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/fs_struct.h>
  43 #include <linux/jiffies.h>
  44 #include <linux/kernel.h>
  45 #include <linux/kthread.h>
  46 #include <libcfs/libcfs.h>
  47 #include <obd.h>
  48 #include <obd_class.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /*
  55  * @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         u32 refnet = LNET_NET_ANY;
  64         int rc = 0;
  65
  66         ENTRY;
  67
  68         if (!create && !priority) {
  69                 CDEBUG(D_HA, "Nothing to do\n");
  70                 RETURN(-EINVAL);
  71         }
  72
  73         if (imp->imp_connection &&
  74             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  75                 /* refnet is used to restrict network connections */
  76                 refnet = LNET_NID_NET(&imp->imp_connection->c_self);
  77         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, refnet);
  78         if (!ptlrpc_conn) {
  79                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  80                 RETURN(-ENOENT);
  81         }
  82
  83         if (create) {
  84                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  85                 if (!imp_conn)
  86                         GOTO(out_put, rc = -ENOMEM);
  87         }
  88
  89         spin_lock(&imp->imp_lock);
  90         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  91                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  92                         if (priority) {
  93                                 list_move(&item->oic_item,
  94                                           &imp->imp_conn_list);
  95                                 item->oic_last_attempt = 0;
  96                         }
  97                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  98                                imp, imp->imp_obd->obd_name, uuid->uuid,
  99                                (priority ? ", moved to head" : ""));
 100                         spin_unlock(&imp->imp_lock);
 101                         GOTO(out_free, rc = 0);
 102                 }
 103         }
 104         /* No existing import connection found for \a uuid. */
 105         if (create) {
 106                 imp_conn->oic_conn = ptlrpc_conn;
 107                 imp_conn->oic_uuid = *uuid;
 108                 imp_conn->oic_last_attempt = 0;
 109                 if (priority)
 110                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 111                 else
 112                         list_add_tail(&imp_conn->oic_item,
 113                                       &imp->imp_conn_list);
 114                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 115                        imp, imp->imp_obd->obd_name, uuid->uuid,
 116                        (priority ? "head" : "tail"));
 117         } else {
 118                 spin_unlock(&imp->imp_lock);
 119                 GOTO(out_free, rc = -ENOENT);
 120         }
 121
 122         spin_unlock(&imp->imp_lock);
 123         RETURN(0);
 124 out_free:
 125         if (imp_conn)
 126                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 127 out_put:
 128         ptlrpc_connection_put(ptlrpc_conn);
 129         RETURN(rc);
 130 }
 131
 132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 133 {
 134         return import_set_conn(imp, uuid, 1, 0);
 135 }
 136
 137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 138                            int priority)
 139 {
 140         return import_set_conn(imp, uuid, priority, 1);
 141 }
 142 EXPORT_SYMBOL(client_import_add_conn);
 143
 144 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 145                                lnet_nid_t prim_nid, int priority)
 146 {
 147         struct ptlrpc_connection *ptlrpc_conn;
 148         int rc;
 149
 150         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, LNET_NIDNET(prim_nid));
 151         if (!ptlrpc_conn) {
 152                 const char *str_uuid = obd_uuid2str(uuid);
 153
 154                 rc = class_add_uuid(str_uuid, prim_nid);
 155                 if (rc) {
 156                         CERROR("%s: failed to add UUID '%s': rc = %d\n",
 157                                imp->imp_obd->obd_name, str_uuid, rc);
 158                         return rc;
 159                 }
 160         }
 161         return import_set_conn(imp, uuid, priority, 1);
 162 }
 163 EXPORT_SYMBOL(client_import_dyn_add_conn);
 164
 165 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
 166                                    int nid_count, struct obd_uuid *uuid)
 167 {
 168         struct obd_import_conn *conn;
 169         int rc = -ENOENT;
 170
 171         ENTRY;
 172         if (nid_count <= 0 || !nids)
 173                 return rc;
 174
 175         spin_lock(&imp->imp_lock);
 176         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 177                 if (class_check_uuid(&conn->oic_uuid, nids[0])) {
 178                         *uuid = conn->oic_uuid;
 179                         spin_unlock(&imp->imp_lock);
 180                         rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
 181                                                     nid_count);
 182                         RETURN(rc);
 183                 }
 184         }
 185         spin_unlock(&imp->imp_lock);
 186         RETURN(rc);
 187 }
 188 EXPORT_SYMBOL(client_import_add_nids_to_conn);
 189
 190 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 191 {
 192         struct obd_import_conn *imp_conn;
 193         struct obd_export *dlmexp;
 194         int rc = -ENOENT;
 195
 196         ENTRY;
 197
 198         spin_lock(&imp->imp_lock);
 199         if (list_empty(&imp->imp_conn_list)) {
 200                 LASSERT(!imp->imp_connection);
 201                 GOTO(out, rc);
 202         }
 203
 204         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 205                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 206                         continue;
 207                 LASSERT(imp_conn->oic_conn);
 208
 209                 if (imp_conn == imp->imp_conn_current) {
 210                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 211
 212                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 213                             imp->imp_state != LUSTRE_IMP_DISCON) {
 214                                 CERROR("can't remove current connection\n");
 215                                 GOTO(out, rc = -EBUSY);
 216                         }
 217
 218                         ptlrpc_connection_put(imp->imp_connection);
 219                         imp->imp_connection = NULL;
 220
 221                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 222                         if (dlmexp && dlmexp->exp_connection) {
 223                                 LASSERT(dlmexp->exp_connection ==
 224                                         imp_conn->oic_conn);
 225                                 ptlrpc_connection_put(dlmexp->exp_connection);
 226                                 dlmexp->exp_connection = NULL;
 227                         }
 228
 229                         if (dlmexp != NULL)
 230                                 class_export_put(dlmexp);
 231                 }
 232
 233                 list_del(&imp_conn->oic_item);
 234                 ptlrpc_connection_put(imp_conn->oic_conn);
 235                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 236                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 237                        imp, imp->imp_obd->obd_name, uuid->uuid);
 238                 rc = 0;
 239                 break;
 240         }
 241 out:
 242         spin_unlock(&imp->imp_lock);
 243         if (rc == -ENOENT)
 244                 CERROR("connection %s not found\n", uuid->uuid);
 245         RETURN(rc);
 246 }
 247 EXPORT_SYMBOL(client_import_del_conn);
 248
 249 /**
 250  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 251  * to find a conn uuid of \a imp which can reach \a peer.
 252  */
 253 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 254                             struct obd_uuid *uuid)
 255 {
 256         struct obd_import_conn *conn;
 257         int rc = -ENOENT;
 258
 259         ENTRY;
 260
 261         spin_lock(&imp->imp_lock);
 262         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 263                 /* Check if conn UUID does have this peer NID. */
 264                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 265                         *uuid = conn->oic_uuid;
 266                         rc = 0;
 267                         break;
 268                 }
 269         }
 270         spin_unlock(&imp->imp_lock);
 271         RETURN(rc);
 272 }
 273 EXPORT_SYMBOL(client_import_find_conn);
 274
 275 void client_destroy_import(struct obd_import *imp)
 276 {
 277         /*
 278          * Drop security policy instance after all RPCs have finished/aborted
 279          * to let all busy contexts be released.
 280          */
 281         class_import_get(imp);
 282         class_destroy_import(imp);
 283         sptlrpc_import_sec_put(imp);
 284         class_import_put(imp);
 285 }
 286 EXPORT_SYMBOL(client_destroy_import);
 287
 288 /**
 289  * Check whether or not the OSC is on MDT.
 290  * In the config log,
 291  * osc on MDT
 292  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 293  * osc on client
 294  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 295  *
 296  **/
 297 static int osc_on_mdt(char *obdname)
 298 {
 299         char *ptr;
 300
 301         ptr = strrchr(obdname, '-');
 302         if (ptr == NULL)
 303                 return 0;
 304
 305         if (strncmp(ptr + 1, "MDT", 3) == 0)
 306                 return 1;
 307
 308         return 0;
 309 }
 310
 311 /*
 312  * Configure an RPC client OBD device.
 313  *
 314  * lcfg parameters:
 315  * 1 - client UUID
 316  * 2 - server UUID
 317  * 3 - inactive-on-startup
 318  * 4 - restrictive net
 319  */
 320 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 321 {
 322         struct client_obd *cli = &obd->u.cli;
 323         struct obd_import *imp;
 324         struct obd_uuid server_uuid;
 325         int rq_portal, rp_portal, connect_op;
 326         const char *name = obd->obd_type->typ_name;
 327         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 328         char *cli_name = lustre_cfg_buf(lcfg, 0);
 329         struct ptlrpc_connection fake_conn = { .c_self = {},
 330                                                .c_remote_uuid.uuid[0] = 0 };
 331         int rc;
 332
 333         ENTRY;
 334
 335         /*
 336          * In a more perfect world, we would hang a ptlrpc_client off of
 337          * obd_type and just use the values from there.
 338          */
 339         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 340                 rq_portal = OST_REQUEST_PORTAL;
 341                 rp_portal = OSC_REPLY_PORTAL;
 342                 connect_op = OST_CONNECT;
 343                 cli->cl_sp_me = LUSTRE_SP_CLI;
 344                 cli->cl_sp_to = LUSTRE_SP_OST;
 345                 ns_type = LDLM_NS_TYPE_OSC;
 346         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 347                    !strcmp(name, LUSTRE_LWP_NAME)) {
 348                 rq_portal = MDS_REQUEST_PORTAL;
 349                 rp_portal = MDC_REPLY_PORTAL;
 350                 connect_op = MDS_CONNECT;
 351                 if (is_lwp_on_ost(cli_name))
 352                         cli->cl_sp_me = LUSTRE_SP_OST;
 353                 else if (is_lwp_on_mdt(cli_name))
 354                         cli->cl_sp_me = LUSTRE_SP_MDT;
 355                 else
 356                         cli->cl_sp_me = LUSTRE_SP_CLI;
 357                 cli->cl_sp_to = LUSTRE_SP_MDT;
 358                 ns_type = LDLM_NS_TYPE_MDC;
 359         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 360                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 361                         /* OSP_on_MDT for other MDTs */
 362                         connect_op = MDS_CONNECT;
 363                         cli->cl_sp_to = LUSTRE_SP_MDT;
 364                         ns_type = LDLM_NS_TYPE_MDC;
 365                         rq_portal = OUT_PORTAL;
 366                 } else {
 367                         /* OSP on MDT for OST */
 368                         connect_op = OST_CONNECT;
 369                         cli->cl_sp_to = LUSTRE_SP_OST;
 370                         ns_type = LDLM_NS_TYPE_OSC;
 371                         rq_portal = OST_REQUEST_PORTAL;
 372                 }
 373                 rp_portal = OSC_REPLY_PORTAL;
 374                 cli->cl_sp_me = LUSTRE_SP_MDT;
 375         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 376                 rq_portal = MGS_REQUEST_PORTAL;
 377                 rp_portal = MGC_REPLY_PORTAL;
 378                 connect_op = MGS_CONNECT;
 379                 cli->cl_sp_me = LUSTRE_SP_MGC;
 380                 cli->cl_sp_to = LUSTRE_SP_MGS;
 381                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 382                 ns_type = LDLM_NS_TYPE_MGC;
 383         } else {
 384                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 385                        name);
 386                 RETURN(-EINVAL);
 387         }
 388
 389         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 390                 CERROR("requires a TARGET UUID\n");
 391                 RETURN(-EINVAL);
 392         }
 393
 394         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 395                 CERROR("client UUID must be less than 38 characters\n");
 396                 RETURN(-EINVAL);
 397         }
 398
 399         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 400                 CERROR("setup requires a SERVER UUID\n");
 401                 RETURN(-EINVAL);
 402         }
 403
 404         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 405                 CERROR("target UUID must be less than 38 characters\n");
 406                 RETURN(-EINVAL);
 407         }
 408
 409         init_rwsem(&cli->cl_sem);
 410         mutex_init(&cli->cl_mgc_mutex);
 411         cli->cl_seq = NULL;
 412         init_rwsem(&cli->cl_seq_rwsem);
 413         cli->cl_conn_count = 0;
 414         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 415                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 416                      sizeof(server_uuid)));
 417
 418         cli->cl_dirty_pages = 0;
 419         cli->cl_dirty_max_pages = 0;
 420         cli->cl_avail_grant = 0;
 421         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 422         /*
 423          * cl_dirty_max_pages may be changed at connect time in
 424          * ptlrpc_connect_interpret().
 425          */
 426         client_adjust_max_dirty(cli);
 427         init_waitqueue_head(&cli->cl_cache_waiters);
 428         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 429         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 430         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 431         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 432         spin_lock_init(&cli->cl_loi_list_lock);
 433         atomic_set(&cli->cl_pending_w_pages, 0);
 434         atomic_set(&cli->cl_pending_r_pages, 0);
 435         cli->cl_r_in_flight = 0;
 436         cli->cl_w_in_flight = 0;
 437
 438         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 439         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 440         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 441         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 442         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 443         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 444
 445         /* lru for osc. */
 446         INIT_LIST_HEAD(&cli->cl_lru_osc);
 447         atomic_set(&cli->cl_lru_shrinkers, 0);
 448         atomic_long_set(&cli->cl_lru_busy, 0);
 449         atomic_long_set(&cli->cl_lru_in_list, 0);
 450         INIT_LIST_HEAD(&cli->cl_lru_list);
 451         spin_lock_init(&cli->cl_lru_list_lock);
 452         atomic_long_set(&cli->cl_unstable_count, 0);
 453         INIT_LIST_HEAD(&cli->cl_shrink_list);
 454         INIT_LIST_HEAD(&cli->cl_grant_chain);
 455
 456         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 457         cli->cl_rpcs_in_flight = 0;
 458
 459         init_waitqueue_head(&cli->cl_destroy_waitq);
 460         atomic_set(&cli->cl_destroy_in_flight, 0);
 461
 462
 463         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 464         cli->cl_preferred_cksum_type = 0;
 465 #ifdef ENABLE_CHECKSUM
 466         /* Turn on checksumming by default. */
 467         cli->cl_checksum = 1;
 468         /*
 469          * The supported checksum types will be worked out at connect time
 470          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 471          * through procfs.
 472          */
 473         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 474 #endif
 475         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 476
 477         /*
 478          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 479          * from OFD after connecting.
 480          */
 481         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 482
 483         cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
 484
 485         /*
 486          * set cl_chunkbits default value to PAGE_SHIFT,
 487          * it will be updated at OSC connection time.
