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  */
  31
  32 /**
  33  * This file deals with various client/target related logic including recovery.
  34  *
  35  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  36  * should be moved.
  37  */
  38
  39 #define DEBUG_SUBSYSTEM S_LDLM
  40
  41 #include <cl_object.h>
  42 #include <linux/fs_struct.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 <uapi/linux/lustre/lustre_ioctl.h>
  53 #include "ldlm_internal.h"
  54
  55 /*
  56  * @priority: If non-zero, move the selected connection to the list head.
  57  * @create: If zero, only search in existing connections.
  58  */
  59 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  60                            int priority, int create)
  61 {
  62         struct ptlrpc_connection *ptlrpc_conn;
  63         struct obd_import_conn *imp_conn = NULL, *item;
  64         u32 refnet = imp->imp_conn_restricted_net;
  65         int rc = 0;
  66
  67         ENTRY;
  68
  69         if (!create && !priority) {
  70                 CDEBUG(D_HA, "Nothing to do\n");
  71                 RETURN(-EINVAL);
  72         }
  73
  74         /* refnet is used to restrict network connections */
  75         if (refnet != LNET_NET_ANY)
  76                 CDEBUG(D_HA, "imp %s: restrict %s to %s net\n",
  77                        imp->imp_obd->obd_name, uuid->uuid,
  78                        libcfs_net2str(refnet));
  79
  80         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, refnet);
  81         if (!ptlrpc_conn) {
  82                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  83                 RETURN(-ENOENT);
  84         }
  85
  86         if (create) {
  87                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  88                 if (!imp_conn)
  89                         GOTO(out_put, rc = -ENOMEM);
  90         }
  91
  92         spin_lock(&imp->imp_lock);
  93         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  94                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  95                         if (priority) {
  96                                 list_move(&item->oic_item,
  97                                           &imp->imp_conn_list);
  98                                 item->oic_last_attempt = 0;
  99                         }
 100                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
 101                                imp, imp->imp_obd->obd_name, uuid->uuid,
 102                                (priority ? ", moved to head" : ""));
 103                         spin_unlock(&imp->imp_lock);
 104                         GOTO(out_free, rc = 0);
 105                 }
 106         }
 107         /* No existing import connection found for \a uuid. */
 108         if (create) {
 109                 imp_conn->oic_conn = ptlrpc_conn;
 110                 imp_conn->oic_uuid = *uuid;
 111                 imp_conn->oic_last_attempt = 0;
 112                 if (priority)
 113                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 114                 else
 115                         list_add_tail(&imp_conn->oic_item,
 116                                       &imp->imp_conn_list);
 117                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 118                        imp, imp->imp_obd->obd_name, uuid->uuid,
 119                        (priority ? "head" : "tail"));
 120         } else {
 121                 spin_unlock(&imp->imp_lock);
 122                 GOTO(out_free, rc = -ENOENT);
 123         }
 124
 125         spin_unlock(&imp->imp_lock);
 126         RETURN(0);
 127 out_free:
 128         if (imp_conn)
 129                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 130 out_put:
 131         ptlrpc_connection_put(ptlrpc_conn);
 132         RETURN(rc);
 133 }
 134
 135 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 136 {
 137         return import_set_conn(imp, uuid, 1, 0);
 138 }
 139
 140 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 141                            int priority)
 142 {
 143         return import_set_conn(imp, uuid, priority, 1);
 144 }
 145 EXPORT_SYMBOL(client_import_add_conn);
 146
 147 int client_import_dyn_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 148                                lnet_nid_t prim_nid4, int priority)
 149 {
 150         struct ptlrpc_connection *ptlrpc_conn;
 151         struct lnet_nid prim_nid;
 152         int rc;
 153
 154         lnet_nid4_to_nid(prim_nid4, &prim_nid);
 155         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid, LNET_NID_NET(&prim_nid));
 156         if (!ptlrpc_conn) {
 157                 const char *str_uuid = obd_uuid2str(uuid);
 158
 159                 rc = class_add_uuid(str_uuid, &prim_nid);
 160                 if (rc) {
 161                         CERROR("%s: failed to add UUID '%s': rc = %d\n",
 162                                imp->imp_obd->obd_name, str_uuid, rc);
 163                         return rc;
 164                 }
 165         }
 166         return import_set_conn(imp, uuid, priority, 1);
 167 }
 168 EXPORT_SYMBOL(client_import_dyn_add_conn);
 169
 170 int client_import_add_nids_to_conn(struct obd_import *imp, lnet_nid_t *nids,
 171                                    int nid_count, struct obd_uuid *uuid)
 172 {
 173         struct obd_import_conn *conn;
 174         int rc = -ENOENT;
 175
 176         ENTRY;
 177         if (nid_count <= 0 || !nids)
 178                 return rc;
 179
 180         spin_lock(&imp->imp_lock);
 181         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 182                 struct lnet_nid nid;
 183
 184                 lnet_nid4_to_nid(nids[0], &nid);
 185                 if (class_check_uuid(&conn->oic_uuid, &nid)) {
 186                         *uuid = conn->oic_uuid;
 187                         spin_unlock(&imp->imp_lock);
 188                         rc = class_add_nids_to_uuid(&conn->oic_uuid, nids,
 189                                                     nid_count);
 190                         RETURN(rc);
 191                 }
 192         }
 193         spin_unlock(&imp->imp_lock);
 194         RETURN(rc);
 195 }
 196 EXPORT_SYMBOL(client_import_add_nids_to_conn);
 197
 198 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 199 {
 200         struct obd_import_conn *imp_conn;
 201         struct obd_export *dlmexp;
 202         int rc = -ENOENT;
 203
 204         ENTRY;
 205
 206         spin_lock(&imp->imp_lock);
 207         if (list_empty(&imp->imp_conn_list)) {
 208                 LASSERT(!imp->imp_connection);
 209                 GOTO(out, rc);
 210         }
 211
 212         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 213                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 214                         continue;
 215                 LASSERT(imp_conn->oic_conn);
 216
 217                 if (imp_conn == imp->imp_conn_current) {
 218                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 219
 220                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 221                             imp->imp_state != LUSTRE_IMP_DISCON) {
 222                                 CERROR("can't remove current connection\n");
 223                                 GOTO(out, rc = -EBUSY);
 224                         }
 225
 226                         ptlrpc_connection_put(imp->imp_connection);
 227                         imp->imp_connection = NULL;
 228
 229                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 230                         if (dlmexp && dlmexp->exp_connection) {
 231                                 LASSERT(dlmexp->exp_connection ==
 232                                         imp_conn->oic_conn);
 233                                 ptlrpc_connection_put(dlmexp->exp_connection);
 234                                 dlmexp->exp_connection = NULL;
 235                         }
 236
 237                         if (dlmexp != NULL)
 238                                 class_export_put(dlmexp);
 239                 }
 240
 241                 list_del(&imp_conn->oic_item);
 242                 ptlrpc_connection_put(imp_conn->oic_conn);
 243                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 244                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 245                        imp, imp->imp_obd->obd_name, uuid->uuid);
 246                 rc = 0;
 247                 break;
 248         }
 249 out:
 250         spin_unlock(&imp->imp_lock);
 251         if (rc == -ENOENT)
 252                 CERROR("connection %s not found\n", uuid->uuid);
 253         RETURN(rc);
 254 }
 255 EXPORT_SYMBOL(client_import_del_conn);
 256
 257 void client_destroy_import(struct obd_import *imp)
 258 {
 259         /*
 260          * Drop security policy instance after all RPCs have finished/aborted
 261          * to let all busy contexts be released.
 262          */
 263         class_import_get(imp);
 264         class_destroy_import(imp);
 265         sptlrpc_import_sec_put(imp);
 266         class_import_put(imp);
 267 }
 268 EXPORT_SYMBOL(client_destroy_import);
 269
 270 /**
 271  * Check whether or not the OSC is on MDT.
 272  * In the config log,
 273  * osc on MDT
 274  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 275  * osc on client
 276  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 277  *
 278  **/
 279 static int osc_on_mdt(char *obdname)
 280 {
 281         char *ptr;
 282
 283         ptr = strrchr(obdname, '-');
 284         if (ptr == NULL)
 285                 return 0;
 286
 287         if (strncmp(ptr + 1, "MDT", 3) == 0)
 288                 return 1;
 289
 290         return 0;
 291 }
 292
 293 /*
 294  * Configure an RPC client OBD device.
 295  *
 296  * lcfg parameters:
 297  * 1 - client UUID
 298  * 2 - server UUID
 299  * 3 - inactive-on-startup
 300  * 4 - restrictive net
 301  */
 302 int client_obd_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
 303 {
 304         struct client_obd *cli = &obd->u.cli;
 305         struct obd_import *imp;
 306         struct obd_uuid server_uuid;
 307         int rq_portal, rp_portal, connect_op;
 308         const char *name = obd->obd_type->typ_name;
 309         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 310         char *cli_name = lustre_cfg_buf(lcfg, 0);
 311         int rc;
 312
 313         ENTRY;
 314
 315         /*
 316          * In a more perfect world, we would hang a ptlrpc_client off of
 317          * obd_type and just use the values from there.
 318          */
 319         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 320                 rq_portal = OST_REQUEST_PORTAL;
 321                 rp_portal = OSC_REPLY_PORTAL;
 322                 connect_op = OST_CONNECT;
 323                 cli->cl_sp_me = LUSTRE_SP_CLI;
 324                 cli->cl_sp_to = LUSTRE_SP_OST;
 325                 ns_type = LDLM_NS_TYPE_OSC;
 326         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 327                    !strcmp(name, LUSTRE_LWP_NAME)) {
 328                 rq_portal = MDS_REQUEST_PORTAL;
 329                 rp_portal = MDC_REPLY_PORTAL;
 330                 connect_op = MDS_CONNECT;
 331                 if (is_lwp_on_ost(cli_name))
 332                         cli->cl_sp_me = LUSTRE_SP_OST;
 333                 else if (is_lwp_on_mdt(cli_name))
 334                         cli->cl_sp_me = LUSTRE_SP_MDT;
 335                 else
 336                         cli->cl_sp_me = LUSTRE_SP_CLI;
 337                 cli->cl_sp_to = LUSTRE_SP_MDT;
 338                 ns_type = LDLM_NS_TYPE_MDC;
 339         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 340                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 341                         /* OSP_on_MDT for other MDTs */
 342                         connect_op = MDS_CONNECT;
 343                         cli->cl_sp_to = LUSTRE_SP_MDT;
 344                         ns_type = LDLM_NS_TYPE_MDC;
 345                         rq_portal = OUT_PORTAL;
 346                 } else {
 347                         /* OSP on MDT for OST */
 348                         connect_op = OST_CONNECT;
 349                         cli->cl_sp_to = LUSTRE_SP_OST;
 350                         ns_type = LDLM_NS_TYPE_OSC;
 351                         rq_portal = OST_REQUEST_PORTAL;
 352                 }
 353                 rp_portal = OSC_REPLY_PORTAL;
 354                 cli->cl_sp_me = LUSTRE_SP_MDT;
 355         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 356                 rq_portal = MGS_REQUEST_PORTAL;
 357                 rp_portal = MGC_REPLY_PORTAL;
 358                 connect_op = MGS_CONNECT;
 359                 cli->cl_sp_me = LUSTRE_SP_MGC;
 360                 cli->cl_sp_to = LUSTRE_SP_MGS;
 361                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 362                 ns_type = LDLM_NS_TYPE_MGC;
 363         } else {
 364                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 365                        name);
 366                 RETURN(-EINVAL);
 367         }
 368
 369         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 370                 CERROR("requires a TARGET UUID\n");
 371                 RETURN(-EINVAL);
 372         }
 373
 374         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 375                 CERROR("client UUID must be less than 38 characters\n");
 376                 RETURN(-EINVAL);
 377         }
 378
 379         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 380                 CERROR("setup requires a SERVER UUID\n");
 381                 RETURN(-EINVAL);
 382         }
 383
 384         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 385                 CERROR("target UUID must be less than 38 characters\n");
 386                 RETURN(-EINVAL);
 387         }
 388
 389         init_rwsem(&cli->cl_sem);
 390         mutex_init(&cli->cl_mgc_mutex);
 391         cli->cl_seq = NULL;
 392         init_rwsem(&cli->cl_seq_rwsem);
 393         cli->cl_conn_count = 0;
 394         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 395                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 396                      sizeof(server_uuid)));
 397
 398         cli->cl_dirty_pages = 0;
 399         cli->cl_dirty_max_pages = 0;
 400         cli->cl_avail_grant = 0;
 401         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 402         /*
 403          * cl_dirty_max_pages may be changed at connect time in
 404          * ptlrpc_connect_interpret().
 405          */
 406         client_adjust_max_dirty(cli);
 407         init_waitqueue_head(&cli->cl_cache_waiters);
 408         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 409         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 410         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 411         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 412         spin_lock_init(&cli->cl_loi_list_lock);
 413         atomic_set(&cli->cl_pending_w_pages, 0);
 414         atomic_set(&cli->cl_pending_r_pages, 0);
 415         cli->cl_r_in_flight = 0;
 416         cli->cl_w_in_flight = 0;
 417
 418         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 419         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 420         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 421         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 422         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 423         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 424         spin_lock_init(&cli->cl_batch_rpc_hist.oh_lock);
 425
 426         /* lru for osc. */
 427         INIT_LIST_HEAD(&cli->cl_lru_osc);
 428         atomic_set(&cli->cl_lru_shrinkers, 0);
 429         atomic_long_set(&cli->cl_lru_busy, 0);
 430         atomic_long_set(&cli->cl_lru_in_list, 0);
 431         INIT_LIST_HEAD(&cli->cl_lru_list);
 432         spin_lock_init(&cli->cl_lru_list_lock);
 433         atomic_long_set(&cli->cl_unstable_count, 0);
 434         INIT_LIST_HEAD(&cli->cl_shrink_list);
 435         INIT_LIST_HEAD(&cli->cl_grant_chain);
 436
 437         INIT_LIST_HEAD(&cli->cl_flight_waiters);
 438         cli->cl_rpcs_in_flight = 0;
 439
 440         init_waitqueue_head(&cli->cl_destroy_waitq);
 441         atomic_set(&cli->cl_destroy_in_flight, 0);
 442
 443
 444         cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 445         cli->cl_preferred_cksum_type = 0;
 446 #ifdef ENABLE_CHECKSUM
 447         /* Turn on checksumming by default. */
 448         cli->cl_checksum = 1;
 449         /*
 450          * The supported checksum types will be worked out at connect time
 451          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 452          * through procfs.
 453          */
 454         cli->cl_cksum_type = cli->cl_supp_cksum_types;
 455 #endif
 456         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 457
 458         /*
 459          * Set it to possible maximum size. It may be reduced by ocd_brw_size
 460          * from OFD after connecting.
 461          */
 462         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 463
 464         cli->cl_max_short_io_bytes = OBD_DEF_SHORT_IO_BYTES;
 465
 466         /*
 467          * set cl_chunkbits default value to PAGE_SHIFT,
 468          * it will be updated at OSC connection time.
