lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 /**
  34  * This file deals with various client/target related logic including recovery.
  35  *
  36  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  37  * should be moved.
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #include <cl_object.h>
  43 #include <linux/jiffies.h>
  44 #include <linux/kernel.h>
  45 #include <linux/kthread.h>
  46 #include <libcfs/libcfs.h>
  47 #include <obd.h>
  48 #include <obd_class.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /*
  55  * @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         lnet_nid_t nid4refnet = LNET_NID_ANY;
  64         int rc = 0;
  65
  66         ENTRY;
  67
  68         if (!create && !priority) {
  69                 CDEBUG(D_HA, "Nothing to do\n");
  70                 RETURN(-EINVAL);
  71         }
  72
  73         if (imp->imp_connection &&
  74             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  75                 /* nid4refnet is used to restrict network connections */
  76                 nid4refnet = imp->imp_connection->c_self;
  77         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
  78         if (!ptlrpc_conn) {
  79                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  80                 RETURN(-ENOENT);
  81         }
  82
  83         if (create) {
  84                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  85                 if (!imp_conn)
  86                         GOTO(out_put, rc = -ENOMEM);
  87         }
  88
  89         spin_lock(&imp->imp_lock);
  90         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  91                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  92                         if (priority) {
  93                                 list_move(&item->oic_item,
  94                                           &imp->imp_conn_list);
  95                                 item->oic_last_attempt = 0;
  96                         }
  97                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  98                                imp, imp->imp_obd->obd_name, uuid->uuid,
  99                                (priority ? ", moved to head" : ""));
 100                         spin_unlock(&imp->imp_lock);
 101                         GOTO(out_free, rc = 0);
 102                 }
 103         }
 104         /* No existing import connection found for \a uuid. */
 105         if (create) {
 106                 imp_conn->oic_conn = ptlrpc_conn;
 107                 imp_conn->oic_uuid = *uuid;
 108                 imp_conn->oic_last_attempt = 0;
 109                 if (priority)
 110                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 111                 else
 112                         list_add_tail(&imp_conn->oic_item,
 113                                       &imp->imp_conn_list);
 114                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 115                        imp, imp->imp_obd->obd_name, uuid->uuid,
 116                        (priority ? "head" : "tail"));
 117         } else {
 118                 spin_unlock(&imp->imp_lock);
 119                 GOTO(out_free, rc = -ENOENT);
 120         }
 121
 122         spin_unlock(&imp->imp_lock);
 123         RETURN(0);
 124 out_free:
 125         if (imp_conn)
 126                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 127 out_put:
 128         ptlrpc_connection_put(ptlrpc_conn);
 129         RETURN(rc);
 130 }
 131
 132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 133 {
 134         return import_set_conn(imp, uuid, 1, 0);
 135 }
 136
 137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 138                            int priority)
 139 {
 140         return import_set_conn(imp, uuid, priority, 1);
 141 }
 142 EXPORT_SYMBOL(client_import_add_conn);
 143
 144 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 145 {
 146         struct obd_import_conn *imp_conn;
 147         struct obd_export *dlmexp;
 148         int rc = -ENOENT;
 149
 150         ENTRY;
 151
 152         spin_lock(&imp->imp_lock);
 153         if (list_empty(&imp->imp_conn_list)) {
 154                 LASSERT(!imp->imp_connection);
 155                 GOTO(out, rc);
 156         }
 157
 158         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 159                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 160                         continue;
 161                 LASSERT(imp_conn->oic_conn);
 162
 163                 if (imp_conn == imp->imp_conn_current) {
 164                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 165
 166                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 167                             imp->imp_state != LUSTRE_IMP_DISCON) {
 168                                 CERROR("can't remove current connection\n");
 169                                 GOTO(out, rc = -EBUSY);
 170                         }
 171
 172                         ptlrpc_connection_put(imp->imp_connection);
 173                         imp->imp_connection = NULL;
 174
 175                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 176                         if (dlmexp && dlmexp->exp_connection) {
 177                                 LASSERT(dlmexp->exp_connection ==
 178                                         imp_conn->oic_conn);
 179                                 ptlrpc_connection_put(dlmexp->exp_connection);
 180                                 dlmexp->exp_connection = NULL;
 181                         }
 182
 183                         if (dlmexp != NULL)
 184                                 class_export_put(dlmexp);
 185                 }
 186
 187                 list_del(&imp_conn->oic_item);
 188                 ptlrpc_connection_put(imp_conn->oic_conn);
 189                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 190                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 191                        imp, imp->imp_obd->obd_name, uuid->uuid);
 192                 rc = 0;
 193                 break;
 194         }
 195 out:
 196         spin_unlock(&imp->imp_lock);
 197         if (rc == -ENOENT)
 198                 CERROR("connection %s not found\n", uuid->uuid);
 199         RETURN(rc);
 200 }
 201 EXPORT_SYMBOL(client_import_del_conn);
 202
 203 /**
 204  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 205  * to find a conn uuid of \a imp which can reach \a peer.
 206  */
 207 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 208                             struct obd_uuid *uuid)
 209 {
 210         struct obd_import_conn *conn;
 211         int rc = -ENOENT;
 212
 213         ENTRY;
 214
 215         spin_lock(&imp->imp_lock);
 216         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 217                 /* Check if conn UUID does have this peer NID. */
 218                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 219                         *uuid = conn->oic_uuid;
 220                         rc = 0;
 221                         break;
 222                 }
 223         }
 224         spin_unlock(&imp->imp_lock);
 225         RETURN(rc);
 226 }
 227 EXPORT_SYMBOL(client_import_find_conn);
 228
 229 void client_destroy_import(struct obd_import *imp)
 230 {
 231         /*
 232          * Drop security policy instance after all RPCs have finished/aborted
 233          * to let all busy contexts be released.
 234          */
 235         class_import_get(imp);
 236         class_destroy_import(imp);
 237         sptlrpc_import_sec_put(imp);
 238         class_import_put(imp);
 239 }
 240 EXPORT_SYMBOL(client_destroy_import);
 241
 242 /**
 243  * Check whether or not the OSC is on MDT.
 244  * In the config log,
 245  * osc on MDT
 246  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 247  * osc on client
 248  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 249  *
 250  **/
 251 static int osc_on_mdt(char *obdname)
 252 {
 253         char *ptr;
 254
 255         ptr = strrchr(obdname, '-');
 256         if (ptr == NULL)
 257                 return 0;
 258
 259         if (strncmp(ptr + 1, "MDT", 3) == 0)
 260                 return 1;
 261
 262         return 0;
 263 }
 264
 265 /*
 266  * Configure an RPC client OBD device.
 267  *
 268  * lcfg parameters:
 269  * 1 - client UUID
 270  * 2 - server UUID
 271  * 3 - inactive-on-startup
 272  * 4 - restrictive net
 273  */
 274 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 275 {
 276         struct client_obd *cli = &obddev->u.cli;
 277         struct obd_import *imp;
 278         struct obd_uuid server_uuid;
 279         int rq_portal, rp_portal, connect_op;
 280         const char *name = obddev->obd_type->typ_name;
 281         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 282         char *cli_name = lustre_cfg_buf(lcfg, 0);
 283         struct ptlrpc_connection fake_conn = { .c_self = 0,
 284                                                .c_remote_uuid.uuid[0] = 0 };
 285         int rc;
 286
 287         ENTRY;
 288
 289         /*
 290          * In a more perfect world, we would hang a ptlrpc_client off of
 291          * obd_type and just use the values from there.
 292          */
 293         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 294                 rq_portal = OST_REQUEST_PORTAL;
 295                 rp_portal = OSC_REPLY_PORTAL;
 296                 connect_op = OST_CONNECT;
 297                 cli->cl_sp_me = LUSTRE_SP_CLI;
 298                 cli->cl_sp_to = LUSTRE_SP_OST;
 299                 ns_type = LDLM_NS_TYPE_OSC;
 300         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 301                    !strcmp(name, LUSTRE_LWP_NAME)) {
 302                 rq_portal = MDS_REQUEST_PORTAL;
 303                 rp_portal = MDC_REPLY_PORTAL;
 304                 connect_op = MDS_CONNECT;
 305                 if (is_lwp_on_ost(cli_name))
 306                         cli->cl_sp_me = LUSTRE_SP_OST;
 307                 else if (is_lwp_on_mdt(cli_name))
 308                         cli->cl_sp_me = LUSTRE_SP_MDT;
 309                 else
 310                         cli->cl_sp_me = LUSTRE_SP_CLI;
 311                 cli->cl_sp_to = LUSTRE_SP_MDT;
 312                 ns_type = LDLM_NS_TYPE_MDC;
 313         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 314                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 315                         /* OSP_on_MDT for other MDTs */
 316                         connect_op = MDS_CONNECT;
 317                         cli->cl_sp_to = LUSTRE_SP_MDT;
 318                         ns_type = LDLM_NS_TYPE_MDC;
 319                         rq_portal = OUT_PORTAL;
 320                 } else {
 321                         /* OSP on MDT for OST */
 322                         connect_op = OST_CONNECT;
 323                         cli->cl_sp_to = LUSTRE_SP_OST;
 324                         ns_type = LDLM_NS_TYPE_OSC;
 325                         rq_portal = OST_REQUEST_PORTAL;
 326                 }
 327                 rp_portal = OSC_REPLY_PORTAL;
 328                 cli->cl_sp_me = LUSTRE_SP_MDT;
 329         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 330                 rq_portal = MGS_REQUEST_PORTAL;
 331                 rp_portal = MGC_REPLY_PORTAL;
 332                 connect_op = MGS_CONNECT;
 333                 cli->cl_sp_me = LUSTRE_SP_MGC;
 334                 cli->cl_sp_to = LUSTRE_SP_MGS;
 335                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 336                 ns_type = LDLM_NS_TYPE_MGC;
 337         } else {
 338                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 339                        name);
 340                 RETURN(-EINVAL);
 341         }
 342
 343         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 344                 CERROR("requires a TARGET UUID\n");
 345                 RETURN(-EINVAL);
 346         }
 347
 348         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 349                 CERROR("client UUID must be less than 38 characters\n");
 350                 RETURN(-EINVAL);
 351         }
 352
 353         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 354                 CERROR("setup requires a SERVER UUID\n");
 355                 RETURN(-EINVAL);
 356         }
 357
 358         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 359                 CERROR("target UUID must be less than 38 characters\n");
 360                 RETURN(-EINVAL);
 361         }
 362
 363         init_rwsem(&cli->cl_sem);
 364         mutex_init(&cli->cl_mgc_mutex);
 365         cli->cl_seq = NULL;
 366         init_rwsem(&cli->cl_seq_rwsem);
 367         cli->cl_conn_count = 0;
 368         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 369                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 370                      sizeof(server_uuid)));
 371
 372         cli->cl_dirty_pages = 0;
 373         cli->cl_dirty_max_pages = 0;
 374         cli->cl_avail_grant = 0;
 375         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 376         /*
 377          * cl_dirty_max_pages may be changed at connect time in
 378          * ptlrpc_connect_interpret().
 379          */
 380         client_adjust_max_dirty(cli);
 381         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 382         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 383         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 384         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 385         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 386         spin_lock_init(&cli->cl_loi_list_lock);
 387         atomic_set(&cli->cl_pending_w_pages, 0);
 388         atomic_set(&cli->cl_pending_r_pages, 0);
 389         cli->cl_r_in_flight = 0;
 390         cli->cl_w_in_flight = 0;
 391
 392         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 393         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 394         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 395         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 396         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 397         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 398
 399         /* lru for osc. */
 400         INIT_LIST_HEAD(&cli->cl_lru_osc);
 401         atomic_set(&cli->cl_lru_shrinkers, 0);
 402         atomic_long_set(&cli->cl_lru_busy, 0);
 403         atomic_long_set(&cli->cl_lru_in_list, 0);
 404         INIT_LIST_HEAD(&cli->cl_lru_list);
 405         spin_lock_init(&cli->cl_lru_list_lock);
 406         atomic_long_set(&cli->cl_unstable_count, 0);
 407         INIT_LIST_HEAD(&cli->cl_shrink_list);
 408         INIT_LIST_HEAD(&cli->cl_grant_chain);
 409
 410         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 411         cli->cl_rpcs_in_flight = 0;
 412
 413         init_waitqueue_head(&cli->cl_destroy_waitq);
 414         atomic_set(&cli->cl_destroy_in_flight, 0);
 415
 416
 417         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 418         cli->cl_preferred_cksum_type = 0;
 419 #ifdef ENABLE_CHECKSUM
 420         /* Turn on checksumming by default. */
 421         cli->cl_checksum = 1;
 422         /*
 423          * The supported checksum types will be worked out at connect time
 424          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 425          * through procfs.
 426          */
 427         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 428 #endif
 429         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 430
 431         /*
 432          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 433          * from OFD after connecting.
 434          */
 435         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 436
 437         cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
 438
 439         /*
 440          * set cl_chunkbits default value to PAGE_SHIFT,
 441          * it will be updated at OSC connection time.
