lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 /**
  34  * This file deals with various client/target related logic including recovery.
  35  *
  36  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  37  * should be moved.
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #include <cl_object.h>
  43 #include <linux/jiffies.h>
  44 #include <linux/kernel.h>
  45 #include <linux/kthread.h>
  46 #include <libcfs/libcfs.h>
  47 #include <obd.h>
  48 #include <obd_class.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /*
  55  * @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         lnet_nid_t nid4refnet = LNET_NID_ANY;
  64         int rc = 0;
  65
  66         ENTRY;
  67
  68         if (!create && !priority) {
  69                 CDEBUG(D_HA, "Nothing to do\n");
  70                 RETURN(-EINVAL);
  71         }
  72
  73         if (imp->imp_connection &&
  74             imp->imp_connection->c_remote_uuid.uuid[0] == 0)
  75                 /* nid4refnet is used to restrict network connections */
  76                 nid4refnet = imp->imp_connection->c_self;
  77         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, nid4refnet);
  78         if (!ptlrpc_conn) {
  79                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  80                 RETURN(-ENOENT);
  81         }
  82
  83         if (create) {
  84                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  85                 if (!imp_conn)
  86                         GOTO(out_put, rc = -ENOMEM);
  87         }
  88
  89         spin_lock(&imp->imp_lock);
  90         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  91                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  92                         if (priority) {
  93                                 list_move(&item->oic_item,
  94                                           &imp->imp_conn_list);
  95                                 item->oic_last_attempt = 0;
  96                         }
  97                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  98                                imp, imp->imp_obd->obd_name, uuid->uuid,
  99                                (priority ? ", moved to head" : ""));
 100                         spin_unlock(&imp->imp_lock);
 101                         GOTO(out_free, rc = 0);
 102                 }
 103         }
 104         /* No existing import connection found for \a uuid. */
 105         if (create) {
 106                 imp_conn->oic_conn = ptlrpc_conn;
 107                 imp_conn->oic_uuid = *uuid;
 108                 imp_conn->oic_last_attempt = 0;
 109                 if (priority)
 110                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 111                 else
 112                         list_add_tail(&imp_conn->oic_item,
 113                                       &imp->imp_conn_list);
 114                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 115                        imp, imp->imp_obd->obd_name, uuid->uuid,
 116                        (priority ? "head" : "tail"));
 117         } else {
 118                 spin_unlock(&imp->imp_lock);
 119                 GOTO(out_free, rc = -ENOENT);
 120         }
 121
 122         spin_unlock(&imp->imp_lock);
 123         RETURN(0);
 124 out_free:
 125         if (imp_conn)
 126                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 127 out_put:
 128         ptlrpc_connection_put(ptlrpc_conn);
 129         RETURN(rc);
 130 }
 131
 132 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 133 {
 134         return import_set_conn(imp, uuid, 1, 0);
 135 }
 136
 137 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 138                            int priority)
 139 {
 140         return import_set_conn(imp, uuid, priority, 1);
 141 }
 142 EXPORT_SYMBOL(client_import_add_conn);
 143
 144 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 145 {
 146         struct obd_import_conn *imp_conn;
 147         struct obd_export *dlmexp;
 148         int rc = -ENOENT;
 149
 150         ENTRY;
 151
 152         spin_lock(&imp->imp_lock);
 153         if (list_empty(&imp->imp_conn_list)) {
 154                 LASSERT(!imp->imp_connection);
 155                 GOTO(out, rc);
 156         }
 157
 158         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 159                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 160                         continue;
 161                 LASSERT(imp_conn->oic_conn);
 162
 163                 if (imp_conn == imp->imp_conn_current) {
 164                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 165
 166                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 167                             imp->imp_state != LUSTRE_IMP_DISCON) {
 168                                 CERROR("can't remove current connection\n");
 169                                 GOTO(out, rc = -EBUSY);
 170                         }
 171
 172                         ptlrpc_connection_put(imp->imp_connection);
 173                         imp->imp_connection = NULL;
 174
 175                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 176                         if (dlmexp && dlmexp->exp_connection) {
 177                                 LASSERT(dlmexp->exp_connection ==
 178                                         imp_conn->oic_conn);
 179                                 ptlrpc_connection_put(dlmexp->exp_connection);
 180                                 dlmexp->exp_connection = NULL;
 181                         }
 182
 183                         if (dlmexp != NULL)
 184                                 class_export_put(dlmexp);
 185                 }
 186
 187                 list_del(&imp_conn->oic_item);
 188                 ptlrpc_connection_put(imp_conn->oic_conn);
 189                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 190                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 191                        imp, imp->imp_obd->obd_name, uuid->uuid);
 192                 rc = 0;
 193                 break;
 194         }
 195 out:
 196         spin_unlock(&imp->imp_lock);
 197         if (rc == -ENOENT)
 198                 CERROR("connection %s not found\n", uuid->uuid);
 199         RETURN(rc);
 200 }
 201 EXPORT_SYMBOL(client_import_del_conn);
 202
 203 /**
 204  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 205  * to find a conn uuid of \a imp which can reach \a peer.
 206  */
 207 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 208                             struct obd_uuid *uuid)
 209 {
 210         struct obd_import_conn *conn;
 211         int rc = -ENOENT;
 212
 213         ENTRY;
 214
 215         spin_lock(&imp->imp_lock);
 216         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 217                 /* Check if conn UUID does have this peer NID. */
 218                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 219                         *uuid = conn->oic_uuid;
 220                         rc = 0;
 221                         break;
 222                 }
 223         }
 224         spin_unlock(&imp->imp_lock);
 225         RETURN(rc);
 226 }
 227 EXPORT_SYMBOL(client_import_find_conn);
 228
 229 void client_destroy_import(struct obd_import *imp)
 230 {
 231         /*
 232          * Drop security policy instance after all RPCs have finished/aborted
 233          * to let all busy contexts be released.
 234          */
 235         class_import_get(imp);
 236         class_destroy_import(imp);
 237         sptlrpc_import_sec_put(imp);
 238         class_import_put(imp);
 239 }
 240 EXPORT_SYMBOL(client_destroy_import);
 241
 242 /**
 243  * Check whether or not the OSC is on MDT.
 244  * In the config log,
 245  * osc on MDT
 246  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 247  * osc on client
 248  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 249  *
 250  **/
 251 static int osc_on_mdt(char *obdname)
 252 {
 253         char *ptr;
 254
 255         ptr = strrchr(obdname, '-');
 256         if (ptr == NULL)
 257                 return 0;
 258
 259         if (strncmp(ptr + 1, "MDT", 3) == 0)
 260                 return 1;
 261
 262         return 0;
 263 }
 264
 265 /*
 266  * Configure an RPC client OBD device.
 267  *
 268  * lcfg parameters:
 269  * 1 - client UUID
 270  * 2 - server UUID
 271  * 3 - inactive-on-startup
 272  * 4 - restrictive net
 273  */
 274 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 275 {
 276         struct client_obd *cli = &obddev->u.cli;
 277         struct obd_import *imp;
 278         struct obd_uuid server_uuid;
 279         int rq_portal, rp_portal, connect_op;
 280         const char *name = obddev->obd_type->typ_name;
 281         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 282         char *cli_name = lustre_cfg_buf(lcfg, 0);
 283         struct ptlrpc_connection fake_conn = { .c_self = 0,
 284                                                .c_remote_uuid.uuid[0] = 0 };
 285         int rc;
 286
 287         ENTRY;
 288
 289         /*
 290          * In a more perfect world, we would hang a ptlrpc_client off of
 291          * obd_type and just use the values from there.
 292          */
 293         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 294                 rq_portal = OST_REQUEST_PORTAL;
 295                 rp_portal = OSC_REPLY_PORTAL;
 296                 connect_op = OST_CONNECT;
 297                 cli->cl_sp_me = LUSTRE_SP_CLI;
 298                 cli->cl_sp_to = LUSTRE_SP_OST;
 299                 ns_type = LDLM_NS_TYPE_OSC;
 300         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 301                    !strcmp(name, LUSTRE_LWP_NAME)) {
 302                 rq_portal = MDS_REQUEST_PORTAL;
 303                 rp_portal = MDC_REPLY_PORTAL;
 304                 connect_op = MDS_CONNECT;
 305                 if (is_lwp_on_ost(cli_name))
 306                         cli->cl_sp_me = LUSTRE_SP_OST;
 307                 else if (is_lwp_on_mdt(cli_name))
 308                         cli->cl_sp_me = LUSTRE_SP_MDT;
 309                 else
 310                         cli->cl_sp_me = LUSTRE_SP_CLI;
 311                 cli->cl_sp_to = LUSTRE_SP_MDT;
 312                 ns_type = LDLM_NS_TYPE_MDC;
 313         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 314                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 315                         /* OSP_on_MDT for other MDTs */
 316                         connect_op = MDS_CONNECT;
 317                         cli->cl_sp_to = LUSTRE_SP_MDT;
 318                         ns_type = LDLM_NS_TYPE_MDC;
 319                         rq_portal = OUT_PORTAL;
 320                 } else {
 321                         /* OSP on MDT for OST */
 322                         connect_op = OST_CONNECT;
 323                         cli->cl_sp_to = LUSTRE_SP_OST;
 324                         ns_type = LDLM_NS_TYPE_OSC;
 325                         rq_portal = OST_REQUEST_PORTAL;
 326                 }
 327                 rp_portal = OSC_REPLY_PORTAL;
 328                 cli->cl_sp_me = LUSTRE_SP_MDT;
 329         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 330                 rq_portal = MGS_REQUEST_PORTAL;
 331                 rp_portal = MGC_REPLY_PORTAL;
 332                 connect_op = MGS_CONNECT;
 333                 cli->cl_sp_me = LUSTRE_SP_MGC;
 334                 cli->cl_sp_to = LUSTRE_SP_MGS;
 335                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 336                 ns_type = LDLM_NS_TYPE_MGC;
 337         } else {
 338                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 339                        name);
 340                 RETURN(-EINVAL);
 341         }
 342
 343         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 344                 CERROR("requires a TARGET UUID\n");
 345                 RETURN(-EINVAL);
 346         }
 347
 348         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 349                 CERROR("client UUID must be less than 38 characters\n");
 350                 RETURN(-EINVAL);
 351         }
 352
 353         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 354                 CERROR("setup requires a SERVER UUID\n");
 355                 RETURN(-EINVAL);
 356         }
 357
 358         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 359                 CERROR("target UUID must be less than 38 characters\n");
 360                 RETURN(-EINVAL);
 361         }
 362
 363         init_rwsem(&cli->cl_sem);
 364         mutex_init(&cli->cl_mgc_mutex);
 365         cli->cl_seq = NULL;
 366         init_rwsem(&cli->cl_seq_rwsem);
 367         cli->cl_conn_count = 0;
 368         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 369                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 370                      sizeof(server_uuid)));
 371
 372         cli->cl_dirty_pages = 0;
 373         cli->cl_dirty_max_pages = 0;
 374         cli->cl_avail_grant = 0;
 375         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 376         /*
 377          * cl_dirty_max_pages may be changed at connect time in
 378          * ptlrpc_connect_interpret().
 379          */
 380         client_adjust_max_dirty(cli);
 381         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 382         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 383         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 384         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 385         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 386         spin_lock_init(&cli->cl_loi_list_lock);
 387         atomic_set(&cli->cl_pending_w_pages, 0);
 388         atomic_set(&cli->cl_pending_r_pages, 0);
 389         cli->cl_r_in_flight = 0;
 390         cli->cl_w_in_flight = 0;
 391
 392         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 393         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 394         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 395         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 396         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 397         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 398
 399         /* lru for osc. */
 400         INIT_LIST_HEAD(&cli->cl_lru_osc);
 401         atomic_set(&cli->cl_lru_shrinkers, 0);
 402         atomic_long_set(&cli->cl_lru_busy, 0);
 403         atomic_long_set(&cli->cl_lru_in_list, 0);
 404         INIT_LIST_HEAD(&cli->cl_lru_list);
 405         spin_lock_init(&cli->cl_lru_list_lock);
 406         atomic_long_set(&cli->cl_unstable_count, 0);
 407         INIT_LIST_HEAD(&cli->cl_shrink_list);
 408         INIT_LIST_HEAD(&cli->cl_grant_chain);
 409
 410         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 411         cli->cl_rpcs_in_flight = 0;
 412
 413         init_waitqueue_head(&cli->cl_destroy_waitq);
 414         atomic_set(&cli->cl_destroy_in_flight, 0);
 415
 416
 417         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 418         cli->cl_preferred_cksum_type = 0;
 419 #ifdef ENABLE_CHECKSUM
 420         /* Turn on checksumming by default. */
 421         cli->cl_checksum = 1;
 422         /*
 423          * The supported checksum types will be worked out at connect time
 424          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 425          * through procfs.
 426          */
 427         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 428 #endif
 429         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 430
 431         /*
 432          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 433          * from OFD after connecting.
