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