 442          */
 443         cli->cl_chunkbits = PAGE_SHIFT;
 444
 445         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 446                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 447         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 448                 cli->cl_max_rpcs_in_flight = 2;
 449         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 450                 cli->cl_max_rpcs_in_flight = 3;
 451         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 452                 cli->cl_max_rpcs_in_flight = 4;
 453         } else {
 454                 if (osc_on_mdt(obddev->obd_name))
 455                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 456                 else
 457                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 458         }
 459
 460         spin_lock_init(&cli->cl_mod_rpcs_lock);
 461         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 462         cli->cl_max_mod_rpcs_in_flight = 0;
 463         cli->cl_mod_rpcs_in_flight = 0;
 464         cli->cl_close_rpcs_in_flight = 0;
 465         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 466         cli->cl_mod_tag_bitmap = NULL;
 467
 468         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 469
 470         if (connect_op == MDS_CONNECT) {
 471                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 472                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 473                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 474                 if (cli->cl_mod_tag_bitmap == NULL)
 475                         GOTO(err, rc = -ENOMEM);
 476         }
 477
 478         rc = ldlm_get_ref();
 479         if (rc) {
 480                 CERROR("ldlm_get_ref failed: %d\n", rc);
 481                 GOTO(err, rc);
 482         }
 483
 484         ptlrpc_init_client(rq_portal, rp_portal, name,
 485                            &obddev->obd_ldlm_client);
 486
 487         imp = class_new_import(obddev);
 488         if (imp == NULL)
 489                 GOTO(err_ldlm, rc = -ENOENT);
 490         imp->imp_client = &obddev->obd_ldlm_client;
 491         imp->imp_connect_op = connect_op;
 492         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 493                LUSTRE_CFG_BUFLEN(lcfg, 1));
 494         class_import_put(imp);
 495
 496         if (lustre_cfg_buf(lcfg, 4)) {
 497                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 498
 499                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 500                         rc = -EINVAL;
 501                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 502                                obddev->obd_name, lustre_cfg_string(lcfg, 4),
 503                                rc);
 504                         GOTO(err_import, rc);
 505                 }
 506                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 507                 imp->imp_connection = &fake_conn;
 508         }
 509
 510         rc = client_import_add_conn(imp, &server_uuid, 1);
 511         if (rc) {
 512                 CERROR("can't add initial connection\n");
 513                 GOTO(err_import, rc);
 514         }
 515         imp->imp_connection = NULL;
 516
 517         cli->cl_import = imp;
 518         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 519         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 520
 521         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 522                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 523                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 524                                name, obddev->obd_name,
 525                                cli->cl_target_uuid.uuid);
 526                         spin_lock(&imp->imp_lock);
 527                         imp->imp_deactive = 1;
 528                         spin_unlock(&imp->imp_lock);
 529                 }
 530         }
 531
 532         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 533                                                    LDLM_NAMESPACE_CLIENT,
 534                                                    LDLM_NAMESPACE_GREEDY,
 535                                                    ns_type);
 536         if (obddev->obd_namespace == NULL) {
 537                 CERROR("Unable to create client namespace - %s\n",
 538                        obddev->obd_name);
 539                 GOTO(err_import, rc = -ENOMEM);
 540         }
 541
 542         RETURN(rc);
 543
 544 err_import:
 545         class_destroy_import(imp);
 546 err_ldlm:
 547         ldlm_put_ref();
 548 err:
 549         if (cli->cl_mod_tag_bitmap != NULL)
 550                 OBD_FREE(cli->cl_mod_tag_bitmap,
 551                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 552         cli->cl_mod_tag_bitmap = NULL;
 553         RETURN(rc);
 554
 555 }
 556 EXPORT_SYMBOL(client_obd_setup);
 557
 558 int client_obd_cleanup(struct obd_device *obddev)
 559 {
 560         struct client_obd *cli = &obddev->u.cli;
 561
 562         ENTRY;
 563
 564         ldlm_namespace_free_post(obddev->obd_namespace);
 565         obddev->obd_namespace = NULL;
 566
 567         obd_cleanup_client_import(obddev);
 568         LASSERT(obddev->u.cli.cl_import == NULL);
 569
 570         ldlm_put_ref();
 571
 572         if (cli->cl_mod_tag_bitmap != NULL)
 573                 OBD_FREE(cli->cl_mod_tag_bitmap,
 574                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 575         cli->cl_mod_tag_bitmap = NULL;
 576
 577         RETURN(0);
 578 }
 579 EXPORT_SYMBOL(client_obd_cleanup);
 580
 581 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 582 int client_connect_import(const struct lu_env *env,
 583                           struct obd_export **exp,
 584                           struct obd_device *obd, struct obd_uuid *cluuid,
 585                           struct obd_connect_data *data, void *localdata)
 586 {
 587         struct client_obd *cli = &obd->u.cli;
 588         struct obd_import *imp = cli->cl_import;
 589         struct obd_connect_data *ocd;
 590         struct lustre_handle conn = { 0 };
 591         int rc;
 592
 593         ENTRY;
 594
 595         *exp = NULL;
 596         down_write(&cli->cl_sem);
 597         if (cli->cl_conn_count > 0)
 598                 GOTO(out_sem, rc = -EALREADY);
 599
 600         rc = class_connect(&conn, obd, cluuid);
 601         if (rc)
 602                 GOTO(out_sem, rc);
 603
 604         cli->cl_conn_count++;
 605         *exp = class_conn2export(&conn);
 606
 607         LASSERT(obd->obd_namespace);
 608
 609         imp->imp_dlm_handle = conn;
 610         rc = ptlrpc_init_import(imp);
 611         if (rc != 0)
 612                 GOTO(out_ldlm, rc);
 613
 614         ocd = &imp->imp_connect_data;
 615         if (data) {
 616                 *ocd = *data;
 617                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 618                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 619         }
 620
 621         rc = ptlrpc_connect_import(imp);
 622         if (rc != 0) {
 623                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 624                 GOTO(out_ldlm, rc);
 625         }
 626         LASSERT(*exp != NULL && (*exp)->exp_connection);
 627
 628         if (data) {
 629                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 630                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 631                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 632                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 633                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 634         }
 635
 636         ptlrpc_pinger_add_import(imp);
 637
 638         EXIT;
 639
 640         if (rc) {
 641 out_ldlm:
 642                 cli->cl_conn_count--;
 643                 class_disconnect(*exp);
 644                 *exp = NULL;
 645         }
 646 out_sem:
 647         up_write(&cli->cl_sem);
 648
 649         if (!rc && localdata) {
 650                 LASSERT(cli->cl_cache == NULL); /* only once */
 651                 cli->cl_cache = (struct cl_client_cache *)localdata;
 652                 cl_cache_incref(cli->cl_cache);
 653                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 654
 655                 /* add this osc into entity list */
 656                 LASSERT(list_empty(&cli->cl_lru_osc));
 657                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 658                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 659                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 660         }
 661
 662         return rc;
 663 }
 664 EXPORT_SYMBOL(client_connect_import);
 665
 666 int client_disconnect_export(struct obd_export *exp)
 667 {
 668         struct obd_device *obd = class_exp2obd(exp);
 669         struct client_obd *cli;
 670         struct obd_import *imp;
 671         int rc = 0, err;
 672
 673         ENTRY;
 674
 675         if (!obd) {
 676                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 677                        exp, exp ? exp->exp_handle.h_cookie : -1);
 678                 RETURN(-EINVAL);
 679         }
 680
 681         cli = &obd->u.cli;
 682         imp = cli->cl_import;
 683
 684         down_write(&cli->cl_sem);
 685         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 686                 cli->cl_conn_count);
 687
 688         if (cli->cl_conn_count == 0) {
 689                 CERROR("disconnecting disconnected device (%s)\n",
 690                        obd->obd_name);
 691                 GOTO(out_disconnect, rc = -EINVAL);
 692         }
 693
 694         cli->cl_conn_count--;
 695         if (cli->cl_conn_count != 0)
 696                 GOTO(out_disconnect, rc = 0);
 697
 698         /*
 699          * Mark import deactivated now, so we don't try to reconnect if any
 700          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 701          * fully deactivate the import, or that would drop all requests.
 702          */
 703         spin_lock(&imp->imp_lock);
 704         imp->imp_deactive = 1;
 705         spin_unlock(&imp->imp_lock);
 706
 707         /*
 708          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 709          * delete it regardless.  (It's safe to delete an import that was
 710          * never added.)
 711          */
 712         (void)ptlrpc_pinger_del_import(imp);
 713
 714         if (obd->obd_namespace != NULL) {
 715                 /* obd_force == local only */
 716                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 717                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 718                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 719                                           obd->obd_force);
 720         }
 721
 722         /*
 723          * There's no need to hold sem while disconnecting an import,
 724          * and it may actually cause deadlock in GSS.
 725          */
 726         up_write(&cli->cl_sem);
 727         rc = ptlrpc_disconnect_import(imp, 0);
 728         down_write(&cli->cl_sem);
 729
 730         ptlrpc_invalidate_import(imp);
 731
 732         EXIT;
 733
 734 out_disconnect:
 735         /*
 736          * Use server style - class_disconnect should be always called for
 737          * o_disconnect.
 738          */
 739         err = class_disconnect(exp);
 740         if (!rc && err)
 741                 rc = err;
 742
 743         up_write(&cli->cl_sem);
 744
 745         RETURN(rc);
 746 }
 747 EXPORT_SYMBOL(client_disconnect_export);
 748
 749 #ifdef HAVE_SERVER_SUPPORT
 750 int server_disconnect_export(struct obd_export *exp)
 751 {
 752         int rc;
 753
 754         ENTRY;
 755
 756         /* Disconnect early so that clients can't keep using export. */
 757         rc = class_disconnect(exp);
 758         /* Close import to avoid sending any requests. */
 759         if (exp->exp_imp_reverse)
 760                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 761
 762         ldlm_bl_thread_wakeup();
 763
 764         /* complete all outstanding replies */
 765         spin_lock(&exp->exp_lock);
 766         while (!list_empty(&exp->exp_outstanding_replies)) {
 767                 struct ptlrpc_reply_state *rs =
 768                         list_entry(exp->exp_outstanding_replies.next,
 769                                        struct ptlrpc_reply_state, rs_exp_list);
 770                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 771
 772                 spin_lock(&svcpt->scp_rep_lock);
 773
 774                 list_del_init(&rs->rs_exp_list);
 775
 776                 spin_lock(&rs->rs_lock);
 777                 /* clear rs_convert_lock to make sure rs is handled and put */
 778                 rs->rs_convert_lock = 0;
 779                 ptlrpc_schedule_difficult_reply(rs);
 780                 spin_unlock(&rs->rs_lock);
 781
 782                 spin_unlock(&svcpt->scp_rep_lock);
 783         }
 784         spin_unlock(&exp->exp_lock);
 785
 786         RETURN(rc);
 787 }
 788 EXPORT_SYMBOL(server_disconnect_export);
 789
 790 /*
 791  * --------------------------------------------------------------------------
 792  * from old lib/target.c
 793  * --------------------------------------------------------------------------
 794  */
 795
 796 static int target_handle_reconnect(struct lustre_handle *conn,
 797                                    struct obd_export *exp,
 798                                    struct obd_uuid *cluuid)
 799 {
 800         struct obd_device *target;
 801         struct lustre_handle *hdl;
 802         ktime_t remaining;
 803         s64 timeout;
 804         int rc = 0;
 805
 806         ENTRY;
 807         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 808         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 809                 conn->cookie = exp->exp_handle.h_cookie;
 810                 CDEBUG(D_HA,
 811                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 812                        cluuid->uuid, exp, conn->cookie);
 813                 RETURN(0);
 814         }
 815
 816         target = exp->exp_obd;
 817
 818         /* Might be a re-connect after a partition. */
 819         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 820                 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",
 821                               target->obd_name,
 822                               obd_uuid2str(&exp->exp_client_uuid),
 823                               obd_export_nid2str(exp),
 824                               hdl->cookie, conn->cookie);
 825                 memset(conn, 0, sizeof(*conn));
 826                 /*
 827                  * target_handle_connect() treats EALREADY and
 828                  * -EALREADY differently.  -EALREADY is an error
 829                  * (same UUID, different handle).
 830                  */
 831                 RETURN(-EALREADY);
 832         }
 833
 834         if (!target->obd_recovering) {
 835                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 836                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 837                         obd_export_nid2str(exp));
 838                 GOTO(out_already, rc);
 839         }
 840
 841         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 842         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 843         if (timeout > 0) {
 844                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 845                               target->obd_name,
 846                               obd_uuid2str(&exp->exp_client_uuid),
 847                               obd_export_nid2str(exp),
 848                               target->obd_max_recoverable_clients,
 849                               timeout / 60, timeout % 60);
 850         } else {
 851                 struct target_distribute_txn_data *tdtd;
 852                 int size = 0;
 853                 int count = 0;
 854                 char *buf = NULL;
 855
 856                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 857                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 858                         buf = tdtd->tdtd_show_update_logs_retrievers(
 859                                 tdtd->tdtd_show_retrievers_cbdata,
 860                                 &size, &count);
 861
 862                 if (count > 0)
 863                         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",
 864                                       target->obd_name,
 865                                       obd_uuid2str(&exp->exp_client_uuid),
 866                                       obd_export_nid2str(exp), count,
 867                                       buf ? buf : "unknown (not enough RAM)",
 868                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 869                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 870                                       target->obd_name);
 871                 else
 872                         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",
 873                                       target->obd_name, abs(timeout) / 60,
 874                                       abs(timeout) % 60, target->obd_name);
 875
 876                 if (buf != NULL)
 877                         OBD_FREE(buf, size);
 878         }
 879
 880 out_already:
 881         conn->cookie = exp->exp_handle.h_cookie;
 882         /*
 883          * target_handle_connect() treats EALREADY and
 884          * -EALREADY differently.  EALREADY means we are
 885          * doing a valid reconnect from the same client.
 886          */
 887         RETURN(EALREADY);
 888 }
 889
 890 static void
 891 check_and_start_recovery_timer(struct obd_device *obd,
 892                                struct ptlrpc_request *req, int new_client);
 893
 894 /**
 895  * update flags for import during reconnect process
 896  */
 897 static int rev_import_flags_update(struct obd_import *revimp,
 898                                    struct ptlrpc_request *req)
 899 {
 900         int rc;
 901         struct obd_connect_data *data;
 902
 903         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 904
 905         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 906                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 907         else
 908                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 909
 910         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 911
 912         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 913         if (rc) {
 914                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 915                         revimp->imp_client->cli_name,
 916                         libcfs_id2str(req->rq_peer), rc);
 917                 return rc;
 918         }
 919
 920         return 0;
 921 }
 922
 923 /**
 924  * Allocate a new reverse import for an export.
