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 *obd, struct lustre_cfg *lcfg)
 275 {
 276         struct client_obd *cli = &obd->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 = obd->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_waitqueue_head(&cli->cl_cache_waiters);
 382         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 383         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 384         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 385         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 386         spin_lock_init(&cli->cl_loi_list_lock);
 387         atomic_set(&cli->cl_pending_w_pages, 0);
 388         atomic_set(&cli->cl_pending_r_pages, 0);
 389         cli->cl_r_in_flight = 0;
 390         cli->cl_w_in_flight = 0;
 391
 392         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 393         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 394         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 395         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 396         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 397         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 398
 399         /* lru for osc. */
 400         INIT_LIST_HEAD(&cli->cl_lru_osc);
 401         atomic_set(&cli->cl_lru_shrinkers, 0);
 402         atomic_long_set(&cli->cl_lru_busy, 0);
 403         atomic_long_set(&cli->cl_lru_in_list, 0);
 404         INIT_LIST_HEAD(&cli->cl_lru_list);
 405         spin_lock_init(&cli->cl_lru_list_lock);
 406         atomic_long_set(&cli->cl_unstable_count, 0);
 407         INIT_LIST_HEAD(&cli->cl_shrink_list);
 408         INIT_LIST_HEAD(&cli->cl_grant_chain);
 409
 410         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 411         cli->cl_rpcs_in_flight = 0;
 412
 413         init_waitqueue_head(&cli->cl_destroy_waitq);
 414         atomic_set(&cli->cl_destroy_in_flight, 0);
 415
 416
 417         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 418         cli->cl_preferred_cksum_type = 0;
 419 #ifdef ENABLE_CHECKSUM
 420         /* Turn on checksumming by default. */
 421         cli->cl_checksum = 1;
 422         /*
 423          * The supported checksum types will be worked out at connect time
 424          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 425          * through procfs.
 426          */
 427         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 428 #endif
 429         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 430
 431         /*
 432          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 433          * from OFD after connecting.
 434          */
 435         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 436
 437         cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
 438
 439         /*
 440          * set cl_chunkbits default value to PAGE_SHIFT,
 441          * it will be updated at OSC connection time.
 442          */
 443         cli->cl_chunkbits = PAGE_SHIFT;
 444
 445         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 446                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 447         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 448                 cli->cl_max_rpcs_in_flight = 2;
 449         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 450                 cli->cl_max_rpcs_in_flight = 3;
 451         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 452                 cli->cl_max_rpcs_in_flight = 4;
 453         } else {
 454                 if (osc_on_mdt(obd->obd_name))
 455                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 456                 else
 457                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 458         }
 459
 460         spin_lock_init(&cli->cl_mod_rpcs_lock);
 461         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 462         cli->cl_max_mod_rpcs_in_flight = 0;
 463         cli->cl_mod_rpcs_in_flight = 0;
 464         cli->cl_close_rpcs_in_flight = 0;
 465         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 466         cli->cl_mod_tag_bitmap = NULL;
 467
 468         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 469
 470         if (connect_op == MDS_CONNECT) {
 471                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 472                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 473                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 474                 if (cli->cl_mod_tag_bitmap == NULL)
 475                         GOTO(err, rc = -ENOMEM);
 476         }
 477
 478         rc = ldlm_get_ref();
 479         if (rc) {
 480                 CERROR("ldlm_get_ref failed: %d\n", rc);
 481                 GOTO(err, rc);
 482         }
 483
 484         ptlrpc_init_client(rq_portal, rp_portal, name,
 485                            &obd->obd_ldlm_client);
 486
 487         imp = class_new_import(obd);
 488         if (imp == NULL)
 489                 GOTO(err_ldlm, rc = -ENOENT);
 490         imp->imp_client = &obd->obd_ldlm_client;
 491         imp->imp_connect_op = connect_op;
 492         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 493                LUSTRE_CFG_BUFLEN(lcfg, 1));
 494         class_import_put(imp);
 495
 496         if (lustre_cfg_buf(lcfg, 4)) {
 497                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 498
 499                 if (refnet == LNET_NIDNET(LNET_NID_ANY)) {
 500                         rc = -EINVAL;
 501                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 502                                obd->obd_name, lustre_cfg_string(lcfg, 4),
 503                                rc);
 504                         GOTO(err_import, rc);
 505                 }
 506                 fake_conn.c_self = LNET_MKNID(refnet, 0);
 507                 imp->imp_connection = &fake_conn;
 508         }
 509
 510         rc = client_import_add_conn(imp, &server_uuid, 1);
 511         if (rc) {
 512                 CERROR("can't add initial connection\n");
 513                 GOTO(err_import, rc);
 514         }
 515         imp->imp_connection = NULL;
 516
 517         cli->cl_import = imp;
 518         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 519         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 520
 521         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 522                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 523                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 524                                name, obd->obd_name,
 525                                cli->cl_target_uuid.uuid);
 526                         spin_lock(&imp->imp_lock);
 527                         imp->imp_deactive = 1;
 528                         spin_unlock(&imp->imp_lock);
 529                 }
 530         }
 531
 532         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 533                                                 LDLM_NAMESPACE_CLIENT,
 534                                                 LDLM_NAMESPACE_GREEDY,
 535                                                 ns_type);
 536         if (obd->obd_namespace == NULL) {
 537                 CERROR("Unable to create client namespace - %s\n",
 538                        obd->obd_name);
 539                 GOTO(err_import, rc = -ENOMEM);
 540         }
 541
 542         RETURN(rc);
 543
 544 err_import:
 545         class_destroy_import(imp);
 546 err_ldlm:
 547         ldlm_put_ref();
 548 err:
 549         if (cli->cl_mod_tag_bitmap != NULL)
 550                 OBD_FREE(cli->cl_mod_tag_bitmap,
 551                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 552         cli->cl_mod_tag_bitmap = NULL;
 553         RETURN(rc);
 554
 555 }
 556 EXPORT_SYMBOL(client_obd_setup);
 557
 558 int client_obd_cleanup(struct obd_device *obd)
 559 {
 560         struct client_obd *cli = &obd->u.cli;
 561
 562         ENTRY;
 563
 564         ldlm_namespace_free_post(obd->obd_namespace);
 565         obd->obd_namespace = NULL;
 566
 567         obd_cleanup_client_import(obd);
 568         LASSERT(obd->u.cli.cl_import == NULL);
 569
 570         ldlm_put_ref();
 571
 572         if (cli->cl_mod_tag_bitmap != NULL)
 573                 OBD_FREE(cli->cl_mod_tag_bitmap,
 574                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 575         cli->cl_mod_tag_bitmap = NULL;
 576
 577         RETURN(0);
 578 }
 579 EXPORT_SYMBOL(client_obd_cleanup);
 580
 581 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 582 int client_connect_import(const struct lu_env *env,
 583                           struct obd_export **exp,
 584                           struct obd_device *obd, struct obd_uuid *cluuid,
 585                           struct obd_connect_data *data, void *localdata)
 586 {
 587         struct client_obd *cli = &obd->u.cli;
 588         struct obd_import *imp = cli->cl_import;
 589         struct obd_connect_data *ocd;
 590         struct lustre_handle conn = { 0 };
 591         int rc;
 592
 593         ENTRY;
 594
 595         *exp = NULL;
 596         down_write(&cli->cl_sem);
 597         if (cli->cl_conn_count > 0)
 598                 GOTO(out_sem, rc = -EALREADY);
 599
 600         rc = class_connect(&conn, obd, cluuid);
 601         if (rc)
 602                 GOTO(out_sem, rc);
 603
 604         cli->cl_conn_count++;
 605         *exp = class_conn2export(&conn);
 606
 607         LASSERT(obd->obd_namespace);
 608
 609         imp->imp_dlm_handle = conn;
 610         rc = ptlrpc_init_import(imp);
 611         if (rc != 0)
 612                 GOTO(out_ldlm, rc);
 613
 614         ocd = &imp->imp_connect_data;
 615         if (data) {
 616                 *ocd = *data;
 617                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 618                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 619         }
 620
 621         rc = ptlrpc_connect_import(imp);
 622         if (rc != 0) {
 623                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 624                 GOTO(out_ldlm, rc);
 625         }
 626         LASSERT(*exp != NULL && (*exp)->exp_connection);
 627
 628         if (data) {
 629                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 630                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 631                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 632                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 633                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 634         }
 635
 636         ptlrpc_pinger_add_import(imp);
 637
 638         EXIT;
 639
 640         if (rc) {
 641 out_ldlm:
 642                 cli->cl_conn_count--;
 643                 class_disconnect(*exp);
 644                 *exp = NULL;
 645         }
 646 out_sem:
 647         up_write(&cli->cl_sem);
 648
 649         if (!rc && localdata) {
 650                 LASSERT(cli->cl_cache == NULL); /* only once */
 651                 cli->cl_cache = (struct cl_client_cache *)localdata;
 652                 cl_cache_incref(cli->cl_cache);
 653                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 654
 655                 /* add this osc into entity list */
 656                 LASSERT(list_empty(&cli->cl_lru_osc));
 657                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 658                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 659                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 660         }
 661
 662         return rc;
 663 }
 664 EXPORT_SYMBOL(client_connect_import);
 665
 666 int client_disconnect_export(struct obd_export *exp)
 667 {
 668         struct obd_device *obd = class_exp2obd(exp);
 669         struct client_obd *cli;
 670         struct obd_import *imp;
 671         int rc = 0, err;
 672
 673         ENTRY;
 674
 675         if (!obd) {
 676                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 677                        exp, exp ? exp->exp_handle.h_cookie : -1);
 678                 RETURN(-EINVAL);
 679         }
 680
 681         cli = &obd->u.cli;
 682         imp = cli->cl_import;
 683
 684         down_write(&cli->cl_sem);
 685         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 686                 cli->cl_conn_count);
 687
 688         if (cli->cl_conn_count == 0) {
 689                 CERROR("disconnecting disconnected device (%s)\n",
 690                        obd->obd_name);
 691                 GOTO(out_disconnect, rc = -EINVAL);
 692         }
 693
 694         cli->cl_conn_count--;
 695         if (cli->cl_conn_count != 0)
 696                 GOTO(out_disconnect, rc = 0);
 697
 698         /*
 699          * Mark import deactivated now, so we don't try to reconnect if any
 700          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 701          * fully deactivate the import, or that would drop all requests.
 702          */
 703         spin_lock(&imp->imp_lock);
 704         imp->imp_deactive = 1;
 705         spin_unlock(&imp->imp_lock);
 706
 707         /*
 708          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 709          * delete it regardless.  (It's safe to delete an import that was
 710          * never added.)
 711          */
 712         (void)ptlrpc_pinger_del_import(imp);
 713
 714         if (obd->obd_namespace != NULL) {
 715                 /* obd_force == local only */
 716                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 717                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 718                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 719                                           obd->obd_force);
 720         }
 721
 722         /*
 723          * There's no need to hold sem while disconnecting an import,
 724          * and it may actually cause deadlock in GSS.
 725          */
 726         up_write(&cli->cl_sem);
 727         rc = ptlrpc_disconnect_import(imp, 0);
 728         down_write(&cli->cl_sem);
 729
 730         ptlrpc_invalidate_import(imp);
 731
 732         EXIT;
 733
 734 out_disconnect:
 735         /*
 736          * Use server style - class_disconnect should be always called for
 737          * o_disconnect.
