
3027                          * timed out, probably reorganize the list as
3028                          * a hash list later
3029                          */
3030                         list_for_each_entry_safe(reqiter, tmp,
3031                                                  &obd->obd_final_req_queue,
3032                                                  rq_list) {
3033                                 if (reqiter->rq_export == req->rq_export) {
3034                                         list_del_init(&reqiter->rq_list);
3035                                         duplicate = reqiter;
3036                                         break;
3037                                 }
3038                         }
3039
3040                         list_add_tail(&req->rq_list,
3041                                       &obd->obd_final_req_queue);
3042                         req->rq_export->exp_replay_done = 1;
3043                         spin_unlock(&obd->obd_recovery_task_lock);
3044
3045                         if (duplicate != NULL) {
3046                                 DEBUG_REQ(D_HA, duplicate,
3047                                           "put prev final req");
3048                                 target_request_copy_put(duplicate);
3049                         }
3050                         RETURN(0);
3051                 } else {
3052                         spin_unlock(&obd->obd_recovery_task_lock);
3053                         target_request_copy_put(req);
3054                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3055                 }
3056         }
3057         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3058                 /* client declares he's ready to replay locks */
3059                 target_request_copy_get(req);
3060                 DEBUG_REQ(D_HA, req, "queue lock replay req");
3061                 wake_up(&obd->obd_next_transno_waitq);
3062                 spin_lock(&obd->obd_recovery_task_lock);
3063                 LASSERT(obd->obd_recovering);
3064                 /* usually due to recovery abort */
3065                 if (!req->rq_export->exp_in_recovery) {
3066                         spin_unlock(&obd->obd_recovery_task_lock);
3067                         target_request_copy_put(req);
3068                         RETURN(-ENOTCONN);
3069                 }
3070                 LASSERT(req->rq_export->exp_lock_replay_needed);
3071                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3072                 spin_unlock(&obd->obd_recovery_task_lock);
3073                 RETURN(0);
3074         }
3075
3076         /*
3077          * CAVEAT EMPTOR: The incoming request message has been swabbed
3078          * (i.e. buflens etc are in my own byte order), but type-dependent
3079          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3080          */
3081
3082         if (!transno) {
3083                 INIT_LIST_HEAD(&req->rq_list);
3084                 DEBUG_REQ(D_HA, req, "not queueing");
3085                 RETURN(1);
3086         }
3087
3088         /*
3089          * If we're processing the queue, we want don't want to queue this
3090          * message.
3091          *
3092          * Also, if this request has a transno less than the one we're waiting
3093          * for, we should process it now.  It could (and currently always will)
3094          * be an open request for a descriptor that was opened some time ago.
3095          *
3096          * Also, a resent, replayed request that has already been
3097          * handled will pass through here and be processed immediately.
3098          */
3099         CDEBUG(D_HA,
3100                "Next recovery transno: %llu, current: %llu, replaying\n",
3101                obd->obd_next_recovery_transno, transno);
3102
3103         /*
3104          * If the request has been replayed by update replay, then sends this
3105          * request to the recovery thread (replay_request_or_update()), where
3106          * it will be handled
3107          */
3108         spin_lock(&obd->obd_recovery_task_lock);
3109         if (transno < obd->obd_next_recovery_transno &&
3110             !is_req_replayed_by_update(req)) {
3111                 /* Processing the queue right now, don't re-add. */
3112                 LASSERT(list_empty(&req->rq_list));
3113                 spin_unlock(&obd->obd_recovery_task_lock);
3114                 RETURN(1);
3115         }
3116         spin_unlock(&obd->obd_recovery_task_lock);
3117
3118         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3119                 RETURN(0);
3120
3121         target_request_copy_get(req);
3122         if (!req->rq_export->exp_in_recovery) {
3123                 target_request_copy_put(req);
3124                 RETURN(-ENOTCONN);
3125         }
3126         LASSERT(req->rq_export->exp_req_replay_needed);
3127
3128         if (target_exp_enqueue_req_replay(req)) {
3129                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3130                 target_request_copy_put(req);
3131                 RETURN(0);
3132         }
3133
3134         /* XXX O(n^2) */
3135         spin_lock(&obd->obd_recovery_task_lock);
3136         LASSERT(obd->obd_recovering);
3137         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3138                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3139                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3140                         inserted = 1;
3141                         goto added;
3142                 }
3143
3144                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3145                              transno)) {
3146                         DEBUG_REQ(D_ERROR, req,
3147                                   "dropping replay: transno has been claimed by another client");
3148                         spin_unlock(&obd->obd_recovery_task_lock);
3149                         target_exp_dequeue_req_replay(req);
3150                         target_request_copy_put(req);
3151                         RETURN(0);
3152                 }
3153         }
3154 added:
3155         if (!inserted)
3156                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3157
3158         obd->obd_requests_queued_for_recovery++;
3159         spin_unlock(&obd->obd_recovery_task_lock);
3160         wake_up(&obd->obd_next_transno_waitq);
3161         RETURN(0);
3162 }
3163
3164 void target_committed_to_req(struct ptlrpc_request *req)
3165 {
3166         struct obd_export *exp = req->rq_export;
3167
3168         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3169                 lustre_msg_set_last_committed(req->rq_repmsg,
3170                                               exp->exp_last_committed);
3171         else
3172                 DEBUG_REQ(D_IOCTL, req,
3173                           "not sending last_committed update (%d/%d)",
3174                           exp->exp_obd->obd_no_transno,
3175                           req->rq_repmsg == NULL);
3176
3177         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3178                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3179 }
3180
3181 #endif /* HAVE_SERVER_SUPPORT */
3182
3183 /**
3184  * Packs current SLV and Limit into \a req.