 488          */
 489         cli->cl_chunkbits = PAGE_SHIFT;
 490
 491         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 492                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 493         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 494                 cli->cl_max_rpcs_in_flight = 2;
 495         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 496                 cli->cl_max_rpcs_in_flight = 3;
 497         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 498                 cli->cl_max_rpcs_in_flight = 4;
 499         } else {
 500                 if (osc_on_mdt(obd->obd_name))
 501                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 502                 else
 503                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 504         }
 505
 506         spin_lock_init(&cli->cl_mod_rpcs_lock);
 507         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 508         cli->cl_max_mod_rpcs_in_flight = 0;
 509         cli->cl_mod_rpcs_in_flight = 0;
 510         cli->cl_close_rpcs_in_flight = 0;
 511         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 512         cli->cl_mod_tag_bitmap = NULL;
 513
 514         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 515
 516         if (connect_op == MDS_CONNECT) {
 517                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 518                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 519                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 520                 if (cli->cl_mod_tag_bitmap == NULL)
 521                         GOTO(err, rc = -ENOMEM);
 522         }
 523
 524         rc = ldlm_get_ref();
 525         if (rc) {
 526                 CERROR("ldlm_get_ref failed: %d\n", rc);
 527                 GOTO(err, rc);
 528         }
 529
 530         ptlrpc_init_client(rq_portal, rp_portal, name,
 531                            &obd->obd_ldlm_client);
 532
 533         imp = class_new_import(obd);
 534         if (imp == NULL)
 535                 GOTO(err_ldlm, rc = -ENOENT);
 536         imp->imp_client = &obd->obd_ldlm_client;
 537         imp->imp_connect_op = connect_op;
 538         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 539                LUSTRE_CFG_BUFLEN(lcfg, 1));
 540         class_import_put(imp);
 541
 542         if (lustre_cfg_buf(lcfg, 4)) {
 543                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 544
 545                 if (refnet == LNET_NET_ANY) {
 546                         rc = -EINVAL;
 547                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 548                                obd->obd_name, lustre_cfg_string(lcfg, 4),
 549                                rc);
 550                         GOTO(err_import, rc);
 551                 }
 552                 lnet_nid4_to_nid(LNET_MKNID(refnet, 0), &fake_conn.c_self);
 553                 imp->imp_connection = &fake_conn;
 554         }
 555
 556         rc = client_import_add_conn(imp, &server_uuid, 1);
 557         if (rc) {
 558                 CERROR("can't add initial connection\n");
 559                 GOTO(err_import, rc);
 560         }
 561         imp->imp_connection = NULL;
 562
 563         cli->cl_import = imp;
 564         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 565         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 566
 567         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 568                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 569                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 570                                name, obd->obd_name,
 571                                cli->cl_target_uuid.uuid);
 572                         spin_lock(&imp->imp_lock);
 573                         imp->imp_deactive = 1;
 574                         spin_unlock(&imp->imp_lock);
 575                 }
 576         }
 577
 578         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 579                                                 LDLM_NAMESPACE_CLIENT,
 580                                                 LDLM_NAMESPACE_GREEDY,
 581                                                 ns_type);
 582         if (IS_ERR(obd->obd_namespace)) {
 583                 rc = PTR_ERR(obd->obd_namespace);
 584                 CERROR("%s: unable to create client namespace: rc = %d\n",
 585                        obd->obd_name, rc);
 586                 obd->obd_namespace = NULL;
 587                 GOTO(err_import, rc);
 588         }
 589
 590         RETURN(rc);
 591
 592 err_import:
 593         class_destroy_import(imp);
 594 err_ldlm:
 595         ldlm_put_ref();
 596 err:
 597         if (cli->cl_mod_tag_bitmap != NULL)
 598                 OBD_FREE(cli->cl_mod_tag_bitmap,
 599                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 600         cli->cl_mod_tag_bitmap = NULL;
 601
 602         RETURN(rc);
 603 }
 604 EXPORT_SYMBOL(client_obd_setup);
 605
 606 int client_obd_cleanup(struct obd_device *obd)
 607 {
 608         struct client_obd *cli = &obd->u.cli;
 609
 610         ENTRY;
 611
 612         ldlm_namespace_free_post(obd->obd_namespace);
 613         obd->obd_namespace = NULL;
 614
 615         obd_cleanup_client_import(obd);
 616         LASSERT(obd->u.cli.cl_import == NULL);
 617
 618         ldlm_put_ref();
 619
 620         if (cli->cl_mod_tag_bitmap != NULL)
 621                 OBD_FREE(cli->cl_mod_tag_bitmap,
 622                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 623         cli->cl_mod_tag_bitmap = NULL;
 624
 625         RETURN(0);
 626 }
 627 EXPORT_SYMBOL(client_obd_cleanup);
 628
 629 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 630 int client_connect_import(const struct lu_env *env,
 631                           struct obd_export **exp,
 632                           struct obd_device *obd, struct obd_uuid *cluuid,
 633                           struct obd_connect_data *data, void *localdata)
 634 {
 635         struct client_obd *cli = &obd->u.cli;
 636         struct obd_import *imp = cli->cl_import;
 637         struct obd_connect_data *ocd;
 638         struct lustre_handle conn = { 0 };
 639         int rc;
 640
 641         ENTRY;
 642
 643         *exp = NULL;
 644         down_write(&cli->cl_sem);
 645         if (cli->cl_conn_count > 0)
 646                 GOTO(out_sem, rc = -EALREADY);
 647
 648         rc = class_connect(&conn, obd, cluuid);
 649         if (rc)
 650                 GOTO(out_sem, rc);
 651
 652         cli->cl_conn_count++;
 653         *exp = class_conn2export(&conn);
 654
 655         LASSERT(obd->obd_namespace);
 656
 657         imp->imp_dlm_handle = conn;
 658         rc = ptlrpc_init_import(imp);
 659         if (rc != 0)
 660                 GOTO(out_ldlm, rc);
 661
 662         ocd = &imp->imp_connect_data;
 663         if (data) {
 664                 *ocd = *data;
 665                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 666                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 667         }
 668
 669         rc = ptlrpc_connect_import(imp);
 670         if (rc != 0) {
 671                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 672                 GOTO(out_ldlm, rc);
 673         }
 674         LASSERT(*exp != NULL && (*exp)->exp_connection);
 675
 676         if (data) {
 677                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 678                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 679                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 680                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 681                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 682         }
 683
 684         ptlrpc_pinger_add_import(imp);
 685
 686         EXIT;
 687
 688         if (rc) {
 689 out_ldlm:
 690                 cli->cl_conn_count--;
 691                 class_disconnect(*exp);
 692                 *exp = NULL;
 693         }
 694 out_sem:
 695         up_write(&cli->cl_sem);
 696
 697         if (!rc && localdata) {
 698                 LASSERT(cli->cl_cache == NULL); /* only once */
 699                 cli->cl_cache = (struct cl_client_cache *)localdata;
 700                 cl_cache_incref(cli->cl_cache);
 701                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 702
 703                 /* add this osc into entity list */
 704                 LASSERT(list_empty(&cli->cl_lru_osc));
 705                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 706                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 707                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 708         }
 709
 710         return rc;
 711 }
 712 EXPORT_SYMBOL(client_connect_import);
 713
 714 int client_disconnect_export(struct obd_export *exp)
 715 {
 716         struct obd_device *obd = class_exp2obd(exp);
 717         struct client_obd *cli;
 718         struct obd_import *imp;
 719         int rc = 0, err;
 720
 721         ENTRY;
 722
 723         if (!obd) {
 724                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 725                        exp, exp ? exp->exp_handle.h_cookie : -1);
 726                 RETURN(-EINVAL);
 727         }
 728
 729         cli = &obd->u.cli;
 730         imp = cli->cl_import;
 731
 732         down_write(&cli->cl_sem);
 733         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 734                 cli->cl_conn_count);
 735
 736         if (cli->cl_conn_count == 0) {
 737                 CERROR("disconnecting disconnected device (%s)\n",
 738                        obd->obd_name);
 739                 GOTO(out_disconnect, rc = -EINVAL);
 740         }
 741
 742         cli->cl_conn_count--;
 743         if (cli->cl_conn_count != 0)
 744                 GOTO(out_disconnect, rc = 0);
 745
 746         /*
 747          * Mark import deactivated now, so we don't try to reconnect if any
 748          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 749          * fully deactivate the import, or that would drop all requests.
 750          */
 751         spin_lock(&imp->imp_lock);
 752         imp->imp_deactive = 1;
 753         spin_unlock(&imp->imp_lock);
 754
 755         /*
 756          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 757          * delete it regardless.  (It's safe to delete an import that was
 758          * never added.)
 759          */
 760         (void)ptlrpc_pinger_del_import(imp);
 761
 762         if (obd->obd_namespace != NULL) {
 763                 /* obd_force == local only */
 764                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 765                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 766                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 767                                           obd->obd_force);
 768         }
 769
 770         /*
 771          * There's no need to hold sem while disconnecting an import,
 772          * and it may actually cause deadlock in GSS.
 773          */
 774         up_write(&cli->cl_sem);
 775         rc = ptlrpc_disconnect_import(imp, 0);
 776         down_write(&cli->cl_sem);
 777
 778         ptlrpc_invalidate_import(imp);
 779
 780         EXIT;
 781
 782 out_disconnect:
 783         /*
 784          * Use server style - class_disconnect should be always called for
 785          * o_disconnect.
 786          */
 787         err = class_disconnect(exp);
 788         if (!rc && err)
 789                 rc = err;
 790
 791         up_write(&cli->cl_sem);
 792
 793         RETURN(rc);
 794 }
 795 EXPORT_SYMBOL(client_disconnect_export);
 796
 797 #ifdef HAVE_SERVER_SUPPORT
 798 int server_disconnect_export(struct obd_export *exp)
 799 {
 800         int rc;
 801
 802         ENTRY;
 803
 804         /* Disconnect early so that clients can't keep using export. */
 805         rc = class_disconnect(exp);
 806         /* Close import to avoid sending any requests. */
 807         if (exp->exp_imp_reverse)
 808                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 809
 810         ldlm_bl_thread_wakeup();
 811
 812         /* complete all outstanding replies */
 813         spin_lock(&exp->exp_lock);
 814         while (!list_empty(&exp->exp_outstanding_replies)) {
 815                 struct ptlrpc_reply_state *rs =
 816                         list_entry(exp->exp_outstanding_replies.next,
 817                                        struct ptlrpc_reply_state, rs_exp_list);
 818                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 819
 820                 spin_lock(&svcpt->scp_rep_lock);
 821
 822                 list_del_init(&rs->rs_exp_list);
 823
 824                 spin_lock(&rs->rs_lock);
 825                 /* clear rs_convert_lock to make sure rs is handled and put */
 826                 rs->rs_convert_lock = 0;
 827                 ptlrpc_schedule_difficult_reply(rs);
 828                 spin_unlock(&rs->rs_lock);
 829
 830                 spin_unlock(&svcpt->scp_rep_lock);
 831         }
 832         spin_unlock(&exp->exp_lock);
 833
 834         RETURN(rc);
 835 }
 836 EXPORT_SYMBOL(server_disconnect_export);
 837
 838 static inline int target_check_recovery_timer(struct obd_device *target)
 839 {
 840         ktime_t remaining;
 841         s64 timeout;
 842
 843         if (!target->obd_recovering || target->obd_recovery_start == 0)
 844                 return 0;
 845
 846         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
 847         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 848         if (timeout > -30)
 849                 return 0;
 850
 851         /* the recovery timer should expire, but it isn't triggered,
 852          * it's better to abort the recovery of this target to speed up
 853          * the recovery of the whole cluster. */
 854         spin_lock(&target->obd_dev_lock);
 855         if (target->obd_recovering) {
 856                 CERROR("%s: Aborting recovery\n", target->obd_name);
 857                 target->obd_abort_recovery = 1;
 858                 wake_up(&target->obd_next_transno_waitq);
 859         }
 860         spin_unlock(&target->obd_dev_lock);
 861         return 0;
 862 }
 863
 864 /*
 865  * --------------------------------------------------------------------------
 866  * from old lib/target.c
 867  * --------------------------------------------------------------------------
 868  */
 869 static int target_handle_reconnect(struct lustre_handle *conn,
 870                                    struct obd_export *exp,
 871                                    struct obd_uuid *cluuid)
 872 {
 873         struct obd_device *target;
 874         struct lustre_handle *hdl;
 875         ktime_t remaining;
 876         s64 timeout;
 877         int rc = 0;
 878
 879         ENTRY;
 880         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 881         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 882                 conn->cookie = exp->exp_handle.h_cookie;
 883                 CDEBUG(D_HA,
 884                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 885                        cluuid->uuid, exp, conn->cookie);
 886                 RETURN(0);
 887         }
 888
 889         target = exp->exp_obd;
 890
 891         /* Might be a re-connect after a partition. */
 892         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 893                 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",
 894                               target->obd_name,
 895                               obd_uuid2str(&exp->exp_client_uuid),
 896                               obd_export_nid2str(exp),
 897                               hdl->cookie, conn->cookie);
 898                 memset(conn, 0, sizeof(*conn));
 899                 /*
 900                  * target_handle_connect() treats EALREADY and
 901                  * -EALREADY differently.  -EALREADY is an error
 902                  * (same UUID, different handle).
 903                  */
 904                 RETURN(-EALREADY);
 905         }
 906
 907         if (!target->obd_recovering) {
 908                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 909                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 910                         obd_export_nid2str(exp));
 911                 GOTO(out_already, rc);
 912         }
 913
 914         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
 915         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 916         if (timeout > 0) {
 917                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 918                               target->obd_name,
 919                               obd_uuid2str(&exp->exp_client_uuid),
 920                               obd_export_nid2str(exp),
 921                               atomic_read(&target->obd_max_recoverable_clients),
 922                               timeout / 60, timeout % 60);
 923         } else {
 924                 struct target_distribute_txn_data *tdtd;
 925                 int size = 0;
 926                 int count = 0;
 927                 char *buf = NULL;
 928
 929                 target_check_recovery_timer(target);
 930
 931                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 932                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 933                         buf = tdtd->tdtd_show_update_logs_retrievers(
 934                                 tdtd->tdtd_show_retrievers_cbdata,
 935                                 &size, &count);
 936
 937                 if (count > 0)
 938                         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",
 939                                       target->obd_name,
 940                                       obd_uuid2str(&exp->exp_client_uuid),
 941                                       obd_export_nid2str(exp), count,
 942                                       buf ? buf : "unknown (not enough RAM)",
 943                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 944                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 945                                       target->obd_name);
 946                 else
 947                         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",
 948                                       target->obd_name, abs(timeout) / 60,
 949                                       abs(timeout) % 60, target->obd_name);
 950
 951                 if (buf != NULL)
 952                         OBD_FREE(buf, size);
 953         }
 954
 955 out_already:
 956         conn->cookie = exp->exp_handle.h_cookie;
 957         /*
 958          * target_handle_connect() treats EALREADY and
 959          * -EALREADY differently.  EALREADY means we are
 960          * doing a valid reconnect from the same client.