 469          */
 470         cli->cl_chunkbits = PAGE_SHIFT;
 471
 472         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 473                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 474         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 475                 cli->cl_max_rpcs_in_flight = 2;
 476         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 477                 cli->cl_max_rpcs_in_flight = 3;
 478         } else if (cfs_totalram_pages() >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 479                 cli->cl_max_rpcs_in_flight = 4;
 480         } else {
 481                 if (osc_on_mdt(obd->obd_name))
 482                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 483                 else
 484                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 485         }
 486
 487         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 488         cli->cl_max_mod_rpcs_in_flight = 0;
 489         cli->cl_mod_rpcs_in_flight = 0;
 490         cli->cl_close_rpcs_in_flight = 0;
 491         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 492         cli->cl_mod_rpcs_init = ktime_get_real();
 493         cli->cl_mod_tag_bitmap = NULL;
 494
 495         INIT_LIST_HEAD(&cli->cl_chg_dev_linkage);
 496
 497         if (connect_op == MDS_CONNECT) {
 498                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 499                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 500                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 501                 if (cli->cl_mod_tag_bitmap == NULL)
 502                         GOTO(err, rc = -ENOMEM);
 503         }
 504
 505         rc = ldlm_get_ref();
 506         if (rc) {
 507                 CERROR("ldlm_get_ref failed: %d\n", rc);
 508                 GOTO(err, rc);
 509         }
 510
 511         ptlrpc_init_client(rq_portal, rp_portal, name,
 512                            &obd->obd_ldlm_client);
 513
 514         imp = class_new_import(obd);
 515         if (imp == NULL)
 516                 GOTO(err_ldlm, rc = -ENOENT);
 517         imp->imp_client = &obd->obd_ldlm_client;
 518         imp->imp_connect_op = connect_op;
 519         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 520                LUSTRE_CFG_BUFLEN(lcfg, 1));
 521         class_import_put(imp);
 522
 523         if (lustre_cfg_buf(lcfg, 4)) {
 524                 __u32 refnet = libcfs_str2net(lustre_cfg_string(lcfg, 4));
 525
 526                 if (refnet == LNET_NET_ANY) {
 527                         rc = -EINVAL;
 528                         CERROR("%s: bad mount option 'network=%s': rc = %d\n",
 529                                obd->obd_name, lustre_cfg_string(lcfg, 4),
 530                                rc);
 531                         GOTO(err_import, rc);
 532                 }
 533                 imp->imp_conn_restricted_net = refnet;
 534         } else {
 535                 imp->imp_conn_restricted_net = LNET_NET_ANY;
 536         }
 537
 538         rc = client_import_add_conn(imp, &server_uuid, 1);
 539         if (rc) {
 540                 CERROR("can't add initial connection\n");
 541                 GOTO(err_import, rc);
 542         }
 543         imp->imp_connection = NULL;
 544
 545         cli->cl_import = imp;
 546         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 547         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 548
 549         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 550                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 551                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 552                                name, obd->obd_name,
 553                                cli->cl_target_uuid.uuid);
 554                         spin_lock(&imp->imp_lock);
 555                         imp->imp_deactive = 1;
 556                         spin_unlock(&imp->imp_lock);
 557                 }
 558         }
 559
 560         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 561                                                 LDLM_NAMESPACE_CLIENT,
 562                                                 LDLM_NAMESPACE_GREEDY,
 563                                                 ns_type);
 564         if (IS_ERR(obd->obd_namespace)) {
 565                 rc = PTR_ERR(obd->obd_namespace);
 566                 CERROR("%s: unable to create client namespace: rc = %d\n",
 567                        obd->obd_name, rc);
 568                 obd->obd_namespace = NULL;
 569                 GOTO(err_import, rc);
 570         }
 571
 572         RETURN(rc);
 573
 574 err_import:
 575         class_destroy_import(imp);
 576 err_ldlm:
 577         ldlm_put_ref();
 578 err:
 579         if (cli->cl_mod_tag_bitmap != NULL)
 580                 OBD_FREE(cli->cl_mod_tag_bitmap,
 581                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 582         cli->cl_mod_tag_bitmap = NULL;
 583
 584         RETURN(rc);
 585 }
 586 EXPORT_SYMBOL(client_obd_setup);
 587
 588 int client_obd_cleanup(struct obd_device *obd)
 589 {
 590         struct client_obd *cli = &obd->u.cli;
 591
 592         ENTRY;
 593
 594         ldlm_namespace_free_post(obd->obd_namespace);
 595         obd->obd_namespace = NULL;
 596
 597         obd_cleanup_client_import(obd);
 598         LASSERT(obd->u.cli.cl_import == NULL);
 599
 600         ldlm_put_ref();
 601
 602         if (cli->cl_mod_tag_bitmap != NULL)
 603                 OBD_FREE(cli->cl_mod_tag_bitmap,
 604                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 605         cli->cl_mod_tag_bitmap = NULL;
 606
 607         RETURN(0);
 608 }
 609 EXPORT_SYMBOL(client_obd_cleanup);
 610
 611 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 612 int client_connect_import(const struct lu_env *env,
 613                           struct obd_export **exp,
 614                           struct obd_device *obd, struct obd_uuid *cluuid,
 615                           struct obd_connect_data *data, void *localdata)
 616 {
 617         struct client_obd *cli = &obd->u.cli;
 618         struct obd_import *imp = cli->cl_import;
 619         struct obd_connect_data *ocd;
 620         struct lustre_handle conn = { 0 };
 621         int rc;
 622
 623         ENTRY;
 624
 625         *exp = NULL;
 626         down_write(&cli->cl_sem);
 627         if (cli->cl_conn_count > 0)
 628                 GOTO(out_sem, rc = -EALREADY);
 629
 630         rc = class_connect(&conn, obd, cluuid);
 631         if (rc)
 632                 GOTO(out_sem, rc);
 633
 634         cli->cl_conn_count++;
 635         *exp = class_conn2export(&conn);
 636
 637         LASSERT(obd->obd_namespace);
 638
 639         spin_lock(&imp->imp_lock);
 640         if (imp->imp_state == LUSTRE_IMP_CLOSED && imp->imp_deactive) {
 641                 /* need to reactivate import if trying to connect
 642                  * to a previously disconnected
 643                  */
 644                 imp->imp_deactive = 0;
 645                 imp->imp_invalid = 0;
 646         }
 647         spin_unlock(&imp->imp_lock);
 648
 649         imp->imp_dlm_handle = conn;
 650         rc = ptlrpc_init_import(imp);
 651         if (rc != 0)
 652                 GOTO(out_ldlm, rc);
 653
 654         ocd = &imp->imp_connect_data;
 655         if (data) {
 656                 *ocd = *data;
 657                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 658                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 659         }
 660
 661         rc = ptlrpc_connect_import(imp);
 662         if (rc != 0) {
 663                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 664                 GOTO(out_ldlm, rc);
 665         }
 666         LASSERT(*exp != NULL && (*exp)->exp_connection);
 667
 668         if (data) {
 669                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 670                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 671                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 672                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 673                 data->ocd_connect_flags2 = ocd->ocd_connect_flags2;
 674         }
 675
 676         ptlrpc_pinger_add_import(imp);
 677
 678         EXIT;
 679
 680         if (rc) {
 681 out_ldlm:
 682                 cli->cl_conn_count--;
 683                 class_disconnect(*exp);
 684                 *exp = NULL;
 685         }
 686 out_sem:
 687         up_write(&cli->cl_sem);
 688
 689         if (!rc && localdata && !cli->cl_cache) {
 690                 cli->cl_cache = (struct cl_client_cache *)localdata;
 691                 cl_cache_incref(cli->cl_cache);
 692                 cli->cl_lru_left = &cli->cl_cache->ccc_lru_left;
 693
 694                 /* add this osc into entity list */
 695                 LASSERT(list_empty(&cli->cl_lru_osc));
 696                 spin_lock(&cli->cl_cache->ccc_lru_lock);
 697                 list_add(&cli->cl_lru_osc, &cli->cl_cache->ccc_lru);
 698                 spin_unlock(&cli->cl_cache->ccc_lru_lock);
 699         }
 700
 701         return rc;
 702 }
 703 EXPORT_SYMBOL(client_connect_import);
 704
 705 int client_disconnect_export(struct obd_export *exp)
 706 {
 707         struct obd_device *obd = class_exp2obd(exp);
 708         struct client_obd *cli;
 709         struct obd_import *imp;
 710         int rc = 0, err;
 711
 712         ENTRY;
 713
 714         if (!obd) {
 715                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 716                        exp, exp ? exp->exp_handle.h_cookie : -1);
 717                 RETURN(-EINVAL);
 718         }
 719
 720         cli = &obd->u.cli;
 721         imp = cli->cl_import;
 722
 723         down_write(&cli->cl_sem);
 724         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 725                 cli->cl_conn_count);
 726
 727         if (cli->cl_conn_count == 0) {
 728                 CERROR("disconnecting disconnected device (%s)\n",
 729                        obd->obd_name);
 730                 GOTO(out_disconnect, rc = -EINVAL);
 731         }
 732
 733         cli->cl_conn_count--;
 734         if (cli->cl_conn_count != 0)
 735                 GOTO(out_disconnect, rc = 0);
 736
 737         /*
 738          * Mark import deactivated now, so we don't try to reconnect if any
 739          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 740          * fully deactivate the import, or that would drop all requests.
 741          */
 742         spin_lock(&imp->imp_lock);
 743         imp->imp_deactive = 1;
 744         spin_unlock(&imp->imp_lock);
 745
 746         /*
 747          * Some non-replayable imports (MDS's OSCs) are pinged, so just
 748          * delete it regardless.  (It's safe to delete an import that was
 749          * never added.)
 750          */
 751         (void)ptlrpc_pinger_del_import(imp);
 752
 753         if (obd->obd_namespace != NULL) {
 754                 /* obd_force == local only */
 755                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 756                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 757                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 758                                           obd->obd_force);
 759         }
 760
 761         /*
 762          * There's no need to hold sem while disconnecting an import,
 763          * and it may actually cause deadlock in GSS.
 764          */
 765         up_write(&cli->cl_sem);
 766         rc = ptlrpc_disconnect_import(imp, 0);
 767         down_write(&cli->cl_sem);
 768
 769         ptlrpc_invalidate_import(imp);
 770
 771         EXIT;
 772
 773 out_disconnect:
 774         /*
 775          * Use server style - class_disconnect should be always called for
 776          * o_disconnect.
 777          */
 778         err = class_disconnect(exp);
 779         if (!rc && err)
 780                 rc = err;
 781
 782         up_write(&cli->cl_sem);
 783
 784         RETURN(rc);
 785 }
 786 EXPORT_SYMBOL(client_disconnect_export);
 787
 788 #ifdef HAVE_SERVER_SUPPORT
 789 int server_disconnect_export(struct obd_export *exp)
 790 {
 791         int rc;
 792
 793         ENTRY;
 794
 795         /* Disconnect early so that clients can't keep using export. */
 796         rc = class_disconnect(exp);
 797         /* Close import to avoid sending any requests. */
 798         if (exp->exp_imp_reverse)
 799                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 800
 801         ldlm_bl_thread_wakeup();
 802
 803         /* complete all outstanding replies */
 804         spin_lock(&exp->exp_lock);
 805         while (!list_empty(&exp->exp_outstanding_replies)) {
 806                 struct ptlrpc_reply_state *rs =
 807                         list_entry(exp->exp_outstanding_replies.next,
 808                                        struct ptlrpc_reply_state, rs_exp_list);
 809                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 810
 811                 spin_lock(&svcpt->scp_rep_lock);
 812
 813                 list_del_init(&rs->rs_exp_list);
 814
 815                 spin_lock(&rs->rs_lock);
 816                 /* clear rs_convert_lock to make sure rs is handled and put */
 817                 rs->rs_convert_lock = 0;
 818                 ptlrpc_schedule_difficult_reply(rs);
 819                 spin_unlock(&rs->rs_lock);
 820
 821                 spin_unlock(&svcpt->scp_rep_lock);
 822         }
 823         spin_unlock(&exp->exp_lock);
 824
 825         RETURN(rc);
 826 }
 827 EXPORT_SYMBOL(server_disconnect_export);
 828
 829 static inline int target_check_recovery_timer(struct obd_device *target)
 830 {
 831         ktime_t remaining;
 832         s64 timeout;
 833
 834         if (!target->obd_recovering || target->obd_recovery_start == 0)
 835                 return 0;
 836
 837         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
 838         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 839         if (timeout > -30)
 840                 return 0;
 841
 842         /* the recovery timer should expire, but it isn't triggered,
 843          * it's better to abort the recovery of this target to speed up
 844          * the recovery of the whole cluster. */
 845         spin_lock(&target->obd_dev_lock);
 846         if (target->obd_recovering) {
 847                 CERROR("%s: Aborting recovery\n", target->obd_name);
 848                 target->obd_abort_recovery = 1;
 849                 wake_up(&target->obd_next_transno_waitq);
 850         }
 851         spin_unlock(&target->obd_dev_lock);
 852         return 0;
 853 }
 854
 855 /*
 856  * --------------------------------------------------------------------------
 857  * from old lib/target.c
 858  * --------------------------------------------------------------------------
 859  */
 860 static int target_handle_reconnect(struct lustre_handle *conn,
 861                                    struct obd_export *exp,
 862                                    struct obd_uuid *cluuid)
 863 {
 864         struct obd_device *target;
 865         struct lustre_handle *hdl;
 866         ktime_t remaining;
 867         s64 timeout;
 868         int rc = 0;
 869
 870         ENTRY;
 871         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 872         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 873                 conn->cookie = exp->exp_handle.h_cookie;
 874                 CDEBUG(D_HA,
 875                        "connect export for UUID '%s' at %p, cookie %#llx\n",
 876                        cluuid->uuid, exp, conn->cookie);
 877                 RETURN(0);
 878         }
 879
 880         target = exp->exp_obd;
 881
 882         /* Might be a re-connect after a partition. */
 883         if (memcmp(&conn->cookie, &hdl->cookie, sizeof(conn->cookie))) {
 884                 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",
 885                               target->obd_name,
 886                               obd_uuid2str(&exp->exp_client_uuid),
 887                               obd_export_nid2str(exp),
 888                               hdl->cookie, conn->cookie);
 889                 memset(conn, 0, sizeof(*conn));
 890                 /*
 891                  * target_handle_connect() treats EALREADY and
 892                  * -EALREADY differently.  -EALREADY is an error
 893                  * (same UUID, different handle).
 894                  */
 895                 RETURN(-EALREADY);
 896         }
 897
 898         if (!target->obd_recovering) {
 899                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 900                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 901                         obd_export_nid2str(exp));
 902                 GOTO(out_already, rc);
 903         }
 904
 905         remaining = hrtimer_get_remaining(&target->obd_recovery_timer);
 906         timeout = ktime_divns(remaining, NSEC_PER_SEC);
 907         if (timeout > 0) {
 908                 LCONSOLE_WARN("%s: Client %s (at %s) reconnected, waiting for %d clients in recovery for %lld:%.02lld\n",
 909                               target->obd_name,
 910                               obd_uuid2str(&exp->exp_client_uuid),
 911                               obd_export_nid2str(exp),
 912                               atomic_read(&target->obd_max_recoverable_clients),
 913                               timeout / 60, timeout % 60);
 914         } else {
 915                 struct target_distribute_txn_data *tdtd;
 916                 int size = 0;
 917                 int count = 0;
 918                 char *buf = NULL;
 919
 920                 target_check_recovery_timer(target);
 921
 922                 tdtd = class_exp2tgt(exp)->lut_tdtd;
 923                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 924                         buf = tdtd->tdtd_show_update_logs_retrievers(
 925                                 tdtd->tdtd_show_retrievers_cbdata,
 926                                 &size, &count);
 927
 928                 if (count > 0)
 929                         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",
 930                                       target->obd_name,
 931                                       obd_uuid2str(&exp->exp_client_uuid),
 932                                       obd_export_nid2str(exp), count,
 933                                       buf ? buf : "unknown (not enough RAM)",
 934                                       (abs(timeout) + target->obd_recovery_timeout) / 60,
 935                                       (abs(timeout) + target->obd_recovery_timeout) % 60,
 936                                       target->obd_name);
 937                 else
 938                         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",
 939                                       target->obd_name, abs(timeout) / 60,
 940                                       abs(timeout) % 60, target->obd_name);
 941
 942                 if (buf != NULL)
 943                         OBD_FREE(buf, size);
 944         }
 945
 946 out_already:
 947         conn->cookie = exp->exp_handle.h_cookie;
 948         /*
 949          * target_handle_connect() treats EALREADY and
 950          * -EALREADY differently.  EALREADY means we are
 951          * doing a valid reconnect from the same client.
