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_id2str(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 = NULL;
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_nid2str(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_nid2str(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_nid2str(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_nid2str(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_nid2str(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 (req->rq_peer.nid !=
1274                     lnet_nid_to_nid4(&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_nid2str(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_nid2str(req->rq_peer.nid),
1287                                       OCD_HAS_FLAG(data, MDS_MDS) ?
1288                                       "keep" : "remove");
1289                 }
1290
1291                 if (req->rq_peer.nid ==
1292                     lnet_nid_to_nid4(&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                    req->rq_peer.nid != lnet_nid_to_nid4(&export->exp_connection->c_peer.nid)) {
1315                 spin_unlock(&export->exp_lock);
1316                 /* In MDS failover we have static UUID but NID can change. */
1317                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1318                               target->obd_name, cluuid.uuid,
1319                               libcfs_nid2str(req->rq_peer.nid),
1320                               libcfs_nidstr(
1321                                       &export->exp_connection->c_peer.nid));
1322                 rc = -EALREADY;
1323                 class_export_put(export);
1324                 export = NULL;
1325         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1326                    !lw_client) {
1327                 spin_unlock(&export->exp_lock);
1328                 rc = -EAGAIN;
1329         } else {
1330                 export->exp_connecting = 1;
1331                 spin_unlock(&export->exp_lock);
1332                 LASSERT(export->exp_obd == target);
1333
1334                 rc = target_handle_reconnect(&conn, export, &cluuid);
1335         }
1336
1337         /* If we found an export, we already unlocked. */
1338         if (!export) {
1339 no_export:
1340                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1341         } else if (req->rq_export == NULL &&
1342                    atomic_read(&export->exp_rpc_count) > 0) {
1343                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1344                               target->obd_name, cluuid.uuid,
1345                               libcfs_nid2str(req->rq_peer.nid),
1346                               refcount_read(&export->exp_handle.h_ref));
1347                         GOTO(out, rc = -EBUSY);
1348         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1349                    rc != EALREADY) {
1350                 if (!strstr(cluuid.uuid, "mdt"))
1351                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1352                                       target->obd_name, cluuid.uuid,
1353                                       libcfs_nid2str(req->rq_peer.nid));
1354                 GOTO(out, rc = -EALREADY);
1355         } else {
1356                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1357         }
1358
1359         if (rc < 0)
1360                 GOTO(out, rc);
1361
1362         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1363                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1364                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1365                export, ktime_get_seconds(),
1366                export ? export->exp_last_request_time : 0);
1367
1368         /*
1369          * If this is the first time a client connects, reset the recovery
1370          * timer. Discard lightweight connections which might be local.
1371          */
1372         if (!lw_client && rc == 0 && target->obd_recovering)
1373                 check_and_start_recovery_timer(target, req, export == NULL);
1374
1375         /*
1376          * We want to handle EALREADY but *not* -EALREADY from
1377          * target_handle_reconnect(), return reconnection state in a flag.
1378          */
1379         if (rc == EALREADY) {
1380                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1381                 rc = 0;
1382         } else {
1383                 LASSERT(rc == 0);
1384         }
1385
1386         /* Tell the client if we support replayable requests. */
1387         if (target->obd_replayable)
1388                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1389         client_nid = &req->rq_peer.nid;
1390
1391         if (export == NULL) {
1392                 /* allow lightweight connections during recovery */
1393                 /*
1394                  * allow "new" MDT to be connected during recovery, since we
1395                  * need retrieve recovery update records from it
1396                  */
1397                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1398                         struct hrtimer *timer = &target->obd_recovery_timer;
1399                         ktime_t remaining;
1400                         s64 timeout, left;
1401                         int in_progress;
1402                         int connected;
1403                         int known;
1404                         int stale;
1405                         char *msg;
1406
1407                         connected = atomic_read(&target->obd_connected_clients);
1408                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1409                         known =
1410                            atomic_read(&target->obd_max_recoverable_clients);
1411                         stale = target->obd_stale_clients;
1412                         remaining = hrtimer_get_remaining(timer);
1413                         left = ktime_divns(remaining, NSEC_PER_SEC);
1414
1415                         if (ktime_to_ns(remaining) > 0) {
1416                                 msg = "to recover in";
1417                                 timeout = left;
1418                         } else {
1419                                 msg = "already passed deadline";
1420                                 timeout = -left;
1421
1422                                 target_check_recovery_timer(target);
1423                         }
1424
1425                         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",
1426                                       target->obd_name, cluuid.uuid,
1427                                       libcfs_nid2str(req->rq_peer.nid), known,
1428                                       connected - in_progress, in_progress,
1429                                       stale, msg, timeout / 60, timeout % 60);
1430                         rc = -EBUSY;
1431                 } else {
1432 dont_check_exports:
1433                         rc = obd_connect(req->rq_svc_thread->t_env,
1434                                          &export, target, &cluuid, data,
1435                                          client_nid);
1436                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1437                                 lustre_msg_add_op_flags(req->rq_repmsg,
1438                                                         MSG_CONNECT_RECOVERING);
1439                         if (rc == 0) {
1440                                 conn.cookie = export->exp_handle.h_cookie;
1441                                 rc = rev_import_init(export);
1442                         }
1443
1444                         if (mds_mds_conn)
1445                                 new_mds_mds_conn = true;
1446                 }
1447         } else {
1448                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1449                                    export, target, &cluuid, data, client_nid);
1450         }
1451         if (rc)
1452                 GOTO(out, rc);
1453
1454         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1455         data->ocd_instance = target->u.obt.obt_instance;
1456
1457         /*
1458          * Return only the parts of obd_connect_data that we understand, so the
1459          * client knows that we don't understand the rest.
1460          */
1461         if (data) {
1462                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1463                                                RCL_SERVER);
1464                 tmpdata = req_capsule_server_get(&req->rq_pill,
1465                                                  &RMF_CONNECT_DATA);
1466                 /*
1467                  * Don't use struct assignment here, because the client reply
1468                  * buffer may be smaller/larger than the local struct
1469                  * obd_connect_data.
1470                  */
1471                 memcpy(tmpdata, data, min(tmpsize, size));
1472         }
1473
1474         /*
1475          * If the client and the server are the same node, we will already
1476          * have an export that really points to the client's DLM export,
1477          * because we have a shared handles table.
1478          *
1479          * XXX this will go away when shaver stops sending the "connect" handle
1480          * in the real "remote handle" field of the request --phik 24 Apr 2003
1481          */
1482         ptlrpc_request_change_export(req, export);
1483
1484         pcon = ptlrpc_connection_get(req->rq_peer,
1485                                      &req->rq_self, &cluuid);
1486         if (pcon == NULL)
1487                 GOTO(out, rc = -ENOTCONN);
1488
1489         spin_lock(&export->exp_lock);
1490
1491         if (export->exp_disconnected) {
1492                 spin_unlock(&export->exp_lock);
1493                 GOTO(out, rc = -ENODEV);
1494         }
1495         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1496                 spin_unlock(&export->exp_lock);
1497                 CDEBUG(D_RPCTRACE,
1498                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1499                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1500                        export->exp_conn_cnt,
1501                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1502
1503                 GOTO(out, rc = -EALREADY);
1504         }
1505         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1506         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1507
1508         /* Check to see if connection came from another NID. */
1509         if (export->exp_connection != NULL &&
1510             lnet_nid_to_nid4(&export->exp_connection->c_peer.nid) !=
1511             req->rq_peer.nid) {
1512                 obd_nid_del(export->exp_obd, export);
1513                 ptlrpc_connection_put(export->exp_connection);
1514                 export->exp_connection = NULL;
1515         }
1516
1517         if (export->exp_connection == NULL) {
1518                 export->exp_connection = pcon;
1519                 pcon = NULL;
1520         }
1521         obd_nid_add(export->exp_obd, export);
1522
1523         spin_unlock(&export->exp_lock);
1524
1525         lustre_msg_set_handle(req->rq_repmsg, &conn);
1526
1527         rc = rev_import_reconnect(export, req);
1528         if (rc != 0)
1529                 GOTO(out, rc);
1530
1531         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1532                 int has_transno;
1533                 __u64 transno = data->ocd_transno;
1534
1535                 spin_lock(&export->exp_lock);
1536                 /*
1537                  * possible race with class_disconnect_stale_exports,
1538                  * export may be already in the eviction process
1539                  */
1540                 if (export->exp_failed) {
1541                         spin_unlock(&export->exp_lock);
1542                         GOTO(out, rc = -ENODEV);
1543                 }
1544                 export->exp_in_recovery = 1;
1545                 export->exp_req_replay_needed = 1;
1546                 export->exp_lock_replay_needed = 1;
1547                 spin_unlock(&export->exp_lock);
1548
1549                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1550                                  MSG_CONNECT_TRANSNO);
1551                 if (has_transno && transno == 0)
1552                         CWARN("Connect with zero transno!\n");
1553
1554                 if (has_transno && transno > 0 &&
1555                     transno < target->obd_next_recovery_transno &&
1556                     transno > target->obd_last_committed) {
1557                         /* Another way is to use cmpxchg() to be lock-free. */
1558                         spin_lock(&target->obd_recovery_task_lock);
1559                         if (transno < target->obd_next_recovery_transno)
1560                                 target->obd_next_recovery_transno = transno;
1561                         spin_unlock(&target->obd_recovery_task_lock);
1562                 }
1563
1564                 atomic_inc(&target->obd_req_replay_clients);
1565                 atomic_inc(&target->obd_lock_replay_clients);
1566                 /*
1567                  * Note: MDS-MDS connection is allowed to be connected during
1568                  * recovery, no matter if the exports needs to be recoveried.
1569                  * Because we need retrieve updates logs from all other MDTs.
1570                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1571                  * also needs to be increased to match other recovery checking
1572                  * condition.
