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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 /**
  38  * This file deals with various client/target related logic including recovery.
  39  *
  40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  41  * should be moved.
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LDLM
  45
  46 #include <linux/kthread.h>
  47 #include <libcfs/libcfs.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_dlm.h>
  51 #include <lustre_net.h>
  52 #include <lustre_sec.h>
  53 #include "ldlm_internal.h"
  54
  55 /* @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         int rc = 0;
  64         ENTRY;
  65
  66         if (!create && !priority) {
  67                 CDEBUG(D_HA, "Nothing to do\n");
  68                 RETURN(-EINVAL);
  69         }
  70
  71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  72         if (!ptlrpc_conn) {
  73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  74                 RETURN (-ENOENT);
  75         }
  76
  77         if (create) {
  78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  79                 if (!imp_conn) {
  80                         GOTO(out_put, rc = -ENOMEM);
  81                 }
  82         }
  83
  84         spin_lock(&imp->imp_lock);
  85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  87                         if (priority) {
  88                                 list_del(&item->oic_item);
  89                                 list_add(&item->oic_item,
  90                                              &imp->imp_conn_list);
  91                                 item->oic_last_attempt = 0;
  92                         }
  93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  94                                imp, imp->imp_obd->obd_name, uuid->uuid,
  95                                (priority ? ", moved to head" : ""));
  96                         spin_unlock(&imp->imp_lock);
  97                         GOTO(out_free, rc = 0);
  98                 }
  99         }
 100         /* No existing import connection found for \a uuid. */
 101         if (create) {
 102                 imp_conn->oic_conn = ptlrpc_conn;
 103                 imp_conn->oic_uuid = *uuid;
 104                 imp_conn->oic_last_attempt = 0;
 105                 if (priority)
 106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 else
 108                         list_add_tail(&imp_conn->oic_item,
 109                                           &imp->imp_conn_list);
 110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 111                        imp, imp->imp_obd->obd_name, uuid->uuid,
 112                        (priority ? "head" : "tail"));
 113         } else {
 114                 spin_unlock(&imp->imp_lock);
 115                 GOTO(out_free, rc = -ENOENT);
 116         }
 117
 118         spin_unlock(&imp->imp_lock);
 119         RETURN(0);
 120 out_free:
 121         if (imp_conn)
 122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 123 out_put:
 124         ptlrpc_connection_put(ptlrpc_conn);
 125         RETURN(rc);
 126 }
 127
 128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 129 {
 130         return import_set_conn(imp, uuid, 1, 0);
 131 }
 132
 133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 134                            int priority)
 135 {
 136         return import_set_conn(imp, uuid, priority, 1);
 137 }
 138 EXPORT_SYMBOL(client_import_add_conn);
 139
 140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 141 {
 142         struct obd_import_conn *imp_conn;
 143         struct obd_export *dlmexp;
 144         int rc = -ENOENT;
 145         ENTRY;
 146
 147         spin_lock(&imp->imp_lock);
 148         if (list_empty(&imp->imp_conn_list)) {
 149                 LASSERT(!imp->imp_connection);
 150                 GOTO(out, rc);
 151         }
 152
 153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 155                         continue;
 156                 LASSERT(imp_conn->oic_conn);
 157
 158                 if (imp_conn == imp->imp_conn_current) {
 159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 160
 161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 162                             imp->imp_state != LUSTRE_IMP_DISCON) {
 163                                 CERROR("can't remove current connection\n");
 164                                 GOTO(out, rc = -EBUSY);
 165                         }
 166
 167                         ptlrpc_connection_put(imp->imp_connection);
 168                         imp->imp_connection = NULL;
 169
 170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 171                         if (dlmexp && dlmexp->exp_connection) {
 172                                 LASSERT(dlmexp->exp_connection ==
 173                                         imp_conn->oic_conn);
 174                                 ptlrpc_connection_put(dlmexp->exp_connection);
 175                                 dlmexp->exp_connection = NULL;
 176                         }
 177                 }
 178
 179                 list_del(&imp_conn->oic_item);
 180                 ptlrpc_connection_put(imp_conn->oic_conn);
 181                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 182                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 183                        imp, imp->imp_obd->obd_name, uuid->uuid);
 184                 rc = 0;
 185                 break;
 186         }
 187 out:
 188         spin_unlock(&imp->imp_lock);
 189         if (rc == -ENOENT)
 190                 CERROR("connection %s not found\n", uuid->uuid);
 191         RETURN(rc);
 192 }
 193 EXPORT_SYMBOL(client_import_del_conn);
 194
 195 /**
 196  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 197  * to find a conn uuid of \a imp which can reach \a peer.