 434          */
 435         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 436
 437         cli->cl_max_short_io_bytes = OBD_MAX_SHORT_IO_BYTES;
 438
 439         /*
 440          * set cl_chunkbits default value to PAGE_SHIFT,
 441          * it will be updated at OSC connection time.
 442          */
 443         cli->cl_chunkbits = PAGE_SHIFT;
 444
 445         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 446                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 447         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 448                 cli->cl_max_rpcs_in_flight = 2;
 449         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 450                 cli->cl_max_rpcs_in_flight = 3;
 451         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 452                 cli->cl_max_rpcs_in_flight = 4;
 453         } else {
 454                 if (osc_on_mdt(obddev->obd_name))
 455                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 456                 else
 457                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 458         }
 459
 460         spin_lock_init(&cli->cl_mod_rpcs_lock);
 461         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 462         cli->cl_max_mod_rpcs_in_flight = 0;
 463         cli->cl_mod_rpcs_in_flight = 0;
 464         cli->cl_close_rpcs_in_flight = 0;
 465         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 466         cli->cl_mod_tag_bitmap = NULL;
 467
 468         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 469
 470         if (connect_op == MDS_CONNECT) {
 471                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 472                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 473                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 474                 if (cli->cl_mod_tag_bitmap == NULL)
 475                         GOTO(err, rc = -ENOMEM);
 476         }
 477
 478         rc = ldlm_get_ref();
 479         if (rc) {
 480                 CERROR("ldlm_get_ref failed: %d\n", rc);
 481                 GOTO(err, rc);
 482         }
 483
 484         ptlrpc_init_client(rq_portal, rp_portal, name,
 485                            &obddev->obd_ldlm_client);
 486
 487         imp = class_new_import(obddev);
 488         if (imp == NULL)
 489                 GOTO(err_ldlm, rc = -ENOENT);
 490         imp->imp_client = &obddev->obd_ldlm_client;
 491         imp->imp_connect_op = connect_op;
 492         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 493                LUSTRE_CFG_BUFLEN(lcfg, 1));
 494         class_import_put(imp);
 495
 496         if (lustre_cfg_buf(lcfg, 4)) {
 497                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 498
 499                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 500                         rc = -EINVAL;
 501                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 502                                obddev->obd_name, lustre_cfg_string(lcfg, 4),
 503                                rc);
 504                         GOTO(err_import, rc);
 505                 }
 506                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 507                 imp->imp_connection = &fake_conn;
 508         }
 509
 510         rc = client_import_add_conn(imp, &server_uuid, 1);
 511         if (rc) {
 512                 CERROR("can't add initial connection\n");
 513                 GOTO(err_import, rc);
 514         }
 515         imp->imp_connection = NULL;
 516
 517         cli->cl_import = imp;
 518         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 519         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 520
 521         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 522                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 523                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 524                                name, obddev->obd_name,
 525                                cli->cl_target_uuid.uuid);
 526                         spin_lock(&imp->imp_lock);
 527                         imp->imp_deactive = 1;
 528                         spin_unlock(&imp->imp_lock);
 529                 }
 530         }
 531
 532         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 533                                                    LDLM_NAMESPACE_CLIENT,
 534                                                    LDLM_NAMESPACE_GREEDY,
 535                                                    ns_type);
 536         if (obddev->obd_namespace == NULL) {
 537                 CERROR("Unable to create client namespace - %s\n",
 538                        obddev->obd_name);
 539                 GOTO(err_import, rc = -ENOMEM);
 540         }
 541
 542         RETURN(rc);
 543
 544 err_import:
 545         class_destroy_import(imp);
 546 err_ldlm:
 547         ldlm_put_ref();
 548 err:
 549         if (cli->cl_mod_tag_bitmap != NULL)
 550                 OBD_FREE(cli->cl_mod_tag_bitmap,
 551                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 552         cli->cl_mod_tag_bitmap = NULL;
 553         RETURN(rc);
 554
 555 }
 556 EXPORT_SYMBOL(client_obd_setup);
 557
 558 int client_obd_cleanup(struct obd_device *obddev)
 559 {
 560         struct client_obd *cli = &obddev->u.cli;
 561
 562         ENTRY;
 563
 564         ldlm_namespace_free_post(obddev->obd_namespace);
 565         obddev->obd_namespace = NULL;
 566
 567         obd_cleanup_client_import(obddev);
 568         LASSERT(obddev->u.cli.cl_import == NULL);
 569
 570         ldlm_put_ref();
 571
 572         if (cli->cl_mod_tag_bitmap != NULL)
 573                 OBD_FREE(cli->cl_mod_tag_bitmap,
 574                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 575         cli->cl_mod_tag_bitmap = NULL;
 576
 577         RETURN(0);
 578 }
 579 EXPORT_SYMBOL(client_obd_cleanup);
 580
 581 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 582 int client_connect_import(const struct lu_env *env,
 583                           struct obd_export **exp,
 584                           struct obd_device *obd, struct obd_uuid *cluuid,
 585                           struct obd_connect_data *data, void *localdata)
 586 {
 587         struct client_obd *cli = &obd->u.cli;
 588         struct obd_import *imp = cli->cl_import;
 589         struct obd_connect_data *ocd;
 590         struct lustre_handle conn = { 0 };
 591         int rc;
 592
 593         ENTRY;
 594
 595         *exp = NULL;
 596         down_write(&cli->cl_sem);
 597         if (cli->cl_conn_count > 0)
 598                 GOTO(out_sem, rc = -EALREADY);
 599
 600         rc = class_connect(&conn, obd, cluuid);
 601         if (rc)
 602                 GOTO(out_sem, rc);
 603
 604         cli->cl_conn_count++;
 605         *exp = class_conn2export(&conn);
 606
 607         LASSERT(obd->obd_namespace);
 608
 609         imp->imp_dlm_handle = conn;
 610         rc = ptlrpc_init_import(imp);
 611         if (rc != 0)
 612                 GOTO(out_ldlm, rc);
 613
 614         ocd = &imp->imp_connect_data;
 615         if (data) {
 616                 *ocd = *data;
 617                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 618                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 619         }
 620
 621         rc = ptlrpc_connect_import(imp);
 622         if (rc != 0) {
 623                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 624                 GOTO(out_ldlm, rc);
 625         }
 626         LASSERT(*exp != NULL && (*exp)->exp_connection);
 627
 628         if (data) {
 629                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 630                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 631                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 632                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 633                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 634         }
 635
 636         ptlrpc_pinger_add_import(imp);
 637
 638         EXIT;
 639
 640         if (rc) {
 641 out_ldlm:
 642                 cli->cl_conn_count--;
 643                 class_disconnect(*exp);
 644                 *exp = NULL;
 645         }
 646 out_sem:
 647         up_write(&cli->cl_sem);
 648
 649         if (!rc && localdata) {
 650                 LASSERT(cli->cl_cache == NULL); /* only once */
 651                 cli->cl_cache = (struct cl_client_cache *)localdata;
 652                 cl_cache_incref(cli->cl_cache);
 653                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 654
 655                 /* add this osc into entity list */
 656                 LASSERT(list_empty(&cli->cl_lru_osc));
 657                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 658                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 659                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 660         }
 661
 662         return rc;
 663 }
 664 EXPORT_SYMBOL(client_connect_import);
 665
 666 int client_disconnect_export(struct obd_export *exp)
 667 {
 668         struct obd_device *obd = class_exp2obd(exp);
 669         struct client_obd *cli;
 670         struct obd_import *imp;
 671         int rc = 0, err;
 672
 673         ENTRY;
 674
 675         if (!obd) {
 676                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 677                        exp, exp ? exp->exp_handle.h_cookie : -1);
 678                 RETURN(-EINVAL);
 679         }
 680
 681         cli = &obd->u.cli;
 682         imp = cli->cl_import;
 683
 684         down_write(&cli->cl_sem);
 685         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 686                 cli->cl_conn_count);
 687
 688         if (cli->cl_conn_count == 0) {
 689                 CERROR("disconnecting disconnected device (%s)\n",
 690                        obd->obd_name);
 691                 GOTO(out_disconnect, rc = -EINVAL);
 692         }
 693
 694         cli->cl_conn_count--;
 695         if (cli->cl_conn_count != 0)
 696                 GOTO(out_disconnect, rc = 0);
 697
 698         /*
 699          * Mark import deactivated now, so we don't try to reconnect if any
 700          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 701          * fully deactivate the import, or that would drop all requests.
 702          */
 703         spin_lock(&imp->imp_lock);
 704         imp->imp_deactive = 1;
 705         spin_unlock(&imp->imp_lock);
 706
 707         /*
 708          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 709          * delete it regardless.  (It's safe to delete an import that was
 710          * never added.)
 711          */
 712         (void)ptlrpc_pinger_del_import(imp);
 713
 714         if (obd->obd_namespace != NULL) {
 715                 /* obd_force == local only */
 716                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 717                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 718                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 719                                           obd->obd_force);
 720         }
 721
 722         /*
 723          * There's no need to hold sem while disconnecting an import,
 724          * and it may actually cause deadlock in GSS.
 725          */
 726         up_write(&cli->cl_sem);
 727         rc = ptlrpc_disconnect_import(imp, 0);
 728         down_write(&cli->cl_sem);
 729
 730         ptlrpc_invalidate_import(imp);
 731
 732         EXIT;
 733
 734 out_disconnect:
 735         /*
 736          * Use server style - class_disconnect should be always called for
 737          * o_disconnect.
 738          */
 739         err = class_disconnect(exp);
 740         if (!rc && err)
 741                 rc = err;
 742
 743         up_write(&cli->cl_sem);
 744
 745         RETURN(rc);
 746 }
 747 EXPORT_SYMBOL(client_disconnect_export);
 748
 749 #ifdef HAVE_SERVER_SUPPORT
 750 int server_disconnect_export(struct obd_export *exp)
 751 {
 752         int rc;
 753
 754         ENTRY;
 755
 756         /* Disconnect early so that clients can't keep using export. */
 757         rc = class_disconnect(exp);
 758         /* Close import to avoid sending any requests. */
 759         if (exp->exp_imp_reverse)
 760                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 761
 762         ldlm_bl_thread_wakeup();
 763
 764         /* complete all outstanding replies */
 765         spin_lock(&exp->exp_lock);
 766         while (!list_empty(&exp->exp_outstanding_replies)) {
 767                 struct ptlrpc_reply_state *rs =
 768                         list_entry(exp->exp_outstanding_replies.next,
 769                                        struct ptlrpc_reply_state, rs_exp_list);
 770                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 771
 772                 spin_lock(&svcpt->scp_rep_lock);
 773
 774                 list_del_init(&rs->rs_exp_list);
 775
 776                 spin_lock(&rs->rs_lock);
 777                 /* clear rs_convert_lock to make sure rs is handled and put */
 778                 rs->rs_convert_lock = 0;
 779                 ptlrpc_schedule_difficult_reply(rs);
 780                 spin_unlock(&rs->rs_lock);
 781
 782                 spin_unlock(&svcpt->scp_rep_lock);
 783         }
 784         spin_unlock(&exp->exp_lock);
 785
 786         RETURN(rc);
 787 }
 788 EXPORT_SYMBOL(server_disconnect_export);
 789
 790 /*
 791  * --------------------------------------------------------------------------
 792  * from old lib/target.c
 793  * --------------------------------------------------------------------------
 794  */
 795
 796 static int target_handle_reconnect(struct lustre_handle *conn,
 797                                    struct obd_export *exp,
 798                                    struct obd_uuid *cluuid)
 799 {
 800         struct obd_device *target;
 801         struct lustre_handle *hdl;
 802         ktime_t remaining;
 803         s64 timeout;
 804         int rc = 0;
 805
 806         ENTRY;
 807         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 808         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 809                 conn->cookie = exp->exp_handle.h_cookie;
 810                 CDEBUG(D_HA,
 811                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 812                        cluuid->uuid, exp, conn->cookie);
 813                 RETURN(0);
 814         }
 815
 816         target = exp->exp_obd;
 817
 818         /* Might be a re-connect after a partition. */
 819         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 820                 LCONSOLE_WARN("%s: already connected client %s (at %s) with handle %#llx. Rejecting client with the same UUID trying to reconnect with handle %#llx\n",
 821                               target->obd_name,
 822                               obd_uuid2str(&exp->exp_client_uuid),
 823                               obd_export_nid2str(exp),
 824                               hdl->cookie, conn->cookie);
 825                 memset(conn, 0, sizeof(*conn));
 826                 /*
 827                  * target_handle_connect() treats EALREADY and
 828                  * -EALREADY differently.  -EALREADY is an error
 829                  * (same UUID, different handle).