1573                  */
1574                 if (new_mds_mds_conn)
1575                         atomic_inc(&target->obd_max_recoverable_clients);
1576
1577                 if (atomic_inc_return(&target->obd_connected_clients) ==
1578                     atomic_read(&target->obd_max_recoverable_clients))
1579                         wake_up(&target->obd_next_transno_waitq);
1580         }
1581
1582         /* Tell the client we're in recovery, when client is involved in it. */
1583         if (target->obd_recovering && !lw_client)
1584                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1585
1586 out:
1587         if (export) {
1588                 spin_lock(&export->exp_lock);
1589                 export->exp_connecting = 0;
1590                 spin_unlock(&export->exp_lock);
1591
1592                 class_export_put(export);
1593         }
1594         if (target != NULL) {
1595                 spin_lock(&target->obd_dev_lock);
1596                 target->obd_conn_inprogress--;
1597                 spin_unlock(&target->obd_dev_lock);
1598                 class_decref(target, "find", current);
1599         }
1600         if (pcon)
1601                 ptlrpc_connection_put(pcon);
1602         req->rq_status = rc;
1603         RETURN(rc);
1604 }
1605
1606 int target_handle_disconnect(struct ptlrpc_request *req)
1607 {
1608         int rc;
1609
1610         ENTRY;
1611
1612         rc = req_capsule_server_pack(&req->rq_pill);
1613         if (rc)
1614                 RETURN(rc);
1615
1616         /* In case of target disconnect, updating sec ctx immediately is
1617          * required in order to record latest sequence number used.
1618          * Sequence is normally updated on export destroy, but this event
1619          * can occur too late, ie after a new target connect request has
1620          * been processed.
1621          * Maintaining correct sequence when client connection becomes idle
1622          * ensures that GSS does not erroneously consider requests as replays.
1623          */
1624         rc = sptlrpc_export_update_ctx(req->rq_export);
1625         if (rc)
1626                 RETURN(rc);
1627
1628         /* Keep the rq_export around so we can send the reply. */
1629         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1630
1631         RETURN(0);
1632 }
1633
1634 void target_destroy_export(struct obd_export *exp)
1635 {
1636         struct obd_import *imp = NULL;
1637         /*
1638          * exports created from last_rcvd data, and "fake"
1639          * exports created by lctl don't have an import
1640          */
1641         spin_lock(&exp->exp_lock);
1642         if (exp->exp_imp_reverse != NULL) {
1643                 imp = exp->exp_imp_reverse;
1644                 exp->exp_imp_reverse = NULL;
1645         }
1646         spin_unlock(&exp->exp_lock);
1647         if (imp != NULL)
1648                 client_destroy_import(imp);
1649
1650         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1651         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1652         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1653         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1654 }
1655 EXPORT_SYMBOL(target_destroy_export);
1656
1657 /*
1658  * Recovery functions
1659  */
1660 static void target_request_copy_get(struct ptlrpc_request *req)
1661 {
1662         class_export_rpc_inc(req->rq_export);
1663         LASSERT(list_empty(&req->rq_list));
1664         INIT_LIST_HEAD(&req->rq_replay_list);
1665
1666         /* Increase refcount to keep request in queue. */
1667         atomic_inc(&req->rq_refcount);
1668         /* Let export know it has replays to be handled. */
1669         atomic_inc(&req->rq_export->exp_replay_count);
1670 }
1671
1672 static void target_request_copy_put(struct ptlrpc_request *req)
1673 {
1674         LASSERT(list_empty(&req->rq_replay_list));
1675         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1676
1677         atomic_dec(&req->rq_export->exp_replay_count);
1678         class_export_rpc_dec(req->rq_export);
1679         ptlrpc_server_drop_request(req);
1680 }
1681
1682 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1683 {
1684         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1685         struct obd_export *exp = req->rq_export;
1686         struct ptlrpc_request *reqiter;
1687         struct ptlrpc_request *dup_req = NULL;
1688         int dup = 0;
1689
1690         LASSERT(exp);
1691
1692         spin_lock(&exp->exp_lock);
1693         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1694                             rq_replay_list) {
1695                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1696                         dup_req = reqiter;
1697                         dup = 1;
1698                         break;
1699                 }
1700         }
1701
1702         if (dup) {
1703                 /* We expect it with RESENT and REPLAY flags. */
1704                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1705                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1706                         CERROR("invalid flags %x of resent replay\n",
1707                                lustre_msg_get_flags(req->rq_reqmsg));
1708
1709                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1710                         __u32 new_conn;
1711
1712                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1713                         if (new_conn >
1714                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1715                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1716                                                         new_conn);
1717                 }
1718         } else {
1719                 list_add_tail(&req->rq_replay_list,
1720                               &exp->exp_req_replay_queue);
1721         }
1722
1723         spin_unlock(&exp->exp_lock);
1724         return dup;
1725 }
1726
1727 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1728 {
1729         LASSERT(!list_empty(&req->rq_replay_list));
1730         LASSERT(req->rq_export);
1731
1732         spin_lock(&req->rq_export->exp_lock);
1733         list_del_init(&req->rq_replay_list);
1734         spin_unlock(&req->rq_export->exp_lock);
1735 }
1736
1737 static void target_finish_recovery(struct lu_target *lut)
1738 {
1739         struct obd_device *obd = lut->lut_obd;
1740
1741         ENTRY;
1742
1743         /* Only log a recovery message when recovery has occurred. */
1744         if (obd->obd_recovery_start) {
1745                 time64_t now = ktime_get_seconds();
1746                 time64_t elapsed_time;
1747
1748                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1749                                      1);
1750                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1751                               obd->obd_name, elapsed_time / 60,
1752                               elapsed_time % 60,
1753                               atomic_read(&obd->obd_max_recoverable_clients),
1754                               atomic_read(&obd->obd_connected_clients),
1755                               obd->obd_stale_clients,
1756                               obd->obd_stale_clients == 1 ? "was" : "were");
1757                 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1758                         libcfs_debug_dumplog();
1759         }
1760
1761         ldlm_reprocess_recovery_done(obd->obd_namespace);
1762         spin_lock(&obd->obd_recovery_task_lock);
1763         if (!list_empty(&obd->obd_req_replay_queue) ||
1764             !list_empty(&obd->obd_lock_replay_queue) ||
1765             !list_empty(&obd->obd_final_req_queue)) {
1766                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1767                        obd->obd_name,
1768                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1769                        list_empty(&obd->obd_lock_replay_queue) ? \
1770                                   "" : "lock ",
1771                        list_empty(&obd->obd_final_req_queue) ? \
1772                                   "" : "final ");
1773                 spin_unlock(&obd->obd_recovery_task_lock);
1774                 LBUG();
1775         }
1776         spin_unlock(&obd->obd_recovery_task_lock);
1777
1778         obd->obd_recovery_end = ktime_get_seconds();
1779
1780         /* When recovery finished, cleanup orphans on MDS and OST. */
1781         if (obd->obd_type && OBP(obd, postrecov)) {
1782                 int rc = OBP(obd, postrecov)(obd);
1783
1784                 if (rc < 0)
1785                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1786                                       obd->obd_name, rc);
1787         }
1788         EXIT;
1789 }
1790
1791 static void abort_req_replay_queue(struct obd_device *obd)
1792 {
1793         struct ptlrpc_request *req, *n;
1794         LIST_HEAD(abort_list);
1795
1796         spin_lock(&obd->obd_recovery_task_lock);
1797         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1798         spin_unlock(&obd->obd_recovery_task_lock);
1799         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1800                 DEBUG_REQ(D_WARNING, req, "aborted:");
1801                 req->rq_status = -ENOTCONN;
1802                 if (ptlrpc_error(req)) {
1803                         DEBUG_REQ(D_ERROR, req,
1804                                   "failed abort_req_reply; skipping");
1805                 }
1806                 target_exp_dequeue_req_replay(req);
1807                 target_request_copy_put(req);
1808         }
1809 }
1810
1811 static void abort_lock_replay_queue(struct obd_device *obd)
1812 {
1813         struct ptlrpc_request *req, *n;
1814         LIST_HEAD(abort_list);
1815
1816         spin_lock(&obd->obd_recovery_task_lock);
1817         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1818         spin_unlock(&obd->obd_recovery_task_lock);
1819         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1820                 DEBUG_REQ(D_ERROR, req, "aborted:");
1821                 req->rq_status = -ENOTCONN;
1822                 if (ptlrpc_error(req)) {
1823                         DEBUG_REQ(D_ERROR, req,
1824                                   "failed abort_lock_reply; skipping");
1825                 }
1826                 target_request_copy_put(req);
1827         }
1828 }
1829
1830 /*
1831  * Called from a cleanup function if the device is being cleaned up
1832  * forcefully.  The exports should all have been disconnected already,
1833  * the only thing left to do is
1834  * - clear the recovery flags
1835  * - cancel the timer
1836  * - free queued requests and replies, but don't send replies
1837  * Because the obd_stopping flag is set, no new requests should be received.