 925  *
 926  * \retval -errno in case error hit
 927  * \retval 0 if reverse import correctly init
 928  **/
 929 int rev_import_init(struct obd_export *export)
 930 {
 931         struct obd_device *obd = export->exp_obd;
 932         struct obd_import *revimp;
 933
 934         LASSERT(export->exp_imp_reverse == NULL);
 935
 936         revimp = class_new_import(obd);
 937         if (revimp == NULL)
 938                 return -ENOMEM;
 939
 940         revimp->imp_remote_handle.cookie = 0ULL;
 941         revimp->imp_client = &obd->obd_ldlm_client;
 942         revimp->imp_dlm_fake = 1;
 943
 944         /* it is safe to connect import in new state as no sends possible */
 945         spin_lock(&export->exp_lock);
 946         export->exp_imp_reverse = revimp;
 947         spin_unlock(&export->exp_lock);
 948         class_import_put(revimp);
 949
 950         return 0;
 951 }
 952 EXPORT_SYMBOL(rev_import_init);
 953
 954 /**
 955  * Handle reconnect for an export.
 956  *
 957  * \param exp export to handle reconnect process
 958  * \param req client reconnect request
 959  *
 960  * \retval -rc in case securitfy flavor can't be changed
 961  * \retval 0 in case none problems
 962  */
 963 static int rev_import_reconnect(struct obd_export *exp,
 964                                 struct ptlrpc_request *req)
 965 {
 966         struct obd_import *revimp = exp->exp_imp_reverse;
 967         struct lustre_handle *lh;
 968         int rc;
 969
 970         /* avoid sending a request until import flags are changed */
 971         ptlrpc_import_enter_resend(revimp);
 972
 973         if (revimp->imp_connection != NULL)
 974                 ptlrpc_connection_put(revimp->imp_connection);
 975
 976         /*
 977          * client from recovery don't have a handle so we need to take from
 978          * request. it may produce situation when wrong client connected
 979          * to recovery as we trust a client uuid
 980          */
 981         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 982         revimp->imp_remote_handle = *lh;
 983
 984         /*
 985          * unknown versions will be caught in
 986          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
 987          */
 988         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 989
 990         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 991
 992         rc = rev_import_flags_update(revimp, req);
 993         if (rc != 0) {
 994                 /*
 995                  * it is safe to still be in RECOVERY phase as we are not able
 996                  * to setup correct security flavor so requests are not able to
 997                  * be delivered correctly
 998                  */
 999                 return rc;
1000         }
1001
1002         /* resend all rpc's via new connection */
1003         return ptlrpc_import_recovery_state_machine(revimp);
1004 }
1005
1006 int target_handle_connect(struct ptlrpc_request *req)
1007 {
1008         struct obd_device *target = NULL;
1009         struct obd_export *export = NULL;
1010         /*
1011          * connect handle - filled from target_handle_reconnect in
1012          * reconnect case
1013          */
1014         struct lustre_handle conn;
1015         struct lustre_handle *tmp;
1016         struct obd_uuid cluuid;
1017         char *str;
1018         int rc = 0;
1019         char *target_start;
1020         int target_len;
1021         bool mds_conn = false, lw_client = false, initial_conn = false;
1022         bool mds_mds_conn = false;
1023         bool new_mds_mds_conn = false;
1024         struct obd_connect_data *data, *tmpdata;
1025         int size, tmpsize;
1026         lnet_nid_t *client_nid = NULL;
1027
1028         ENTRY;
1029
1030         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1031
1032         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1033         if (str == NULL) {
1034                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1035                 GOTO(out, rc = -EINVAL);
1036         }
1037
1038         target = class_dev_by_str(str);
1039         if (!target) {
1040                 deuuidify(str, NULL, &target_start, &target_len);
1041                 LCONSOLE_ERROR_MSG(0x137,
1042                                    "%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",
1043                                    str, libcfs_nid2str(req->rq_peer.nid));
1044                 GOTO(out, rc = -ENODEV);
1045         }
1046
1047         spin_lock(&target->obd_dev_lock);
1048
1049         target->obd_conn_inprogress++;
1050
1051         if (target->obd_stopping || !target->obd_set_up) {
1052                 spin_unlock(&target->obd_dev_lock);
1053
1054                 deuuidify(str, NULL, &target_start, &target_len);
1055                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1056                               target_len, target_start,
1057                               libcfs_nid2str(req->rq_peer.nid),
1058                               (target->obd_stopping ?
1059                                "stopping" : "not set up"));
1060                 GOTO(out, rc = -ENODEV);
1061         }
1062
1063         if (target->obd_no_conn) {
1064                 spin_unlock(&target->obd_dev_lock);
1065
1066                 CDEBUG(D_INFO,
1067                        "%s: Temporarily refusing client connection from %s\n",
1068                        target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1069                 GOTO(out, rc = -EAGAIN);
1070         }
1071
1072         spin_unlock(&target->obd_dev_lock);
1073
1074         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1075         if (str == NULL) {
1076                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1077                 GOTO(out, rc = -EINVAL);
1078         }
1079
1080         obd_str2uuid(&cluuid, str);
1081
1082         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1083         if (tmp == NULL)
1084                 GOTO(out, rc = -EPROTO);
1085
1086         conn = *tmp;
1087
1088         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1089                                     RCL_CLIENT);
1090         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1091                 GOTO(out, rc = -EPROTO);
1092         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1093         if (!data)
1094                 GOTO(out, rc = -EPROTO);
1095
1096         rc = req_capsule_server_pack(&req->rq_pill);
1097         if (rc)
1098                 GOTO(out, rc);
1099
1100 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1101         /*
1102          * Don't allow clients to connect that are using old 1.8 format
1103          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1104          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1105          * FULL20 flag should be set on all connections since 2.0, but no
1106          * longer affects behaviour.
1107          *
1108          * Later this check will be disabled and the flag can be retired
1109          * completely once interop with 3.0 is no longer needed.
1110          */
1111         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1112                 GOTO(out, rc = -EPROTO);
1113
1114         /*
1115          * Don't allow liblustre clients to connect.
1116          * - testing was disabled in v2_2_50_0-61-g6a75d65
1117          * - building was disabled in v2_5_58_0-28-g7277179
1118          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1119          * - clients were refused connect for version difference > 0.0.1.32
1120          */
1121         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1122                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1123                 GOTO(out, rc = -EPROTO);
1124         }
1125 #endif
1126
1127         /*
1128          * Note: lw_client is needed in MDS-MDS failover during update log
1129          * processing, so we needs to allow lw_client to be connected at
1130          * anytime, instead of only the initial connection
1131          */
1132         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1133
1134         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1135                 initial_conn = true;
1136                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1137                 mds_mds_conn = (data->ocd_connect_flags &
1138                                 OBD_CONNECT_MDS_MDS) != 0;
1139
1140                 /*
1141                  * OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1142                  * for Imperative Recovery connection from MGC to MGS.
1143                  *
1144                  * Via check OBD_CONNECT_FID, we can distinguish whether
1145                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1146                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1147                  */
1148                 if (!lw_client &&
1149                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1150                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1151                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1152                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1153                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1154                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1155
1156                         /*
1157                          * We do not support the MDT-MDT interoperations with
1158                          * different version MDT because of protocol changes.
1159                          */
1160                         if (unlikely(major != LUSTRE_MAJOR ||
1161                                      minor != LUSTRE_MINOR ||
1162                                      abs(patch - LUSTRE_PATCH) > 3)) {
1163                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1164                                               target->obd_name, LUSTRE_MAJOR,
1165                                               LUSTRE_MINOR, LUSTRE_PATCH,
1166                                               LUSTRE_FIX, major, minor, patch,
1167                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1168                                               libcfs_nid2str(req->rq_peer.nid),
1169                                               str);
1170                                 GOTO(out, rc = -EPROTO);
1171                         }
1172                 }
1173         }
1174
1175         /* lctl gets a backstage, all-access pass. */
1176         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1177                 goto dont_check_exports;
1178
1179         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1180         if (!export)
1181                 goto no_export;
1182
1183         /* We've found an export in the hash. */
1184
1185         spin_lock(&export->exp_lock);
1186
1187         if (export->exp_connecting) { /* b=9635, et. al. */
1188                 spin_unlock(&export->exp_lock);
1189                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1190                               export->exp_obd->obd_name, export,
1191                               libcfs_nid2str(req->rq_peer.nid));
1192                 class_export_put(export);
1193                 export = NULL;
1194                 rc = -EALREADY;
1195         } else if ((mds_conn || (lw_client && initial_conn) ||
1196                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1197                    export->exp_connection != NULL) {
1198                 spin_unlock(&export->exp_lock);
1199                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1200                         /* MDS or LWP reconnected after failover. */
1201                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1202                                       target->obd_name,
1203                                       mds_conn ? "MDS" : "LWP",
1204                                       libcfs_nid2str(req->rq_peer.nid),
1205                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
1206                 } else {
1207                         /* New MDS connection from the same NID. */
1208                         LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1209                                       target->obd_name,
1210                                       mds_conn ? "MDS" : "LWP",
1211                                       libcfs_nid2str(req->rq_peer.nid));
1212                 }
1213
1214                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1215                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1216                         /*
1217                          * Because exports between MDTs will always be
1218                          * kept, let's do not fail such export if they
1219                          * come from the same NID, otherwise it might
1220                          * cause eviction between MDTs, which might
1221                          * cause namespace inconsistency
1222                          */
1223                         spin_lock(&export->exp_lock);
1224                         export->exp_connecting = 1;
1225                         export->exp_conn_cnt = 0;
1226                         spin_unlock(&export->exp_lock);
1227                         conn.cookie = export->exp_handle.h_cookie;
1228                         rc = EALREADY;
1229                 } else {
1230                         class_fail_export(export);
1231                         class_export_put(export);
1232                         export = NULL;
1233                         rc = 0;
1234                 }
1235         } else if (export->exp_connection != NULL && initial_conn &&
1236                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1237                 spin_unlock(&export->exp_lock);
1238                 /* In MDS failover we have static UUID but NID can change. */
1239                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1240                               target->obd_name, cluuid.uuid,
1241                               libcfs_nid2str(req->rq_peer.nid),
1242                               libcfs_nid2str(
1243                                         export->exp_connection->c_peer.nid));
1244                 rc = -EALREADY;
1245                 class_export_put(export);
1246                 export = NULL;
1247         } else {
1248                 export->exp_connecting = 1;
1249                 spin_unlock(&export->exp_lock);
1250                 LASSERT(export->exp_obd == target);
1251
1252                 rc = target_handle_reconnect(&conn, export, &cluuid);
1253         }
1254
1255         /* If we found an export, we already unlocked. */
1256         if (!export) {
1257 no_export:
1258                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1259         } else if (req->rq_export == NULL &&
1260                    atomic_read(&export->exp_rpc_count) > 0) {
1261                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1262                               target->obd_name, cluuid.uuid,
1263                               libcfs_nid2str(req->rq_peer.nid),
1264                               atomic_read(&export->exp_refcount));
1265                         GOTO(out, rc = -EBUSY);
1266         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1267                    rc != EALREADY) {
1268                 if (!strstr(cluuid.uuid, "mdt"))
1269                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1270                                       target->obd_name, cluuid.uuid,
1271                                       libcfs_nid2str(req->rq_peer.nid));
1272                 GOTO(out, rc = -EALREADY);
1273         } else {
1274                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1275         }
1276
1277         if (rc < 0)
1278                 GOTO(out, rc);
1279
1280         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1281                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1282                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1283                export, ktime_get_real_seconds(),
1284                export ? export->exp_last_request_time : 0);
1285
1286         /*
1287          * If this is the first time a client connects, reset the recovery
1288          * timer. Discard lightweight connections which might be local.
1289          */
1290         if (!lw_client && rc == 0 && target->obd_recovering)
1291                 check_and_start_recovery_timer(target, req, export == NULL);
1292
1293         /*
1294          * We want to handle EALREADY but *not* -EALREADY from
1295          * target_handle_reconnect(), return reconnection state in a flag.
1296          */
1297         if (rc == EALREADY) {
1298                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1299                 rc = 0;
1300         } else {
1301                 LASSERT(rc == 0);
1302         }
1303
1304         /* Tell the client if we support replayable requests. */
1305         if (target->obd_replayable)
1306                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1307         client_nid = &req->rq_peer.nid;
1308
1309         if (export == NULL) {
1310                 /* allow lightweight connections during recovery */
1311                 /*
1312                  * allow "new" MDT to be connected during recovery, since we
1313                  * need retrieve recovery update records from it
1314                  */
1315                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1316                         struct hrtimer *timer = &target->obd_recovery_timer;
1317                         ktime_t remaining;
1318                         s64 timeout, left;
1319                         int in_progress;
1320                         int connected;
1321                         int known;
1322                         int stale;
1323                         char *msg;
1324
1325                         connected = atomic_read(&target->obd_connected_clients);
1326                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1327                         known = target->obd_max_recoverable_clients;
1328                         stale = target->obd_stale_clients;
1329                         remaining = hrtimer_expires_remaining(timer);
1330                         left = ktime_divns(remaining, NSEC_PER_SEC);
1331
1332                         if (ktime_to_ns(remaining) > 0) {
1333                                 msg = "to recover in";
1334                                 timeout = left;
1335                         } else {
1336                                 msg = "already passed deadline";
1337                                 timeout = -left;
1338                         }
1339
1340                         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",
1341                                       target->obd_name, cluuid.uuid,
1342                                       libcfs_nid2str(req->rq_peer.nid), known,
1343                                       connected - in_progress, in_progress,
1344                                       stale, msg, timeout / 60, timeout % 60);
1345                         rc = -EBUSY;
1346                 } else {
1347 dont_check_exports:
1348                         rc = obd_connect(req->rq_svc_thread->t_env,
1349                                          &export, target, &cluuid, data,
1350                                          client_nid);
1351                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1352                                 lustre_msg_add_op_flags(req->rq_repmsg,
1353                                                         MSG_CONNECT_RECOVERING);
1354                         if (rc == 0) {
1355                                 conn.cookie = export->exp_handle.h_cookie;
1356                                 rc = rev_import_init(export);
1357                         }
1358
1359                         if (mds_mds_conn)
1360                                 new_mds_mds_conn = true;
1361                 }
1362         } else {
1363                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1364                                    export, target, &cluuid, data, client_nid);
1365         }
1366         if (rc)
1367                 GOTO(out, rc);
1368
1369         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1370         data->ocd_instance = target->u.obt.obt_instance;
1371
1372         /*
1373          * Return only the parts of obd_connect_data that we understand, so the
1374          * client knows that we don't understand the rest.