 830                  */
 831                 RETURN(-EALREADY);
 832         }
 833
 834         if (!target->obd_recovering) {
 835                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 836                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 837                         obd_export_nid2str(exp));
 838                 GOTO(out_already, rc);
 839         }
 840
 841         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 842         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 843         if (timeout > 0) {
 844                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 845                               target->obd_name,
 846                               obd_uuid2str(&exp->exp_client_uuid),
 847                               obd_export_nid2str(exp),
 848                               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 = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1180         if (!export)
1181                 goto no_export;
1182
1183         /* We've found an export in the hash. */
1184
1185         spin_lock(&export->exp_lock);
1186
1187         if (export->exp_connecting) { /* b=9635, et. al. */
1188                 spin_unlock(&export->exp_lock);
1189                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1190                               export->exp_obd->obd_name, export,
1191                               libcfs_nid2str(req->rq_peer.nid));
1192                 class_export_put(export);
1193                 export = NULL;
1194                 rc = -EALREADY;
1195         } else if ((mds_conn || (lw_client && initial_conn) ||
1196                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1197                    export->exp_connection != NULL) {
1198                 spin_unlock(&export->exp_lock);
1199                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1200                         /* MDS or LWP reconnected after failover. */
1201                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1202                                       target->obd_name,
1203                                       mds_conn ? "MDS" : "LWP",
1204                                       libcfs_nid2str(req->rq_peer.nid),
1205                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
1206                 } else {
1207                         /* New MDS connection from the same NID. */
1208                         LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1209                                       target->obd_name,
1210                                       mds_conn ? "MDS" : "LWP",
1211                                       libcfs_nid2str(req->rq_peer.nid));
1212                 }
1213
1214                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1215                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1216                         /*
1217                          * Because exports between MDTs will always be
1218                          * kept, let's do not fail such export if they
1219                          * come from the same NID, otherwise it might
1220                          * cause eviction between MDTs, which might
1221                          * cause namespace inconsistency
1222                          */
1223                         spin_lock(&export->exp_lock);
1224                         export->exp_connecting = 1;
1225                         export->exp_conn_cnt = 0;
1226                         spin_unlock(&export->exp_lock);
1227                         conn.cookie = export->exp_handle.h_cookie;
1228                         rc = EALREADY;
1229                 } else {
1230                         class_fail_export(export);
1231                         class_export_put(export);
1232                         export = NULL;
1233                         rc = 0;
1234                 }
1235         } else if (export->exp_connection != NULL && initial_conn &&
1236                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1237                 spin_unlock(&export->exp_lock);
1238                 /* In MDS failover we have static UUID but NID can change. */
1239                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1240                               target->obd_name, cluuid.uuid,
1241                               libcfs_nid2str(req->rq_peer.nid),
1242                               libcfs_nid2str(
1243                                         export->exp_connection->c_peer.nid));
1244                 rc = -EALREADY;
1245                 class_export_put(export);
1246                 export = NULL;
1247         } else {
1248                 export->exp_connecting = 1;
1249                 spin_unlock(&export->exp_lock);
1250                 LASSERT(export->exp_obd == target);
1251
1252                 rc = target_handle_reconnect(&conn, export, &cluuid);
1253         }
1254
1255         /* If we found an export, we already unlocked. */
1256         if (!export) {
1257 no_export:
1258                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1259         } else if (req->rq_export == NULL &&
1260                    atomic_read(&export->exp_rpc_count) > 0) {
1261                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1262                               target->obd_name, cluuid.uuid,
1263                               libcfs_nid2str(req->rq_peer.nid),
1264                               atomic_read(&export->exp_refcount));
1265                         GOTO(out, rc = -EBUSY);
1266         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1267                    rc != EALREADY) {
1268                 if (!strstr(cluuid.uuid, "mdt"))
1269                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1270                                       target->obd_name, cluuid.uuid,
1271                                       libcfs_nid2str(req->rq_peer.nid));
1272                 GOTO(out, rc = -EALREADY);
1273         } else {
1274                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1275         }
1276
1277         if (rc < 0)
1278                 GOTO(out, rc);
1279
1280         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1281                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1282                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1283                export, ktime_get_seconds(),
1284                export ? export->exp_last_request_time : 0);
1285
1286         /*
1287          * If this is the first time a client connects, reset the recovery
1288          * timer. Discard lightweight connections which might be local.
1289          */
1290         if (!lw_client && rc == 0 && target->obd_recovering)
1291                 check_and_start_recovery_timer(target, req, export == NULL);
1292
1293         /*
1294          * We want to handle EALREADY but *not* -EALREADY from
1295          * target_handle_reconnect(), return reconnection state in a flag.
1296          */
1297         if (rc == EALREADY) {
1298                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1299                 rc = 0;
1300         } else {
1301                 LASSERT(rc == 0);
1302         }
1303
1304         /* Tell the client if we support replayable requests. */
1305         if (target->obd_replayable)
1306                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1307         client_nid = &req->rq_peer.nid;
1308
1309         if (export == NULL) {
1310                 /* allow lightweight connections during recovery */
1311                 /*
1312                  * allow "new" MDT to be connected during recovery, since we
1313                  * need retrieve recovery update records from it
1314                  */
1315                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1316                         struct hrtimer *timer = &target->obd_recovery_timer;
1317                         ktime_t remaining;
1318                         s64 timeout, left;
1319                         int in_progress;
1320                         int connected;
1321                         int known;
1322                         int stale;
1323                         char *msg;
1324
1325                         connected = atomic_read(&target->obd_connected_clients);
1326                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1327                         known =
1328                            atomic_read(&target->obd_max_recoverable_clients);
1329                         stale = target->obd_stale_clients;
1330                         remaining = hrtimer_expires_remaining(timer);
1331                         left = ktime_divns(remaining, NSEC_PER_SEC);
1332
1333                         if (ktime_to_ns(remaining) > 0) {
1334                                 msg = "to recover in";
1335                                 timeout = left;
1336                         } else {
1337                                 msg = "already passed deadline";
1338                                 timeout = -left;
1339                         }
1340
1341                         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",
1342                                       target->obd_name, cluuid.uuid,
1343                                       libcfs_nid2str(req->rq_peer.nid), known,
1344                                       connected - in_progress, in_progress,
1345                                       stale, msg, timeout / 60, timeout % 60);
1346                         rc = -EBUSY;
1347                 } else {
1348 dont_check_exports:
1349                         rc = obd_connect(req->rq_svc_thread->t_env,
1350                                          &export, target, &cluuid, data,
1351                                          client_nid);
1352                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1353                                 lustre_msg_add_op_flags(req->rq_repmsg,
1354                                                         MSG_CONNECT_RECOVERING);
1355                         if (rc == 0) {
1356                                 conn.cookie = export->exp_handle.h_cookie;
1357                                 rc = rev_import_init(export);
1358                         }
1359
1360                         if (mds_mds_conn)
1361                                 new_mds_mds_conn = true;
1362                 }
1363         } else {
1364                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1365                                    export, target, &cluuid, data, client_nid);
1366         }
1367         if (rc)
1368                 GOTO(out, rc);
1369
1370         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1371         data->ocd_instance = target->u.obt.obt_instance;
1372
1373         /*
1374          * Return only the parts of obd_connect_data that we understand, so the
1375          * client knows that we don't understand the rest.
1376          */
1377         if (data) {
1378                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1379                                                RCL_SERVER);
1380                 tmpdata = req_capsule_server_get(&req->rq_pill,
1381                                                  &RMF_CONNECT_DATA);
1382                 /*
1383                  * Don't use struct assignment here, because the client reply
1384                  * buffer may be smaller/larger than the local struct
1385                  * obd_connect_data.
1386                  */
1387                 memcpy(tmpdata, data, min(tmpsize, size));
1388         }
1389
1390         /*
1391          * If the client and the server are the same node, we will already
1392          * have an export that really points to the client's DLM export,
1393          * because we have a shared handles table.
1394          *
1395          * XXX this will go away when shaver stops sending the "connect" handle
1396          * in the real "remote handle" field of the request --phik 24 Apr 2003
1397          */
1398         ptlrpc_request_change_export(req, export);
1399
1400         spin_lock(&export->exp_lock);
1401         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1402                 spin_unlock(&export->exp_lock);
1403                 CDEBUG(D_RPCTRACE,
1404                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1405                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1406                        export->exp_conn_cnt,
1407                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1408
1409                 GOTO(out, rc = -EALREADY);
1410         }
1411         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1412         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1413         spin_unlock(&export->exp_lock);
1414
1415         if (export->exp_connection != NULL) {
1416                 /* Check to see if connection came from another NID. */
1417                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1418                     !hlist_unhashed(&export->exp_nid_hash))
1419                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1420                                      &export->exp_connection->c_peer.nid,
1421                                      &export->exp_nid_hash);
1422
1423                 ptlrpc_connection_put(export->exp_connection);
1424         }
1425
1426         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1427                                                        req->rq_self,
1428                                                        &cluuid);
1429         if (hlist_unhashed(&export->exp_nid_hash))
1430                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1431                              &export->exp_connection->c_peer.nid,
1432                              &export->exp_nid_hash);
1433
1434         lustre_msg_set_handle(req->rq_repmsg, &conn);
1435
1436         rc = rev_import_reconnect(export, req);
1437         if (rc != 0)
1438                 GOTO(out, rc);
1439
1440         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1441                 int has_transno;
1442                 __u64 transno = data->ocd_transno;
1443
1444                 spin_lock(&export->exp_lock);
1445                 /*
1446                  * possible race with class_disconnect_stale_exports,
1447                  * export may be already in the eviction process
1448                  */
1449                 if (export->exp_failed) {
1450                         spin_unlock(&export->exp_lock);
1451                         GOTO(out, rc = -ENODEV);
1452                 }
1453                 export->exp_in_recovery = 1;
1454                 export->exp_req_replay_needed = 1;
1455                 export->exp_lock_replay_needed = 1;
1456                 spin_unlock(&export->exp_lock);
1457
1458                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1459                                  MSG_CONNECT_TRANSNO);
1460                 if (has_transno && transno == 0)
1461                         CWARN("Connect with zero transno!\n");
1462
1463                 if (has_transno && transno > 0 &&
1464                     transno < target->obd_next_recovery_transno &&
1465                     transno > target->obd_last_committed) {
1466                         /* Another way is to use cmpxchg() to be lock-free. */
1467                         spin_lock(&target->obd_recovery_task_lock);
1468                         if (transno < target->obd_next_recovery_transno)
1469                                 target->obd_next_recovery_transno = transno;
1470                         spin_unlock(&target->obd_recovery_task_lock);
1471                 }
1472
1473                 atomic_inc(&target->obd_req_replay_clients);
1474                 atomic_inc(&target->obd_lock_replay_clients);
1475                 /*
1476                  * Note: MDS-MDS connection is allowed to be connected during
1477                  * recovery, no matter if the exports needs to be recoveried.
1478                  * Because we need retrieve updates logs from all other MDTs.
1479                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1480                  * also needs to be increased to match other recovery checking
1481                  * condition.
