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_DEF_SHORT_IO_BYTES;
 438
 439         /*
 440          * set cl_chunkbits default value to PAGE_SHIFT,
 441          * it will be updated at OSC connection time.
 442          */
 443         cli->cl_chunkbits = PAGE_SHIFT;
 444
 445         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 446                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 447         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 448                 cli->cl_max_rpcs_in_flight = 2;
 449         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 450                 cli->cl_max_rpcs_in_flight = 3;
 451         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 452                 cli->cl_max_rpcs_in_flight = 4;
 453         } else {
 454                 if (osc_on_mdt(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                               atomic_read(&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 = obd_uuid_lookup(target, &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 if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1248                    !lw_client) {
1249                 spin_unlock(&export->exp_lock);
1250                 rc = -EAGAIN;
1251         } else {
1252                 export->exp_connecting = 1;
1253                 spin_unlock(&export->exp_lock);
1254                 LASSERT(export->exp_obd == target);
1255
1256                 rc = target_handle_reconnect(&conn, export, &cluuid);
1257         }
1258
1259         /* If we found an export, we already unlocked. */
1260         if (!export) {
1261 no_export:
1262                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1263         } else if (req->rq_export == NULL &&
1264                    atomic_read(&export->exp_rpc_count) > 0) {
1265                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1266                               target->obd_name, cluuid.uuid,
1267                               libcfs_nid2str(req->rq_peer.nid),
1268                               refcount_read(&export->exp_handle.h_ref));
1269                         GOTO(out, rc = -EBUSY);
1270         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1271                    rc != EALREADY) {
1272                 if (!strstr(cluuid.uuid, "mdt"))
1273                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1274                                       target->obd_name, cluuid.uuid,
1275                                       libcfs_nid2str(req->rq_peer.nid));
1276                 GOTO(out, rc = -EALREADY);
1277         } else {
1278                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1279         }
1280
1281         if (rc < 0)
1282                 GOTO(out, rc);
1283
1284         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1285                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1286                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1287                export, ktime_get_seconds(),
1288                export ? export->exp_last_request_time : 0);
1289
1290         /*
1291          * If this is the first time a client connects, reset the recovery
1292          * timer. Discard lightweight connections which might be local.
1293          */
1294         if (!lw_client && rc == 0 && target->obd_recovering)
1295                 check_and_start_recovery_timer(target, req, export == NULL);
1296
1297         /*
1298          * We want to handle EALREADY but *not* -EALREADY from
1299          * target_handle_reconnect(), return reconnection state in a flag.
1300          */
1301         if (rc == EALREADY) {
1302                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1303                 rc = 0;
1304         } else {
1305                 LASSERT(rc == 0);
1306         }
1307
1308         /* Tell the client if we support replayable requests. */
1309         if (target->obd_replayable)
1310                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1311         client_nid = &req->rq_peer.nid;
1312
1313         if (export == NULL) {
1314                 /* allow lightweight connections during recovery */
1315                 /*
1316                  * allow "new" MDT to be connected during recovery, since we
1317                  * need retrieve recovery update records from it
1318                  */
1319                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1320                         struct hrtimer *timer = &target->obd_recovery_timer;
1321                         ktime_t remaining;
1322                         s64 timeout, left;
1323                         int in_progress;
1324                         int connected;
1325                         int known;
1326                         int stale;
1327                         char *msg;
1328
1329                         connected = atomic_read(&target->obd_connected_clients);
1330                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1331                         known =
1332                            atomic_read(&target->obd_max_recoverable_clients);
1333                         stale = target->obd_stale_clients;
1334                         remaining = hrtimer_expires_remaining(timer);
1335                         left = ktime_divns(remaining, NSEC_PER_SEC);
1336
1337                         if (ktime_to_ns(remaining) > 0) {
1338                                 msg = "to recover in";
1339                                 timeout = left;
1340                         } else {
1341                                 msg = "already passed deadline";
1342                                 timeout = -left;
1343                         }
1344
1345                         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",
1346                                       target->obd_name, cluuid.uuid,
1347                                       libcfs_nid2str(req->rq_peer.nid), known,
1348                                       connected - in_progress, in_progress,
1349                                       stale, msg, timeout / 60, timeout % 60);
1350                         rc = -EBUSY;
1351                 } else {
1352 dont_check_exports:
1353                         rc = obd_connect(req->rq_svc_thread->t_env,
1354                                          &export, target, &cluuid, data,
1355                                          client_nid);
1356                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1357                                 lustre_msg_add_op_flags(req->rq_repmsg,
1358                                                         MSG_CONNECT_RECOVERING);
1359                         if (rc == 0) {
1360                                 conn.cookie = export->exp_handle.h_cookie;
1361                                 rc = rev_import_init(export);
1362                         }
1363
1364                         if (mds_mds_conn)
1365                                 new_mds_mds_conn = true;
1366                 }
1367         } else {
1368                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1369                                    export, target, &cluuid, data, client_nid);
1370         }
1371         if (rc)
1372                 GOTO(out, rc);
1373
1374         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1375         data->ocd_instance = target->u.obt.obt_instance;
1376
1377         /*
1378          * Return only the parts of obd_connect_data that we understand, so the
1379          * client knows that we don't understand the rest.
1380          */
1381         if (data) {
1382                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1383                                                RCL_SERVER);
1384                 tmpdata = req_capsule_server_get(&req->rq_pill,
1385                                                  &RMF_CONNECT_DATA);
1386                 /*
1387                  * Don't use struct assignment here, because the client reply
1388                  * buffer may be smaller/larger than the local struct
1389                  * obd_connect_data.
1390                  */
1391                 memcpy(tmpdata, data, min(tmpsize, size));
1392         }
1393
1394         /*
1395          * If the client and the server are the same node, we will already
1396          * have an export that really points to the client's DLM export,
1397          * because we have a shared handles table.
1398          *
1399          * XXX this will go away when shaver stops sending the "connect" handle
1400          * in the real "remote handle" field of the request --phik 24 Apr 2003
1401          */
1402         ptlrpc_request_change_export(req, export);
1403
1404         spin_lock(&export->exp_lock);
1405         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1406                 spin_unlock(&export->exp_lock);
1407                 CDEBUG(D_RPCTRACE,
1408                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1409                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1410                        export->exp_conn_cnt,
1411                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1412
1413                 GOTO(out, rc = -EALREADY);
1414         }
1415         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1416         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1417         spin_unlock(&export->exp_lock);
1418
1419         if (export->exp_connection != NULL) {
1420                 /* Check to see if connection came from another NID. */
1421                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1422                     !hlist_unhashed(&export->exp_nid_hash))
1423                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1424                                      &export->exp_connection->c_peer.nid,
1425                                      &export->exp_nid_hash);
1426
1427                 ptlrpc_connection_put(export->exp_connection);
1428         }
1429
1430         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1431                                                        req->rq_self,
1432                                                        &cluuid);
1433         if (hlist_unhashed(&export->exp_nid_hash))
1434                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1435                              &export->exp_connection->c_peer.nid,
1436                              &export->exp_nid_hash);
1437
1438         lustre_msg_set_handle(req->rq_repmsg, &conn);
1439
1440         rc = rev_import_reconnect(export, req);
1441         if (rc != 0)
1442                 GOTO(out, rc);
1443
1444         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1445                 int has_transno;
1446                 __u64 transno = data->ocd_transno;
1447
1448                 spin_lock(&export->exp_lock);
1449                 /*
1450                  * possible race with class_disconnect_stale_exports,
1451                  * export may be already in the eviction process
1452                  */
1453                 if (export->exp_failed) {
1454                         spin_unlock(&export->exp_lock);
1455                         GOTO(out, rc = -ENODEV);
1456                 }
1457                 export->exp_in_recovery = 1;
1458                 export->exp_req_replay_needed = 1;
1459                 export->exp_lock_replay_needed = 1;
1460                 spin_unlock(&export->exp_lock);
1461
1462                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1463                                  MSG_CONNECT_TRANSNO);
1464                 if (has_transno && transno == 0)
1465                         CWARN("Connect with zero transno!\n");
1466
1467                 if (has_transno && transno > 0 &&
1468                     transno < target->obd_next_recovery_transno &&
1469                     transno > target->obd_last_committed) {
1470                         /* Another way is to use cmpxchg() to be lock-free. */
1471                         spin_lock(&target->obd_recovery_task_lock);
1472                         if (transno < target->obd_next_recovery_transno)
1473                                 target->obd_next_recovery_transno = transno;
1474                         spin_unlock(&target->obd_recovery_task_lock);
1475                 }
1476
1477                 atomic_inc(&target->obd_req_replay_clients);
1478                 atomic_inc(&target->obd_lock_replay_clients);
1479                 /*
1480                  * Note: MDS-MDS connection is allowed to be connected during
1481                  * recovery, no matter if the exports needs to be recoveried.
1482                  * Because we need retrieve updates logs from all other MDTs.
1483                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1484                  * also needs to be increased to match other recovery checking
1485                  * condition.