1375          */
1376         if (data) {
1377                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1378                                                RCL_SERVER);
1379                 tmpdata = req_capsule_server_get(&req->rq_pill,
1380                                                  &RMF_CONNECT_DATA);
1381                 /*
1382                  * Don't use struct assignment here, because the client reply
1383                  * buffer may be smaller/larger than the local struct
1384                  * obd_connect_data.
1385                  */
1386                 memcpy(tmpdata, data, min(tmpsize, size));
1387         }
1388
1389         /*
1390          * If the client and the server are the same node, we will already
1391          * have an export that really points to the client's DLM export,
1392          * because we have a shared handles table.
1393          *
1394          * XXX this will go away when shaver stops sending the "connect" handle
1395          * in the real "remote handle" field of the request --phik 24 Apr 2003
1396          */
1397         ptlrpc_request_change_export(req, export);
1398
1399         spin_lock(&export->exp_lock);
1400         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1401                 spin_unlock(&export->exp_lock);
1402                 CDEBUG(D_RPCTRACE,
1403                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1404                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1405                        export->exp_conn_cnt,
1406                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1407
1408                 GOTO(out, rc = -EALREADY);
1409         }
1410         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1411         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1412         spin_unlock(&export->exp_lock);
1413
1414         if (export->exp_connection != NULL) {
1415                 /* Check to see if connection came from another NID. */
1416                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1417                     !hlist_unhashed(&export->exp_nid_hash))
1418                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1419                                      &export->exp_connection->c_peer.nid,
1420                                      &export->exp_nid_hash);
1421
1422                 ptlrpc_connection_put(export->exp_connection);
1423         }
1424
1425         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1426                                                        req->rq_self,
1427                                                        &cluuid);
1428         if (hlist_unhashed(&export->exp_nid_hash))
1429                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1430                              &export->exp_connection->c_peer.nid,
1431                              &export->exp_nid_hash);
1432
1433         lustre_msg_set_handle(req->rq_repmsg, &conn);
1434
1435         rc = rev_import_reconnect(export, req);
1436         if (rc != 0)
1437                 GOTO(out, rc);
1438
1439         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1440                 int has_transno;
1441                 __u64 transno = data->ocd_transno;
1442
1443                 spin_lock(&export->exp_lock);
1444                 /*
1445                  * possible race with class_disconnect_stale_exports,
1446                  * export may be already in the eviction process
1447                  */
1448                 if (export->exp_failed) {
1449                         spin_unlock(&export->exp_lock);
1450                         GOTO(out, rc = -ENODEV);
1451                 }
1452                 export->exp_in_recovery = 1;
1453                 export->exp_req_replay_needed = 1;
1454                 export->exp_lock_replay_needed = 1;
1455                 spin_unlock(&export->exp_lock);
1456
1457                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1458                                  MSG_CONNECT_TRANSNO);
1459                 if (has_transno && transno == 0)
1460                         CWARN("Connect with zero transno!\n");
1461
1462                 if (has_transno && transno > 0 &&
1463                     transno < target->obd_next_recovery_transno &&
1464                     transno > target->obd_last_committed) {
1465                         /* Another way is to use cmpxchg() to be lock-free. */
1466                         spin_lock(&target->obd_recovery_task_lock);
1467                         if (transno < target->obd_next_recovery_transno)
1468                                 target->obd_next_recovery_transno = transno;
1469                         spin_unlock(&target->obd_recovery_task_lock);
1470                 }
1471
1472                 atomic_inc(&target->obd_req_replay_clients);
1473                 atomic_inc(&target->obd_lock_replay_clients);
1474                 /*
1475                  * Note: MDS-MDS connection is allowed to be connected during
1476                  * recovery, no matter if the exports needs to be recoveried.
1477                  * Because we need retrieve updates logs from all other MDTs.
1478                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1479                  * also needs to be increased to match other recovery checking
1480                  * condition.
1481                  */
1482                 if (new_mds_mds_conn)
1483                         target->obd_max_recoverable_clients++;
1484                 if (atomic_inc_return(&target->obd_connected_clients) ==
1485                     target->obd_max_recoverable_clients)
1486                         wake_up(&target->obd_next_transno_waitq);
1487         }
1488
1489         /* Tell the client we're in recovery, when client is involved in it. */
1490         if (target->obd_recovering && !lw_client)
1491                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1492
1493 out:
1494         if (export) {
1495                 spin_lock(&export->exp_lock);
1496                 export->exp_connecting = 0;
1497                 spin_unlock(&export->exp_lock);
1498
1499                 class_export_put(export);
1500         }
1501         if (target != NULL) {
1502                 spin_lock(&target->obd_dev_lock);
1503                 target->obd_conn_inprogress--;
1504                 spin_unlock(&target->obd_dev_lock);
1505                 class_decref(target, "find", current);
1506         }
1507         req->rq_status = rc;
1508         RETURN(rc);
1509 }
1510
1511 int target_handle_disconnect(struct ptlrpc_request *req)
1512 {
1513         int rc;
1514
1515         ENTRY;
1516
1517         rc = req_capsule_server_pack(&req->rq_pill);
1518         if (rc)
1519                 RETURN(rc);
1520
1521         /* Keep the rq_export around so we can send the reply. */
1522         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1523
1524         RETURN(0);
1525 }
1526
1527 void target_destroy_export(struct obd_export *exp)
1528 {
1529         struct obd_import *imp = NULL;
1530         /*
1531          * exports created from last_rcvd data, and "fake"
1532          * exports created by lctl don't have an import
1533          */
1534         spin_lock(&exp->exp_lock);
1535         if (exp->exp_imp_reverse != NULL) {
1536                 imp = exp->exp_imp_reverse;
1537                 exp->exp_imp_reverse = NULL;
1538         }
1539         spin_unlock(&exp->exp_lock);
1540         if (imp != NULL)
1541                 client_destroy_import(imp);
1542
1543         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1544         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1545         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1546         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1547 }
1548 EXPORT_SYMBOL(target_destroy_export);
1549
1550 /*
1551  * Recovery functions
1552  */
1553 static void target_request_copy_get(struct ptlrpc_request *req)
1554 {
1555         class_export_rpc_inc(req->rq_export);
1556         LASSERT(list_empty(&req->rq_list));
1557         INIT_LIST_HEAD(&req->rq_replay_list);
1558
1559         /* Increase refcount to keep request in queue. */
1560         atomic_inc(&req->rq_refcount);
1561         /* Let export know it has replays to be handled. */
1562         atomic_inc(&req->rq_export->exp_replay_count);
1563 }
1564
1565 static void target_request_copy_put(struct ptlrpc_request *req)
1566 {
1567         LASSERT(list_empty(&req->rq_replay_list));
1568         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1569
1570         atomic_dec(&req->rq_export->exp_replay_count);
1571         class_export_rpc_dec(req->rq_export);
1572         ptlrpc_server_drop_request(req);
1573 }
1574
1575 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1576 {
1577         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1578         struct obd_export *exp = req->rq_export;
1579         struct ptlrpc_request *reqiter;
1580         struct ptlrpc_request *dup_req = NULL;
1581         int dup = 0;
1582
1583         LASSERT(exp);
1584
1585         spin_lock(&exp->exp_lock);
1586         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1587                             rq_replay_list) {
1588                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1589                         dup_req = reqiter;
1590                         dup = 1;
1591                         break;
1592                 }
1593         }
1594
1595         if (dup) {
1596                 /* We expect it with RESENT and REPLAY flags. */
1597                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1598                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1599                         CERROR("invalid flags %x of resent replay\n",
1600                                lustre_msg_get_flags(req->rq_reqmsg));
1601
1602                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1603                         __u32 new_conn;
1604
1605                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1606                         if (new_conn >
1607                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1608                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1609                                                         new_conn);
1610                 }
1611         } else {
1612                 list_add_tail(&req->rq_replay_list,
1613                               &exp->exp_req_replay_queue);
1614         }
1615
1616         spin_unlock(&exp->exp_lock);
1617         return dup;
1618 }
1619
1620 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1621 {
1622         LASSERT(!list_empty(&req->rq_replay_list));
1623         LASSERT(req->rq_export);
1624
1625         spin_lock(&req->rq_export->exp_lock);
1626         list_del_init(&req->rq_replay_list);
1627         spin_unlock(&req->rq_export->exp_lock);
1628 }
1629
1630 static void target_finish_recovery(struct lu_target *lut)
1631 {
1632         struct obd_device *obd = lut->lut_obd;
1633
1634         ENTRY;
1635
1636         /* Only log a recovery message when recovery has occurred. */
1637         if (obd->obd_recovery_start) {
1638                 time64_t now = ktime_get_real_seconds();
1639                 time64_t elapsed_time;
1640
1641                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1642                                      1);
1643                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1644                               obd->obd_name, (s64)elapsed_time / 60,
1645                               (s64)elapsed_time % 60,
1646                               obd->obd_max_recoverable_clients,
1647                               atomic_read(&obd->obd_connected_clients),
1648                               obd->obd_stale_clients,
1649                               obd->obd_stale_clients == 1 ? "was" : "were");
1650         }
1651
1652         ldlm_reprocess_recovery_done(obd->obd_namespace);
1653         spin_lock(&obd->obd_recovery_task_lock);
1654         if (!list_empty(&obd->obd_req_replay_queue) ||
1655             !list_empty(&obd->obd_lock_replay_queue) ||
1656             !list_empty(&obd->obd_final_req_queue)) {
1657                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1658                        obd->obd_name,
1659                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1660                        list_empty(&obd->obd_lock_replay_queue) ? \
1661                                   "" : "lock ",
1662                        list_empty(&obd->obd_final_req_queue) ? \
1663                                   "" : "final ");
1664                 spin_unlock(&obd->obd_recovery_task_lock);
1665                 LBUG();
1666         }
1667         spin_unlock(&obd->obd_recovery_task_lock);
1668
1669         obd->obd_recovery_end = ktime_get_real_seconds();
1670
1671         /* When recovery finished, cleanup orphans on MDS and OST. */
1672         if (obd->obd_type && OBP(obd, postrecov)) {
1673                 int rc = OBP(obd, postrecov)(obd);
1674
1675                 if (rc < 0)
1676                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1677                                       obd->obd_name, rc);
1678         }
1679         EXIT;
1680 }
1681
1682 static void abort_req_replay_queue(struct obd_device *obd)
1683 {
1684         struct ptlrpc_request *req, *n;
1685         struct list_head abort_list;
1686
1687         INIT_LIST_HEAD(&abort_list);
1688         spin_lock(&obd->obd_recovery_task_lock);
1689         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1690         spin_unlock(&obd->obd_recovery_task_lock);
1691         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1692                 DEBUG_REQ(D_WARNING, req, "aborted:");
1693                 req->rq_status = -ENOTCONN;
1694                 if (ptlrpc_error(req)) {
1695                         DEBUG_REQ(D_ERROR, req,
1696                                   "failed abort_req_reply; skipping");
1697                 }
1698                 target_exp_dequeue_req_replay(req);
1699                 target_request_copy_put(req);
1700         }
1701 }
1702
1703 static void abort_lock_replay_queue(struct obd_device *obd)
1704 {
1705         struct ptlrpc_request *req, *n;
1706         struct list_head abort_list;
1707
1708         INIT_LIST_HEAD(&abort_list);
1709         spin_lock(&obd->obd_recovery_task_lock);
1710         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1711         spin_unlock(&obd->obd_recovery_task_lock);
1712         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1713                 DEBUG_REQ(D_ERROR, req, "aborted:");
1714                 req->rq_status = -ENOTCONN;
1715                 if (ptlrpc_error(req)) {
1716                         DEBUG_REQ(D_ERROR, req,
1717                                   "failed abort_lock_reply; skipping");
1718                 }
1719                 target_request_copy_put(req);
1720         }
1721 }
1722
1723 /*
1724  * Called from a cleanup function if the device is being cleaned up
1725  * forcefully.  The exports should all have been disconnected already,
1726  * the only thing left to do is
1727  * - clear the recovery flags
1728  * - cancel the timer
1729  * - free queued requests and replies, but don't send replies
1730  * Because the obd_stopping flag is set, no new requests should be received.
1731  */
1732 void target_cleanup_recovery(struct obd_device *obd)
1733 {
1734         struct ptlrpc_request *req, *n;
1735         struct list_head clean_list;
1736
1737         INIT_LIST_HEAD(&clean_list);
1738         spin_lock(&obd->obd_dev_lock);
1739         if (!obd->obd_recovering) {
1740                 spin_unlock(&obd->obd_dev_lock);
1741                 EXIT;
1742                 return;
1743         }
1744         obd->obd_recovering = obd->obd_abort_recovery = 0;
1745         spin_unlock(&obd->obd_dev_lock);
1746
1747         spin_lock(&obd->obd_recovery_task_lock);
1748         target_cancel_recovery_timer(obd);
1749         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1750         spin_unlock(&obd->obd_recovery_task_lock);
1751
1752         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1753                 LASSERT(req->rq_reply_state == NULL);
1754                 target_exp_dequeue_req_replay(req);
1755                 target_request_copy_put(req);
1756         }
1757
1758         spin_lock(&obd->obd_recovery_task_lock);
1759         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1760         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1761         spin_unlock(&obd->obd_recovery_task_lock);
1762
1763         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1764                 LASSERT(req->rq_reply_state == NULL);
1765                 target_request_copy_put(req);
1766         }
1767
1768         EXIT;
1769 }
1770 EXPORT_SYMBOL(target_cleanup_recovery);
1771
1772 /* obd_recovery_task_lock should be held */
1773 void target_cancel_recovery_timer(struct obd_device *obd)
1774 {
1775         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1776         hrtimer_cancel(&obd->obd_recovery_timer);
1777 }
1778
1779 static void target_start_recovery_timer(struct obd_device *obd)
1780 {
1781         ktime_t delay;
1782
1783         if (obd->obd_recovery_start != 0)
1784                 return;
1785
1786         spin_lock(&obd->obd_dev_lock);
1787         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1788                 spin_unlock(&obd->obd_dev_lock);
1789                 return;
1790         }
1791
1792         LASSERT(obd->obd_recovery_timeout != 0);
1793
1794         if (obd->obd_recovery_start != 0) {
1795                 spin_unlock(&obd->obd_dev_lock);
1796                 return;
1797         }
1798
1799         delay = ktime_set(obd->obd_recovery_timeout, 0);
1800         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_REL);
1801         obd->obd_recovery_start = ktime_get_real_seconds();
1802         spin_unlock(&obd->obd_dev_lock);
1803
1804         LCONSOLE_WARN("%s: Will be in recovery for at least %lu:%02lu, or until %d client%s reconnect%s\n",
1805                       obd->obd_name,
1806                       obd->obd_recovery_timeout / 60,
1807                       obd->obd_recovery_timeout % 60,
1808                       obd->obd_max_recoverable_clients,
1809                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1810                       (obd->obd_max_recoverable_clients == 1) ? "s" : "");
1811 }
1812
1813 /**
1814  * extend recovery window.