1482                  */
1483                 if (new_mds_mds_conn)
1484                         atomic_inc(&target->obd_max_recoverable_clients);
1485
1486                 if (atomic_inc_return(&target->obd_connected_clients) ==
1487                     atomic_read(&target->obd_max_recoverable_clients))
1488                         wake_up(&target->obd_next_transno_waitq);
1489         }
1490
1491         /* Tell the client we're in recovery, when client is involved in it. */
1492         if (target->obd_recovering && !lw_client)
1493                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1494
1495 out:
1496         if (export) {
1497                 spin_lock(&export->exp_lock);
1498                 export->exp_connecting = 0;
1499                 spin_unlock(&export->exp_lock);
1500
1501                 class_export_put(export);
1502         }
1503         if (target != NULL) {
1504                 spin_lock(&target->obd_dev_lock);
1505                 target->obd_conn_inprogress--;
1506                 spin_unlock(&target->obd_dev_lock);
1507                 class_decref(target, "find", current);
1508         }
1509         req->rq_status = rc;
1510         RETURN(rc);
1511 }
1512
1513 int target_handle_disconnect(struct ptlrpc_request *req)
1514 {
1515         int rc;
1516
1517         ENTRY;
1518
1519         rc = req_capsule_server_pack(&req->rq_pill);
1520         if (rc)
1521                 RETURN(rc);
1522
1523         /* Keep the rq_export around so we can send the reply. */
1524         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1525
1526         RETURN(0);
1527 }
1528
1529 void target_destroy_export(struct obd_export *exp)
1530 {
1531         struct obd_import *imp = NULL;
1532         /*
1533          * exports created from last_rcvd data, and "fake"
1534          * exports created by lctl don't have an import
1535          */
1536         spin_lock(&exp->exp_lock);
1537         if (exp->exp_imp_reverse != NULL) {
1538                 imp = exp->exp_imp_reverse;
1539                 exp->exp_imp_reverse = NULL;
1540         }
1541         spin_unlock(&exp->exp_lock);
1542         if (imp != NULL)
1543                 client_destroy_import(imp);
1544
1545         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1546         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1547         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1548         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1549 }
1550 EXPORT_SYMBOL(target_destroy_export);
1551
1552 /*
1553  * Recovery functions
1554  */
1555 static void target_request_copy_get(struct ptlrpc_request *req)
1556 {
1557         class_export_rpc_inc(req->rq_export);
1558         LASSERT(list_empty(&req->rq_list));
1559         INIT_LIST_HEAD(&req->rq_replay_list);
1560
1561         /* Increase refcount to keep request in queue. */
1562         atomic_inc(&req->rq_refcount);
1563         /* Let export know it has replays to be handled. */
1564         atomic_inc(&req->rq_export->exp_replay_count);
1565 }
1566
1567 static void target_request_copy_put(struct ptlrpc_request *req)
1568 {
1569         LASSERT(list_empty(&req->rq_replay_list));
1570         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1571
1572         atomic_dec(&req->rq_export->exp_replay_count);
1573         class_export_rpc_dec(req->rq_export);
1574         ptlrpc_server_drop_request(req);
1575 }
1576
1577 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1578 {
1579         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1580         struct obd_export *exp = req->rq_export;
1581         struct ptlrpc_request *reqiter;
1582         struct ptlrpc_request *dup_req = NULL;
1583         int dup = 0;
1584
1585         LASSERT(exp);
1586
1587         spin_lock(&exp->exp_lock);
1588         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1589                             rq_replay_list) {
1590                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1591                         dup_req = reqiter;
1592                         dup = 1;
1593                         break;
1594                 }
1595         }
1596
1597         if (dup) {
1598                 /* We expect it with RESENT and REPLAY flags. */
1599                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1600                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1601                         CERROR("invalid flags %x of resent replay\n",
1602                                lustre_msg_get_flags(req->rq_reqmsg));
1603
1604                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1605                         __u32 new_conn;
1606
1607                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1608                         if (new_conn >
1609                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1610                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1611                                                         new_conn);
1612                 }
1613         } else {
1614                 list_add_tail(&req->rq_replay_list,
1615                               &exp->exp_req_replay_queue);
1616         }
1617
1618         spin_unlock(&exp->exp_lock);
1619         return dup;
1620 }
1621
1622 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1623 {
1624         LASSERT(!list_empty(&req->rq_replay_list));
1625         LASSERT(req->rq_export);
1626
1627         spin_lock(&req->rq_export->exp_lock);
1628         list_del_init(&req->rq_replay_list);
1629         spin_unlock(&req->rq_export->exp_lock);
1630 }
1631
1632 static void target_finish_recovery(struct lu_target *lut)
1633 {
1634         struct obd_device *obd = lut->lut_obd;
1635
1636         ENTRY;
1637
1638         /* Only log a recovery message when recovery has occurred. */
1639         if (obd->obd_recovery_start) {
1640                 time64_t now = ktime_get_seconds();
1641                 time64_t elapsed_time;
1642
1643                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1644                                      1);
1645                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1646                               obd->obd_name, (s64)elapsed_time / 60,
1647                               (s64)elapsed_time % 60,
1648                               atomic_read(&obd->obd_max_recoverable_clients),
1649                               atomic_read(&obd->obd_connected_clients),
1650                               obd->obd_stale_clients,
1651                               obd->obd_stale_clients == 1 ? "was" : "were");
1652         }
1653
1654         ldlm_reprocess_recovery_done(obd->obd_namespace);
1655         spin_lock(&obd->obd_recovery_task_lock);
1656         if (!list_empty(&obd->obd_req_replay_queue) ||
1657             !list_empty(&obd->obd_lock_replay_queue) ||
1658             !list_empty(&obd->obd_final_req_queue)) {
1659                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1660                        obd->obd_name,
1661                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1662                        list_empty(&obd->obd_lock_replay_queue) ? \
1663                                   "" : "lock ",
1664                        list_empty(&obd->obd_final_req_queue) ? \
1665                                   "" : "final ");
1666                 spin_unlock(&obd->obd_recovery_task_lock);
1667                 LBUG();
1668         }
1669         spin_unlock(&obd->obd_recovery_task_lock);
1670
1671         obd->obd_recovery_end = ktime_get_seconds();
1672
1673         /* When recovery finished, cleanup orphans on MDS and OST. */
1674         if (obd->obd_type && OBP(obd, postrecov)) {
1675                 int rc = OBP(obd, postrecov)(obd);
1676
1677                 if (rc < 0)
1678                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1679                                       obd->obd_name, rc);
1680         }
1681         EXIT;
1682 }
1683
1684 static void abort_req_replay_queue(struct obd_device *obd)
1685 {
1686         struct ptlrpc_request *req, *n;
1687         struct list_head abort_list;
1688
1689         INIT_LIST_HEAD(&abort_list);
1690         spin_lock(&obd->obd_recovery_task_lock);
1691         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1692         spin_unlock(&obd->obd_recovery_task_lock);
1693         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1694                 DEBUG_REQ(D_WARNING, req, "aborted:");
1695                 req->rq_status = -ENOTCONN;
1696                 if (ptlrpc_error(req)) {
1697                         DEBUG_REQ(D_ERROR, req,
1698                                   "failed abort_req_reply; skipping");
1699                 }
1700                 target_exp_dequeue_req_replay(req);
1701                 target_request_copy_put(req);
1702         }
1703 }
1704
1705 static void abort_lock_replay_queue(struct obd_device *obd)
1706 {
1707         struct ptlrpc_request *req, *n;
1708         struct list_head abort_list;
1709
1710         INIT_LIST_HEAD(&abort_list);
1711         spin_lock(&obd->obd_recovery_task_lock);
1712         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1713         spin_unlock(&obd->obd_recovery_task_lock);
1714         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1715                 DEBUG_REQ(D_ERROR, req, "aborted:");
1716                 req->rq_status = -ENOTCONN;
1717                 if (ptlrpc_error(req)) {
1718                         DEBUG_REQ(D_ERROR, req,
1719                                   "failed abort_lock_reply; skipping");
1720                 }
1721                 target_request_copy_put(req);
1722         }
1723 }
1724
1725 /*
1726  * Called from a cleanup function if the device is being cleaned up
1727  * forcefully.  The exports should all have been disconnected already,
1728  * the only thing left to do is
1729  * - clear the recovery flags
1730  * - cancel the timer
1731  * - free queued requests and replies, but don't send replies
1732  * Because the obd_stopping flag is set, no new requests should be received.
1733  */
1734 void target_cleanup_recovery(struct obd_device *obd)
1735 {
1736         struct ptlrpc_request *req, *n;
1737         struct list_head clean_list;
1738
1739         INIT_LIST_HEAD(&clean_list);
1740         spin_lock(&obd->obd_dev_lock);
1741         if (!obd->obd_recovering) {
1742                 spin_unlock(&obd->obd_dev_lock);
1743                 EXIT;
1744                 return;
1745         }
1746         obd->obd_recovering = obd->obd_abort_recovery = 0;
1747         spin_unlock(&obd->obd_dev_lock);
1748
1749         spin_lock(&obd->obd_recovery_task_lock);
1750         target_cancel_recovery_timer(obd);
1751         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1752         spin_unlock(&obd->obd_recovery_task_lock);
1753
1754         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1755                 LASSERT(req->rq_reply_state == NULL);
1756                 target_exp_dequeue_req_replay(req);
1757                 target_request_copy_put(req);
1758         }
1759
1760         spin_lock(&obd->obd_recovery_task_lock);
1761         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1762         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1763         spin_unlock(&obd->obd_recovery_task_lock);
1764
1765         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1766                 LASSERT(req->rq_reply_state == NULL);
1767                 target_request_copy_put(req);
1768         }
1769
1770         EXIT;
1771 }
1772 EXPORT_SYMBOL(target_cleanup_recovery);
1773
1774 /* obd_recovery_task_lock should be held */
1775 void target_cancel_recovery_timer(struct obd_device *obd)
1776 {
1777         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1778         hrtimer_cancel(&obd->obd_recovery_timer);
1779 }
1780
1781 static void target_start_recovery_timer(struct obd_device *obd)
1782 {
1783         ktime_t delay;
1784
1785         if (obd->obd_recovery_start != 0)
1786                 return;
1787
1788         spin_lock(&obd->obd_dev_lock);
1789         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1790                 spin_unlock(&obd->obd_dev_lock);
1791                 return;
1792         }
1793
1794         LASSERT(obd->obd_recovery_timeout != 0);
1795
1796         if (obd->obd_recovery_start != 0) {
1797                 spin_unlock(&obd->obd_dev_lock);
1798                 return;
1799         }
1800
1801         obd->obd_recovery_start = ktime_get_seconds();
1802         delay = ktime_set(obd->obd_recovery_start +
1803                           obd->obd_recovery_timeout, 0);
1804         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1805         spin_unlock(&obd->obd_dev_lock);
1806
1807         LCONSOLE_WARN("%s: Will be in recovery for at least %lu:%02lu, or until %d client%s reconnect%s\n",
1808                       obd->obd_name,
1809                       obd->obd_recovery_timeout / 60,
1810                       obd->obd_recovery_timeout % 60,
1811                       atomic_read(&obd->obd_max_recoverable_clients),
1812                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1813                       "" : "s",
1814                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1815                       "s" : "");
1816 }
1817
1818 /**
1819  * extend recovery window.
1820  *
1821  * if @extend is true, extend recovery window to have @dr_timeout remaining
1822  * at least; otherwise, make sure the recovery timeout value is not less
1823  * than @dr_timeout.
1824  */
1825 static void extend_recovery_timer(struct obd_device *obd, time_t dr_timeout,
1826                                   bool extend)
1827 {
1828         ktime_t left_ns;
1829         time_t timeout;
1830         time_t left;
1831
1832         spin_lock(&obd->obd_dev_lock);
1833         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1834                 spin_unlock(&obd->obd_dev_lock);
1835                 return;
1836         }
1837         LASSERT(obd->obd_recovery_start != 0);
1838
1839         left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1840         left = ktime_divns(left_ns, NSEC_PER_SEC);
1841
1842         if (extend) {
1843                 timeout = obd->obd_recovery_timeout;
1844                 /* dr_timeout will happen after the hrtimer has expired.
1845                  * Add the excess time to the soft recovery timeout without
1846                  * exceeding the hard recovery timeout.
1847                  */
1848                 if (dr_timeout > left) {
1849                         timeout += dr_timeout - left;
1850                         timeout = min_t(time_t, obd->obd_recovery_time_hard,
1851                                         timeout);
1852                 }
1853         } else {
1854                 timeout = clamp_t(time_t, dr_timeout, obd->obd_recovery_timeout,
1855                                   obd->obd_recovery_time_hard);
1856         }
1857
1858         if (timeout == obd->obd_recovery_time_hard)
1859                 CWARN("%s: extended recovery timer reached hard limit: %ld, extend: %d\n",
1860                       obd->obd_name, timeout, extend);
1861
1862         if (obd->obd_recovery_timeout < timeout) {
1863                 ktime_t end, now;
1864
1865                 obd->obd_recovery_timeout = timeout;
1866                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1867                 now = ktime_set(ktime_get_seconds(), 0);
1868                 left_ns = ktime_sub(end, now);
1869                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1870                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1871         }
1872         spin_unlock(&obd->obd_dev_lock);
1873
1874         CDEBUG(D_HA, "%s: recovery timer will expire in %ld seconds\n",
1875                 obd->obd_name, left);
1876 }
1877
1878 /* Reset the timer with each new client connection */
1879 /*
1880  * This timer is actually reconnect_timer, which is for making sure
1881  * the total recovery window is at least as big as my reconnect
1882  * attempt timing. So the initial recovery time_out will be set to
1883  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1884  * from client is bigger than this, then the recovery time_out will
1885  * be extended to make sure the client could be reconnected, in the
1886  * process, the timeout from the new client should be ignored.
1887  */
1888 static void
1889 check_and_start_recovery_timer(struct obd_device *obd,
1890                                struct ptlrpc_request *req,
1891                                int new_client)
1892 {
1893         time_t service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1894         struct obd_device_target *obt = &obd->u.obt;
1895
1896         if (!new_client && service_time)
1897                 /*
1898                  * Teach server about old server's estimates, as first guess
1899                  * at how long new requests will take.
1900                  */
1901                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1902                             service_time);
1903
1904         target_start_recovery_timer(obd);
1905
1906         /*
1907          * Convert the service time to RPC timeout,
1908          * and reuse service_time to limit stack usage.