 738          */
 739         err = class_disconnect(exp);
 740         if (!rc && err)
 741                 rc = err;
 742
 743         up_write(&cli->cl_sem);
 744
 745         RETURN(rc);
 746 }
 747 EXPORT_SYMBOL(client_disconnect_export);
 748
 749 #ifdef HAVE_SERVER_SUPPORT
 750 int server_disconnect_export(struct obd_export *exp)
 751 {
 752         int rc;
 753
 754         ENTRY;
 755
 756         /* Disconnect early so that clients can't keep using export. */
 757         rc = class_disconnect(exp);
 758         /* Close import to avoid sending any requests. */
 759         if (exp->exp_imp_reverse)
 760                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 761
 762         ldlm_bl_thread_wakeup();
 763
 764         /* complete all outstanding replies */
 765         spin_lock(&exp->exp_lock);
 766         while (!list_empty(&exp->exp_outstanding_replies)) {
 767                 struct ptlrpc_reply_state *rs =
 768                         list_entry(exp->exp_outstanding_replies.next,
 769                                        struct ptlrpc_reply_state, rs_exp_list);
 770                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 771
 772                 spin_lock(&svcpt->scp_rep_lock);
 773
 774                 list_del_init(&rs->rs_exp_list);
 775
 776                 spin_lock(&rs->rs_lock);
 777                 /* clear rs_convert_lock to make sure rs is handled and put */
 778                 rs->rs_convert_lock = 0;
 779                 ptlrpc_schedule_difficult_reply(rs);
 780                 spin_unlock(&rs->rs_lock);
 781
 782                 spin_unlock(&svcpt->scp_rep_lock);
 783         }
 784         spin_unlock(&exp->exp_lock);
 785
 786         RETURN(rc);
 787 }
 788 EXPORT_SYMBOL(server_disconnect_export);
 789
 790 static inline int target_check_recovery_timer(struct obd_device *target)
 791 {
 792         ktime_t remaining;
 793         s64 timeout;
 794
 795         if (!target->obd_recovering || target->obd_recovery_start == 0)
 796                 return 0;
 797
 798         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 799         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 800         if (timeout > -30)
 801                 return 0;
 802
 803         /* the recovery timer should expire, but it isn't triggered,
 804          * it's better to abort the recovery of this target to speed up
 805          * the recovery of the whole cluster. */
 806         spin_lock(&target->obd_dev_lock);
 807         if (target->obd_recovering) {
 808                 CERROR("%s: Aborting recovery\n", target->obd_name);
 809                 target->obd_abort_recovery = 1;
 810                 wake_up(&target->obd_next_transno_waitq);
 811         }
 812         spin_unlock(&target->obd_dev_lock);
 813         return 0;
 814 }
 815
 816 /*
 817  * --------------------------------------------------------------------------
 818  * from old lib/target.c
 819  * --------------------------------------------------------------------------
 820  */
 821 static int target_handle_reconnect(struct lustre_handle *conn,
 822                                    struct obd_export *exp,
 823                                    struct obd_uuid *cluuid)
 824 {
 825         struct obd_device *target;
 826         struct lustre_handle *hdl;
 827         ktime_t remaining;
 828         s64 timeout;
 829         int rc = 0;
 830
 831         ENTRY;
 832         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 833         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 834                 conn->cookie = exp->exp_handle.h_cookie;
 835                 CDEBUG(D_HA,
 836                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 837                        cluuid->uuid, exp, conn->cookie);
 838                 RETURN(0);
 839         }
 840
 841         target = exp->exp_obd;
 842
 843         /* Might be a re-connect after a partition. */
 844         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 845                 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",
 846                               target->obd_name,
 847                               obd_uuid2str(&exp->exp_client_uuid),
 848                               obd_export_nid2str(exp),
 849                               hdl->cookie, conn->cookie);
 850                 memset(conn, 0, sizeof(*conn));
 851                 /*
 852                  * target_handle_connect() treats EALREADY and
 853                  * -EALREADY differently.  -EALREADY is an error
 854                  * (same UUID, different handle).
 855                  */
 856                 RETURN(-EALREADY);
 857         }
 858
 859         if (!target->obd_recovering) {
 860                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 861                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 862                         obd_export_nid2str(exp));
 863                 GOTO(out_already, rc);
 864         }
 865
 866         remaining = hrtimer_expires_remaining(&target->obd_recovery_timer);
 867         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 868         if (timeout > 0) {
 869                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 870                               target->obd_name,
 871                               obd_uuid2str(&exp->exp_client_uuid),
 872                               obd_export_nid2str(exp),
 873                               atomic_read(&target->obd_max_recoverable_clients),
 874                               timeout / 60, timeout % 60);
 875         } else {
 876                 struct target_distribute_txn_data *tdtd;
 877                 int size = 0;
 878                 int count = 0;
 879                 char *buf = NULL;
 880
 881                 target_check_recovery_timer(target);
 882
 883                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 884                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 885                         buf = tdtd->tdtd_show_update_logs_retrievers(
 886                                 tdtd->tdtd_show_retrievers_cbdata,
 887                                 &size, &count);
 888
 889                 if (count > 0)
 890                         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",
 891                                       target->obd_name,
 892                                       obd_uuid2str(&exp->exp_client_uuid),
 893                                       obd_export_nid2str(exp), count,
 894                                       buf ? buf : "unknown (not enough RAM)",
 895                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 896                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 897                                       target->obd_name);
 898                 else
 899                         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",
 900                                       target->obd_name, abs(timeout) / 60,
 901                                       abs(timeout) % 60, target->obd_name);
 902
 903                 if (buf != NULL)
 904                         OBD_FREE(buf, size);
 905         }
 906
 907 out_already:
 908         conn->cookie = exp->exp_handle.h_cookie;
 909         /*
 910          * target_handle_connect() treats EALREADY and
 911          * -EALREADY differently.  EALREADY means we are
 912          * doing a valid reconnect from the same client.
 913          */
 914         RETURN(EALREADY);
 915 }
 916
 917 static void
 918 check_and_start_recovery_timer(struct obd_device *obd,
 919                                struct ptlrpc_request *req, int new_client);
 920
 921 /**
 922  * update flags for import during reconnect process
 923  */
 924 static int rev_import_flags_update(struct obd_import *revimp,
 925                                    struct ptlrpc_request *req)
 926 {
 927         int rc;
 928         struct obd_connect_data *data;
 929
 930         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 931
 932         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 933                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 934         else
 935                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 936
 937         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 938
 939         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 940         if (rc) {
 941                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 942                         revimp->imp_client->cli_name,
 943                         libcfs_id2str(req->rq_peer), rc);
 944                 return rc;
 945         }
 946
 947         return 0;
 948 }
 949
 950 /**
 951  * Allocate a new reverse import for an export.
 952  *
 953  * \retval -errno in case error hit
 954  * \retval 0 if reverse import correctly init
 955  **/
 956 int rev_import_init(struct obd_export *export)
 957 {
 958         struct obd_device *obd = export->exp_obd;
 959         struct obd_import *revimp;
 960
 961         LASSERT(export->exp_imp_reverse == NULL);
 962
 963         revimp = class_new_import(obd);
 964         if (revimp == NULL)
 965                 return -ENOMEM;
 966
 967         revimp->imp_remote_handle.cookie = 0ULL;
 968         revimp->imp_client = &obd->obd_ldlm_client;
 969         revimp->imp_dlm_fake = 1;
 970
 971         /* it is safe to connect import in new state as no sends possible */
 972         spin_lock(&export->exp_lock);
 973         export->exp_imp_reverse = revimp;
 974         spin_unlock(&export->exp_lock);
 975         class_import_put(revimp);
 976
 977         return 0;
 978 }
 979 EXPORT_SYMBOL(rev_import_init);
 980
 981 /**
 982  * Handle reconnect for an export.
 983  *
 984  * \param exp export to handle reconnect process
 985  * \param req client reconnect request
 986  *
 987  * \retval -rc in case securitfy flavor can't be changed
 988  * \retval 0 in case none problems
 989  */
 990 static int rev_import_reconnect(struct obd_export *exp,
 991                                 struct ptlrpc_request *req)
 992 {
 993         struct obd_import *revimp = exp->exp_imp_reverse;
 994         struct lustre_handle *lh;
 995         int rc;
 996
 997         /* avoid sending a request until import flags are changed */
 998         ptlrpc_import_enter_resend(revimp);
 999
1000         ptlrpc_connection_put(revimp->imp_connection);
1001
1002         /*
1003          * client from recovery don't have a handle so we need to take from
1004          * request. it may produce situation when wrong client connected
1005          * to recovery as we trust a client uuid
1006          */
1007         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1008         revimp->imp_remote_handle = *lh;
1009
1010         /*
1011          * unknown versions will be caught in
1012          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1013          */
1014         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1015
1016         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1017
1018         rc = rev_import_flags_update(revimp, req);
1019         if (rc != 0) {
1020                 /*
1021                  * it is safe to still be in RECOVERY phase as we are not able
1022                  * to setup correct security flavor so requests are not able to
1023                  * be delivered correctly
1024                  */
1025                 return rc;
1026         }
1027
1028         /* resend all rpc's via new connection */
1029         return ptlrpc_import_recovery_state_machine(revimp);
1030 }
1031
1032 int target_handle_connect(struct ptlrpc_request *req)
1033 {
1034         struct obd_device *target = NULL;
1035         struct obd_export *export = NULL;
1036         /*
1037          * connect handle - filled from target_handle_reconnect in
1038          * reconnect case
1039          */
1040         struct lustre_handle conn;
1041         struct lustre_handle *tmp;
1042         struct obd_uuid cluuid;
1043         char *str;
1044         int rc = 0;
1045         char *target_start;
1046         int target_len;
1047         bool mds_conn = false, lw_client = false, initial_conn = false;
1048         bool mds_mds_conn = false;
1049         bool new_mds_mds_conn = false;
1050         struct obd_connect_data *data, *tmpdata;
1051         int size, tmpsize;
1052         lnet_nid_t *client_nid = NULL;
1053
1054         ENTRY;
1055
1056         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1057
1058         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1059         if (str == NULL) {
1060                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1061                 GOTO(out, rc = -EINVAL);
1062         }
1063
1064         target = class_dev_by_str(str);
1065         if (!target) {
1066                 deuuidify(str, NULL, &target_start, &target_len);
1067                 LCONSOLE_ERROR_MSG(0x137,
1068                                    "%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",
1069                                    str, libcfs_nid2str(req->rq_peer.nid));
1070                 GOTO(out, rc = -ENODEV);
1071         }
1072
1073         spin_lock(&target->obd_dev_lock);
1074
1075         target->obd_conn_inprogress++;
1076
1077         if (target->obd_stopping || !target->obd_set_up) {
1078                 spin_unlock(&target->obd_dev_lock);
1079
1080                 deuuidify(str, NULL, &target_start, &target_len);
1081                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1082                               target_len, target_start,
1083                               libcfs_nid2str(req->rq_peer.nid),
1084                               (target->obd_stopping ?
1085                                "stopping" : "not set up"));
1086                 GOTO(out, rc = -ENODEV);
1087         }
1088
1089         if (target->obd_no_conn) {
1090                 spin_unlock(&target->obd_dev_lock);
1091
1092                 CDEBUG(D_INFO,
1093                        "%s: Temporarily refusing client connection from %s\n",
1094                        target->obd_name, libcfs_nid2str(req->rq_peer.nid));
1095                 GOTO(out, rc = -EAGAIN);
1096         }
1097
1098         spin_unlock(&target->obd_dev_lock);
1099
1100         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1101         if (str == NULL) {
1102                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1103                 GOTO(out, rc = -EINVAL);
1104         }
1105
1106         obd_str2uuid(&cluuid, str);
1107
1108         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1109         if (tmp == NULL)
1110                 GOTO(out, rc = -EPROTO);
1111
1112         conn = *tmp;
1113
1114         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1115                                     RCL_CLIENT);
1116         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1117                 GOTO(out, rc = -EPROTO);
1118         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1119         if (!data)
1120                 GOTO(out, rc = -EPROTO);
1121
1122         rc = req_capsule_server_pack(&req->rq_pill);
1123         if (rc)
1124                 GOTO(out, rc);
1125
1126 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1127         /*
1128          * Don't allow clients to connect that are using old 1.8 format
1129          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1130          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1131          * FULL20 flag should be set on all connections since 2.0, but no
1132          * longer affects behaviour.
1133          *
1134          * Later this check will be disabled and the flag can be retired
1135          * completely once interop with 3.0 is no longer needed.
1136          */
1137         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1138                 GOTO(out, rc = -EPROTO);
1139
1140         /*
1141          * Don't allow liblustre clients to connect.
1142          * - testing was disabled in v2_2_50_0-61-g6a75d65
1143          * - building was disabled in v2_5_58_0-28-g7277179
1144          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1145          * - clients were refused connect for version difference > 0.0.1.32
1146          */
1147         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1148                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1149                 GOTO(out, rc = -EPROTO);
1150         }
1151 #endif
1152
1153         /*
1154          * Note: lw_client is needed in MDS-MDS failover during update log
1155          * processing, so we needs to allow lw_client to be connected at
1156          * anytime, instead of only the initial connection
1157          */
1158         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1159
1160         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1161                 initial_conn = true;
1162                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1163                 mds_mds_conn = (data->ocd_connect_flags &
1164                                 OBD_CONNECT_MDS_MDS) != 0;
1165
1166                 /*
1167                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1168                  * Checking OBD_CONNECT_FID can be removed in the future.
1169                  *
1170                  * Via check OBD_CONNECT_FID, we can distinguish whether
1171                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1172                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1173                  */
1174                 if (!lw_client &&
1175                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1176                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1177                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1178                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1179                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1180                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1181
1182                         /*
1183                          * We do not support the MDT-MDT interoperations with
1184                          * different version MDT because of protocol changes.