 198  */
 199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 200                             struct obd_uuid *uuid)
 201 {
 202         struct obd_import_conn *conn;
 203         int rc = -ENOENT;
 204         ENTRY;
 205
 206         spin_lock(&imp->imp_lock);
 207         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 208                 /* Check if conn UUID does have this peer NID. */
 209                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 210                         *uuid = conn->oic_uuid;
 211                         rc = 0;
 212                         break;
 213                 }
 214         }
 215         spin_unlock(&imp->imp_lock);
 216         RETURN(rc);
 217 }
 218 EXPORT_SYMBOL(client_import_find_conn);
 219
 220 void client_destroy_import(struct obd_import *imp)
 221 {
 222         /* Drop security policy instance after all RPCs have finished/aborted
 223          * to let all busy contexts be released. */
 224         class_import_get(imp);
 225         class_destroy_import(imp);
 226         sptlrpc_import_sec_put(imp);
 227         class_import_put(imp);
 228 }
 229 EXPORT_SYMBOL(client_destroy_import);
 230
 231 /**
 232  * Check whether or not the OSC is on MDT.
 233  * In the config log,
 234  * osc on MDT
 235  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 236  * osc on client
 237  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 238  *
 239  **/
 240 static int osc_on_mdt(char *obdname)
 241 {
 242         char *ptr;
 243
 244         ptr = strrchr(obdname, '-');
 245         if (ptr == NULL)
 246                 return 0;
 247
 248         if (strncmp(ptr + 1, "MDT", 3) == 0)
 249                 return 1;
 250
 251         return 0;
 252 }
 253
 254 /* Configure an RPC client OBD device.
 255  *
 256  * lcfg parameters:
 257  * 1 - client UUID
 258  * 2 - server UUID
 259  * 3 - inactive-on-startup
 260  */
 261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 262 {
 263         struct client_obd *cli = &obddev->u.cli;
 264         struct obd_import *imp;
 265         struct obd_uuid server_uuid;
 266         int rq_portal, rp_portal, connect_op;
 267         char *name = obddev->obd_type->typ_name;
 268         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 269         char *cli_name = lustre_cfg_buf(lcfg, 0);
 270         int rc;
 271         ENTRY;
 272
 273         /* In a more perfect world, we would hang a ptlrpc_client off of
 274          * obd_type and just use the values from there. */
 275         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 276                 rq_portal = OST_REQUEST_PORTAL;
 277                 rp_portal = OSC_REPLY_PORTAL;
 278                 connect_op = OST_CONNECT;
 279                 cli->cl_sp_me = LUSTRE_SP_CLI;
 280                 cli->cl_sp_to = LUSTRE_SP_OST;
 281                 ns_type = LDLM_NS_TYPE_OSC;
 282         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 283                    !strcmp(name, LUSTRE_LWP_NAME)) {
 284                 rq_portal = MDS_REQUEST_PORTAL;
 285                 rp_portal = MDC_REPLY_PORTAL;
 286                 connect_op = MDS_CONNECT;
 287                 if (is_lwp_on_ost(cli_name))
 288                         cli->cl_sp_me = LUSTRE_SP_OST;
 289                 else if (is_lwp_on_mdt(cli_name))
 290                         cli->cl_sp_me = LUSTRE_SP_MDT;
 291                 else
 292                         cli->cl_sp_me = LUSTRE_SP_CLI;
 293                 cli->cl_sp_to = LUSTRE_SP_MDT;
 294                 ns_type = LDLM_NS_TYPE_MDC;
 295         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 296                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 297                         /* OSP_on_MDT for other MDTs */
 298                         connect_op = MDS_CONNECT;
 299                         cli->cl_sp_to = LUSTRE_SP_MDT;
 300                         ns_type = LDLM_NS_TYPE_MDC;