 952          */
 953         RETURN(EALREADY);
 954 }
 955
 956 static void
 957 check_and_start_recovery_timer(struct obd_device *obd,
 958                                struct ptlrpc_request *req, int new_client);
 959
 960 /**
 961  * update flags for import during reconnect process
 962  */
 963 static int rev_import_flags_update(struct obd_import *revimp,
 964                                    struct ptlrpc_request *req)
 965 {
 966         int rc;
 967         struct obd_connect_data *data;
 968
 969         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 970
 971         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 972                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 973         else
 974                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 975
 976         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 977
 978         revimp->imp_connect_data = *data;
 979         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 980         if (rc) {
 981                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 982                         revimp->imp_client->cli_name,
 983                         libcfs_idstr(&req->rq_peer), rc);
 984                 return rc;
 985         }
 986
 987         return 0;
 988 }
 989
 990 /**
 991  * Allocate a new reverse import for an export.
 992  *
 993  * \retval -errno in case error hit
 994  * \retval 0 if reverse import correctly init
 995  **/
 996 int rev_import_init(struct obd_export *export)
 997 {
 998         struct obd_device *obd = export->exp_obd;
 999         struct obd_import *revimp;
1000
1001         LASSERT(export->exp_imp_reverse == NULL);
1002
1003         revimp = class_new_import(obd);
1004         if (revimp == NULL)
1005                 return -ENOMEM;
1006
1007         revimp->imp_remote_handle.cookie = 0ULL;
1008         revimp->imp_client = &obd->obd_ldlm_client;
1009         revimp->imp_dlm_fake = 1;
1010
1011         /* it is safe to connect import in new state as no sends possible */
1012         spin_lock(&export->exp_lock);
1013         export->exp_imp_reverse = revimp;
1014         spin_unlock(&export->exp_lock);
1015         class_import_put(revimp);
1016
1017         return 0;
1018 }
1019 EXPORT_SYMBOL(rev_import_init);
1020
1021 /**
1022  * Handle reconnect for an export.
1023  *
1024  * \param exp export to handle reconnect process
1025  * \param req client reconnect request
1026  *
1027  * \retval -rc in case securitfy flavor can't be changed
1028  * \retval 0 in case none problems
1029  */
1030 static int rev_import_reconnect(struct obd_export *exp,
1031                                 struct ptlrpc_request *req)
1032 {
1033         struct obd_import *revimp = exp->exp_imp_reverse;
1034         struct lustre_handle *lh;
1035         int rc;
1036
1037         /* avoid sending a request until import flags are changed */
1038         ptlrpc_import_enter_resend(revimp);
1039
1040         ptlrpc_connection_put(revimp->imp_connection);
1041
1042         /*
1043          * client from recovery don't have a handle so we need to take from
1044          * request. it may produce situation when wrong client connected
1045          * to recovery as we trust a client uuid
1046          */
1047         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1048         revimp->imp_remote_handle = *lh;
1049
1050         /*
1051          * unknown versions will be caught in
1052          * ptlrpc_handle_server_req_in->lustre_unpack_msg()
1053          */
1054         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1055
1056         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
1057
1058         rc = rev_import_flags_update(revimp, req);
1059         if (rc != 0) {
1060                 /*
1061                  * it is safe to still be in RECOVERY phase as we are not able
1062                  * to setup correct security flavor so requests are not able to
1063                  * be delivered correctly
1064                  */
1065                 return rc;
1066         }
1067
1068         /* resend all rpc's via new connection */
1069         return ptlrpc_import_recovery_state_machine(revimp);
1070 }
1071
1072 int target_handle_connect(struct ptlrpc_request *req)
1073 {
1074         struct obd_device *target = NULL;
1075         struct obd_export *export = NULL;
1076         /*
1077          * connect handle - filled from target_handle_reconnect in
1078          * reconnect case
1079          */
1080         struct lustre_handle conn;
1081         struct lustre_handle *tmp;
1082         struct obd_uuid cluuid;
1083         char *str;
1084         int rc = 0;
1085         char *target_start;
1086         int target_len;
1087         bool mds_conn = false, lw_client = false, initial_conn = false;
1088         bool mds_mds_conn = false;
1089         bool new_mds_mds_conn = false;
1090         struct obd_connect_data *data, *tmpdata;
1091         int size, tmpsize;
1092         lnet_nid_t client_nid;
1093 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1094         int tmp_exp_old_falloc;
1095 #endif
1096         struct ptlrpc_connection *pcon = NULL;
1097         bool reconnected = false;
1098
1099         ENTRY;
1100
1101         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
1102
1103         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
1104         if (str == NULL) {
1105                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
1106                 GOTO(out, rc = -EINVAL);
1107         }
1108
1109         target = class_dev_by_str(str);
1110         if (!target) {
1111                 deuuidify(str, NULL, &target_start, &target_len);
1112                 LCONSOLE_ERROR_MSG(0x137,
1113                                    "%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",
1114                                    str, libcfs_nidstr(&req->rq_peer.nid));
1115                 GOTO(out, rc = -ENODEV);
1116         }
1117
1118         spin_lock(&target->obd_dev_lock);
1119
1120         target->obd_conn_inprogress++;
1121
1122         if (target->obd_stopping || !target->obd_set_up) {
1123                 spin_unlock(&target->obd_dev_lock);
1124
1125                 deuuidify(str, NULL, &target_start, &target_len);
1126                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
1127                               target_len, target_start,
1128                               libcfs_nidstr(&req->rq_peer.nid),
1129                               (target->obd_stopping ?
1130                                "stopping" : "not set up"));
1131                 GOTO(out, rc = -ENODEV);
1132         }
1133
1134         if (target->obd_no_conn) {
1135                 spin_unlock(&target->obd_dev_lock);
1136
1137                 CDEBUG(D_INFO,
1138                        "%s: Temporarily refusing client connection from %s\n",
1139                        target->obd_name, libcfs_nidstr(&req->rq_peer.nid));
1140                 GOTO(out, rc = -EAGAIN);
1141         }
1142
1143         spin_unlock(&target->obd_dev_lock);
1144
1145         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1146         if (str == NULL) {
1147                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1148                 GOTO(out, rc = -EINVAL);
1149         }
1150
1151         obd_str2uuid(&cluuid, str);
1152
1153         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1154         if (tmp == NULL)
1155                 GOTO(out, rc = -EPROTO);
1156
1157         conn = *tmp;
1158
1159         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1160                                     RCL_CLIENT);
1161         if (size < 0 || size > 8 * sizeof(struct obd_connect_data))
1162                 GOTO(out, rc = -EPROTO);
1163         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1164         if (!data)
1165                 GOTO(out, rc = -EPROTO);
1166
1167         rc = req_capsule_server_pack(&req->rq_pill);
1168         if (rc)
1169                 GOTO(out, rc);
1170
1171 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1172         /*
1173          * Don't allow clients to connect that are using old 1.8 format
1174          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1175          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1176          * FULL20 flag should be set on all connections since 2.0, but no
1177          * longer affects behaviour.
1178          *
1179          * Later this check will be disabled and the flag can be retired
1180          * completely once interop with 3.0 is no longer needed.
1181          */
1182         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1183                 GOTO(out, rc = -EPROTO);
1184
1185         /* Old clients will have 'tmp_exp_old_falloc' as 1.
1186          * Newer clients (2.15) and beyond will have it set as 0
1187          */
1188         tmp_exp_old_falloc =
1189                 !!(data->ocd_connect_flags & OBD_CONNECT_OLD_FALLOC);
1190
1191         CDEBUG(D_INFO, "%s: ocd_connect_flags: %#llx tmp_exp_old_falloc: %d\n",
1192                target->obd_name, data->ocd_connect_flags, tmp_exp_old_falloc);
1193
1194         /*
1195          * Don't allow liblustre clients to connect.
1196          * - testing was disabled in v2_2_50_0-61-g6a75d65
1197          * - building was disabled in v2_5_58_0-28-g7277179
1198          * - client code was deleted in v2_6_50_0-101-gcdfbc72,
1199          * - clients were refused connect for version difference > 0.0.1.32
1200          */
1201         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1202                 DEBUG_REQ(D_WARNING, req, "Refusing libclient connection");
1203                 GOTO(out, rc = -EPROTO);
1204         }
1205 #endif
1206
1207         /*
1208          * Note: lw_client is needed in MDS-MDS failover during update log
1209          * processing, so we needs to allow lw_client to be connected at
1210          * anytime, instead of only the initial connection
1211          */
1212         lw_client = OCD_HAS_FLAG(data, LIGHTWEIGHT);
1213
1214         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1215                 initial_conn = true;
1216                 mds_conn = OCD_HAS_FLAG(data, MDS);
1217                 mds_mds_conn = OCD_HAS_FLAG(data, MDS_MDS);
1218
1219                 /*
1220                  * OBD_CONNECT_MNE_SWAB is removed at 2.14
1221                  * Checking OBD_CONNECT_FID can be removed in the future.
1222                  *
1223                  * Via check OBD_CONNECT_FID, we can distinguish whether
1224                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1225                  * MGC or MDT, since MGC does not use OBD_CONNECT_FID.
1226                  */
1227                 if (!lw_client &&
1228                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1229                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1230                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1231                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1232                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1233                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1234
1235                         /*
1236                          * We do not support the MDT-MDT interoperations with
1237                          * different version MDT because of protocol changes.
1238                          */
1239                         if (unlikely(major != LUSTRE_MAJOR ||
1240                                      minor != LUSTRE_MINOR ||
1241                                      abs(patch - LUSTRE_PATCH) > 3)) {
1242                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the connection from different version MDT (%d.%d.%d.%d) %s %s\n",
1243                                               target->obd_name, LUSTRE_MAJOR,
1244                                               LUSTRE_MINOR, LUSTRE_PATCH,
1245                                               LUSTRE_FIX, major, minor, patch,
1246                                               OBD_OCD_VERSION_FIX(data->ocd_version),
1247                                               libcfs_nidstr(&req->rq_peer.nid),
1248                                               str);
1249                                 GOTO(out, rc = -EPROTO);
1250                         }
1251                 }
1252         }
1253
1254         /* lctl gets a backstage, all-access pass. */
1255         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1256                 goto dont_check_exports;
1257
1258         export = obd_uuid_lookup(target, &cluuid);
1259         if (!export)
1260                 goto no_export;
1261
1262         /* We've found an export in the hash. */
1263
1264         spin_lock(&export->exp_lock);
1265
1266         if (export->exp_connecting) { /* b=9635, et. al. */
1267                 spin_unlock(&export->exp_lock);
1268                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1269                               export->exp_obd->obd_name, export,
1270                               libcfs_nidstr(&req->rq_peer.nid));
1271                 class_export_put(export);
1272                 export = NULL;
1273                 rc = -EALREADY;
1274         } else if ((mds_conn || (lw_client && initial_conn) ||
1275                    OCD_HAS_FLAG(data, MDS_MDS)) && export->exp_connection) {
1276                 spin_unlock(&export->exp_lock);
1277                 if (!nid_same(&req->rq_peer.nid,
1278                               &export->exp_connection->c_peer.nid)) {
1279                         /* MDS or LWP reconnected after failover. */
1280                         LCONSOLE_WARN("%s: Received %s connection from %s, removing former export from %s\n",
1281                                       target->obd_name,
1282                                       lw_client ? "LWP" : "MDS",
1283                                       libcfs_nidstr(&req->rq_peer.nid),
1284                                       libcfs_nidstr(&export->exp_connection->c_peer.nid));
1285                 } else {
1286                         /* New connection from the same NID. */
1287                         LCONSOLE_WARN("%s: Received new %s connection from %s, %s former export from same NID\n",
1288                                       target->obd_name,
1289                                       lw_client ? "LWP" : "MDS",
1290                                       libcfs_nidstr(&req->rq_peer.nid),
1291                                       OCD_HAS_FLAG(data, MDS_MDS) ?
1292                                       "keep" : "remove");
1293                 }
1294
1295                 if (nid_same(&req->rq_peer.nid,
1296                              &export->exp_connection->c_peer.nid) &&
1297                     OCD_HAS_FLAG(data, MDS_MDS)) {
1298                         /*
1299                          * Because exports between MDTs will always be
1300                          * kept, let's do not fail such export if they
1301                          * come from the same NID, otherwise it might
1302                          * cause eviction between MDTs, which might
1303                          * cause namespace inconsistency
1304                          */
1305                         spin_lock(&export->exp_lock);
1306                         export->exp_connecting = 1;
1307                         export->exp_conn_cnt = 0;
1308                         spin_unlock(&export->exp_lock);
1309                         conn.cookie = export->exp_handle.h_cookie;
1310                         rc = EALREADY;
1311                 } else {
1312                         class_fail_export(export);
1313                         class_export_put(export);
1314                         export = NULL;
1315                         rc = 0;
1316                 }
1317         } else if (export->exp_connection != NULL && initial_conn &&
1318                    !nid_same(&req->rq_peer.nid,
1319                              &export->exp_connection->c_peer.nid)) {
1320                 spin_unlock(&export->exp_lock);
1321                 /* In MDS failover we have static UUID but NID can change. */
1322                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when existing nid %s is already connected\n",
1323                               target->obd_name, cluuid.uuid,
1324                               libcfs_nidstr(&req->rq_peer.nid),
1325                               libcfs_nidstr(
1326                                       &export->exp_connection->c_peer.nid));
1327                 rc = -EALREADY;
1328                 class_export_put(export);
1329                 export = NULL;
1330         } else if (OBD_FAIL_PRECHECK(OBD_FAIL_TGT_RECOVERY_CONNECT) &&
1331                    !lw_client) {
1332                 spin_unlock(&export->exp_lock);
1333                 rc = -EAGAIN;
1334         } else {
1335                 export->exp_connecting = 1;
1336                 spin_unlock(&export->exp_lock);
1337                 LASSERT(export->exp_obd == target);
1338
1339                 rc = target_handle_reconnect(&conn, export, &cluuid);
1340         }
1341
1342         /* If we found an export, we already unlocked. */
1343         if (!export) {
1344 no_export:
1345                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1346         } else if (req->rq_export == NULL &&
1347                    atomic_read(&export->exp_rpc_count) > 0) {
1348                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, still busy with %d references\n",
1349                               target->obd_name, cluuid.uuid,
1350                               libcfs_nidstr(&req->rq_peer.nid),
1351                               refcount_read(&export->exp_handle.h_ref));
1352                         GOTO(out, rc = -EBUSY);
1353         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
1354                    rc != EALREADY) {
1355                 if (!strstr(cluuid.uuid, "mdt"))
1356                         LCONSOLE_WARN("%s: Rejecting reconnect from the known client %s (at %s) because it is indicating it is a new client\n",
1357                                       target->obd_name, cluuid.uuid,
1358                                       libcfs_nidstr(&req->rq_peer.nid));
1359                 GOTO(out, rc = -EALREADY);
1360         } else {
1361                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1362         }
1363
1364         if (rc < 0)
1365                 GOTO(out, rc);
1366
1367         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %lld last %lld\n",
1368                target->obd_name, cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1369                target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1370                export, ktime_get_seconds(),
1371                export ? export->exp_last_request_time : 0);
1372
1373         /*
1374          * If this is the first time a client connects, reset the recovery
1375          * timer. Discard lightweight connections which might be local.
1376          */
1377         if (!lw_client && rc == 0 && target->obd_recovering)
1378                 check_and_start_recovery_timer(target, req, export == NULL);
1379
1380         /*
1381          * We want to handle EALREADY but *not* -EALREADY from
1382          * target_handle_reconnect(), return reconnection state in a flag.