1185                          */
1186                         if (unlikely(major != LUSTRE_MAJOR ||
1187                                      minor != LUSTRE_MINOR ||
1188                                      abs(patch - LUSTRE_PATCH) > 3)) {
1189                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1190                                               target->obd_name, LUSTRE_MAJOR,
1191                                               LUSTRE_MINOR, LUSTRE_PATCH,
1192                                               LUSTRE_FIX, major, minor, patch,
1193                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1194                                               libcfs_nid2str(req->rq_peer.nid),
1195                                               str);
1196                                 GOTO(out, rc = -EPROTO);
1197                         }
1198                 }
1199         }
1200
1201         /* lctl gets a backstage, all-access pass. */
1202         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1203                 goto dont_check_exports;
1204
1205         export = obd_uuid_lookup(target, &cluuid);
1206         if (!export)
1207                 goto no_export;
1208
1209         /* We've found an export in the hash. */
1210
1211         spin_lock(&export->exp_lock);
1212
1213         if (export->exp_connecting) { /* b=9635, et. al. */
1214                 spin_unlock(&export->exp_lock);
1215                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1216                               export->exp_obd->obd_name, export,
1217                               libcfs_nid2str(req->rq_peer.nid));
1218                 class_export_put(export);
1219                 export = NULL;
1220                 rc = -EALREADY;
1221         } else if ((mds_conn || (lw_client && initial_conn) ||
1222                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1223                    export->exp_connection != NULL) {
1224                 spin_unlock(&export->exp_lock);
1225                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1226                         /* MDS or LWP reconnected after failover. */
1227                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1228                                       target->obd_name,
1229                                       mds_conn ? "MDS" : "LWP",
1230                                       libcfs_nid2str(req->rq_peer.nid),
1231                                       libcfs_nid2str(export->exp_connection->c_peer.nid));
1232                 } else {
1233                         /* New MDS connection from the same NID. */
1234                         LCONSOLE_WARN("%s: Received new %s connection from %s, removing former export from same NID\n",
1235                                       target->obd_name,
1236                                       mds_conn ? "MDS" : "LWP",
1237                                       libcfs_nid2str(req->rq_peer.nid));
1238                 }
1239
1240                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1241                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1242                         /*
1243                          * Because exports between MDTs will always be
1244                          * kept, let's do not fail such export if they
1245                          * come from the same NID, otherwise it might
1246                          * cause eviction between MDTs, which might
1247                          * cause namespace inconsistency
1248                          */
1249                         spin_lock(&export->exp_lock);
1250                         export->exp_connecting = 1;
1251                         export->exp_conn_cnt = 0;
1252                         spin_unlock(&export->exp_lock);
1253                         conn.cookie = export->exp_handle.h_cookie;
1254                         rc = EALREADY;
1255                 } else {
1256                         class_fail_export(export);
1257                         class_export_put(export);
1258                         export = NULL;
1259                         rc = 0;
1260                 }
1261         } else if (export->exp_connection != NULL && initial_conn &&
1262                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1263                 spin_unlock(&export->exp_lock);
1264                 /* In MDS failover we have static UUID but NID can change. */
1265                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1266                               target->obd_name, cluuid.uuid,
1267                               libcfs_nid2str(req->rq_peer.nid),
1268                               libcfs_nid2str(
1269                                         export->exp_connection->c_peer.nid));
1270                 rc = -EALREADY;
1271                 class_export_put(export);
1272                 export = NULL;
1273         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1274                    !lw_client) {
1275                 spin_unlock(&export->exp_lock);
1276                 rc = -EAGAIN;
1277         } else {
1278                 export->exp_connecting = 1;
1279                 spin_unlock(&export->exp_lock);
1280                 LASSERT(export->exp_obd == target);
1281
1282                 rc = target_handle_reconnect(&conn, export, &cluuid);
1283         }
1284
1285         /* If we found an export, we already unlocked. */
1286         if (!export) {
1287 no_export:
1288                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1289         } else if (req->rq_export == NULL &&
1290                    atomic_read(&export->exp_rpc_count) > 0) {
1291                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1292                               target->obd_name, cluuid.uuid,
1293                               libcfs_nid2str(req->rq_peer.nid),
1294                               refcount_read(&export->exp_handle.h_ref));
1295                         GOTO(out, rc = -EBUSY);
1296         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1297                    rc != EALREADY) {
1298                 if (!strstr(cluuid.uuid, "mdt"))
1299                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client",
1300                                       target->obd_name, cluuid.uuid,
1301                                       libcfs_nid2str(req->rq_peer.nid));
1302                 GOTO(out, rc = -EALREADY);
1303         } else {
1304                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1305         }
1306
1307         if (rc < 0)
1308                 GOTO(out, rc);
1309
1310         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1311                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1312                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1313                export, ktime_get_seconds(),
1314                export ? export->exp_last_request_time : 0);
1315
1316         /*
1317          * If this is the first time a client connects, reset the recovery
1318          * timer. Discard lightweight connections which might be local.
1319          */
1320         if (!lw_client && rc == 0 && target->obd_recovering)
1321                 check_and_start_recovery_timer(target, req, export == NULL);
1322
1323         /*
1324          * We want to handle EALREADY but *not* -EALREADY from
1325          * target_handle_reconnect(), return reconnection state in a flag.
1326          */
1327         if (rc == EALREADY) {
1328                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1329                 rc = 0;
1330         } else {
1331                 LASSERT(rc == 0);
1332         }
1333
1334         /* Tell the client if we support replayable requests. */
1335         if (target->obd_replayable)
1336                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1337         client_nid = &req->rq_peer.nid;
1338
1339         if (export == NULL) {
1340                 /* allow lightweight connections during recovery */
1341                 /*
1342                  * allow "new" MDT to be connected during recovery, since we
1343                  * need retrieve recovery update records from it
1344                  */
1345                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1346                         struct hrtimer *timer = &target->obd_recovery_timer;
1347                         ktime_t remaining;
1348                         s64 timeout, left;
1349                         int in_progress;
1350                         int connected;
1351                         int known;
1352                         int stale;
1353                         char *msg;
1354
1355                         connected = atomic_read(&target->obd_connected_clients);
1356                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1357                         known =
1358                            atomic_read(&target->obd_max_recoverable_clients);
1359                         stale = target->obd_stale_clients;
1360                         remaining = hrtimer_expires_remaining(timer);
1361                         left = ktime_divns(remaining, NSEC_PER_SEC);
1362
1363                         if (ktime_to_ns(remaining) > 0) {
1364                                 msg = "to recover in";
1365                                 timeout = left;
1366                         } else {
1367                                 msg = "already passed deadline";
1368                                 timeout = -left;
1369
1370                                 target_check_recovery_timer(target);
1371                         }
1372
1373                         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",
1374                                       target->obd_name, cluuid.uuid,
1375                                       libcfs_nid2str(req->rq_peer.nid), known,
1376                                       connected - in_progress, in_progress,
1377                                       stale, msg, timeout / 60, timeout % 60);
1378                         rc = -EBUSY;
1379                 } else {
1380 dont_check_exports:
1381                         rc = obd_connect(req->rq_svc_thread->t_env,
1382                                          &export, target, &cluuid, data,
1383                                          client_nid);
1384                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1385                                 lustre_msg_add_op_flags(req->rq_repmsg,
1386                                                         MSG_CONNECT_RECOVERING);
1387                         if (rc == 0) {
1388                                 conn.cookie = export->exp_handle.h_cookie;
1389                                 rc = rev_import_init(export);
1390                         }
1391
1392                         if (mds_mds_conn)
1393                                 new_mds_mds_conn = true;
1394                 }
1395         } else {
1396                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1397                                    export, target, &cluuid, data, client_nid);
1398         }
1399         if (rc)
1400                 GOTO(out, rc);
1401
1402         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1403         data->ocd_instance = target->u.obt.obt_instance;
1404
1405         /*
1406          * Return only the parts of obd_connect_data that we understand, so the
1407          * client knows that we don't understand the rest.
1408          */
1409         if (data) {
1410                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1411                                                RCL_SERVER);
1412                 tmpdata = req_capsule_server_get(&req->rq_pill,
1413                                                  &RMF_CONNECT_DATA);
1414                 /*
1415                  * Don't use struct assignment here, because the client reply
1416                  * buffer may be smaller/larger than the local struct
1417                  * obd_connect_data.
1418                  */
1419                 memcpy(tmpdata, data, min(tmpsize, size));
1420         }
1421
1422         /*
1423          * If the client and the server are the same node, we will already
1424          * have an export that really points to the client's DLM export,
1425          * because we have a shared handles table.
1426          *
1427          * XXX this will go away when shaver stops sending the "connect" handle
1428          * in the real "remote handle" field of the request --phik 24 Apr 2003
1429          */
1430         ptlrpc_request_change_export(req, export);
1431
1432         spin_lock(&export->exp_lock);
1433         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1434                 spin_unlock(&export->exp_lock);
1435                 CDEBUG(D_RPCTRACE,
1436                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1437                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1438                        export->exp_conn_cnt,
1439                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1440
1441                 GOTO(out, rc = -EALREADY);
1442         }
1443         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1444         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1445         spin_unlock(&export->exp_lock);
1446
1447         if (export->exp_connection != NULL) {
1448                 /* Check to see if connection came from another NID. */
1449                 if (export->exp_connection->c_peer.nid != req->rq_peer.nid)
1450                         obd_nid_del(export->exp_obd, export);
1451
1452                 ptlrpc_connection_put(export->exp_connection);
1453         }
1454
1455         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1456                                                        req->rq_self,
1457                                                        &cluuid);
1458         if (!export->exp_connection)
1459                 GOTO(out, rc = -ENOTCONN);
1460
1461         obd_nid_add(export->exp_obd, export);
1462
1463         lustre_msg_set_handle(req->rq_repmsg, &conn);
1464
1465         rc = rev_import_reconnect(export, req);
1466         if (rc != 0)
1467                 GOTO(out, rc);
1468
1469         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1470                 int has_transno;
1471                 __u64 transno = data->ocd_transno;
1472
1473                 spin_lock(&export->exp_lock);
1474                 /*
1475                  * possible race with class_disconnect_stale_exports,
1476                  * export may be already in the eviction process
1477                  */
1478                 if (export->exp_failed) {
1479                         spin_unlock(&export->exp_lock);
1480                         GOTO(out, rc = -ENODEV);
1481                 }
1482                 export->exp_in_recovery = 1;
1483                 export->exp_req_replay_needed = 1;
1484                 export->exp_lock_replay_needed = 1;
1485                 spin_unlock(&export->exp_lock);
1486
1487                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1488                                  MSG_CONNECT_TRANSNO);
1489                 if (has_transno && transno == 0)
1490                         CWARN("Connect with zero transno!\n");
1491
1492                 if (has_transno && transno > 0 &&
1493                     transno < target->obd_next_recovery_transno &&
1494                     transno > target->obd_last_committed) {
1495                         /* Another way is to use cmpxchg() to be lock-free. */
1496                         spin_lock(&target->obd_recovery_task_lock);
1497                         if (transno < target->obd_next_recovery_transno)
1498                                 target->obd_next_recovery_transno = transno;
1499                         spin_unlock(&target->obd_recovery_task_lock);
1500                 }
1501
1502                 atomic_inc(&target->obd_req_replay_clients);
1503                 atomic_inc(&target->obd_lock_replay_clients);
1504                 /*
1505                  * Note: MDS-MDS connection is allowed to be connected during
1506                  * recovery, no matter if the exports needs to be recoveried.
1507                  * Because we need retrieve updates logs from all other MDTs.
1508                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1509                  * also needs to be increased to match other recovery checking
1510                  * condition.