1486                  */
1487                 if (new_mds_mds_conn)
1488                         atomic_inc(&target->obd_max_recoverable_clients);
1489
1490                 if (atomic_inc_return(&target->obd_connected_clients) ==
1491                     atomic_read(&target->obd_max_recoverable_clients))
1492                         wake_up(&target->obd_next_transno_waitq);
1493         }
1494
1495         /* Tell the client we're in recovery, when client is involved in it. */
1496         if (target->obd_recovering && !lw_client)
1497                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1498
1499 out:
1500         if (export) {
1501                 spin_lock(&export->exp_lock);
1502                 export->exp_connecting = 0;
1503                 spin_unlock(&export->exp_lock);
1504
1505                 class_export_put(export);
1506         }
1507         if (target != NULL) {
1508                 spin_lock(&target->obd_dev_lock);
1509                 target->obd_conn_inprogress--;
1510                 spin_unlock(&target->obd_dev_lock);
1511                 class_decref(target, "find", current);
1512         }
1513         req->rq_status = rc;
1514         RETURN(rc);
1515 }
1516
1517 int target_handle_disconnect(struct ptlrpc_request *req)
1518 {
1519         int rc;
1520
1521         ENTRY;
1522
1523         rc = req_capsule_server_pack(&req->rq_pill);
1524         if (rc)
1525                 RETURN(rc);
1526
1527         /* Keep the rq_export around so we can send the reply. */
1528         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1529
1530         RETURN(0);
1531 }
1532
1533 void target_destroy_export(struct obd_export *exp)
1534 {
1535         struct obd_import *imp = NULL;
1536         /*
1537          * exports created from last_rcvd data, and "fake"
1538          * exports created by lctl don't have an import
1539          */
1540         spin_lock(&exp->exp_lock);
1541         if (exp->exp_imp_reverse != NULL) {
1542                 imp = exp->exp_imp_reverse;
1543                 exp->exp_imp_reverse = NULL;
1544         }
1545         spin_unlock(&exp->exp_lock);
1546         if (imp != NULL)
1547                 client_destroy_import(imp);
1548
1549         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1550         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1551         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1552         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1553 }
1554 EXPORT_SYMBOL(target_destroy_export);
1555
1556 /*
1557  * Recovery functions
1558  */
1559 static void target_request_copy_get(struct ptlrpc_request *req)
1560 {
1561         class_export_rpc_inc(req->rq_export);
1562         LASSERT(list_empty(&req->rq_list));
1563         INIT_LIST_HEAD(&req->rq_replay_list);
1564
1565         /* Increase refcount to keep request in queue. */
1566         atomic_inc(&req->rq_refcount);
1567         /* Let export know it has replays to be handled. */
1568         atomic_inc(&req->rq_export->exp_replay_count);
1569 }
1570
1571 static void target_request_copy_put(struct ptlrpc_request *req)
1572 {
1573         LASSERT(list_empty(&req->rq_replay_list));
1574         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1575
1576         atomic_dec(&req->rq_export->exp_replay_count);
1577         class_export_rpc_dec(req->rq_export);
1578         ptlrpc_server_drop_request(req);
1579 }
1580
1581 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1582 {
1583         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1584         struct obd_export *exp = req->rq_export;
1585         struct ptlrpc_request *reqiter;
1586         struct ptlrpc_request *dup_req = NULL;
1587         int dup = 0;
1588
1589         LASSERT(exp);
1590
1591         spin_lock(&exp->exp_lock);
1592         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1593                             rq_replay_list) {
1594                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1595                         dup_req = reqiter;
1596                         dup = 1;
1597                         break;
1598                 }
1599         }
1600
1601         if (dup) {
1602                 /* We expect it with RESENT and REPLAY flags. */
1603                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1604                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1605                         CERROR("invalid flags %x of resent replay\n",
1606                                lustre_msg_get_flags(req->rq_reqmsg));
1607
1608                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1609                         __u32 new_conn;
1610
1611                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1612                         if (new_conn >
1613                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1614                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1615                                                         new_conn);
1616                 }
1617         } else {
1618                 list_add_tail(&req->rq_replay_list,
1619                               &exp->exp_req_replay_queue);
1620         }
1621
1622         spin_unlock(&exp->exp_lock);
1623         return dup;
1624 }
1625
1626 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1627 {
1628         LASSERT(!list_empty(&req->rq_replay_list));
1629         LASSERT(req->rq_export);
1630
1631         spin_lock(&req->rq_export->exp_lock);
1632         list_del_init(&req->rq_replay_list);
1633         spin_unlock(&req->rq_export->exp_lock);
1634 }
1635
1636 static void target_finish_recovery(struct lu_target *lut)
1637 {
1638         struct obd_device *obd = lut->lut_obd;
1639
1640         ENTRY;
1641
1642         /* Only log a recovery message when recovery has occurred. */
1643         if (obd->obd_recovery_start) {
1644                 time64_t now = ktime_get_seconds();
1645                 time64_t elapsed_time;
1646
1647                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1648                                      1);
1649                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1650                               obd->obd_name, (s64)elapsed_time / 60,
1651                               (s64)elapsed_time % 60,
1652                               atomic_read(&obd->obd_max_recoverable_clients),
1653                               atomic_read(&obd->obd_connected_clients),
1654                               obd->obd_stale_clients,
1655                               obd->obd_stale_clients == 1 ? "was" : "were");
1656         }
1657
1658         ldlm_reprocess_recovery_done(obd->obd_namespace);
1659         spin_lock(&obd->obd_recovery_task_lock);
1660         if (!list_empty(&obd->obd_req_replay_queue) ||
1661             !list_empty(&obd->obd_lock_replay_queue) ||
1662             !list_empty(&obd->obd_final_req_queue)) {
1663                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1664                        obd->obd_name,
1665                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1666                        list_empty(&obd->obd_lock_replay_queue) ? \
1667                                   "" : "lock ",
1668                        list_empty(&obd->obd_final_req_queue) ? \
1669                                   "" : "final ");
1670                 spin_unlock(&obd->obd_recovery_task_lock);
1671                 LBUG();
1672         }
1673         spin_unlock(&obd->obd_recovery_task_lock);
1674
1675         obd->obd_recovery_end = ktime_get_seconds();
1676
1677         /* When recovery finished, cleanup orphans on MDS and OST. */
1678         if (obd->obd_type && OBP(obd, postrecov)) {
1679                 int rc = OBP(obd, postrecov)(obd);
1680
1681                 if (rc < 0)
1682                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1683                                       obd->obd_name, rc);
1684         }
1685         EXIT;
1686 }
1687
1688 static void abort_req_replay_queue(struct obd_device *obd)
1689 {
1690         struct ptlrpc_request *req, *n;
1691         struct list_head abort_list;
1692
1693         INIT_LIST_HEAD(&abort_list);
1694         spin_lock(&obd->obd_recovery_task_lock);
1695         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1696         spin_unlock(&obd->obd_recovery_task_lock);
1697         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1698                 DEBUG_REQ(D_WARNING, req, "aborted:");
1699                 req->rq_status = -ENOTCONN;
1700                 if (ptlrpc_error(req)) {
1701                         DEBUG_REQ(D_ERROR, req,
1702                                   "failed abort_req_reply; skipping");
1703                 }
1704                 target_exp_dequeue_req_replay(req);
1705                 target_request_copy_put(req);
1706         }
1707 }
1708
1709 static void abort_lock_replay_queue(struct obd_device *obd)
1710 {
1711         struct ptlrpc_request *req, *n;
1712         struct list_head abort_list;
1713
1714         INIT_LIST_HEAD(&abort_list);
1715         spin_lock(&obd->obd_recovery_task_lock);
1716         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1717         spin_unlock(&obd->obd_recovery_task_lock);
1718         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1719                 DEBUG_REQ(D_ERROR, req, "aborted:");
1720                 req->rq_status = -ENOTCONN;
1721                 if (ptlrpc_error(req)) {
1722                         DEBUG_REQ(D_ERROR, req,
1723                                   "failed abort_lock_reply; skipping");
1724                 }
1725                 target_request_copy_put(req);
1726         }
1727 }
1728
1729 /*
1730  * Called from a cleanup function if the device is being cleaned up
1731  * forcefully.  The exports should all have been disconnected already,
1732  * the only thing left to do is
1733  * - clear the recovery flags
1734  * - cancel the timer
1735  * - free queued requests and replies, but don't send replies
1736  * Because the obd_stopping flag is set, no new requests should be received.
1737  */
1738 void target_cleanup_recovery(struct obd_device *obd)
1739 {
1740         struct ptlrpc_request *req, *n;
1741         struct list_head clean_list;
1742
1743         INIT_LIST_HEAD(&clean_list);
1744         spin_lock(&obd->obd_dev_lock);
1745         if (!obd->obd_recovering) {
1746                 spin_unlock(&obd->obd_dev_lock);
1747                 EXIT;
1748                 return;
1749         }
1750         obd->obd_recovering = obd->obd_abort_recovery = 0;
1751         spin_unlock(&obd->obd_dev_lock);
1752
1753         spin_lock(&obd->obd_recovery_task_lock);
1754         target_cancel_recovery_timer(obd);
1755         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1756         spin_unlock(&obd->obd_recovery_task_lock);
1757
1758         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1759                 LASSERT(req->rq_reply_state == NULL);
1760                 target_exp_dequeue_req_replay(req);
1761                 target_request_copy_put(req);
1762         }
1763
1764         spin_lock(&obd->obd_recovery_task_lock);
1765         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1766         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1767         spin_unlock(&obd->obd_recovery_task_lock);
1768
1769         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1770                 LASSERT(req->rq_reply_state == NULL);
1771                 target_request_copy_put(req);
1772         }
1773
1774         EXIT;
1775 }
1776 EXPORT_SYMBOL(target_cleanup_recovery);
1777
1778 /* obd_recovery_task_lock should be held */
1779 void target_cancel_recovery_timer(struct obd_device *obd)
1780 {
1781         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1782         hrtimer_cancel(&obd->obd_recovery_timer);
1783 }
1784
1785 static void target_start_recovery_timer(struct obd_device *obd)
1786 {
1787         ktime_t delay;
1788
1789         if (obd->obd_recovery_start != 0)
1790                 return;
1791
1792         spin_lock(&obd->obd_dev_lock);
1793         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1794                 spin_unlock(&obd->obd_dev_lock);
1795                 return;
1796         }
1797
1798         LASSERT(obd->obd_recovery_timeout != 0);
1799
1800         if (obd->obd_recovery_start != 0) {
1801                 spin_unlock(&obd->obd_dev_lock);
1802                 return;
1803         }
1804
1805         obd->obd_recovery_start = ktime_get_seconds();
1806         delay = ktime_set(obd->obd_recovery_start +
1807                           obd->obd_recovery_timeout, 0);
1808         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1809         spin_unlock(&obd->obd_dev_lock);
1810
1811         LCONSOLE_WARN("%s: Will be in recovery for at least %lu:%02lu, or until %d client%s reconnect%s\n",
1812                       obd->obd_name,
1813                       obd->obd_recovery_timeout / 60,
1814                       obd->obd_recovery_timeout % 60,
1815                       atomic_read(&obd->obd_max_recoverable_clients),
1816                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1817                       "" : "s",
1818                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1819                       "s" : "");
1820 }
1821
1822 /**
1823  * extend recovery window.