1815  *
1816  * if @extend is true, extend recovery window to have @dr_timeout remaining
1817  * at least; otherwise, make sure the recovery timeout value is not less
1818  * than @dr_timeout.
1819  */
1820 static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout,
1821                                   bool extend)
1822 {
1823         ktime_t left_ns;
1824         time_t timeout;
1825         time_t left;
1826
1827         spin_lock(&obd->obd_dev_lock);
1828         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1829                 spin_unlock(&obd->obd_dev_lock);
1830                 return;
1831         }
1832         LASSERT(obd->obd_recovery_start != 0);
1833
1834         left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1835         left = ktime_divns(left_ns, NSEC_PER_SEC);
1836
1837         if (extend) {
1838                 timeout = obd->obd_recovery_timeout;
1839                 /* dr_timeout will happen after the hrtimer has expired.
1840                  * Add the excess time to the soft recovery timeout without
1841                  * exceeding the hard recovery timeout.
1842                  */
1843                 if (dr_timeout > left) {
1844                         timeout += dr_timeout - left;
1845                         timeout = min_t(time_t, obd->obd_recovery_time_hard,
1846                                         timeout);
1847                 }
1848         } else {
1849                 timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout,
1850                                   obd->obd_recovery_time_hard);
1851         }
1852
1853         if (timeout == obd->obd_recovery_time_hard)
1854                 CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n",
1855                       obd->obd_name, timeout, extend);
1856
1857         if (obd->obd_recovery_timeout < timeout) {
1858                 ktime_t now = ktime_get_real();
1859                 ktime_t end;
1860
1861                 obd->obd_recovery_timeout = timeout;
1862                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1863                 left_ns = ktime_sub(end, now);
1864                 hrtimer_forward_now(&obd->obd_recovery_timer, left_ns);
1865                 left = ktime_divns(left_ns, NSEC_PER_MSEC);
1866         }
1867         spin_unlock(&obd->obd_dev_lock);
1868
1869         CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1870                 obd->obd_name, left);
1871 }
1872
1873 /* Reset the timer with each new client connection */
1874 /*
1875  * This timer is actually reconnect_timer, which is for making sure
1876  * the total recovery window is at least as big as my reconnect
1877  * attempt timing. So the initial recovery time_out will be set to
1878  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1879  * from client is bigger than this, then the recovery time_out will
1880  * be extended to make sure the client could be reconnected, in the
1881  * process, the timeout from the new client should be ignored.
1882  */
1883 static void
1884 check_and_start_recovery_timer(struct obd_device *obd,
1885                                struct ptlrpc_request *req,
1886                                int new_client)
1887 {
1888         time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1889         struct obd_device_target *obt = &obd->u.obt;
1890
1891         if (!new_client && service_time)
1892                 /*
1893                  * Teach server about old server's estimates, as first guess
1894                  * at how long new requests will take.
1895                  */
1896                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1897                             service_time);
1898
1899         target_start_recovery_timer(obd);
1900
1901         /*
1902          * Convert the service time to RPC timeout,
1903          * and reuse service_time to limit stack usage.
1904          */
1905         service_time = at_est2timeout(service_time);
1906
1907         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1908             service_time < at_extra)
1909                 service_time = at_extra;
1910
1911         /*
1912          * We expect other clients to timeout within service_time, then try
1913          * to reconnect, then try the failover server.  The max delay between
1914          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1915          */
1916         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1917
1918         LASSERT(obt->obt_magic == OBT_MAGIC);
1919         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1920         if (service_time > obd->obd_recovery_timeout && !new_client)
1921                 extend_recovery_timer(obd, service_time, false);
1922 }
1923
1924 /** Health checking routines */
1925 static inline int exp_connect_healthy(struct obd_export *exp)
1926 {
1927         return exp->exp_in_recovery;
1928 }
1929
1930 /** if export done req_replay or has replay in queue */
1931 static inline int exp_req_replay_healthy(struct obd_export *exp)
1932 {
1933         return (!exp->exp_req_replay_needed ||
1934                 atomic_read(&exp->exp_replay_count) > 0);
1935 }
1936
1937
1938 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1939 {
1940         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1941                exp_req_replay_healthy(exp);
1942 }
1943
1944 /** if export done lock_replay or has replay in queue */
1945 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1946 {
1947         return (!exp->exp_lock_replay_needed ||
1948                 atomic_read(&exp->exp_replay_count) > 0);
1949 }
1950
1951 static inline int exp_vbr_healthy(struct obd_export *exp)
1952 {
1953         return !exp->exp_vbr_failed;
1954 }
1955
1956 static inline int exp_finished(struct obd_export *exp)
1957 {
1958         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1959 }
1960
1961 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1962 {
1963         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1964                 exp_finished(exp);
1965 }
1966
1967 static int check_for_next_transno(struct lu_target *lut)
1968 {
1969         struct ptlrpc_request *req = NULL;
1970         struct obd_device *obd = lut->lut_obd;
1971         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1972         int wake_up = 0, connected, completed, queue_len;
1973         __u64 req_transno = 0;
1974         __u64 update_transno = 0;
1975         __u64 next_transno = 0;
1976
1977         ENTRY;
1978
1979         spin_lock(&obd->obd_recovery_task_lock);
1980         if (!list_empty(&obd->obd_req_replay_queue)) {
1981                 req = list_entry(obd->obd_req_replay_queue.next,
1982                                      struct ptlrpc_request, rq_list);
1983                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1984         }
1985
1986         if (tdtd != NULL)
1987                 update_transno = distribute_txn_get_next_transno(tdtd);
1988
1989         connected = atomic_read(&obd->obd_connected_clients);
1990         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1991         queue_len = obd->obd_requests_queued_for_recovery;
1992         next_transno = obd->obd_next_recovery_transno;
1993
1994         CDEBUG(D_HA,
1995                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
1996                obd->obd_max_recoverable_clients, connected, completed,
1997                queue_len, req_transno, next_transno);
1998
1999         if (obd->obd_abort_recovery) {
2000                 CDEBUG(D_HA, "waking for aborted recovery\n");
2001                 wake_up = 1;
2002         } else if (obd->obd_recovery_expired) {
2003                 CDEBUG(D_HA, "waking for expired recovery\n");
2004                 wake_up = 1;
2005         } else if (tdtd != NULL && req != NULL &&
2006                    is_req_replayed_by_update(req)) {
2007                 LASSERTF(req_transno < next_transno,
2008                          "req_transno %llu next_transno%llu\n", req_transno,
2009                          next_transno);
2010                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2011                        req_transno);
2012                 wake_up = 1;
2013         } else if (req_transno == next_transno ||
2014                    (update_transno != 0 && update_transno <= next_transno)) {
2015                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2016                 wake_up = 1;
2017         } else if (queue_len > 0 &&
2018                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2019                 /** handle gaps occured due to lost reply or VBR */
2020                 LASSERTF(req_transno >= next_transno,
2021                          "req_transno: %llu, next_transno: %llu\n",
2022                          req_transno, next_transno);
2023                 CDEBUG(D_HA,
2024                        "%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",
2025                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2026                        next_transno, queue_len, completed, connected,
2027                        req_transno, update_transno, obd->obd_last_committed);
2028                 obd->obd_next_recovery_transno = req_transno;
2029                 wake_up = 1;
2030         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2031                 CDEBUG(D_HA, "waking for completed recovery\n");
2032                 wake_up = 1;
2033         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2034                 CDEBUG(D_HA,
2035                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2036                        next_transno);
2037                 obd->obd_next_recovery_transno = req_transno;
2038                 wake_up = 1;
2039         }
2040         spin_unlock(&obd->obd_recovery_task_lock);
2041         return wake_up;
2042 }
2043
2044 static int check_for_next_lock(struct lu_target *lut)
2045 {
2046         struct obd_device *obd = lut->lut_obd;
2047         int wake_up = 0;
2048
2049         spin_lock(&obd->obd_recovery_task_lock);
2050         if (!list_empty(&obd->obd_lock_replay_queue)) {
2051                 CDEBUG(D_HA, "waking for next lock\n");
2052                 wake_up = 1;
2053         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2054                 CDEBUG(D_HA, "waking for completed lock replay\n");
2055                 wake_up = 1;
2056         } else if (obd->obd_abort_recovery) {
2057                 CDEBUG(D_HA, "waking for aborted recovery\n");
2058                 wake_up = 1;
2059         } else if (obd->obd_recovery_expired) {
2060                 CDEBUG(D_HA, "waking for expired recovery\n");
2061                 wake_up = 1;
2062         }
2063         spin_unlock(&obd->obd_recovery_task_lock);
2064
2065         return wake_up;
2066 }
2067
2068 /**
2069  * wait for recovery events,
2070  * check its status with help of check_routine
2071  * evict dead clients via health_check
2072  */
2073 static int target_recovery_overseer(struct lu_target *lut,
2074                                     int (*check_routine)(struct lu_target *),
2075                                     int (*health_check)(struct obd_export *))
2076 {
2077         struct obd_device *obd = lut->lut_obd;
2078         struct target_distribute_txn_data *tdtd;
2079         time64_t last = 0;
2080         time64_t now;
2081 repeat:
2082         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2083                 now = ktime_get_seconds();
2084                 if (now - last > 600) {
2085                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2086                                       obd->obd_name);
2087                         last = now;
2088                 }
2089         }
2090         if (obd->obd_recovery_start != 0 && ktime_get_real_seconds() >=
2091               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2092                 __u64 next_update_transno = 0;
2093
2094                 /*
2095                  * Only abort the recovery if there are no update recovery
2096                  * left in the queue
2097                  */
2098                 spin_lock(&obd->obd_recovery_task_lock);
2099                 if (lut->lut_tdtd != NULL) {
2100                         next_update_transno =
2101                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2102
2103                         tdtd = lut->lut_tdtd;
2104                         /*
2105                          * If next_update_transno == 0, it probably because
2106                          * updatelog retrieve threads did not get any records
2107                          * yet, let's wait those threads stopped
2108                          */
2109                         if (next_update_transno == 0) {
2110                                 struct l_wait_info lwi = { 0 };
2111
2112                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2113                                        atomic_read(
2114                                        &tdtd->tdtd_recovery_threads_count) == 0,
2115                                        &lwi);
2116
2117                                 next_update_transno =
2118                                         distribute_txn_get_next_transno(
2119                                                                 lut->lut_tdtd);
2120                         }
2121                 }
2122
2123                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2124                         obd->obd_next_recovery_transno = next_update_transno;
2125                         spin_unlock(&obd->obd_recovery_task_lock);
2126                         /*
2127                          * Disconnect unfinished exports from clients, and
2128                          * keep connection from MDT to make sure the update
2129                          * recovery will still keep trying until some one
2130                          * manually abort the recovery
2131                          */
2132                         class_disconnect_stale_exports(obd,
2133                                                 exp_finished_or_from_mdt);
2134                         /* Abort all of replay & replay lock req from clients */
2135                         abort_req_replay_queue(obd);
2136                         abort_lock_replay_queue(obd);
2137                         CDEBUG(D_HA,
2138                                "%s: there are still update replay (%#llx)in the queue.\n",
2139                                obd->obd_name, next_update_transno);
2140                 } else {
2141                         obd->obd_abort_recovery = 1;
2142                         spin_unlock(&obd->obd_recovery_task_lock);
2143                         CWARN("%s recovery is aborted by hard timeout\n",
2144                               obd->obd_name);
2145                 }
2146         }
2147
2148         while (wait_event_timeout(obd->obd_next_transno_waitq,
2149                                   check_routine(lut),
2150                                   cfs_time_seconds(60)) == 0)
2151                 ; /* wait indefinitely for event, but don't trigger watchdog */
2152
2153         if (obd->obd_abort_recovery) {
2154                 CWARN("recovery is aborted, evict exports in recovery\n");
2155                 if (lut->lut_tdtd != NULL) {
2156                         struct l_wait_info lwi = { 0 };
2157
2158                         tdtd = lut->lut_tdtd;
2159                         /*
2160                          * Let's wait all of the update log recovery thread
2161                          * finished
2162                          */
2163                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2164                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2165                              &lwi);
2166                         /* Then abort the update recovery list */
2167                         dtrq_list_destroy(lut->lut_tdtd);
2168                 }
2169
2170                 /** evict exports which didn't finish recovery yet */
2171                 class_disconnect_stale_exports(obd, exp_finished);
2172                 return 1;
2173         } else if (obd->obd_recovery_expired) {
2174                 obd->obd_recovery_expired = 0;
2175                 /** If some clients died being recovered, evict them */
2176                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2177                               obd->obd_name);
2178                 /** evict cexports with no replay in queue, they are stalled */
2179                 class_disconnect_stale_exports(obd, health_check);
2180
2181                 /** continue with VBR */
2182                 spin_lock(&obd->obd_dev_lock);
2183                 obd->obd_version_recov = 1;
2184                 spin_unlock(&obd->obd_dev_lock);
2185                 /**
2186                  * reset timer, recovery will proceed with versions now,
2187                  * timeout is set just to handle reconnection delays
2188                  */
2189                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2190                 /**
2191                  * Wait for recovery events again, after evicting bad clients
2192                  */
2193                 goto repeat;
2194         }
2195         return 0;
2196 }
2197
2198 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2199 {
2200         struct obd_device *obd = lut->lut_obd;
2201         struct ptlrpc_request *req = NULL;
2202
2203         CDEBUG(D_HA, "Waiting for lock\n");
2204         if (target_recovery_overseer(lut, check_for_next_lock,
2205                                      exp_lock_replay_healthy))
2206                 abort_lock_replay_queue(obd);
2207
2208         spin_lock(&obd->obd_recovery_task_lock);
2209         if (!list_empty(&obd->obd_lock_replay_queue)) {
2210                 req = list_entry(obd->obd_lock_replay_queue.next,
2211                                      struct ptlrpc_request, rq_list);
2212                 list_del_init(&req->rq_list);
2213                 spin_unlock(&obd->obd_recovery_task_lock);
2214         } else {
2215                 spin_unlock(&obd->obd_recovery_task_lock);