1511                  */
1512                 if (new_mds_mds_conn)
1513                         atomic_inc(&target->obd_max_recoverable_clients);
1514
1515                 if (atomic_inc_return(&target->obd_connected_clients) ==
1516                     atomic_read(&target->obd_max_recoverable_clients))
1517                         wake_up(&target->obd_next_transno_waitq);
1518         }
1519
1520         /* Tell the client we're in recovery, when client is involved in it. */
1521         if (target->obd_recovering && !lw_client)
1522                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1523
1524 out:
1525         if (export) {
1526                 spin_lock(&export->exp_lock);
1527                 export->exp_connecting = 0;
1528                 spin_unlock(&export->exp_lock);
1529
1530                 class_export_put(export);
1531         }
1532         if (target != NULL) {
1533                 spin_lock(&target->obd_dev_lock);
1534                 target->obd_conn_inprogress--;
1535                 spin_unlock(&target->obd_dev_lock);
1536                 class_decref(target, "find", current);
1537         }
1538         req->rq_status = rc;
1539         RETURN(rc);
1540 }
1541
1542 int target_handle_disconnect(struct ptlrpc_request *req)
1543 {
1544         int rc;
1545
1546         ENTRY;
1547
1548         rc = req_capsule_server_pack(&req->rq_pill);
1549         if (rc)
1550                 RETURN(rc);
1551
1552         /* Keep the rq_export around so we can send the reply. */
1553         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1554
1555         RETURN(0);
1556 }
1557
1558 void target_destroy_export(struct obd_export *exp)
1559 {
1560         struct obd_import *imp = NULL;
1561         /*
1562          * exports created from last_rcvd data, and "fake"
1563          * exports created by lctl don't have an import
1564          */
1565         spin_lock(&exp->exp_lock);
1566         if (exp->exp_imp_reverse != NULL) {
1567                 imp = exp->exp_imp_reverse;
1568                 exp->exp_imp_reverse = NULL;
1569         }
1570         spin_unlock(&exp->exp_lock);
1571         if (imp != NULL)
1572                 client_destroy_import(imp);
1573
1574         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1575         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1576         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1577         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1578 }
1579 EXPORT_SYMBOL(target_destroy_export);
1580
1581 /*
1582  * Recovery functions
1583  */
1584 static void target_request_copy_get(struct ptlrpc_request *req)
1585 {
1586         class_export_rpc_inc(req->rq_export);
1587         LASSERT(list_empty(&req->rq_list));
1588         INIT_LIST_HEAD(&req->rq_replay_list);
1589
1590         /* Increase refcount to keep request in queue. */
1591         atomic_inc(&req->rq_refcount);
1592         /* Let export know it has replays to be handled. */
1593         atomic_inc(&req->rq_export->exp_replay_count);
1594 }
1595
1596 static void target_request_copy_put(struct ptlrpc_request *req)
1597 {
1598         LASSERT(list_empty(&req->rq_replay_list));
1599         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1600
1601         atomic_dec(&req->rq_export->exp_replay_count);
1602         class_export_rpc_dec(req->rq_export);
1603         ptlrpc_server_drop_request(req);
1604 }
1605
1606 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1607 {
1608         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1609         struct obd_export *exp = req->rq_export;
1610         struct ptlrpc_request *reqiter;
1611         struct ptlrpc_request *dup_req = NULL;
1612         int dup = 0;
1613
1614         LASSERT(exp);
1615
1616         spin_lock(&exp->exp_lock);
1617         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1618                             rq_replay_list) {
1619                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1620                         dup_req = reqiter;
1621                         dup = 1;
1622                         break;
1623                 }
1624         }
1625
1626         if (dup) {
1627                 /* We expect it with RESENT and REPLAY flags. */
1628                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1629                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1630                         CERROR("invalid flags %x of resent replay\n",
1631                                lustre_msg_get_flags(req->rq_reqmsg));
1632
1633                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1634                         __u32 new_conn;
1635
1636                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1637                         if (new_conn >
1638                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1639                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1640                                                         new_conn);
1641                 }
1642         } else {
1643                 list_add_tail(&req->rq_replay_list,
1644                               &exp->exp_req_replay_queue);
1645         }
1646
1647         spin_unlock(&exp->exp_lock);
1648         return dup;
1649 }
1650
1651 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1652 {
1653         LASSERT(!list_empty(&req->rq_replay_list));
1654         LASSERT(req->rq_export);
1655
1656         spin_lock(&req->rq_export->exp_lock);
1657         list_del_init(&req->rq_replay_list);
1658         spin_unlock(&req->rq_export->exp_lock);
1659 }
1660
1661 static void target_finish_recovery(struct lu_target *lut)
1662 {
1663         struct obd_device *obd = lut->lut_obd;
1664
1665         ENTRY;
1666
1667         /* Only log a recovery message when recovery has occurred. */
1668         if (obd->obd_recovery_start) {
1669                 time64_t now = ktime_get_seconds();
1670                 time64_t elapsed_time;
1671
1672                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1673                                      1);
1674                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1675                               obd->obd_name, elapsed_time / 60,
1676                               elapsed_time % 60,
1677                               atomic_read(&obd->obd_max_recoverable_clients),
1678                               atomic_read(&obd->obd_connected_clients),
1679                               obd->obd_stale_clients,
1680                               obd->obd_stale_clients == 1 ? "was" : "were");
1681         }
1682
1683         ldlm_reprocess_recovery_done(obd->obd_namespace);
1684         spin_lock(&obd->obd_recovery_task_lock);
1685         if (!list_empty(&obd->obd_req_replay_queue) ||
1686             !list_empty(&obd->obd_lock_replay_queue) ||
1687             !list_empty(&obd->obd_final_req_queue)) {
1688                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1689                        obd->obd_name,
1690                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1691                        list_empty(&obd->obd_lock_replay_queue) ? \
1692                                   "" : "lock ",
1693                        list_empty(&obd->obd_final_req_queue) ? \
1694                                   "" : "final ");
1695                 spin_unlock(&obd->obd_recovery_task_lock);
1696                 LBUG();
1697         }
1698         spin_unlock(&obd->obd_recovery_task_lock);
1699
1700         obd->obd_recovery_end = ktime_get_seconds();
1701
1702         /* When recovery finished, cleanup orphans on MDS and OST. */
1703         if (obd->obd_type && OBP(obd, postrecov)) {
1704                 int rc = OBP(obd, postrecov)(obd);
1705
1706                 if (rc < 0)
1707                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1708                                       obd->obd_name, rc);
1709         }
1710         EXIT;
1711 }
1712
1713 static void abort_req_replay_queue(struct obd_device *obd)
1714 {
1715         struct ptlrpc_request *req, *n;
1716         LIST_HEAD(abort_list);
1717
1718         spin_lock(&obd->obd_recovery_task_lock);
1719         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1720         spin_unlock(&obd->obd_recovery_task_lock);
1721         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1722                 DEBUG_REQ(D_WARNING, req, "aborted:");
1723                 req->rq_status = -ENOTCONN;
1724                 if (ptlrpc_error(req)) {
1725                         DEBUG_REQ(D_ERROR, req,
1726                                   "failed abort_req_reply; skipping");
1727                 }
1728                 target_exp_dequeue_req_replay(req);
1729                 target_request_copy_put(req);
1730         }
1731 }
1732
1733 static void abort_lock_replay_queue(struct obd_device *obd)
1734 {
1735         struct ptlrpc_request *req, *n;
1736         LIST_HEAD(abort_list);
1737
1738         spin_lock(&obd->obd_recovery_task_lock);
1739         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1740         spin_unlock(&obd->obd_recovery_task_lock);
1741         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1742                 DEBUG_REQ(D_ERROR, req, "aborted:");
1743                 req->rq_status = -ENOTCONN;
1744                 if (ptlrpc_error(req)) {
1745                         DEBUG_REQ(D_ERROR, req,
1746                                   "failed abort_lock_reply; skipping");
1747                 }
1748                 target_request_copy_put(req);
1749         }
1750 }
1751
1752 /*
1753  * Called from a cleanup function if the device is being cleaned up
1754  * forcefully.  The exports should all have been disconnected already,
1755  * the only thing left to do is
1756  * - clear the recovery flags
1757  * - cancel the timer
1758  * - free queued requests and replies, but don't send replies
1759  * Because the obd_stopping flag is set, no new requests should be received.
1760  */
1761 void target_cleanup_recovery(struct obd_device *obd)
1762 {
1763         struct ptlrpc_request *req, *n;
1764         LIST_HEAD(clean_list);
1765
1766         spin_lock(&obd->obd_dev_lock);
1767         if (!obd->obd_recovering) {
1768                 spin_unlock(&obd->obd_dev_lock);
1769                 EXIT;
1770                 return;
1771         }
1772         obd->obd_recovering = obd->obd_abort_recovery = 0;
1773         spin_unlock(&obd->obd_dev_lock);
1774
1775         spin_lock(&obd->obd_recovery_task_lock);
1776         target_cancel_recovery_timer(obd);
1777         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1778         spin_unlock(&obd->obd_recovery_task_lock);
1779
1780         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1781                 LASSERT(req->rq_reply_state == NULL);
1782                 target_exp_dequeue_req_replay(req);
1783                 target_request_copy_put(req);
1784         }
1785
1786         spin_lock(&obd->obd_recovery_task_lock);
1787         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1788         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1789         spin_unlock(&obd->obd_recovery_task_lock);
1790
1791         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1792                 LASSERT(req->rq_reply_state == NULL);
1793                 target_request_copy_put(req);
1794         }
1795
1796         EXIT;
1797 }
1798 EXPORT_SYMBOL(target_cleanup_recovery);
1799
1800 /* obd_recovery_task_lock should be held */
1801 void target_cancel_recovery_timer(struct obd_device *obd)
1802 {
1803         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1804         hrtimer_cancel(&obd->obd_recovery_timer);
1805 }
1806
1807 static void target_start_recovery_timer(struct obd_device *obd)
1808 {
1809         ktime_t delay;
1810
1811         if (obd->obd_recovery_start != 0)
1812                 return;
1813
1814         spin_lock(&obd->obd_dev_lock);
1815         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1816                 spin_unlock(&obd->obd_dev_lock);
1817                 return;
1818         }
1819
1820         LASSERT(obd->obd_recovery_timeout != 0);
1821
1822         if (obd->obd_recovery_start != 0) {
1823                 spin_unlock(&obd->obd_dev_lock);
1824                 return;
1825         }
1826
1827         obd->obd_recovery_start = ktime_get_seconds();
1828         delay = ktime_set(obd->obd_recovery_start +
1829                           obd->obd_recovery_timeout, 0);
1830         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1831         spin_unlock(&obd->obd_dev_lock);
1832
1833         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1834                       obd->obd_name,
1835                       obd->obd_recovery_timeout / 60,
1836                       obd->obd_recovery_timeout % 60,
1837                       atomic_read(&obd->obd_max_recoverable_clients),
1838                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1839                       "" : "s",
1840                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1841                       "s" : "");
1842 }
1843
1844 /**
1845  * extend recovery window.
1846  *
1847  * if @extend is true, extend recovery window to have @dr_timeout remaining
1848  * at least; otherwise, make sure the recovery timeout value is not less
1849  * than @dr_timeout.
1850  */
1851 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1852                                   bool extend)
1853 {
1854         ktime_t left_ns;
1855         timeout_t timeout;
1856         timeout_t left;
1857
1858         spin_lock(&obd->obd_dev_lock);
1859         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1860             obd->obd_stopping) {
1861                 spin_unlock(&obd->obd_dev_lock);
1862                 return;
1863         }
1864         LASSERT(obd->obd_recovery_start != 0);
1865
1866         left_ns = hrtimer_expires_remaining(&obd->obd_recovery_timer);
1867         left = ktime_divns(left_ns, NSEC_PER_SEC);
1868
1869         if (extend) {
1870                 timeout = obd->obd_recovery_timeout;
1871                 /* dr_timeout will happen after the hrtimer has expired.
1872                  * Add the excess time to the soft recovery timeout without
1873                  * exceeding the hard recovery timeout.
1874                  */
1875                 if (dr_timeout > left) {
1876                         timeout += dr_timeout - left;
1877                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1878                                         timeout);
1879                 }
1880         } else {
1881                 timeout = clamp_t(timeout_t, dr_timeout,
1882                                   obd->obd_recovery_timeout,
1883                                   obd->obd_recovery_time_hard);
1884         }
1885
1886         if (timeout == obd->obd_recovery_time_hard)
1887                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1888                       obd->obd_name, timeout, extend);
1889
1890         if (obd->obd_recovery_timeout < timeout) {
1891                 ktime_t end, now;
1892
1893                 obd->obd_recovery_timeout = timeout;
1894                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
1895                 now = ktime_set(ktime_get_seconds(), 0);
1896                 left_ns = ktime_sub(end, now);
1897                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
1898                 left = ktime_divns(left_ns, NSEC_PER_SEC);
1899         }
1900         spin_unlock(&obd->obd_dev_lock);
1901
1902         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
1903                 obd->obd_name, left);
1904 }
1905
1906 /* Reset the timer with each new client connection */
1907 /*
1908  * This timer is actually reconnect_timer, which is for making sure
1909  * the total recovery window is at least as big as my reconnect
1910  * attempt timing. So the initial recovery time_out will be set to
1911  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1912  * from client is bigger than this, then the recovery time_out will
1913  * be extended to make sure the client could be reconnected, in the
1914  * process, the timeout from the new client should be ignored.
1915  */
1916 static void
1917 check_and_start_recovery_timer(struct obd_device *obd,
1918                                struct ptlrpc_request *req,
1919                                int new_client)
1920 {
1921         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
1922         struct obd_device_target *obt = &obd->u.obt;
1923
1924         if (!new_client && service_timeout)
1925                 /*
1926                  * Teach server about old server's estimates, as first guess
1927                  * at how long new requests will take.
1928                  */
1929                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1930                             service_timeout);
1931
1932         target_start_recovery_timer(obd);
1933
1934         /*
1935          * Convert the service time to RPC timeout,
1936          * and reuse service_timeout to limit stack usage.