1824  *
1825  * if @extend is true, extend recovery window to have @dr_timeout remaining
1826  * at least; otherwise, make sure the recovery timeout value is not less
1827  * than @dr_timeout.
1828  */
1829 static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout,
1830                                   bool extend)
1831 {
1832         ktime_t left_ns;
1833         time_t timeout;
1834         time_t left;
1835
1836         spin_lock(&obd->obd_dev_lock);
1837         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1838             obd->obd_stopping) {
1839                 spin_unlock(&obd->obd_dev_lock);
1840                 return;
1841         }
1842         LASSERT(obd->obd_recovery_start != 0);
1843
1844         left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1845         left = ktime_divns(left_ns, NSEC_PER_SEC);
1846
1847         if (extend) {
1848                 timeout = obd->obd_recovery_timeout;
1849                 /* dr_timeout will happen after the hrtimer has expired.
1850                  * Add the excess time to the soft recovery timeout without
1851                  * exceeding the hard recovery timeout.
1852                  */
1853                 if (dr_timeout > left) {
1854                         timeout += dr_timeout - left;
1855                         timeout = min_t(time_t, obd->obd_recovery_time_hard,
1856                                         timeout);
1857                 }
1858         } else {
1859                 timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout,
1860                                   obd->obd_recovery_time_hard);
1861         }
1862
1863         if (timeout == obd->obd_recovery_time_hard)
1864                 CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n",
1865                       obd->obd_name, timeout, extend);
1866
1867         if (obd->obd_recovery_timeout < timeout) {
1868                 ktime_t end, now;
1869
1870                 obd->obd_recovery_timeout = timeout;
1871                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1872                 now = ktime_set(ktime_get_seconds(), 0);
1873                 left_ns = ktime_sub(end, now);
1874                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1875                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1876         }
1877         spin_unlock(&obd->obd_dev_lock);
1878
1879         CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1880                 obd->obd_name, left);
1881 }
1882
1883 /* Reset the timer with each new client connection */
1884 /*
1885  * This timer is actually reconnect_timer, which is for making sure
1886  * the total recovery window is at least as big as my reconnect
1887  * attempt timing. So the initial recovery time_out will be set to
1888  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1889  * from client is bigger than this, then the recovery time_out will
1890  * be extended to make sure the client could be reconnected, in the
1891  * process, the timeout from the new client should be ignored.
1892  */
1893 static void
1894 check_and_start_recovery_timer(struct obd_device *obd,
1895                                struct ptlrpc_request *req,
1896                                int new_client)
1897 {
1898         time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1899         struct obd_device_target *obt = &obd->u.obt;
1900
1901         if (!new_client && service_time)
1902                 /*
1903                  * Teach server about old server's estimates, as first guess
1904                  * at how long new requests will take.
1905                  */
1906                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1907                             service_time);
1908
1909         target_start_recovery_timer(obd);
1910
1911         /*
1912          * Convert the service time to RPC timeout,
1913          * and reuse service_time to limit stack usage.
1914          */
1915         service_time = at_est2timeout(service_time);
1916
1917         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1918             service_time < at_extra)
1919                 service_time = at_extra;
1920
1921         /*
1922          * We expect other clients to timeout within service_time, then try
1923          * to reconnect, then try the failover server.  The max delay between
1924          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1925          */
1926         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1927
1928         LASSERT(obt->obt_magic == OBT_MAGIC);
1929         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1930         if (service_time > obd->obd_recovery_timeout && !new_client)
1931                 extend_recovery_timer(obd, service_time, false);
1932 }
1933
1934 /** Health checking routines */
1935 static inline int exp_connect_healthy(struct obd_export *exp)
1936 {
1937         return exp->exp_in_recovery;
1938 }
1939
1940 /** if export done req_replay or has replay in queue */
1941 static inline int exp_req_replay_healthy(struct obd_export *exp)
1942 {
1943         return (!exp->exp_req_replay_needed ||
1944                 atomic_read(&exp->exp_replay_count) > 0);
1945 }
1946
1947
1948 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1949 {
1950         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1951                exp_req_replay_healthy(exp);
1952 }
1953
1954 /** if export done lock_replay or has replay in queue */
1955 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1956 {
1957         return (!exp->exp_lock_replay_needed ||
1958                 atomic_read(&exp->exp_replay_count) > 0);
1959 }
1960
1961 static inline int exp_vbr_healthy(struct obd_export *exp)
1962 {
1963         return !exp->exp_vbr_failed;
1964 }
1965
1966 static inline int exp_finished(struct obd_export *exp)
1967 {
1968         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1969 }
1970
1971 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1972 {
1973         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1974                 exp_finished(exp);
1975 }
1976
1977 static int check_for_next_transno(struct lu_target *lut)
1978 {
1979         struct ptlrpc_request *req = NULL;
1980         struct obd_device *obd = lut->lut_obd;
1981         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1982         int wake_up = 0, connected, completed, queue_len;
1983         __u64 req_transno = 0;
1984         __u64 update_transno = 0;
1985         __u64 next_transno = 0;
1986
1987         ENTRY;
1988
1989         spin_lock(&obd->obd_recovery_task_lock);
1990         if (!list_empty(&obd->obd_req_replay_queue)) {
1991                 req = list_entry(obd->obd_req_replay_queue.next,
1992                                      struct ptlrpc_request, rq_list);
1993                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1994         }
1995
1996         if (tdtd != NULL)
1997                 update_transno = distribute_txn_get_next_transno(tdtd);
1998
1999         connected = atomic_read(&obd->obd_connected_clients);
2000         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2001         queue_len = obd->obd_requests_queued_for_recovery;
2002         next_transno = obd->obd_next_recovery_transno;
2003
2004         CDEBUG(D_HA,
2005                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2006                atomic_read(&obd->obd_max_recoverable_clients),
2007                connected, completed,
2008                queue_len, req_transno, next_transno);
2009
2010         if (obd->obd_abort_recovery) {
2011                 CDEBUG(D_HA, "waking for aborted recovery\n");
2012                 wake_up = 1;
2013         } else if (obd->obd_recovery_expired) {
2014                 CDEBUG(D_HA, "waking for expired recovery\n");
2015                 wake_up = 1;
2016         } else if (tdtd != NULL && req != NULL &&
2017                    is_req_replayed_by_update(req)) {
2018                 LASSERTF(req_transno < next_transno,
2019                          "req_transno %llu next_transno%llu\n", req_transno,
2020                          next_transno);
2021                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2022                        req_transno);
2023                 wake_up = 1;
2024         } else if (req_transno == next_transno ||
2025                    (update_transno != 0 && update_transno <= next_transno)) {
2026                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2027                 wake_up = 1;
2028         } else if (queue_len > 0 &&
2029                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2030                 /** handle gaps occured due to lost reply or VBR */
2031                 LASSERTF(req_transno >= next_transno,
2032                          "req_transno: %llu, next_transno: %llu\n",
2033                          req_transno, next_transno);
2034                 CDEBUG(D_HA,
2035                        "%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",
2036                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2037                        next_transno, queue_len, completed, connected,
2038                        req_transno, update_transno, obd->obd_last_committed);
2039                 obd->obd_next_recovery_transno = req_transno;
2040                 wake_up = 1;
2041         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2042                 CDEBUG(D_HA, "waking for completed recovery\n");
2043                 wake_up = 1;
2044         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2045                 CDEBUG(D_HA,
2046                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2047                        next_transno);
2048                 obd->obd_next_recovery_transno = req_transno;
2049                 wake_up = 1;
2050         }
2051         spin_unlock(&obd->obd_recovery_task_lock);
2052         return wake_up;
2053 }
2054
2055 static int check_for_next_lock(struct lu_target *lut)
2056 {
2057         struct obd_device *obd = lut->lut_obd;
2058         int wake_up = 0;
2059
2060         spin_lock(&obd->obd_recovery_task_lock);
2061         if (!list_empty(&obd->obd_lock_replay_queue)) {
2062                 CDEBUG(D_HA, "waking for next lock\n");
2063                 wake_up = 1;
2064         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2065                 CDEBUG(D_HA, "waking for completed lock replay\n");
2066                 wake_up = 1;
2067         } else if (obd->obd_abort_recovery) {
2068                 CDEBUG(D_HA, "waking for aborted recovery\n");
2069                 wake_up = 1;
2070         } else if (obd->obd_recovery_expired) {
2071                 CDEBUG(D_HA, "waking for expired recovery\n");
2072                 wake_up = 1;
2073         }
2074         spin_unlock(&obd->obd_recovery_task_lock);
2075
2076         return wake_up;
2077 }
2078
2079 /**
2080  * wait for recovery events,
2081  * check its status with help of check_routine
2082  * evict dead clients via health_check
2083  */
2084 static int target_recovery_overseer(struct lu_target *lut,
2085                                     int (*check_routine)(struct lu_target *),
2086                                     int (*health_check)(struct obd_export *))
2087 {
2088         struct obd_device *obd = lut->lut_obd;
2089         struct target_distribute_txn_data *tdtd;
2090         time64_t last = 0;
2091         time64_t now;
2092 repeat:
2093         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2094                 now = ktime_get_seconds();
2095                 if (now - last > 600) {
2096                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2097                                       obd->obd_name);
2098                         last = now;
2099                 }
2100         }
2101         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2102               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2103                 __u64 next_update_transno = 0;
2104
2105                 /*
2106                  * Only abort the recovery if there are no update recovery
2107                  * left in the queue
2108                  */
2109                 spin_lock(&obd->obd_recovery_task_lock);
2110                 if (lut->lut_tdtd != NULL) {
2111                         next_update_transno =
2112                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2113
2114                         tdtd = lut->lut_tdtd;
2115                         /*
2116                          * If next_update_transno == 0, it probably because
2117                          * updatelog retrieve threads did not get any records
2118                          * yet, let's wait those threads stopped
2119                          */