 961          */
 962         RETURN(EALREADY);
 963 }
 964
 965 static void
 966 check_and_start_recovery_timer(struct obd_device *obd,
 967                                struct ptlrpc_request *req, int new_client);
 968
 969 /**
 970  * update flags for import during reconnect process
 971  */
 972 static int rev_import_flags_update(struct obd_import *revimp,
 973                                    struct ptlrpc_request *req)
 974 {
 975         int rc;
 976         struct obd_connect_data *data;
 977
 978         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 979
 980         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 981                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 982         else
 983                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 984
 985         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 986
 987         revimp->imp_connect_data = *data;
 988         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 989         if (rc) {
 990                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 991                         revimp->imp_client->cli_name,
 992                         libcfs_idstr(&req->rq_peer), rc);
 993                 return rc;
 994         }
 995
 996         return 0;
 997 }
 998
 999 /**
1000  * Allocate a new reverse import for an export.
1001  *
1002  * \retval -errno in case error hit
1003  * \retval 0 if reverse import correctly init
1004  **/
1005 int rev_import_init(struct obd_export *export)
1006 {
1007         struct obd_device *obd = export->exp_obd;
1008         struct obd_import *revimp;
1009
1010         LASSERT(export->exp_imp_reverse == NULL);
1011
1012         revimp = class_new_import(obd);
1013         if (revimp == NULL)
1014                 return -ENOMEM;
1015
1016         revimp->imp_remote_handle.cookie = 0ULL;
1017         revimp->imp_client = &obd->obd_ldlm_client;
1018         revimp->imp_dlm_fake = 1;
1019
1020         /* it is safe to connect import in new state as no sends possible */
1021         spin_lock(&export->exp_lock);
1022         export->exp_imp_reverse = revimp;
1023         spin_unlock(&export->exp_lock);
1024         class_import_put(revimp);
1025
1026         return 0;
1027 }
1028 EXPORT_SYMBOL(rev_import_init);
1029
1030 /**
1031  * Handle reconnect for an export.
1032  *
1033  * \param exp export to handle reconnect process
1034  * \param req client reconnect request
1035  *
1036  * \retval -rc in case securitfy flavor can't be changed
1037  * \retval 0 in case none problems
1038  */
1039 static int rev_import_reconnect(struct obd_export *exp,
1040                                 struct ptlrpc_request *req)
1041 {
1042         struct obd_import *revimp = exp->exp_imp_reverse;
1043         struct lustre_handle *lh;
1044         int rc;
1045
1046         /* avoid sending a request until import flags are changed */
1047         ptlrpc_import_enter_resend(revimp);
1048
1049         ptlrpc_connection_put(revimp->imp_connection);
1050
1051         /*
1052          * client from recovery don't have a handle so we need to take from
1053          * request. it may produce situation when wrong client connected
1054          * to recovery as we trust a client uuid
1055          */
1056         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1057         revimp->imp_remote_handle = *lh;
1058
1059         /*
1060          * unknown versions will be caught in
1061          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1062          */
1063         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1064
1065         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1066
1067         rc = rev_import_flags_update(revimp, req);
1068         if (rc != 0) {
1069                 /*
1070                  * it is safe to still be in RECOVERY phase as we are not able
1071                  * to setup correct security flavor so requests are not able to
1072                  * be delivered correctly
1073                  */
1074                 return rc;
1075         }
1076
1077         /* resend all rpc's via new connection */
1078         return ptlrpc_import_recovery_state_machine(revimp);
1079 }
1080
1081 int target_handle_connect(struct ptlrpc_request *req)
1082 {
1083         struct obd_device *target = NULL;
1084         struct obd_export *export = NULL;
1085         /*
1086          * connect handle - filled from target_handle_reconnect in
1087          * reconnect case
1088          */
1089         struct lustre_handle conn;
1090         struct lustre_handle *tmp;
1091         struct obd_uuid cluuid;
1092         char *str;
1093         int rc = 0;
1094         char *target_start;
1095         int target_len;
1096         bool mds_conn = false, lw_client = false, initial_conn = false;
1097         bool mds_mds_conn = false;
1098         bool new_mds_mds_conn = false;
1099         struct obd_connect_data *data, *tmpdata;
1100         int size, tmpsize;
1101         lnet_nid_t client_nid;
1102         struct ptlrpc_connection *pcon = NULL;
1103
1104         ENTRY;
1105
1106         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1107
1108         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1109         if (str == NULL) {
1110                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1111                 GOTO(out, rc = -EINVAL);
1112         }
1113
1114         target = class_dev_by_str(str);
1115         if (!target) {
1116                 deuuidify(str, NULL, &target_start, &target_len);
1117                 LCONSOLE_ERROR_MSG(0x137,
1118                                    "%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",
1119                                    str, libcfs_nidstr(&req->rq_peer.nid));
1120                 GOTO(out, rc = -ENODEV);
1121         }
1122
1123         spin_lock(&target->obd_dev_lock);
1124
1125         target->obd_conn_inprogress++;
1126
1127         if (target->obd_stopping || !target->obd_set_up) {
1128                 spin_unlock(&target->obd_dev_lock);
1129
1130                 deuuidify(str, NULL, &target_start, &target_len);
1131                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1132                               target_len, target_start,
1133                               libcfs_nidstr(&req->rq_peer.nid),
1134                               (target->obd_stopping ?
1135                                "stopping" : "not set up"));
1136                 GOTO(out, rc = -ENODEV);
1137         }
1138
1139         if (target->obd_no_conn) {
1140                 spin_unlock(&target->obd_dev_lock);
1141
1142                 CDEBUG(D_INFO,
1143                        "%s: Temporarily refusing client connection from %s\n",
1144                        target->obd_name, libcfs_nidstr(&req->rq_peer.nid));
1145                 GOTO(out, rc = -EAGAIN);
1146         }
1147
1148         spin_unlock(&target->obd_dev_lock);
1149
1150         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1151         if (str == NULL) {
1152                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1153                 GOTO(out, rc = -EINVAL);
1154         }
1155
1156         obd_str2uuid(&cluuid, str);
1157
1158         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1159         if (tmp == NULL)
1160                 GOTO(out, rc = -EPROTO);
1161
1162         conn = *tmp;
1163
1164         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1165                                     RCL_CLIENT);
1166         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1167                 GOTO(out, rc = -EPROTO);
1168         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1169         if (!data)
1170                 GOTO(out, rc = -EPROTO);
1171
1172         rc = req_capsule_server_pack(&req->rq_pill);
1173         if (rc)
1174                 GOTO(out, rc);
1175
1176 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1177         /*
1178          * Don't allow clients to connect that are using old 1.8 format
1179          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1180          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1181          * FULL20 flag should be set on all connections since 2.0, but no
1182          * longer affects behaviour.
1183          *
1184          * Later this check will be disabled and the flag can be retired
1185          * completely once interop with 3.0 is no longer needed.
1186          */
1187         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1188                 GOTO(out, rc = -EPROTO);
1189
1190         /*
1191          * Don't allow liblustre clients to connect.
1192          * - testing was disabled in v2_2_50_0-61-g6a75d65
1193          * - building was disabled in v2_5_58_0-28-g7277179
1194          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1195          * - clients were refused connect for version difference > 0.0.1.32
1196          */
1197         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1198                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1199                 GOTO(out, rc = -EPROTO);
1200         }
1201 #endif
1202
1203         /*
1204          * Note: lw_client is needed in MDS-MDS failover during update log
1205          * processing, so we needs to allow lw_client to be connected at
1206          * anytime, instead of only the initial connection
1207          */
1208         lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1209
1210         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1211                 initial_conn = true;
1212                 mds_conn = OCD_HAS_FLAG(data, MDS);
1213                 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1214
1215                 /*
1216                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1217                  * Checking OBD_CONNECT_FID can be removed in the future.
1218                  *
1219                  * Via check OBD_CONNECT_FID, we can distinguish whether
1220                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1221                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1222                  */
1223                 if (!lw_client &&
1224                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1225                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1226                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1227                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1228                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1229                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1230
1231                         /*
1232                          * We do not support the MDT-MDT interoperations with
1233                          * different version MDT because of protocol changes.
1234                          */
1235                         if (unlikely(major != LUSTRE_MAJOR ||
1236                                      minor != LUSTRE_MINOR ||
1237                                      abs(patch - LUSTRE_PATCH) > 3)) {
1238                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1239                                               target->obd_name, LUSTRE_MAJOR,
1240                                               LUSTRE_MINOR, LUSTRE_PATCH,
1241                                               LUSTRE_FIX, major, minor, patch,
1242                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1243                                               libcfs_nidstr(&req->rq_peer.nid),
1244                                               str);
1245                                 GOTO(out, rc = -EPROTO);
1246                         }
1247                 }
1248         }
1249
1250         /* lctl gets a backstage, all-access pass. */
1251         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1252                 goto dont_check_exports;
1253
1254         export = obd_uuid_lookup(target, &cluuid);
1255         if (!export)
1256                 goto no_export;
1257
1258         /* We've found an export in the hash. */
1259
1260         spin_lock(&export->exp_lock);
1261
1262         if (export->exp_connecting) { /* b=9635, et. al. */
1263                 spin_unlock(&export->exp_lock);
1264                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1265                               export->exp_obd->obd_name, export,
1266                               libcfs_nidstr(&req->rq_peer.nid));
1267                 class_export_put(export);
1268                 export = NULL;
1269                 rc = -EALREADY;
1270         } else if ((mds_conn || (lw_client && initial_conn) ||
1271                    OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1272                 spin_unlock(&export->exp_lock);
1273                 if (!nid_same(&req->rq_peer.nid,
1274                               &export->exp_connection->c_peer.nid)) {
1275                         /* MDS or LWP reconnected after failover. */
1276                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1277                                       target->obd_name,
1278                                       lw_client ? "LWP" : "MDS",
1279                                       libcfs_nidstr(&req->rq_peer.nid),
1280                                       libcfs_nidstr(&export->exp_connection->c_peer.nid));
1281                 } else {
1282                         /* New connection from the same NID. */
1283                         LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1284                                       target->obd_name,
1285                                       lw_client ? "LWP" : "MDS",
1286                                       libcfs_nidstr(&req->rq_peer.nid),
1287                                       OCD_HAS_FLAG(data, MDS_MDS) ?
1288                                       "keep" : "remove");
1289                 }
1290
1291                 if (nid_same(&req->rq_peer.nid,
1292                              &export->exp_connection->c_peer.nid) &&
1293                     OCD_HAS_FLAG(data, MDS_MDS)) {
1294                         /*
1295                          * Because exports between MDTs will always be
1296                          * kept, let's do not fail such export if they
1297                          * come from the same NID, otherwise it might
1298                          * cause eviction between MDTs, which might
1299                          * cause namespace inconsistency
1300                          */
1301                         spin_lock(&export->exp_lock);
1302                         export->exp_connecting = 1;
1303                         export->exp_conn_cnt = 0;
1304                         spin_unlock(&export->exp_lock);
1305                         conn.cookie = export->exp_handle.h_cookie;
1306                         rc = EALREADY;
1307                 } else {
1308                         class_fail_export(export);
1309                         class_export_put(export);
1310                         export = NULL;
1311                         rc = 0;
1312                 }
1313         } else if (export->exp_connection != NULL && initial_conn &&
1314                    !nid_same(&req->rq_peer.nid,
1315                              &export->exp_connection->c_peer.nid)) {
1316                 spin_unlock(&export->exp_lock);
1317                 /* In MDS failover we have static UUID but NID can change. */
1318                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1319                               target->obd_name, cluuid.uuid,
1320                               libcfs_nidstr(&req->rq_peer.nid),
1321                               libcfs_nidstr(
1322                                       &export->exp_connection->c_peer.nid));
1323                 rc = -EALREADY;
1324                 class_export_put(export);
1325                 export = NULL;
1326         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1327                    !lw_client) {
1328                 spin_unlock(&export->exp_lock);
1329                 rc = -EAGAIN;
1330         } else {
1331                 export->exp_connecting = 1;
1332                 spin_unlock(&export->exp_lock);
1333                 LASSERT(export->exp_obd == target);
1334
1335                 rc = target_handle_reconnect(&conn, export, &cluuid);
1336         }
1337
1338         /* If we found an export, we already unlocked. */
1339         if (!export) {
1340 no_export:
1341                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1342         } else if (req->rq_export == NULL &&
1343                    atomic_read(&export->exp_rpc_count) > 0) {
1344                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1345                               target->obd_name, cluuid.uuid,
1346                               libcfs_nidstr(&req->rq_peer.nid),
1347                               refcount_read(&export->exp_handle.h_ref));
1348                         GOTO(out, rc = -EBUSY);
1349         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1350                    rc != EALREADY) {
1351                 if (!strstr(cluuid.uuid, "mdt"))
1352                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1353                                       target->obd_name, cluuid.uuid,
1354                                       libcfs_nidstr(&req->rq_peer.nid));
1355                 GOTO(out, rc = -EALREADY);
1356         } else {
1357                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1358         }
1359
1360         if (rc < 0)
1361                 GOTO(out, rc);
1362
1363         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1364                target->obd_name, cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1365                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1366                export, ktime_get_seconds(),
1367                export ? export->exp_last_request_time : 0);
1368
1369         /*
1370          * If this is the first time a client connects, reset the recovery
1371          * timer. Discard lightweight connections which might be local.
1372          */
1373         if (!lw_client && rc == 0 && target->obd_recovering)
1374                 check_and_start_recovery_timer(target, req, export == NULL);
1375
1376         /*
1377          * We want to handle EALREADY but *not* -EALREADY from
1378          * target_handle_reconnect(), return reconnection state in a flag.