1937          */
1938         service_timeout = at_est2timeout(service_timeout);
1939
1940         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
1941             service_timeout < at_extra)
1942                 service_timeout = at_extra;
1943
1944         /*
1945          * We expect other clients to timeout within service_timeout, then try
1946          * to reconnect, then try the failover server.  The max delay between
1947          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
1948          */
1949         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
1950
1951         LASSERT(obt->obt_magic == OBT_MAGIC);
1952         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1953         if (service_timeout > obd->obd_recovery_timeout && !new_client)
1954                 extend_recovery_timer(obd, service_timeout, false);
1955 }
1956
1957 /** Health checking routines */
1958 static inline int exp_connect_healthy(struct obd_export *exp)
1959 {
1960         return exp->exp_in_recovery;
1961 }
1962
1963 /** if export done req_replay or has replay in queue */
1964 static inline int exp_req_replay_healthy(struct obd_export *exp)
1965 {
1966         return (!exp->exp_req_replay_needed ||
1967                 atomic_read(&exp->exp_replay_count) > 0);
1968 }
1969
1970
1971 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1972 {
1973         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1974                exp_req_replay_healthy(exp);
1975 }
1976
1977 /** if export done lock_replay or has replay in queue */
1978 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1979 {
1980         return (!exp->exp_lock_replay_needed ||
1981                 atomic_read(&exp->exp_replay_count) > 0);
1982 }
1983
1984 static inline int exp_vbr_healthy(struct obd_export *exp)
1985 {
1986         return !exp->exp_vbr_failed;
1987 }
1988
1989 static inline int exp_finished(struct obd_export *exp)
1990 {
1991         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
1992 }
1993
1994 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1995 {
1996         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1997                 exp_finished(exp);
1998 }
1999
2000 static int check_for_next_transno(struct lu_target *lut)
2001 {
2002         struct ptlrpc_request *req = NULL;
2003         struct obd_device *obd = lut->lut_obd;
2004         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2005         int wake_up = 0, connected, completed, queue_len;
2006         __u64 req_transno = 0;
2007         __u64 update_transno = 0;
2008         __u64 next_transno = 0;
2009
2010         ENTRY;
2011
2012         spin_lock(&obd->obd_recovery_task_lock);
2013         if (!list_empty(&obd->obd_req_replay_queue)) {
2014                 req = list_entry(obd->obd_req_replay_queue.next,
2015                                      struct ptlrpc_request, rq_list);
2016                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2017         }
2018
2019         if (tdtd != NULL)
2020                 update_transno = distribute_txn_get_next_transno(tdtd);
2021
2022         connected = atomic_read(&obd->obd_connected_clients);
2023         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2024         queue_len = obd->obd_requests_queued_for_recovery;
2025         next_transno = obd->obd_next_recovery_transno;
2026
2027         CDEBUG(D_HA,
2028                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2029                atomic_read(&obd->obd_max_recoverable_clients),
2030                connected, completed,
2031                queue_len, req_transno, next_transno);
2032
2033         if (obd->obd_abort_recovery) {
2034                 CDEBUG(D_HA, "waking for aborted recovery\n");
2035                 wake_up = 1;
2036         } else if (obd->obd_recovery_expired) {
2037                 CDEBUG(D_HA, "waking for expired recovery\n");
2038                 wake_up = 1;
2039         } else if (tdtd != NULL && req != NULL &&
2040                    is_req_replayed_by_update(req)) {
2041                 LASSERTF(req_transno < next_transno,
2042                          "req_transno %llu next_transno%llu\n", req_transno,
2043                          next_transno);
2044                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2045                        req_transno);
2046                 wake_up = 1;
2047         } else if (req_transno == next_transno ||
2048                    (update_transno != 0 && update_transno <= next_transno)) {
2049                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2050                 wake_up = 1;
2051         } else if (queue_len > 0 &&
2052                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2053                 /** handle gaps occured due to lost reply or VBR */
2054                 LASSERTF(req_transno >= next_transno,
2055                          "req_transno: %llu, next_transno: %llu\n",
2056                          req_transno, next_transno);
2057                 CDEBUG(D_HA,
2058                        "%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",
2059                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2060                        next_transno, queue_len, completed, connected,
2061                        req_transno, update_transno, obd->obd_last_committed);
2062                 obd->obd_next_recovery_transno = req_transno;
2063                 wake_up = 1;
2064         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2065                 CDEBUG(D_HA, "waking for completed recovery\n");
2066                 wake_up = 1;
2067         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2068                 CDEBUG(D_HA,
2069                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2070                        next_transno);
2071                 obd->obd_next_recovery_transno = req_transno;
2072                 wake_up = 1;
2073         }
2074         spin_unlock(&obd->obd_recovery_task_lock);
2075         return wake_up;
2076 }
2077
2078 static int check_for_next_lock(struct lu_target *lut)
2079 {
2080         struct obd_device *obd = lut->lut_obd;
2081         int wake_up = 0;
2082
2083         spin_lock(&obd->obd_recovery_task_lock);
2084         if (!list_empty(&obd->obd_lock_replay_queue)) {
2085                 CDEBUG(D_HA, "waking for next lock\n");
2086                 wake_up = 1;
2087         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2088                 CDEBUG(D_HA, "waking for completed lock replay\n");
2089                 wake_up = 1;
2090         } else if (obd->obd_abort_recovery) {
2091                 CDEBUG(D_HA, "waking for aborted recovery\n");
2092                 wake_up = 1;
2093         } else if (obd->obd_recovery_expired) {
2094                 CDEBUG(D_HA, "waking for expired recovery\n");
2095                 wake_up = 1;
2096         }
2097         spin_unlock(&obd->obd_recovery_task_lock);
2098
2099         return wake_up;
2100 }
2101
2102 static int check_update_llog(struct lu_target *lut)
2103 {
2104         struct obd_device *obd = lut->lut_obd;
2105         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2106
2107         if (obd->obd_abort_recovery) {
2108                 CDEBUG(D_HA, "waking for aborted recovery\n");
2109                 return 1;
2110         }
2111
2112         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2113                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2114                 return 1;
2115         }
2116
2117         return 0;
2118 }
2119
2120 /**
2121  * wait for recovery events,
2122  * check its status with help of check_routine
2123  * evict dead clients via health_check
2124  */
2125 static int target_recovery_overseer(struct lu_target *lut,
2126                                     int (*check_routine)(struct lu_target *),
2127                                     int (*health_check)(struct obd_export *))
2128 {
2129         struct obd_device *obd = lut->lut_obd;
2130         struct target_distribute_txn_data *tdtd;
2131         time64_t last = 0;
2132         time64_t now;
2133 repeat:
2134         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2135                 now = ktime_get_seconds();
2136                 if (now - last > 600) {
2137                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2138                                       obd->obd_name);
2139                         last = now;
2140                 }
2141         }
2142         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2143               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2144                 __u64 next_update_transno = 0;
2145
2146                 /*
2147                  * Only abort the recovery if there are no update recovery
2148                  * left in the queue
2149                  */
2150                 spin_lock(&obd->obd_recovery_task_lock);
2151                 if (lut->lut_tdtd != NULL) {
2152                         next_update_transno =
2153                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2154
2155                         tdtd = lut->lut_tdtd;
2156                         /*
2157                          * If next_update_transno == 0, it probably because
2158                          * updatelog retrieve threads did not get any records
2159                          * yet, let's wait those threads stopped
2160                          */
2161                         if (next_update_transno == 0) {
2162                                 spin_unlock(&obd->obd_recovery_task_lock);
2163
2164                                 while (wait_event_timeout(
2165                                         tdtd->tdtd_recovery_threads_waitq,
2166                                         check_update_llog(lut),
2167                                         cfs_time_seconds(60)) == 0);
2168
2169                                 spin_lock(&obd->obd_recovery_task_lock);
2170                                 next_update_transno =
2171                                         distribute_txn_get_next_transno(tdtd);
2172                         }
2173                 }
2174
2175                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2176                         obd->obd_next_recovery_transno = next_update_transno;
2177                         spin_unlock(&obd->obd_recovery_task_lock);
2178                         /*
2179                          * Disconnect unfinished exports from clients, and
2180                          * keep connection from MDT to make sure the update
2181                          * recovery will still keep trying until some one
2182                          * manually abort the recovery
2183                          */
2184                         class_disconnect_stale_exports(obd,
2185                                                 exp_finished_or_from_mdt);
2186                         /* Abort all of replay & replay lock req from clients */
2187                         abort_req_replay_queue(obd);
2188                         abort_lock_replay_queue(obd);
2189                         CDEBUG(D_HA,
2190                                "%s: there are still update replay (%#llx)in the queue.\n",
2191                                obd->obd_name, next_update_transno);
2192                 } else {
2193                         obd->obd_abort_recovery = 1;
2194                         spin_unlock(&obd->obd_recovery_task_lock);
2195                         CWARN("%s recovery is aborted by hard timeout\n",
2196                               obd->obd_name);
2197                 }
2198         }
2199
2200         while (wait_event_timeout(obd->obd_next_transno_waitq,
2201                                   check_routine(lut),
2202                                   cfs_time_seconds(60)) == 0)
2203                 ; /* wait indefinitely for event, but don't trigger watchdog */
2204
2205         if (obd->obd_abort_recovery) {
2206                 CWARN("recovery is aborted, evict exports in recovery\n");
2207                 if (lut->lut_tdtd != NULL) {
2208                         tdtd = lut->lut_tdtd;
2209                         /*
2210                          * Let's wait all of the update log recovery thread
2211                          * finished
2212                          */
2213                         wait_event_idle(
2214                                 tdtd->tdtd_recovery_threads_waitq,
2215                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2216                                 == 0);
2217                         /* Then abort the update recovery list */
2218                         dtrq_list_destroy(lut->lut_tdtd);
2219                 }
2220
2221                 /** evict exports which didn't finish recovery yet */
2222                 class_disconnect_stale_exports(obd, exp_finished);
2223                 return 1;
2224         } else if (obd->obd_recovery_expired &&
2225                    obd->obd_recovery_timeout < obd->obd_recovery_time_hard) {
2226                 obd->obd_recovery_expired = 0;
2227
2228                 /** If some clients died being recovered, evict them */
2229                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2230                               obd->obd_name);
2231                 /** evict cexports with no replay in queue, they are stalled */
2232                 class_disconnect_stale_exports(obd, health_check);
2233
2234                 /** continue with VBR */
2235                 spin_lock(&obd->obd_dev_lock);
2236                 obd->obd_version_recov = 1;
2237                 spin_unlock(&obd->obd_dev_lock);
2238                 /**
2239                  * reset timer, recovery will proceed with versions now,
2240                  * timeout is set just to handle reconnection delays
2241                  */
2242                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2243                 /**
2244                  * Wait for recovery events again, after evicting bad clients
2245                  */
2246                 goto repeat;
2247         }
2248         return 0;
2249 }
2250
2251 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2252 {
2253         struct obd_device *obd = lut->lut_obd;
2254         struct ptlrpc_request *req = NULL;
2255
2256         CDEBUG(D_HA, "Waiting for lock\n");
2257         if (target_recovery_overseer(lut, check_for_next_lock,
2258                                      exp_lock_replay_healthy))
2259                 abort_lock_replay_queue(obd);
2260
2261         spin_lock(&obd->obd_recovery_task_lock);
2262         if (!list_empty(&obd->obd_lock_replay_queue)) {
2263                 req = list_entry(obd->obd_lock_replay_queue.next,
2264                                      struct ptlrpc_request, rq_list);
2265                 list_del_init(&req->rq_list);
2266                 spin_unlock(&obd->obd_recovery_task_lock);
2267         } else {
2268                 spin_unlock(&obd->obd_recovery_task_lock);
2269                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2270                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2271                 /** evict exports failed VBR */
2272                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2273         }
2274         return req;
2275 }
2276
2277 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2278 {
2279         struct ptlrpc_request *req = NULL;
2280
2281         spin_lock(&obd->obd_recovery_task_lock);
2282         if (!list_empty(&obd->obd_final_req_queue)) {
2283                 req = list_entry(obd->obd_final_req_queue.next,
2284                                      struct ptlrpc_request, rq_list);
2285                 list_del_init(&req->rq_list);
2286                 spin_unlock(&obd->obd_recovery_task_lock);
2287                 if (req->rq_export->exp_in_recovery) {
2288                         spin_lock(&req->rq_export->exp_lock);
2289                         req->rq_export->exp_in_recovery = 0;
2290                         spin_unlock(&req->rq_export->exp_lock);
2291                 }
2292         } else {
2293                 spin_unlock(&obd->obd_recovery_task_lock);
2294         }
2295         return req;
2296 }
2297
2298 static void handle_recovery_req(struct ptlrpc_thread *thread,
2299                                 struct ptlrpc_request *req,
2300                                 svc_handler_t handler)
2301 {
2302         ENTRY;
2303
2304         /**
2305          * export can be evicted during recovery, no need to handle replays for
2306          * it after that, discard such request silently
2307          */
2308         if (req->rq_export->exp_disconnected)
2309                 RETURN_EXIT;
2310
2311         req->rq_session.lc_thread = thread;
2312         req->rq_svc_thread = thread;
2313         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2314
2315         /* thread context */
2316         lu_context_enter(&thread->t_env->le_ctx);
2317         (void)handler(req);
2318         lu_context_exit(&thread->t_env->le_ctx);
2319
2320         req->rq_svc_thread->t_env->le_ses = NULL;
2321
2322         /* don't reset timer for final stage */
2323         if (!exp_finished(req->rq_export)) {
2324                 timeout_t timeout = obd_timeout;
2325
2326                 /**
2327                  * Add request @timeout to the recovery time so next request from
2328                  * this client may come in recovery time
2329                  */
2330                 if (!AT_OFF) {
2331                         struct ptlrpc_service_part *svcpt;
2332                         timeout_t est_timeout;
2333
2334                         svcpt = req->rq_rqbd->rqbd_svcpt;
2335                         /*
2336                          * If the server sent early reply for this request,
2337                          * the client will recalculate the timeout according to
2338                          * current server estimate service time, so we will
2339                          * use the maxium timeout here for waiting the client
2340                          * sending the next req
2341                          */
2342                         est_timeout = at_get(&svcpt->scp_at_estimate);
2343                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2344                                         lustre_msg_get_timeout(req->rq_reqmsg));
2345                         /*
2346                          * Add 2 net_latency, one for balance rq_deadline
2347                          * (see ptl_send_rpc), one for resend the req to server,
2348                          * Note: client will pack net_latency in replay req
2349                          * (see ptlrpc_replay_req)
2350                          */
2351                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2352                 }
2353                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2354                                       true);
2355         }
2356         EXIT;
2357 }
2358
2359 /** Checking routines for recovery */
2360 static int check_for_recovery_ready(struct lu_target *lut)
2361 {
2362         struct obd_device *obd = lut->lut_obd;
2363         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2364
2365         CDEBUG(D_HA,
2366                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2367                clnts, obd->obd_stale_clients,
2368                atomic_read(&obd->obd_max_recoverable_clients),
2369                obd->obd_abort_recovery, obd->obd_recovery_expired);
2370
2371         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2372                 LASSERT(clnts <=
2373                         atomic_read(&obd->obd_max_recoverable_clients));
2374                 if (clnts + obd->obd_stale_clients <
2375                     atomic_read(&obd->obd_max_recoverable_clients))
2376                         return 0;
2377         }
2378
2379         if (lut->lut_tdtd != NULL) {
2380                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2381                     !obd->obd_abort_recovery && !obd->obd_stopping) {
2382                         /*
2383                          * Let's extend recovery timer, in case the recovery
2384                          * timer expired, and some clients got evicted
2385                          */
2386                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2387                                               true);
2388                         CDEBUG(D_HA,
2389                                "%s update recovery is not ready, extend recovery %d\n",
2390                                obd->obd_name, obd->obd_recovery_timeout);
2391                         return 0;
2392                 }
2393         }
2394
2395         return 1;
2396 }
2397
2398 enum {
2399         REQUEST_RECOVERY = 1,
2400         UPDATE_RECOVERY = 2,
2401 };
2402
2403 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2404 {
2405         __u64 transno = 0;
2406
2407         if (!list_empty(&obd->obd_req_replay_queue)) {
2408                 struct ptlrpc_request *req;
2409
2410                 req = list_entry(obd->obd_req_replay_queue.next,
2411                                  struct ptlrpc_request, rq_list);
2412                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2413         }
2414
2415         return transno;
2416 }
2417
2418 static __u64 get_next_transno(struct lu_target *lut, int *type)
2419 {
2420         struct obd_device *obd = lut->lut_obd;
2421         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2422         __u64 transno = 0;
2423         __u64 update_transno;
2424
2425         ENTRY;
2426
2427         transno = get_next_replay_req_transno(obd);
2428         if (type != NULL)
2429                 *type = REQUEST_RECOVERY;
2430
2431         if (tdtd == NULL)
2432                 RETURN(transno);
2433
2434         update_transno = distribute_txn_get_next_transno(tdtd);
2435         if (transno == 0 || (transno >= update_transno &&
2436                              update_transno != 0)) {
2437                 transno = update_transno;
2438                 if (type != NULL)
2439                         *type = UPDATE_RECOVERY;
2440         }
2441
2442         RETURN(transno);
2443 }
2444
2445 /**
2446  * drop duplicate replay request
2447  *
2448  * Because the operation has been replayed by update recovery, the request
2449  * with the same transno will be dropped and also notify the client to send
2450  * next replay request.