3185  */
3186 int target_pack_pool_reply(struct ptlrpc_request *req)
3187 {
3188         struct obd_device *obd;
3189
3190         ENTRY;
3191
3192         /*
3193          * Check that we still have all structures alive as this may
3194          * be some late RPC at shutdown time.
3195          */
3196         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3197                      !exp_connect_lru_resize(req->rq_export))) {
3198                 lustre_msg_set_slv(req->rq_repmsg, 0);
3199                 lustre_msg_set_limit(req->rq_repmsg, 0);
3200                 RETURN(0);
3201         }
3202
3203         /* OBD is alive here as export is alive, which we checked above. */
3204         obd = req->rq_export->exp_obd;
3205
3206         read_lock(&obd->obd_pool_lock);
3207         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3208         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3209         read_unlock(&obd->obd_pool_lock);
3210
3211         RETURN(0);
3212 }
3213
3214 static int target_send_reply_msg(struct ptlrpc_request *req,
3215                                  int rc, int fail_id)
3216 {
3217         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3218                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3219                 return -ECOMM;
3220         }
3221         /*
3222          * We can have a null rq_reqmsg in the event of bad signature or
3223          * no context when unwrapping
3224          */
3225         if (req->rq_reqmsg &&
3226             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3227             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3228                 return -ECOMM;
3229
3230         if (unlikely(rc)) {
3231                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3232                 req->rq_status = rc;
3233                 return ptlrpc_send_error(req, 1);
3234         }
3235         DEBUG_REQ(D_NET, req, "sending reply");
3236
3237         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3238 }
3239
3240 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3241 {
3242         struct ptlrpc_service_part *svcpt;
3243         int netrc;
3244         struct ptlrpc_reply_state *rs;
3245         struct obd_export *exp;
3246
3247         ENTRY;
3248
3249         if (req->rq_no_reply) {
3250                 EXIT;
3251                 return;
3252         }
3253
3254         svcpt = req->rq_rqbd->rqbd_svcpt;
3255         rs = req->rq_reply_state;
3256         if (rs == NULL || !rs->rs_difficult) {
3257                 /* no notifiers */
3258                 target_send_reply_msg(req, rc, fail_id);
3259                 EXIT;
3260                 return;
3261         }
3262
3263         /* must be an export if locks saved */
3264         LASSERT(req->rq_export != NULL);
3265         /* req/reply consistent */
3266         LASSERT(rs->rs_svcpt == svcpt);
3267
3268         /* "fresh" reply */
3269         LASSERT(!rs->rs_scheduled);
3270         LASSERT(!rs->rs_scheduled_ever);
3271         LASSERT(!rs->rs_handled);
3272         LASSERT(!rs->rs_sent);
3273         LASSERT(!rs->rs_unlinked);
3274         LASSERT(rs->rs_export == NULL);
3275         LASSERT(list_empty(&rs->rs_obd_list));
3276         LASSERT(list_empty(&rs->rs_exp_list));
3277
3278         exp = class_export_get(req->rq_export);
3279
3280         /* disable reply scheduling while I'm setting up */
3281         rs->rs_scheduled = 1;
3282         rs->rs_sent      = 0;
3283         rs->rs_unlinked  = 0;
3284         rs->rs_xid       = req->rq_xid;
3285         rs->rs_transno   = req->rq_transno;
3286         rs->rs_export    = exp;
3287         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3288
3289         spin_lock(&exp->exp_uncommitted_replies_lock);
3290         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3291                rs->rs_transno, exp->exp_last_committed);
3292         if (rs->rs_transno > exp->exp_last_committed) {
3293                 /* not committed already */
3294                 list_add_tail(&rs->rs_obd_list,
3295                                   &exp->exp_uncommitted_replies);
3296         }
3297         spin_unlock(&exp->exp_uncommitted_replies_lock);
3298
3299         spin_lock(&exp->exp_lock);
3300         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3301         spin_unlock(&exp->exp_lock);
3302
3303         netrc = target_send_reply_msg(req, rc, fail_id);
3304
3305         spin_lock(&svcpt->scp_rep_lock);
3306
3307         atomic_inc(&svcpt->scp_nreps_difficult);
3308
3309         if (netrc != 0) {
3310                 /*
3311                  * error sending: reply is off the net.  Also we need +1
3312                  * reply ref until ptlrpc_handle_rs() is done
3313                  * with the reply state (if the send was successful, there