1379          */
1380         if (rc == EALREADY) {
1381                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1382                 rc = 0;
1383         } else {
1384                 LASSERT(rc == 0);
1385         }
1386
1387         /* Tell the client if we support replayable requests. */
1388         if (target->obd_replayable)
1389                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1390         client_nid = lnet_nid_to_nid4(&req->rq_peer.nid);
1391
1392         if (export == NULL) {
1393                 /* allow lightweight connections during recovery */
1394                 /*
1395                  * allow "new" MDT to be connected during recovery, since we
1396                  * need retrieve recovery update records from it
1397                  */
1398                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1399                         struct hrtimer *timer = &target->obd_recovery_timer;
1400                         ktime_t remaining;
1401                         s64 timeout, left;
1402                         int in_progress;
1403                         int connected;
1404                         int known;
1405                         int stale;
1406                         char *msg;
1407
1408                         connected = atomic_read(&target->obd_connected_clients);
1409                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1410                         known =
1411                            atomic_read(&target->obd_max_recoverable_clients);
1412                         stale = target->obd_stale_clients;
1413                         remaining = hrtimer_get_remaining(timer);
1414                         left = ktime_divns(remaining, NSEC_PER_SEC);
1415
1416                         if (ktime_to_ns(remaining) > 0) {
1417                                 msg = "to recover in";
1418                                 timeout = left;
1419                         } else {
1420                                 msg = "already passed deadline";
1421                                 timeout = -left;
1422
1423                                 target_check_recovery_timer(target);
1424                         }
1425
1426                         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",
1427                                       target->obd_name, cluuid.uuid,
1428                                       libcfs_nidstr(&req->rq_peer.nid), known,
1429                                       connected - in_progress, in_progress,
1430                                       stale, msg, timeout / 60, timeout % 60);
1431                         rc = -EBUSY;
1432                 } else {
1433 dont_check_exports:
1434                         rc = obd_connect(req->rq_svc_thread->t_env,
1435                                          &export, target, &cluuid, data,
1436                                          &client_nid);
1437                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1438                                 lustre_msg_add_op_flags(req->rq_repmsg,
1439                                                         MSG_CONNECT_RECOVERING);
1440                         if (rc == 0) {
1441                                 conn.cookie = export->exp_handle.h_cookie;
1442                                 rc = rev_import_init(export);
1443                         }
1444
1445                         if (mds_mds_conn)
1446                                 new_mds_mds_conn = true;
1447                 }
1448         } else {
1449                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1450                                    export, target, &cluuid, data, &client_nid);
1451         }
1452         if (rc)
1453                 GOTO(out, rc);
1454
1455         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1456         data->ocd_instance = target->u.obt.obt_instance;
1457
1458         /*
1459          * Return only the parts of obd_connect_data that we understand, so the
1460          * client knows that we don't understand the rest.
1461          */
1462         if (data) {
1463                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1464                                                RCL_SERVER);
1465                 tmpdata = req_capsule_server_get(&req->rq_pill,
1466                                                  &RMF_CONNECT_DATA);
1467                 /*
1468                  * Don't use struct assignment here, because the client reply
1469                  * buffer may be smaller/larger than the local struct
1470                  * obd_connect_data.
1471                  */
1472                 memcpy(tmpdata, data, min(tmpsize, size));
1473         }
1474
1475         /*
1476          * If the client and the server are the same node, we will already
1477          * have an export that really points to the client's DLM export,
1478          * because we have a shared handles table.
1479          *
1480          * XXX this will go away when shaver stops sending the "connect" handle
1481          * in the real "remote handle" field of the request --phik 24 Apr 2003
1482          */
1483         ptlrpc_request_change_export(req, export);
1484
1485         pcon = ptlrpc_connection_get(&req->rq_peer,
1486                                      &req->rq_self, &cluuid);
1487         if (pcon == NULL)
1488                 GOTO(out, rc = -ENOTCONN);
1489
1490         spin_lock(&export->exp_lock);
1491
1492         if (export->exp_disconnected) {
1493                 spin_unlock(&export->exp_lock);
1494                 GOTO(out, rc = -ENODEV);
1495         }
1496         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1497                 spin_unlock(&export->exp_lock);
1498                 CDEBUG(D_RPCTRACE,
1499                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1500                        cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1501                        export->exp_conn_cnt,
1502                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1503
1504                 GOTO(out, rc = -EALREADY);
1505         }
1506         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1507         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1508
1509         /* Check to see if connection came from another NID. */
1510         if (export->exp_connection != NULL &&
1511             !nid_same(&export->exp_connection->c_peer.nid,
1512                       &req->rq_peer.nid)) {
1513                 obd_nid_del(export->exp_obd, export);
1514                 ptlrpc_connection_put(export->exp_connection);
1515                 export->exp_connection = NULL;
1516         }
1517
1518         if (export->exp_connection == NULL) {
1519                 export->exp_connection = pcon;
1520                 pcon = NULL;
1521         }
1522         obd_nid_add(export->exp_obd, export);
1523
1524         spin_unlock(&export->exp_lock);
1525
1526         lustre_msg_set_handle(req->rq_repmsg, &conn);
1527
1528         rc = rev_import_reconnect(export, req);
1529         if (rc != 0)
1530                 GOTO(out, rc);
1531
1532         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1533                 int has_transno;
1534                 __u64 transno = data->ocd_transno;
1535
1536                 spin_lock(&export->exp_lock);
1537                 /*
1538                  * possible race with class_disconnect_stale_exports,
1539                  * export may be already in the eviction process
1540                  */
1541                 if (export->exp_failed) {
1542                         spin_unlock(&export->exp_lock);
1543                         GOTO(out, rc = -ENODEV);
1544                 }
1545                 export->exp_in_recovery = 1;
1546                 export->exp_req_replay_needed = 1;
1547                 export->exp_lock_replay_needed = 1;
1548                 spin_unlock(&export->exp_lock);
1549
1550                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1551                                  MSG_CONNECT_TRANSNO);
1552                 if (has_transno && transno == 0)
1553                         CWARN("Connect with zero transno!\n");
1554
1555                 if (has_transno && transno > 0 &&
1556                     transno < target->obd_next_recovery_transno &&
1557                     transno > target->obd_last_committed) {
1558                         /* Another way is to use cmpxchg() to be lock-free. */
1559                         spin_lock(&target->obd_recovery_task_lock);
1560                         if (transno < target->obd_next_recovery_transno)
1561                                 target->obd_next_recovery_transno = transno;
1562                         spin_unlock(&target->obd_recovery_task_lock);
1563                 }
1564
1565                 atomic_inc(&target->obd_req_replay_clients);
1566                 atomic_inc(&target->obd_lock_replay_clients);
1567                 /*
1568                  * Note: MDS-MDS connection is allowed to be connected during
1569                  * recovery, no matter if the exports needs to be recoveried.
1570                  * Because we need retrieve updates logs from all other MDTs.
1571                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1572                  * also needs to be increased to match other recovery checking
1573                  * condition.
1574                  */
1575                 if (new_mds_mds_conn)
1576                         atomic_inc(&target->obd_max_recoverable_clients);
1577
1578                 if (atomic_inc_return(&target->obd_connected_clients) ==
1579                     atomic_read(&target->obd_max_recoverable_clients))
1580                         wake_up(&target->obd_next_transno_waitq);
1581         }
1582
1583         /* Tell the client we're in recovery, when client is involved in it. */
1584         if (target->obd_recovering && !lw_client)
1585                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1586
1587 out:
1588         if (export) {
1589                 spin_lock(&export->exp_lock);
1590                 export->exp_connecting = 0;
1591                 spin_unlock(&export->exp_lock);
1592
1593                 class_export_put(export);
1594         }
1595         if (target != NULL) {
1596                 spin_lock(&target->obd_dev_lock);
1597                 target->obd_conn_inprogress--;
1598                 spin_unlock(&target->obd_dev_lock);
1599                 class_decref(target, "find", current);
1600         }
1601         if (pcon)
1602                 ptlrpc_connection_put(pcon);
1603         req->rq_status = rc;
1604         RETURN(rc);
1605 }
1606
1607 int target_handle_disconnect(struct ptlrpc_request *req)
1608 {
1609         int rc;
1610
1611         ENTRY;
1612
1613         rc = req_capsule_server_pack(&req->rq_pill);
1614         if (rc)
1615                 RETURN(rc);
1616
1617         /* In case of target disconnect, updating sec ctx immediately is
1618          * required in order to record latest sequence number used.
1619          * Sequence is normally updated on export destroy, but this event
1620          * can occur too late, ie after a new target connect request has
1621          * been processed.
1622          * Maintaining correct sequence when client connection becomes idle
1623          * ensures that GSS does not erroneously consider requests as replays.
1624          */
1625         rc = sptlrpc_export_update_ctx(req->rq_export);
1626         if (rc)
1627                 RETURN(rc);
1628
1629         /* Keep the rq_export around so we can send the reply. */
1630         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1631
1632         RETURN(0);
1633 }
1634
1635 void target_destroy_export(struct obd_export *exp)
1636 {
1637         struct obd_import *imp = NULL;
1638         /*
1639          * exports created from last_rcvd data, and "fake"
1640          * exports created by lctl don't have an import
1641          */
1642         spin_lock(&exp->exp_lock);
1643         if (exp->exp_imp_reverse != NULL) {
1644                 imp = exp->exp_imp_reverse;
1645                 exp->exp_imp_reverse = NULL;
1646         }
1647         spin_unlock(&exp->exp_lock);
1648         if (imp != NULL)
1649                 client_destroy_import(imp);
1650
1651         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1652         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1653         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1654         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1655 }
1656 EXPORT_SYMBOL(target_destroy_export);
1657
1658 /*
1659  * Recovery functions
1660  */
1661 static void target_request_copy_get(struct ptlrpc_request *req)
1662 {
1663         class_export_rpc_inc(req->rq_export);
1664         LASSERT(list_empty(&req->rq_list));
1665         INIT_LIST_HEAD(&req->rq_replay_list);
1666
1667         /* Increase refcount to keep request in queue. */
1668         atomic_inc(&req->rq_refcount);
1669         /* Let export know it has replays to be handled. */
1670         atomic_inc(&req->rq_export->exp_replay_count);
1671 }
1672
1673 static void target_request_copy_put(struct ptlrpc_request *req)
1674 {
1675         LASSERT(list_empty(&req->rq_replay_list));
1676         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1677
1678         atomic_dec(&req->rq_export->exp_replay_count);
1679         class_export_rpc_dec(req->rq_export);
1680         ptlrpc_server_drop_request(req);
1681 }
1682
1683 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1684 {
1685         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1686         struct obd_export *exp = req->rq_export;
1687         struct ptlrpc_request *reqiter;
1688         struct ptlrpc_request *dup_req = NULL;
1689         int dup = 0;
1690
1691         LASSERT(exp);
1692
1693         spin_lock(&exp->exp_lock);
1694         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1695                             rq_replay_list) {
1696                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1697                         dup_req = reqiter;
1698                         dup = 1;
1699                         break;
1700                 }
1701         }
1702
1703         if (dup) {
1704                 /* We expect it with RESENT and REPLAY flags. */
1705                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1706                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1707                         CERROR("invalid flags %x of resent replay\n",
1708                                lustre_msg_get_flags(req->rq_reqmsg));
1709
1710                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1711                         __u32 new_conn;
1712
1713                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1714                         if (new_conn >
1715                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1716                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1717                                                         new_conn);
1718                 }
1719         } else {
1720                 list_add_tail(&req->rq_replay_list,
1721                               &exp->exp_req_replay_queue);
1722         }
1723
1724         spin_unlock(&exp->exp_lock);
1725         return dup;
1726 }
1727
1728 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1729 {
1730         LASSERT(!list_empty(&req->rq_replay_list));
1731         LASSERT(req->rq_export);
1732
1733         spin_lock(&req->rq_export->exp_lock);
1734         list_del_init(&req->rq_replay_list);
1735         spin_unlock(&req->rq_export->exp_lock);
1736 }
1737
1738 static void target_finish_recovery(struct lu_target *lut)
1739 {
1740         struct obd_device *obd = lut->lut_obd;
1741
1742         ENTRY;
1743
1744         /* Only log a recovery message when recovery has occurred. */
1745         if (obd->obd_recovery_start) {
1746                 time64_t now = ktime_get_seconds();
1747                 time64_t elapsed_time;
1748
1749                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1750                                      1);
1751                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1752                               obd->obd_name, elapsed_time / 60,
1753                               elapsed_time % 60,
1754                               atomic_read(&obd->obd_max_recoverable_clients),
1755                               atomic_read(&obd->obd_connected_clients),
1756                               obd->obd_stale_clients,
1757                               obd->obd_stale_clients == 1 ? "was" : "were");
1758                 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1759                         libcfs_debug_dumplog();
1760         }
1761
1762         ldlm_reprocess_recovery_done(obd->obd_namespace);
1763         spin_lock(&obd->obd_recovery_task_lock);
1764         if (!list_empty(&obd->obd_req_replay_queue) ||
1765             !list_empty(&obd->obd_lock_replay_queue) ||
1766             !list_empty(&obd->obd_final_req_queue)) {
1767                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1768                        obd->obd_name,
1769                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1770                        list_empty(&obd->obd_lock_replay_queue) ? \
1771                                   "" : "lock ",
1772                        list_empty(&obd->obd_final_req_queue) ? \
1773                                   "" : "final ");
1774                 spin_unlock(&obd->obd_recovery_task_lock);
1775                 LBUG();
1776         }
1777         spin_unlock(&obd->obd_recovery_task_lock);
1778
1779         obd->obd_recovery_end = ktime_get_seconds();
1780
1781         /* When recovery finished, cleanup orphans on MDS and OST. */
1782         if (obd->obd_type && OBP(obd, postrecov)) {
1783                 int rc = OBP(obd, postrecov)(obd);
1784
1785                 if (rc < 0)
1786                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1787                                       obd->obd_name, rc);
1788         }
1789         EXIT;
1790 }
1791
1792 static void abort_req_replay_queue(struct obd_device *obd)
1793 {
1794         struct ptlrpc_request *req, *n;
1795         LIST_HEAD(abort_list);
1796
1797         spin_lock(&obd->obd_recovery_task_lock);
1798         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1799         spin_unlock(&obd->obd_recovery_task_lock);
1800         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1801                 DEBUG_REQ(D_WARNING, req, "aborted:");
1802                 req->rq_status = -ENOTCONN;
1803                 if (ptlrpc_error(req)) {
1804                         DEBUG_REQ(D_ERROR, req,
1805                                   "failed abort_req_reply; skipping");
1806                 }
1807                 target_exp_dequeue_req_replay(req);
1808                 target_request_copy_put(req);
1809         }
1810 }
1811
1812 static void abort_lock_replay_queue(struct obd_device *obd)
1813 {
1814         struct ptlrpc_request *req, *n;
1815         LIST_HEAD(abort_list);
1816
1817         spin_lock(&obd->obd_recovery_task_lock);
1818         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1819         spin_unlock(&obd->obd_recovery_task_lock);
1820         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1821                 DEBUG_REQ(D_ERROR, req, "aborted:");
1822                 req->rq_status = -ENOTCONN;
1823                 if (ptlrpc_error(req)) {
1824                         DEBUG_REQ(D_ERROR, req,
1825                                   "failed abort_lock_reply; skipping");
1826                 }
1827                 target_request_copy_put(req);
1828         }
1829 }
1830
1831 /*
1832  * Called from a cleanup function if the device is being cleaned up
1833  * forcefully.  The exports should all have been disconnected already,
1834  * the only thing left to do is
1835  * - clear the recovery flags
1836  * - cancel the timer
1837  * - free queued requests and replies, but don't send replies
1838  * Because the obd_stopping flag is set, no new requests should be received.