2451  *
2452  * \param[in] env       execution environment
2453  * \param[in] obd       failover obd device
2454  * \param[in] req       request to be dropped
2455  */
2456 static void drop_duplicate_replay_req(struct lu_env *env,
2457                                       struct obd_device *obd,
2458                                       struct ptlrpc_request *req)
2459 {
2460         DEBUG_REQ(D_HA, req,
2461                   "remove t%lld from %s because duplicate update records found",
2462                   lustre_msg_get_transno(req->rq_reqmsg),
2463                   libcfs_nid2str(req->rq_peer.nid));
2464
2465         /*
2466          * Right now, only for MDS reint operation update replay and
2467          * normal request replay can have the same transno
2468          */
2469         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2470                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2471                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2472                 if (likely(req->rq_export))
2473                         target_committed_to_req(req);
2474                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2475                 target_send_reply(req, req->rq_status, 0);
2476         } else {
2477                 DEBUG_REQ(D_ERROR, req, "wrong opc from %s",
2478                 libcfs_nid2str(req->rq_peer.nid));
2479         }
2480         target_exp_dequeue_req_replay(req);
2481         target_request_copy_put(req);
2482         obd->obd_replayed_requests++;
2483 }
2484
2485 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2486
2487 static void replay_request_or_update(struct lu_env *env,
2488                                      struct lu_target *lut,
2489                                      struct target_recovery_data *trd,
2490                                      struct ptlrpc_thread *thread)
2491 {
2492         struct obd_device *obd = lut->lut_obd;
2493         struct ptlrpc_request *req = NULL;
2494         int type;
2495         __u64 transno;
2496
2497         ENTRY;
2498
2499         CDEBUG(D_HA, "Waiting for transno %lld\n",
2500                obd->obd_next_recovery_transno);
2501
2502         /* Replay all of request and update by transno */
2503         do {
2504                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2505
2506                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2507
2508                 /**
2509                  * It is needed to extend recovery window above
2510                  *  recovery_time_soft. Extending is possible only in the
2511                  *  end of recovery window (see more details in
2512                  *  handle_recovery_req()).
2513                  */
2514                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2515
2516                 if (target_recovery_overseer(lut, check_for_next_transno,
2517                                         exp_req_replay_healthy_or_from_mdt)) {
2518                         abort_req_replay_queue(obd);
2519                         abort_lock_replay_queue(obd);
2520                         goto abort;
2521                 }
2522
2523                 spin_lock(&obd->obd_recovery_task_lock);
2524                 transno = get_next_transno(lut, &type);
2525                 if (type == REQUEST_RECOVERY && transno != 0) {
2526                         /*
2527                          * Drop replay request from client side, if the
2528                          * replay has been executed by update with the
2529                          * same transno
2530                          */
2531                         req = list_entry(obd->obd_req_replay_queue.next,
2532                                         struct ptlrpc_request, rq_list);
2533
2534                         list_del_init(&req->rq_list);
2535                         obd->obd_requests_queued_for_recovery--;
2536                         spin_unlock(&obd->obd_recovery_task_lock);
2537
2538                         /*
2539                          * Let's check if the request has been redone by
2540                          * update replay
2541                          */
2542                         if (is_req_replayed_by_update(req)) {
2543                                 struct distribute_txn_replay_req *dtrq;
2544
2545                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2546                                                                       transno);
2547                                 LASSERT(dtrq != NULL);
2548                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2549                                 list_del_init(&dtrq->dtrq_list);
2550                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2551                                 dtrq_destroy(dtrq);
2552
2553                                 drop_duplicate_replay_req(env, obd, req);
2554
2555                                 continue;
2556                         }
2557
2558                         LASSERT(trd->trd_processing_task == current->pid);
2559                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2560                                   req->rq_xid,
2561                                   lustre_msg_get_transno(req->rq_reqmsg),
2562                                   libcfs_nid2str(req->rq_peer.nid));
2563
2564                         ptlrpc_watchdog_init(&thread->t_watchdog,
2565                                              WATCHDOG_TIMEOUT);
2566                         handle_recovery_req(thread, req,
2567                                             trd->trd_recovery_handler);
2568                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2569
2570                         /**
2571                          * bz18031: increase next_recovery_transno before
2572                          * target_request_copy_put() will drop exp_rpc reference
2573                          */
2574                         spin_lock(&obd->obd_recovery_task_lock);
2575                         obd->obd_next_recovery_transno++;
2576                         spin_unlock(&obd->obd_recovery_task_lock);
2577                         target_exp_dequeue_req_replay(req);
2578                         target_request_copy_put(req);
2579                         obd->obd_replayed_requests++;
2580                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2581                         struct distribute_txn_replay_req *dtrq;
2582                         int rc;
2583
2584                         spin_unlock(&obd->obd_recovery_task_lock);
2585
2586                         LASSERT(tdtd != NULL);
2587                         dtrq = distribute_txn_get_next_req(tdtd);
2588                         lu_context_enter(&thread->t_env->le_ctx);
2589                         ptlrpc_watchdog_init(&thread->t_watchdog,
2590                                              WATCHDOG_TIMEOUT);
2591                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2592                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2593                         lu_context_exit(&thread->t_env->le_ctx);
2594                         extend_recovery_timer(obd, obd_timeout, true);
2595
2596                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2597                                 CDEBUG(D_HA,
2598                                        "Move x%llu t%llu to finish list\n",
2599                                        dtrq->dtrq_xid,
2600                                        dtrq->dtrq_master_transno);
2601
2602                                 /* Add it to the replay finish list */
2603                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2604                                 list_add(&dtrq->dtrq_list,
2605                                          &tdtd->tdtd_replay_finish_list);
2606                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2607
2608                                 spin_lock(&obd->obd_recovery_task_lock);
2609                                 if (transno == obd->obd_next_recovery_transno)
2610                                         obd->obd_next_recovery_transno++;
2611                                 else if (transno >
2612                                          obd->obd_next_recovery_transno)
2613                                         obd->obd_next_recovery_transno =
2614                                                                 transno + 1;
2615                                 spin_unlock(&obd->obd_recovery_task_lock);
2616                         } else {
2617                                 dtrq_destroy(dtrq);
2618                         }
2619                 } else {
2620                         spin_unlock(&obd->obd_recovery_task_lock);
2621 abort:
2622                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2623                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2624                         /** evict exports failed VBR */
2625                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2626                         break;
2627                 }
2628         } while (1);
2629 }
2630
2631 static int target_recovery_thread(void *arg)
2632 {
2633         struct lu_target *lut = arg;
2634         struct obd_device *obd = lut->lut_obd;
2635         struct ptlrpc_request *req;
2636         struct target_recovery_data *trd = &obd->obd_recovery_data;
2637         unsigned long delta;
2638         struct lu_env *env;
2639         struct ptlrpc_thread *thread = NULL;
2640         int rc = 0;
2641
2642         ENTRY;
2643         OBD_ALLOC_PTR(thread);
2644         if (thread == NULL)
2645                 RETURN(-ENOMEM);
2646
2647         OBD_ALLOC_PTR(env);
2648         if (env == NULL)
2649                 GOTO(out_thread, rc = -ENOMEM);
2650         rc = lu_env_add(env);
2651         if (rc)
2652                 GOTO(out_env, rc);
2653
2654         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2655         if (rc)
2656                 GOTO(out_env_remove, rc);
2657
2658         thread->t_env = env;
2659         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2660         env->le_ctx.lc_thread = thread;
2661         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2662
2663         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2664                current->pid);
2665         trd->trd_processing_task = current->pid;
2666
2667         spin_lock(&obd->obd_dev_lock);
2668         obd->obd_recovering = 1;
2669         spin_unlock(&obd->obd_dev_lock);
2670         complete(&trd->trd_starting);
2671
2672         /* first of all, we have to know the first transno to replay */
2673         if (target_recovery_overseer(lut, check_for_recovery_ready,
2674                                      exp_connect_healthy)) {
2675                 abort_req_replay_queue(obd);
2676                 abort_lock_replay_queue(obd);
2677                 if (lut->lut_tdtd != NULL)
2678                         dtrq_list_destroy(lut->lut_tdtd);
2679         }
2680
2681         /* next stage: replay requests or update */
2682         delta = jiffies;
2683         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2684                atomic_read(&obd->obd_req_replay_clients),
2685                obd->obd_next_recovery_transno);
2686         replay_request_or_update(env, lut, trd, thread);
2687
2688         /**
2689          * The second stage: replay locks
2690          */
2691         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2692                atomic_read(&obd->obd_lock_replay_clients));
2693         while ((req = target_next_replay_lock(lut))) {
2694                 LASSERT(trd->trd_processing_task == current->pid);
2695                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2696                           libcfs_nid2str(req->rq_peer.nid));
2697                 handle_recovery_req(thread, req,
2698                                     trd->trd_recovery_handler);
2699                 target_request_copy_put(req);
2700                 obd->obd_replayed_locks++;
2701         }
2702
2703         /**
2704          * The third stage: reply on final pings, at this moment all clients
2705          * must have request in final queue
2706          */
2707         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2708         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2709         /** Update server last boot epoch */
2710         tgt_boot_epoch_update(lut);
2711         /*
2712          * We drop recoverying flag to forward all new requests
2713          * to regular mds_handle() since now
2714          */
2715         spin_lock(&obd->obd_dev_lock);
2716         obd->obd_recovering = obd->obd_abort_recovery = 0;
2717         spin_unlock(&obd->obd_dev_lock);
2718         spin_lock(&obd->obd_recovery_task_lock);
2719         target_cancel_recovery_timer(obd);
2720         spin_unlock(&obd->obd_recovery_task_lock);
2721         while ((req = target_next_final_ping(obd))) {
2722                 LASSERT(trd->trd_processing_task == current->pid);
2723                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2724                           libcfs_nid2str(req->rq_peer.nid));
2725                 handle_recovery_req(thread, req,
2726                                     trd->trd_recovery_handler);
2727                 /*
2728                  * Because the waiting client can not send ping to server,
2729                  * so we need refresh the last_request_time, to avoid the
2730                  * export is being evicted
2731                  */
2732                 ptlrpc_update_export_timer(req->rq_export, 0);
2733                 target_request_copy_put(req);
2734         }
2735
2736         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2737         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2738                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2739         if (delta > OBD_RECOVERY_TIME_SOFT) {
2740                 CWARN("too long recovery - read logs\n");
2741                 libcfs_debug_dumplog();
2742         }
2743
2744         target_finish_recovery(lut);
2745         lu_context_fini(&env->le_ctx);
2746         trd->trd_processing_task = 0;
2747         complete_all(&trd->trd_finishing);
2748         tgt_io_thread_done(thread);
2749 out_env_remove:
2750         lu_env_remove(env);
2751 out_env:
2752         OBD_FREE_PTR(env);
2753 out_thread:
2754         OBD_FREE_PTR(thread);
2755         RETURN(rc);
2756 }
2757
2758 static int target_start_recovery_thread(struct lu_target *lut,
2759                                         svc_handler_t handler)
2760 {
2761         struct obd_device *obd = lut->lut_obd;
2762         int rc = 0;
2763         struct target_recovery_data *trd = &obd->obd_recovery_data;