2120                         if (next_update_transno == 0) {
2121                                 wait_event_idle(
2122                                         tdtd->tdtd_recovery_threads_waitq,
2123                                         atomic_read(&tdtd->tdtd_recovery_threads_count)
2124                                         == 0);
2125
2126                                 next_update_transno =
2127                                         distribute_txn_get_next_transno(
2128                                                                 lut->lut_tdtd);
2129                         }
2130                 }
2131
2132                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2133                         obd->obd_next_recovery_transno = next_update_transno;
2134                         spin_unlock(&obd->obd_recovery_task_lock);
2135                         /*
2136                          * Disconnect unfinished exports from clients, and
2137                          * keep connection from MDT to make sure the update
2138                          * recovery will still keep trying until some one
2139                          * manually abort the recovery
2140                          */
2141                         class_disconnect_stale_exports(obd,
2142                                                 exp_finished_or_from_mdt);
2143                         /* Abort all of replay & replay lock req from clients */
2144                         abort_req_replay_queue(obd);
2145                         abort_lock_replay_queue(obd);
2146                         CDEBUG(D_HA,
2147                                "%s: there are still update replay (%#llx)in the queue.\n",
2148                                obd->obd_name, next_update_transno);
2149                 } else {
2150                         obd->obd_abort_recovery = 1;
2151                         spin_unlock(&obd->obd_recovery_task_lock);
2152                         CWARN("%s recovery is aborted by hard timeout\n",
2153                               obd->obd_name);
2154                 }
2155         }
2156
2157         while (wait_event_timeout(obd->obd_next_transno_waitq,
2158                                   check_routine(lut),
2159                                   cfs_time_seconds(60)) == 0)
2160                 ; /* wait indefinitely for event, but don't trigger watchdog */
2161
2162         if (obd->obd_abort_recovery) {
2163                 CWARN("recovery is aborted, evict exports in recovery\n");
2164                 if (lut->lut_tdtd != NULL) {
2165                         tdtd = lut->lut_tdtd;
2166                         /*
2167                          * Let's wait all of the update log recovery thread
2168                          * finished
2169                          */
2170                         wait_event_idle(
2171                                 tdtd->tdtd_recovery_threads_waitq,
2172                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2173                                 == 0);
2174                         /* Then abort the update recovery list */
2175                         dtrq_list_destroy(lut->lut_tdtd);
2176                 }
2177
2178                 /** evict exports which didn't finish recovery yet */
2179                 class_disconnect_stale_exports(obd, exp_finished);
2180                 return 1;
2181         } else if (obd->obd_recovery_expired &&
2182                    obd->obd_recovery_timeout < obd->obd_recovery_time_hard) {
2183                 obd->obd_recovery_expired = 0;
2184
2185                 /** If some clients died being recovered, evict them */
2186                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2187                               obd->obd_name);
2188                 /** evict cexports with no replay in queue, they are stalled */
2189                 class_disconnect_stale_exports(obd, health_check);
2190
2191                 /** continue with VBR */
2192                 spin_lock(&obd->obd_dev_lock);
2193                 obd->obd_version_recov = 1;
2194                 spin_unlock(&obd->obd_dev_lock);
2195                 /**
2196                  * reset timer, recovery will proceed with versions now,
2197                  * timeout is set just to handle reconnection delays
2198                  */
2199                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2200                 /**
2201                  * Wait for recovery events again, after evicting bad clients
2202                  */
2203                 goto repeat;
2204         }
2205         return 0;
2206 }
2207
2208 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2209 {
2210         struct obd_device *obd = lut->lut_obd;
2211         struct ptlrpc_request *req = NULL;
2212
2213         CDEBUG(D_HA, "Waiting for lock\n");
2214         if (target_recovery_overseer(lut, check_for_next_lock,
2215                                      exp_lock_replay_healthy))
2216                 abort_lock_replay_queue(obd);
2217
2218         spin_lock(&obd->obd_recovery_task_lock);
2219         if (!list_empty(&obd->obd_lock_replay_queue)) {
2220                 req = list_entry(obd->obd_lock_replay_queue.next,
2221                                      struct ptlrpc_request, rq_list);
2222                 list_del_init(&req->rq_list);
2223                 spin_unlock(&obd->obd_recovery_task_lock);
2224         } else {
2225                 spin_unlock(&obd->obd_recovery_task_lock);
2226                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2227                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2228                 /** evict exports failed VBR */
2229                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2230         }
2231         return req;
2232 }
2233
2234 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2235 {
2236         struct ptlrpc_request *req = NULL;
2237
2238         spin_lock(&obd->obd_recovery_task_lock);
2239         if (!list_empty(&obd->obd_final_req_queue)) {
2240                 req = list_entry(obd->obd_final_req_queue.next,
2241                                      struct ptlrpc_request, rq_list);
2242                 list_del_init(&req->rq_list);
2243                 spin_unlock(&obd->obd_recovery_task_lock);
2244                 if (req->rq_export->exp_in_recovery) {
2245                         spin_lock(&req->rq_export->exp_lock);
2246                         req->rq_export->exp_in_recovery = 0;
2247                         spin_unlock(&req->rq_export->exp_lock);
2248                 }
2249         } else {
2250                 spin_unlock(&obd->obd_recovery_task_lock);
2251         }
2252         return req;
2253 }
2254
2255 static void handle_recovery_req(struct ptlrpc_thread *thread,
2256                                 struct ptlrpc_request *req,
2257                                 svc_handler_t handler)
2258 {
2259         ENTRY;
2260
2261         /**
2262          * export can be evicted during recovery, no need to handle replays for
2263          * it after that, discard such request silently
2264          */
2265         if (req->rq_export->exp_disconnected)
2266                 RETURN_EXIT;
2267
2268         req->rq_session.lc_thread = thread;
2269         req->rq_svc_thread = thread;
2270         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2271
2272         /* thread context */
2273         lu_context_enter(&thread->t_env->le_ctx);
2274         (void)handler(req);
2275         lu_context_exit(&thread->t_env->le_ctx);
2276
2277         req->rq_svc_thread->t_env->le_ses = NULL;
2278
2279         /* don't reset timer for final stage */
2280         if (!exp_finished(req->rq_export)) {
2281                 time_t to = obd_timeout;
2282
2283                 /**
2284                  * Add request timeout to the recovery time so next request from
2285                  * this client may come in recovery time
2286                  */
2287                 if (!AT_OFF) {
2288                         struct ptlrpc_service_part *svcpt;
2289
2290                         svcpt = req->rq_rqbd->rqbd_svcpt;
2291                         /*
2292                          * If the server sent early reply for this request,
2293                          * the client will recalculate the timeout according to
2294                          * current server estimate service time, so we will
2295                          * use the maxium timeout here for waiting the client
2296                          * sending the next req
2297                          */
2298                         to = max_t(time_t,
2299                                    at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2300                                    lustre_msg_get_timeout(req->rq_reqmsg));
2301                         /*
2302                          * Add 2 net_latency, one for balance rq_deadline
2303                          * (see ptl_send_rpc), one for resend the req to server,
2304                          * Note: client will pack net_latency in replay req
2305                          * (see ptlrpc_replay_req)
2306                          */
2307                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2308                 }
2309                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2310         }
2311         EXIT;
2312 }
2313
2314 /** Checking routines for recovery */
2315 static int check_for_recovery_ready(struct lu_target *lut)
2316 {
2317         struct obd_device *obd = lut->lut_obd;
2318         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2319
2320         CDEBUG(D_HA,
2321                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2322                clnts, obd->obd_stale_clients,
2323                atomic_read(&obd->obd_max_recoverable_clients),
2324                obd->obd_abort_recovery, obd->obd_recovery_expired);
2325
2326         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2327                 LASSERT(clnts <=
2328                         atomic_read(&obd->obd_max_recoverable_clients));
2329                 if (clnts + obd->obd_stale_clients <
2330                     atomic_read(&obd->obd_max_recoverable_clients))
2331                         return 0;
2332         }
2333
2334         if (lut->lut_tdtd != NULL) {
2335                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2336                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2337                         /*
2338                          * Let's extend recovery timer, in case the recovery
2339                          * timer expired, and some clients got evicted
2340                          */
2341                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2342                                               true);
2343                         CDEBUG(D_HA,
2344                                "%s update recovery is not ready, extend recovery %lu\n",
2345                                obd->obd_name, obd->obd_recovery_timeout);
2346                         return 0;
2347                 }
2348         }
2349
2350         return 1;
2351 }
2352
2353 enum {
2354         REQUEST_RECOVERY = 1,
2355         UPDATE_RECOVERY = 2,
2356 };
2357
2358 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2359 {
2360         __u64 transno = 0;
2361
2362         if (!list_empty(&obd->obd_req_replay_queue)) {
2363                 struct ptlrpc_request *req;
2364
2365                 req = list_entry(obd->obd_req_replay_queue.next,
2366                                  struct ptlrpc_request, rq_list);
2367                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2368         }
2369
2370         return transno;
2371 }
2372
2373 static __u64 get_next_transno(struct lu_target *lut, int *type)
2374 {
2375         struct obd_device *obd = lut->lut_obd;
2376         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2377         __u64 transno = 0;
2378         __u64 update_transno;
2379
2380         ENTRY;
2381
2382         transno = get_next_replay_req_transno(obd);
2383         if (type != NULL)
2384                 *type = REQUEST_RECOVERY;
2385
2386         if (tdtd == NULL)
2387                 RETURN(transno);
2388
2389         update_transno = distribute_txn_get_next_transno(tdtd);
2390         if (transno == 0 || (transno >= update_transno &&
2391                              update_transno != 0)) {
2392                 transno = update_transno;
2393                 if (type != NULL)
2394                         *type = UPDATE_RECOVERY;
2395         }
2396
2397         RETURN(transno);
2398 }
2399
2400 /**
2401  * drop duplicate replay request
2402  *
2403  * Because the operation has been replayed by update recovery, the request
2404  * with the same transno will be dropped and also notify the client to send
2405  * next replay request.