1838  */
1839 void target_cleanup_recovery(struct obd_device *obd)
1840 {
1841         struct ptlrpc_request *req, *n;
1842         LIST_HEAD(clean_list);
1843
1844         spin_lock(&obd->obd_dev_lock);
1845         if (!obd->obd_recovering) {
1846                 spin_unlock(&obd->obd_dev_lock);
1847                 EXIT;
1848                 return;
1849         }
1850         obd->obd_recovering = obd->obd_abort_recovery = 0;
1851         obd->obd_abort_recov_mdt = 0;
1852         spin_unlock(&obd->obd_dev_lock);
1853
1854         spin_lock(&obd->obd_recovery_task_lock);
1855         target_cancel_recovery_timer(obd);
1856         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1857         spin_unlock(&obd->obd_recovery_task_lock);
1858
1859         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1860                 LASSERT(req->rq_reply_state == NULL);
1861                 target_exp_dequeue_req_replay(req);
1862                 target_request_copy_put(req);
1863         }
1864
1865         spin_lock(&obd->obd_recovery_task_lock);
1866         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1867         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1868         spin_unlock(&obd->obd_recovery_task_lock);
1869
1870         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1871                 LASSERT(req->rq_reply_state == NULL);
1872                 target_request_copy_put(req);
1873         }
1874
1875         EXIT;
1876 }
1877 EXPORT_SYMBOL(target_cleanup_recovery);
1878
1879 /* obd_recovery_task_lock should be held */
1880 void target_cancel_recovery_timer(struct obd_device *obd)
1881 {
1882         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1883         hrtimer_cancel(&obd->obd_recovery_timer);
1884 }
1885
1886 static void target_start_recovery_timer(struct obd_device *obd)
1887 {
1888         ktime_t delay;
1889
1890         if (obd->obd_recovery_start != 0)
1891                 return;
1892
1893         spin_lock(&obd->obd_dev_lock);
1894         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1895                 spin_unlock(&obd->obd_dev_lock);
1896                 return;
1897         }
1898
1899         LASSERT(obd->obd_recovery_timeout != 0);
1900
1901         if (obd->obd_recovery_start != 0) {
1902                 spin_unlock(&obd->obd_dev_lock);
1903                 return;
1904         }
1905
1906         obd->obd_recovery_start = ktime_get_seconds();
1907         delay = ktime_set(obd->obd_recovery_start +
1908                           obd->obd_recovery_timeout, 0);
1909         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1910         spin_unlock(&obd->obd_dev_lock);
1911
1912         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1913                       obd->obd_name,
1914                       obd->obd_recovery_timeout / 60,
1915                       obd->obd_recovery_timeout % 60,
1916                       atomic_read(&obd->obd_max_recoverable_clients),
1917                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1918                       "" : "s",
1919                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1920                       "s" : "");
1921 }
1922
1923 /**
1924  * extend recovery window.
1925  *
1926  * if @extend is true, extend recovery window to have @dr_timeout remaining
1927  * at least; otherwise, make sure the recovery timeout value is not less
1928  * than @dr_timeout.
1929  */
1930 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1931                                   bool extend)
1932 {
1933         ktime_t left_ns;
1934         timeout_t timeout;
1935         timeout_t left;
1936
1937         spin_lock(&obd->obd_dev_lock);
1938         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1939             obd->obd_stopping) {
1940                 spin_unlock(&obd->obd_dev_lock);
1941                 return;
1942         }
1943         LASSERT(obd->obd_recovery_start != 0);
1944
1945         left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1946         left = ktime_divns(left_ns, NSEC_PER_SEC);
1947
1948         if (extend) {
1949                 timeout = obd->obd_recovery_timeout;
1950                 /* dr_timeout will happen after the hrtimer has expired.
1951                  * Add the excess time to the soft recovery timeout without
1952                  * exceeding the hard recovery timeout.
1953                  */
1954                 if (dr_timeout > left) {
1955                         timeout += dr_timeout - left;
1956                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1957                                         timeout);
1958                 }
1959         } else {
1960                 timeout = clamp_t(timeout_t, dr_timeout,
1961                                   obd->obd_recovery_timeout,
1962                                   obd->obd_recovery_time_hard);
1963         }
1964
1965         if (timeout == obd->obd_recovery_time_hard)
1966                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1967                       obd->obd_name, timeout, extend);
1968
1969         if (obd->obd_recovery_timeout < timeout) {
1970                 ktime_t end, now;
1971
1972                 obd->obd_recovery_timeout = timeout;
1973                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1974                 now = ktime_set(ktime_get_seconds(), 0);
1975                 left_ns = ktime_sub(end, now);
1976                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1977                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1978         }
1979         spin_unlock(&obd->obd_dev_lock);
1980
1981         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1982                 obd->obd_name, left);
1983 }
1984
1985 /* Reset the timer with each new client connection */
1986 /*
1987  * This timer is actually reconnect_timer, which is for making sure
1988  * the total recovery window is at least as big as my reconnect
1989  * attempt timing. So the initial recovery time_out will be set to
1990  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1991  * from client is bigger than this, then the recovery time_out will
1992  * be extended to make sure the client could be reconnected, in the
1993  * process, the timeout from the new client should be ignored.
1994  */
1995 static void
1996 check_and_start_recovery_timer(struct obd_device *obd,
1997                                struct ptlrpc_request *req,
1998                                int new_client)
1999 {
2000         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2001         struct obd_device_target *obt = &obd->u.obt;
2002
2003         if (!new_client && service_timeout)
2004                 /*
2005                  * Teach server about old server's estimates, as first guess
2006                  * at how long new requests will take.
2007                  */
2008                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2009                             service_timeout);
2010
2011         target_start_recovery_timer(obd);
2012
2013         /*
2014          * Convert the service time to RPC timeout,
2015          * and reuse service_timeout to limit stack usage.
2016          */
2017         service_timeout = at_est2timeout(service_timeout);
2018
2019         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2020             service_timeout < at_extra)
2021                 service_timeout = at_extra;
2022
2023         /*
2024          * We expect other clients to timeout within service_timeout, then try
2025          * to reconnect, then try the failover server.  The max delay between
2026          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2027          */
2028         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2029
2030         LASSERT(obt->obt_magic == OBT_MAGIC);
2031         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2032         if (service_timeout > obd->obd_recovery_timeout && !new_client)
2033                 extend_recovery_timer(obd, service_timeout, false);
2034 }
2035
2036 /** Health checking routines */
2037 static inline int exp_connect_healthy(struct obd_export *exp)
2038 {
2039         return exp->exp_in_recovery;
2040 }
2041
2042 /** if export done req_replay or has replay in queue */
2043 static inline int exp_req_replay_healthy(struct obd_export *exp)
2044 {
2045         return (!exp->exp_req_replay_needed ||
2046                 atomic_read(&exp->exp_replay_count) > 0);
2047 }
2048
2049
2050 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2051 {
2052         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2053                exp_req_replay_healthy(exp);
2054 }
2055
2056 /** if export done lock_replay or has replay in queue */
2057 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2058 {
2059         return (!exp->exp_lock_replay_needed ||
2060                 atomic_read(&exp->exp_replay_count) > 0);
2061 }
2062
2063 static inline int exp_vbr_healthy(struct obd_export *exp)
2064 {
2065         return !exp->exp_vbr_failed;
2066 }
2067
2068 static inline int exp_finished(struct obd_export *exp)
2069 {
2070         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2071 }
2072
2073 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2074 {
2075         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2076                 exp_finished(exp);
2077 }
2078
2079 static int check_for_next_transno(struct lu_target *lut)
2080 {
2081         struct ptlrpc_request *req = NULL;
2082         struct obd_device *obd = lut->lut_obd;
2083         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2084         int wake_up = 0, connected, completed, queue_len;
2085         __u64 req_transno = 0;
2086         __u64 update_transno = 0;
2087         __u64 next_transno = 0;
2088
2089         ENTRY;
2090
2091         spin_lock(&obd->obd_recovery_task_lock);
2092         if (!list_empty(&obd->obd_req_replay_queue)) {
2093                 req = list_entry(obd->obd_req_replay_queue.next,
2094                                      struct ptlrpc_request, rq_list);
2095                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2096         }
2097
2098         if (!obd->obd_abort_recov_mdt && tdtd)
2099                 update_transno = distribute_txn_get_next_transno(tdtd);
2100
2101         connected = atomic_read(&obd->obd_connected_clients);
2102         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2103         queue_len = obd->obd_requests_queued_for_recovery;
2104         next_transno = obd->obd_next_recovery_transno;
2105
2106         CDEBUG(D_HA,
2107                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2108                atomic_read(&obd->obd_max_recoverable_clients),
2109                connected, completed,
2110                queue_len, req_transno, next_transno);
2111
2112         if (obd->obd_abort_recovery) {
2113                 CDEBUG(D_HA, "waking for aborted recovery\n");
2114                 wake_up = 1;
2115         } else if (obd->obd_recovery_expired) {
2116                 CDEBUG(D_HA, "waking for expired recovery\n");
2117                 wake_up = 1;
2118         } else if (!obd->obd_abort_recov_mdt && tdtd && req &&
2119                    is_req_replayed_by_update(req)) {
2120                 LASSERTF(req_transno < next_transno,
2121                          "req_transno %llu next_transno%llu\n", req_transno,
2122                          next_transno);
2123                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2124                        req_transno);
2125                 wake_up = 1;
2126         } else if (req_transno == next_transno ||
2127                    (update_transno != 0 && update_transno <= next_transno)) {
2128                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2129                 wake_up = 1;
2130         } else if (queue_len > 0 &&
2131                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2132                 /** handle gaps occured due to lost reply or VBR */