2764         int index;
2765
2766         memset(trd, 0, sizeof(*trd));
2767         init_completion(&trd->trd_starting);
2768         init_completion(&trd->trd_finishing);
2769         trd->trd_recovery_handler = handler;
2770
2771         rc = server_name2index(obd->obd_name, &index, NULL);
2772         if (rc < 0)
2773                 return rc;
2774
2775         if (!IS_ERR(kthread_run(target_recovery_thread,
2776                                 lut, "tgt_recover_%d", index))) {
2777                 wait_for_completion(&trd->trd_starting);
2778                 LASSERT(obd->obd_recovering != 0);
2779         } else {
2780                 rc = -ECHILD;
2781         }
2782
2783         return rc;
2784 }
2785
2786 void target_stop_recovery_thread(struct obd_device *obd)
2787 {
2788         if (obd->obd_recovery_data.trd_processing_task > 0) {
2789                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2790                 /** recovery can be done but postrecovery is not yet */
2791                 spin_lock(&obd->obd_dev_lock);
2792                 if (obd->obd_recovering) {
2793                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2794                         obd->obd_abort_recovery = 1;
2795                         wake_up(&obd->obd_next_transno_waitq);
2796                 }
2797                 spin_unlock(&obd->obd_dev_lock);
2798                 wait_for_completion(&trd->trd_finishing);
2799         }
2800 }
2801 EXPORT_SYMBOL(target_stop_recovery_thread);
2802
2803 void target_recovery_fini(struct obd_device *obd)
2804 {
2805         class_disconnect_exports(obd);
2806         target_stop_recovery_thread(obd);
2807         target_cleanup_recovery(obd);
2808 }
2809 EXPORT_SYMBOL(target_recovery_fini);
2810
2811 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2812 {
2813         struct obd_device *obd = container_of(timer, struct obd_device,
2814                                               obd_recovery_timer);
2815
2816         CDEBUG(D_HA,
2817                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2818                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2819                ktime_get_seconds() - obd->obd_recovery_start,
2820                atomic_read(&obd->obd_connected_clients));
2821
2822         obd->obd_recovery_expired = 1;
2823         wake_up(&obd->obd_next_transno_waitq);
2824         return HRTIMER_NORESTART;
2825 }
2826
2827 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2828 {
2829         struct obd_device *obd = lut->lut_obd;
2830
2831         if (lut->lut_bottom->dd_rdonly)
2832                 return;
2833
2834         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2835                 /** Update server last boot epoch */
2836                 tgt_boot_epoch_update(lut);
2837                 return;
2838         }
2839
2840         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2841                "last_transno %llu\n", obd->obd_name,
2842                atomic_read(&obd->obd_max_recoverable_clients),
2843                obd->obd_last_committed);
2844         LASSERT(obd->obd_stopping == 0);
2845         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2846         obd->obd_recovery_start = 0;
2847         obd->obd_recovery_end = 0;
2848
2849         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2850                      HRTIMER_MODE_ABS);
2851         obd->obd_recovery_timer.function = &target_recovery_expired;
2852         target_start_recovery_thread(lut, handler);
2853 }
2854 EXPORT_SYMBOL(target_recovery_init);
2855
2856 static int target_process_req_flags(struct obd_device *obd,
2857                                     struct ptlrpc_request *req)
2858 {
2859         struct obd_export *exp = req->rq_export;
2860
2861         LASSERT(exp != NULL);
2862         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2863                 /* client declares he's ready to replay locks */
2864                 spin_lock(&exp->exp_lock);
2865                 if (exp->exp_req_replay_needed) {
2866                         exp->exp_req_replay_needed = 0;
2867                         spin_unlock(&exp->exp_lock);
2868
2869                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2870                         atomic_dec(&obd->obd_req_replay_clients);
2871                 } else {
2872                         spin_unlock(&exp->exp_lock);
2873                 }
2874         }
2875         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2876                 /*
2877                  * client declares he's ready to complete recovery
2878                  * so, we put the request on th final queue
2879                  */
2880                 spin_lock(&exp->exp_lock);
2881                 if (exp->exp_lock_replay_needed) {
2882                         exp->exp_lock_replay_needed = 0;
2883                         spin_unlock(&exp->exp_lock);
2884
2885                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2886                         atomic_dec(&obd->obd_lock_replay_clients);
2887                 } else {
2888                         spin_unlock(&exp->exp_lock);
2889                 }
2890         }
2891         return 0;
2892 }
2893
2894 int target_queue_recovery_request(struct ptlrpc_request *req,
2895                                   struct obd_device *obd)
2896 {
2897         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2898         struct ptlrpc_request *reqiter;
2899         int inserted = 0;
2900
2901         ENTRY;
2902
2903         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
2904                 /* Processing the queue right now, don't re-add. */
2905                 RETURN(1);
2906         }
2907
2908         target_process_req_flags(obd, req);
2909
2910         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2911                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
2912                         if (cfs_fail_val == 1) {
2913                                 cfs_race_state = 1;
2914                                 cfs_fail_val = 0;
2915                                 wake_up(&cfs_race_waitq);
2916
2917                                 schedule_timeout_interruptible(
2918                                         cfs_time_seconds(1));
2919                         }
2920                 }
2921
2922                 /*
2923                  * client declares he's ready to complete recovery
2924                  * so, we put the request on th final queue
2925                  */
2926                 target_request_copy_get(req);
2927                 DEBUG_REQ(D_HA, req, "queue final req");
2928                 wake_up(&obd->obd_next_transno_waitq);
2929                 spin_lock(&obd->obd_recovery_task_lock);
2930                 if (obd->obd_recovering) {
2931                         struct ptlrpc_request *tmp;
2932                         struct ptlrpc_request *duplicate = NULL;
2933
2934                         if (likely(!req->rq_export->exp_replay_done)) {
2935                                 req->rq_export->exp_replay_done = 1;
2936                                 list_add_tail(&req->rq_list,
2937                                               &obd->obd_final_req_queue);
2938                                 spin_unlock(&obd->obd_recovery_task_lock);
2939                                 RETURN(0);
2940                         }
2941
2942                         /*
2943                          * XXX O(n), but only happens if final ping is
2944                          * timed out, probably reorganize the list as
2945                          * a hash list later
2946                          */
2947                         list_for_each_entry_safe(reqiter, tmp,
2948                                                  &obd->obd_final_req_queue,
2949                                                  rq_list) {
2950                                 if (reqiter->rq_export == req->rq_export) {
2951                                         list_del_init(&reqiter->rq_list);
2952                                         duplicate = reqiter;
2953                                         break;
2954                                 }
2955                         }
2956
2957                         list_add_tail(&req->rq_list,
2958                                       &obd->obd_final_req_queue);
2959                         req->rq_export->exp_replay_done = 1;
2960                         spin_unlock(&obd->obd_recovery_task_lock);
2961
2962                         if (duplicate != NULL) {
2963                                 DEBUG_REQ(D_HA, duplicate,
2964                                           "put prev final req");
2965                                 target_request_copy_put(duplicate);
2966                         }
2967                         RETURN(0);
2968                 } else {
2969                         spin_unlock(&obd->obd_recovery_task_lock);
2970                         target_request_copy_put(req);
2971                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2972                 }
2973         }
2974         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2975                 /* client declares he's ready to replay locks */
2976                 target_request_copy_get(req);
2977                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2978                 wake_up(&obd->obd_next_transno_waitq);
2979                 spin_lock(&obd->obd_recovery_task_lock);
2980                 LASSERT(obd->obd_recovering);
2981                 /* usually due to recovery abort */
2982                 if (!req->rq_export->exp_in_recovery) {
2983                         spin_unlock(&obd->obd_recovery_task_lock);
2984                         target_request_copy_put(req);
2985                         RETURN(-ENOTCONN);
2986                 }
2987                 LASSERT(req->rq_export->exp_lock_replay_needed);
2988                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2989                 spin_unlock(&obd->obd_recovery_task_lock);
2990                 RETURN(0);
2991         }
2992
2993         /*
2994          * CAVEAT EMPTOR: The incoming request message has been swabbed
2995          * (i.e. buflens etc are in my own byte order), but type-dependent
2996          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
2997          */
2998
2999         if (!transno) {
3000                 INIT_LIST_HEAD(&req->rq_list);
3001                 DEBUG_REQ(D_HA, req, "not queueing");
3002                 RETURN(1);
3003         }
3004
3005         /*
3006          * If we're processing the queue, we want don't want to queue this
3007          * message.
3008          *
3009          * Also, if this request has a transno less than the one we're waiting
3010          * for, we should process it now.  It could (and currently always will)
3011          * be an open request for a descriptor that was opened some time ago.
3012          *
3013          * Also, a resent, replayed request that has already been
3014          * handled will pass through here and be processed immediately.
3015          */
3016         CDEBUG(D_HA,
3017                "Next recovery transno: %llu, current: %llu, replaying\n",
3018                obd->obd_next_recovery_transno, transno);
3019
3020         /*
3021          * If the request has been replayed by update replay, then sends this
3022          * request to the recovery thread (replay_request_or_update()), where
3023          * it will be handled
3024          */
3025         spin_lock(&obd->obd_recovery_task_lock);
3026         if (transno < obd->obd_next_recovery_transno &&
3027             !is_req_replayed_by_update(req)) {
3028                 /* Processing the queue right now, don't re-add. */
3029                 LASSERT(list_empty(&req->rq_list));
3030                 spin_unlock(&obd->obd_recovery_task_lock);
3031                 RETURN(1);
3032         }
3033         spin_unlock(&obd->obd_recovery_task_lock);
3034
3035         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3036                 RETURN(0);
3037
3038         target_request_copy_get(req);
3039         if (!req->rq_export->exp_in_recovery) {
3040                 target_request_copy_put(req);
3041                 RETURN(-ENOTCONN);
3042         }
3043         LASSERT(req->rq_export->exp_req_replay_needed);
3044
3045         if (target_exp_enqueue_req_replay(req)) {
3046                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3047                 target_request_copy_put(req);
3048                 RETURN(0);
3049         }
3050
3051         /* XXX O(n^2) */
3052         spin_lock(&obd->obd_recovery_task_lock);
3053         LASSERT(obd->obd_recovering);
3054         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3055                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3056                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3057                         inserted = 1;
3058                         goto added;
3059                 }
3060
3061                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3062                              transno)) {
3063                         DEBUG_REQ(D_ERROR, req,
3064                                   "dropping replay: transno has been claimed by another client");
3065                         spin_unlock(&obd->obd_recovery_task_lock);
3066                         target_exp_dequeue_req_replay(req);
3067                         target_request_copy_put(req);
3068                         RETURN(0);
3069                 }
3070         }
3071 added:
3072         if (!inserted)
3073                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3074
3075         obd->obd_requests_queued_for_recovery++;
3076         spin_unlock(&obd->obd_recovery_task_lock);
3077         wake_up(&obd->obd_next_transno_waitq);
3078         RETURN(0);
3079 }
3080
3081 void target_committed_to_req(struct ptlrpc_request *req)
3082 {
3083         struct obd_export *exp = req->rq_export;
3084
3085         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3086                 lustre_msg_set_last_committed(req->rq_repmsg,
3087                                               exp->exp_last_committed);
3088         else
3089                 DEBUG_REQ(D_IOCTL, req,
3090                           "not sending last_committed update (%d/%d)",
3091                           exp->exp_obd->obd_no_transno,
3092                           req->rq_repmsg == NULL);
3093
3094         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3095                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3096 }
3097
3098 #endif /* HAVE_SERVER_SUPPORT */
3099
3100 /**
3101  * Packs current SLV and Limit into \a req.