3314                  * would have been +1 ref for the net, which
3315                  * reply_out_callback leaves alone)
3316                  */
3317                 rs->rs_sent = 1;
3318                 rs->rs_unlinked = 1;
3319                 ptlrpc_rs_addref(rs);
3320         }
3321
3322         spin_lock(&rs->rs_lock);
3323         if (rs->rs_transno <= exp->exp_last_committed ||
3324             (rs->rs_unlinked && !rs->rs_no_ack) ||
3325             list_empty(&rs->rs_exp_list) ||     /* completed already */
3326             list_empty(&rs->rs_obd_list)) {
3327                 CDEBUG(D_HA, "Schedule reply immediately\n");
3328                 ptlrpc_dispatch_difficult_reply(rs);
3329         } else {
3330                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3331                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3332         }
3333         spin_unlock(&rs->rs_lock);
3334         spin_unlock(&svcpt->scp_rep_lock);
3335         EXIT;
3336 }
3337
3338 enum ldlm_mode lck_compat_array[] = {
3339         [LCK_EX]    = LCK_COMPAT_EX,
3340         [LCK_PW]    = LCK_COMPAT_PW,
3341         [LCK_PR]    = LCK_COMPAT_PR,
3342         [LCK_CW]    = LCK_COMPAT_CW,
3343         [LCK_CR]    = LCK_COMPAT_CR,
3344         [LCK_NL]    = LCK_COMPAT_NL,
3345         [LCK_GROUP] = LCK_COMPAT_GROUP,
3346         [LCK_COS]   = LCK_COMPAT_COS,
3347 };
3348
3349 /**
3350  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3351  * not escape to the user level.
3352  */
3353 int ldlm_error2errno(enum ldlm_error error)
3354 {
3355         int result;
3356
3357         switch (error) {
3358         case ELDLM_OK:
3359         case ELDLM_LOCK_MATCHED:
3360                 result = 0;
3361                 break;
3362         case ELDLM_LOCK_CHANGED:
3363                 result = -ESTALE;
3364                 break;
3365         case ELDLM_LOCK_ABORTED:
3366                 result = -ENAVAIL;
3367                 break;
3368         case ELDLM_LOCK_REPLACED:
3369                 result = -ESRCH;
3370                 break;
3371         case ELDLM_NO_LOCK_DATA:
3372                 result = -ENOENT;
3373                 break;
3374         case ELDLM_NAMESPACE_EXISTS:
3375                 result = -EEXIST;
3376                 break;
3377         case ELDLM_BAD_NAMESPACE:
3378                 result = -EBADF;
3379                 break;
3380         default:
3381                 if (((int)error) < 0) { /* cast to signed type */
3382                         result = error; /* as ldlm_error can be unsigned */
3383                 } else {
3384                         CERROR("Invalid DLM result code: %d\n", error);
3385                         result = -EPROTO;
3386                 }
3387         }
3388         return result;
3389 }
3390 EXPORT_SYMBOL(ldlm_error2errno);
3391
3392 /**
3393  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3394  */
3395 enum ldlm_error ldlm_errno2error(int err_no)
3396 {
3397         int error;
3398
3399         switch (err_no) {
3400         case 0:
3401                 error = ELDLM_OK;
3402                 break;
3403         case -ESTALE:
3404                 error = ELDLM_LOCK_CHANGED;
3405                 break;
3406         case -ENAVAIL:
3407                 error = ELDLM_LOCK_ABORTED;
3408                 break;
3409         case -ESRCH:
3410                 error = ELDLM_LOCK_REPLACED;
3411                 break;
3412         case -ENOENT:
3413                 error = ELDLM_NO_LOCK_DATA;
3414                 break;
3415         case -EEXIST:
3416                 error = ELDLM_NAMESPACE_EXISTS;
3417                 break;
3418         case -EBADF:
3419                 error = ELDLM_BAD_NAMESPACE;
3420                 break;
3421         default:
3422                 error = err_no;
3423         }
3424         return error;
3425 }
3426
3427 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3428 void ldlm_dump_export_locks(struct obd_export *exp)
3429 {
3430         spin_lock(&exp->exp_locks_list_guard);
3431         if (!list_empty(&exp->exp_locks_list)) {
3432                 struct ldlm_lock *lock;
3433
3434                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3435                        exp);
3436                 list_for_each_entry(lock, &exp->exp_locks_list,
3437                                         l_exp_refs_link)
3438                         LDLM_ERROR(lock, "lock:");
3439         }
3440         spin_unlock(&exp->exp_locks_list_guard);
3441 }
3442 #endif
3443
3444 #ifdef HAVE_SERVER_SUPPORT
3445 static inline const char *bulk2type(struct ptlrpc_request *req)
3446 {
3447         if (req->rq_bulk_read)
3448                 return "READ";
3449         if (req->rq_bulk_write)
3450                 return "WRITE";
3451         return "UNKNOWN";
3452 }
3453