2133                 LASSERTF(req_transno >= next_transno,
2134                          "req_transno: %llu, next_transno: %llu\n",
2135                          req_transno, next_transno);
2136                 CDEBUG(D_HA,
2137                        "%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",
2138                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2139                        next_transno, queue_len, completed, connected,
2140                        req_transno, update_transno, obd->obd_last_committed);
2141                 obd->obd_next_recovery_transno = req_transno;
2142                 wake_up = 1;
2143         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2144                 CDEBUG(D_HA, "waking for completed recovery\n");
2145                 wake_up = 1;
2146         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2147                 CDEBUG(D_HA,
2148                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2149                        next_transno);
2150                 obd->obd_next_recovery_transno = req_transno;
2151                 wake_up = 1;
2152         }
2153         spin_unlock(&obd->obd_recovery_task_lock);
2154         return wake_up;
2155 }
2156
2157 static int check_for_next_lock(struct lu_target *lut)
2158 {
2159         struct obd_device *obd = lut->lut_obd;
2160         int wake_up = 0;
2161
2162         spin_lock(&obd->obd_recovery_task_lock);
2163         if (!list_empty(&obd->obd_lock_replay_queue)) {
2164                 CDEBUG(D_HA, "waking for next lock\n");
2165                 wake_up = 1;
2166         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2167                 CDEBUG(D_HA, "waking for completed lock replay\n");
2168                 wake_up = 1;
2169         } else if (obd->obd_abort_recovery) {
2170                 CDEBUG(D_HA, "waking for aborted recovery\n");
2171                 wake_up = 1;
2172         } else if (obd->obd_recovery_expired) {
2173                 CDEBUG(D_HA, "waking for expired recovery\n");
2174                 wake_up = 1;
2175         }
2176         spin_unlock(&obd->obd_recovery_task_lock);
2177
2178         return wake_up;
2179 }
2180
2181 static int check_update_llog(struct lu_target *lut)
2182 {
2183         struct obd_device *obd = lut->lut_obd;
2184         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2185
2186         if (obd->obd_abort_recovery) {
2187                 CDEBUG(D_HA, "waking for aborted recovery\n");
2188                 return 1;
2189         }
2190
2191         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2192                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2193                 return 1;
2194         }
2195
2196         return 0;
2197 }
2198
2199 /**
2200  * wait for recovery events,
2201  * check its status with help of check_routine
2202  * evict dead clients via health_check
2203  */
2204 static int target_recovery_overseer(struct lu_target *lut,
2205                                     int (*check_routine)(struct lu_target *),
2206                                     int (*health_check)(struct obd_export *))
2207 {
2208         struct obd_device *obd = lut->lut_obd;
2209         struct target_distribute_txn_data *tdtd;
2210         time64_t last = 0;
2211         time64_t now;
2212 repeat:
2213         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2214                 now = ktime_get_seconds();
2215                 if (now - last > 600) {
2216                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2217                                       obd->obd_name);
2218                         last = now;
2219                 }
2220         }
2221         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2222               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2223                 __u64 next_update_transno = 0;
2224
2225                 /*
2226                  * Only abort the recovery if there are no update recovery
2227                  * left in the queue
2228                  */
2229                 spin_lock(&obd->obd_recovery_task_lock);
2230                 if (!obd->obd_abort_recov_mdt && lut->lut_tdtd) {
2231                         next_update_transno =
2232                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2233
2234                         tdtd = lut->lut_tdtd;
2235                         /*
2236                          * If next_update_transno == 0, it probably because
2237                          * updatelog retrieve threads did not get any records
2238                          * yet, let's wait those threads stopped
2239                          */
2240                         if (next_update_transno == 0) {
2241                                 spin_unlock(&obd->obd_recovery_task_lock);
2242
2243                                 while (wait_event_timeout(
2244                                         tdtd->tdtd_recovery_threads_waitq,
2245                                         check_update_llog(lut),
2246                                         cfs_time_seconds(60)) == 0);
2247
2248                                 spin_lock(&obd->obd_recovery_task_lock);
2249                                 next_update_transno =
2250                                         distribute_txn_get_next_transno(tdtd);
2251                         }
2252                 }
2253
2254                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2255                         obd->obd_next_recovery_transno = next_update_transno;
2256                         spin_unlock(&obd->obd_recovery_task_lock);
2257                         /*
2258                          * Disconnect unfinished exports from clients, and
2259                          * keep connection from MDT to make sure the update
2260                          * recovery will still keep trying until some one
2261                          * manually abort the recovery
2262                          */
2263                         class_disconnect_stale_exports(obd,
2264                                                 exp_finished_or_from_mdt);
2265                         /* Abort all of replay & replay lock req from clients */
2266                         abort_req_replay_queue(obd);
2267                         abort_lock_replay_queue(obd);
2268                         CDEBUG(D_HA,
2269                                "%s: there are still update replay (%#llx)in the queue.\n",
2270                                obd->obd_name, next_update_transno);
2271                 } else {
2272                         obd->obd_abort_recovery = 1;
2273                         spin_unlock(&obd->obd_recovery_task_lock);
2274                         CWARN("%s recovery is aborted by hard timeout\n",
2275                               obd->obd_name);
2276                 }
2277         }
2278
2279         while (wait_event_timeout(obd->obd_next_transno_waitq,
2280                                   check_routine(lut),
2281                                   cfs_time_seconds(60)) == 0)
2282                 ; /* wait indefinitely for event, but don't trigger watchdog */
2283
2284         if (obd->obd_abort_recovery) {
2285                 CWARN("recovery is aborted, evict exports in recovery\n");
2286                 if (lut->lut_tdtd != NULL) {
2287                         tdtd = lut->lut_tdtd;
2288                         /*
2289                          * Let's wait all of the update log recovery thread
2290                          * finished
2291                          */
2292                         wait_event_idle(
2293                                 tdtd->tdtd_recovery_threads_waitq,
2294                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2295                                 == 0);
2296                         /* Then abort the update recovery list */
2297                         dtrq_list_destroy(lut->lut_tdtd);
2298                 }
2299
2300                 /** evict exports which didn't finish recovery yet */
2301                 class_disconnect_stale_exports(obd, exp_finished);
2302                 return 1;
2303         } else if (obd->obd_recovery_expired) {
2304                 obd->obd_recovery_expired = 0;
2305
2306                 /** If some clients died being recovered, evict them */
2307                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2308                               obd->obd_name);
2309                 /** evict cexports with no replay in queue, they are stalled */
2310                 class_disconnect_stale_exports(obd, health_check);
2311
2312                 /** continue with VBR */
2313                 spin_lock(&obd->obd_dev_lock);
2314                 obd->obd_version_recov = 1;
2315                 spin_unlock(&obd->obd_dev_lock);
2316                 /**
2317                  * reset timer, recovery will proceed with versions now,
2318                  * timeout is set just to handle reconnection delays
2319                  */
2320                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2321                 /**
2322                  * Wait for recovery events again, after evicting bad clients
2323                  */
2324                 goto repeat;
2325         }
2326         return 0;
2327 }
2328
2329 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2330 {
2331         struct obd_device *obd = lut->lut_obd;
2332         struct ptlrpc_request *req = NULL;
2333
2334         CDEBUG(D_HA, "Waiting for lock\n");
2335         if (target_recovery_overseer(lut, check_for_next_lock,
2336                                      exp_lock_replay_healthy))
2337                 abort_lock_replay_queue(obd);
2338
2339         spin_lock(&obd->obd_recovery_task_lock);
2340         if (!list_empty(&obd->obd_lock_replay_queue)) {
2341                 req = list_entry(obd->obd_lock_replay_queue.next,
2342                                      struct ptlrpc_request, rq_list);
2343                 list_del_init(&req->rq_list);
2344                 spin_unlock(&obd->obd_recovery_task_lock);
2345         } else {
2346                 spin_unlock(&obd->obd_recovery_task_lock);
2347                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2348                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2349                 /** evict exports failed VBR */
2350                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2351         }
2352         return req;
2353 }
2354
2355 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2356 {
2357         struct ptlrpc_request *req = NULL;
2358
2359         spin_lock(&obd->obd_recovery_task_lock);
2360         if (!list_empty(&obd->obd_final_req_queue)) {
2361                 req = list_entry(obd->obd_final_req_queue.next,
2362                                      struct ptlrpc_request, rq_list);
2363                 list_del_init(&req->rq_list);
2364                 spin_unlock(&obd->obd_recovery_task_lock);
2365                 if (req->rq_export->exp_in_recovery) {
2366                         spin_lock(&req->rq_export->exp_lock);
2367                         req->rq_export->exp_in_recovery = 0;
2368                         spin_unlock(&req->rq_export->exp_lock);
2369                 }
2370         } else {
2371                 spin_unlock(&obd->obd_recovery_task_lock);
2372         }
2373         return req;
2374 }
2375
2376 static void handle_recovery_req(struct ptlrpc_thread *thread,
2377                                 struct ptlrpc_request *req,
2378                                 svc_handler_t handler)
2379 {
2380         ENTRY;
2381
2382         /**
2383          * export can be evicted during recovery, no need to handle replays for
2384          * it after that, discard such request silently
2385          */
2386         if (req->rq_export->exp_disconnected)
2387                 RETURN_EXIT;
2388
2389         req->rq_session.lc_thread = thread;