1383          */
1384         if (rc == EALREADY) {
1385                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1386                 rc = 0;
1387         } else {
1388                 LASSERT(rc == 0);
1389         }
1390
1391         /* Tell the client if we support replayable requests. */
1392         if (target->obd_replayable)
1393                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1394         client_nid = lnet_nid_to_nid4(&req->rq_peer.nid);
1395
1396         if (export == NULL) {
1397                 /* allow lightweight connections during recovery */
1398                 /*
1399                  * allow "new" MDT to be connected during recovery, since we
1400                  * need retrieve recovery update records from it
1401                  */
1402                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1403                         struct hrtimer *timer = &target->obd_recovery_timer;
1404                         ktime_t remaining;
1405                         s64 timeout, left;
1406                         int in_progress;
1407                         int connected;
1408                         int known;
1409                         int stale;
1410                         char *msg;
1411
1412                         connected = atomic_read(&target->obd_connected_clients);
1413                         in_progress = atomic_read(&target->obd_lock_replay_clients);
1414                         known =
1415                            atomic_read(&target->obd_max_recoverable_clients);
1416                         stale = target->obd_stale_clients;
1417                         remaining = hrtimer_get_remaining(timer);
1418                         left = ktime_divns(remaining, NSEC_PER_SEC);
1419
1420                         if (ktime_to_ns(remaining) > 0) {
1421                                 msg = "to recover in";
1422                                 timeout = left;
1423                         } else {
1424                                 msg = "already passed deadline";
1425                                 timeout = -left;
1426
1427                                 target_check_recovery_timer(target);
1428                         }
1429
1430                         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",
1431                                       target->obd_name, cluuid.uuid,
1432                                       libcfs_nidstr(&req->rq_peer.nid), known,
1433                                       connected - in_progress, in_progress,
1434                                       stale, msg, timeout / 60, timeout % 60);
1435                         rc = -EBUSY;
1436                 } else {
1437 dont_check_exports:
1438                         rc = obd_connect(req->rq_svc_thread->t_env,
1439                                          &export, target, &cluuid, data,
1440                                          &client_nid);
1441                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1442                                 lustre_msg_add_op_flags(req->rq_repmsg,
1443                                                         MSG_CONNECT_RECOVERING);
1444                         if (rc == 0) {
1445                                 conn.cookie = export->exp_handle.h_cookie;
1446                                 rc = rev_import_init(export);
1447                         }
1448
1449                         if (mds_mds_conn)
1450                                 new_mds_mds_conn = true;
1451                 }
1452         } else {
1453                 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_CONNECT_VS_EVICT)) {
1454                         class_export_get(export);
1455                         class_fail_export(export);
1456                         class_export_put(export);
1457                 }
1458                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1459                                    export, target, &cluuid, data, &client_nid);
1460                 if (rc == 0)
1461                         reconnected = true;
1462         }
1463         if (rc)
1464                 GOTO(out, rc);
1465
1466         data->ocd_instance = obd2obt(target)->obt_instance;
1467
1468         /*
1469          * Return only the parts of obd_connect_data that we understand, so the
1470          * client knows that we don't understand the rest.
1471          */
1472         if (data) {
1473                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1474                                                RCL_SERVER);
1475                 tmpdata = req_capsule_server_get(&req->rq_pill,
1476                                                  &RMF_CONNECT_DATA);
1477                 /*
1478                  * Don't use struct assignment here, because the client reply
1479                  * buffer may be smaller/larger than the local struct
1480                  * obd_connect_data.
1481                  */
1482                 memcpy(tmpdata, data, min(tmpsize, size));
1483         }
1484
1485         /*
1486          * If the client and the server are the same node, we will already
1487          * have an export that really points to the client's DLM export,
1488          * because we have a shared handles table.
1489          *
1490          * XXX this will go away when shaver stops sending the "connect" handle
1491          * in the real "remote handle" field of the request --phik 24 Apr 2003
1492          */
1493         ptlrpc_request_change_export(req, export);
1494
1495         pcon = ptlrpc_connection_get(&req->rq_peer,
1496                                      &req->rq_self, &cluuid);
1497         if (pcon == NULL)
1498                 GOTO(out, rc = -ENOTCONN);
1499
1500         spin_lock(&export->exp_lock);
1501
1502         if (export->exp_disconnected) {
1503                 spin_unlock(&export->exp_lock);
1504                 if (reconnected) {
1505                         /*
1506                          * for each connect called disconnect
1507                          * should be called to cleanup stuff
1508                          */
1509                         class_export_get(export);
1510                         obd_disconnect(export);
1511                 }
1512
1513                 GOTO(out, rc = -ENODEV);
1514         }
1515         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1516                 spin_unlock(&export->exp_lock);
1517                 CDEBUG(D_RPCTRACE,
1518                        "%s: %s already connected at greater or equal conn_cnt: %d >= %d\n",
1519                        cluuid.uuid, libcfs_nidstr(&req->rq_peer.nid),
1520                        export->exp_conn_cnt,
1521                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1522
1523                 GOTO(out, rc = -EALREADY);
1524         }
1525         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1526         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1527
1528 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1529         /* make 'tmp_exp_old_falloc' persistent by saving it into
1530          * server side export object(obd_export)
1531          */
1532         export->exp_old_falloc = tmp_exp_old_falloc;
1533 #endif
1534
1535         /* Check to see if connection came from another NID. */
1536         if (export->exp_connection != NULL &&
1537             !nid_same(&export->exp_connection->c_peer.nid,
1538                       &req->rq_peer.nid)) {
1539                 obd_nid_del(export->exp_obd, export);
1540                 ptlrpc_connection_put(export->exp_connection);
1541                 export->exp_connection = NULL;
1542         }
1543
1544         if (export->exp_connection == NULL) {
1545                 export->exp_connection = pcon;
1546                 pcon = NULL;
1547         }
1548         obd_nid_add(export->exp_obd, export);
1549
1550         spin_unlock(&export->exp_lock);
1551
1552         lustre_msg_set_handle(req->rq_repmsg, &conn);
1553
1554         rc = rev_import_reconnect(export, req);
1555         if (rc != 0)
1556                 GOTO(out, rc);
1557
1558         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1559                 int has_transno;
1560                 __u64 transno = data->ocd_transno;
1561
1562                 spin_lock(&export->exp_lock);
1563                 /*
1564                  * possible race with class_disconnect_stale_exports,
1565                  * export may be already in the eviction process
1566                  */
1567                 if (export->exp_failed) {
1568                         spin_unlock(&export->exp_lock);
1569                         GOTO(out, rc = -ENODEV);
1570                 }
1571                 export->exp_in_recovery = 1;
1572                 export->exp_req_replay_needed = 1;
1573                 export->exp_lock_replay_needed = 1;
1574                 spin_unlock(&export->exp_lock);
1575
1576                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1577                                  MSG_CONNECT_TRANSNO);
1578                 if (has_transno && transno == 0)
1579                         CWARN("Connect with zero transno!\n");
1580
1581                 if (has_transno && transno > 0 &&
1582                     transno < target->obd_next_recovery_transno &&
1583                     transno > target->obd_last_committed) {
1584                         /* Another way is to use cmpxchg() to be lock-free. */
1585                         spin_lock(&target->obd_recovery_task_lock);
1586                         if (transno < target->obd_next_recovery_transno)
1587                                 target->obd_next_recovery_transno = transno;
1588                         spin_unlock(&target->obd_recovery_task_lock);
1589                 }
1590
1591                 atomic_inc(&target->obd_req_replay_clients);
1592                 atomic_inc(&target->obd_lock_replay_clients);
1593                 /*
1594                  * Note: MDS-MDS connection is allowed to be connected during
1595                  * recovery, no matter if the exports needs to be recoveried.
1596                  * Because we need retrieve updates logs from all other MDTs.
1597                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1598                  * also needs to be increased to match other recovery checking
1599                  * condition.
1600                  */
1601                 if (new_mds_mds_conn)
1602                         atomic_inc(&target->obd_max_recoverable_clients);
1603
1604                 if (atomic_inc_return(&target->obd_connected_clients) ==
1605                     atomic_read(&target->obd_max_recoverable_clients))
1606                         wake_up(&target->obd_next_transno_waitq);
1607         }
1608
1609         /* Tell the client we're in recovery, when client is involved in it. */
1610         if (target->obd_recovering && !lw_client)
1611                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1612
1613 out:
1614         if (export) {
1615                 spin_lock(&export->exp_lock);
1616                 export->exp_connecting = 0;
1617                 spin_unlock(&export->exp_lock);
1618
1619                 class_export_put(export);
1620         }
1621         if (target != NULL) {
1622                 spin_lock(&target->obd_dev_lock);
1623                 target->obd_conn_inprogress--;
1624                 spin_unlock(&target->obd_dev_lock);
1625                 class_decref(target, "find", current);
1626         }
1627         if (pcon)
1628                 ptlrpc_connection_put(pcon);
1629         req->rq_status = rc;
1630         RETURN(rc);
1631 }
1632
1633 int target_handle_disconnect(struct ptlrpc_request *req)
1634 {
1635         int rc;
1636
1637         ENTRY;
1638
1639         rc = req_capsule_server_pack(&req->rq_pill);
1640         if (rc)
1641                 RETURN(rc);
1642
1643         /* In case of target disconnect, updating sec ctx immediately is
1644          * required in order to record latest sequence number used.
1645          * Sequence is normally updated on export destroy, but this event
1646          * can occur too late, ie after a new target connect request has
1647          * been processed.
1648          * Maintaining correct sequence when client connection becomes idle
1649          * ensures that GSS does not erroneously consider requests as replays.
1650          */
1651         rc = sptlrpc_export_update_ctx(req->rq_export);
1652         if (rc)
1653                 RETURN(rc);
1654
1655         /* Keep the rq_export around so we can send the reply. */
1656         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1657
1658         RETURN(0);
1659 }
1660
1661 void target_destroy_export(struct obd_export *exp)
1662 {
1663         struct obd_import *imp = NULL;
1664         /*
1665          * exports created from last_rcvd data, and "fake"
1666          * exports created by lctl don't have an import
1667          */
1668         spin_lock(&exp->exp_lock);
1669         if (exp->exp_imp_reverse != NULL) {
1670                 imp = exp->exp_imp_reverse;
1671                 exp->exp_imp_reverse = NULL;
1672         }
1673         spin_unlock(&exp->exp_lock);
1674         if (imp != NULL)
1675                 client_destroy_import(imp);
1676
1677         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1678         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1679         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1680         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1681 }
1682 EXPORT_SYMBOL(target_destroy_export);
1683
1684 /*
1685  * Recovery functions
1686  */
1687 static void target_request_copy_get(struct ptlrpc_request *req)
1688 {
1689         class_export_rpc_inc(req->rq_export);
1690         LASSERT(list_empty(&req->rq_list));
1691         INIT_LIST_HEAD(&req->rq_replay_list);
1692
1693         /* Increase refcount to keep request in queue. */
1694         atomic_inc(&req->rq_refcount);
1695         /* Let export know it has replays to be handled. */
1696         atomic_inc(&req->rq_export->exp_replay_count);
1697 }
1698
1699 static void target_request_copy_put(struct ptlrpc_request *req)
1700 {
1701         LASSERT(list_empty(&req->rq_replay_list));
1702         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1703
1704         atomic_dec(&req->rq_export->exp_replay_count);
1705         class_export_rpc_dec(req->rq_export);
1706         ptlrpc_server_drop_request(req);
1707 }
1708
1709 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1710 {
1711         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1712         struct obd_export *exp = req->rq_export;
1713         struct ptlrpc_request *reqiter;
1714         struct ptlrpc_request *dup_req = NULL;
1715         int dup = 0;
1716
1717         LASSERT(exp);
1718
1719         spin_lock(&exp->exp_lock);
1720         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1721                             rq_replay_list) {
1722                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1723                         dup_req = reqiter;
1724                         dup = 1;
1725                         break;
1726                 }
1727         }
1728
1729         if (dup) {
1730                 /* We expect it with RESENT and REPLAY flags. */
1731                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1732                     (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1733                         CERROR("invalid flags %x of resent replay\n",
1734                                lustre_msg_get_flags(req->rq_reqmsg));
1735
1736                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1737                         __u32 new_conn;
1738
1739                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1740                         if (new_conn >
1741                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1742                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1743                                                         new_conn);
1744                 }
1745         } else {
1746                 list_add_tail(&req->rq_replay_list,
1747                               &exp->exp_req_replay_queue);
1748         }
1749
1750         spin_unlock(&exp->exp_lock);
1751         return dup;
1752 }
1753
1754 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1755 {
1756         LASSERT(!list_empty(&req->rq_replay_list));
1757         LASSERT(req->rq_export);
1758
1759         spin_lock(&req->rq_export->exp_lock);
1760         list_del_init(&req->rq_replay_list);
1761         spin_unlock(&req->rq_export->exp_lock);
1762 }
1763
1764 static void target_finish_recovery(struct lu_target *lut)
1765 {
1766         struct obd_device *obd = lut->lut_obd;
1767
1768         ENTRY;
1769
1770         /* Only log a recovery message when recovery has occurred. */
1771         if (obd->obd_recovery_start) {
1772                 time64_t now = ktime_get_seconds();
1773                 time64_t elapsed_time;
1774
1775                 elapsed_time = max_t(time64_t, now - obd->obd_recovery_start,
1776                                      1);
1777                 LCONSOLE_INFO("%s: Recovery over after %lld:%.02lld, of %d clients %d recovered and %d %s evicted.\n",
1778                               obd->obd_name, elapsed_time / 60,
1779                               elapsed_time % 60,
1780                               atomic_read(&obd->obd_max_recoverable_clients),
1781                               atomic_read(&obd->obd_connected_clients),
1782                               obd->obd_stale_clients,
1783                               obd->obd_stale_clients == 1 ? "was" : "were");
1784                 if (obd->obd_stale_clients && do_dump_on_eviction(obd))
1785                         libcfs_debug_dumplog();
1786         }
1787
1788         ldlm_reprocess_recovery_done(obd->obd_namespace);
1789         spin_lock(&obd->obd_recovery_task_lock);
1790         if (!list_empty(&obd->obd_req_replay_queue) ||
1791             !list_empty(&obd->obd_lock_replay_queue) ||
1792             !list_empty(&obd->obd_final_req_queue)) {
1793                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1794                        obd->obd_name,
1795                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1796                        list_empty(&obd->obd_lock_replay_queue) ? \
1797                                   "" : "lock ",
1798                        list_empty(&obd->obd_final_req_queue) ? \
1799                                   "" : "final ");
1800                 spin_unlock(&obd->obd_recovery_task_lock);
1801                 LBUG();
1802         }
1803         spin_unlock(&obd->obd_recovery_task_lock);
1804
1805         obd->obd_recovery_end = ktime_get_seconds();
1806
1807         /* When recovery finished, cleanup orphans on MDS and OST. */
1808         if (obd->obd_type && OBP(obd, postrecov)) {
1809                 int rc = OBP(obd, postrecov)(obd);
1810
1811                 if (rc < 0)
1812                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1813                                       obd->obd_name, rc);
1814         }
1815         EXIT;
1816 }
1817
1818 static void abort_req_replay_queue(struct obd_device *obd)
1819 {
1820         struct ptlrpc_request *req, *n;
1821         LIST_HEAD(abort_list);
1822
1823         spin_lock(&obd->obd_recovery_task_lock);
1824         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1825         spin_unlock(&obd->obd_recovery_task_lock);
1826         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1827                 DEBUG_REQ(D_WARNING, req, "aborted:");
1828                 req->rq_status = -ENOTCONN;
1829                 if (ptlrpc_error(req)) {
1830                         DEBUG_REQ(D_ERROR, req,
1831                                   "failed abort_req_reply; skipping");
1832                 }
1833                 target_exp_dequeue_req_replay(req);
1834                 target_request_copy_put(req);
1835         }
1836 }
1837
1838 static void abort_lock_replay_queue(struct obd_device *obd)
1839 {
1840         struct ptlrpc_request *req, *n;
1841         LIST_HEAD(abort_list);
1842
1843         spin_lock(&obd->obd_recovery_task_lock);
1844         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1845         spin_unlock(&obd->obd_recovery_task_lock);
1846         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1847                 DEBUG_REQ(D_ERROR, req, "aborted:");
1848                 req->rq_status = -ENOTCONN;
1849                 if (ptlrpc_error(req)) {
1850                         DEBUG_REQ(D_ERROR, req,
1851                                   "failed abort_lock_reply; skipping");
1852                 }
1853                 target_request_copy_put(req);
1854         }
1855 }
1856
1857 /*
1858  * Called from a cleanup function if the device is being cleaned up
1859  * forcefully.  The exports should all have been disconnected already,
1860  * the only thing left to do is
1861  * - clear the recovery flags
1862  * - cancel the timer
1863  * - free queued requests and replies, but don't send replies
1864  * Because the obd_stopping flag is set, no new requests should be received.