2406  *
2407  * \param[in] env       execution environment
2408  * \param[in] obd       failover obd device
2409  * \param[in] req       request to be dropped
2410  */
2411 static void drop_duplicate_replay_req(struct lu_env *env,
2412                                       struct obd_device *obd,
2413                                       struct ptlrpc_request *req)
2414 {
2415         DEBUG_REQ(D_HA, req,
2416                   "remove t%lld from %s because duplicate update records found",
2417                   lustre_msg_get_transno(req->rq_reqmsg),
2418                   libcfs_nid2str(req->rq_peer.nid));
2419
2420         /*
2421          * Right now, only for MDS reint operation update replay and
2422          * normal request replay can have the same transno
2423          */
2424         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2425                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2426                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2427                 if (likely(req->rq_export))
2428                         target_committed_to_req(req);
2429                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2430                 target_send_reply(req, req->rq_status, 0);
2431         } else {
2432                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2433                 libcfs_nid2str(req->rq_peer.nid));
2434         }
2435         target_exp_dequeue_req_replay(req);
2436         target_request_copy_put(req);
2437         obd->obd_replayed_requests++;
2438 }
2439
2440 static void replay_request_or_update(struct lu_env *env,
2441                                      struct lu_target *lut,
2442                                      struct target_recovery_data *trd,
2443                                      struct ptlrpc_thread *thread)
2444 {
2445         struct obd_device *obd = lut->lut_obd;
2446         struct ptlrpc_request *req = NULL;
2447         int type;
2448         __u64 transno;
2449
2450         ENTRY;
2451
2452         CDEBUG(D_HA, "Waiting for transno %lld\n",
2453                obd->obd_next_recovery_transno);
2454
2455         /* Replay all of request and update by transno */
2456         do {
2457                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2458
2459                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2460
2461                 /**
2462                  * It is needed to extend recovery window above
2463                  *  recovery_time_soft. Extending is possible only in the
2464                  *  end of recovery window (see more details in
2465                  *  handle_recovery_req()).
2466                  */
2467                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2468
2469                 if (target_recovery_overseer(lut, check_for_next_transno,
2470                                         exp_req_replay_healthy_or_from_mdt)) {
2471                         abort_req_replay_queue(obd);
2472                         abort_lock_replay_queue(obd);
2473                         goto abort;
2474                 }
2475
2476                 spin_lock(&obd->obd_recovery_task_lock);
2477                 transno = get_next_transno(lut, &type);
2478                 if (type == REQUEST_RECOVERY && transno != 0) {
2479                         /*
2480                          * Drop replay request from client side, if the
2481                          * replay has been executed by update with the
2482                          * same transno
2483                          */
2484                         req = list_entry(obd->obd_req_replay_queue.next,
2485                                         struct ptlrpc_request, rq_list);
2486
2487                         list_del_init(&req->rq_list);
2488                         obd->obd_requests_queued_for_recovery--;
2489                         spin_unlock(&obd->obd_recovery_task_lock);
2490
2491                         /*
2492                          * Let's check if the request has been redone by
2493                          * update replay
2494                          */
2495                         if (is_req_replayed_by_update(req)) {
2496                                 struct distribute_txn_replay_req *dtrq;
2497
2498                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2499                                                                       transno);
2500                                 LASSERT(dtrq != NULL);
2501                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2502                                 list_del_init(&dtrq->dtrq_list);
2503                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2504                                 dtrq_destroy(dtrq);
2505
2506                                 drop_duplicate_replay_req(env, obd, req);
2507
2508                                 continue;
2509                         }
2510
2511                         LASSERT(trd->trd_processing_task == current_pid());
2512                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2513                                   req->rq_xid,
2514                                   lustre_msg_get_transno(req->rq_reqmsg),
2515                                   libcfs_nid2str(req->rq_peer.nid));
2516
2517                         handle_recovery_req(thread, req,
2518                                             trd->trd_recovery_handler);
2519                         /**
2520                          * bz18031: increase next_recovery_transno before
2521                          * target_request_copy_put() will drop exp_rpc reference
2522                          */
2523                         spin_lock(&obd->obd_recovery_task_lock);
2524                         obd->obd_next_recovery_transno++;
2525                         spin_unlock(&obd->obd_recovery_task_lock);
2526                         target_exp_dequeue_req_replay(req);
2527                         target_request_copy_put(req);
2528                         obd->obd_replayed_requests++;
2529                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2530                         struct distribute_txn_replay_req *dtrq;
2531                         int rc;
2532
2533                         spin_unlock(&obd->obd_recovery_task_lock);
2534
2535                         LASSERT(tdtd != NULL);
2536                         dtrq = distribute_txn_get_next_req(tdtd);
2537                         lu_context_enter(&thread->t_env->le_ctx);
2538                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2539                         lu_context_exit(&thread->t_env->le_ctx);
2540                         extend_recovery_timer(obd, obd_timeout, true);
2541
2542                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2543                                 CDEBUG(D_HA,
2544                                        "Move x%llu t%llu to finish list\n",
2545                                        dtrq->dtrq_xid,
2546                                        dtrq->dtrq_master_transno);
2547
2548                                 /* Add it to the replay finish list */
2549                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2550                                 list_add(&dtrq->dtrq_list,
2551                                          &tdtd->tdtd_replay_finish_list);
2552                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2553
2554                                 spin_lock(&obd->obd_recovery_task_lock);
2555                                 if (transno == obd->obd_next_recovery_transno)
2556                                         obd->obd_next_recovery_transno++;
2557                                 else if (transno >
2558                                          obd->obd_next_recovery_transno)
2559                                         obd->obd_next_recovery_transno =
2560                                                                 transno + 1;
2561                                 spin_unlock(&obd->obd_recovery_task_lock);
2562                         } else {
2563                                 dtrq_destroy(dtrq);
2564                         }
2565                 } else {
2566                         spin_unlock(&obd->obd_recovery_task_lock);
2567 abort:
2568                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2569                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2570                         /** evict exports failed VBR */
2571                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2572                         break;
2573                 }
2574         } while (1);
2575 }
2576
2577 static int target_recovery_thread(void *arg)
2578 {
2579         struct lu_target *lut = arg;
2580         struct obd_device *obd = lut->lut_obd;
2581         struct ptlrpc_request *req;
2582         struct target_recovery_data *trd = &obd->obd_recovery_data;
2583         unsigned long delta;
2584         struct lu_env *env;
2585         struct ptlrpc_thread *thread = NULL;
2586         int rc = 0;
2587
2588         ENTRY;
2589
2590         unshare_fs_struct();
2591         OBD_ALLOC_PTR(thread);
2592         if (thread == NULL)
2593                 RETURN(-ENOMEM);
2594
2595         OBD_ALLOC_PTR(env);
2596         if (env == NULL)
2597                 GOTO(out_thread, rc = -ENOMEM);
2598         rc = lu_env_add(env);
2599         if (rc)
2600                 GOTO(out_env, rc);
2601
2602         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2603         if (rc)
2604                 GOTO(out_env_remove, rc);
2605
2606         thread->t_env = env;
2607         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2608         env->le_ctx.lc_thread = thread;
2609         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2610
2611         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2612                current_pid());
2613         trd->trd_processing_task = current_pid();
2614
2615         spin_lock(&obd->obd_dev_lock);
2616         obd->obd_recovering = 1;
2617         spin_unlock(&obd->obd_dev_lock);
2618         complete(&trd->trd_starting);
2619
2620         /* first of all, we have to know the first transno to replay */
2621         if (target_recovery_overseer(lut, check_for_recovery_ready,
2622                                      exp_connect_healthy)) {
2623                 abort_req_replay_queue(obd);
2624                 abort_lock_replay_queue(obd);
2625                 if (lut->lut_tdtd != NULL)
2626                         dtrq_list_destroy(lut->lut_tdtd);
2627         }
2628
2629         /* next stage: replay requests or update */
2630         delta = jiffies;
2631         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2632                atomic_read(&obd->obd_req_replay_clients),
2633                obd->obd_next_recovery_transno);
2634         replay_request_or_update(env, lut, trd, thread);
2635
2636         /**
2637          * The second stage: replay locks
2638          */
2639         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2640                atomic_read(&obd->obd_lock_replay_clients));
2641         while ((req = target_next_replay_lock(lut))) {
2642                 LASSERT(trd->trd_processing_task == current_pid());
2643                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2644                           libcfs_nid2str(req->rq_peer.nid));
2645                 handle_recovery_req(thread, req,
2646                                     trd->trd_recovery_handler);
2647                 target_request_copy_put(req);
2648                 obd->obd_replayed_locks++;
2649         }
2650
2651         /**
2652          * The third stage: reply on final pings, at this moment all clients
2653          * must have request in final queue
2654          */
2655         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2656         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2657         /** Update server last boot epoch */
2658         tgt_boot_epoch_update(lut);
2659         /*
2660          * We drop recoverying flag to forward all new requests
2661          * to regular mds_handle() since now
2662          */
2663         spin_lock(&obd->obd_dev_lock);
2664         obd->obd_recovering = obd->obd_abort_recovery = 0;
2665         spin_unlock(&obd->obd_dev_lock);
2666         spin_lock(&obd->obd_recovery_task_lock);
2667         target_cancel_recovery_timer(obd);
2668         spin_unlock(&obd->obd_recovery_task_lock);
2669         while ((req = target_next_final_ping(obd))) {