1839  */
1840 void target_cleanup_recovery(struct obd_device *obd)
1841 {
1842         struct ptlrpc_request *req, *n;
1843         LIST_HEAD(clean_list);
1844
1845         spin_lock(&obd->obd_dev_lock);
1846         if (!obd->obd_recovering) {
1847                 spin_unlock(&obd->obd_dev_lock);
1848                 EXIT;
1849                 return;
1850         }
1851         obd->obd_recovering = obd->obd_abort_recovery = 0;
1852         obd->obd_abort_recov_mdt = 0;
1853         spin_unlock(&obd->obd_dev_lock);
1854
1855         spin_lock(&obd->obd_recovery_task_lock);
1856         target_cancel_recovery_timer(obd);
1857         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1858         spin_unlock(&obd->obd_recovery_task_lock);
1859
1860         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1861                 LASSERT(req->rq_reply_state == NULL);
1862                 target_exp_dequeue_req_replay(req);
1863                 target_request_copy_put(req);
1864         }
1865
1866         spin_lock(&obd->obd_recovery_task_lock);
1867         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1868         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1869         spin_unlock(&obd->obd_recovery_task_lock);
1870
1871         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1872                 LASSERT(req->rq_reply_state == NULL);
1873                 target_request_copy_put(req);
1874         }
1875
1876         EXIT;
1877 }
1878 EXPORT_SYMBOL(target_cleanup_recovery);
1879
1880 /* obd_recovery_task_lock should be held */
1881 void target_cancel_recovery_timer(struct obd_device *obd)
1882 {
1883         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1884         hrtimer_cancel(&obd->obd_recovery_timer);
1885 }
1886
1887 static void target_start_recovery_timer(struct obd_device *obd)
1888 {
1889         ktime_t delay;
1890
1891         if (obd->obd_recovery_start != 0)
1892                 return;
1893
1894         spin_lock(&obd->obd_dev_lock);
1895         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1896                 spin_unlock(&obd->obd_dev_lock);
1897                 return;
1898         }
1899
1900         LASSERT(obd->obd_recovery_timeout != 0);
1901
1902         if (obd->obd_recovery_start != 0) {
1903                 spin_unlock(&obd->obd_dev_lock);
1904                 return;
1905         }
1906
1907         obd->obd_recovery_start = ktime_get_seconds();
1908         delay = ktime_set(obd->obd_recovery_start +
1909                           obd->obd_recovery_timeout, 0);
1910         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1911         spin_unlock(&obd->obd_dev_lock);
1912
1913         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1914                       obd->obd_name,
1915                       obd->obd_recovery_timeout / 60,
1916                       obd->obd_recovery_timeout % 60,
1917                       atomic_read(&obd->obd_max_recoverable_clients),
1918                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1919                       "" : "s",
1920                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1921                       "s" : "");
1922 }
1923
1924 /**
1925  * extend recovery window.
1926  *
1927  * if @extend is true, extend recovery window to have @dr_timeout remaining
1928  * at least; otherwise, make sure the recovery timeout value is not less
1929  * than @dr_timeout.
1930  */
1931 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1932                                   bool extend)
1933 {
1934         ktime_t left_ns;
1935         timeout_t timeout;
1936         timeout_t left;
1937
1938         spin_lock(&obd->obd_dev_lock);
1939         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1940             obd->obd_stopping) {
1941                 spin_unlock(&obd->obd_dev_lock);
1942                 return;
1943         }
1944         LASSERT(obd->obd_recovery_start != 0);
1945
1946         left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1947         left = ktime_divns(left_ns, NSEC_PER_SEC);
1948
1949         if (extend) {
1950                 timeout = obd->obd_recovery_timeout;
1951                 /* dr_timeout will happen after the hrtimer has expired.
1952                  * Add the excess time to the soft recovery timeout without
1953                  * exceeding the hard recovery timeout.
1954                  */
1955                 if (dr_timeout > left) {
1956                         timeout += dr_timeout - left;
1957                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1958                                         timeout);
1959                 }
1960         } else {
1961                 timeout = clamp_t(timeout_t, dr_timeout,
1962                                   obd->obd_recovery_timeout,
1963                                   obd->obd_recovery_time_hard);
1964         }
1965
1966         if (timeout == obd->obd_recovery_time_hard)
1967                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1968                       obd->obd_name, timeout, extend);
1969
1970         if (obd->obd_recovery_timeout < timeout) {
1971                 ktime_t end, now;
1972
1973                 obd->obd_recovery_timeout = timeout;
1974                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1975                 now = ktime_set(ktime_get_seconds(), 0);
1976                 left_ns = ktime_sub(end, now);
1977                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1978                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1979         }
1980         spin_unlock(&obd->obd_dev_lock);
1981
1982         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1983                 obd->obd_name, left);
1984 }
1985
1986 /* Reset the timer with each new client connection */
1987 /*
1988  * This timer is actually reconnect_timer, which is for making sure
1989  * the total recovery window is at least as big as my reconnect
1990  * attempt timing. So the initial recovery time_out will be set to
1991  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1992  * from client is bigger than this, then the recovery time_out will
1993  * be extended to make sure the client could be reconnected, in the
1994  * process, the timeout from the new client should be ignored.
1995  */
1996 static void
1997 check_and_start_recovery_timer(struct obd_device *obd,
1998                                struct ptlrpc_request *req,
1999                                int new_client)
2000 {
2001         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2002         struct obd_device_target *obt = &obd->u.obt;
2003
2004         if (!new_client && service_timeout)
2005                 /*
2006                  * Teach server about old server's estimates, as first guess
2007                  * at how long new requests will take.
2008                  */
2009                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2010                             service_timeout);
2011
2012         target_start_recovery_timer(obd);
2013
2014         /*
2015          * Convert the service time to RPC timeout,
2016          * and reuse service_timeout to limit stack usage.
2017          */
2018         service_timeout = at_est2timeout(service_timeout);
2019
2020         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2021             service_timeout < at_extra)
2022                 service_timeout = at_extra;
2023
2024         /*
2025          * We expect other clients to timeout within service_timeout, then try
2026          * to reconnect, then try the failover server.  The max delay between
2027          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2028          */
2029         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2030
2031         LASSERT(obt->obt_magic == OBT_MAGIC);
2032         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2033         if (service_timeout > obd->obd_recovery_timeout && !new_client)
2034                 extend_recovery_timer(obd, service_timeout, false);
2035 }
2036
2037 /** Health checking routines */
2038 static inline int exp_connect_healthy(struct obd_export *exp)
2039 {
2040         return exp->exp_in_recovery;
2041 }
2042
2043 /** if export done req_replay or has replay in queue */
2044 static inline int exp_req_replay_healthy(struct obd_export *exp)
2045 {
2046         return (!exp->exp_req_replay_needed ||
2047                 atomic_read(&exp->exp_replay_count) > 0);
2048 }
2049
2050
2051 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2052 {
2053         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2054                exp_req_replay_healthy(exp);
2055 }
2056
2057 /** if export done lock_replay or has replay in queue */
2058 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2059 {
2060         return (!exp->exp_lock_replay_needed ||
2061                 atomic_read(&exp->exp_replay_count) > 0);
2062 }
2063
2064 static inline int exp_vbr_healthy(struct obd_export *exp)
2065 {
2066         return !exp->exp_vbr_failed;
2067 }
2068
2069 static inline int exp_finished(struct obd_export *exp)
2070 {
2071         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2072 }
2073
2074 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2075 {
2076         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2077                 exp_finished(exp);
2078 }
2079
2080 static int check_for_next_transno(struct lu_target *lut)
2081 {
2082         struct ptlrpc_request *req = NULL;
2083         struct obd_device *obd = lut->lut_obd;
2084         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2085         int wake_up = 0, connected, completed, queue_len;
2086         __u64 req_transno = 0;
2087         __u64 update_transno = 0;
2088         __u64 next_transno = 0;
2089
2090         ENTRY;
2091
2092         spin_lock(&obd->obd_recovery_task_lock);
2093         if (!list_empty(&obd->obd_req_replay_queue)) {
2094                 req = list_entry(obd->obd_req_replay_queue.next,
2095                                      struct ptlrpc_request, rq_list);
2096                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2097         }
2098
2099         if (!obd->obd_abort_recov_mdt && tdtd)
2100                 update_transno = distribute_txn_get_next_transno(tdtd);
2101
2102         connected = atomic_read(&obd->obd_connected_clients);
2103         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2104         queue_len = obd->obd_requests_queued_for_recovery;
2105         next_transno = obd->obd_next_recovery_transno;
2106
2107         CDEBUG(D_HA,
2108                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2109                atomic_read(&obd->obd_max_recoverable_clients),
2110                connected, completed,
2111                queue_len, req_transno, next_transno);
2112
2113         if (obd->obd_abort_recovery) {
2114                 CDEBUG(D_HA, "waking for aborted recovery\n");
2115                 wake_up = 1;
2116         } else if (obd->obd_recovery_expired) {
2117                 CDEBUG(D_HA, "waking for expired recovery\n");
2118                 wake_up = 1;
2119         } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2120                    is_req_replayed_by_update(req)) {
2121                 LASSERTF(req_transno < next_transno,
2122                          "req_transno %llu next_transno%llu\n", req_transno,
2123                          next_transno);
2124                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2125                        req_transno);
2126                 wake_up = 1;
2127         } else if (req_transno == next_transno ||
2128                    (update_transno != 0 && update_transno <= next_transno)) {
2129                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2130                 wake_up = 1;
2131         } else if (queue_len > 0 &&
2132                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2133                 /** handle gaps occured due to lost reply or VBR */
2134                 LASSERTF(req_transno >= next_transno,
2135                          "req_transno: %llu, next_transno: %llu\n",
2136                          req_transno, next_transno);
2137                 CDEBUG(D_HA,
2138                        "%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",
2139                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2140                        next_transno, queue_len, completed, connected,
2141                        req_transno, update_transno, obd->obd_last_committed);
2142                 obd->obd_next_recovery_transno = req_transno;
2143                 wake_up = 1;
2144         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2145                 CDEBUG(D_HA, "waking for completed recovery\n");
2146                 wake_up = 1;
2147         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2148                 CDEBUG(D_HA,
2149                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2150                        next_transno);
2151                 obd->obd_next_recovery_transno = req_transno;
2152                 wake_up = 1;
2153         }
2154         spin_unlock(&obd->obd_recovery_task_lock);
2155         return wake_up;
2156 }
2157
2158 static int check_for_next_lock(struct lu_target *lut)
2159 {
2160         struct obd_device *obd = lut->lut_obd;
2161         int wake_up = 0;
2162
2163         spin_lock(&obd->obd_recovery_task_lock);
2164         if (!list_empty(&obd->obd_lock_replay_queue)) {
2165                 CDEBUG(D_HA, "waking for next lock\n");
2166                 wake_up = 1;
2167         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2168                 CDEBUG(D_HA, "waking for completed lock replay\n");
2169                 wake_up = 1;
2170         } else if (obd->obd_abort_recovery) {
2171                 CDEBUG(D_HA, "waking for aborted recovery\n");
2172                 wake_up = 1;
2173         } else if (obd->obd_recovery_expired) {
2174                 CDEBUG(D_HA, "waking for expired recovery\n");
2175                 wake_up = 1;
2176         }
2177         spin_unlock(&obd->obd_recovery_task_lock);
2178
2179         return wake_up;
2180 }
2181
2182 static int check_update_llog(struct lu_target *lut)
2183 {
2184         struct obd_device *obd = lut->lut_obd;
2185         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2186
2187         if (obd->obd_abort_recovery) {
2188                 CDEBUG(D_HA, "waking for aborted recovery\n");
2189                 return 1;
2190         }
2191
2192         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2193                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2194                 return 1;
2195         }
2196
2197         return 0;
2198 }
2199
2200 /**
2201  * wait for recovery events,
2202  * check its status with help of check_routine
2203  * evict dead clients via health_check
2204  */
2205 static int target_recovery_overseer(struct lu_target *lut,
2206                                     int (*check_routine)(struct lu_target *),
2207                                     int (*health_check)(struct obd_export *))
2208 {
2209         struct obd_device *obd = lut->lut_obd;
2210         struct target_distribute_txn_data *tdtd;
2211         time64_t last = 0;
2212         time64_t now;
2213 repeat:
2214         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2215                 now = ktime_get_seconds();
2216                 if (now - last > 600) {
2217                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2218                                       obd->obd_name);
2219                         last = now;
2220                 }
2221         }
2222         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2223               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2224                 __u64 next_update_transno = 0;
2225
2226                 /*
2227                  * Only abort the recovery if there are no update recovery
2228                  * left in the queue
2229                  */
2230                 spin_lock(&obd->obd_recovery_task_lock);