2390         req->rq_svc_thread = thread;
2391         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2392
2393         /* thread context */
2394         lu_context_enter(&thread->t_env->le_ctx);
2395         (void)handler(req);
2396         lu_context_exit(&thread->t_env->le_ctx);
2397
2398         req->rq_svc_thread->t_env->le_ses = NULL;
2399
2400         /* don't reset timer for final stage */
2401         if (!exp_finished(req->rq_export)) {
2402                 timeout_t timeout = obd_timeout;
2403
2404                 /**
2405                  * Add request @timeout to the recovery time so next request from
2406                  * this client may come in recovery time
2407                  */
2408                 if (!AT_OFF) {
2409                         struct ptlrpc_service_part *svcpt;
2410                         timeout_t est_timeout;
2411
2412                         svcpt = req->rq_rqbd->rqbd_svcpt;
2413                         /*
2414                          * If the server sent early reply for this request,
2415                          * the client will recalculate the timeout according to
2416                          * current server estimate service time, so we will
2417                          * use the maxium timeout here for waiting the client
2418                          * sending the next req
2419                          */
2420                         est_timeout = at_get(&svcpt->scp_at_estimate);
2421                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2422                                         lustre_msg_get_timeout(req->rq_reqmsg));
2423                         /*
2424                          * Add 2 net_latency, one for balance rq_deadline
2425                          * (see ptl_send_rpc), one for resend the req to server,
2426                          * Note: client will pack net_latency in replay req
2427                          * (see ptlrpc_replay_req)
2428                          */
2429                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2430                 }
2431                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2432                                       true);
2433         }
2434         EXIT;
2435 }
2436
2437 /** Checking routines for recovery */
2438 static int check_for_recovery_ready(struct lu_target *lut)
2439 {
2440         struct obd_device *obd = lut->lut_obd;
2441         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2442
2443         CDEBUG(D_HA,
2444                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2445                clnts, obd->obd_stale_clients,
2446                atomic_read(&obd->obd_max_recoverable_clients),
2447                obd->obd_abort_recovery, obd->obd_recovery_expired);
2448
2449         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2450                 LASSERT(clnts <=
2451                         atomic_read(&obd->obd_max_recoverable_clients));
2452                 if (clnts + obd->obd_stale_clients <
2453                     atomic_read(&obd->obd_max_recoverable_clients))
2454                         return 0;
2455         }
2456
2457         if (!obd->obd_abort_recov_mdt && lut->lut_tdtd != NULL) {
2458                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2459                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2460                         /*
2461                          * Let's extend recovery timer, in case the recovery
2462                          * timer expired, and some clients got evicted
2463                          */
2464                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2465                                               true);
2466                         CDEBUG(D_HA,
2467                                "%s update recovery is not ready, extend recovery %d\n",
2468                                obd->obd_name, obd->obd_recovery_timeout);
2469                         return 0;
2470                 }
2471         }
2472
2473         return 1;
2474 }
2475
2476 enum {
2477         REQUEST_RECOVERY = 1,
2478         UPDATE_RECOVERY = 2,
2479 };
2480
2481 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2482 {
2483         __u64 transno = 0;
2484
2485         if (!list_empty(&obd->obd_req_replay_queue)) {
2486                 struct ptlrpc_request *req;
2487
2488                 req = list_entry(obd->obd_req_replay_queue.next,
2489                                  struct ptlrpc_request, rq_list);
2490                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2491         }
2492
2493         return transno;
2494 }
2495
2496 static __u64 get_next_transno(struct lu_target *lut, int *type)
2497 {
2498         struct obd_device *obd = lut->lut_obd;
2499         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2500         __u64 transno = 0;
2501         __u64 update_transno;
2502
2503         ENTRY;
2504
2505         transno = get_next_replay_req_transno(obd);
2506         if (type != NULL)
2507                 *type = REQUEST_RECOVERY;
2508
2509         if (!tdtd || obd->obd_abort_recov_mdt)
2510                 RETURN(transno);
2511
2512         update_transno = distribute_txn_get_next_transno(tdtd);
2513         if (transno == 0 || (transno >= update_transno &&
2514                              update_transno != 0)) {
2515                 transno = update_transno;
2516                 if (type != NULL)
2517                         *type = UPDATE_RECOVERY;
2518         }
2519
2520         RETURN(transno);
2521 }
2522
2523 /**
2524  * drop duplicate replay request
2525  *
2526  * Because the operation has been replayed by update recovery, the request
2527  * with the same transno will be dropped and also notify the client to send
2528  * next replay request.
2529  *
2530  * \param[in] env       execution environment
2531  * \param[in] obd       failover obd device
2532  * \param[in] req       request to be dropped
2533  */
2534 static void drop_duplicate_replay_req(struct lu_env *env,
2535                                       struct obd_device *obd,
2536                                       struct ptlrpc_request *req)
2537 {
2538         DEBUG_REQ(D_HA, req,
2539                   "remove t%lld from %s because duplicate update records found",
2540                   lustre_msg_get_transno(req->rq_reqmsg),
2541                   libcfs_nid2str(req->rq_peer.nid));
2542
2543         /*
2544          * Right now, only for MDS reint operation update replay and
2545          * normal request replay can have the same transno
2546          */
2547         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2548                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2549                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2550                 if (likely(req->rq_export))
2551                         target_committed_to_req(req);
2552                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2553                 target_send_reply(req, req->rq_status, 0);
2554         } else {
2555                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2556                 libcfs_nid2str(req->rq_peer.nid));
2557         }
2558         target_exp_dequeue_req_replay(req);
2559         target_request_copy_put(req);
2560         obd->obd_replayed_requests++;
2561 }
2562
2563 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2564
2565 static void replay_request_or_update(struct lu_env *env,
2566                                      struct lu_target *lut,
2567                                      struct target_recovery_data *trd,
2568                                      struct ptlrpc_thread *thread)
2569 {
2570         struct obd_device *obd = lut->lut_obd;
2571         struct ptlrpc_request *req = NULL;
2572         int type;
2573         __u64 transno;
2574
2575         ENTRY;
2576
2577         CDEBUG(D_HA, "Waiting for transno %lld\n",
2578                obd->obd_next_recovery_transno);
2579
2580         /* Replay all of request and update by transno */
2581         do {
2582                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2583
2584                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2585
2586                 /**
2587                  * It is needed to extend recovery window above
2588                  *  recovery_time_soft. Extending is possible only in the
2589                  *  end of recovery window (see more details in
2590                  *  handle_recovery_req()).
2591                  */
2592                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2593
2594                 if (target_recovery_overseer(lut, check_for_next_transno,
2595                                         exp_req_replay_healthy_or_from_mdt)) {
2596                         abort_req_replay_queue(obd);
2597                         abort_lock_replay_queue(obd);
2598                         goto abort;
2599                 }
2600
2601                 spin_lock(&obd->obd_recovery_task_lock);
2602                 transno = get_next_transno(lut, &type);
2603                 if (type == REQUEST_RECOVERY && transno != 0) {
2604                         /*
2605                          * Drop replay request from client side, if the
2606                          * replay has been executed by update with the
2607                          * same transno
2608                          */
2609                         req = list_entry(obd->obd_req_replay_queue.next,
2610                                         struct ptlrpc_request, rq_list);
2611
2612                         list_del_init(&req->rq_list);
2613                         obd->obd_requests_queued_for_recovery--;
2614                         spin_unlock(&obd->obd_recovery_task_lock);
2615
2616                         /*
2617                          * Let's check if the request has been redone by
2618                          * update replay
2619                          */
2620                         if (is_req_replayed_by_update(req)) {
2621                                 struct distribute_txn_replay_req *dtrq;
2622
2623                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2624                                                                       transno);
2625                                 LASSERT(dtrq != NULL);
2626                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2627                                 list_del_init(&dtrq->dtrq_list);
2628                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2629                                 dtrq_destroy(dtrq);
2630
2631                                 drop_duplicate_replay_req(env, obd, req);
2632
2633                                 continue;
2634                         }
2635
2636                         LASSERT(trd->trd_processing_task == current->pid);
2637                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2638                                   req->rq_xid,
2639                                   lustre_msg_get_transno(req->rq_reqmsg),
2640                                   libcfs_nid2str(req->rq_peer.nid));
2641
2642                         ptlrpc_watchdog_init(&thread->t_watchdog,
2643                                              WATCHDOG_TIMEOUT);
2644                         handle_recovery_req(thread, req,
2645                                             trd->trd_recovery_handler);
2646                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2647
2648                         /**
2649                          * bz18031: increase next_recovery_transno before
2650                          * target_request_copy_put() will drop exp_rpc reference
2651                          */
2652                         spin_lock(&obd->obd_recovery_task_lock);
2653                         obd->obd_next_recovery_transno++;
2654                         spin_unlock(&obd->obd_recovery_task_lock);
2655                         target_exp_dequeue_req_replay(req);