1865  */
1866 void target_cleanup_recovery(struct obd_device *obd)
1867 {
1868         struct ptlrpc_request *req, *n;
1869         LIST_HEAD(clean_list);
1870
1871         spin_lock(&obd->obd_dev_lock);
1872         if (!obd->obd_recovering) {
1873                 spin_unlock(&obd->obd_dev_lock);
1874                 EXIT;
1875                 return;
1876         }
1877         obd->obd_recovering = 0;
1878         obd->obd_abort_recovery = 0;
1879         obd->obd_abort_mdt_recovery = 0;
1880         spin_unlock(&obd->obd_dev_lock);
1881
1882         spin_lock(&obd->obd_recovery_task_lock);
1883         target_cancel_recovery_timer(obd);
1884         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1885         spin_unlock(&obd->obd_recovery_task_lock);
1886
1887         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1888                 LASSERT(req->rq_reply_state == NULL);
1889                 target_exp_dequeue_req_replay(req);
1890                 target_request_copy_put(req);
1891         }
1892
1893         spin_lock(&obd->obd_recovery_task_lock);
1894         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1895         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1896         spin_unlock(&obd->obd_recovery_task_lock);
1897
1898         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1899                 LASSERT(req->rq_reply_state == NULL);
1900                 target_request_copy_put(req);
1901         }
1902
1903         EXIT;
1904 }
1905 EXPORT_SYMBOL(target_cleanup_recovery);
1906
1907 /* obd_recovery_task_lock should be held */
1908 void target_cancel_recovery_timer(struct obd_device *obd)
1909 {
1910         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1911         hrtimer_cancel(&obd->obd_recovery_timer);
1912 }
1913
1914 static void target_start_recovery_timer(struct obd_device *obd)
1915 {
1916         ktime_t delay;
1917
1918         if (obd->obd_recovery_start != 0)
1919                 return;
1920
1921         spin_lock(&obd->obd_dev_lock);
1922         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1923                 spin_unlock(&obd->obd_dev_lock);
1924                 return;
1925         }
1926
1927         LASSERT(obd->obd_recovery_timeout != 0);
1928
1929         if (obd->obd_recovery_start != 0) {
1930                 spin_unlock(&obd->obd_dev_lock);
1931                 return;
1932         }
1933
1934         obd->obd_recovery_start = ktime_get_seconds();
1935         delay = ktime_set(obd->obd_recovery_start +
1936                           obd->obd_recovery_timeout, 0);
1937         hrtimer_start(&obd->obd_recovery_timer, delay, HRTIMER_MODE_ABS);
1938         spin_unlock(&obd->obd_dev_lock);
1939
1940         LCONSOLE_WARN("%s: Will be in recovery for at least %u:%02u, or until %d client%s reconnect%s\n",
1941                       obd->obd_name,
1942                       obd->obd_recovery_timeout / 60,
1943                       obd->obd_recovery_timeout % 60,
1944                       atomic_read(&obd->obd_max_recoverable_clients),
1945                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1946                       "" : "s",
1947                       (atomic_read(&obd->obd_max_recoverable_clients) == 1) ?
1948                       "s" : "");
1949 }
1950
1951 /**
1952  * extend recovery window.
1953  *
1954  * if @extend is true, extend recovery window to have @dr_timeout remaining
1955  * at least; otherwise, make sure the recovery timeout value is not less
1956  * than @dr_timeout.
1957  */
1958 static void extend_recovery_timer(struct obd_device *obd, timeout_t dr_timeout,
1959                                   bool extend)
1960 {
1961         ktime_t left_ns;
1962         timeout_t timeout;
1963         timeout_t left;
1964
1965         spin_lock(&obd->obd_dev_lock);
1966         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1967             obd->obd_stopping) {
1968                 spin_unlock(&obd->obd_dev_lock);
1969                 return;
1970         }
1971         LASSERT(obd->obd_recovery_start != 0);
1972
1973         left_ns = hrtimer_get_remaining(&obd->obd_recovery_timer);
1974         left = ktime_divns(left_ns, NSEC_PER_SEC);
1975
1976         if (extend) {
1977                 timeout = obd->obd_recovery_timeout;
1978                 /* dr_timeout will happen after the hrtimer has expired.
1979                  * Add the excess time to the soft recovery timeout without
1980                  * exceeding the hard recovery timeout.
1981                  */
1982                 if (dr_timeout > left) {
1983                         timeout += dr_timeout - left;
1984                         timeout = min_t(timeout_t, obd->obd_recovery_time_hard,
1985                                         timeout);
1986                 }
1987         } else {
1988                 timeout = clamp_t(timeout_t, dr_timeout,
1989                                   obd->obd_recovery_timeout,
1990                                   obd->obd_recovery_time_hard);
1991         }
1992
1993         if (timeout == obd->obd_recovery_time_hard)
1994                 CWARN("%s: extended recovery timer reached hard limit: %d, extend: %d\n",
1995                       obd->obd_name, timeout, extend);
1996
1997         if (obd->obd_recovery_timeout < timeout) {
1998                 ktime_t end, now;
1999
2000                 obd->obd_recovery_timeout = timeout;
2001                 end = ktime_set(obd->obd_recovery_start + timeout, 0);
2002                 now = ktime_set(ktime_get_seconds(), 0);
2003                 left_ns = ktime_sub(end, now);
2004                 hrtimer_start(&obd->obd_recovery_timer, end, HRTIMER_MODE_ABS);
2005                 left = ktime_divns(left_ns, NSEC_PER_SEC);
2006         }
2007         spin_unlock(&obd->obd_dev_lock);
2008
2009         CDEBUG(D_HA, "%s: recovery timer will expire in %d seconds\n",
2010                 obd->obd_name, left);
2011 }
2012
2013 /* Reset the timer with each new client connection */
2014 /*
2015  * This timer is actually reconnect_timer, which is for making sure
2016  * the total recovery window is at least as big as my reconnect
2017  * attempt timing. So the initial recovery time_out will be set to
2018  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
2019  * from client is bigger than this, then the recovery time_out will
2020  * be extended to make sure the client could be reconnected, in the
2021  * process, the timeout from the new client should be ignored.
2022  */
2023 static void
2024 check_and_start_recovery_timer(struct obd_device *obd,
2025                                struct ptlrpc_request *req,
2026                                int new_client)
2027 {
2028         timeout_t service_timeout = lustre_msg_get_service_timeout(req->rq_reqmsg);
2029         struct obd_device_target *obt = obd2obt(obd);
2030
2031         if (!new_client && service_timeout)
2032                 /*
2033                  * Teach server about old server's estimates, as first guess
2034                  * at how long new requests will take.
2035                  */
2036                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
2037                             service_timeout);
2038
2039         target_start_recovery_timer(obd);
2040
2041         /*
2042          * Convert the service time to RPC timeout,
2043          * and reuse service_timeout to limit stack usage.
2044          */
2045         service_timeout = at_est2timeout(service_timeout);
2046
2047         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_SLUGGISH_NET) &&
2048             service_timeout < at_extra)
2049                 service_timeout = at_extra;
2050
2051         /*
2052          * We expect other clients to timeout within service_timeout, then try
2053          * to reconnect, then try the failover server.  The max delay between
2054          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL.
2055          */
2056         service_timeout += 2 * INITIAL_CONNECT_TIMEOUT;
2057
2058         LASSERT(obt->obt_magic == OBT_MAGIC);
2059         service_timeout += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
2060         if (service_timeout > obd->obd_recovery_timeout && !new_client)
2061                 extend_recovery_timer(obd, service_timeout, false);
2062 }
2063
2064 /** Health checking routines */
2065 static inline int exp_connect_healthy(struct obd_export *exp)
2066 {
2067         return exp->exp_in_recovery;
2068 }
2069
2070 /** if export done req_replay or has replay in queue */
2071 static inline int exp_req_replay_healthy(struct obd_export *exp)
2072 {
2073         return (!exp->exp_req_replay_needed ||
2074                 atomic_read(&exp->exp_replay_count) > 0);
2075 }
2076
2077
2078 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
2079 {
2080         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2081                exp_req_replay_healthy(exp);
2082 }
2083
2084 /** if export done lock_replay or has replay in queue */
2085 static inline int exp_lock_replay_healthy(struct obd_export *exp)
2086 {
2087         return (!exp->exp_lock_replay_needed ||
2088                 atomic_read(&exp->exp_replay_count) > 0);
2089 }
2090
2091 static inline int exp_vbr_healthy(struct obd_export *exp)
2092 {
2093         return !exp->exp_vbr_failed;
2094 }
2095
2096 static inline int exp_finished(struct obd_export *exp)
2097 {
2098         return exp->exp_in_recovery && !exp->exp_lock_replay_needed;
2099 }
2100
2101 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
2102 {
2103         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
2104                 exp_finished(exp);
2105 }
2106
2107 static int check_for_next_transno(struct lu_target *lut)
2108 {
2109         struct ptlrpc_request *req = NULL;
2110         struct obd_device *obd = lut->lut_obd;
2111         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2112         int wake_up = 0, connected, completed, queue_len;
2113         __u64 req_transno = 0;
2114         __u64 update_transno = 0;
2115         __u64 next_transno = 0;
2116
2117         ENTRY;
2118
2119         spin_lock(&obd->obd_recovery_task_lock);
2120         if (!list_empty(&obd->obd_req_replay_queue)) {
2121                 req = list_entry(obd->obd_req_replay_queue.next,
2122                                      struct ptlrpc_request, rq_list);
2123                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
2124         }
2125
2126         if (!obd_mdt_recovery_abort(obd) && tdtd)
2127                 update_transno = distribute_txn_get_next_transno(tdtd);
2128
2129         connected = atomic_read(&obd->obd_connected_clients);
2130         completed = connected - atomic_read(&obd->obd_req_replay_clients);
2131         queue_len = obd->obd_requests_queued_for_recovery;
2132         next_transno = obd->obd_next_recovery_transno;
2133
2134         CDEBUG(D_HA,
2135                "max: %d, connected: %d, completed: %d, queue_len: %d, req_transno: %llu, next_transno: %llu\n",
2136                atomic_read(&obd->obd_max_recoverable_clients),
2137                connected, completed,
2138                queue_len, req_transno, next_transno);
2139
2140         if (obd_recovery_abort(obd)) {
2141                 CDEBUG(D_HA, "waking for aborted recovery\n");
2142                 wake_up = 1;
2143         } else if (obd->obd_recovery_expired) {
2144                 CDEBUG(D_HA, "waking for expired recovery\n");
2145                 wake_up = 1;
2146         } else if (!obd_mdt_recovery_abort(obd) && tdtd && req &&
2147                    is_req_replayed_by_update(req)) {
2148                 LASSERTF(req_transno < next_transno,
2149                          "req_transno %llu next_transno%llu\n", req_transno,
2150                          next_transno);
2151                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
2152                        req_transno);
2153                 wake_up = 1;
2154         } else if (req_transno == next_transno ||
2155                    (update_transno != 0 && update_transno <= next_transno)) {
2156                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
2157                 wake_up = 1;
2158         } else if (queue_len > 0 &&
2159                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
2160                 /** handle gaps occured due to lost reply or VBR */
2161                 LASSERTF(req_transno >= next_transno,
2162                          "req_transno: %llu, next_transno: %llu\n",
2163                          req_transno, next_transno);
2164                 CDEBUG(D_HA,
2165                        "%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",
2166                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
2167                        next_transno, queue_len, completed, connected,
2168                        req_transno, update_transno, obd->obd_last_committed);
2169                 obd->obd_next_recovery_transno = req_transno;
2170                 wake_up = 1;
2171         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
2172                 CDEBUG(D_HA, "waking for completed recovery\n");
2173                 wake_up = 1;
2174         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
2175                 CDEBUG(D_HA,
2176                        "accepting transno gaps is explicitly allowed by fail_lock, waking up (%lld)\n",
2177                        next_transno);
2178                 obd->obd_next_recovery_transno = req_transno;
2179                 wake_up = 1;
2180         }
2181         spin_unlock(&obd->obd_recovery_task_lock);
2182         return wake_up;
2183 }
2184
2185 static int check_for_next_lock(struct lu_target *lut)
2186 {
2187         struct obd_device *obd = lut->lut_obd;
2188         int wake_up = 0;
2189
2190         spin_lock(&obd->obd_recovery_task_lock);
2191         if (!list_empty(&obd->obd_lock_replay_queue)) {
2192                 CDEBUG(D_HA, "waking for next lock\n");
2193                 wake_up = 1;
2194         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
2195                 CDEBUG(D_HA, "waking for completed lock replay\n");
2196                 wake_up = 1;
2197         } else if (obd->obd_abort_recovery) {
2198                 CDEBUG(D_HA, "waking for aborted recovery\n");
2199                 wake_up = 1;
2200         } else if (obd->obd_recovery_expired) {
2201                 CDEBUG(D_HA, "waking for expired recovery\n");
2202                 wake_up = 1;
2203         }
2204         spin_unlock(&obd->obd_recovery_task_lock);
2205
2206         return wake_up;
2207 }
2208
2209 static int check_update_llog(struct lu_target *lut)
2210 {
2211         struct obd_device *obd = lut->lut_obd;
2212         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2213
2214         if (obd_mdt_recovery_abort(obd)) {
2215                 CDEBUG(D_HA, "waking for aborted recovery\n");
2216                 return 1;
2217         }
2218
2219         if (atomic_read(&tdtd->tdtd_recovery_threads_count) == 0) {
2220                 CDEBUG(D_HA, "waking for completion of reading update log\n");
2221                 return 1;
2222         }
2223
2224         return 0;
2225 }
2226
2227 /**
2228  * wait for recovery events,
2229  * check its status with help of check_routine
2230  * evict dead clients via health_check
2231  */
2232 static int target_recovery_overseer(struct lu_target *lut,
2233                                     int (*check_routine)(struct lu_target *),
2234                                     int (*health_check)(struct obd_export *))
2235 {
2236         struct obd_device *obd = lut->lut_obd;
2237         struct target_distribute_txn_data *tdtd;
2238         time64_t last = 0;
2239         time64_t now;
2240 repeat:
2241         if (obd->obd_recovering && obd->obd_recovery_start == 0) {
2242                 now = ktime_get_seconds();
2243                 if (now - last > 600) {
2244                         LCONSOLE_INFO("%s: in recovery but waiting for the first client to connect\n",
2245                                       obd->obd_name);
2246                         last = now;
2247                 }
2248         }
2249         if (obd->obd_recovery_start != 0 && ktime_get_seconds() >=
2250               (obd->obd_recovery_start + obd->obd_recovery_time_hard)) {
2251                 __u64 next_update_transno = 0;
2252
2253                 /*
2254                  * Only abort the recovery if there are no update recovery
2255                  * left in the queue
2256                  */
2257                 spin_lock(&obd->obd_recovery_task_lock);