2216                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2217                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2218                 /** evict exports failed VBR */
2219                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2220         }
2221         return req;
2222 }
2223
2224 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2225 {
2226         struct ptlrpc_request *req = NULL;
2227
2228         spin_lock(&obd->obd_recovery_task_lock);
2229         if (!list_empty(&obd->obd_final_req_queue)) {
2230                 req = list_entry(obd->obd_final_req_queue.next,
2231                                      struct ptlrpc_request, rq_list);
2232                 list_del_init(&req->rq_list);
2233                 spin_unlock(&obd->obd_recovery_task_lock);
2234                 if (req->rq_export->exp_in_recovery) {
2235                         spin_lock(&req->rq_export->exp_lock);
2236                         req->rq_export->exp_in_recovery = 0;
2237                         spin_unlock(&req->rq_export->exp_lock);
2238                 }
2239         } else {
2240                 spin_unlock(&obd->obd_recovery_task_lock);
2241         }
2242         return req;
2243 }
2244
2245 static void handle_recovery_req(struct ptlrpc_thread *thread,
2246                                 struct ptlrpc_request *req,
2247                                 svc_handler_t handler)
2248 {
2249         ENTRY;
2250
2251         /**
2252          * export can be evicted during recovery, no need to handle replays for
2253          * it after that, discard such request silently
2254          */
2255         if (req->rq_export->exp_disconnected)
2256                 RETURN_EXIT;
2257
2258         req->rq_session.lc_thread = thread;
2259         req->rq_svc_thread = thread;
2260         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2261
2262         /* thread context */
2263         lu_context_enter(&thread->t_env->le_ctx);
2264         (void)handler(req);
2265         lu_context_exit(&thread->t_env->le_ctx);
2266
2267         /* don't reset timer for final stage */
2268         if (!exp_finished(req->rq_export)) {
2269                 time_t to = obd_timeout;
2270
2271                 /**
2272                  * Add request timeout to the recovery time so next request from
2273                  * this client may come in recovery time
2274                  */
2275                 if (!AT_OFF) {
2276                         struct ptlrpc_service_part *svcpt;
2277
2278                         svcpt = req->rq_rqbd->rqbd_svcpt;
2279                         /*
2280                          * If the server sent early reply for this request,
2281                          * the client will recalculate the timeout according to
2282                          * current server estimate service time, so we will
2283                          * use the maxium timeout here for waiting the client
2284                          * sending the next req
2285                          */
2286                         to = max_t(time_t,
2287                                    at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2288                                    lustre_msg_get_timeout(req->rq_reqmsg));
2289                         /*
2290                          * Add 2 net_latency, one for balance rq_deadline
2291                          * (see ptl_send_rpc), one for resend the req to server,
2292                          * Note: client will pack net_latency in replay req
2293                          * (see ptlrpc_replay_req)
2294                          */
2295                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2296                 }
2297                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2298         }
2299         EXIT;
2300 }
2301
2302 /** Checking routines for recovery */
2303 static int check_for_recovery_ready(struct lu_target *lut)
2304 {
2305         struct obd_device *obd = lut->lut_obd;
2306         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2307
2308         CDEBUG(D_HA,
2309                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2310                clnts, obd->obd_stale_clients, obd->obd_max_recoverable_clients,
2311                obd->obd_abort_recovery, obd->obd_recovery_expired);
2312
2313         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2314                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2315                 if (clnts + obd->obd_stale_clients <
2316                     obd->obd_max_recoverable_clients)
2317                         return 0;
2318         }
2319
2320         if (lut->lut_tdtd != NULL) {
2321                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2322                     !obd->obd_abort_recovery) {
2323                         /*
2324                          * Let's extend recovery timer, in case the recovery
2325                          * timer expired, and some clients got evicted
2326                          */
2327                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2328                                               true);
2329                         CDEBUG(D_HA,
2330                                "%s update recovery is not ready, extend recovery %lu\n",
2331                                obd->obd_name, obd->obd_recovery_timeout);
2332                         return 0;
2333                 }
2334         }
2335
2336         return 1;
2337 }
2338
2339 enum {
2340         REQUEST_RECOVERY = 1,
2341         UPDATE_RECOVERY = 2,
2342 };
2343
2344 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2345 {
2346         __u64 transno = 0;
2347
2348         if (!list_empty(&obd->obd_req_replay_queue)) {
2349                 struct ptlrpc_request *req;
2350
2351                 req = list_entry(obd->obd_req_replay_queue.next,
2352                                  struct ptlrpc_request, rq_list);
2353                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2354         }
2355
2356         return transno;
2357 }
2358
2359 static __u64 get_next_transno(struct lu_target *lut, int *type)
2360 {
2361         struct obd_device *obd = lut->lut_obd;
2362         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2363         __u64 transno = 0;
2364         __u64 update_transno;
2365
2366         ENTRY;
2367
2368         transno = get_next_replay_req_transno(obd);
2369         if (type != NULL)
2370                 *type = REQUEST_RECOVERY;
2371
2372         if (tdtd == NULL)
2373                 RETURN(transno);
2374
2375         update_transno = distribute_txn_get_next_transno(tdtd);
2376         if (transno == 0 || (transno >= update_transno &&
2377                              update_transno != 0)) {
2378                 transno = update_transno;
2379                 if (type != NULL)
2380                         *type = UPDATE_RECOVERY;
2381         }
2382
2383         RETURN(transno);
2384 }
2385
2386 /**
2387  * drop duplicate replay request
2388  *
2389  * Because the operation has been replayed by update recovery, the request
2390  * with the same transno will be dropped and also notify the client to send
2391  * next replay request.
2392  *
2393  * \param[in] env       execution environment
2394  * \param[in] obd       failover obd device
2395  * \param[in] req       request to be dropped
2396  */
2397 static void drop_duplicate_replay_req(struct lu_env *env,
2398                                       struct obd_device *obd,
2399                                       struct ptlrpc_request *req)
2400 {
2401         DEBUG_REQ(D_HA, req,
2402                   "remove t%lld from %s because of duplicate update records are found.\n",
2403                   lustre_msg_get_transno(req->rq_reqmsg),
2404                   libcfs_nid2str(req->rq_peer.nid));
2405
2406         /*
2407          * Right now, only for MDS reint operation update replay and
2408          * normal request replay can have the same transno
2409          */
2410         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2411                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2412                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2413                 if (likely(req->rq_export))
2414                         target_committed_to_req(req);
2415                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2416                 target_send_reply(req, req->rq_status, 0);
2417         } else {
2418                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s\n",
2419                 libcfs_nid2str(req->rq_peer.nid));
2420         }
2421         target_exp_dequeue_req_replay(req);
2422         target_request_copy_put(req);
2423         obd->obd_replayed_requests++;
2424 }
2425
2426 static void replay_request_or_update(struct lu_env *env,
2427                                      struct lu_target *lut,
2428                                      struct target_recovery_data *trd,
2429                                      struct ptlrpc_thread *thread)
2430 {
2431         struct obd_device *obd = lut->lut_obd;
2432         struct ptlrpc_request *req = NULL;
2433         int type;
2434         __u64 transno;
2435
2436         ENTRY;
2437
2438         CDEBUG(D_HA, "Waiting for transno %lld\n",
2439                obd->obd_next_recovery_transno);
2440
2441         /* Replay all of request and update by transno */
2442         do {
2443                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2444
2445                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2446
2447                 /**
2448                  * It is needed to extend recovery window above
2449                  *  recovery_time_soft. Extending is possible only in the
2450                  *  end of recovery window (see more details in
2451                  *  handle_recovery_req()).
2452                  */
2453                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2454
2455                 if (target_recovery_overseer(lut, check_for_next_transno,
2456                                         exp_req_replay_healthy_or_from_mdt)) {
2457                         abort_req_replay_queue(obd);
2458                         abort_lock_replay_queue(obd);
2459                         goto abort;
2460                 }
2461
2462                 spin_lock(&obd->obd_recovery_task_lock);
2463                 transno = get_next_transno(lut, &type);
2464                 if (type == REQUEST_RECOVERY && transno != 0) {
2465                         /*
2466                          * Drop replay request from client side, if the
2467                          * replay has been executed by update with the
2468                          * same transno
2469                          */
2470                         req = list_entry(obd->obd_req_replay_queue.next,
2471                                         struct ptlrpc_request, rq_list);
2472
2473                         list_del_init(&req->rq_list);
2474                         obd->obd_requests_queued_for_recovery--;
2475                         spin_unlock(&obd->obd_recovery_task_lock);
2476
2477                         /*
2478                          * Let's check if the request has been redone by
2479                          * update replay
2480                          */
2481                         if (is_req_replayed_by_update(req)) {
2482                                 struct distribute_txn_replay_req *dtrq;
2483
2484                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2485                                                                       transno);
2486                                 LASSERT(dtrq != NULL);
2487                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2488                                 list_del_init(&dtrq->dtrq_list);
2489                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2490                                 dtrq_destroy(dtrq);
2491
2492                                 drop_duplicate_replay_req(env, obd, req);
2493
2494                                 continue;
2495                         }
2496
2497                         LASSERT(trd->trd_processing_task == current_pid());
2498                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2499                                   req->rq_xid,
2500                                   lustre_msg_get_transno(req->rq_reqmsg),
2501                                   libcfs_nid2str(req->rq_peer.nid));
2502
2503                         handle_recovery_req(thread, req,
2504                                             trd->trd_recovery_handler);
2505                         /**
2506                          * bz18031: increase next_recovery_transno before
2507                          * target_request_copy_put() will drop exp_rpc reference
2508                          */
2509                         spin_lock(&obd->obd_recovery_task_lock);
2510                         obd->obd_next_recovery_transno++;
2511                         spin_unlock(&obd->obd_recovery_task_lock);
2512                         target_exp_dequeue_req_replay(req);
2513                         target_request_copy_put(req);
2514                         obd->obd_replayed_requests++;
2515                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2516                         struct distribute_txn_replay_req *dtrq;
2517                         int rc;
2518
2519                         spin_unlock(&obd->obd_recovery_task_lock);
2520
2521                         LASSERT(tdtd != NULL);
2522                         dtrq = distribute_txn_get_next_req(tdtd);
2523                         lu_context_enter(&thread->t_env->le_ctx);
2524                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2525                         lu_context_exit(&thread->t_env->le_ctx);
2526                         extend_recovery_timer(obd, obd_timeout, true);
2527
2528                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2529                                 CDEBUG(D_HA,
2530                                        "Move x%llu t%llu to finish list\n",
2531                                        dtrq->dtrq_xid,
2532                                        dtrq->dtrq_master_transno);
2533
2534                                 /* Add it to the replay finish list */
2535                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2536                                 list_add(&dtrq->dtrq_list,
2537                                          &tdtd->tdtd_replay_finish_list);
2538                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2539
2540                                 spin_lock(&obd->obd_recovery_task_lock);
2541                                 if (transno == obd->obd_next_recovery_transno)
2542                                         obd->obd_next_recovery_transno++;
2543                                 else if (transno >
2544                                          obd->obd_next_recovery_transno)
2545                                         obd->obd_next_recovery_transno =
2546                                                                 transno + 1;
2547                                 spin_unlock(&obd->obd_recovery_task_lock);
2548                         } else {
2549                                 dtrq_destroy(dtrq);
2550                         }
2551                 } else {
2552                         spin_unlock(&obd->obd_recovery_task_lock);
2553 abort:
2554                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2555                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2556                         /** evict exports failed VBR */
2557                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2558                         break;
2559                 }
2560         } while (1);
2561 }
2562
2563 static int target_recovery_thread(void *arg)
2564 {
2565         struct lu_target *lut = arg;
2566         struct obd_device *obd = lut->lut_obd;
2567         struct ptlrpc_request *req;
2568         struct target_recovery_data *trd = &obd->obd_recovery_data;
2569         unsigned long delta;
2570         struct lu_env *env;
2571         struct ptlrpc_thread *thread = NULL;
2572         int rc = 0;
2573
2574         ENTRY;
2575
2576         unshare_fs_struct();
2577         OBD_ALLOC_PTR(thread);
2578         if (thread == NULL)
2579                 RETURN(-ENOMEM);
2580
2581         OBD_ALLOC_PTR(env);
2582         if (env == NULL)
2583                 GOTO(out_thread, rc = -ENOMEM);
2584         rc = lu_env_add(env);
2585         if (rc)
2586                 GOTO(out_env, rc);
2587
2588         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2589         if (rc)
2590                 GOTO(out_env_remove, rc);
2591
2592         thread->t_env = env;
2593         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2594         env->le_ctx.lc_thread = thread;
2595         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2596
2597         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2598                current_pid());
2599         trd->trd_processing_task = current_pid();
2600
2601         spin_lock(&obd->obd_dev_lock);
2602         obd->obd_recovering = 1;
2603         spin_unlock(&obd->obd_dev_lock);
2604         complete(&trd->trd_starting);
2605
2606         /* first of all, we have to know the first transno to replay */
2607         if (target_recovery_overseer(lut, check_for_recovery_ready,
2608                                      exp_connect_healthy)) {
2609                 abort_req_replay_queue(obd);
2610                 abort_lock_replay_queue(obd);
2611                 if (lut->lut_tdtd != NULL)
2612                         dtrq_list_destroy(lut->lut_tdtd);
2613         }
2614
2615         /* next stage: replay requests or update */
2616         delta = jiffies;
2617         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2618                atomic_read(&obd->obd_req_replay_clients),
2619                obd->obd_next_recovery_transno);
2620         replay_request_or_update(env, lut, trd, thread);
2621