2670                 LASSERT(trd->trd_processing_task == current_pid());
2671                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2672                           libcfs_nid2str(req->rq_peer.nid));
2673                 handle_recovery_req(thread, req,
2674                                     trd->trd_recovery_handler);
2675                 /*
2676                  * Because the waiting client can not send ping to server,
2677                  * so we need refresh the last_request_time, to avoid the
2678                  * export is being evicted
2679                  */
2680                 ptlrpc_update_export_timer(req->rq_export, 0);
2681                 target_request_copy_put(req);
2682         }
2683
2684         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2685         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2686                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2687         if (delta > OBD_RECOVERY_TIME_SOFT) {
2688                 CWARN("too long recovery - read logs\n");
2689                 libcfs_debug_dumplog();
2690         }
2691
2692         target_finish_recovery(lut);
2693         lu_context_fini(&env->le_ctx);
2694         trd->trd_processing_task = 0;
2695         complete_all(&trd->trd_finishing);
2696         tgt_io_thread_done(thread);
2697 out_env_remove:
2698         lu_env_remove(env);
2699 out_env:
2700         OBD_FREE_PTR(env);
2701 out_thread:
2702         OBD_FREE_PTR(thread);
2703         RETURN(rc);
2704 }
2705
2706 static int target_start_recovery_thread(struct lu_target *lut,
2707                                         svc_handler_t handler)
2708 {
2709         struct obd_device *obd = lut->lut_obd;
2710         int rc = 0;
2711         struct target_recovery_data *trd = &obd->obd_recovery_data;
2712         int index;
2713
2714         memset(trd, 0, sizeof(*trd));
2715         init_completion(&trd->trd_starting);
2716         init_completion(&trd->trd_finishing);
2717         trd->trd_recovery_handler = handler;
2718
2719         rc = server_name2index(obd->obd_name, &index, NULL);
2720         if (rc < 0)
2721                 return rc;
2722
2723         if (!IS_ERR(kthread_run(target_recovery_thread,
2724                                 lut, "tgt_recover_%d", index))) {
2725                 wait_for_completion(&trd->trd_starting);
2726                 LASSERT(obd->obd_recovering != 0);
2727         } else {
2728                 rc = -ECHILD;
2729         }
2730
2731         return rc;
2732 }
2733
2734 void target_stop_recovery_thread(struct obd_device *obd)
2735 {
2736         if (obd->obd_recovery_data.trd_processing_task > 0) {
2737                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2738                 /** recovery can be done but postrecovery is not yet */
2739                 spin_lock(&obd->obd_dev_lock);
2740                 if (obd->obd_recovering) {
2741                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2742                         obd->obd_abort_recovery = 1;
2743                         wake_up(&obd->obd_next_transno_waitq);
2744                 }
2745                 spin_unlock(&obd->obd_dev_lock);
2746                 wait_for_completion(&trd->trd_finishing);
2747         }
2748 }
2749 EXPORT_SYMBOL(target_stop_recovery_thread);
2750
2751 void target_recovery_fini(struct obd_device *obd)
2752 {
2753         class_disconnect_exports(obd);
2754         target_stop_recovery_thread(obd);
2755         target_cleanup_recovery(obd);
2756 }
2757 EXPORT_SYMBOL(target_recovery_fini);
2758
2759 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2760 {
2761         struct obd_device *obd = container_of(timer, struct obd_device,
2762                                               obd_recovery_timer);
2763
2764         CDEBUG(D_HA,
2765                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2766                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2767                ktime_get_real_seconds() - obd->obd_recovery_start,
2768                atomic_read(&obd->obd_connected_clients));
2769
2770         obd->obd_recovery_expired = 1;
2771         wake_up(&obd->obd_next_transno_waitq);
2772         return HRTIMER_NORESTART;
2773 }
2774
2775 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2776 {
2777         struct obd_device *obd = lut->lut_obd;
2778
2779         if (lut->lut_bottom->dd_rdonly)
2780                 return;
2781
2782         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2783                 /** Update server last boot epoch */
2784                 tgt_boot_epoch_update(lut);
2785                 return;
2786         }
2787
2788         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2789                "last_transno %llu\n", obd->obd_name,
2790                atomic_read(&obd->obd_max_recoverable_clients),
2791                obd->obd_last_committed);
2792         LASSERT(obd->obd_stopping == 0);
2793         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2794         obd->obd_recovery_start = 0;
2795         obd->obd_recovery_end = 0;
2796
2797         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2798                      HRTIMER_MODE_ABS);
2799         obd->obd_recovery_timer.function = &target_recovery_expired;
2800         target_start_recovery_thread(lut, handler);
2801 }
2802 EXPORT_SYMBOL(target_recovery_init);
2803
2804 static int target_process_req_flags(struct obd_device *obd,
2805                                     struct ptlrpc_request *req)
2806 {
2807         struct obd_export *exp = req->rq_export;
2808
2809         LASSERT(exp != NULL);
2810         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2811                 /* client declares he's ready to replay locks */
2812                 spin_lock(&exp->exp_lock);
2813                 if (exp->exp_req_replay_needed) {
2814                         exp->exp_req_replay_needed = 0;
2815                         spin_unlock(&exp->exp_lock);
2816
2817                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2818                         atomic_dec(&obd->obd_req_replay_clients);
2819                 } else {
2820                         spin_unlock(&exp->exp_lock);
2821                 }
2822         }
2823         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2824                 /*
2825                  * client declares he's ready to complete recovery
2826                  * so, we put the request on th final queue
2827                  */
2828                 spin_lock(&exp->exp_lock);
2829                 if (exp->exp_lock_replay_needed) {
2830                         exp->exp_lock_replay_needed = 0;
2831                         spin_unlock(&exp->exp_lock);
2832
2833                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2834                         atomic_dec(&obd->obd_lock_replay_clients);
2835                 } else {
2836                         spin_unlock(&exp->exp_lock);
2837                 }
2838         }
2839         return 0;
2840 }
2841
2842 int target_queue_recovery_request(struct ptlrpc_request *req,
2843                                   struct obd_device *obd)
2844 {
2845         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2846         struct ptlrpc_request *reqiter;
2847         int inserted = 0;
2848
2849         ENTRY;
2850
2851         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2852                 /* Processing the queue right now, don't re-add. */
2853                 RETURN(1);
2854         }
2855
2856         target_process_req_flags(obd, req);
2857
2858         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2859                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2860                         if (cfs_fail_val == 1) {
2861                                 cfs_race_state = 1;
2862                                 cfs_fail_val = 0;
2863                                 wake_up(&cfs_race_waitq);
2864
2865                                 set_current_state(TASK_INTERRUPTIBLE);
2866                                 schedule_timeout(cfs_time_seconds(1));
2867                         }
2868                 }
2869
2870                 /*
2871                  * client declares he's ready to complete recovery
2872                  * so, we put the request on th final queue
2873                  */
2874                 target_request_copy_get(req);
2875                 DEBUG_REQ(D_HA, req, "queue final req");
2876                 wake_up(&obd->obd_next_transno_waitq);
2877                 spin_lock(&obd->obd_recovery_task_lock);
2878                 if (obd->obd_recovering) {
2879                         struct ptlrpc_request *tmp;
2880                         struct ptlrpc_request *duplicate = NULL;
2881
2882                         if (likely(!req->rq_export->exp_replay_done)) {
2883                                 req->rq_export->exp_replay_done = 1;
2884                                 list_add_tail(&req->rq_list,
2885                                               &obd->obd_final_req_queue);
2886                                 spin_unlock(&obd->obd_recovery_task_lock);
2887                                 RETURN(0);
2888                         }
2889
2890                         /*
2891                          * XXX O(n), but only happens if final ping is
2892                          * timed out, probably reorganize the list as
2893                          * a hash list later
2894                          */
2895                         list_for_each_entry_safe(reqiter, tmp,
2896                                                  &obd->obd_final_req_queue,
2897                                                  rq_list) {
2898                                 if (reqiter->rq_export == req->rq_export) {
2899                                         list_del_init(&reqiter->rq_list);
2900                                         duplicate = reqiter;
2901                                         break;
2902                                 }
2903                         }
2904
2905                         list_add_tail(&req->rq_list,
2906                                       &obd->obd_final_req_queue);
2907                         req->rq_export->exp_replay_done = 1;
2908                         spin_unlock(&obd->obd_recovery_task_lock);
2909
2910                         if (duplicate != NULL) {
2911                                 DEBUG_REQ(D_HA, duplicate,
2912                                           "put prev final req");
2913                                 target_request_copy_put(duplicate);
2914                         }
2915                         RETURN(0);
2916                 } else {
2917                         spin_unlock(&obd->obd_recovery_task_lock);
2918                         target_request_copy_put(req);
2919                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2920                 }
2921         }
2922         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2923                 /* client declares he's ready to replay locks */
2924                 target_request_copy_get(req);
2925                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2926                 wake_up(&obd->obd_next_transno_waitq);
2927                 spin_lock(&obd->obd_recovery_task_lock);
2928                 LASSERT(obd->obd_recovering);
2929                 /* usually due to recovery abort */
2930                 if (!req->rq_export->exp_in_recovery) {
2931                         spin_unlock(&obd->obd_recovery_task_lock);
2932                         target_request_copy_put(req);
2933                         RETURN(-ENOTCONN);
2934                 }
2935                 LASSERT(req->rq_export->exp_lock_replay_needed);
2936                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2937                 spin_unlock(&obd->obd_recovery_task_lock);
2938                 RETURN(0);
2939         }
2940
2941         /*
2942          * CAVEAT EMPTOR: The incoming request message has been swabbed
2943          * (i.e. buflens etc are in my own byte order), but type-dependent
2944          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2945          */
2946
2947         if (!transno) {
2948                 INIT_LIST_HEAD(&req->rq_list);
2949                 DEBUG_REQ(D_HA, req, "not queueing");
2950                 RETURN(1);
2951         }
2952
2953         /*
2954          * If we're processing the queue, we want don't want to queue this
2955          * message.