2258                 if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd) {
2259                         next_update_transno =
2260                                 distribute_txn_get_next_transno(lut->lut_tdtd);
2261
2262                         tdtd = lut->lut_tdtd;
2263                         /*
2264                          * If next_update_transno == 0, it probably because
2265                          * updatelog retrieve threads did not get any records
2266                          * yet, let's wait those threads stopped
2267                          */
2268                         if (next_update_transno == 0) {
2269                                 spin_unlock(&obd->obd_recovery_task_lock);
2270
2271                                 while (wait_event_timeout(
2272                                         tdtd->tdtd_recovery_threads_waitq,
2273                                         check_update_llog(lut),
2274                                         cfs_time_seconds(60)) == 0);
2275
2276                                 spin_lock(&obd->obd_recovery_task_lock);
2277                                 next_update_transno =
2278                                         distribute_txn_get_next_transno(tdtd);
2279                         }
2280                 }
2281
2282                 if (next_update_transno != 0 && !obd_recovery_abort(obd)) {
2283                         obd->obd_next_recovery_transno = next_update_transno;
2284                         spin_unlock(&obd->obd_recovery_task_lock);
2285                         /*
2286                          * Disconnect unfinished exports from clients, and
2287                          * keep connection from MDT to make sure the update
2288                          * recovery will still keep trying until some one
2289                          * manually abort the recovery
2290                          */
2291                         class_disconnect_stale_exports(obd,
2292                                                 exp_finished_or_from_mdt);
2293                         /* Abort all of replay & replay lock req from clients */
2294                         abort_req_replay_queue(obd);
2295                         abort_lock_replay_queue(obd);
2296                         CDEBUG(D_HA,
2297                                "%s: there are still update replay (%#llx)in the queue.\n",
2298                                obd->obd_name, next_update_transno);
2299                 } else {
2300                         obd->obd_abort_recovery = 1;
2301                         spin_unlock(&obd->obd_recovery_task_lock);
2302                         CWARN("%s recovery is aborted by hard timeout\n",
2303                               obd->obd_name);
2304                 }
2305         }
2306
2307         while (wait_event_timeout(obd->obd_next_transno_waitq,
2308                                   check_routine(lut),
2309                                   cfs_time_seconds(60)) == 0)
2310                 ; /* wait indefinitely for event, but don't trigger watchdog */
2311
2312         if (obd_recovery_abort(obd)) {
2313                 CWARN("recovery is aborted, evict exports in recovery\n");
2314                 if (lut->lut_tdtd != NULL) {
2315                         tdtd = lut->lut_tdtd;
2316                         /*
2317                          * Let's wait all of the update log recovery thread
2318                          * finished
2319                          */
2320                         wait_event_idle(
2321                                 tdtd->tdtd_recovery_threads_waitq,
2322                                 atomic_read(&tdtd->tdtd_recovery_threads_count)
2323                                 == 0);
2324                         /* Then abort the update recovery list */
2325                         dtrq_list_destroy(lut->lut_tdtd);
2326                 }
2327
2328                 /** evict exports which didn't finish recovery yet */
2329                 class_disconnect_stale_exports(obd, exp_finished);
2330                 return 1;
2331         } else if (obd->obd_recovery_expired) {
2332                 obd->obd_recovery_expired = 0;
2333
2334                 /** If some clients died being recovered, evict them */
2335                 LCONSOLE_WARN("%s: recovery is timed out, evict stale exports\n",
2336                               obd->obd_name);
2337                 /** evict cexports with no replay in queue, they are stalled */
2338                 class_disconnect_stale_exports(obd, health_check);
2339
2340                 /** continue with VBR */
2341                 spin_lock(&obd->obd_dev_lock);
2342                 obd->obd_version_recov = 1;
2343                 spin_unlock(&obd->obd_dev_lock);
2344                 /**
2345                  * reset timer, recovery will proceed with versions now,
2346                  * timeout is set just to handle reconnection delays
2347                  */
2348                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2349                 /**
2350                  * Wait for recovery events again, after evicting bad clients
2351                  */
2352                 goto repeat;
2353         }
2354         return 0;
2355 }
2356
2357 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2358 {
2359         struct obd_device *obd = lut->lut_obd;
2360         struct ptlrpc_request *req = NULL;
2361
2362         CDEBUG(D_HA, "Waiting for lock\n");
2363         if (target_recovery_overseer(lut, check_for_next_lock,
2364                                      exp_lock_replay_healthy))
2365                 abort_lock_replay_queue(obd);
2366
2367         spin_lock(&obd->obd_recovery_task_lock);
2368         if (!list_empty(&obd->obd_lock_replay_queue)) {
2369                 req = list_entry(obd->obd_lock_replay_queue.next,
2370                                      struct ptlrpc_request, rq_list);
2371                 list_del_init(&req->rq_list);
2372                 spin_unlock(&obd->obd_recovery_task_lock);
2373         } else {
2374                 spin_unlock(&obd->obd_recovery_task_lock);
2375                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2376                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2377                 /** evict exports failed VBR */
2378                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2379         }
2380         return req;
2381 }
2382
2383 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2384 {
2385         struct ptlrpc_request *req = NULL;
2386
2387         spin_lock(&obd->obd_recovery_task_lock);
2388         if (!list_empty(&obd->obd_final_req_queue)) {
2389                 req = list_entry(obd->obd_final_req_queue.next,
2390                                      struct ptlrpc_request, rq_list);
2391                 list_del_init(&req->rq_list);
2392                 spin_unlock(&obd->obd_recovery_task_lock);
2393                 if (req->rq_export->exp_in_recovery) {
2394                         spin_lock(&req->rq_export->exp_lock);
2395                         req->rq_export->exp_in_recovery = 0;
2396                         spin_unlock(&req->rq_export->exp_lock);
2397                 }
2398         } else {
2399                 spin_unlock(&obd->obd_recovery_task_lock);
2400         }
2401         return req;
2402 }
2403
2404 static void handle_recovery_req(struct ptlrpc_thread *thread,
2405                                 struct ptlrpc_request *req,
2406                                 svc_handler_t handler)
2407 {
2408         ENTRY;
2409
2410         /**
2411          * export can be evicted during recovery, no need to handle replays for
2412          * it after that, discard such request silently
2413          */
2414         if (req->rq_export->exp_disconnected)
2415                 RETURN_EXIT;
2416
2417         req->rq_session.lc_thread = thread;
2418         req->rq_svc_thread = thread;
2419         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2420
2421         /* thread context */
2422         lu_context_enter(&thread->t_env->le_ctx);
2423         (void)handler(req);
2424         lu_context_exit(&thread->t_env->le_ctx);
2425
2426         req->rq_svc_thread->t_env->le_ses = NULL;
2427
2428         /* don't reset timer for final stage */
2429         if (!exp_finished(req->rq_export)) {
2430                 timeout_t timeout = obd_timeout;
2431
2432                 /**
2433                  * Add request @timeout to the recovery time so next request from
2434                  * this client may come in recovery time
2435                  */
2436                 if (!AT_OFF) {
2437                         struct ptlrpc_service_part *svcpt;
2438                         timeout_t est_timeout;
2439
2440                         svcpt = req->rq_rqbd->rqbd_svcpt;
2441                         /*
2442                          * If the server sent early reply for this request,
2443                          * the client will recalculate the timeout according to
2444                          * current server estimate service time, so we will
2445                          * use the maxium timeout here for waiting the client
2446                          * sending the next req
2447                          */
2448                         est_timeout = at_get(&svcpt->scp_at_estimate);
2449                         timeout = max_t(timeout_t, at_est2timeout(est_timeout),
2450                                         lustre_msg_get_timeout(req->rq_reqmsg));
2451                         /*
2452                          * Add 2 net_latency, one for balance rq_deadline
2453                          * (see ptl_send_rpc), one for resend the req to server,
2454                          * Note: client will pack net_latency in replay req
2455                          * (see ptlrpc_replay_req)
2456                          */
2457                         timeout += 2 * lustre_msg_get_service_timeout(req->rq_reqmsg);
2458                 }
2459                 extend_recovery_timer(class_exp2obd(req->rq_export), timeout,
2460                                       true);
2461         }
2462         EXIT;
2463 }
2464
2465 /** Checking routines for recovery */
2466 static int check_for_recovery_ready(struct lu_target *lut)
2467 {
2468         struct obd_device *obd = lut->lut_obd;
2469         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2470
2471         CDEBUG(D_HA,
2472                "connected %d stale %d max_recoverable_clients %d abort %d expired %d\n",
2473                clnts, obd->obd_stale_clients,
2474                atomic_read(&obd->obd_max_recoverable_clients),
2475                obd->obd_abort_recovery, obd->obd_recovery_expired);
2476
2477         if (!obd_recovery_abort(obd) && !obd->obd_recovery_expired) {
2478                 LASSERT(clnts <=
2479                         atomic_read(&obd->obd_max_recoverable_clients));
2480                 if (clnts + obd->obd_stale_clients <
2481                     atomic_read(&obd->obd_max_recoverable_clients))
2482                         return 0;
2483         }
2484
2485         if (!obd_mdt_recovery_abort(obd) && lut->lut_tdtd &&
2486             !lut->lut_tdtd->tdtd_replay_ready) {
2487                 /* Let's extend recovery timer, in case the recovery timer
2488                  * expired, and some clients got evicted
2489                  */
2490                 extend_recovery_timer(obd, obd->obd_recovery_timeout, true);
2491                 CDEBUG(D_HA,
2492                        "%s update recovery is not ready, extend recovery %d\n",
2493                        obd->obd_name, obd->obd_recovery_timeout);
2494                 return 0;
2495         }
2496
2497         return 1;
2498 }
2499
2500 enum {
2501         REQUEST_RECOVERY = 1,
2502         UPDATE_RECOVERY = 2,
2503 };
2504
2505 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2506 {
2507         __u64 transno = 0;
2508
2509         if (!list_empty(&obd->obd_req_replay_queue)) {
2510                 struct ptlrpc_request *req;
2511
2512                 req = list_entry(obd->obd_req_replay_queue.next,
2513                                  struct ptlrpc_request, rq_list);
2514                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2515         }
2516
2517         return transno;
2518 }
2519
2520 static __u64 get_next_transno(struct lu_target *lut, int *type)
2521 {
2522         struct obd_device *obd = lut->lut_obd;
2523         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2524         __u64 transno = 0;
2525         __u64 update_transno;
2526
2527         ENTRY;
2528
2529         transno = get_next_replay_req_transno(obd);
2530         if (type != NULL)
2531                 *type = REQUEST_RECOVERY;
2532
2533         if (!tdtd || obd_mdt_recovery_abort(obd))
2534                 RETURN(transno);
2535
2536         update_transno = distribute_txn_get_next_transno(tdtd);
2537         if (transno == 0 || (transno >= update_transno &&
2538                              update_transno != 0)) {
2539                 transno = update_transno;
2540                 if (type != NULL)
2541                         *type = UPDATE_RECOVERY;
2542         }
2543
2544         RETURN(transno);
2545 }
2546
2547 /**
2548  * drop duplicate replay request
2549  *
2550  * Because the operation has been replayed by update recovery, the request
2551  * with the same transno will be dropped and also notify the client to send
2552  * next replay request.
2553  *
2554  * \param[in] env       execution environment
2555  * \param[in] obd       failover obd device
2556  * \param[in] req       request to be dropped
2557  *
2558  * \retval true         duplicate replay update is dropped
2559  * \retval false        duplicate replay update is not dropped
2560  */
2561 static bool drop_duplicate_replay_req(struct lu_env *env,
2562                                       struct obd_device *obd,
2563                                       struct ptlrpc_request *req)
2564 {
2565         __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
2566
2567         DEBUG_REQ(D_HA, req,
2568                   "remove t%lld from %s because duplicate update records found",
2569                   lustre_msg_get_transno(req->rq_reqmsg),
2570                   libcfs_nidstr(&req->rq_peer.nid));
2571
2572         /*
2573          * Right now, only for MDS reint operation update replay and
2574          * normal request replay can have the same transno
2575          */
2576         if (opc == MDS_REINT) {
2577                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2578                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2579                 if (likely(req->rq_export))
2580                         target_committed_to_req(req);
2581                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2582                 target_send_reply(req, req->rq_status, 0);
2583         } else if (opc == MDS_CLOSE) {
2584                 DEBUG_REQ(D_HA, req, "duplicate close replay from %s\n",
2585                                 libcfs_nidstr(&req->rq_peer.nid));
2586                 return false;
2587         } else {
2588                 DEBUG_REQ(D_ERROR, req, "wrong opc %d from %s\n", opc,
2589                           libcfs_nidstr(&req->rq_peer.nid));
2590         }
2591         target_exp_dequeue_req_replay(req);
2592         target_request_copy_put(req);
2593         obd->obd_replayed_requests++;
2594
2595         return true;
2596 }
2597
2598 #define WATCHDOG_TIMEOUT (obd_timeout * 10)
2599
2600 static void replay_request_or_update(struct lu_env *env,
2601                                      struct lu_target *lut,
2602                                      struct target_recovery_data *trd,
2603                                      struct ptlrpc_thread *thread)
2604 {
2605         struct obd_device *obd = lut->lut_obd;
2606         struct ptlrpc_request *req = NULL;
2607         int type;
2608         __u64 transno;
2609
2610         ENTRY;
2611
2612         CDEBUG(D_HA, "Waiting for transno %lld\n",
2613                obd->obd_next_recovery_transno);
2614
2615         /* Replay all of request and update by transno */
2616         do {
2617                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2618
2619                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2620
2621                 /**
2622                  * It is needed to extend recovery window above
2623                  *  recovery_time_soft. Extending is possible only in the
2624                  *  end of recovery window (see more details in
2625                  *  handle_recovery_req()).