2656                         target_request_copy_put(req);
2657                         obd->obd_replayed_requests++;
2658                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2659                         struct distribute_txn_replay_req *dtrq;
2660                         int rc;
2661
2662                         spin_unlock(&obd->obd_recovery_task_lock);
2663
2664                         LASSERT(tdtd != NULL);
2665                         dtrq = distribute_txn_get_next_req(tdtd);
2666                         lu_context_enter(&thread->t_env->le_ctx);
2667                         ptlrpc_watchdog_init(&thread->t_watchdog,
2668                                              WATCHDOG_TIMEOUT);
2669                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2670                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2671                         lu_context_exit(&thread->t_env->le_ctx);
2672                         extend_recovery_timer(obd, obd_timeout, true);
2673
2674                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2675                                 CDEBUG(D_HA,
2676                                        "Move x%llu t%llu to finish list\n",
2677                                        dtrq->dtrq_xid,
2678                                        dtrq->dtrq_master_transno);
2679
2680                                 /* Add it to the replay finish list */
2681                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2682                                 list_add(&dtrq->dtrq_list,
2683                                          &tdtd->tdtd_replay_finish_list);
2684                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2685
2686                                 spin_lock(&obd->obd_recovery_task_lock);
2687                                 if (transno == obd->obd_next_recovery_transno)
2688                                         obd->obd_next_recovery_transno++;
2689                                 else if (transno >
2690                                          obd->obd_next_recovery_transno)
2691                                         obd->obd_next_recovery_transno =
2692                                                                 transno + 1;
2693                                 spin_unlock(&obd->obd_recovery_task_lock);
2694                         } else {
2695                                 dtrq_destroy(dtrq);
2696                         }
2697                 } else {
2698                         spin_unlock(&obd->obd_recovery_task_lock);
2699 abort:
2700                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2701                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2702                         /** evict exports failed VBR */
2703                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2704                         break;
2705                 }
2706         } while (1);
2707 }
2708
2709 static int target_recovery_thread(void *arg)
2710 {
2711         struct lu_target *lut = arg;
2712         struct obd_device *obd = lut->lut_obd;
2713         struct ptlrpc_request *req;
2714         struct target_recovery_data *trd = &obd->obd_recovery_data;
2715         unsigned long delta;
2716         struct lu_env *env;
2717         struct ptlrpc_thread *thread = NULL;
2718         int rc = 0;
2719
2720         ENTRY;
2721         unshare_fs_struct();
2722         OBD_ALLOC_PTR(thread);
2723         if (thread == NULL)
2724                 RETURN(-ENOMEM);
2725
2726         OBD_ALLOC_PTR(env);
2727         if (env == NULL)
2728                 GOTO(out_thread, rc = -ENOMEM);
2729         rc = lu_env_add(env);
2730         if (rc)
2731                 GOTO(out_env, rc);
2732
2733         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2734         if (rc)
2735                 GOTO(out_env_remove, rc);
2736
2737         thread->t_env = env;
2738         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2739         thread->t_task = current;
2740         env->le_ctx.lc_thread = thread;
2741         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2742
2743         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2744                current->pid);
2745         trd->trd_processing_task = current->pid;
2746
2747         spin_lock(&obd->obd_dev_lock);
2748         obd->obd_recovering = 1;
2749         spin_unlock(&obd->obd_dev_lock);
2750         complete(&trd->trd_starting);
2751
2752         /* first of all, we have to know the first transno to replay */
2753         if (target_recovery_overseer(lut, check_for_recovery_ready,
2754                                      exp_connect_healthy)) {
2755                 abort_req_replay_queue(obd);
2756                 abort_lock_replay_queue(obd);
2757                 if (lut->lut_tdtd != NULL)
2758                         dtrq_list_destroy(lut->lut_tdtd);
2759         }
2760
2761         /* next stage: replay requests or update */
2762         delta = jiffies;
2763         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2764                atomic_read(&obd->obd_req_replay_clients),
2765                obd->obd_next_recovery_transno);
2766         replay_request_or_update(env, lut, trd, thread);
2767
2768         /**
2769          * The second stage: replay locks
2770          */
2771         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2772                atomic_read(&obd->obd_lock_replay_clients));
2773         while ((req = target_next_replay_lock(lut))) {
2774                 LASSERT(trd->trd_processing_task == current->pid);
2775                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2776                           libcfs_nid2str(req->rq_peer.nid));
2777                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2778                         req->rq_status = -ENODEV;
2779                         target_request_copy_put(req);
2780                         continue;
2781                 }
2782                 handle_recovery_req(thread, req,
2783                                     trd->trd_recovery_handler);
2784                 target_request_copy_put(req);
2785                 obd->obd_replayed_locks++;
2786         }
2787
2788         /**
2789          * The third stage: reply on final pings, at this moment all clients
2790          * must have request in final queue
2791          */
2792         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2793         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2794         /** Update server last boot epoch */
2795         tgt_boot_epoch_update(lut);
2796         /*
2797          * We drop recoverying flag to forward all new requests
2798          * to regular mds_handle() since now
2799          */
2800         spin_lock(&obd->obd_dev_lock);
2801         obd->obd_recovering = obd->obd_abort_recovery = 0;
2802         obd->obd_abort_recov_mdt = 0;
2803         spin_unlock(&obd->obd_dev_lock);
2804         spin_lock(&obd->obd_recovery_task_lock);
2805         target_cancel_recovery_timer(obd);
2806         spin_unlock(&obd->obd_recovery_task_lock);
2807         while ((req = target_next_final_ping(obd))) {
2808                 LASSERT(trd->trd_processing_task == current->pid);
2809                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2810                           libcfs_nid2str(req->rq_peer.nid));
2811                 handle_recovery_req(thread, req,
2812                                     trd->trd_recovery_handler);
2813                 /*
2814                  * Because the waiting client can not send ping to server,
2815                  * so we need refresh the last_request_time, to avoid the
2816                  * export is being evicted
2817                  */
2818                 ptlrpc_update_export_timer(req->rq_export, 0);
2819                 target_request_copy_put(req);
2820         }
2821
2822         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2823         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2824                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2825         if (delta > OBD_RECOVERY_TIME_SOFT) {
2826                 CWARN("too long recovery - read logs\n");
2827                 libcfs_debug_dumplog();
2828         }
2829
2830         target_finish_recovery(lut);
2831         lu_context_fini(&env->le_ctx);
2832         trd->trd_processing_task = 0;
2833         complete_all(&trd->trd_finishing);
2834         tgt_io_thread_done(thread);
2835 out_env_remove:
2836         lu_env_remove(env);
2837 out_env:
2838         OBD_FREE_PTR(env);
2839 out_thread:
2840         OBD_FREE_PTR(thread);
2841         RETURN(rc);
2842 }
2843
2844 static int target_start_recovery_thread(struct lu_target *lut,
2845                                         svc_handler_t handler)
2846 {
2847         struct obd_device *obd = lut->lut_obd;
2848         int rc = 0;
2849         struct target_recovery_data *trd = &obd->obd_recovery_data;
2850         int index;
2851
2852         memset(trd, 0, sizeof(*trd));
2853         init_completion(&trd->trd_starting);
2854         init_completion(&trd->trd_finishing);
2855         trd->trd_recovery_handler = handler;
2856
2857         rc = server_name2index(obd->obd_name, &index, NULL);
2858         if (rc < 0)
2859                 return rc;
2860
2861         if (!IS_ERR(kthread_run(target_recovery_thread,
2862                                 lut, "tgt_recover_%d", index))) {
2863                 wait_for_completion(&trd->trd_starting);
2864                 LASSERT(obd->obd_recovering != 0);
2865         } else {
2866                 rc = -ECHILD;
2867         }
2868
2869         return rc;
2870 }
2871
2872 void target_stop_recovery_thread(struct obd_device *obd)
2873 {
2874         if (obd->obd_recovery_data.trd_processing_task > 0) {
2875                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2876                 /** recovery can be done but postrecovery is not yet */
2877                 spin_lock(&obd->obd_dev_lock);
2878                 if (obd->obd_recovering) {
2879                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2880                         obd->obd_abort_recovery = 1;
2881                         wake_up(&obd->obd_next_transno_waitq);
2882                 }
2883                 spin_unlock(&obd->obd_dev_lock);
2884                 wait_for_completion(&trd->trd_finishing);
2885         }
2886 }
2887 EXPORT_SYMBOL(target_stop_recovery_thread);
2888
2889 void target_recovery_fini(struct obd_device *obd)
2890 {
2891         class_disconnect_exports(obd);
2892         target_stop_recovery_thread(obd);
2893         target_cleanup_recovery(obd);
2894 }
2895 EXPORT_SYMBOL(target_recovery_fini);
2896
2897 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2898 {
2899         struct obd_device *obd = container_of(timer, struct obd_device,
2900                                               obd_recovery_timer);
2901
2902         CDEBUG(D_HA,
2903                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2904                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2905                ktime_get_seconds() - obd->obd_recovery_start,
2906                atomic_read(&obd->obd_connected_clients));
2907
2908         obd->obd_recovery_expired = 1;
2909         wake_up(&obd->obd_next_transno_waitq);
2910         return HRTIMER_NORESTART;
2911 }
2912
2913 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2914 {
2915         struct obd_device *obd = lut->lut_obd;
2916
2917         if (lut->lut_bottom->dd_rdonly)
2918                 return;
2919
2920         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2921                 /** Update server last boot epoch */
2922                 tgt_boot_epoch_update(lut);
2923                 return;
2924         }
2925
2926         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2927                "last_transno %llu\n", obd->obd_name,
2928                atomic_read(&obd->obd_max_recoverable_clients),