2956          *
2957          * Also, if this request has a transno less than the one we're waiting
2958          * for, we should process it now.  It could (and currently always will)
2959          * be an open request for a descriptor that was opened some time ago.
2960          *
2961          * Also, a resent, replayed request that has already been
2962          * handled will pass through here and be processed immediately.
2963          */
2964         CDEBUG(D_HA,
2965                "Next recovery transno: %llu, current: %llu, replaying\n",
2966                obd->obd_next_recovery_transno, transno);
2967
2968         /*
2969          * If the request has been replayed by update replay, then sends this
2970          * request to the recovery thread (replay_request_or_update()), where
2971          * it will be handled
2972          */
2973         spin_lock(&obd->obd_recovery_task_lock);
2974         if (transno < obd->obd_next_recovery_transno &&
2975             !is_req_replayed_by_update(req)) {
2976                 /* Processing the queue right now, don't re-add. */
2977                 LASSERT(list_empty(&req->rq_list));
2978                 spin_unlock(&obd->obd_recovery_task_lock);
2979                 RETURN(1);
2980         }
2981         spin_unlock(&obd->obd_recovery_task_lock);
2982
2983         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2984                 RETURN(0);
2985
2986         target_request_copy_get(req);
2987         if (!req->rq_export->exp_in_recovery) {
2988                 target_request_copy_put(req);
2989                 RETURN(-ENOTCONN);
2990         }
2991         LASSERT(req->rq_export->exp_req_replay_needed);
2992
2993         if (target_exp_enqueue_req_replay(req)) {
2994                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2995                 target_request_copy_put(req);
2996                 RETURN(0);
2997         }
2998
2999         /* XXX O(n^2) */
3000         spin_lock(&obd->obd_recovery_task_lock);
3001         LASSERT(obd->obd_recovering);
3002         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3003                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3004                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3005                         inserted = 1;
3006                         goto added;
3007                 }
3008
3009                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3010                              transno)) {
3011                         DEBUG_REQ(D_ERROR, req,
3012                                   "dropping replay: transno has been claimed by another client");
3013                         spin_unlock(&obd->obd_recovery_task_lock);
3014                         target_exp_dequeue_req_replay(req);
3015                         target_request_copy_put(req);
3016                         RETURN(0);
3017                 }
3018         }
3019 added:
3020         if (!inserted)
3021                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3022
3023         obd->obd_requests_queued_for_recovery++;
3024         spin_unlock(&obd->obd_recovery_task_lock);
3025         wake_up(&obd->obd_next_transno_waitq);
3026         RETURN(0);
3027 }
3028
3029 void target_committed_to_req(struct ptlrpc_request *req)
3030 {
3031         struct obd_export *exp = req->rq_export;
3032
3033         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3034                 lustre_msg_set_last_committed(req->rq_repmsg,
3035                                               exp->exp_last_committed);
3036         else
3037                 DEBUG_REQ(D_IOCTL, req,
3038                           "not sending last_committed update (%d/%d)",
3039                           exp->exp_obd->obd_no_transno,
3040                           req->rq_repmsg == NULL);
3041
3042         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3043                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3044 }
3045
3046 #endif /* HAVE_SERVER_SUPPORT */
3047
3048 /**
3049  * Packs current SLV and Limit into \a req.
3050  */
3051 int target_pack_pool_reply(struct ptlrpc_request *req)
3052 {
3053         struct obd_device *obd;
3054
3055         ENTRY;
3056
3057         /*
3058          * Check that we still have all structures alive as this may
3059          * be some late RPC at shutdown time.
3060          */
3061         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3062                      !exp_connect_lru_resize(req->rq_export))) {
3063                 lustre_msg_set_slv(req->rq_repmsg, 0);
3064                 lustre_msg_set_limit(req->rq_repmsg, 0);
3065                 RETURN(0);
3066         }
3067
3068         /* OBD is alive here as export is alive, which we checked above. */
3069         obd = req->rq_export->exp_obd;
3070
3071         read_lock(&obd->obd_pool_lock);
3072         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3073         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3074         read_unlock(&obd->obd_pool_lock);
3075
3076         RETURN(0);
3077 }
3078
3079 static int target_send_reply_msg(struct ptlrpc_request *req,
3080                                  int rc, int fail_id)
3081 {
3082         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3083                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3084                 return -ECOMM;
3085         }
3086         /*
3087          * We can have a null rq_reqmsg in the event of bad signature or
3088          * no context when unwrapping
3089          */
3090         if (req->rq_reqmsg &&
3091             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3092             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3093                 return -ECOMM;
3094
3095         if (unlikely(rc)) {
3096                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3097                 req->rq_status = rc;
3098                 return ptlrpc_send_error(req, 1);
3099         }
3100         DEBUG_REQ(D_NET, req, "sending reply");
3101
3102         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3103 }
3104
3105 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3106 {
3107         struct ptlrpc_service_part *svcpt;
3108         int netrc;
3109         struct ptlrpc_reply_state *rs;
3110         struct obd_export *exp;
3111
3112         ENTRY;
3113
3114         if (req->rq_no_reply) {
3115                 EXIT;
3116                 return;
3117         }
3118
3119         svcpt = req->rq_rqbd->rqbd_svcpt;
3120         rs = req->rq_reply_state;
3121         if (rs == NULL || !rs->rs_difficult) {
3122                 /* no notifiers */
3123                 target_send_reply_msg(req, rc, fail_id);
3124                 EXIT;
3125                 return;
3126         }
3127
3128         /* must be an export if locks saved */
3129         LASSERT(req->rq_export != NULL);
3130         /* req/reply consistent */
3131         LASSERT(rs->rs_svcpt == svcpt);
3132
3133         /* "fresh" reply */
3134         LASSERT(!rs->rs_scheduled);
3135         LASSERT(!rs->rs_scheduled_ever);
3136         LASSERT(!rs->rs_handled);
3137         LASSERT(!rs->rs_on_net);
3138         LASSERT(rs->rs_export == NULL);
3139         LASSERT(list_empty(&rs->rs_obd_list));
3140         LASSERT(list_empty(&rs->rs_exp_list));
3141
3142         exp = class_export_get(req->rq_export);
3143
3144         /* disable reply scheduling while I'm setting up */
3145         rs->rs_scheduled = 1;
3146         rs->rs_on_net    = 1;
3147         rs->rs_xid       = req->rq_xid;
3148         rs->rs_transno   = req->rq_transno;
3149         rs->rs_export    = exp;
3150         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3151
3152         spin_lock(&exp->exp_uncommitted_replies_lock);
3153         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3154                rs->rs_transno, exp->exp_last_committed);
3155         if (rs->rs_transno > exp->exp_last_committed) {
3156                 /* not committed already */
3157                 list_add_tail(&rs->rs_obd_list,
3158                                   &exp->exp_uncommitted_replies);
3159         }
3160         spin_unlock(&exp->exp_uncommitted_replies_lock);
3161
3162         spin_lock(&exp->exp_lock);
3163         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3164         spin_unlock(&exp->exp_lock);
3165
3166         netrc = target_send_reply_msg(req, rc, fail_id);
3167
3168         spin_lock(&svcpt->scp_rep_lock);
3169
3170         atomic_inc(&svcpt->scp_nreps_difficult);
3171
3172         if (netrc != 0) {
3173                 /*
3174                  * error sending: reply is off the net.  Also we need +1
3175                  * reply ref until ptlrpc_handle_rs() is done
3176                  * with the reply state (if the send was successful, there
3177                  * would have been +1 ref for the net, which
3178                  * reply_out_callback leaves alone)
3179                  */
3180                 rs->rs_on_net = 0;
3181                 ptlrpc_rs_addref(rs);
3182         }
3183
3184         spin_lock(&rs->rs_lock);
3185         if (rs->rs_transno <= exp->exp_last_committed ||
3186             (!rs->rs_on_net && !rs->rs_no_ack) ||
3187             list_empty(&rs->rs_exp_list) ||     /* completed already */
3188             list_empty(&rs->rs_obd_list)) {
3189                 CDEBUG(D_HA, "Schedule reply immediately\n");
3190                 ptlrpc_dispatch_difficult_reply(rs);
3191         } else {
3192                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3193                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3194         }
3195         spin_unlock(&rs->rs_lock);
3196         spin_unlock(&svcpt->scp_rep_lock);
3197         EXIT;
3198 }
3199
3200 enum ldlm_mode lck_compat_array[] = {
3201         [LCK_EX]    = LCK_COMPAT_EX,
3202         [LCK_PW]    = LCK_COMPAT_PW,
3203         [LCK_PR]    = LCK_COMPAT_PR,
3204         [LCK_CW]    = LCK_COMPAT_CW,
3205         [LCK_CR]    = LCK_COMPAT_CR,
3206         [LCK_NL]    = LCK_COMPAT_NL,
3207         [LCK_GROUP] = LCK_COMPAT_GROUP,
3208         [LCK_COS]   = LCK_COMPAT_COS,
3209 };
3210
3211 /**
3212  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3213  * not escape to the user level.