2622         /**
2623          * The second stage: replay locks
2624          */
2625         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2626                atomic_read(&obd->obd_lock_replay_clients));
2627         while ((req = target_next_replay_lock(lut))) {
2628                 LASSERT(trd->trd_processing_task == current_pid());
2629                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2630                           libcfs_nid2str(req->rq_peer.nid));
2631                 handle_recovery_req(thread, req,
2632                                     trd->trd_recovery_handler);
2633                 target_request_copy_put(req);
2634                 obd->obd_replayed_locks++;
2635         }
2636
2637         /**
2638          * The third stage: reply on final pings, at this moment all clients
2639          * must have request in final queue
2640          */
2641         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2642         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2643         /** Update server last boot epoch */
2644         tgt_boot_epoch_update(lut);
2645         /*
2646          * We drop recoverying flag to forward all new requests
2647          * to regular mds_handle() since now
2648          */
2649         spin_lock(&obd->obd_dev_lock);
2650         obd->obd_recovering = obd->obd_abort_recovery = 0;
2651         spin_unlock(&obd->obd_dev_lock);
2652         spin_lock(&obd->obd_recovery_task_lock);
2653         target_cancel_recovery_timer(obd);
2654         spin_unlock(&obd->obd_recovery_task_lock);
2655         while ((req = target_next_final_ping(obd))) {
2656                 LASSERT(trd->trd_processing_task == current_pid());
2657                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2658                           libcfs_nid2str(req->rq_peer.nid));
2659                 handle_recovery_req(thread, req,
2660                                     trd->trd_recovery_handler);
2661                 /*
2662                  * Because the waiting client can not send ping to server,
2663                  * so we need refresh the last_request_time, to avoid the
2664                  * export is being evicted
2665                  */
2666                 ptlrpc_update_export_timer(req->rq_export, 0);
2667                 target_request_copy_put(req);
2668         }
2669
2670         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2671         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2672                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2673         if (delta > OBD_RECOVERY_TIME_SOFT) {
2674                 CWARN("too long recovery - read logs\n");
2675                 libcfs_debug_dumplog();
2676         }
2677
2678         target_finish_recovery(lut);
2679         lu_context_fini(&env->le_ctx);
2680         trd->trd_processing_task = 0;
2681         complete_all(&trd->trd_finishing);
2682         tgt_io_thread_done(thread);
2683 out_env_remove:
2684         lu_env_remove(env);
2685 out_env:
2686         OBD_FREE_PTR(env);
2687 out_thread:
2688         OBD_FREE_PTR(thread);
2689         RETURN(rc);
2690 }
2691
2692 static int target_start_recovery_thread(struct lu_target *lut,
2693                                         svc_handler_t handler)
2694 {
2695         struct obd_device *obd = lut->lut_obd;
2696         int rc = 0;
2697         struct target_recovery_data *trd = &obd->obd_recovery_data;
2698         int index;
2699
2700         memset(trd, 0, sizeof(*trd));
2701         init_completion(&trd->trd_starting);
2702         init_completion(&trd->trd_finishing);
2703         trd->trd_recovery_handler = handler;
2704
2705         rc = server_name2index(obd->obd_name, &index, NULL);
2706         if (rc < 0)
2707                 return rc;
2708
2709         if (!IS_ERR(kthread_run(target_recovery_thread,
2710                                 lut, "tgt_recover_%d", index))) {
2711                 wait_for_completion(&trd->trd_starting);
2712                 LASSERT(obd->obd_recovering != 0);
2713         } else {
2714                 rc = -ECHILD;
2715         }
2716
2717         return rc;
2718 }
2719
2720 void target_stop_recovery_thread(struct obd_device *obd)
2721 {
2722         if (obd->obd_recovery_data.trd_processing_task > 0) {
2723                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2724                 /** recovery can be done but postrecovery is not yet */
2725                 spin_lock(&obd->obd_dev_lock);
2726                 if (obd->obd_recovering) {
2727                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2728                         obd->obd_abort_recovery = 1;
2729                         wake_up(&obd->obd_next_transno_waitq);
2730                 }
2731                 spin_unlock(&obd->obd_dev_lock);
2732                 wait_for_completion(&trd->trd_finishing);
2733         }
2734 }
2735 EXPORT_SYMBOL(target_stop_recovery_thread);
2736
2737 void target_recovery_fini(struct obd_device *obd)
2738 {
2739         class_disconnect_exports(obd);
2740         target_stop_recovery_thread(obd);
2741         target_cleanup_recovery(obd);
2742 }
2743 EXPORT_SYMBOL(target_recovery_fini);
2744
2745 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2746 {
2747         struct obd_device *obd = container_of(timer, struct obd_device,
2748                                               obd_recovery_timer);
2749
2750         CDEBUG(D_HA,
2751                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2752                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2753                ktime_get_real_seconds() - obd->obd_recovery_start,
2754                atomic_read(&obd->obd_connected_clients));
2755
2756         obd->obd_recovery_expired = 1;
2757         wake_up(&obd->obd_next_transno_waitq);
2758         return HRTIMER_NORESTART;
2759 }
2760
2761 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2762 {
2763         struct obd_device *obd = lut->lut_obd;
2764
2765         if (lut->lut_bottom->dd_rdonly)
2766                 return;
2767
2768         if (obd->obd_max_recoverable_clients == 0) {
2769                 /** Update server last boot epoch */
2770                 tgt_boot_epoch_update(lut);
2771                 return;
2772         }
2773
2774         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2775                "last_transno %llu\n", obd->obd_name,
2776                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2777         LASSERT(obd->obd_stopping == 0);
2778         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2779         obd->obd_recovery_start = 0;
2780         obd->obd_recovery_end = 0;
2781
2782         hrtimer_init(&obd->obd_recovery_timer, CLOCK_REALTIME,
2783                      HRTIMER_MODE_REL);
2784         obd->obd_recovery_timer.function = &target_recovery_expired;
2785         target_start_recovery_thread(lut, handler);
2786 }
2787 EXPORT_SYMBOL(target_recovery_init);
2788
2789 static int target_process_req_flags(struct obd_device *obd,
2790                                     struct ptlrpc_request *req)
2791 {
2792         struct obd_export *exp = req->rq_export;
2793
2794         LASSERT(exp != NULL);
2795         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2796                 /* client declares he's ready to replay locks */
2797                 spin_lock(&exp->exp_lock);
2798                 if (exp->exp_req_replay_needed) {
2799                         exp->exp_req_replay_needed = 0;
2800                         spin_unlock(&exp->exp_lock);
2801
2802                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2803                         atomic_dec(&obd->obd_req_replay_clients);
2804                 } else {
2805                         spin_unlock(&exp->exp_lock);
2806                 }
2807         }
2808         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2809                 /*
2810                  * client declares he's ready to complete recovery
2811                  * so, we put the request on th final queue
2812                  */
2813                 spin_lock(&exp->exp_lock);
2814                 if (exp->exp_lock_replay_needed) {
2815                         exp->exp_lock_replay_needed = 0;
2816                         spin_unlock(&exp->exp_lock);
2817
2818                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2819                         atomic_dec(&obd->obd_lock_replay_clients);
2820                 } else {
2821                         spin_unlock(&exp->exp_lock);
2822                 }
2823         }
2824         return 0;
2825 }
2826
2827 int target_queue_recovery_request(struct ptlrpc_request *req,
2828                                   struct obd_device *obd)
2829 {
2830         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2831         struct ptlrpc_request *reqiter;
2832         int inserted = 0;
2833
2834         ENTRY;
2835
2836         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2837                 /* Processing the queue right now, don't re-add. */
2838                 RETURN(1);
2839         }
2840
2841         target_process_req_flags(obd, req);
2842
2843         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2844                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2845                         if (cfs_fail_val == 1) {
2846                                 cfs_race_state = 1;
2847                                 cfs_fail_val = 0;
2848                                 wake_up(&cfs_race_waitq);
2849
2850                                 set_current_state(TASK_INTERRUPTIBLE);
2851                                 schedule_timeout(cfs_time_seconds(1));
2852                         }
2853                 }
2854
2855                 /*
2856                  * client declares he's ready to complete recovery
2857                  * so, we put the request on th final queue
2858                  */
2859                 target_request_copy_get(req);
2860                 DEBUG_REQ(D_HA, req, "queue final req");
2861                 wake_up(&obd->obd_next_transno_waitq);
2862                 spin_lock(&obd->obd_recovery_task_lock);
2863                 if (obd->obd_recovering) {
2864                         struct ptlrpc_request *tmp;
2865                         struct ptlrpc_request *duplicate = NULL;
2866
2867                         if (likely(!req->rq_export->exp_replay_done)) {
2868                                 req->rq_export->exp_replay_done = 1;
2869                                 list_add_tail(&req->rq_list,
2870                                               &obd->obd_final_req_queue);
2871                                 spin_unlock(&obd->obd_recovery_task_lock);
2872                                 RETURN(0);
2873                         }
2874
2875                         /*
2876                          * XXX O(n), but only happens if final ping is
2877                          * timed out, probably reorganize the list as
2878                          * a hash list later
2879                          */
2880                         list_for_each_entry_safe(reqiter, tmp,
2881                                                  &obd->obd_final_req_queue,
2882                                                  rq_list) {
2883                                 if (reqiter->rq_export == req->rq_export) {
2884                                         list_del_init(&reqiter->rq_list);
2885                                         duplicate = reqiter;
2886                                         break;
2887                                 }
2888                         }
2889
2890                         list_add_tail(&req->rq_list,
2891                                       &obd->obd_final_req_queue);
2892                         req->rq_export->exp_replay_done = 1;
2893                         spin_unlock(&obd->obd_recovery_task_lock);
2894
2895                         if (duplicate != NULL) {
2896                                 DEBUG_REQ(D_HA, duplicate,
2897                                           "put prev final req\n");
2898                                 target_request_copy_put(duplicate);
2899                         }
2900                         RETURN(0);
2901                 } else {
2902                         spin_unlock(&obd->obd_recovery_task_lock);
2903                         target_request_copy_put(req);
2904                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2905                 }
2906         }
2907         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2908                 /* client declares he's ready to replay locks */
2909                 target_request_copy_get(req);
2910                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2911                 wake_up(&obd->obd_next_transno_waitq);
2912                 spin_lock(&obd->obd_recovery_task_lock);
2913                 LASSERT(obd->obd_recovering);
2914                 /* usually due to recovery abort */
2915                 if (!req->rq_export->exp_in_recovery) {
2916                         spin_unlock(&obd->obd_recovery_task_lock);
2917                         target_request_copy_put(req);
2918                         RETURN(-ENOTCONN);
2919                 }
2920                 LASSERT(req->rq_export->exp_lock_replay_needed);
2921                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2922                 spin_unlock(&obd->obd_recovery_task_lock);
2923                 RETURN(0);
2924         }
2925
2926         /*
2927          * CAVEAT EMPTOR: The incoming request message has been swabbed
2928          * (i.e. buflens etc are in my own byte order), but type-dependent
2929          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2930          */
2931
2932         if (!transno) {
2933                 INIT_LIST_HEAD(&req->rq_list);
2934                 DEBUG_REQ(D_HA, req, "not queueing");
2935                 RETURN(1);
2936         }
2937
2938         /*
2939          * If we're processing the queue, we want don't want to queue this
2940          * message.
2941          *
2942          * Also, if this request has a transno less than the one we're waiting
2943          * for, we should process it now.  It could (and currently always will)
2944          * be an open request for a descriptor that was opened some time ago.
2945          *
2946          * Also, a resent, replayed request that has already been
2947          * handled will pass through here and be processed immediately.
2948          */
2949         CDEBUG(D_HA,
2950                "Next recovery transno: %llu, current: %llu, replaying\n",
2951                obd->obd_next_recovery_transno, transno);
2952
2953         /*
2954          * If the request has been replayed by update replay, then sends this
2955          * request to the recovery thread (replay_request_or_update()), where
2956          * it will be handled
2957          */
2958         spin_lock(&obd->obd_recovery_task_lock);
2959         if (transno < obd->obd_next_recovery_transno &&
2960             !is_req_replayed_by_update(req)) {
2961                 /* Processing the queue right now, don't re-add. */
2962                 LASSERT(list_empty(&req->rq_list));
2963                 spin_unlock(&obd->obd_recovery_task_lock);
2964                 RETURN(1);
2965         }
2966         spin_unlock(&obd->obd_recovery_task_lock);
2967
2968         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2969                 RETURN(0);
2970
2971         target_request_copy_get(req);
2972         if (!req->rq_export->exp_in_recovery) {
2973                 target_request_copy_put(req);
2974                 RETURN(-ENOTCONN);
2975         }
2976         LASSERT(req->rq_export->exp_req_replay_needed);
2977
2978         if (target_exp_enqueue_req_replay(req)) {
2979                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2980                 target_request_copy_put(req);
2981                 RETURN(0);
2982         }
2983
2984         /* XXX O(n^2) */
2985         spin_lock(&obd->obd_recovery_task_lock);
2986         LASSERT(obd->obd_recovering);
2987         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2988                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2989                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2990                         inserted = 1;
2991                         goto added;
2992                 }
2993
2994                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2995                              transno)) {
2996                         DEBUG_REQ(D_ERROR, req,
2997                                   "dropping replay: transno has been claimed by another client");
2998                         spin_unlock(&obd->obd_recovery_task_lock);
2999                         target_exp_dequeue_req_replay(req);
3000                         target_request_copy_put(req);
3001                         RETURN(0);
3002                 }
3003         }
3004 added:
3005         if (!inserted)
3006                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3007
3008         obd->obd_requests_queued_for_recovery++;
3009         spin_unlock(&obd->obd_recovery_task_lock);
3010         wake_up(&obd->obd_next_transno_waitq);
3011         RETURN(0);
3012 }
3013
3014 void target_committed_to_req(struct ptlrpc_request *req)
3015 {
3016         struct obd_export *exp = req->rq_export;
3017
3018         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3019                 lustre_msg_set_last_committed(req->rq_repmsg,
3020                                               exp->exp_last_committed);
3021         else
3022                 DEBUG_REQ(D_IOCTL, req,
3023                           "not sending last_committed update (%d/%d)",
3024                           exp->exp_obd->obd_no_transno,
3025                           req->rq_repmsg == NULL);
3026
3027         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3028                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3029 }
3030
3031 #endif /* HAVE_SERVER_SUPPORT */
3032
3033 /**
3034  * Packs current SLV and Limit into \a req.