2929                obd->obd_last_committed);
2930         LASSERT(obd->obd_stopping == 0);
2931         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2932         obd->obd_recovery_start = 0;
2933         obd->obd_recovery_end = 0;
2934
2935         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2936                      HRTIMER_MODE_ABS);
2937         obd->obd_recovery_timer.function = &target_recovery_expired;
2938         target_start_recovery_thread(lut, handler);
2939 }
2940 EXPORT_SYMBOL(target_recovery_init);
2941
2942 static int target_process_req_flags(struct obd_device *obd,
2943                                     struct ptlrpc_request *req)
2944 {
2945         struct obd_export *exp = req->rq_export;
2946
2947         LASSERT(exp != NULL);
2948         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2949                 /* client declares he's ready to replay locks */
2950                 spin_lock(&exp->exp_lock);
2951                 if (exp->exp_req_replay_needed) {
2952                         exp->exp_req_replay_needed = 0;
2953                         spin_unlock(&exp->exp_lock);
2954
2955                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2956                         atomic_dec(&obd->obd_req_replay_clients);
2957                 } else {
2958                         spin_unlock(&exp->exp_lock);
2959                 }
2960         }
2961         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2962                 /*
2963                  * client declares he's ready to complete recovery
2964                  * so, we put the request on th final queue
2965                  */
2966                 spin_lock(&exp->exp_lock);
2967                 if (exp->exp_lock_replay_needed) {
2968                         exp->exp_lock_replay_needed = 0;
2969                         spin_unlock(&exp->exp_lock);
2970
2971                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2972                         atomic_dec(&obd->obd_lock_replay_clients);
2973                 } else {
2974                         spin_unlock(&exp->exp_lock);
2975                 }
2976         }
2977         return 0;
2978 }
2979
2980 int target_queue_recovery_request(struct ptlrpc_request *req,
2981                                   struct obd_device *obd)
2982 {
2983         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2984         struct ptlrpc_request *reqiter;
2985         int inserted = 0;
2986
2987         ENTRY;
2988
2989         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2990                 /* Processing the queue right now, don't re-add. */
2991                 RETURN(1);
2992         }
2993
2994         target_process_req_flags(obd, req);
2995
2996         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2997                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2998                         if (cfs_fail_val == 1) {
2999                                 cfs_race_state = 1;
3000                                 cfs_fail_val = 0;
3001                                 wake_up(&cfs_race_waitq);
3002
3003                                 schedule_timeout_interruptible(
3004                                         cfs_time_seconds(1));
3005                         }
3006                 }
3007
3008                 /*
3009                  * client declares he's ready to complete recovery
3010                  * so, we put the request on th final queue
3011                  */
3012                 target_request_copy_get(req);
3013                 DEBUG_REQ(D_HA, req, "queue final req");
3014                 wake_up(&obd->obd_next_transno_waitq);
3015                 spin_lock(&obd->obd_recovery_task_lock);
3016                 if (obd->obd_recovering) {
3017                         struct ptlrpc_request *tmp;
3018                         struct ptlrpc_request *duplicate = NULL;
3019
3020                         if (likely(!req->rq_export->exp_replay_done)) {
3021                                 req->rq_export->exp_replay_done = 1;
3022                                 list_add_tail(&req->rq_list,
3023                                               &obd->obd_final_req_queue);
3024                                 spin_unlock(&obd->obd_recovery_task_lock);
3025                                 RETURN(0);
3026                         }
3027
3028                         /*
3029                          * XXX O(n), but only happens if final ping is
3030                          * timed out, probably reorganize the list as
3031                          * a hash list later
3032                          */
3033                         list_for_each_entry_safe(reqiter, tmp,
3034                                                  &obd->obd_final_req_queue,
3035                                                  rq_list) {
3036                                 if (reqiter->rq_export == req->rq_export) {
3037                                         list_del_init(&reqiter->rq_list);
3038                                         duplicate = reqiter;
3039                                         break;
3040                                 }
3041                         }
3042
3043                         list_add_tail(&req->rq_list,
3044                                       &obd->obd_final_req_queue);
3045                         req->rq_export->exp_replay_done = 1;
3046                         spin_unlock(&obd->obd_recovery_task_lock);
3047
3048                         if (duplicate != NULL) {
3049                                 DEBUG_REQ(D_HA, duplicate,
3050                                           "put prev final req");
3051                                 target_request_copy_put(duplicate);
3052                         }
3053                         RETURN(0);
3054                 } else {
3055                         spin_unlock(&obd->obd_recovery_task_lock);
3056                         target_request_copy_put(req);
3057                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3058                 }
3059         }
3060         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3061                 /* client declares he's ready to replay locks */
3062                 target_request_copy_get(req);
3063                 DEBUG_REQ(D_HA, req, "queue lock replay req");
3064                 wake_up(&obd->obd_next_transno_waitq);
3065                 spin_lock(&obd->obd_recovery_task_lock);
3066                 LASSERT(obd->obd_recovering);
3067                 /* usually due to recovery abort */
3068                 if (!req->rq_export->exp_in_recovery) {
3069                         spin_unlock(&obd->obd_recovery_task_lock);
3070                         target_request_copy_put(req);
3071                         RETURN(-ENOTCONN);
3072                 }
3073                 LASSERT(req->rq_export->exp_lock_replay_needed);
3074                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3075                 spin_unlock(&obd->obd_recovery_task_lock);
3076                 RETURN(0);
3077         }
3078
3079         /*
3080          * CAVEAT EMPTOR: The incoming request message has been swabbed
3081          * (i.e. buflens etc are in my own byte order), but type-dependent
3082          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3083          */
3084
3085         if (!transno) {
3086                 INIT_LIST_HEAD(&req->rq_list);
3087                 DEBUG_REQ(D_HA, req, "not queueing");
3088                 RETURN(1);
3089         }
3090
3091         /*
3092          * If we're processing the queue, we want don't want to queue this
3093          * message.
3094          *
3095          * Also, if this request has a transno less than the one we're waiting
3096          * for, we should process it now.  It could (and currently always will)
3097          * be an open request for a descriptor that was opened some time ago.
3098          *
3099          * Also, a resent, replayed request that has already been
3100          * handled will pass through here and be processed immediately.
3101          */
3102         CDEBUG(D_HA,
3103                "Next recovery transno: %llu, current: %llu, replaying\n",
3104                obd->obd_next_recovery_transno, transno);
3105
3106         /*
3107          * If the request has been replayed by update replay, then sends this
3108          * request to the recovery thread (replay_request_or_update()), where
3109          * it will be handled
3110          */
3111         spin_lock(&obd->obd_recovery_task_lock);
3112         if (transno < obd->obd_next_recovery_transno &&
3113             !is_req_replayed_by_update(req)) {
3114                 /* Processing the queue right now, don't re-add. */
3115                 LASSERT(list_empty(&req->rq_list));
3116                 spin_unlock(&obd->obd_recovery_task_lock);
3117                 RETURN(1);
3118         }
3119         spin_unlock(&obd->obd_recovery_task_lock);
3120
3121         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3122                 RETURN(0);
3123
3124         target_request_copy_get(req);
3125         if (!req->rq_export->exp_in_recovery) {
3126                 target_request_copy_put(req);
3127                 RETURN(-ENOTCONN);
3128         }
3129         LASSERT(req->rq_export->exp_req_replay_needed);
3130
3131         if (target_exp_enqueue_req_replay(req)) {
3132                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3133                 target_request_copy_put(req);
3134                 RETURN(0);
3135         }
3136
3137         /* XXX O(n^2) */
3138         spin_lock(&obd->obd_recovery_task_lock);
3139         LASSERT(obd->obd_recovering);
3140         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3141                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3142                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3143                         inserted = 1;
3144                         goto added;
3145                 }
3146
3147                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3148                              transno)) {
3149                         DEBUG_REQ(D_ERROR, req,
3150                                   "dropping replay: transno has been claimed by another client");
3151                         spin_unlock(&obd->obd_recovery_task_lock);
3152                         target_exp_dequeue_req_replay(req);
3153                         target_request_copy_put(req);
3154                         RETURN(0);
3155                 }
3156         }
3157 added:
3158         if (!inserted)
3159                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3160
3161         obd->obd_requests_queued_for_recovery++;
3162         spin_unlock(&obd->obd_recovery_task_lock);
3163         wake_up(&obd->obd_next_transno_waitq);
3164         RETURN(0);
3165 }
3166
3167 void target_committed_to_req(struct ptlrpc_request *req)
3168 {
3169         struct obd_export *exp = req->rq_export;
3170
3171         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3172                 lustre_msg_set_last_committed(req->rq_repmsg,
3173                                               exp->exp_last_committed);
3174         else
3175                 DEBUG_REQ(D_IOCTL, req,
3176                           "not sending last_committed update (%d/%d)",
3177                           exp->exp_obd->obd_no_transno,
3178                           req->rq_repmsg == NULL);
3179
3180         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3181                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3182 }
3183
3184 #endif /* HAVE_SERVER_SUPPORT */
3185
3186 /**
3187  * Packs current SLV and Limit into \a req.
3188  */
3189 int target_pack_pool_reply(struct ptlrpc_request *req)
3190 {
3191         struct obd_device *obd;
3192
3193         ENTRY;
3194
3195         /*
3196          * Check that we still have all structures alive as this may
3197          * be some late RPC at shutdown time.