3214  */
3215 int ldlm_error2errno(enum ldlm_error error)
3216 {
3217         int result;
3218
3219         switch (error) {
3220         case ELDLM_OK:
3221         case ELDLM_LOCK_MATCHED:
3222                 result = 0;
3223                 break;
3224         case ELDLM_LOCK_CHANGED:
3225                 result = -ESTALE;
3226                 break;
3227         case ELDLM_LOCK_ABORTED:
3228                 result = -ENAVAIL;
3229                 break;
3230         case ELDLM_LOCK_REPLACED:
3231                 result = -ESRCH;
3232                 break;
3233         case ELDLM_NO_LOCK_DATA:
3234                 result = -ENOENT;
3235                 break;
3236         case ELDLM_NAMESPACE_EXISTS:
3237                 result = -EEXIST;
3238                 break;
3239         case ELDLM_BAD_NAMESPACE:
3240                 result = -EBADF;
3241                 break;
3242         default:
3243                 if (((int)error) < 0) { /* cast to signed type */
3244                         result = error; /* as ldlm_error can be unsigned */
3245                 } else {
3246                         CERROR("Invalid DLM result code: %d\n", error);
3247                         result = -EPROTO;
3248                 }
3249         }
3250         return result;
3251 }
3252 EXPORT_SYMBOL(ldlm_error2errno);
3253
3254 /**
3255  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3256  */
3257 enum ldlm_error ldlm_errno2error(int err_no)
3258 {
3259         int error;
3260
3261         switch (err_no) {
3262         case 0:
3263                 error = ELDLM_OK;
3264                 break;
3265         case -ESTALE:
3266                 error = ELDLM_LOCK_CHANGED;
3267                 break;
3268         case -ENAVAIL:
3269                 error = ELDLM_LOCK_ABORTED;
3270                 break;
3271         case -ESRCH:
3272                 error = ELDLM_LOCK_REPLACED;
3273                 break;
3274         case -ENOENT:
3275                 error = ELDLM_NO_LOCK_DATA;
3276                 break;
3277         case -EEXIST:
3278                 error = ELDLM_NAMESPACE_EXISTS;
3279                 break;
3280         case -EBADF:
3281                 error = ELDLM_BAD_NAMESPACE;
3282                 break;
3283         default:
3284                 error = err_no;
3285         }
3286         return error;
3287 }
3288
3289 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3290 void ldlm_dump_export_locks(struct obd_export *exp)
3291 {
3292         spin_lock(&exp->exp_locks_list_guard);
3293         if (!list_empty(&exp->exp_locks_list)) {
3294                 struct ldlm_lock *lock;
3295
3296                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3297                        exp);
3298                 list_for_each_entry(lock, &exp->exp_locks_list,
3299                                         l_exp_refs_link)
3300                         LDLM_ERROR(lock, "lock:");
3301         }
3302         spin_unlock(&exp->exp_locks_list_guard);
3303 }
3304 #endif
3305
3306 #ifdef HAVE_SERVER_SUPPORT
3307 static int target_bulk_timeout(void *data)
3308 {
3309         ENTRY;
3310         /*
3311          * We don't fail the connection here, because having the export
3312          * killed makes the (vital) call to commitrw very sad.
3313          */
3314         RETURN(1);
3315 }
3316
3317 static inline const char *bulk2type(struct ptlrpc_request *req)
3318 {
3319         if (req->rq_bulk_read)
3320                 return "READ";
3321         if (req->rq_bulk_write)
3322                 return "WRITE";
3323         return "UNKNOWN";
3324 }
3325
3326 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3327 {
3328         struct ptlrpc_request *req = desc->bd_req;
3329         time64_t start = ktime_get_seconds();
3330         time64_t deadline;
3331         struct l_wait_info lwi;
3332         int rc = 0;
3333
3334         ENTRY;
3335
3336         /* If there is eviction in progress, wait for it to finish. */
3337         wait_event_idle(
3338                 exp->exp_obd->obd_evict_inprogress_waitq,
3339                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3340
3341         /* Check if client was evicted or reconnected already. */
3342         if (exp->exp_failed ||
3343             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3344                 rc = -ENOTCONN;
3345         } else {
3346                 if (req->rq_bulk_read)
3347                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3348
3349                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3350                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3351                 else /* old version, bulk matchbits is rq_xid */
3352                         req->rq_mbits = req->rq_xid;
3353
3354                 if (rc == 0)
3355                         rc = ptlrpc_start_bulk_transfer(desc);
3356         }
3357
3358         if (rc < 0) {
3359                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3360                           bulk2type(req), rc);
3361                 RETURN(rc);
3362         }
3363
3364         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3365                 ptlrpc_abort_bulk(desc);
3366                 RETURN(0);
3367         }
3368
3369         /* limit actual bulk transfer to bulk_timeout seconds */
3370         deadline = start + bulk_timeout;
3371         if (deadline > req->rq_deadline)
3372                 deadline = req->rq_deadline;
3373
3374         do {
3375                 time64_t timeoutl = deadline - ktime_get_seconds();
3376                 long timeout_jiffies = timeoutl <= 0 ?
3377                                        1 : cfs_time_seconds(timeoutl);
3378                 time64_t rq_deadline;
3379
3380                 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3381                                            cfs_time_seconds(1),
3382                                            target_bulk_timeout, desc);
3383                 rc = l_wait_event(desc->bd_waitq,
3384                                   !ptlrpc_server_bulk_active(desc) ||
3385                                   exp->exp_failed ||
3386                                   exp->exp_conn_cnt >
3387                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3388                                   &lwi);
3389                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3390                 /* Wait again if we changed rq_deadline. */
3391                 rq_deadline = READ_ONCE(req->rq_deadline);
3392                 deadline = start + bulk_timeout;
3393                 if (deadline > rq_deadline)
3394                         deadline = rq_deadline;
3395         } while (rc == -ETIMEDOUT &&
3396                  deadline > ktime_get_seconds());
3397
3398         if (rc == -ETIMEDOUT) {
3399                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3400                           bulk2type(req), deadline - start,
3401                           ktime_get_real_seconds() - deadline);
3402                 ptlrpc_abort_bulk(desc);
3403         } else if (exp->exp_failed) {
3404                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3405                           bulk2type(req));
3406                 rc = -ENOTCONN;
3407                 ptlrpc_abort_bulk(desc);
3408         } else if (exp->exp_conn_cnt >
3409                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3410                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3411                           bulk2type(req));
3412                 /* We don't reply anyway. */
3413                 rc = -ETIMEDOUT;
3414                 ptlrpc_abort_bulk(desc);
3415         } else if (desc->bd_failure) {
3416                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3417                           bulk2type(req));
3418                 /* XXX should this be a different errno? */
3419                 rc = -ETIMEDOUT;
3420         } else {
3421                 if (req->rq_bulk_write)
3422                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3423                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3424                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3425                                   bulk2type(req), desc->bd_nob_transferred,
3426                                   desc->bd_nob);
3427                         /* XXX should this be a different errno? */
3428                         rc = -ETIMEDOUT;
3429                 }
3430         }
3431
3432         RETURN(rc);
3433 }
3434 EXPORT_SYMBOL(target_bulk_io);
3435
3436 #endif /* HAVE_SERVER_SUPPORT */