3454 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3455 {
3456         struct ptlrpc_request *req = desc->bd_req;
3457         time64_t start = ktime_get_seconds();
3458         time64_t deadline;
3459         int rc = 0;
3460
3461         ENTRY;
3462
3463         /* If there is eviction in progress, wait for it to finish. */
3464         wait_event_idle(
3465                 exp->exp_obd->obd_evict_inprogress_waitq,
3466                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3467
3468         /* Check if client was evicted or reconnected already. */
3469         if (exp->exp_failed ||
3470             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3471                 rc = -ENOTCONN;
3472         } else {
3473                 if (req->rq_bulk_read)
3474                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3475
3476                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3477                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3478                 else /* old version, bulk matchbits is rq_xid */
3479                         req->rq_mbits = req->rq_xid;
3480
3481                 if (rc == 0)
3482                         rc = ptlrpc_start_bulk_transfer(desc);
3483         }
3484
3485         if (rc < 0) {
3486                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3487                           bulk2type(req), rc);
3488                 RETURN(rc);
3489         }
3490
3491         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3492                 ptlrpc_abort_bulk(desc);
3493                 RETURN(0);
3494         }
3495
3496         /* limit actual bulk transfer to bulk_timeout seconds */
3497         deadline = start + bulk_timeout;
3498         if (deadline > req->rq_deadline)
3499                 deadline = req->rq_deadline;
3500
3501         do {
3502                 time64_t timeoutl = deadline - ktime_get_seconds();
3503                 time64_t rq_deadline;
3504
3505                 while (timeoutl >= 0 &&
3506                        wait_event_idle_timeout(
3507                                desc->bd_waitq,
3508                                !ptlrpc_server_bulk_active(desc) ||
3509                                exp->exp_failed ||
3510                                exp->exp_conn_cnt >
3511                                lustre_msg_get_conn_cnt(req->rq_reqmsg),
3512                                timeoutl ? cfs_time_seconds(1) : 1) == 0)
3513                         timeoutl -= 1;
3514                 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3515
3516                 /* Wait again if we changed rq_deadline. */
3517                 rq_deadline = READ_ONCE(req->rq_deadline);
3518                 deadline = start + bulk_timeout;
3519                 if (deadline > rq_deadline)
3520                         deadline = rq_deadline;
3521         } while (rc == -ETIMEDOUT &&
3522                  deadline > ktime_get_seconds());
3523
3524         if (rc == -ETIMEDOUT) {
3525                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3526                           bulk2type(req), deadline - start,
3527                           ktime_get_real_seconds() - deadline);
3528                 ptlrpc_abort_bulk(desc);
3529         } else if (exp->exp_failed) {
3530                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3531                           bulk2type(req));
3532                 rc = -ENOTCONN;
3533                 ptlrpc_abort_bulk(desc);
3534         } else if (exp->exp_conn_cnt >
3535                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3536                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3537                           bulk2type(req));
3538                 /* We don't reply anyway. */
3539                 rc = -ETIMEDOUT;
3540                 ptlrpc_abort_bulk(desc);
3541         } else if (desc->bd_failure) {
3542                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3543                           bulk2type(req));
3544                 /* XXX should this be a different errno? */
3545                 rc = -ETIMEDOUT;
3546         } else {
3547                 if (req->rq_bulk_write)
3548                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3549                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3550                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3551                                   bulk2type(req), desc->bd_nob_transferred,
3552                                   desc->bd_nob);
3553                         /* XXX should this be a different errno? */
3554                         rc = -ETIMEDOUT;
3555                 }
3556         }
3557
3558         RETURN(rc);
3559 }
3560 EXPORT_SYMBOL(target_bulk_io);
3561
3562 #endif /* HAVE_SERVER_SUPPORT */