2231                 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2232                         next_update_transno =
2233                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2234
2235                         tdtd = lut->lut_tdtd;
2236                         /*
2237                          * If next_update_transno == 0, it probably because
2238                          * updatelog retrieve threads did not get any records
2239                          * yet, let's wait those threads stopped
2240                          */
2241                         if (next_update_transno == 0) {
2242                                 spin_unlock(&obd->obd_recovery_task_lock);
2243
2244                                 while (wait_event_timeout(
2245                                         tdtd->tdtd_recovery_threads_waitq,
2246                                         check_update_llog(lut),
2247                                         cfs_time_seconds(60)) == 0);
2248
2249                                 spin_lock(&obd->obd_recovery_task_lock);
2250                                 next_update_transno =
2251                                         distribute_txn_get_next_transno(tdtd);
2252                         }
2253                 }
2254
2255                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2256                         obd->obd_next_recovery_transno = next_update_transno;
2257                         spin_unlock(&obd->obd_recovery_task_lock);
2258                         /*
2259                          * Disconnect unfinished exports from clients, and
2260                          * keep connection from MDT to make sure the update
2261                          * recovery will still keep trying until some one
2262                          * manually abort the recovery
2263                          */
2264                         class_disconnect_stale_exports(obd,
2265                                                 exp_finished_or_from_mdt);
2266                         /* Abort all of replay & replay lock req from clients */
2267                         abort_req_replay_queue(obd);
2268                         abort_lock_replay_queue(obd);
2269                         CDEBUG(D_HA,
2270                                "%s: there are still update replay (%#llx)in the queue.\n",
2271                                obd->obd_name, next_update_transno);
2272                 } else {
2273                         obd->obd_abort_recovery = 1;
2274                         spin_unlock(&obd->obd_recovery_task_lock);
2275                         CWARN("%s recovery is aborted by hard timeout\n",
2276                               obd->obd_name);
2277                 }
2278         }
2279
2280         while (wait_event_timeout(obd->obd_next_transno_waitq,
2281                                   check_routine(lut),
2282                                   cfs_time_seconds(60)) == 0)
2283                 ; /* wait indefinitely for event, but don't trigger watchdog */
2284
2285         if (obd->obd_abort_recovery) {
2286                 CWARN("recovery is aborted, evict exports in recovery\n");
2287                 if (lut->lut_tdtd != NULL) {
2288                         tdtd = lut->lut_tdtd;
2289                         /*
2290                          * Let's wait all of the update log recovery thread
2291                          * finished
2292                          */
2293                         wait_event_idle(
2294                                 tdtd->tdtd_recovery_threads_waitq,
2295                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2296                                 == 0);
2297                         /* Then abort the update recovery list */
2298                         dtrq_list_destroy(lut->lut_tdtd);
2299                 }
2300
2301                 /** evict exports which didn't finish recovery yet */
2302                 class_disconnect_stale_exports(obd, exp_finished);
2303                 return 1;
2304         } else if (obd->obd_recovery_expired) {
2305                 obd->obd_recovery_expired = 0;
2306
2307                 /** If some clients died being recovered, evict them */
2308                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2309                               obd->obd_name);
2310                 /** evict cexports with no replay in queue, they are stalled */
2311                 class_disconnect_stale_exports(obd, health_check);
2312
2313                 /** continue with VBR */
2314                 spin_lock(&obd->obd_dev_lock);
2315                 obd->obd_version_recov = 1;
2316                 spin_unlock(&obd->obd_dev_lock);
2317                 /**
2318                  * reset timer, recovery will proceed with versions now,
2319                  * timeout is set just to handle reconnection delays
2320                  */
2321                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2322                 /**
2323                  * Wait for recovery events again, after evicting bad clients
2324                  */
2325                 goto repeat;
2326         }
2327         return 0;
2328 }
2329
2330 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2331 {
2332         struct obd_device *obd = lut->lut_obd;
2333         struct ptlrpc_request *req = NULL;
2334
2335         CDEBUG(D_HA, "Waiting for lock\n");
2336         if (target_recovery_overseer(lut, check_for_next_lock,
2337                                      exp_lock_replay_healthy))
2338                 abort_lock_replay_queue(obd);
2339
2340         spin_lock(&obd->obd_recovery_task_lock);
2341         if (!list_empty(&obd->obd_lock_replay_queue)) {
2342                 req = list_entry(obd->obd_lock_replay_queue.next,
2343                                      struct ptlrpc_request, rq_list);
2344                 list_del_init(&req->rq_list);
2345                 spin_unlock(&obd->obd_recovery_task_lock);
2346         } else {
2347                 spin_unlock(&obd->obd_recovery_task_lock);
2348                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2349                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2350                 /** evict exports failed VBR */
2351                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2352         }
2353         return req;
2354 }
2355
2356 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2357 {
2358         struct ptlrpc_request *req = NULL;
2359
2360         spin_lock(&obd->obd_recovery_task_lock);
2361         if (!list_empty(&obd->obd_final_req_queue)) {
2362                 req = list_entry(obd->obd_final_req_queue.next,
2363                                      struct ptlrpc_request, rq_list);
2364                 list_del_init(&req->rq_list);
2365                 spin_unlock(&obd->obd_recovery_task_lock);
2366                 if (req->rq_export->exp_in_recovery) {
2367                         spin_lock(&req->rq_export->exp_lock);
2368                         req->rq_export->exp_in_recovery = 0;
2369                         spin_unlock(&req->rq_export->exp_lock);
2370                 }
2371         } else {
2372                 spin_unlock(&obd->obd_recovery_task_lock);
2373         }
2374         return req;
2375 }
2376
2377 static void handle_recovery_req(struct ptlrpc_thread *thread,
2378                                 struct ptlrpc_request *req,
2379                                 svc_handler_t handler)
2380 {
2381         ENTRY;
2382
2383         /**
2384          * export can be evicted during recovery, no need to handle replays for
2385          * it after that, discard such request silently
2386          */
2387         if (req->rq_export->exp_disconnected)
2388                 RETURN_EXIT;
2389
2390         req->rq_session.lc_thread = thread;
2391         req->rq_svc_thread = thread;
2392         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2393
2394         /* thread context */
2395         lu_context_enter(&thread->t_env->le_ctx);
2396         (void)handler(req);
2397         lu_context_exit(&thread->t_env->le_ctx);
2398
2399         req->rq_svc_thread->t_env->le_ses = NULL;
2400
2401         /* don't reset timer for final stage */
2402         if (!exp_finished(req->rq_export)) {
2403                 timeout_t timeout = obd_timeout;
2404
2405                 /**
2406                  * Add request @timeout to the recovery time so next request from
2407                  * this client may come in recovery time
2408                  */
2409                 if (!AT_OFF) {
2410                         struct ptlrpc_service_part *svcpt;
2411                         timeout_t est_timeout;
2412
2413                         svcpt = req->rq_rqbd->rqbd_svcpt;
2414                         /*
2415                          * If the server sent early reply for this request,
2416                          * the client will recalculate the timeout according to
2417                          * current server estimate service time, so we will
2418                          * use the maxium timeout here for waiting the client
2419                          * sending the next req
2420                          */
2421                         est_timeout = at_get(&svcpt->scp_at_estimate);
2422                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2423                                         lustre_msg_get_timeout(req->rq_reqmsg));
2424                         /*
2425                          * Add 2 net_latency, one for balance rq_deadline
2426                          * (see ptl_send_rpc), one for resend the req to server,
2427                          * Note: client will pack net_latency in replay req
2428                          * (see ptlrpc_replay_req)
2429                          */
2430                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2431                 }
2432                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2433                                       true);
2434         }
2435         EXIT;
2436 }
2437
2438 /** Checking routines for recovery */
2439 static int check_for_recovery_ready(struct lu_target *lut)
2440 {
2441         struct obd_device *obd = lut->lut_obd;
2442         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2443
2444         CDEBUG(D_HA,
2445                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2446                clnts, obd->obd_stale_clients,
2447                atomic_read(&obd->obd_max_recoverable_clients),
2448                obd->obd_abort_recovery, obd->obd_recovery_expired);
2449
2450         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2451                 LASSERT(clnts <=
2452                         atomic_read(&obd->obd_max_recoverable_clients));
2453                 if (clnts + obd->obd_stale_clients <
2454                     atomic_read(&obd->obd_max_recoverable_clients))
2455                         return 0;
2456         }
2457
2458         if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2459                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2460                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2461                         /*
2462                          * Let's extend recovery timer, in case the recovery
2463                          * timer expired, and some clients got evicted
2464                          */
2465                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2466                                               true);
2467                         CDEBUG(D_HA,
2468                                "%s update recovery is not ready, extend recovery %d\n",
2469                                obd->obd_name, obd->obd_recovery_timeout);
2470                         return 0;
2471                 }
2472         }
2473
2474         return 1;
2475 }
2476
2477 enum {
2478         REQUEST_RECOVERY = 1,
2479         UPDATE_RECOVERY = 2,
2480 };
2481
2482 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2483 {
2484         __u64 transno = 0;
2485
2486         if (!list_empty(&obd->obd_req_replay_queue)) {
2487                 struct ptlrpc_request *req;
2488
2489                 req = list_entry(obd->obd_req_replay_queue.next,
2490                                  struct ptlrpc_request, rq_list);
2491                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2492         }
2493
2494         return transno;
2495 }
2496
2497 static __u64 get_next_transno(struct lu_target *lut, int *type)
2498 {
2499         struct obd_device *obd = lut->lut_obd;
2500         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2501         __u64 transno = 0;
2502         __u64 update_transno;
2503
2504         ENTRY;
2505
2506         transno = get_next_replay_req_transno(obd);
2507         if (type != NULL)
2508                 *type = REQUEST_RECOVERY;
2509
2510         if (!tdtd || obd->obd_abort_recov_mdt)
2511                 RETURN(transno);
2512
2513         update_transno = distribute_txn_get_next_transno(tdtd);
2514         if (transno == 0 || (transno >= update_transno &&
2515                              update_transno != 0)) {
2516                 transno = update_transno;
2517                 if (type != NULL)
2518                         *type = UPDATE_RECOVERY;
2519         }
2520
2521         RETURN(transno);
2522 }
2523
2524 /**
2525  * drop duplicate replay request
2526  *
2527  * Because the operation has been replayed by update recovery, the request
2528  * with the same transno will be dropped and also notify the client to send
2529  * next replay request.
2530  *
2531  * \param[in] env       execution environment
2532  * \param[in] obd       failover obd device
2533  * \param[in] req       request to be dropped
2534  */
2535 static void drop_duplicate_replay_req(struct lu_env *env,
2536                                       struct obd_device *obd,
2537                                       struct ptlrpc_request *req)
2538 {
2539         DEBUG_REQ(D_HA, req,
2540                   "remove t%lld from %s because duplicate update records found",
2541                   lustre_msg_get_transno(req->rq_reqmsg),
2542                   libcfs_nidstr(&req->rq_peer.nid));
2543
2544         /*
2545          * Right now, only for MDS reint operation update replay and
2546          * normal request replay can have the same transno
2547          */
2548         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2549                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2550                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2551                 if (likely(req->rq_export))
2552                         target_committed_to_req(req);
2553                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2554                 target_send_reply(req, req->rq_status, 0);
2555         } else {
2556                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2557                 libcfs_nidstr(&req->rq_peer.nid));
2558         }
2559         target_exp_dequeue_req_replay(req);
2560         target_request_copy_put(req);
2561         obd->obd_replayed_requests++;
2562 }
2563
2564 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2565
2566 static void replay_request_or_update(struct lu_env *env,
2567                                      struct lu_target *lut,
2568                                      struct target_recovery_data *trd,
2569                                      struct ptlrpc_thread *thread)
2570 {
2571         struct obd_device *obd = lut->lut_obd;
2572         struct ptlrpc_request *req = NULL;
2573         int type;
2574         __u64 transno;
2575
2576         ENTRY;
2577
2578         CDEBUG(D_HA, "Waiting for transno %lld\n",
2579                obd->obd_next_recovery_transno);
2580
2581         /* Replay all of request and update by transno */
2582         do {
2583                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2584
2585                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2586
2587                 /**
2588                  * It is needed to extend recovery window above
2589                  *  recovery_time_soft. Extending is possible only in the
2590                  *  end of recovery window (see more details in
2591                  *  handle_recovery_req()).