1909          */
1910         service_time = at_est2timeout(service_time);
1911
1912         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1913             service_time < at_extra)
1914                 service_time = at_extra;
1915
1916         /*
1917          * We expect other clients to timeout within service_time, then try
1918          * to reconnect, then try the failover server.  The max delay between
1919          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1920          */
1921         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1922
1923         LASSERT(obt->obt_magic == OBT_MAGIC);
1924         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1925         if (service_time > obd->obd_recovery_timeout && !new_client)
1926                 extend_recovery_timer(obd, service_time, false);
1927 }
1928
1929 /** Health checking routines */
1930 static inline int exp_connect_healthy(struct obd_export *exp)
1931 {
1932         return exp->exp_in_recovery;
1933 }
1934
1935 /** if export done req_replay or has replay in queue */
1936 static inline int exp_req_replay_healthy(struct obd_export *exp)
1937 {
1938         return (!exp->exp_req_replay_needed ||
1939                 atomic_read(&exp->exp_replay_count) > 0);
1940 }
1941
1942
1943 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1944 {
1945         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1946                exp_req_replay_healthy(exp);
1947 }
1948
1949 /** if export done lock_replay or has replay in queue */
1950 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1951 {
1952         return (!exp->exp_lock_replay_needed ||
1953                 atomic_read(&exp->exp_replay_count) > 0);
1954 }
1955
1956 static inline int exp_vbr_healthy(struct obd_export *exp)
1957 {
1958         return !exp->exp_vbr_failed;
1959 }
1960
1961 static inline int exp_finished(struct obd_export *exp)
1962 {
1963         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1964 }
1965
1966 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1967 {
1968         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1969                 exp_finished(exp);
1970 }
1971
1972 static int check_for_next_transno(struct lu_target *lut)
1973 {
1974         struct ptlrpc_request *req = NULL;
1975         struct obd_device *obd = lut->lut_obd;
1976         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1977         int wake_up = 0, connected, completed, queue_len;
1978         __u64 req_transno = 0;
1979         __u64 update_transno = 0;
1980         __u64 next_transno = 0;
1981
1982         ENTRY;
1983
1984         spin_lock(&obd->obd_recovery_task_lock);
1985         if (!list_empty(&obd->obd_req_replay_queue)) {
1986                 req = list_entry(obd->obd_req_replay_queue.next,
1987                                      struct ptlrpc_request, rq_list);
1988                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1989         }
1990
1991         if (tdtd != NULL)
1992                 update_transno = distribute_txn_get_next_transno(tdtd);
1993
1994         connected = atomic_read(&obd->obd_connected_clients);
1995         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1996         queue_len = obd->obd_requests_queued_for_recovery;
1997         next_transno = obd->obd_next_recovery_transno;
1998
1999         CDEBUG(D_HA,
2000                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2001                atomic_read(&obd->obd_max_recoverable_clients),
2002                connected, completed,
2003                queue_len, req_transno, next_transno);
2004
2005         if (obd->obd_abort_recovery) {
2006                 CDEBUG(D_HA, "waking for aborted recovery\n");
2007                 wake_up = 1;
2008         } else if (obd->obd_recovery_expired) {
2009                 CDEBUG(D_HA, "waking for expired recovery\n");
2010                 wake_up = 1;
2011         } else if (tdtd != NULL && req != NULL &&
2012                    is_req_replayed_by_update(req)) {
2013                 LASSERTF(req_transno < next_transno,
2014                          "req_transno %llu next_transno%llu\n", req_transno,
2015                          next_transno);
2016                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2017                        req_transno);
2018                 wake_up = 1;
2019         } else if (req_transno == next_transno ||
2020                    (update_transno != 0 && update_transno <= next_transno)) {
2021                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2022                 wake_up = 1;
2023         } else if (queue_len > 0 &&
2024                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2025                 /** handle gaps occured due to lost reply or VBR */
2026                 LASSERTF(req_transno >= next_transno,
2027                          "req_transno: %llu, next_transno: %llu\n",
2028                          req_transno, next_transno);
2029                 CDEBUG(D_HA,
2030                        "%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",
2031                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2032                        next_transno, queue_len, completed, connected,
2033                        req_transno, update_transno, obd->obd_last_committed);
2034                 obd->obd_next_recovery_transno = req_transno;
2035                 wake_up = 1;
2036         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2037                 CDEBUG(D_HA, "waking for completed recovery\n");
2038                 wake_up = 1;
2039         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2040                 CDEBUG(D_HA,
2041                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2042                        next_transno);
2043                 obd->obd_next_recovery_transno = req_transno;
2044                 wake_up = 1;
2045         }
2046         spin_unlock(&obd->obd_recovery_task_lock);
2047         return wake_up;
2048 }
2049
2050 static int check_for_next_lock(struct lu_target *lut)
2051 {
2052         struct obd_device *obd = lut->lut_obd;
2053         int wake_up = 0;
2054
2055         spin_lock(&obd->obd_recovery_task_lock);
2056         if (!list_empty(&obd->obd_lock_replay_queue)) {
2057                 CDEBUG(D_HA, "waking for next lock\n");
2058                 wake_up = 1;
2059         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2060                 CDEBUG(D_HA, "waking for completed lock replay\n");
2061                 wake_up = 1;
2062         } else if (obd->obd_abort_recovery) {
2063                 CDEBUG(D_HA, "waking for aborted recovery\n");
2064                 wake_up = 1;
2065         } else if (obd->obd_recovery_expired) {
2066                 CDEBUG(D_HA, "waking for expired recovery\n");
2067                 wake_up = 1;
2068         }
2069         spin_unlock(&obd->obd_recovery_task_lock);
2070
2071         return wake_up;
2072 }
2073
2074 /**
2075  * wait for recovery events,
2076  * check its status with help of check_routine
2077  * evict dead clients via health_check
2078  */
2079 static int target_recovery_overseer(struct lu_target *lut,
2080                                     int (*check_routine)(struct lu_target *),
2081                                     int (*health_check)(struct obd_export *))
2082 {
2083         struct obd_device *obd = lut->lut_obd;
2084         struct target_distribute_txn_data *tdtd;
2085         time64_t last = 0;
2086         time64_t now;
2087 repeat:
2088         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2089                 now = ktime_get_seconds();
2090                 if (now - last > 600) {
2091                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2092                                       obd->obd_name);
2093                         last = now;
2094                 }
2095         }
2096         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2097               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2098                 __u64 next_update_transno = 0;
2099
2100                 /*
2101                  * Only abort the recovery if there are no update recovery
2102                  * left in the queue
2103                  */
2104                 spin_lock(&obd->obd_recovery_task_lock);
2105                 if (lut->lut_tdtd != NULL) {
2106                         next_update_transno =
2107                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2108
2109                         tdtd = lut->lut_tdtd;
2110                         /*
2111                          * If next_update_transno == 0, it probably because
2112                          * updatelog retrieve threads did not get any records
2113                          * yet, let's wait those threads stopped
2114                          */
2115                         if (next_update_transno == 0) {
2116                                 struct l_wait_info lwi = { 0 };
2117
2118                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2119                                        atomic_read(
2120                                        &tdtd->tdtd_recovery_threads_count) == 0,
2121                                        &lwi);
2122
2123                                 next_update_transno =
2124                                         distribute_txn_get_next_transno(
2125                                                                 lut->lut_tdtd);
2126                         }
2127                 }
2128
2129                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2130                         obd->obd_next_recovery_transno = next_update_transno;
2131                         spin_unlock(&obd->obd_recovery_task_lock);
2132                         /*
2133                          * Disconnect unfinished exports from clients, and
2134                          * keep connection from MDT to make sure the update
2135                          * recovery will still keep trying until some one
2136                          * manually abort the recovery
2137                          */
2138                         class_disconnect_stale_exports(obd,
2139                                                 exp_finished_or_from_mdt);
2140                         /* Abort all of replay & replay lock req from clients */
2141                         abort_req_replay_queue(obd);
2142                         abort_lock_replay_queue(obd);
2143                         CDEBUG(D_HA,
2144                                "%s: there are still update replay (%#llx)in the queue.\n",
2145                                obd->obd_name, next_update_transno);
2146                 } else {
2147                         obd->obd_abort_recovery = 1;
2148                         spin_unlock(&obd->obd_recovery_task_lock);
2149                         CWARN("%s recovery is aborted by hard timeout\n",
2150                               obd->obd_name);
2151                 }
2152         }
2153
2154         while (wait_event_timeout(obd->obd_next_transno_waitq,
2155                                   check_routine(lut),
2156                                   cfs_time_seconds(60)) == 0)
2157                 ; /* wait indefinitely for event, but don't trigger watchdog */
2158
2159         if (obd->obd_abort_recovery) {
2160                 CWARN("recovery is aborted, evict exports in recovery\n");
2161                 if (lut->lut_tdtd != NULL) {
2162                         struct l_wait_info lwi = { 0 };
2163
2164                         tdtd = lut->lut_tdtd;
2165                         /*
2166                          * Let's wait all of the update log recovery thread
2167                          * finished
2168                          */
2169                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2170                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2171                              &lwi);
2172                         /* Then abort the update recovery list */
2173                         dtrq_list_destroy(lut->lut_tdtd);
2174                 }
2175
2176                 /** evict exports which didn't finish recovery yet */
2177                 class_disconnect_stale_exports(obd, exp_finished);
2178                 return 1;
2179         } else if (obd->obd_recovery_expired) {
2180                 obd->obd_recovery_expired = 0;
2181
2182                 /** If some clients died being recovered, evict them */
2183                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2184                               obd->obd_name);
2185                 /** evict cexports with no replay in queue, they are stalled */
2186                 class_disconnect_stale_exports(obd, health_check);
2187
2188                 /** continue with VBR */
2189                 spin_lock(&obd->obd_dev_lock);
2190                 obd->obd_version_recov = 1;
2191                 spin_unlock(&obd->obd_dev_lock);
2192                 /**
2193                  * reset timer, recovery will proceed with versions now,
2194                  * timeout is set just to handle reconnection delays
2195                  */
2196                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2197                 /**
2198                  * Wait for recovery events again, after evicting bad clients
2199                  */
2200                 goto repeat;
2201         }
2202         return 0;
2203 }
2204
2205 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2206 {
2207         struct obd_device *obd = lut->lut_obd;
2208         struct ptlrpc_request *req = NULL;
2209
2210         CDEBUG(D_HA, "Waiting for lock\n");
2211         if (target_recovery_overseer(lut, check_for_next_lock,
2212                                      exp_lock_replay_healthy))
2213                 abort_lock_replay_queue(obd);
2214
2215         spin_lock(&obd->obd_recovery_task_lock);
2216         if (!list_empty(&obd->obd_lock_replay_queue)) {
2217                 req = list_entry(obd->obd_lock_replay_queue.next,
2218                                      struct ptlrpc_request, rq_list);
2219                 list_del_init(&req->rq_list);
2220                 spin_unlock(&obd->obd_recovery_task_lock);
2221         } else {
2222                 spin_unlock(&obd->obd_recovery_task_lock);
2223                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2224                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2225                 /** evict exports failed VBR */
2226                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2227         }
2228         return req;
2229 }
2230
2231 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2232 {
2233         struct ptlrpc_request *req = NULL;
2234
2235         spin_lock(&obd->obd_recovery_task_lock);
2236         if (!list_empty(&obd->obd_final_req_queue)) {
2237                 req = list_entry(obd->obd_final_req_queue.next,
2238                                      struct ptlrpc_request, rq_list);
2239                 list_del_init(&req->rq_list);
2240                 spin_unlock(&obd->obd_recovery_task_lock);
2241                 if (req->rq_export->exp_in_recovery) {
2242                         spin_lock(&req->rq_export->exp_lock);
2243                         req->rq_export->exp_in_recovery = 0;
2244                         spin_unlock(&req->rq_export->exp_lock);
2245                 }
2246         } else {
2247                 spin_unlock(&obd->obd_recovery_task_lock);
2248         }
2249         return req;
2250 }
2251
2252 static void handle_recovery_req(struct ptlrpc_thread *thread,
2253                                 struct ptlrpc_request *req,
2254                                 svc_handler_t handler)
2255 {
2256         ENTRY;
2257
2258         /**
2259          * export can be evicted during recovery, no need to handle replays for
2260          * it after that, discard such request silently
2261          */
2262         if (req->rq_export->exp_disconnected)
2263                 RETURN_EXIT;
2264
2265         req->rq_session.lc_thread = thread;
2266         req->rq_svc_thread = thread;
2267         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2268
2269         /* thread context */
2270         lu_context_enter(&thread->t_env->le_ctx);
2271         (void)handler(req);
2272         lu_context_exit(&thread->t_env->le_ctx);
2273
2274         /* don't reset timer for final stage */
2275         if (!exp_finished(req->rq_export)) {
2276                 time_t to = obd_timeout;
2277
2278                 /**
2279                  * Add request timeout to the recovery time so next request from
2280                  * this client may come in recovery time
2281                  */
2282                 if (!AT_OFF) {
2283                         struct ptlrpc_service_part *svcpt;
2284
2285                         svcpt = req->rq_rqbd->rqbd_svcpt;
2286                         /*
2287                          * If the server sent early reply for this request,
2288                          * the client will recalculate the timeout according to
2289                          * current server estimate service time, so we will
2290                          * use the maxium timeout here for waiting the client
2291                          * sending the next req
2292                          */
2293                         to = max_t(time_t,
2294                                    at_est2timeout(at_get(&svcpt->scp_at_estimate)),
2295                                    lustre_msg_get_timeout(req->rq_reqmsg));
2296                         /*
2297                          * Add 2 net_latency, one for balance rq_deadline
2298                          * (see ptl_send_rpc), one for resend the req to server,
2299                          * Note: client will pack net_latency in replay req
2300                          * (see ptlrpc_replay_req)
2301                          */
2302                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2303                 }
2304                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2305         }
2306         EXIT;
2307 }
2308
2309 /** Checking routines for recovery */
2310 static int check_for_recovery_ready(struct lu_target *lut)
2311 {
2312         struct obd_device *obd = lut->lut_obd;
2313         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2314
2315         CDEBUG(D_HA,
2316                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2317                clnts, obd->obd_stale_clients,
2318                atomic_read(&obd->obd_max_recoverable_clients),
2319                obd->obd_abort_recovery, obd->obd_recovery_expired);
2320
2321         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2322                 LASSERT(clnts <=
2323                         atomic_read(&obd->obd_max_recoverable_clients));
2324                 if (clnts + obd->obd_stale_clients <
2325                     atomic_read(&obd->obd_max_recoverable_clients))
2326                         return 0;
2327         }
2328
2329         if (lut->lut_tdtd != NULL) {
2330                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2331                     !obd->obd_abort_recovery) {
2332                         /*
2333                          * Let's extend recovery timer, in case the recovery
2334                          * timer expired, and some clients got evicted
2335                          */
2336                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2337                                               true);
2338                         CDEBUG(D_HA,
2339                                "%s update recovery is not ready, extend recovery %lu\n",
2340                                obd->obd_name, obd->obd_recovery_timeout);
2341                         return 0;
2342                 }
2343         }
2344
2345         return 1;
2346 }
2347
2348 enum {
2349         REQUEST_RECOVERY = 1,
2350         UPDATE_RECOVERY = 2,
2351 };
2352
2353 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2354 {
2355         __u64 transno = 0;
2356
2357         if (!list_empty(&obd->obd_req_replay_queue)) {
2358                 struct ptlrpc_request *req;
2359
2360                 req = list_entry(obd->obd_req_replay_queue.next,
2361                                  struct ptlrpc_request, rq_list);
2362                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2363         }
2364
2365         return transno;
2366 }
2367
2368 static __u64 get_next_transno(struct lu_target *lut, int *type)
2369 {
2370         struct obd_device *obd = lut->lut_obd;
2371         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2372         __u64 transno = 0;
2373         __u64 update_transno;
2374
2375         ENTRY;
2376
2377         transno = get_next_replay_req_transno(obd);
2378         if (type != NULL)
2379                 *type = REQUEST_RECOVERY;
2380
2381         if (tdtd == NULL)
2382                 RETURN(transno);
2383
2384         update_transno = distribute_txn_get_next_transno(tdtd);
2385         if (transno == 0 || (transno >= update_transno &&
2386                              update_transno != 0)) {
2387                 transno = update_transno;
2388                 if (type != NULL)
2389                         *type = UPDATE_RECOVERY;
2390         }
2391
2392         RETURN(transno);
2393 }
2394
2395 /**
2396  * drop duplicate replay request
2397  *
2398  * Because the operation has been replayed by update recovery, the request
2399  * with the same transno will be dropped and also notify the client to send
2400  * next replay request.