3102  */
3103 int target_pack_pool_reply(struct ptlrpc_request *req)
3104 {
3105         struct obd_device *obd;
3106
3107         ENTRY;
3108
3109         /*
3110          * Check that we still have all structures alive as this may
3111          * be some late RPC at shutdown time.
3112          */
3113         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3114                      !exp_connect_lru_resize(req->rq_export))) {
3115                 lustre_msg_set_slv(req->rq_repmsg, 0);
3116                 lustre_msg_set_limit(req->rq_repmsg, 0);
3117                 RETURN(0);
3118         }
3119
3120         /* OBD is alive here as export is alive, which we checked above. */
3121         obd = req->rq_export->exp_obd;
3122
3123         read_lock(&obd->obd_pool_lock);
3124         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3125         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3126         read_unlock(&obd->obd_pool_lock);
3127
3128         RETURN(0);
3129 }
3130
3131 static int target_send_reply_msg(struct ptlrpc_request *req,
3132                                  int rc, int fail_id)
3133 {
3134         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3135                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3136                 return -ECOMM;
3137         }
3138         /*
3139          * We can have a null rq_reqmsg in the event of bad signature or
3140          * no context when unwrapping
3141          */
3142         if (req->rq_reqmsg &&
3143             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3144             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3145                 return -ECOMM;
3146
3147         if (unlikely(rc)) {
3148                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3149                 req->rq_status = rc;
3150                 return ptlrpc_send_error(req, 1);
3151         }
3152         DEBUG_REQ(D_NET, req, "sending reply");
3153
3154         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3155 }
3156
3157 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3158 {
3159         struct ptlrpc_service_part *svcpt;
3160         int netrc;
3161         struct ptlrpc_reply_state *rs;
3162         struct obd_export *exp;
3163
3164         ENTRY;
3165
3166         if (req->rq_no_reply) {
3167                 EXIT;
3168                 return;
3169         }
3170
3171         svcpt = req->rq_rqbd->rqbd_svcpt;
3172         rs = req->rq_reply_state;
3173         if (rs == NULL || !rs->rs_difficult) {
3174                 /* no notifiers */
3175                 target_send_reply_msg(req, rc, fail_id);
3176                 EXIT;
3177                 return;
3178         }
3179
3180         /* must be an export if locks saved */
3181         LASSERT(req->rq_export != NULL);
3182         /* req/reply consistent */
3183         LASSERT(rs->rs_svcpt == svcpt);
3184
3185         /* "fresh" reply */
3186         LASSERT(!rs->rs_scheduled);
3187         LASSERT(!rs->rs_scheduled_ever);
3188         LASSERT(!rs->rs_handled);
3189         LASSERT(!rs->rs_on_net);
3190         LASSERT(rs->rs_export == NULL);
3191         LASSERT(list_empty(&rs->rs_obd_list));
3192         LASSERT(list_empty(&rs->rs_exp_list));
3193
3194         exp = class_export_get(req->rq_export);
3195
3196         /* disable reply scheduling while I'm setting up */
3197         rs->rs_scheduled = 1;
3198         rs->rs_on_net    = 1;
3199         rs->rs_xid       = req->rq_xid;
3200         rs->rs_transno   = req->rq_transno;
3201         rs->rs_export    = exp;
3202         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3203
3204         spin_lock(&exp->exp_uncommitted_replies_lock);
3205         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3206                rs->rs_transno, exp->exp_last_committed);
3207         if (rs->rs_transno > exp->exp_last_committed) {
3208                 /* not committed already */
3209                 list_add_tail(&rs->rs_obd_list,
3210                                   &exp->exp_uncommitted_replies);
3211         }
3212         spin_unlock(&exp->exp_uncommitted_replies_lock);
3213
3214         spin_lock(&exp->exp_lock);
3215         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3216         spin_unlock(&exp->exp_lock);
3217
3218         netrc = target_send_reply_msg(req, rc, fail_id);
3219
3220         spin_lock(&svcpt->scp_rep_lock);
3221
3222         atomic_inc(&svcpt->scp_nreps_difficult);
3223
3224         if (netrc != 0) {
3225                 /*
3226                  * error sending: reply is off the net.  Also we need +1
3227                  * reply ref until ptlrpc_handle_rs() is done
3228                  * with the reply state (if the send was successful, there
3229                  * would have been +1 ref for the net, which
3230                  * reply_out_callback leaves alone)
3231                  */
3232                 rs->rs_on_net = 0;
3233                 ptlrpc_rs_addref(rs);
3234         }
3235
3236         spin_lock(&rs->rs_lock);
3237         if (rs->rs_transno <= exp->exp_last_committed ||
3238             (!rs->rs_on_net && !rs->rs_no_ack) ||
3239             list_empty(&rs->rs_exp_list) ||     /* completed already */
3240             list_empty(&rs->rs_obd_list)) {
3241                 CDEBUG(D_HA, "Schedule reply immediately\n");
3242                 ptlrpc_dispatch_difficult_reply(rs);
3243         } else {
3244                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3245                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3246         }
3247         spin_unlock(&rs->rs_lock);
3248         spin_unlock(&svcpt->scp_rep_lock);
3249         EXIT;
3250 }
3251
3252 enum ldlm_mode lck_compat_array[] = {
3253         [LCK_EX]    = LCK_COMPAT_EX,
3254         [LCK_PW]    = LCK_COMPAT_PW,
3255         [LCK_PR]    = LCK_COMPAT_PR,
3256         [LCK_CW]    = LCK_COMPAT_CW,
3257         [LCK_CR]    = LCK_COMPAT_CR,
3258         [LCK_NL]    = LCK_COMPAT_NL,
3259         [LCK_GROUP] = LCK_COMPAT_GROUP,
3260         [LCK_COS]   = LCK_COMPAT_COS,
3261 };
3262
3263 /**
3264  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3265  * not escape to the user level.
3266  */
3267 int ldlm_error2errno(enum ldlm_error error)
3268 {
3269         int result;
3270
3271         switch (error) {
3272         case ELDLM_OK:
3273         case ELDLM_LOCK_MATCHED:
3274                 result = 0;
3275                 break;
3276         case ELDLM_LOCK_CHANGED:
3277                 result = -ESTALE;
3278                 break;
3279         case ELDLM_LOCK_ABORTED:
3280                 result = -ENAVAIL;
3281                 break;
3282         case ELDLM_LOCK_REPLACED:
3283                 result = -ESRCH;
3284                 break;
3285         case ELDLM_NO_LOCK_DATA:
3286                 result = -ENOENT;
3287                 break;
3288         case ELDLM_NAMESPACE_EXISTS:
3289                 result = -EEXIST;
3290                 break;
3291         case ELDLM_BAD_NAMESPACE:
3292                 result = -EBADF;
3293                 break;
3294         default:
3295                 if (((int)error) < 0) { /* cast to signed type */
3296                         result = error; /* as ldlm_error can be unsigned */
3297                 } else {
3298                         CERROR("Invalid DLM result code: %d\n", error);
3299                         result = -EPROTO;
3300                 }
3301         }
3302         return result;
3303 }
3304 EXPORT_SYMBOL(ldlm_error2errno);
3305
3306 /**
3307  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3308  */
3309 enum ldlm_error ldlm_errno2error(int err_no)
3310 {
3311         int error;
3312
3313         switch (err_no) {
3314         case 0:
3315                 error = ELDLM_OK;
3316                 break;
3317         case -ESTALE:
3318                 error = ELDLM_LOCK_CHANGED;
3319                 break;
3320         case -ENAVAIL:
3321                 error = ELDLM_LOCK_ABORTED;
3322                 break;
3323         case -ESRCH:
3324                 error = ELDLM_LOCK_REPLACED;
3325                 break;
3326         case -ENOENT:
3327                 error = ELDLM_NO_LOCK_DATA;
3328                 break;
3329         case -EEXIST:
3330                 error = ELDLM_NAMESPACE_EXISTS;
3331                 break;
3332         case -EBADF:
3333                 error = ELDLM_BAD_NAMESPACE;
3334                 break;
3335         default:
3336                 error = err_no;
3337         }
3338         return error;
3339 }
3340
3341 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3342 void ldlm_dump_export_locks(struct obd_export *exp)
3343 {
3344         spin_lock(&exp->exp_locks_list_guard);
3345         if (!list_empty(&exp->exp_locks_list)) {
3346                 struct ldlm_lock *lock;
3347
3348                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3349                        exp);
3350                 list_for_each_entry(lock, &exp->exp_locks_list,
3351                                         l_exp_refs_link)
3352                         LDLM_ERROR(lock, "lock:");
3353         }
3354         spin_unlock(&exp->exp_locks_list_guard);
3355 }
3356 #endif
3357
3358 #ifdef HAVE_SERVER_SUPPORT
3359 static inline const char *bulk2type(struct ptlrpc_request *req)
3360 {
3361         if (req->rq_bulk_read)
3362                 return "READ";
3363         if (req->rq_bulk_write)
3364                 return "WRITE";
3365         return "UNKNOWN";
3366 }
3367
3368 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3369 {
3370         struct ptlrpc_request *req = desc->bd_req;
3371         time64_t start = ktime_get_seconds();
3372         time64_t deadline;
3373         int rc = 0;
3374
3375         ENTRY;
3376
3377         /* If there is eviction in progress, wait for it to finish. */
3378         wait_event_idle(
3379                 exp->exp_obd->obd_evict_inprogress_waitq,
3380                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3381
3382         /* Check if client was evicted or reconnected already. */
3383         if (exp->exp_failed ||
3384             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3385                 rc = -ENOTCONN;
3386         } else {
3387                 if (req->rq_bulk_read)
3388                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3389
3390                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3391                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3392                 else /* old version, bulk matchbits is rq_xid */
3393                         req->rq_mbits = req->rq_xid;
3394
3395                 if (rc == 0)
3396                         rc = ptlrpc_start_bulk_transfer(desc);
3397         }
3398
3399         if (rc < 0) {
3400                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3401                           bulk2type(req), rc);
3402                 RETURN(rc);
3403         }
3404
3405         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3406                 ptlrpc_abort_bulk(desc);
3407                 RETURN(0);
3408         }
3409
3410         /* limit actual bulk transfer to bulk_timeout seconds */
3411         deadline = start + bulk_timeout;
3412         if (deadline > req->rq_deadline)
3413                 deadline = req->rq_deadline;
3414
3415         do {
3416                 time64_t timeoutl = deadline - ktime_get_seconds();
3417                 time64_t rq_deadline;
3418
3419                 while (timeoutl >= 0 &&
3420                        wait_event_idle_timeout(
3421                                desc->bd_waitq,
3422                                !ptlrpc_server_bulk_active(desc) ||
3423                                exp->exp_failed ||
3424                                exp->exp_conn_cnt >
3425                                lustre_msg_get_conn_cnt(req->rq_reqmsg),
3426                                timeoutl ? cfs_time_seconds(1) : 1) == 0)
3427                         timeoutl -= 1;
3428                 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3429
3430                 /* Wait again if we changed rq_deadline. */
3431                 rq_deadline = READ_ONCE(req->rq_deadline);
3432                 deadline = start + bulk_timeout;
3433                 if (deadline > rq_deadline)
3434                         deadline = rq_deadline;
3435         } while (rc == -ETIMEDOUT &&
3436                  deadline > ktime_get_seconds());
3437
3438         if (rc == -ETIMEDOUT) {
3439                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3440                           bulk2type(req), deadline - start,
3441                           ktime_get_real_seconds() - deadline);
3442                 ptlrpc_abort_bulk(desc);
3443         } else if (exp->exp_failed) {
3444                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3445                           bulk2type(req));
3446                 rc = -ENOTCONN;
3447                 ptlrpc_abort_bulk(desc);
3448         } else if (exp->exp_conn_cnt >
3449                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3450                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3451                           bulk2type(req));
3452                 /* We don't reply anyway. */
3453                 rc = -ETIMEDOUT;
3454                 ptlrpc_abort_bulk(desc);
3455         } else if (desc->bd_failure) {
3456                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3457                           bulk2type(req));
3458                 /* XXX should this be a different errno? */
3459                 rc = -ETIMEDOUT;
3460         } else {
3461                 if (req->rq_bulk_write)
3462                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3463                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3464                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3465                                   bulk2type(req), desc->bd_nob_transferred,
3466                                   desc->bd_nob);
3467                         /* XXX should this be a different errno? */
3468                         rc = -ETIMEDOUT;
3469                 }
3470         }
3471
3472         RETURN(rc);
3473 }
3474 EXPORT_SYMBOL(target_bulk_io);
3475
3476 #endif /* HAVE_SERVER_SUPPORT */