3035  */
3036 int target_pack_pool_reply(struct ptlrpc_request *req)
3037 {
3038         struct obd_device *obd;
3039
3040         ENTRY;
3041
3042         /*
3043          * Check that we still have all structures alive as this may
3044          * be some late RPC at shutdown time.
3045          */
3046         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3047                      !exp_connect_lru_resize(req->rq_export))) {
3048                 lustre_msg_set_slv(req->rq_repmsg, 0);
3049                 lustre_msg_set_limit(req->rq_repmsg, 0);
3050                 RETURN(0);
3051         }
3052
3053         /* OBD is alive here as export is alive, which we checked above. */
3054         obd = req->rq_export->exp_obd;
3055
3056         read_lock(&obd->obd_pool_lock);
3057         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3058         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3059         read_unlock(&obd->obd_pool_lock);
3060
3061         RETURN(0);
3062 }
3063
3064 static int target_send_reply_msg(struct ptlrpc_request *req,
3065                                  int rc, int fail_id)
3066 {
3067         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3068                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3069                 return -ECOMM;
3070         }
3071         /*
3072          * We can have a null rq_reqmsg in the event of bad signature or
3073          * no context when unwrapping
3074          */
3075         if (req->rq_reqmsg &&
3076             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3077             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3078                 return -ECOMM;
3079
3080         if (unlikely(rc)) {
3081                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3082                 req->rq_status = rc;
3083                 return ptlrpc_send_error(req, 1);
3084         }
3085         DEBUG_REQ(D_NET, req, "sending reply");
3086
3087         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3088 }
3089
3090 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3091 {
3092         struct ptlrpc_service_part *svcpt;
3093         int netrc;
3094         struct ptlrpc_reply_state *rs;
3095         struct obd_export *exp;
3096
3097         ENTRY;
3098
3099         if (req->rq_no_reply) {
3100                 EXIT;
3101                 return;
3102         }
3103
3104         svcpt = req->rq_rqbd->rqbd_svcpt;
3105         rs = req->rq_reply_state;
3106         if (rs == NULL || !rs->rs_difficult) {
3107                 /* no notifiers */
3108                 target_send_reply_msg(req, rc, fail_id);
3109                 EXIT;
3110                 return;
3111         }
3112
3113         /* must be an export if locks saved */
3114         LASSERT(req->rq_export != NULL);
3115         /* req/reply consistent */
3116         LASSERT(rs->rs_svcpt == svcpt);
3117
3118         /* "fresh" reply */
3119         LASSERT(!rs->rs_scheduled);
3120         LASSERT(!rs->rs_scheduled_ever);
3121         LASSERT(!rs->rs_handled);
3122         LASSERT(!rs->rs_on_net);
3123         LASSERT(rs->rs_export == NULL);
3124         LASSERT(list_empty(&rs->rs_obd_list));
3125         LASSERT(list_empty(&rs->rs_exp_list));
3126
3127         exp = class_export_get(req->rq_export);
3128
3129         /* disable reply scheduling while I'm setting up */
3130         rs->rs_scheduled = 1;
3131         rs->rs_on_net    = 1;
3132         rs->rs_xid       = req->rq_xid;
3133         rs->rs_transno   = req->rq_transno;
3134         rs->rs_export    = exp;
3135         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3136
3137         spin_lock(&exp->exp_uncommitted_replies_lock);
3138         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3139                rs->rs_transno, exp->exp_last_committed);
3140         if (rs->rs_transno > exp->exp_last_committed) {
3141                 /* not committed already */
3142                 list_add_tail(&rs->rs_obd_list,
3143                                   &exp->exp_uncommitted_replies);
3144         }
3145         spin_unlock(&exp->exp_uncommitted_replies_lock);
3146
3147         spin_lock(&exp->exp_lock);
3148         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3149         spin_unlock(&exp->exp_lock);
3150
3151         netrc = target_send_reply_msg(req, rc, fail_id);
3152
3153         spin_lock(&svcpt->scp_rep_lock);
3154
3155         atomic_inc(&svcpt->scp_nreps_difficult);
3156
3157         if (netrc != 0) {
3158                 /*
3159                  * error sending: reply is off the net.  Also we need +1
3160                  * reply ref until ptlrpc_handle_rs() is done
3161                  * with the reply state (if the send was successful, there
3162                  * would have been +1 ref for the net, which
3163                  * reply_out_callback leaves alone)
3164                  */
3165                 rs->rs_on_net = 0;
3166                 ptlrpc_rs_addref(rs);
3167         }
3168
3169         spin_lock(&rs->rs_lock);
3170         if (rs->rs_transno <= exp->exp_last_committed ||
3171             (!rs->rs_on_net && !rs->rs_no_ack) ||
3172             list_empty(&rs->rs_exp_list) ||     /* completed already */
3173             list_empty(&rs->rs_obd_list)) {
3174                 CDEBUG(D_HA, "Schedule reply immediately\n");
3175                 ptlrpc_dispatch_difficult_reply(rs);
3176         } else {
3177                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3178                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3179         }
3180         spin_unlock(&rs->rs_lock);
3181         spin_unlock(&svcpt->scp_rep_lock);
3182         EXIT;
3183 }
3184
3185 enum ldlm_mode lck_compat_array[] = {
3186         [LCK_EX]    = LCK_COMPAT_EX,
3187         [LCK_PW]    = LCK_COMPAT_PW,
3188         [LCK_PR]    = LCK_COMPAT_PR,
3189         [LCK_CW]    = LCK_COMPAT_CW,
3190         [LCK_CR]    = LCK_COMPAT_CR,
3191         [LCK_NL]    = LCK_COMPAT_NL,
3192         [LCK_GROUP] = LCK_COMPAT_GROUP,
3193         [LCK_COS]   = LCK_COMPAT_COS,
3194 };
3195
3196 /**
3197  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3198  * not escape to the user level.
3199  */
3200 int ldlm_error2errno(enum ldlm_error error)
3201 {
3202         int result;
3203
3204         switch (error) {
3205         case ELDLM_OK:
3206         case ELDLM_LOCK_MATCHED:
3207                 result = 0;
3208                 break;
3209         case ELDLM_LOCK_CHANGED:
3210                 result = -ESTALE;
3211                 break;
3212         case ELDLM_LOCK_ABORTED:
3213                 result = -ENAVAIL;
3214                 break;
3215         case ELDLM_LOCK_REPLACED:
3216                 result = -ESRCH;
3217                 break;
3218         case ELDLM_NO_LOCK_DATA:
3219                 result = -ENOENT;
3220                 break;
3221         case ELDLM_NAMESPACE_EXISTS:
3222                 result = -EEXIST;
3223                 break;
3224         case ELDLM_BAD_NAMESPACE:
3225                 result = -EBADF;
3226                 break;
3227         default:
3228                 if (((int)error) < 0) { /* cast to signed type */
3229                         result = error; /* as ldlm_error can be unsigned */
3230                 } else {
3231                         CERROR("Invalid DLM result code: %d\n", error);
3232                         result = -EPROTO;
3233                 }
3234         }
3235         return result;
3236 }
3237 EXPORT_SYMBOL(ldlm_error2errno);
3238
3239 /**
3240  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3241  */
3242 enum ldlm_error ldlm_errno2error(int err_no)
3243 {
3244         int error;
3245
3246         switch (err_no) {
3247         case 0:
3248                 error = ELDLM_OK;
3249                 break;
3250         case -ESTALE:
3251                 error = ELDLM_LOCK_CHANGED;
3252                 break;
3253         case -ENAVAIL:
3254                 error = ELDLM_LOCK_ABORTED;
3255                 break;
3256         case -ESRCH:
3257                 error = ELDLM_LOCK_REPLACED;
3258                 break;
3259         case -ENOENT:
3260                 error = ELDLM_NO_LOCK_DATA;
3261                 break;
3262         case -EEXIST:
3263                 error = ELDLM_NAMESPACE_EXISTS;
3264                 break;
3265         case -EBADF:
3266                 error = ELDLM_BAD_NAMESPACE;
3267                 break;
3268         default:
3269                 error = err_no;
3270         }
3271         return error;
3272 }
3273
3274 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3275 void ldlm_dump_export_locks(struct obd_export *exp)
3276 {
3277         spin_lock(&exp->exp_locks_list_guard);
3278         if (!list_empty(&exp->exp_locks_list)) {
3279                 struct ldlm_lock *lock;
3280
3281                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3282                        exp);
3283                 list_for_each_entry(lock, &exp->exp_locks_list,
3284                                         l_exp_refs_link)
3285                         LDLM_ERROR(lock, "lock:");
3286         }
3287         spin_unlock(&exp->exp_locks_list_guard);
3288 }
3289 #endif
3290
3291 #ifdef HAVE_SERVER_SUPPORT
3292 static int target_bulk_timeout(void *data)
3293 {
3294         ENTRY;
3295         /*
3296          * We don't fail the connection here, because having the export
3297          * killed makes the (vital) call to commitrw very sad.
3298          */
3299         RETURN(1);
3300 }
3301
3302 static inline const char *bulk2type(struct ptlrpc_request *req)
3303 {
3304         if (req->rq_bulk_read)
3305                 return "READ";
3306         if (req->rq_bulk_write)
3307                 return "WRITE";
3308         return "UNKNOWN";
3309 }
3310
3311 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3312                    struct l_wait_info *lwi)
3313 {
3314         struct ptlrpc_request *req = desc->bd_req;
3315         time64_t start = ktime_get_real_seconds();
3316         time64_t deadline;
3317         int rc = 0;
3318
3319         ENTRY;
3320
3321         /* If there is eviction in progress, wait for it to finish. */
3322         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3323                 *lwi = LWI_INTR(NULL, NULL);
3324                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3325                                   !atomic_read(&exp->exp_obd->obd_evict_inprogress),
3326                                   lwi);
3327         }
3328
3329         /* Check if client was evicted or reconnected already. */
3330         if (exp->exp_failed ||
3331             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3332                 rc = -ENOTCONN;
3333         } else {
3334                 if (req->rq_bulk_read)
3335                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3336
3337                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3338                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3339                 else /* old version, bulk matchbits is rq_xid */
3340                         req->rq_mbits = req->rq_xid;
3341
3342                 if (rc == 0)
3343                         rc = ptlrpc_start_bulk_transfer(desc);
3344         }
3345
3346         if (rc < 0) {
3347                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3348                           bulk2type(req), rc);
3349                 RETURN(rc);
3350         }
3351
3352         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3353                 ptlrpc_abort_bulk(desc);
3354                 RETURN(0);
3355         }
3356
3357         /* limit actual bulk transfer to bulk_timeout seconds */
3358         deadline = start + bulk_timeout;
3359         if (deadline > req->rq_deadline)
3360                 deadline = req->rq_deadline;
3361
3362         do {
3363                 time64_t timeoutl = deadline - ktime_get_real_seconds();
3364                 long timeout_jiffies = timeoutl <= 0 ?
3365                                        1 : cfs_time_seconds(timeoutl);
3366                 time64_t rq_deadline;
3367
3368                 *lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3369                                             cfs_time_seconds(1),
3370                                             target_bulk_timeout, desc);
3371                 rc = l_wait_event(desc->bd_waitq,
3372                                   !ptlrpc_server_bulk_active(desc) ||
3373                                   exp->exp_failed ||
3374                                   exp->exp_conn_cnt >
3375                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3376                                   lwi);
3377                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3378                 /* Wait again if we changed rq_deadline. */
3379                 rq_deadline = READ_ONCE(req->rq_deadline);
3380                 deadline = start + bulk_timeout;
3381                 if (deadline > rq_deadline)
3382                         deadline = rq_deadline;
3383         } while (rc == -ETIMEDOUT &&
3384                  deadline > ktime_get_real_seconds());
3385
3386         if (rc == -ETIMEDOUT) {
3387                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3388                           bulk2type(req), deadline - start,
3389                           ktime_get_real_seconds() - deadline);
3390                 ptlrpc_abort_bulk(desc);
3391         } else if (exp->exp_failed) {
3392                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3393                           bulk2type(req));
3394                 rc = -ENOTCONN;
3395                 ptlrpc_abort_bulk(desc);
3396         } else if (exp->exp_conn_cnt >
3397                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3398                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3399                           bulk2type(req));
3400                 /* We don't reply anyway. */
3401                 rc = -ETIMEDOUT;
3402                 ptlrpc_abort_bulk(desc);
3403         } else if (desc->bd_failure) {
3404                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3405                           bulk2type(req));
3406                 /* XXX should this be a different errno? */
3407                 rc = -ETIMEDOUT;
3408         } else {
3409                 if (req->rq_bulk_write)
3410                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3411                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3412                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3413                                   bulk2type(req), desc->bd_nob_transferred,
3414                                   desc->bd_nob);
3415                         /* XXX should this be a different errno? */
3416                         rc = -ETIMEDOUT;
3417                 }
3418         }
3419
3420         RETURN(rc);
3421 }
3422 EXPORT_SYMBOL(target_bulk_io);
3423
3424 #endif /* HAVE_SERVER_SUPPORT */