3198          */
3199         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3200                      !exp_connect_lru_resize(req->rq_export))) {
3201                 lustre_msg_set_slv(req->rq_repmsg, 0);
3202                 lustre_msg_set_limit(req->rq_repmsg, 0);
3203                 RETURN(0);
3204         }
3205
3206         /* OBD is alive here as export is alive, which we checked above. */
3207         obd = req->rq_export->exp_obd;
3208
3209         read_lock(&obd->obd_pool_lock);
3210         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3211         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3212         read_unlock(&obd->obd_pool_lock);
3213
3214         RETURN(0);
3215 }
3216
3217 static int target_send_reply_msg(struct ptlrpc_request *req,
3218                                  int rc, int fail_id)
3219 {
3220         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3221                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3222                 return -ECOMM;
3223         }
3224         /*
3225          * We can have a null rq_reqmsg in the event of bad signature or
3226          * no context when unwrapping
3227          */
3228         if (req->rq_reqmsg &&
3229             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3230             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3231                 return -ECOMM;
3232
3233         if (unlikely(rc)) {
3234                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3235                 req->rq_status = rc;
3236                 return ptlrpc_send_error(req, 1);
3237         }
3238         DEBUG_REQ(D_NET, req, "sending reply");
3239
3240         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3241 }
3242
3243 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3244 {
3245         struct ptlrpc_service_part *svcpt;
3246         int netrc;
3247         struct ptlrpc_reply_state *rs;
3248         struct obd_export *exp;
3249
3250         ENTRY;
3251
3252         if (req->rq_no_reply) {
3253                 EXIT;
3254                 return;
3255         }
3256
3257         svcpt = req->rq_rqbd->rqbd_svcpt;
3258         rs = req->rq_reply_state;
3259         if (rs == NULL || !rs->rs_difficult) {
3260                 /* no notifiers */
3261                 target_send_reply_msg(req, rc, fail_id);
3262                 EXIT;
3263                 return;
3264         }
3265
3266         /* must be an export if locks saved */
3267         LASSERT(req->rq_export != NULL);
3268         /* req/reply consistent */
3269         LASSERT(rs->rs_svcpt == svcpt);
3270
3271         /* "fresh" reply */
3272         LASSERT(!rs->rs_scheduled);
3273         LASSERT(!rs->rs_scheduled_ever);
3274         LASSERT(!rs->rs_handled);
3275         LASSERT(!rs->rs_sent);
3276         LASSERT(!rs->rs_unlinked);
3277         LASSERT(rs->rs_export == NULL);
3278         LASSERT(list_empty(&rs->rs_obd_list));
3279         LASSERT(list_empty(&rs->rs_exp_list));
3280
3281         exp = class_export_get(req->rq_export);
3282
3283         /* disable reply scheduling while I'm setting up */
3284         rs->rs_scheduled = 1;
3285         rs->rs_sent      = 0;
3286         rs->rs_unlinked  = 0;
3287         rs->rs_xid       = req->rq_xid;
3288         rs->rs_transno   = req->rq_transno;
3289         rs->rs_export    = exp;
3290         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3291
3292         spin_lock(&exp->exp_uncommitted_replies_lock);
3293         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3294                rs->rs_transno, exp->exp_last_committed);
3295         if (rs->rs_transno > exp->exp_last_committed) {
3296                 /* not committed already */
3297                 list_add_tail(&rs->rs_obd_list,
3298                                   &exp->exp_uncommitted_replies);
3299         }
3300         spin_unlock(&exp->exp_uncommitted_replies_lock);
3301
3302         spin_lock(&exp->exp_lock);
3303         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3304         spin_unlock(&exp->exp_lock);
3305
3306         netrc = target_send_reply_msg(req, rc, fail_id);
3307
3308         spin_lock(&svcpt->scp_rep_lock);
3309
3310         atomic_inc(&svcpt->scp_nreps_difficult);
3311
3312         if (netrc != 0) {
3313                 /*
3314                  * error sending: reply is off the net.  Also we need +1
3315                  * reply ref until ptlrpc_handle_rs() is done
3316                  * with the reply state (if the send was successful, there
3317                  * would have been +1 ref for the net, which
3318                  * reply_out_callback leaves alone)
3319                  */
3320                 rs->rs_sent = 1;
3321                 rs->rs_unlinked = 1;
3322                 ptlrpc_rs_addref(rs);
3323         }
3324
3325         spin_lock(&rs->rs_lock);
3326         if (rs->rs_transno <= exp->exp_last_committed ||
3327             (rs->rs_unlinked && !rs->rs_no_ack) ||
3328             list_empty(&rs->rs_exp_list) ||     /* completed already */
3329             list_empty(&rs->rs_obd_list)) {
3330                 CDEBUG(D_HA, "Schedule reply immediately\n");
3331                 ptlrpc_dispatch_difficult_reply(rs);
3332         } else {
3333                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3334                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3335         }
3336         spin_unlock(&rs->rs_lock);
3337         spin_unlock(&svcpt->scp_rep_lock);
3338         EXIT;
3339 }
3340
3341 enum ldlm_mode lck_compat_array[] = {
3342         [LCK_EX]    = LCK_COMPAT_EX,
3343         [LCK_PW]    = LCK_COMPAT_PW,
3344         [LCK_PR]    = LCK_COMPAT_PR,
3345         [LCK_CW]    = LCK_COMPAT_CW,
3346         [LCK_CR]    = LCK_COMPAT_CR,
3347         [LCK_NL]    = LCK_COMPAT_NL,
3348         [LCK_GROUP] = LCK_COMPAT_GROUP,
3349         [LCK_COS]   = LCK_COMPAT_COS,
3350 };
3351
3352 /**
3353  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3354  * not escape to the user level.
3355  */
3356 int ldlm_error2errno(enum ldlm_error error)
3357 {
3358         int result;
3359
3360         switch (error) {
3361         case ELDLM_OK:
3362         case ELDLM_LOCK_MATCHED:
3363                 result = 0;
3364                 break;
3365         case ELDLM_LOCK_CHANGED:
3366                 result = -ESTALE;
3367                 break;
3368         case ELDLM_LOCK_ABORTED:
3369                 result = -ENAVAIL;
3370                 break;
3371         case ELDLM_LOCK_REPLACED:
3372                 result = -ESRCH;
3373                 break;
3374         case ELDLM_NO_LOCK_DATA:
3375                 result = -ENOENT;
3376                 break;
3377         case ELDLM_NAMESPACE_EXISTS:
3378                 result = -EEXIST;
3379                 break;
3380         case ELDLM_BAD_NAMESPACE:
3381                 result = -EBADF;
3382                 break;
3383         default:
3384                 if (((int)error) < 0) { /* cast to signed type */
3385                         result = error; /* as ldlm_error can be unsigned */
3386                 } else {
3387                         CERROR("Invalid DLM result code: %d\n", error);
3388                         result = -EPROTO;
3389                 }
3390         }
3391         return result;
3392 }
3393 EXPORT_SYMBOL(ldlm_error2errno);
3394
3395 /**
3396  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3397  */
3398 enum ldlm_error ldlm_errno2error(int err_no)
3399 {
3400         int error;
3401
3402         switch (err_no) {
3403         case 0:
3404                 error = ELDLM_OK;
3405                 break;
3406         case -ESTALE:
3407                 error = ELDLM_LOCK_CHANGED;
3408                 break;
3409         case -ENAVAIL:
3410                 error = ELDLM_LOCK_ABORTED;
3411                 break;
3412         case -ESRCH:
3413                 error = ELDLM_LOCK_REPLACED;
3414                 break;
3415         case -ENOENT:
3416                 error = ELDLM_NO_LOCK_DATA;
3417                 break;
3418         case -EEXIST:
3419                 error = ELDLM_NAMESPACE_EXISTS;
3420                 break;
3421         case -EBADF:
3422                 error = ELDLM_BAD_NAMESPACE;
3423                 break;
3424         default:
3425                 error = err_no;
3426         }
3427         return error;
3428 }
3429
3430 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3431 void ldlm_dump_export_locks(struct obd_export *exp)
3432 {
3433         spin_lock(&exp->exp_locks_list_guard);
3434         if (!list_empty(&exp->exp_locks_list)) {
3435                 struct ldlm_lock *lock;
3436
3437                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3438                        exp);
3439                 list_for_each_entry(lock, &exp->exp_locks_list,
3440                                         l_exp_refs_link)
3441                         LDLM_ERROR(lock, "lock:");
3442         }
3443         spin_unlock(&exp->exp_locks_list_guard);
3444 }
3445 #endif
3446
3447 #ifdef HAVE_SERVER_SUPPORT
3448 static inline const char *bulk2type(struct ptlrpc_request *req)
3449 {
3450         if (req->rq_bulk_read)
3451                 return "READ";
3452         if (req->rq_bulk_write)
3453                 return "WRITE";
3454         return "UNKNOWN";
3455 }
3456
3457 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3458 {
3459         struct ptlrpc_request *req = desc->bd_req;
3460         time64_t start = ktime_get_seconds();
3461         time64_t deadline;
3462         int rc = 0;
3463
3464         ENTRY;
3465
3466         /* If there is eviction in progress, wait for it to finish. */
3467         wait_event_idle(
3468                 exp->exp_obd->obd_evict_inprogress_waitq,
3469                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3470
3471         /* Check if client was evicted or reconnected already. */
3472         if (exp->exp_failed ||
3473             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3474                 rc = -ENOTCONN;
3475         } else {
3476                 if (req->rq_bulk_read)
3477                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3478
3479                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3480                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3481                 else /* old version, bulk matchbits is rq_xid */
3482                         req->rq_mbits = req->rq_xid;
3483
3484                 if (rc == 0)
3485                         rc = ptlrpc_start_bulk_transfer(desc);
3486         }
3487
3488         if (rc < 0) {
3489                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3490                           bulk2type(req), rc);
3491                 RETURN(rc);
3492         }
3493
3494         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3495                 ptlrpc_abort_bulk(desc);
3496                 RETURN(0);
3497         }
3498
3499         /* limit actual bulk transfer to bulk_timeout seconds */
3500         deadline = start + bulk_timeout;
3501         if (deadline > req->rq_deadline)
3502                 deadline = req->rq_deadline;
3503
3504         do {
3505                 time64_t timeoutl = deadline - ktime_get_seconds();
3506                 time64_t rq_deadline;
3507
3508                 while (timeoutl >= 0 &&
3509                        wait_event_idle_timeout(
3510                                desc->bd_waitq,
3511                                !ptlrpc_server_bulk_active(desc) ||
3512                                exp->exp_failed ||
3513                                exp->exp_conn_cnt >
3514                                lustre_msg_get_conn_cnt(req->rq_reqmsg),
3515                                timeoutl ? cfs_time_seconds(1) : 1) == 0)
3516                         timeoutl -= 1;
3517                 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3518
3519                 /* Wait again if we changed rq_deadline. */
3520                 rq_deadline = READ_ONCE(req->rq_deadline);
3521                 deadline = start + bulk_timeout;
3522                 if (deadline > rq_deadline)
3523                         deadline = rq_deadline;
3524         } while (rc == -ETIMEDOUT &&
3525                  deadline > ktime_get_seconds());
3526
3527         if (rc == -ETIMEDOUT) {
3528                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3529                           bulk2type(req), deadline - start,
3530                           ktime_get_real_seconds() - deadline);
3531                 ptlrpc_abort_bulk(desc);
3532         } else if (exp->exp_failed) {
3533                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3534                           bulk2type(req));
3535                 rc = -ENOTCONN;
3536                 ptlrpc_abort_bulk(desc);
3537         } else if (exp->exp_conn_cnt >
3538                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3539                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3540                           bulk2type(req));
3541                 /* We don't reply anyway. */
3542                 rc = -ETIMEDOUT;
3543                 ptlrpc_abort_bulk(desc);
3544         } else if (desc->bd_failure) {
3545                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3546                           bulk2type(req));
3547                 /* XXX should this be a different errno? */
3548                 rc = -ETIMEDOUT;
3549         } else {
3550                 if (req->rq_bulk_write)
3551                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3552                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3553                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3554                                   bulk2type(req), desc->bd_nob_transferred,
3555                                   desc->bd_nob);
3556                         /* XXX should this be a different errno? */
3557                         rc = -ETIMEDOUT;
3558                 }
3559         }
3560
3561         RETURN(rc);
3562 }
3563 EXPORT_SYMBOL(target_bulk_io);
3564
3565 #endif /* HAVE_SERVER_SUPPORT */