2626                  */
2627                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2628
2629                 if (target_recovery_overseer(lut, check_for_next_transno,
2630                                         exp_req_replay_healthy_or_from_mdt)) {
2631                         abort_req_replay_queue(obd);
2632                         abort_lock_replay_queue(obd);
2633                         goto abort;
2634                 }
2635
2636                 spin_lock(&obd->obd_recovery_task_lock);
2637                 transno = get_next_transno(lut, &type);
2638                 if (type == REQUEST_RECOVERY && transno != 0) {
2639                         bool update = false;
2640                         /*
2641                          * Drop replay request from client side, if the
2642                          * replay has been executed by update with the
2643                          * same transno
2644                          */
2645                         req = list_entry(obd->obd_req_replay_queue.next,
2646                                         struct ptlrpc_request, rq_list);
2647
2648                         list_del_init(&req->rq_list);
2649                         obd->obd_requests_queued_for_recovery--;
2650                         spin_unlock(&obd->obd_recovery_task_lock);
2651
2652                         /*
2653                          * Let's check if the request has been redone by
2654                          * update replay
2655                          */
2656                         if (is_req_replayed_by_update(req)) {
2657                                 struct distribute_txn_replay_req *dtrq;
2658
2659                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2660                                                                       transno);
2661                                 LASSERT(dtrq != NULL);
2662                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2663                                 list_del_init(&dtrq->dtrq_list);
2664                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2665                                 dtrq_destroy(dtrq);
2666
2667                                 if (drop_duplicate_replay_req(env, obd, req))
2668                                         continue;
2669                                 /* not dropped yet */
2670                                 update = true;
2671                         }
2672
2673                         LASSERT(trd->trd_processing_task == current->pid);
2674                         DEBUG_REQ(D_HA, req, "processing x%llu t%lld from %s",
2675                                   req->rq_xid,
2676                                   lustre_msg_get_transno(req->rq_reqmsg),
2677                                   libcfs_nidstr(&req->rq_peer.nid));
2678
2679                         ptlrpc_watchdog_init(&thread->t_watchdog,
2680                                              WATCHDOG_TIMEOUT);
2681                         handle_recovery_req(thread, req,
2682                                             trd->trd_recovery_handler);
2683                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2684
2685                         /**
2686                          * bz18031: increase next_recovery_transno before
2687                          * target_request_copy_put() will drop exp_rpc reference
2688                          */
2689                         if (!update) {
2690                                 spin_lock(&obd->obd_recovery_task_lock);
2691                                 obd->obd_next_recovery_transno++;
2692                                 spin_unlock(&obd->obd_recovery_task_lock);
2693                         }
2694                         target_exp_dequeue_req_replay(req);
2695                         target_request_copy_put(req);
2696                         obd->obd_replayed_requests++;
2697                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2698                         struct distribute_txn_replay_req *dtrq;
2699                         int rc;
2700
2701                         spin_unlock(&obd->obd_recovery_task_lock);
2702
2703                         LASSERT(tdtd != NULL);
2704                         dtrq = distribute_txn_get_next_req(tdtd);
2705                         lu_context_enter(&thread->t_env->le_ctx);
2706                         ptlrpc_watchdog_init(&thread->t_watchdog,
2707                                              WATCHDOG_TIMEOUT);
2708                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2709                         ptlrpc_watchdog_disable(&thread->t_watchdog);
2710                         lu_context_exit(&thread->t_env->le_ctx);
2711                         extend_recovery_timer(obd, obd_timeout, true);
2712
2713                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2714                                 CDEBUG(D_HA,
2715                                        "Move x%llu t%llu to finish list\n",
2716                                        dtrq->dtrq_xid,
2717                                        dtrq->dtrq_master_transno);
2718
2719                                 /* Add it to the replay finish list */
2720                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2721                                 list_add(&dtrq->dtrq_list,
2722                                          &tdtd->tdtd_replay_finish_list);
2723                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2724
2725                                 spin_lock(&obd->obd_recovery_task_lock);
2726                                 if (transno == obd->obd_next_recovery_transno)
2727                                         obd->obd_next_recovery_transno++;
2728                                 else if (transno >
2729                                          obd->obd_next_recovery_transno)
2730                                         obd->obd_next_recovery_transno =
2731                                                                 transno + 1;
2732                                 spin_unlock(&obd->obd_recovery_task_lock);
2733                         } else {
2734                                 dtrq_destroy(dtrq);
2735                         }
2736                 } else {
2737                         spin_unlock(&obd->obd_recovery_task_lock);
2738 abort:
2739                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2740                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2741                         /** evict exports failed VBR */
2742                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2743                         break;
2744                 }
2745         } while (1);
2746 }
2747
2748 static int target_recovery_thread(void *arg)
2749 {
2750         struct lu_target *lut = arg;
2751         struct obd_device *obd = lut->lut_obd;
2752         struct ptlrpc_request *req;
2753         struct target_recovery_data *trd = &obd->obd_recovery_data;
2754         unsigned long delta;
2755         struct lu_env *env;
2756         struct ptlrpc_thread *thread = NULL;
2757         int rc = 0;
2758
2759         ENTRY;
2760         unshare_fs_struct();
2761         OBD_ALLOC_PTR(thread);
2762         if (thread == NULL)
2763                 RETURN(-ENOMEM);
2764
2765         OBD_ALLOC_PTR(env);
2766         if (env == NULL)
2767                 GOTO(out_thread, rc = -ENOMEM);
2768         rc = lu_env_add(env);
2769         if (rc)
2770                 GOTO(out_env, rc);
2771
2772         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2773         if (rc)
2774                 GOTO(out_env_remove, rc);
2775
2776         thread->t_env = env;
2777         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2778         thread->t_task = current;
2779         env->le_ctx.lc_thread = thread;
2780         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2781
2782         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2783                current->pid);
2784         trd->trd_processing_task = current->pid;
2785
2786         spin_lock(&obd->obd_dev_lock);
2787         obd->obd_recovering = 1;
2788         spin_unlock(&obd->obd_dev_lock);
2789         complete(&trd->trd_starting);
2790
2791         /* first of all, we have to know the first transno to replay */
2792         if (target_recovery_overseer(lut, check_for_recovery_ready,
2793                                      exp_connect_healthy)) {
2794                 abort_req_replay_queue(obd);
2795                 abort_lock_replay_queue(obd);
2796                 if (lut->lut_tdtd != NULL)
2797                         dtrq_list_destroy(lut->lut_tdtd);
2798         }
2799
2800         /* next stage: replay requests or update */
2801         delta = jiffies;
2802         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2803                atomic_read(&obd->obd_req_replay_clients),
2804                obd->obd_next_recovery_transno);
2805         replay_request_or_update(env, lut, trd, thread);
2806
2807         /**
2808          * The second stage: replay locks
2809          */
2810         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2811                atomic_read(&obd->obd_lock_replay_clients));
2812         while ((req = target_next_replay_lock(lut))) {
2813                 LASSERT(trd->trd_processing_task == current->pid);
2814                 DEBUG_REQ(D_HA, req, "processing lock from %s:",
2815                           libcfs_nidstr(&req->rq_peer.nid));
2816                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_LOCK_REPLAY)) {
2817                         req->rq_status = -ENODEV;
2818                         target_request_copy_put(req);
2819                         continue;
2820                 }
2821                 handle_recovery_req(thread, req,
2822                                     trd->trd_recovery_handler);
2823                 target_request_copy_put(req);
2824                 obd->obd_replayed_locks++;
2825         }
2826
2827         /**
2828          * The third stage: reply on final pings, at this moment all clients
2829          * must have request in final queue
2830          */
2831         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2832         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2833         /** Update server last boot epoch */
2834         tgt_boot_epoch_update(lut);
2835
2836         /* cancel update llogs upon recovery abort */
2837         if (obd->obd_abort_recovery || obd->obd_abort_mdt_recovery)
2838                 OBP(obd, iocontrol)(OBD_IOC_LLOG_CANCEL, obd->obd_self_export,
2839                                     0, NULL, NULL);
2840
2841         list_for_each_entry(req, &obd->obd_final_req_queue, rq_list) {
2842                 /*
2843                  * Because the waiting client can not send ping to server,
2844                  * so we need refresh the last_request_time, to avoid the
2845                  * export is being evicted
2846                  */
2847                 ptlrpc_update_export_timer(req->rq_export, 0);
2848         }
2849
2850         /*
2851          * We drop recoverying flag to forward all new requests
2852          * to regular mds_handle() since now
2853          */
2854         spin_lock(&obd->obd_dev_lock);
2855         obd->obd_recovering = 0;
2856         obd->obd_abort_recovery = 0;
2857         obd->obd_abort_mdt_recovery = 0;
2858         spin_unlock(&obd->obd_dev_lock);
2859         spin_lock(&obd->obd_recovery_task_lock);
2860         target_cancel_recovery_timer(obd);
2861         spin_unlock(&obd->obd_recovery_task_lock);
2862         while ((req = target_next_final_ping(obd))) {
2863                 LASSERT(trd->trd_processing_task == current->pid);
2864                 DEBUG_REQ(D_HA, req, "processing final ping from %s:",
2865                           libcfs_nidstr(&req->rq_peer.nid));
2866                 handle_recovery_req(thread, req,
2867                                     trd->trd_recovery_handler);
2868                 target_request_copy_put(req);
2869         }
2870
2871         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2872         CDEBUG(D_INFO, "4: recovery completed in %lus - %d/%d reqs/locks\n",
2873                delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2874         if (delta > OBD_RECOVERY_TIME_SOFT) {
2875                 CWARN("too long recovery - read logs\n");
2876                 libcfs_debug_dumplog();
2877         }
2878
2879         target_finish_recovery(lut);
2880         lu_context_fini(&env->le_ctx);
2881         trd->trd_processing_task = 0;
2882         complete_all(&trd->trd_finishing);
2883         tgt_io_thread_done(thread);
2884 out_env_remove:
2885         lu_env_remove(env);
2886 out_env:
2887         OBD_FREE_PTR(env);
2888 out_thread:
2889         OBD_FREE_PTR(thread);
2890         RETURN(rc);
2891 }
2892
2893 static int target_start_recovery_thread(struct lu_target *lut,
2894                                         svc_handler_t handler)
2895 {
2896         struct obd_device *obd = lut->lut_obd;
2897         int rc = 0;
2898         struct target_recovery_data *trd = &obd->obd_recovery_data;
2899         int index;
2900
2901         memset(trd, 0, sizeof(*trd));
2902         init_completion(&trd->trd_starting);
2903         init_completion(&trd->trd_finishing);
2904         trd->trd_recovery_handler = handler;
2905
2906         rc = server_name2index(obd->obd_name, &index, NULL);
2907         if (rc < 0)
2908                 return rc;
2909
2910         if (!IS_ERR(kthread_run(target_recovery_thread,
2911                                 lut, "tgt_recover_%d", index))) {
2912                 wait_for_completion(&trd->trd_starting);
2913                 LASSERT(obd->obd_recovering != 0);
2914         } else {
2915                 rc = -ECHILD;
2916         }
2917
2918         return rc;
2919 }
2920
2921 void target_stop_recovery_thread(struct obd_device *obd)
2922 {
2923         if (obd->obd_recovery_data.trd_processing_task > 0) {
2924                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2925                 /** recovery can be done but postrecovery is not yet */
2926                 spin_lock(&obd->obd_dev_lock);
2927                 if (obd->obd_recovering) {
2928                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2929                         obd->obd_abort_recovery = 1;
2930                         wake_up(&obd->obd_next_transno_waitq);
2931                 }
2932                 spin_unlock(&obd->obd_dev_lock);
2933                 wait_for_completion(&trd->trd_finishing);
2934         }
2935 }
2936 EXPORT_SYMBOL(target_stop_recovery_thread);
2937
2938 void target_recovery_fini(struct obd_device *obd)
2939 {
2940         class_disconnect_exports(obd);
2941         target_stop_recovery_thread(obd);
2942         target_cleanup_recovery(obd);
2943 }
2944 EXPORT_SYMBOL(target_recovery_fini);
2945
2946 static enum hrtimer_restart target_recovery_expired(struct hrtimer *timer)
2947 {
2948         struct obd_device *obd = container_of(timer, struct obd_device,
2949                                               obd_recovery_timer);
2950
2951         CDEBUG(D_HA,
2952                "%s: recovery timed out; %d clients are still in recovery after %llu seconds (%d clients connected)\n",
2953                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2954                ktime_get_seconds() - obd->obd_recovery_start,
2955                atomic_read(&obd->obd_connected_clients));
2956
2957         obd->obd_recovery_expired = 1;
2958         wake_up(&obd->obd_next_transno_waitq);
2959         return HRTIMER_NORESTART;
2960 }
2961
2962 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2963 {
2964         struct obd_device *obd = lut->lut_obd;
2965
2966         if (lut->lut_bottom->dd_rdonly)
2967                 return;
2968
2969         if (atomic_read(&obd->obd_max_recoverable_clients) == 0) {
2970                 /** Update server last boot epoch */
2971                 tgt_boot_epoch_update(lut);
2972                 return;
2973         }
2974
2975         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2976                "last_transno %llu\n", obd->obd_name,
2977                atomic_read(&obd->obd_max_recoverable_clients),
2978                obd->obd_last_committed);
2979         LASSERT(obd->obd_stopping == 0);
2980         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2981         obd->obd_recovery_start = 0;
2982         obd->obd_recovery_end = 0;
2983
2984         hrtimer_init(&obd->obd_recovery_timer, CLOCK_MONOTONIC,
2985                      HRTIMER_MODE_ABS);
2986         obd->obd_recovery_timer.function = &target_recovery_expired;
2987         target_start_recovery_thread(lut, handler);
2988 }
2989 EXPORT_SYMBOL(target_recovery_init);
2990
2991 static int target_process_req_flags(struct obd_device *obd,
2992                                     struct ptlrpc_request *req)
2993 {
2994         struct obd_export *exp = req->rq_export;
2995
2996         LASSERT(exp != NULL);
2997         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2998                 /* client declares he's ready to replay locks */
2999                 spin_lock(&exp->exp_lock);
3000                 if (exp->exp_req_replay_needed) {
3001                         exp->exp_req_replay_needed = 0;
3002                         spin_unlock(&exp->exp_lock);
3003
3004                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
3005                         atomic_dec(&obd->obd_req_replay_clients);
3006                 } else {
3007                         spin_unlock(&exp->exp_lock);
3008                 }
3009         }
3010         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3011                 /*
3012                  * client declares he's ready to complete recovery
3013                  * so, we put the request on th final queue
3014                  */
3015                 spin_lock(&exp->exp_lock);
3016                 if (exp->exp_lock_replay_needed) {
3017                         exp->exp_lock_replay_needed = 0;
3018                         spin_unlock(&exp->exp_lock);
3019
3020                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
3021                         atomic_dec(&obd->obd_lock_replay_clients);
3022                 } else {
3023                         spin_unlock(&exp->exp_lock);
3024                 }
3025         }
3026         return 0;
3027 }
3028
3029 int target_queue_recovery_request(struct ptlrpc_request *req,
3030                                   struct obd_device *obd)
3031 {
3032         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
3033         struct ptlrpc_request *reqiter;
3034         int inserted = 0;
3035
3036         ENTRY;
3037
3038         if (obd->obd_recovery_data.trd_processing_task == current->pid) {
3039                 /* Processing the queue right now, don't re-add. */
3040                 RETURN(1);
3041         }
3042
3043         target_process_req_flags(obd, req);
3044
3045         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
3046                 if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_RECOVERY_REQ_RACE))) {
3047                         if (cfs_fail_val == 1) {
3048                                 cfs_race_state = 1;
3049                                 cfs_fail_val = 0;
3050                                 wake_up(&cfs_race_waitq);
3051
3052                                 schedule_timeout_interruptible(
3053                                         cfs_time_seconds(1));
3054                         }
3055                 }
3056
3057                 /*
3058                  * client declares he's ready to complete recovery
3059                  * so, we put the request on th final queue
3060                  */
3061                 target_request_copy_get(req);
3062                 DEBUG_REQ(D_HA, req, "queue final req");
3063                 wake_up(&obd->obd_next_transno_waitq);
3064                 spin_lock(&obd->obd_recovery_task_lock);
3065                 if (obd->obd_recovering) {
3066                         struct ptlrpc_request *tmp;
3067                         struct ptlrpc_request *duplicate = NULL;
3068
3069                         if (likely(!req->rq_export->exp_replay_done)) {
3070                                 req->rq_export->exp_replay_done = 1;
3071                                 list_add_tail(&req->rq_list,
3072                                               &obd->obd_final_req_queue);
3073                                 spin_unlock(&obd->obd_recovery_task_lock);
3074                                 RETURN(0);
3075                         }
3076
3077                         /*
3078                          * XXX O(n), but only happens if final ping is
3079                          * timed out, probably reorganize the list as
3080                          * a hash list later
3081                          */
3082                         list_for_each_entry_safe(reqiter, tmp,
3083                                                  &obd->obd_final_req_queue,
3084                                                  rq_list) {
3085                                 if (reqiter->rq_export == req->rq_export) {
3086                                         list_del_init(&reqiter->rq_list);
3087                                         duplicate = reqiter;
3088                                         break;
3089                                 }
3090                         }
3091
3092                         list_add_tail(&req->rq_list,
3093                                       &obd->obd_final_req_queue);
3094                         req->rq_export->exp_replay_done = 1;
3095                         spin_unlock(&obd->obd_recovery_task_lock);
3096
3097                         if (duplicate != NULL) {
3098                                 DEBUG_REQ(D_HA, duplicate,
3099                                           "put prev final req");
3100                                 target_request_copy_put(duplicate);
3101                         }
3102                         RETURN(0);
3103                 } else {
3104                         spin_unlock(&obd->obd_recovery_task_lock);
3105                         target_request_copy_put(req);
3106                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
3107                 }
3108         }
3109         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
3110                 /* client declares he's ready to replay locks */
3111                 target_request_copy_get(req);
3112                 DEBUG_REQ(D_HA, req, "queue lock replay req");
3113                 wake_up(&obd->obd_next_transno_waitq);
3114                 spin_lock(&obd->obd_recovery_task_lock);
3115                 LASSERT(obd->obd_recovering);
3116                 /* usually due to recovery abort */
3117                 if (!req->rq_export->exp_in_recovery) {
3118                         spin_unlock(&obd->obd_recovery_task_lock);
3119                         target_request_copy_put(req);
3120                         RETURN(-ENOTCONN);
3121                 }
3122                 LASSERT(req->rq_export->exp_lock_replay_needed);
3123                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
3124                 spin_unlock(&obd->obd_recovery_task_lock);
3125                 RETURN(0);
3126         }
3127
3128         /*
3129          * CAVEAT EMPTOR: The incoming request message has been swabbed
3130          * (i.e. buflens etc are in my own byte order), but type-dependent
3131          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed.
3132          */
3133
3134         if (!transno) {
3135                 INIT_LIST_HEAD(&req->rq_list);
3136                 DEBUG_REQ(D_HA, req, "not queueing");
3137                 RETURN(1);
3138         }
3139
3140         /*
3141          * If we're processing the queue, we want don't want to queue this
3142          * message.
3143          *
3144          * Also, if this request has a transno less than the one we're waiting
3145          * for, we should process it now.  It could (and currently always will)
3146          * be an open request for a descriptor that was opened some time ago.
3147          *
3148          * Also, a resent, replayed request that has already been
3149          * handled will pass through here and be processed immediately.