2592                  */
2593                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2594
2595                 if (target_recovery_overseer(lut, check_for_next_transno,
2596                                         exp_req_replay_healthy_or_from_mdt)) {
2597                         abort_req_replay_queue(obd);
2598                         abort_lock_replay_queue(obd);
2599                         goto abort;
2600                 }
2601
2602                 spin_lock(&obd->obd_recovery_task_lock);
2603                 transno = get_next_transno(lut, &type);
2604                 if (type == REQUEST_RECOVERY && transno != 0) {
2605                         /*
2606                          * Drop replay request from client side, if the
2607                          * replay has been executed by update with the
2608                          * same transno
2609                          */
2610                         req = list_entry(obd->obd_req_replay_queue.next,
2611                                         struct ptlrpc_request, rq_list);
2612
2613                         list_del_init(&req->rq_list);
2614                         obd->obd_requests_queued_for_recovery--;
2615                         spin_unlock(&obd->obd_recovery_task_lock);
2616
2617                         /*
2618                          * Let's check if the request has been redone by
2619                          * update replay
2620                          */
2621                         if (is_req_replayed_by_update(req)) {
2622                                 struct distribute_txn_replay_req *dtrq;
2623
2624                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2625                                                                       transno);
2626                                 LASSERT(dtrq != NULL);
2627                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2628                                 list_del_init(&dtrq->dtrq_list);
2629                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2630                                 dtrq_destroy(dtrq);
2631
2632                                 drop_duplicate_replay_req(env, obd, req);
2633
2634                                 continue;
2635                         }
2636
2637                         LASSERT(trd->trd_processing_task == current->pid);
2638                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2639                                   req->rq_xid,
2640                                   lustre_msg_get_transno(req->rq_reqmsg),
2641                                   libcfs_nidstr(&req->rq_peer.nid));
2642
2643                         ptlrpc_watchdog_init(&thread->t_watchdog,
2644                                              WATCHDOG_TIMEOUT);
2645                         handle_recovery_req(thread, req,
2646                                             trd->trd_recovery_handler);
2647                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2648
2649                         /**
2650                          * bz18031: increase next_recovery_transno before
2651                          * target_request_copy_put() will drop exp_rpc reference
2652                          */
2653                         spin_lock(&obd->obd_recovery_task_lock);
2654                         obd->obd_next_recovery_transno++;
2655                         spin_unlock(&obd->obd_recovery_task_lock);
2656                         target_exp_dequeue_req_replay(req);
2657                         target_request_copy_put(req);
2658                         obd->obd_replayed_requests++;
2659                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2660                         struct distribute_txn_replay_req *dtrq;
2661                         int rc;
2662
2663                         spin_unlock(&obd->obd_recovery_task_lock);
2664
2665                         LASSERT(tdtd != NULL);
2666                         dtrq = distribute_txn_get_next_req(tdtd);
2667                         lu_context_enter(&thread->t_env->le_ctx);
2668                         ptlrpc_watchdog_init(&thread->t_watchdog,
2669                                              WATCHDOG_TIMEOUT);
2670                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2671                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2672                         lu_context_exit(&thread->t_env->le_ctx);
2673                         extend_recovery_timer(obd, obd_timeout, true);
2674
2675                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2676                                 CDEBUG(D_HA,
2677                                        "Move x%llu t%llu to finish list\n",
2678                                        dtrq->dtrq_xid,
2679                                        dtrq->dtrq_master_transno);
2680
2681                                 /* Add it to the replay finish list */
2682                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2683                                 list_add(&dtrq->dtrq_list,
2684                                          &tdtd->tdtd_replay_finish_list);
2685                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2686
2687                                 spin_lock(&obd->obd_recovery_task_lock);
2688                                 if (transno == obd->obd_next_recovery_transno)
2689                                         obd->obd_next_recovery_transno++;
2690                                 else if (transno >
2691                                          obd->obd_next_recovery_transno)
2692                                         obd->obd_next_recovery_transno =
2693                                                                 transno + 1;
2694                                 spin_unlock(&obd->obd_recovery_task_lock);
2695                         } else {
2696                                 dtrq_destroy(dtrq);
2697                         }
2698                 } else {
2699                         spin_unlock(&obd->obd_recovery_task_lock);
2700 abort:
2701                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2702                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2703                         /** evict exports failed VBR */
2704                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2705                         break;
2706                 }
2707         } while (1);
2708 }
2709
2710 static int target_recovery_thread(void *arg)
2711 {
2712         struct lu_target *lut = arg;
2713         struct obd_device *obd = lut->lut_obd;
2714         struct ptlrpc_request *req;
2715         struct target_recovery_data *trd = &obd->obd_recovery_data;
2716         unsigned long delta;
2717         struct lu_env *env;
2718         struct ptlrpc_thread *thread = NULL;
2719         int rc = 0;
2720
2721         ENTRY;
2722         unshare_fs_struct();
2723         OBD_ALLOC_PTR(thread);
2724         if (thread == NULL)
2725                 RETURN(-ENOMEM);
2726
2727         OBD_ALLOC_PTR(env);
2728         if (env == NULL)
2729                 GOTO(out_thread, rc = -ENOMEM);
2730         rc = lu_env_add(env);
2731         if (rc)
2732                 GOTO(out_env, rc);
2733
2734         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2735         if (rc)
2736                 GOTO(out_env_remove, rc);
2737
2738         thread->t_env = env;
2739         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2740         thread->t_task = current;
2741         env->le_ctx.lc_thread = thread;
2742         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2743
2744         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2745                current->pid);
2746         trd->trd_processing_task = current->pid;
2747
2748         spin_lock(&obd->obd_dev_lock);
2749         obd->obd_recovering = 1;
2750         spin_unlock(&obd->obd_dev_lock);
2751         complete(&trd->trd_starting);
2752
2753         /* first of all, we have to know the first transno to replay */
2754         if (target_recovery_overseer(lut, check_for_recovery_ready,
2755                                      exp_connect_healthy)) {
2756                 abort_req_replay_queue(obd);
2757                 abort_lock_replay_queue(obd);
2758                 if (lut->lut_tdtd != NULL)
2759                         dtrq_list_destroy(lut->lut_tdtd);
2760         }
2761
2762         /* next stage: replay requests or update */
2763         delta = jiffies;
2764         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2765                atomic_read(&obd->obd_req_replay_clients),
2766                obd->obd_next_recovery_transno);
2767         replay_request_or_update(env, lut, trd, thread);
2768
2769         /**
2770          * The second stage: replay locks
2771          */
2772         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2773                atomic_read(&obd->obd_lock_replay_clients));
2774         while ((req = target_next_replay_lock(lut))) {
2775                 LASSERT(trd->trd_processing_task == current->pid);
2776                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2777                           libcfs_nidstr(&req->rq_peer.nid));
2778                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2779                         req->rq_status = -ENODEV;
2780                         target_request_copy_put(req);
2781                         continue;
2782                 }
2783                 handle_recovery_req(thread, req,
2784                                     trd->trd_recovery_handler);
2785                 target_request_copy_put(req);
2786                 obd->obd_replayed_locks++;
2787         }
2788
2789         /**
2790          * The third stage: reply on final pings, at this moment all clients
2791          * must have request in final queue
2792          */
2793         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2794         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2795         /** Update server last boot epoch */
2796         tgt_boot_epoch_update(lut);
2797         /*
2798          * We drop recoverying flag to forward all new requests
2799          * to regular mds_handle() since now
2800          */
2801         spin_lock(&obd->obd_dev_lock);
2802         obd->obd_recovering = obd->obd_abort_recovery = 0;
2803         obd->obd_abort_recov_mdt = 0;
2804         spin_unlock(&obd->obd_dev_lock);
2805         spin_lock(&obd->obd_recovery_task_lock);
2806         target_cancel_recovery_timer(obd);
2807         spin_unlock(&obd->obd_recovery_task_lock);
2808         while ((req = target_next_final_ping(obd))) {
2809                 LASSERT(trd->trd_processing_task == current->pid);
2810                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2811                           libcfs_nidstr(&req->rq_peer.nid));
2812                 handle_recovery_req(thread, req,
2813                                     trd->trd_recovery_handler);
2814                 /*
2815                  * Because the waiting client can not send ping to server,
2816                  * so we need refresh the last_request_time, to avoid the
2817                  * export is being evicted
2818                  */
2819                 ptlrpc_update_export_timer(req->rq_export, 0);
2820                 target_request_copy_put(req);
2821         }
2822
2823         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2824         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2825                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2826         if (delta > OBD_RECOVERY_TIME_SOFT) {
2827                 CWARN("too long recovery - read logs\n");
2828                 libcfs_debug_dumplog();
2829         }
2830
2831         target_finish_recovery(lut);
2832         lu_context_fini(&env->le_ctx);
2833         trd->trd_processing_task = 0;
2834         complete_all(&trd->trd_finishing);
2835         tgt_io_thread_done(thread);
2836 out_env_remove:
2837         lu_env_remove(env);
2838 out_env:
2839         OBD_FREE_PTR(env);
2840 out_thread:
2841         OBD_FREE_PTR(thread);
2842         RETURN(rc);
2843 }
2844
2845 static int target_start_recovery_thread(struct lu_target *lut,
2846                                         svc_handler_t handler)
2847 {
2848         struct obd_device *obd = lut->lut_obd;
2849         int rc = 0;
2850         struct target_recovery_data *trd = &obd->obd_recovery_data;
2851         int index;
2852
2853         memset(trd, 0, sizeof(*trd));
2854         init_completion(&trd->trd_starting);
2855         init_completion(&trd->trd_finishing);
2856         trd->trd_recovery_handler = handler;
2857
2858         rc = server_name2index(obd->obd_name, &index, NULL);
2859         if (rc < 0)
2860                 return rc;
2861
2862         if (!IS_ERR(kthread_run(target_recovery_thread,
2863                                 lut, "tgt_recover_%d", index))) {
2864                 wait_for_completion(&trd->trd_starting);
2865                 LASSERT(obd->obd_recovering != 0);
2866         } else {
2867                 rc = -ECHILD;
2868         }
2869
2870         return rc;
2871 }
2872
2873 void target_stop_recovery_thread(struct obd_device *obd)
2874 {
2875         if (obd->obd_recovery_data.trd_processing_task > 0) {
2876                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2877                 /** recovery can be done but postrecovery is not yet */
2878                 spin_lock(&obd->obd_dev_lock);
2879                 if (obd->obd_recovering) {
2880                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2881                         obd->obd_abort_recovery = 1;
2882                         wake_up(&obd->obd_next_transno_waitq);
2883                 }
2884                 spin_unlock(&obd->obd_dev_lock);
2885                 wait_for_completion(&trd->trd_finishing);
2886         }
2887 }
2888 EXPORT_SYMBOL(target_stop_recovery_thread);
2889
2890 void target_recovery_fini(struct obd_device *obd)
2891 {
2892         class_disconnect_exports(obd);
2893         target_stop_recovery_thread(obd);
2894         target_cleanup_recovery(obd);
2895 }
2896 EXPORT_SYMBOL(target_recovery_fini);
2897
2898 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2899 {
2900         struct obd_device *obd = container_of(timer, struct obd_device,
2901                                               obd_recovery_timer);
2902
2903         CDEBUG(D_HA,
2904                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2905                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2906                ktime_get_seconds() - obd->obd_recovery_start,
2907                atomic_read(&obd->obd_connected_clients));
2908
2909         obd->obd_recovery_expired = 1;
2910         wake_up(&obd->obd_next_transno_waitq);
2911         return HRTIMER_NORESTART;
2912 }
2913
2914 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2915 {
2916         struct obd_device *obd = lut->lut_obd;
2917
2918         if (lut->lut_bottom->dd_rdonly)
2919                 return;
2920
2921         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2922                 /** Update server last boot epoch */
2923                 tgt_boot_epoch_update(lut);
2924                 return;
2925         }
2926
2927         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2928                "last_transno %llu\n", obd->obd_name,
2929                atomic_read(&obd->obd_max_recoverable_clients),
2930                obd->obd_last_committed);
2931         LASSERT(obd->obd_stopping == 0);
2932         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2933         obd->obd_recovery_start = 0;
2934         obd->obd_recovery_end = 0;
2935
2936         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2937                      HRTIMER_MODE_ABS);
2938         obd->obd_recovery_timer.function = &target_recovery_expired;
2939         target_start_recovery_thread(lut, handler);
2940 }
2941 EXPORT_SYMBOL(target_recovery_init);
2942
2943 static int target_process_req_flags(struct obd_device *obd,
2944                                     struct ptlrpc_request *req)
2945 {
2946         struct obd_export *exp = req->rq_export;
2947
2948         LASSERT(exp != NULL);
2949         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2950                 /* client declares he's ready to replay locks */
2951                 spin_lock(&exp->exp_lock);
2952                 if (exp->exp_req_replay_needed) {
2953                         exp->exp_req_replay_needed = 0;
2954                         spin_unlock(&exp->exp_lock);
2955
2956                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2957                         atomic_dec(&obd->obd_req_replay_clients);
2958                 } else {
2959                         spin_unlock(&exp->exp_lock);
2960                 }
2961         }
2962         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2963                 /*
2964                  * client declares he's ready to complete recovery
2965                  * so, we put the request on th final queue
2966                  */
2967                 spin_lock(&exp->exp_lock);
2968                 if (exp->exp_lock_replay_needed) {
2969                         exp->exp_lock_replay_needed = 0;
2970                         spin_unlock(&exp->exp_lock);
2971
2972                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2973                         atomic_dec(&obd->obd_lock_replay_clients);
2974                 } else {
2975                         spin_unlock(&exp->exp_lock);
2976                 }
2977         }
2978         return 0;
2979 }
2980
2981 int target_queue_recovery_request(struct ptlrpc_request *req,
2982                                   struct obd_device *obd)
2983 {
2984         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2985         struct ptlrpc_request *reqiter;
2986         int inserted = 0;
2987
2988         ENTRY;
2989
2990         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2991                 /* Processing the queue right now, don't re-add. */
2992                 RETURN(1);
2993         }
2994
2995         target_process_req_flags(obd, req);
2996
2997         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2998                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2999                         if (cfs_fail_val == 1) {
3000                                 cfs_race_state = 1;
3001                                 cfs_fail_val = 0;
3002                                 wake_up(&cfs_race_waitq);
3003
3004                                 schedule_timeout_interruptible(
3005                                         cfs_time_seconds(1));
3006                         }
3007                 }
3008
3009                 /*
3010                  * client declares he's ready to complete recovery
3011                  * so, we put the request on th final queue
3012                  */
3013                 target_request_copy_get(req);
3014                 DEBUG_REQ(D_HA, req, "queue final req");
3015                 wake_up(&obd->obd_next_transno_waitq);
3016                 spin_lock(&obd->obd_recovery_task_lock);
3017                 if (obd->obd_recovering) {
3018                         struct ptlrpc_request *tmp;
3019                         struct ptlrpc_request *duplicate = NULL;
3020
3021                         if (likely(!req->rq_export->exp_replay_done)) {
3022                                 req->rq_export->exp_replay_done = 1;
3023                                 list_add_tail(&req->rq_list,
3024                                               &obd->obd_final_req_queue);
3025                                 spin_unlock(&obd->obd_recovery_task_lock);
3026                                 RETURN(0);
3027                         }
3028
3029                         /*
3030                          * XXX O(n), but only happens if final ping is
3031                          * timed out, probably reorganize the list as
3032                          * a hash list later
3033                          */
3034                         list_for_each_entry_safe(reqiter, tmp,
3035                                                  &obd->obd_final_req_queue,
3036                                                  rq_list) {
3037                                 if (reqiter->rq_export == req->rq_export) {
3038                                         list_del_init(&reqiter->rq_list);
3039                                         duplicate = reqiter;
3040                                         break;
3041                                 }
3042                         }
3043
3044                         list_add_tail(&req->rq_list,
3045                                       &obd->obd_final_req_queue);
3046                         req->rq_export->exp_replay_done = 1;
3047                         spin_unlock(&obd->obd_recovery_task_lock);
3048
3049                         if (duplicate != NULL) {
3050                                 DEBUG_REQ(D_HA, duplicate,
3051                                           "put prev final req");
3052                                 target_request_copy_put(duplicate);
3053                         }
3054                         RETURN(0);
3055                 } else {
3056                         spin_unlock(&obd->obd_recovery_task_lock);
3057                         target_request_copy_put(req);
3058                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3059                 }
3060         }
3061         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3062                 /* client declares he's ready to replay locks */
3063                 target_request_copy_get(req);
3064                 DEBUG_REQ(D_HA, req, "queue lock replay req");
3065                 wake_up(&obd->obd_next_transno_waitq);
3066                 spin_lock(&obd->obd_recovery_task_lock);
3067                 LASSERT(obd->obd_recovering);
3068                 /* usually due to recovery abort */
3069                 if (!req->rq_export->exp_in_recovery) {
3070                         spin_unlock(&obd->obd_recovery_task_lock);
3071                         target_request_copy_put(req);
3072                         RETURN(-ENOTCONN);
3073                 }
3074                 LASSERT(req->rq_export->exp_lock_replay_needed);
3075                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3076                 spin_unlock(&obd->obd_recovery_task_lock);
3077                 RETURN(0);
3078         }
3079
3080         /*
3081          * CAVEAT EMPTOR: The incoming request message has been swabbed
3082          * (i.e. buflens etc are in my own byte order), but type-dependent
3083          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3084          */
3085
3086         if (!transno) {
3087                 INIT_LIST_HEAD(&req->rq_list);
3088                 DEBUG_REQ(D_HA, req, "not queueing");
3089                 RETURN(1);
3090         }
3091
3092         /*
3093          * If we're processing the queue, we want don't want to queue this
3094          * message.
3095          *
3096          * Also, if this request has a transno less than the one we're waiting
3097          * for, we should process it now.  It could (and currently always will)
3098          * be an open request for a descriptor that was opened some time ago.
3099          *
3100          * Also, a resent, replayed request that has already been
3101          * handled will pass through here and be processed immediately.