2401  *
2402  * \param[in] env       execution environment
2403  * \param[in] obd       failover obd device
2404  * \param[in] req       request to be dropped
2405  */
2406 static void drop_duplicate_replay_req(struct lu_env *env,
2407                                       struct obd_device *obd,
2408                                       struct ptlrpc_request *req)
2409 {
2410         DEBUG_REQ(D_HA, req,
2411                   "remove t%lld from %s because duplicate update records found",
2412                   lustre_msg_get_transno(req->rq_reqmsg),
2413                   libcfs_nid2str(req->rq_peer.nid));
2414
2415         /*
2416          * Right now, only for MDS reint operation update replay and
2417          * normal request replay can have the same transno
2418          */
2419         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2420                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2421                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2422                 if (likely(req->rq_export))
2423                         target_committed_to_req(req);
2424                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2425                 target_send_reply(req, req->rq_status, 0);
2426         } else {
2427                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2428                 libcfs_nid2str(req->rq_peer.nid));
2429         }
2430         target_exp_dequeue_req_replay(req);
2431         target_request_copy_put(req);
2432         obd->obd_replayed_requests++;
2433 }
2434
2435 static void replay_request_or_update(struct lu_env *env,
2436                                      struct lu_target *lut,
2437                                      struct target_recovery_data *trd,
2438                                      struct ptlrpc_thread *thread)
2439 {
2440         struct obd_device *obd = lut->lut_obd;
2441         struct ptlrpc_request *req = NULL;
2442         int type;
2443         __u64 transno;
2444
2445         ENTRY;
2446
2447         CDEBUG(D_HA, "Waiting for transno %lld\n",
2448                obd->obd_next_recovery_transno);
2449
2450         /* Replay all of request and update by transno */
2451         do {
2452                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2453
2454                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2455
2456                 /**
2457                  * It is needed to extend recovery window above
2458                  *  recovery_time_soft. Extending is possible only in the
2459                  *  end of recovery window (see more details in
2460                  *  handle_recovery_req()).
2461                  */
2462                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2463
2464                 if (target_recovery_overseer(lut, check_for_next_transno,
2465                                         exp_req_replay_healthy_or_from_mdt)) {
2466                         abort_req_replay_queue(obd);
2467                         abort_lock_replay_queue(obd);
2468                         goto abort;
2469                 }
2470
2471                 spin_lock(&obd->obd_recovery_task_lock);
2472                 transno = get_next_transno(lut, &type);
2473                 if (type == REQUEST_RECOVERY && transno != 0) {
2474                         /*
2475                          * Drop replay request from client side, if the
2476                          * replay has been executed by update with the
2477                          * same transno
2478                          */
2479                         req = list_entry(obd->obd_req_replay_queue.next,
2480                                         struct ptlrpc_request, rq_list);
2481
2482                         list_del_init(&req->rq_list);
2483                         obd->obd_requests_queued_for_recovery--;
2484                         spin_unlock(&obd->obd_recovery_task_lock);
2485
2486                         /*
2487                          * Let's check if the request has been redone by
2488                          * update replay
2489                          */
2490                         if (is_req_replayed_by_update(req)) {
2491                                 struct distribute_txn_replay_req *dtrq;
2492
2493                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2494                                                                       transno);
2495                                 LASSERT(dtrq != NULL);
2496                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2497                                 list_del_init(&dtrq->dtrq_list);
2498                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2499                                 dtrq_destroy(dtrq);
2500
2501                                 drop_duplicate_replay_req(env, obd, req);
2502
2503                                 continue;
2504                         }
2505
2506                         LASSERT(trd->trd_processing_task == current_pid());
2507                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2508                                   req->rq_xid,
2509                                   lustre_msg_get_transno(req->rq_reqmsg),
2510                                   libcfs_nid2str(req->rq_peer.nid));
2511
2512                         handle_recovery_req(thread, req,
2513                                             trd->trd_recovery_handler);
2514                         /**
2515                          * bz18031: increase next_recovery_transno before
2516                          * target_request_copy_put() will drop exp_rpc reference
2517                          */
2518                         spin_lock(&obd->obd_recovery_task_lock);
2519                         obd->obd_next_recovery_transno++;
2520                         spin_unlock(&obd->obd_recovery_task_lock);
2521                         target_exp_dequeue_req_replay(req);
2522                         target_request_copy_put(req);
2523                         obd->obd_replayed_requests++;
2524                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2525                         struct distribute_txn_replay_req *dtrq;
2526                         int rc;
2527
2528                         spin_unlock(&obd->obd_recovery_task_lock);
2529
2530                         LASSERT(tdtd != NULL);
2531                         dtrq = distribute_txn_get_next_req(tdtd);
2532                         lu_context_enter(&thread->t_env->le_ctx);
2533                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2534                         lu_context_exit(&thread->t_env->le_ctx);
2535                         extend_recovery_timer(obd, obd_timeout, true);
2536
2537                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2538                                 CDEBUG(D_HA,
2539                                        "Move x%llu t%llu to finish list\n",
2540                                        dtrq->dtrq_xid,
2541                                        dtrq->dtrq_master_transno);
2542
2543                                 /* Add it to the replay finish list */
2544                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2545                                 list_add(&dtrq->dtrq_list,
2546                                          &tdtd->tdtd_replay_finish_list);
2547                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2548
2549                                 spin_lock(&obd->obd_recovery_task_lock);
2550                                 if (transno == obd->obd_next_recovery_transno)
2551                                         obd->obd_next_recovery_transno++;
2552                                 else if (transno >
2553                                          obd->obd_next_recovery_transno)
2554                                         obd->obd_next_recovery_transno =
2555                                                                 transno + 1;
2556                                 spin_unlock(&obd->obd_recovery_task_lock);
2557                         } else {
2558                                 dtrq_destroy(dtrq);
2559                         }
2560                 } else {
2561                         spin_unlock(&obd->obd_recovery_task_lock);
2562 abort:
2563                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2564                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2565                         /** evict exports failed VBR */
2566                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2567                         break;
2568                 }
2569         } while (1);
2570 }
2571
2572 static int target_recovery_thread(void *arg)
2573 {
2574         struct lu_target *lut = arg;
2575         struct obd_device *obd = lut->lut_obd;
2576         struct ptlrpc_request *req;
2577         struct target_recovery_data *trd = &obd->obd_recovery_data;
2578         unsigned long delta;
2579         struct lu_env *env;
2580         struct ptlrpc_thread *thread = NULL;
2581         int rc = 0;
2582
2583         ENTRY;
2584
2585         unshare_fs_struct();
2586         OBD_ALLOC_PTR(thread);
2587         if (thread == NULL)
2588                 RETURN(-ENOMEM);
2589
2590         OBD_ALLOC_PTR(env);
2591         if (env == NULL)
2592                 GOTO(out_thread, rc = -ENOMEM);
2593         rc = lu_env_add(env);
2594         if (rc)
2595                 GOTO(out_env, rc);
2596
2597         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2598         if (rc)
2599                 GOTO(out_env_remove, rc);
2600
2601         thread->t_env = env;
2602         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2603         env->le_ctx.lc_thread = thread;
2604         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2605
2606         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2607                current_pid());
2608         trd->trd_processing_task = current_pid();
2609
2610         spin_lock(&obd->obd_dev_lock);
2611         obd->obd_recovering = 1;
2612         spin_unlock(&obd->obd_dev_lock);
2613         complete(&trd->trd_starting);
2614
2615         /* first of all, we have to know the first transno to replay */
2616         if (target_recovery_overseer(lut, check_for_recovery_ready,
2617                                      exp_connect_healthy)) {
2618                 abort_req_replay_queue(obd);
2619                 abort_lock_replay_queue(obd);
2620                 if (lut->lut_tdtd != NULL)
2621                         dtrq_list_destroy(lut->lut_tdtd);
2622         }
2623
2624         /* next stage: replay requests or update */
2625         delta = jiffies;
2626         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2627                atomic_read(&obd->obd_req_replay_clients),
2628                obd->obd_next_recovery_transno);
2629         replay_request_or_update(env, lut, trd, thread);
2630
2631         /**
2632          * The second stage: replay locks
2633          */
2634         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2635                atomic_read(&obd->obd_lock_replay_clients));
2636         while ((req = target_next_replay_lock(lut))) {
2637                 LASSERT(trd->trd_processing_task == current_pid());
2638                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2639                           libcfs_nid2str(req->rq_peer.nid));
2640                 handle_recovery_req(thread, req,
2641                                     trd->trd_recovery_handler);
2642                 target_request_copy_put(req);
2643                 obd->obd_replayed_locks++;
2644         }
2645
2646         /**
2647          * The third stage: reply on final pings, at this moment all clients
2648          * must have request in final queue
2649          */
2650         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2651         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2652         /** Update server last boot epoch */
2653         tgt_boot_epoch_update(lut);
2654         /*
2655          * We drop recoverying flag to forward all new requests
2656          * to regular mds_handle() since now
2657          */
2658         spin_lock(&obd->obd_dev_lock);
2659         obd->obd_recovering = obd->obd_abort_recovery = 0;
2660         spin_unlock(&obd->obd_dev_lock);
2661         spin_lock(&obd->obd_recovery_task_lock);
2662         target_cancel_recovery_timer(obd);
2663         spin_unlock(&obd->obd_recovery_task_lock);
2664         while ((req = target_next_final_ping(obd))) {
2665                 LASSERT(trd->trd_processing_task == current_pid());
2666                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2667                           libcfs_nid2str(req->rq_peer.nid));
2668                 handle_recovery_req(thread, req,
2669                                     trd->trd_recovery_handler);
2670                 /*
2671                  * Because the waiting client can not send ping to server,
2672                  * so we need refresh the last_request_time, to avoid the
2673                  * export is being evicted
2674                  */
2675                 ptlrpc_update_export_timer(req->rq_export, 0);
2676                 target_request_copy_put(req);
2677         }
2678
2679         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2680         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2681                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2682         if (delta > OBD_RECOVERY_TIME_SOFT) {
2683                 CWARN("too long recovery - read logs\n");
2684                 libcfs_debug_dumplog();
2685         }
2686
2687         target_finish_recovery(lut);
2688         lu_context_fini(&env->le_ctx);
2689         trd->trd_processing_task = 0;
2690         complete_all(&trd->trd_finishing);
2691         tgt_io_thread_done(thread);
2692 out_env_remove:
2693         lu_env_remove(env);
2694 out_env:
2695         OBD_FREE_PTR(env);
2696 out_thread:
2697         OBD_FREE_PTR(thread);
2698         RETURN(rc);
2699 }
2700
2701 static int target_start_recovery_thread(struct lu_target *lut,
2702                                         svc_handler_t handler)
2703 {
2704         struct obd_device *obd = lut->lut_obd;
2705         int rc = 0;
2706         struct target_recovery_data *trd = &obd->obd_recovery_data;
2707         int index;
2708
2709         memset(trd, 0, sizeof(*trd));
2710         init_completion(&trd->trd_starting);
2711         init_completion(&trd->trd_finishing);
2712         trd->trd_recovery_handler = handler;
2713
2714         rc = server_name2index(obd->obd_name, &index, NULL);
2715         if (rc < 0)
2716                 return rc;
2717