3150          */
3151         CDEBUG(D_HA,
3152                "Next recovery transno: %llu, current: %llu, replaying\n",
3153                obd->obd_next_recovery_transno, transno);
3154
3155         /*
3156          * If the request has been replayed by update replay, then sends this
3157          * request to the recovery thread (replay_request_or_update()), where
3158          * it will be handled
3159          */
3160         spin_lock(&obd->obd_recovery_task_lock);
3161         if (transno < obd->obd_next_recovery_transno &&
3162             !is_req_replayed_by_update(req)) {
3163                 /* Processing the queue right now, don't re-add. */
3164                 LASSERT(list_empty(&req->rq_list));
3165                 spin_unlock(&obd->obd_recovery_task_lock);
3166                 RETURN(1);
3167         }
3168         spin_unlock(&obd->obd_recovery_task_lock);
3169
3170         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
3171                 RETURN(0);
3172
3173         target_request_copy_get(req);
3174         if (!req->rq_export->exp_in_recovery) {
3175                 target_request_copy_put(req);
3176                 RETURN(-ENOTCONN);
3177         }
3178         LASSERT(req->rq_export->exp_req_replay_needed);
3179
3180         if (target_exp_enqueue_req_replay(req)) {
3181                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
3182                 target_request_copy_put(req);
3183                 RETURN(0);
3184         }
3185
3186         /* XXX O(n^2) */
3187         spin_lock(&obd->obd_recovery_task_lock);
3188         LASSERT(obd->obd_recovering);
3189         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
3190                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
3191                         list_add_tail(&req->rq_list, &reqiter->rq_list);
3192                         inserted = 1;
3193                         goto added;
3194                 }
3195
3196                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
3197                              transno)) {
3198                         DEBUG_REQ(D_ERROR, req,
3199                                   "dropping replay: transno has been claimed by another client");
3200                         spin_unlock(&obd->obd_recovery_task_lock);
3201                         target_exp_dequeue_req_replay(req);
3202                         target_request_copy_put(req);
3203                         RETURN(0);
3204                 }
3205         }
3206 added:
3207         if (!inserted)
3208                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
3209
3210         obd->obd_requests_queued_for_recovery++;
3211         spin_unlock(&obd->obd_recovery_task_lock);
3212         wake_up(&obd->obd_next_transno_waitq);
3213         RETURN(0);
3214 }
3215
3216 void target_committed_to_req(struct ptlrpc_request *req)
3217 {
3218         struct obd_export *exp = req->rq_export;
3219
3220         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
3221                 lustre_msg_set_last_committed(req->rq_repmsg,
3222                                               exp->exp_last_committed);
3223         else
3224                 DEBUG_REQ(D_IOCTL, req,
3225                           "not sending last_committed update (%d/%d)",
3226                           exp->exp_obd->obd_no_transno,
3227                           req->rq_repmsg == NULL);
3228
3229         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
3230                exp->exp_last_committed, req->rq_transno, req->rq_xid);
3231 }
3232
3233 #endif /* HAVE_SERVER_SUPPORT */
3234
3235 /**
3236  * Packs current SLV and Limit into \a req.
3237  */
3238 int target_pack_pool_reply(struct ptlrpc_request *req)
3239 {
3240         struct obd_device *obd;
3241
3242         ENTRY;
3243
3244         /*
3245          * Check that we still have all structures alive as this may
3246          * be some late RPC at shutdown time.
3247          */
3248         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
3249                      !exp_connect_lru_resize(req->rq_export))) {
3250                 lustre_msg_set_slv(req->rq_repmsg, 0);
3251                 lustre_msg_set_limit(req->rq_repmsg, 0);
3252                 RETURN(0);
3253         }
3254
3255         /* OBD is alive here as export is alive, which we checked above. */
3256         obd = req->rq_export->exp_obd;
3257
3258         read_lock(&obd->obd_pool_lock);
3259         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
3260         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
3261         read_unlock(&obd->obd_pool_lock);
3262
3263         RETURN(0);
3264 }
3265
3266 static int target_send_reply_msg(struct ptlrpc_request *req,
3267                                  int rc, int fail_id)
3268 {
3269         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
3270                 DEBUG_REQ(D_ERROR, req, "dropping reply");
3271                 return -ECOMM;
3272         }
3273         /*
3274          * We can have a null rq_reqmsg in the event of bad signature or
3275          * no context when unwrapping
3276          */
3277         if (req->rq_reqmsg &&
3278             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
3279             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
3280                 return -ECOMM;
3281
3282         if (unlikely(rc)) {
3283                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
3284                 req->rq_status = rc;
3285                 return ptlrpc_send_error(req, 1);
3286         }
3287         DEBUG_REQ(D_NET, req, "sending reply");
3288
3289         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
3290 }
3291
3292 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
3293 {
3294         struct ptlrpc_service_part *svcpt;
3295         int netrc;
3296         struct ptlrpc_reply_state *rs;
3297         struct obd_export *exp;
3298
3299         ENTRY;
3300
3301         if (req->rq_no_reply) {
3302                 EXIT;
3303                 return;
3304         }
3305
3306         svcpt = req->rq_rqbd->rqbd_svcpt;
3307         rs = req->rq_reply_state;
3308         if (rs == NULL || !rs->rs_difficult) {
3309                 /* no notifiers */
3310                 target_send_reply_msg(req, rc, fail_id);
3311                 EXIT;
3312                 return;
3313         }
3314
3315         /* must be an export if locks saved */
3316         LASSERT(req->rq_export != NULL);
3317         /* req/reply consistent */
3318         LASSERT(rs->rs_svcpt == svcpt);
3319
3320         /* "fresh" reply */
3321         LASSERT(!rs->rs_scheduled);
3322         LASSERT(!rs->rs_scheduled_ever);
3323         LASSERT(!rs->rs_handled);
3324         LASSERT(!rs->rs_sent);
3325         LASSERT(!rs->rs_unlinked);
3326         LASSERT(rs->rs_export == NULL);
3327         LASSERT(list_empty(&rs->rs_obd_list));
3328         LASSERT(list_empty(&rs->rs_exp_list));
3329
3330         exp = class_export_get(req->rq_export);
3331
3332         /* disable reply scheduling while I'm setting up */
3333         rs->rs_scheduled = 1;
3334         rs->rs_sent      = 0;
3335         rs->rs_unlinked  = 0;
3336         rs->rs_xid       = req->rq_xid;
3337         rs->rs_transno   = req->rq_transno;
3338         rs->rs_export    = exp;
3339         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
3340
3341         spin_lock(&exp->exp_uncommitted_replies_lock);
3342         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
3343                rs->rs_transno, exp->exp_last_committed);
3344         if (rs->rs_transno > exp->exp_last_committed) {
3345                 /* not committed already */
3346                 list_add_tail(&rs->rs_obd_list,
3347                                   &exp->exp_uncommitted_replies);
3348         }
3349         spin_unlock(&exp->exp_uncommitted_replies_lock);
3350
3351         spin_lock(&exp->exp_lock);
3352         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
3353         spin_unlock(&exp->exp_lock);
3354
3355         netrc = target_send_reply_msg(req, rc, fail_id);
3356
3357         spin_lock(&svcpt->scp_rep_lock);
3358
3359         atomic_inc(&svcpt->scp_nreps_difficult);
3360
3361         if (netrc != 0) {
3362                 /*
3363                  * error sending: reply is off the net.  Also we need +1
3364                  * reply ref until ptlrpc_handle_rs() is done
3365                  * with the reply state (if the send was successful, there
3366                  * would have been +1 ref for the net, which
3367                  * reply_out_callback leaves alone)
3368                  */
3369                 rs->rs_sent = 1;
3370                 rs->rs_unlinked = 1;
3371                 ptlrpc_rs_addref(rs);
3372         }
3373
3374         spin_lock(&rs->rs_lock);
3375         if (rs->rs_transno <= exp->exp_last_committed ||
3376             (rs->rs_unlinked && !rs->rs_no_ack) ||
3377             list_empty(&rs->rs_exp_list) ||     /* completed already */
3378             list_empty(&rs->rs_obd_list)) {
3379                 CDEBUG(D_HA, "Schedule reply immediately\n");
3380                 ptlrpc_dispatch_difficult_reply(rs);
3381         } else {
3382                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3383                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3384         }
3385         spin_unlock(&rs->rs_lock);
3386         spin_unlock(&svcpt->scp_rep_lock);
3387         EXIT;
3388 }
3389
3390 enum ldlm_mode lck_compat_array[] = {
3391         [LCK_EX]    = LCK_COMPAT_EX,
3392         [LCK_PW]    = LCK_COMPAT_PW,
3393         [LCK_PR]    = LCK_COMPAT_PR,
3394         [LCK_CW]    = LCK_COMPAT_CW,
3395         [LCK_CR]    = LCK_COMPAT_CR,
3396         [LCK_NL]    = LCK_COMPAT_NL,
3397         [LCK_GROUP] = LCK_COMPAT_GROUP,
3398         [LCK_COS]   = LCK_COMPAT_COS,
3399 };
3400
3401 /**
3402  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3403  * not escape to the user level.
3404  */
3405 int ldlm_error2errno(enum ldlm_error error)
3406 {
3407         int result;
3408
3409         switch (error) {
3410         case ELDLM_OK:
3411         case ELDLM_LOCK_MATCHED:
3412                 result = 0;
3413                 break;
3414         case ELDLM_LOCK_CHANGED:
3415                 result = -ESTALE;
3416                 break;
3417         case ELDLM_LOCK_ABORTED:
3418                 result = -ENAVAIL;
3419                 break;
3420         case ELDLM_LOCK_REPLACED:
3421                 result = -ESRCH;
3422                 break;
3423         case ELDLM_NO_LOCK_DATA:
3424                 result = -ENOENT;
3425                 break;
3426         case ELDLM_NAMESPACE_EXISTS:
3427                 result = -EEXIST;
3428                 break;
3429         case ELDLM_BAD_NAMESPACE:
3430                 result = -EBADF;
3431                 break;
3432         default:
3433                 if (((int)error) < 0) { /* cast to signed type */
3434                         result = error; /* as ldlm_error can be unsigned */
3435                 } else {
3436                         CERROR("Invalid DLM result code: %d\n", error);
3437                         result = -EPROTO;
3438                 }
3439         }
3440         return result;
3441 }
3442 EXPORT_SYMBOL(ldlm_error2errno);
3443
3444 /**
3445  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3446  */
3447 enum ldlm_error ldlm_errno2error(int err_no)
3448 {
3449         int error;
3450
3451         switch (err_no) {
3452         case 0:
3453                 error = ELDLM_OK;
3454                 break;
3455         case -ESTALE:
3456                 error = ELDLM_LOCK_CHANGED;
3457                 break;
3458         case -ENAVAIL:
3459                 error = ELDLM_LOCK_ABORTED;
3460                 break;
3461         case -ESRCH:
3462                 error = ELDLM_LOCK_REPLACED;
3463                 break;
3464         case -ENOENT:
3465                 error = ELDLM_NO_LOCK_DATA;
3466                 break;
3467         case -EEXIST:
3468                 error = ELDLM_NAMESPACE_EXISTS;
3469                 break;
3470         case -EBADF:
3471                 error = ELDLM_BAD_NAMESPACE;
3472                 break;
3473         default:
3474                 error = err_no;
3475         }
3476         return error;
3477 }
3478
3479 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3480 void ldlm_dump_export_locks(struct obd_export *exp)
3481 {
3482         spin_lock(&exp->exp_locks_list_guard);
3483         if (!list_empty(&exp->exp_locks_list)) {
3484                 struct ldlm_lock *lock;
3485
3486                 CERROR("dumping locks for export %p, ignore if the unmount doesn't hang\n",
3487                        exp);
3488                 list_for_each_entry(lock, &exp->exp_locks_list,
3489                                         l_exp_refs_link)
3490                         LDLM_ERROR(lock, "lock:");
3491         }
3492         spin_unlock(&exp->exp_locks_list_guard);
3493 }
3494 #endif
3495
3496 #ifdef HAVE_SERVER_SUPPORT
3497 static inline const char *bulk2type(struct ptlrpc_request *req)
3498 {
3499         if (req->rq_bulk_read)
3500                 return "READ";
3501         if (req->rq_bulk_write)
3502                 return "WRITE";
3503         return "UNKNOWN";
3504 }
3505
3506 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc)
3507 {
3508         struct ptlrpc_request *req = desc->bd_req;
3509         time64_t start = ktime_get_seconds();
3510         time64_t deadline;
3511         int rc = 0;
3512
3513         ENTRY;
3514
3515         /* If there is eviction in progress, wait for it to finish. */
3516         wait_event_idle(
3517                 exp->exp_obd->obd_evict_inprogress_waitq,
3518                 !atomic_read(&exp->exp_obd->obd_evict_inprogress));
3519
3520         /* Check if client was evicted or reconnected already. */
3521         if (exp->exp_failed ||
3522             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3523                 rc = -ENOTCONN;
3524         } else {
3525                 if (req->rq_bulk_read)
3526                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3527
3528                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3529                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3530                 else /* old version, bulk matchbits is rq_xid */
3531                         req->rq_mbits = req->rq_xid;
3532
3533                 if (rc == 0)
3534                         rc = ptlrpc_start_bulk_transfer(desc);
3535         }
3536
3537         if (rc < 0) {
3538                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc = %d",
3539                           bulk2type(req), rc);
3540                 RETURN(rc);
3541         }
3542
3543         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3544                 ptlrpc_abort_bulk(desc);
3545                 RETURN(0);
3546         }
3547
3548         /* limit actual bulk transfer to bulk_timeout seconds */
3549         deadline = start + bulk_timeout;
3550         if (deadline > req->rq_deadline)
3551                 deadline = req->rq_deadline;
3552
3553         do {
3554                 time64_t timeoutl = deadline - ktime_get_seconds();
3555                 time64_t rq_deadline;
3556
3557                 while (timeoutl >= 0 &&
3558                        wait_event_idle_timeout(
3559                                desc->bd_waitq,
3560                                !ptlrpc_server_bulk_active(desc) ||
3561                                exp->exp_failed ||
3562                                exp->exp_conn_cnt >
3563                                lustre_msg_get_conn_cnt(req->rq_reqmsg),
3564                                timeoutl ? cfs_time_seconds(1) : 1) == 0)
3565                         timeoutl -= 1;
3566                 rc = timeoutl < 0 ? -ETIMEDOUT : 0;
3567
3568                 /* Wait again if we changed rq_deadline. */
3569                 rq_deadline = READ_ONCE(req->rq_deadline);
3570                 deadline = start + bulk_timeout;
3571                 if (deadline > rq_deadline)
3572                         deadline = rq_deadline;
3573         } while (rc == -ETIMEDOUT &&
3574                  deadline > ktime_get_seconds());
3575
3576         if (rc == -ETIMEDOUT) {
3577                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %lld%+llds",
3578                           bulk2type(req), deadline - start,
3579                           ktime_get_real_seconds() - deadline);
3580                 ptlrpc_abort_bulk(desc);
3581         } else if (exp->exp_failed) {
3582                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3583                           bulk2type(req));
3584                 rc = -ENOTCONN;
3585                 ptlrpc_abort_bulk(desc);
3586         } else if (exp->exp_conn_cnt >
3587                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3588                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3589                           bulk2type(req));
3590                 /* We don't reply anyway. */
3591                 rc = -ETIMEDOUT;
3592                 ptlrpc_abort_bulk(desc);
3593         } else if (desc->bd_failure) {
3594                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3595                           bulk2type(req));
3596                 /* XXX should this be a different errno? */
3597                 rc = -ETIMEDOUT;
3598         } else {
3599                 if (req->rq_bulk_write)
3600                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3601                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3602                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3603                                   bulk2type(req), desc->bd_nob_transferred,
3604                                   desc->bd_nob);
3605                         /* XXX should this be a different errno? */
3606                         rc = -ETIMEDOUT;
3607                 }
3608         }
3609
3610         RETURN(rc);
3611 }
3612 EXPORT_SYMBOL(target_bulk_io);
3613
3614 #endif /* HAVE_SERVER_SUPPORT */