3102          */
3103         CDEBUG(D_HA,
3104                "Next recovery transno: %llu, current: %llu, replaying\n",
3105                obd->obd_next_recovery_transno, transno);
3106
3107         /*
3108          * If the request has been replayed by update replay, then sends this
3109          * request to the recovery thread (replay_request_or_update()), where
3110          * it will be handled
3111          */
3112         spin_lock(&obd->obd_recovery_task_lock);
3113         if (transno < obd->obd_next_recovery_transno &&
3114             !is_req_replayed_by_update(req)) {
3115                 /* Processing the queue right now, don't re-add. */
3116                 LASSERT(list_empty(&req->rq_list));
3117                 spin_unlock(&obd->obd_recovery_task_lock);
3118                 RETURN(1);
3119         }
3120         spin_unlock(&obd->obd_recovery_task_lock);
3121
3122         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3123                 RETURN(0);
3124
3125         target_request_copy_get(req);
3126         if (!req->rq_export->exp_in_recovery) {
3127                 target_request_copy_put(req);
3128                 RETURN(-ENOTCONN);
3129         }
3130         LASSERT(req->rq_export->exp_req_replay_needed);
3131
3132         if (target_exp_enqueue_req_replay(req)) {
3133                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3134                 target_request_copy_put(req);
3135                 RETURN(0);
3136         }
3137
3138         /* XXX O(n^2) */
3139         spin_lock(&obd->obd_recovery_task_lock);
3140         LASSERT(obd->obd_recovering);
3141         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3142                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3143                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3144                         inserted = 1;
3145                         goto added;
3146                 }
3147
3148                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3149                              transno)) {
3150                         DEBUG_REQ(D_ERROR, req,
3151                                   "dropping replay: transno has been claimed by another client");
3152                         spin_unlock(&obd->obd_recovery_task_lock);
3153                         target_exp_dequeue_req_replay(req);
3154                         target_request_copy_put(req);
3155                         RETURN(0);
3156                 }
3157         }
3158 added:
3159         if (!inserted)
3160                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3161
3162         obd->obd_requests_queued_for_recovery++;
3163         spin_unlock(&obd->obd_recovery_task_lock);
3164         wake_up(&obd->obd_next_transno_waitq);
3165         RETURN(0);
3166 }
3167
3168 void target_committed_to_req(struct ptlrpc_request *req)
3169 {
3170         struct obd_export *exp = req->rq_export;
3171
3172         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3173                 lustre_msg_set_last_committed(req->rq_repmsg,
3174                                               exp->exp_last_committed);
3175         else
3176                 DEBUG_REQ(D_IOCTL, req,
3177                           "not sending last_committed update (%d/%d)",
3178                           exp->exp_obd->obd_no_transno,
3179                           req->rq_repmsg == NULL);
3180
3181         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3182                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3183 }
3184
3185 #endif /* HAVE_SERVER_SUPPORT */
3186
3187 /**
3188  * Packs current SLV and Limit into \a req.
3189  */
3190 int target_pack_pool_reply(struct ptlrpc_request *req)
3191 {
3192         struct obd_device *obd;
3193
3194         ENTRY;
3195
3196         /*
3197          * Check that we still have all structures alive as this may
3198          * be some late RPC at shutdown time.
3199          */
3200         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3201                      !exp_connect_lru_resize(req->rq_export))) {
3202                 lustre_msg_set_slv(req->rq_repmsg, 0);
3203                 lustre_msg_set_limit(req->rq_repmsg, 0);
3204                 RETURN(0);
3205         }
3206
3207         /* OBD is alive here as export is alive, which we checked above. */
3208         obd = req->rq_export->exp_obd;
3209
3210         read_lock(&obd->obd_pool_lock);
3211         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3212         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3213         read_unlock(&obd->obd_pool_lock);
3214
3215         RETURN(0);
3216 }
3217
3218 static int target_send_reply_msg(struct ptlrpc_request *req,
3219                                  int rc, int fail_id)
3220 {
3221         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3222                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3223                 return -ECOMM;
3224         }
3225         /*
3226          * We can have a null rq_reqmsg in the event of bad signature or
3227          * no context when unwrapping
3228          */
3229         if (req->rq_reqmsg &&
3230             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3231             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3232                 return -ECOMM;
3233
3234         if (unlikely(rc)) {
3235                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3236                 req->rq_status = rc;
3237                 return ptlrpc_send_error(req, 1);
3238         }
3239         DEBUG_REQ(D_NET, req, "sending reply");
3240
3241         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3242 }
3243
3244 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3245 {
3246         struct ptlrpc_service_part *svcpt;
3247         int netrc;
3248         struct ptlrpc_reply_state *rs;
3249         struct obd_export *exp;
3250
3251         ENTRY;
3252
3253         if (req->rq_no_reply) {
3254                 EXIT;
3255                 return;
3256         }
3257
3258         svcpt = req->rq_rqbd->rqbd_svcpt;
3259         rs = req->rq_reply_state;
3260         if (rs == NULL || !rs->rs_difficult) {
3261                 /* no notifiers */
3262                 target_send_reply_msg(req, rc, fail_id);
3263                 EXIT;
3264                 return;
3265         }
3266
3267         /* must be an export if locks saved */
3268         LASSERT(req->rq_export != NULL);
3269         /* req/reply consistent */
3270         LASSERT(rs->rs_svcpt == svcpt);
3271
3272         /* "fresh" reply */
3273         LASSERT(!rs->rs_scheduled);
3274         LASSERT(!rs->rs_scheduled_ever);
3275         LASSERT(!rs->rs_handled);
3276         LASSERT(!rs->rs_sent);
3277         LASSERT(!rs->rs_unlinked);
3278         LASSERT(rs->rs_export == NULL);
3279         LASSERT(list_empty(&rs->rs_obd_list));
3280         LASSERT(list_empty(&rs->rs_exp_list));
3281
3282         exp = class_export_get(req->rq_export);
3283
3284         /* disable reply scheduling while I'm setting up */
3285         rs->rs_scheduled = 1;
3286         rs->rs_sent      = 0;
3287         rs->rs_unlinked  = 0;
3288         rs->rs_xid       = req->rq_xid;
3289         rs->rs_transno   = req->rq_transno;
3290         rs->rs_export    = exp;
3291         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3292
3293         spin_lock(&exp->exp_uncommitted_replies_lock);
3294         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3295                rs->rs_transno, exp->exp_last_committed);
3296         if (rs->rs_transno > exp->exp_last_committed) {
3297                 /* not committed already */
3298                 list_add_tail(&rs->rs_obd_list,
3299                                   &exp->exp_uncommitted_replies);
3300         }
3301         spin_unlock(&exp->exp_uncommitted_replies_lock);
3302
3303         spin_lock(&exp->exp_lock);
3304         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3305         spin_unlock(&exp->exp_lock);
3306
3307         netrc = target_send_reply_msg(req, rc, fail_id);
3308
3309         spin_lock(&svcpt->scp_rep_lock);
3310
3311         atomic_inc(&svcpt->scp_nreps_difficult);
3312
3313         if (netrc != 0) {
3314                 /*
3315                  * error sending: reply is off the net.  Also we need +1
3316                  * reply ref until ptlrpc_handle_rs() is done
3317                  * with the reply state (if the send was successful, there
3318                  * would have been +1 ref for the net, which
3319                  * reply_out_callback leaves alone)
3320                  */
3321                 rs->rs_sent = 1;
3322                 rs->rs_unlinked = 1;
3323                 ptlrpc_rs_addref(rs);
3324         }
3325
3326         spin_lock(&rs->rs_lock);
3327         if (rs->rs_transno <= exp->exp_last_committed ||
3328             (rs->rs_unlinked && !rs->rs_no_ack) ||
3329             list_empty(&rs->rs_exp_list) ||     /* completed already */
3330             list_empty(&rs->rs_obd_list)) {
3331                 CDEBUG(D_HA, "Schedule reply immediately\n");
3332                 ptlrpc_dispatch_difficult_reply(rs);
3333         } else {
3334                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3335                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3336         }
3337         spin_unlock(&rs->rs_lock);
3338         spin_unlock(&svcpt->scp_rep_lock);
3339         EXIT;
3340 }
3341
3342 enum ldlm_mode lck_compat_array[] = {
3343         [LCK_EX]    = LCK_COMPAT_EX,
3344         [LCK_PW]    = LCK_COMPAT_PW,
3345         [LCK_PR]    = LCK_COMPAT_PR,
3346         [LCK_CW]    = LCK_COMPAT_CW,
3347         [LCK_CR]    = LCK_COMPAT_CR,
3348         [LCK_NL]    = LCK_COMPAT_NL,
3349         [LCK_GROUP] = LCK_COMPAT_GROUP,
3350         [LCK_COS]   = LCK_COMPAT_COS,
3351 };
3352
3353 /**
3354  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3355  * not escape to the user level.
3356  */
3357 int ldlm_error2errno(enum ldlm_error error)
3358 {
3359         int result;
3360
3361         switch (error) {
3362         case ELDLM_OK:
3363         case ELDLM_LOCK_MATCHED:
3364                 result = 0;
3365                 break;
3366         case ELDLM_LOCK_CHANGED:
3367                 result = -ESTALE;
3368                 break;
3369         case ELDLM_LOCK_ABORTED:
3370                 result = -ENAVAIL;
3371                 break;
3372         case ELDLM_LOCK_REPLACED:
3373                 result = -ESRCH;
3374                 break;
3375         case ELDLM_NO_LOCK_DATA:
3376                 result = -ENOENT;
3377                 break;
3378         case ELDLM_NAMESPACE_EXISTS:
3379                 result = -EEXIST;
3380                 break;
3381         case ELDLM_BAD_NAMESPACE:
3382                 result = -EBADF;
3383                 break;
3384         default:
3385                 if (((int)error) < 0) { /* cast to signed type */
3386                         result = error; /* as ldlm_error can be unsigned */
3387                 } else {
3388                         CERROR("Invalid DLM result code: %d\n", error);
3389                         result = -EPROTO;
3390                 }
3391         }
3392         return result;
3393 }
3394 EXPORT_SYMBOL(ldlm_error2errno);
3395
3396 /**
3397  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3398  */
3399 enum ldlm_error ldlm_errno2error(int err_no)
3400 {
3401         int error;
3402
3403         switch (err_no) {
3404         case 0:
3405                 error = ELDLM_OK;
3406                 break;
3407         case -ESTALE:
3408                 error = ELDLM_LOCK_CHANGED;
3409                 break;
3410         case -ENAVAIL:
3411                 error = ELDLM_LOCK_ABORTED;
3412                 break;
3413         case -ESRCH:
3414                 error = ELDLM_LOCK_REPLACED;
3415                 break;
3416         case -ENOENT:
3417                 error = ELDLM_NO_LOCK_DATA;
3418                 break;
3419         case -EEXIST:
3420                 error = ELDLM_NAMESPACE_EXISTS;
3421                 break;
3422         case -EBADF:
3423                 error = ELDLM_BAD_NAMESPACE;
3424                 break;
3425         default:
3426                 error = err_no;
3427         }
3428         return error;
3429 }
3430
3431 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3432 void ldlm_dump_export_locks(struct obd_export *exp)
3433 {
3434         spin_lock(&exp->exp_locks_list_guard);
3435         if (!list_empty(&exp->exp_locks_list)) {
3436                 struct ldlm_lock *lock;
3437
3438                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3439                        exp);
3440                 list_for_each_entry(lock, &exp->exp_locks_list,
3441                                         l_exp_refs_link)
3442                         LDLM_ERROR(lock, "lock:");
3443         }
3444         spin_unlock(&exp->exp_locks_list_guard);
3445 }
3446 #endif
3447
3448 #ifdef HAVE_SERVER_SUPPORT
3449 static inline const char *bulk2type(struct ptlrpc_request *req)
3450 {
3451         if (req->rq_bulk_read)
3452                 return "READ";
3453         if (req->rq_bulk_write)
3454                 return "WRITE";
3455         return "UNKNOWN";
3456 }
3457
3458 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3459 {
3460         struct ptlrpc_request *req = desc->bd_req;
3461         time64_t start = ktime_get_seconds();
3462         time64_t deadline;
3463         int rc = 0;
3464
3465         ENTRY;
3466
3467         /* If there is eviction in progress, wait for it to finish. */
3468         wait_event_idle(
3469                 exp->exp_obd->obd_evict_inprogress_waitq,
3470                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3471
3472         /* Check if client was evicted or reconnected already. */
3473         if (exp->exp_failed ||
3474             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3475                 rc = -ENOTCONN;
3476         } else {
3477                 if (req->rq_bulk_read)
3478                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3479
3480                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3481                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3482                 else /* old version, bulk matchbits is rq_xid */
3483                         req->rq_mbits = req->rq_xid;
3484
3485                 if (rc == 0)
3486                         rc = ptlrpc_start_bulk_transfer(desc);
3487         }
3488
3489         if (rc < 0) {
3490                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3491                           bulk2type(req), rc);
3492                 RETURN(rc);
3493         }
3494
3495         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3496                 ptlrpc_abort_bulk(desc);
3497                 RETURN(0);
3498         }
3499
3500         /* limit actual bulk transfer to bulk_timeout seconds */
3501         deadline = start + bulk_timeout;
3502         if (deadline > req->rq_deadline)
3503                 deadline = req->rq_deadline;
3504
3505         do {
3506                 time64_t timeoutl = deadline - ktime_get_seconds();
3507                 time64_t rq_deadline;
3508
3509                 while (timeoutl >= 0 &&
3510                        wait_event_idle_timeout(
3511                                desc->bd_waitq,
3512                                !ptlrpc_server_bulk_active(desc) ||
3513                                exp->exp_failed ||
3514                                exp->exp_conn_cnt >
3515                                lustre_msg_get_conn_cnt(req->rq_reqmsg),
3516                                timeoutl ? cfs_time_seconds(1) : 1) == 0)
3517                         timeoutl -= 1;
3518                 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3519
3520                 /* Wait again if we changed rq_deadline. */
3521                 rq_deadline = READ_ONCE(req->rq_deadline);
3522                 deadline = start + bulk_timeout;
3523                 if (deadline > rq_deadline)
3524                         deadline = rq_deadline;
3525         } while (rc == -ETIMEDOUT &&
3526                  deadline > ktime_get_seconds());
3527
3528         if (rc == -ETIMEDOUT) {
3529                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3530                           bulk2type(req), deadline - start,
3531                           ktime_get_real_seconds() - deadline);
3532                 ptlrpc_abort_bulk(desc);
3533         } else if (exp->exp_failed) {
3534                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3535                           bulk2type(req));
3536                 rc = -ENOTCONN;
3537                 ptlrpc_abort_bulk(desc);
3538         } else if (exp->exp_conn_cnt >
3539                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3540                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3541                           bulk2type(req));
3542                 /* We don't reply anyway. */
3543                 rc = -ETIMEDOUT;
3544                 ptlrpc_abort_bulk(desc);
3545         } else if (desc->bd_failure) {
3546                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3547                           bulk2type(req));
3548                 /* XXX should this be a different errno? */
3549                 rc = -ETIMEDOUT;
3550         } else {
3551                 if (req->rq_bulk_write)
3552                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3553                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3554                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3555                                   bulk2type(req), desc->bd_nob_transferred,
3556                                   desc->bd_nob);
3557                         /* XXX should this be a different errno? */
3558                         rc = -ETIMEDOUT;
3559                 }
3560         }
3561
3562         RETURN(rc);
3563 }
3564 EXPORT_SYMBOL(target_bulk_io);
3565
3566 #endif /* HAVE_SERVER_SUPPORT */