2718         if (!IS_ERR(kthread_run(target_recovery_thread,
2719                                 lut, "tgt_recover_%d", index))) {
2720                 wait_for_completion(&trd->trd_starting);
2721                 LASSERT(obd->obd_recovering != 0);
2722         } else {
2723                 rc = -ECHILD;
2724         }
2725
2726         return rc;
2727 }
2728
2729 void target_stop_recovery_thread(struct obd_device *obd)
2730 {
2731         if (obd->obd_recovery_data.trd_processing_task > 0) {
2732                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2733                 /** recovery can be done but postrecovery is not yet */
2734                 spin_lock(&obd->obd_dev_lock);
2735                 if (obd->obd_recovering) {
2736                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2737                         obd->obd_abort_recovery = 1;
2738                         wake_up(&obd->obd_next_transno_waitq);
2739                 }
2740                 spin_unlock(&obd->obd_dev_lock);
2741                 wait_for_completion(&trd->trd_finishing);
2742         }
2743 }
2744 EXPORT_SYMBOL(target_stop_recovery_thread);
2745
2746 void target_recovery_fini(struct obd_device *obd)
2747 {
2748         class_disconnect_exports(obd);
2749         target_stop_recovery_thread(obd);
2750         target_cleanup_recovery(obd);
2751 }
2752 EXPORT_SYMBOL(target_recovery_fini);
2753
2754 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2755 {
2756         struct obd_device *obd = container_of(timer, struct obd_device,
2757                                               obd_recovery_timer);
2758
2759         CDEBUG(D_HA,
2760                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2761                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2762                ktime_get_real_seconds() - obd->obd_recovery_start,
2763                atomic_read(&obd->obd_connected_clients));
2764
2765         obd->obd_recovery_expired = 1;
2766         wake_up(&obd->obd_next_transno_waitq);
2767         return HRTIMER_NORESTART;
2768 }
2769
2770 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2771 {
2772         struct obd_device *obd = lut->lut_obd;
2773
2774         if (lut->lut_bottom->dd_rdonly)
2775                 return;
2776
2777         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2778                 /** Update server last boot epoch */
2779                 tgt_boot_epoch_update(lut);
2780                 return;
2781         }
2782
2783         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2784                "last_transno %llu\n", obd->obd_name,
2785                atomic_read(&obd->obd_max_recoverable_clients),
2786                obd->obd_last_committed);
2787         LASSERT(obd->obd_stopping == 0);
2788         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2789         obd->obd_recovery_start = 0;
2790         obd->obd_recovery_end = 0;
2791
2792         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2793                      HRTIMER_MODE_ABS);
2794         obd->obd_recovery_timer.function = &target_recovery_expired;
2795         target_start_recovery_thread(lut, handler);
2796 }
2797 EXPORT_SYMBOL(target_recovery_init);
2798
2799 static int target_process_req_flags(struct obd_device *obd,
2800                                     struct ptlrpc_request *req)
2801 {
2802         struct obd_export *exp = req->rq_export;
2803
2804         LASSERT(exp != NULL);
2805         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2806                 /* client declares he's ready to replay locks */
2807                 spin_lock(&exp->exp_lock);
2808                 if (exp->exp_req_replay_needed) {
2809                         exp->exp_req_replay_needed = 0;
2810                         spin_unlock(&exp->exp_lock);
2811
2812                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2813                         atomic_dec(&obd->obd_req_replay_clients);
2814                 } else {
2815                         spin_unlock(&exp->exp_lock);
2816                 }
2817         }
2818         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2819                 /*
2820                  * client declares he's ready to complete recovery
2821                  * so, we put the request on th final queue
2822                  */
2823                 spin_lock(&exp->exp_lock);
2824                 if (exp->exp_lock_replay_needed) {
2825                         exp->exp_lock_replay_needed = 0;
2826                         spin_unlock(&exp->exp_lock);
2827
2828                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2829                         atomic_dec(&obd->obd_lock_replay_clients);
2830                 } else {
2831                         spin_unlock(&exp->exp_lock);
2832                 }
2833         }
2834         return 0;
2835 }
2836
2837 int target_queue_recovery_request(struct ptlrpc_request *req,
2838                                   struct obd_device *obd)
2839 {
2840         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2841         struct ptlrpc_request *reqiter;
2842         int inserted = 0;
2843
2844         ENTRY;
2845
2846         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2847                 /* Processing the queue right now, don't re-add. */
2848                 RETURN(1);
2849         }
2850
2851         target_process_req_flags(obd, req);
2852
2853         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2854                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2855                         if (cfs_fail_val == 1) {
2856                                 cfs_race_state = 1;
2857                                 cfs_fail_val = 0;
2858                                 wake_up(&cfs_race_waitq);
2859
2860                                 set_current_state(TASK_INTERRUPTIBLE);
2861                                 schedule_timeout(cfs_time_seconds(1));
2862                         }
2863                 }
2864
2865                 /*
2866                  * client declares he's ready to complete recovery
2867                  * so, we put the request on th final queue
2868                  */
2869                 target_request_copy_get(req);
2870                 DEBUG_REQ(D_HA, req, "queue final req");
2871                 wake_up(&obd->obd_next_transno_waitq);
2872                 spin_lock(&obd->obd_recovery_task_lock);
2873                 if (obd->obd_recovering) {
2874                         struct ptlrpc_request *tmp;
2875                         struct ptlrpc_request *duplicate = NULL;
2876
2877                         if (likely(!req->rq_export->exp_replay_done)) {
2878                                 req->rq_export->exp_replay_done = 1;
2879                                 list_add_tail(&req->rq_list,
2880                                               &obd->obd_final_req_queue);
2881                                 spin_unlock(&obd->obd_recovery_task_lock);
2882                                 RETURN(0);
2883                         }
2884
2885                         /*
2886                          * XXX O(n), but only happens if final ping is
2887                          * timed out, probably reorganize the list as
2888                          * a hash list later
2889                          */
2890                         list_for_each_entry_safe(reqiter, tmp,
2891                                                  &obd->obd_final_req_queue,
2892                                                  rq_list) {
2893                                 if (reqiter->rq_export == req->rq_export) {
2894                                         list_del_init(&reqiter->rq_list);
2895                                         duplicate = reqiter;
2896                                         break;
2897                                 }
2898                         }
2899
2900                         list_add_tail(&req->rq_list,
2901                                       &obd->obd_final_req_queue);
2902                         req->rq_export->exp_replay_done = 1;
2903                         spin_unlock(&obd->obd_recovery_task_lock);
2904
2905                         if (duplicate != NULL) {
2906                                 DEBUG_REQ(D_HA, duplicate,
2907                                           "put prev final req");
2908                                 target_request_copy_put(duplicate);
2909                         }
2910                         RETURN(0);
2911                 } else {
2912                         spin_unlock(&obd->obd_recovery_task_lock);
2913                         target_request_copy_put(req);
2914                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2915                 }
2916         }
2917         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2918                 /* client declares he's ready to replay locks */
2919                 target_request_copy_get(req);
2920                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2921                 wake_up(&obd->obd_next_transno_waitq);
2922                 spin_lock(&obd->obd_recovery_task_lock);
2923                 LASSERT(obd->obd_recovering);
2924                 /* usually due to recovery abort */
2925                 if (!req->rq_export->exp_in_recovery) {
2926                         spin_unlock(&obd->obd_recovery_task_lock);
2927                         target_request_copy_put(req);
2928                         RETURN(-ENOTCONN);
2929                 }
2930                 LASSERT(req->rq_export->exp_lock_replay_needed);
2931                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2932                 spin_unlock(&obd->obd_recovery_task_lock);
2933                 RETURN(0);
2934         }
2935
2936         /*
2937          * CAVEAT EMPTOR: The incoming request message has been swabbed
2938          * (i.e. buflens etc are in my own byte order), but type-dependent
2939          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2940          */
2941
2942         if (!transno) {
2943                 INIT_LIST_HEAD(&req->rq_list);
2944                 DEBUG_REQ(D_HA, req, "not queueing");
2945                 RETURN(1);
2946         }
2947
2948         /*
2949          * If we're processing the queue, we want don't want to queue this
2950          * message.
2951          *
2952          * Also, if this request has a transno less than the one we're waiting
2953          * for, we should process it now.  It could (and currently always will)
2954          * be an open request for a descriptor that was opened some time ago.
2955          *
2956          * Also, a resent, replayed request that has already been
2957          * handled will pass through here and be processed immediately.
2958          */
2959         CDEBUG(D_HA,
2960                "Next recovery transno: %llu, current: %llu, replaying\n",
2961                obd->obd_next_recovery_transno, transno);
2962
2963         /*
2964          * If the request has been replayed by update replay, then sends this
2965          * request to the recovery thread (replay_request_or_update()), where
2966          * it will be handled
2967          */
2968         spin_lock(&obd->obd_recovery_task_lock);
2969         if (transno < obd->obd_next_recovery_transno &&
2970             !is_req_replayed_by_update(req)) {
2971                 /* Processing the queue right now, don't re-add. */
2972                 LASSERT(list_empty(&req->rq_list));
2973                 spin_unlock(&obd->obd_recovery_task_lock);
2974                 RETURN(1);
2975         }
2976         spin_unlock(&obd->obd_recovery_task_lock);
2977
2978         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2979                 RETURN(0);
2980
2981         target_request_copy_get(req);
2982         if (!req->rq_export->exp_in_recovery) {
2983                 target_request_copy_put(req);
2984                 RETURN(-ENOTCONN);
2985         }
2986         LASSERT(req->rq_export->exp_req_replay_needed);
2987
2988         if (target_exp_enqueue_req_replay(req)) {
2989                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2990                 target_request_copy_put(req);
2991                 RETURN(0);
2992         }
2993
2994         /* XXX O(n^2) */
2995         spin_lock(&obd->obd_recovery_task_lock);
2996         LASSERT(obd->obd_recovering);
2997         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2998                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2999                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3000                         inserted = 1;
3001                         goto added;
3002                 }
3003
3004                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3005                              transno)) {
3006                         DEBUG_REQ(D_ERROR, req,
3007                                   "dropping replay: transno has been claimed by another client");
3008                         spin_unlock(&obd->obd_recovery_task_lock);
3009                         target_exp_dequeue_req_replay(req);
3010                         target_request_copy_put(req);
3011                         RETURN(0);
3012                 }
3013         }
3014 added:
3015         if (!inserted)
3016                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3017
3018         obd->obd_requests_queued_for_recovery++;
3019         spin_unlock(&obd->obd_recovery_task_lock);
3020         wake_up(&obd->obd_next_transno_waitq);
3021         RETURN(0);
3022 }
3023
3024 void target_committed_to_req(struct ptlrpc_request *req)
3025 {
3026         struct obd_export *exp = req->rq_export;
3027
3028         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3029                 lustre_msg_set_last_committed(req->rq_repmsg,
3030                                               exp->exp_last_committed);
3031         else
3032                 DEBUG_REQ(D_IOCTL, req,
3033                           "not sending last_committed update (%d/%d)",
3034                           exp->exp_obd->obd_no_transno,
3035                           req->rq_repmsg == NULL);
3036
3037         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3038                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3039 }
3040
3041 #endif /* HAVE_SERVER_SUPPORT */
3042
3043 /**
3044  * Packs current SLV and Limit into \a req.
3045  */
3046 int target_pack_pool_reply(struct ptlrpc_request *req)
3047 {
3048         struct obd_device *obd;
3049
3050         ENTRY;
3051
3052         /*
3053          * Check that we still have all structures alive as this may
3054          * be some late RPC at shutdown time.
3055          */
3056         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3057                      !exp_connect_lru_resize(req->rq_export))) {
3058                 lustre_msg_set_slv(req->rq_repmsg, 0);
3059                 lustre_msg_set_limit(req->rq_repmsg, 0);
3060                 RETURN(0);
3061         }
3062
3063         /* OBD is alive here as export is alive, which we checked above. */
3064         obd = req->rq_export->exp_obd;
3065
3066         read_lock(&obd->obd_pool_lock);
3067         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3068         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3069         read_unlock(&obd->obd_pool_lock);
3070
3071         RETURN(0);
3072 }
3073
3074 static int target_send_reply_msg(struct ptlrpc_request *req,
3075                                  int rc, int fail_id)
3076 {
3077         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3078                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3079                 return -ECOMM;
3080         }
3081         /*
3082          * We can have a null rq_reqmsg in the event of bad signature or
3083          * no context when unwrapping
3084          */
3085         if (req->rq_reqmsg &&
3086             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3087             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3088                 return -ECOMM;
3089
3090         if (unlikely(rc)) {
3091                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3092                 req->rq_status = rc;
3093                 return ptlrpc_send_error(req, 1);
3094         }
3095         DEBUG_REQ(D_NET, req, "sending reply");
3096
3097         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3098 }
3099
3100 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3101 {
3102         struct ptlrpc_service_part *svcpt;
3103         int netrc;
3104         struct ptlrpc_reply_state *rs;
3105         struct obd_export *exp;
3106
3107         ENTRY;
3108
3109         if (req->rq_no_reply) {
3110                 EXIT;
3111                 return;
3112         }
3113
3114         svcpt = req->rq_rqbd->rqbd_svcpt;
3115         rs = req->rq_reply_state;
3116         if (rs == NULL || !rs->rs_difficult) {
3117                 /* no notifiers */
3118                 target_send_reply_msg(req, rc, fail_id);
3119                 EXIT;
3120                 return;
3121         }
3122
3123         /* must be an export if locks saved */
3124         LASSERT(req->rq_export != NULL);
3125         /* req/reply consistent */
3126         LASSERT(rs->rs_svcpt == svcpt);
3127
3128         /* "fresh" reply */
3129         LASSERT(!rs->rs_scheduled);
3130         LASSERT(!rs->rs_scheduled_ever);
3131         LASSERT(!rs->rs_handled);
3132         LASSERT(!rs->rs_on_net);
3133         LASSERT(rs->rs_export == NULL);
3134         LASSERT(list_empty(&rs->rs_obd_list));
3135         LASSERT(list_empty(&rs->rs_exp_list));
3136
3137         exp = class_export_get(req->rq_export);
3138
3139         /* disable reply scheduling while I'm setting up */
3140         rs->rs_scheduled = 1;
3141         rs->rs_on_net    = 1;
3142         rs->rs_xid       = req->rq_xid;
3143         rs->rs_transno   = req->rq_transno;
3144         rs->rs_export    = exp;
3145         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3146
3147         spin_lock(&exp->exp_uncommitted_replies_lock);
3148         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3149                rs->rs_transno, exp->exp_last_committed);
3150         if (rs->rs_transno > exp->exp_last_committed) {
3151                 /* not committed already */
3152                 list_add_tail(&rs->rs_obd_list,
3153                                   &exp->exp_uncommitted_replies);
3154         }
3155         spin_unlock(&exp->exp_uncommitted_replies_lock);
3156
3157         spin_lock(&exp->exp_lock);
3158         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3159         spin_unlock(&exp->exp_lock);
3160
3161         netrc = target_send_reply_msg(req, rc, fail_id);
3162
3163         spin_lock(&svcpt->scp_rep_lock);
3164
3165         atomic_inc(&svcpt->scp_nreps_difficult);
3166
3167         if (netrc != 0) {
3168                 /*
3169                  * error sending: reply is off the net.  Also we need +1
3170                  * reply ref until ptlrpc_handle_rs() is done
3171                  * with the reply state (if the send was successful, there
3172                  * would have been +1 ref for the net, which
3173                  * reply_out_callback leaves alone)
3174                  */
3175                 rs->rs_on_net = 0;
3176                 ptlrpc_rs_addref(rs);
3177         }
3178
3179         spin_lock(&rs->rs_lock);
3180         if (rs->rs_transno <= exp->exp_last_committed ||
3181             (!rs->rs_on_net && !rs->rs_no_ack) ||
3182             list_empty(&rs->rs_exp_list) ||     /* completed already */
3183             list_empty(&rs->rs_obd_list)) {
3184                 CDEBUG(D_HA, "Schedule reply immediately\n");
3185                 ptlrpc_dispatch_difficult_reply(rs);
3186         } else {
3187                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3188                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3189         }
3190         spin_unlock(&rs->rs_lock);
3191         spin_unlock(&svcpt->scp_rep_lock);
3192         EXIT;
3193 }
3194
3195 enum ldlm_mode lck_compat_array[] = {
3196         [LCK_EX]    = LCK_COMPAT_EX,
3197         [LCK_PW]    = LCK_COMPAT_PW,
3198         [LCK_PR]    = LCK_COMPAT_PR,
3199         [LCK_CW]    = LCK_COMPAT_CW,
3200         [LCK_CR]    = LCK_COMPAT_CR,
3201         [LCK_NL]    = LCK_COMPAT_NL,
3202         [LCK_GROUP] = LCK_COMPAT_GROUP,
3203         [LCK_COS]   = LCK_COMPAT_COS,
3204 };
3205
3206 /**
3207  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3208  * not escape to the user level.
3209  */
3210 int ldlm_error2errno(enum ldlm_error error)
3211 {
3212         int result;
3213
3214         switch (error) {
3215         case ELDLM_OK:
3216         case ELDLM_LOCK_MATCHED:
3217                 result = 0;
3218                 break;
3219         case ELDLM_LOCK_CHANGED:
3220                 result = -ESTALE;
3221                 break;
3222         case ELDLM_LOCK_ABORTED:
3223                 result = -ENAVAIL;
3224                 break;
3225         case ELDLM_LOCK_REPLACED:
3226                 result = -ESRCH;
3227                 break;
3228         case ELDLM_NO_LOCK_DATA:
3229                 result = -ENOENT;
3230                 break;
3231         case ELDLM_NAMESPACE_EXISTS:
3232                 result = -EEXIST;
3233                 break;
3234         case ELDLM_BAD_NAMESPACE:
3235                 result = -EBADF;
3236                 break;
3237         default:
3238                 if (((int)error) < 0) { /* cast to signed type */
3239                         result = error; /* as ldlm_error can be unsigned */
3240                 } else {
3241                         CERROR("Invalid DLM result code: %d\n", error);
3242                         result = -EPROTO;
3243                 }
3244         }
3245         return result;
3246 }
3247 EXPORT_SYMBOL(ldlm_error2errno);
3248
3249 /**
3250  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3251  */
3252 enum ldlm_error ldlm_errno2error(int err_no)
3253 {
3254         int error;
3255
3256         switch (err_no) {
3257         case 0:
3258                 error = ELDLM_OK;
3259                 break;
3260         case -ESTALE:
3261                 error = ELDLM_LOCK_CHANGED;
3262                 break;
3263         case -ENAVAIL:
3264                 error = ELDLM_LOCK_ABORTED;
3265                 break;
3266         case -ESRCH:
3267                 error = ELDLM_LOCK_REPLACED;
3268                 break;
3269         case -ENOENT:
3270                 error = ELDLM_NO_LOCK_DATA;
3271                 break;
3272         case -EEXIST:
3273                 error = ELDLM_NAMESPACE_EXISTS;
3274                 break;
3275         case -EBADF:
3276                 error = ELDLM_BAD_NAMESPACE;
3277                 break;
3278         default:
3279                 error = err_no;
3280         }
3281         return error;
3282 }
3283
3284 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3285 void ldlm_dump_export_locks(struct obd_export *exp)
3286 {
3287         spin_lock(&exp->exp_locks_list_guard);
3288         if (!list_empty(&exp->exp_locks_list)) {
3289                 struct ldlm_lock *lock;
3290
3291                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3292                        exp);
3293                 list_for_each_entry(lock, &exp->exp_locks_list,
3294                                         l_exp_refs_link)
3295                         LDLM_ERROR(lock, "lock:");
3296         }
3297         spin_unlock(&exp->exp_locks_list_guard);
3298 }
3299 #endif
3300
3301 #ifdef HAVE_SERVER_SUPPORT
3302 static int target_bulk_timeout(void *data)
3303 {
3304         ENTRY;
3305         /*
3306          * We don't fail the connection here, because having the export
3307          * killed makes the (vital) call to commitrw very sad.
3308          */
3309         RETURN(1);
3310 }
3311
3312 static inline const char *bulk2type(struct ptlrpc_request *req)
3313 {
3314         if (req->rq_bulk_read)
3315                 return "READ";
3316         if (req->rq_bulk_write)
3317                 return "WRITE";
3318         return "UNKNOWN";
3319 }
3320
3321 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3322 {
3323         struct ptlrpc_request *req = desc->bd_req;
3324         time64_t start = ktime_get_seconds();
3325         time64_t deadline;
3326         struct l_wait_info lwi;
3327         int rc = 0;
3328
3329         ENTRY;
3330
3331         /* If there is eviction in progress, wait for it to finish. */
3332         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3333                 lwi = LWI_INTR(NULL, NULL);
3334                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3335                                   !atomic_read(&exp->exp_obd->obd_evict_inprogress),
3336                                   &lwi);
3337         }
3338
3339         /* Check if client was evicted or reconnected already. */
3340         if (exp->exp_failed ||
3341             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3342                 rc = -ENOTCONN;
3343         } else {
3344                 if (req->rq_bulk_read)
3345                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3346
3347                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3348                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3349                 else /* old version, bulk matchbits is rq_xid */
3350                         req->rq_mbits = req->rq_xid;
3351
3352                 if (rc == 0)
3353                         rc = ptlrpc_start_bulk_transfer(desc);
3354         }
3355
3356         if (rc < 0) {
3357                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3358                           bulk2type(req), rc);
3359                 RETURN(rc);
3360         }
3361
3362         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3363                 ptlrpc_abort_bulk(desc);
3364                 RETURN(0);
3365         }
3366
3367         /* limit actual bulk transfer to bulk_timeout seconds */
3368         deadline = start + bulk_timeout;
3369         if (deadline > req->rq_deadline)
3370                 deadline = req->rq_deadline;
3371
3372         do {
3373                 time64_t timeoutl = deadline - ktime_get_seconds();
3374                 long timeout_jiffies = timeoutl <= 0 ?
3375                                        1 : cfs_time_seconds(timeoutl);
3376                 time64_t rq_deadline;
3377
3378                 lwi = LWI_TIMEOUT_INTERVAL(timeout_jiffies,
3379                                            cfs_time_seconds(1),
3380                                            target_bulk_timeout, desc);
3381                 rc = l_wait_event(desc->bd_waitq,
3382                                   !ptlrpc_server_bulk_active(desc) ||
3383                                   exp->exp_failed ||
3384                                   exp->exp_conn_cnt >
3385                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3386                                   &lwi);
3387                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3388                 /* Wait again if we changed rq_deadline. */
3389                 rq_deadline = READ_ONCE(req->rq_deadline);
3390                 deadline = start + bulk_timeout;
3391                 if (deadline > rq_deadline)
3392                         deadline = rq_deadline;
3393         } while (rc == -ETIMEDOUT &&
3394                  deadline > ktime_get_seconds());
3395
3396         if (rc == -ETIMEDOUT) {
3397                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3398                           bulk2type(req), deadline - start,
3399                           ktime_get_real_seconds() - deadline);
3400                 ptlrpc_abort_bulk(desc);
3401         } else if (exp->exp_failed) {
3402                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3403                           bulk2type(req));
3404                 rc = -ENOTCONN;
3405                 ptlrpc_abort_bulk(desc);
3406         } else if (exp->exp_conn_cnt >
3407                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3408                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3409                           bulk2type(req));
3410                 /* We don't reply anyway. */
3411                 rc = -ETIMEDOUT;
3412                 ptlrpc_abort_bulk(desc);
3413         } else if (desc->bd_failure) {
3414                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3415                           bulk2type(req));
3416                 /* XXX should this be a different errno? */
3417                 rc = -ETIMEDOUT;
3418         } else {
3419                 if (req->rq_bulk_write)
3420                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3421                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3422                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3423                                   bulk2type(req), desc->bd_nob_transferred,
3424                                   desc->bd_nob);
3425                         /* XXX should this be a different errno? */
3426                         rc = -ETIMEDOUT;
3427                 }
3428         }
3429
3430         RETURN(rc);
3431 }
3432 EXPORT_SYMBOL(target_bulk_io);
3433
3434 #endif /* HAVE_SERVER_SUPPORT */