 301                         rq_portal = OUT_PORTAL;
 302                 } else {
 303                         /* OSP on MDT for OST */
 304                         connect_op = OST_CONNECT;
 305                         cli->cl_sp_to = LUSTRE_SP_OST;
 306                         ns_type = LDLM_NS_TYPE_OSC;
 307                         rq_portal = OST_REQUEST_PORTAL;
 308                 }
 309                 rp_portal = OSC_REPLY_PORTAL;
 310                 cli->cl_sp_me = LUSTRE_SP_MDT;
 311         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 312                 rq_portal = MGS_REQUEST_PORTAL;
 313                 rp_portal = MGC_REPLY_PORTAL;
 314                 connect_op = MGS_CONNECT;
 315                 cli->cl_sp_me = LUSTRE_SP_MGC;
 316                 cli->cl_sp_to = LUSTRE_SP_MGS;
 317                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 318                 ns_type = LDLM_NS_TYPE_MGC;
 319         } else {
 320                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 321                        name);
 322                 RETURN(-EINVAL);
 323         }
 324
 325         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 326                 CERROR("requires a TARGET UUID\n");
 327                 RETURN(-EINVAL);
 328         }
 329
 330         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 331                 CERROR("client UUID must be less than 38 characters\n");
 332                 RETURN(-EINVAL);
 333         }
 334
 335         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 336                 CERROR("setup requires a SERVER UUID\n");
 337                 RETURN(-EINVAL);
 338         }
 339
 340         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 341                 CERROR("target UUID must be less than 38 characters\n");
 342                 RETURN(-EINVAL);
 343         }
 344
 345         init_rwsem(&cli->cl_sem);
 346         mutex_init(&cli->cl_mgc_mutex);
 347         cli->cl_conn_count = 0;
 348         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 349                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 350                      sizeof(server_uuid)));
 351
 352         cli->cl_dirty_pages = 0;
 353         cli->cl_avail_grant = 0;
 354         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 355         /* cl_dirty_max_pages may be changed at connect time in
 356          * ptlrpc_connect_interpret(). */
 357         client_adjust_max_dirty(cli);
 358         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 359         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 360         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 361         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 362         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 363         spin_lock_init(&cli->cl_loi_list_lock);
 364         atomic_set(&cli->cl_pending_w_pages, 0);
 365         atomic_set(&cli->cl_pending_r_pages, 0);
 366         cli->cl_r_in_flight = 0;
 367         cli->cl_w_in_flight = 0;
 368
 369         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 370         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 371         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 372         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 373         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 374         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 375
 376         /* lru for osc. */
 377         INIT_LIST_HEAD(&cli->cl_lru_osc);
 378         atomic_set(&cli->cl_lru_shrinkers, 0);
 379         atomic_long_set(&cli->cl_lru_busy, 0);
 380         atomic_long_set(&cli->cl_lru_in_list, 0);
 381         INIT_LIST_HEAD(&cli->cl_lru_list);
 382         spin_lock_init(&cli->cl_lru_list_lock);
 383         atomic_long_set(&cli->cl_unstable_count, 0);
 384
 385         init_waitqueue_head(&cli->cl_destroy_waitq);
 386         atomic_set(&cli->cl_destroy_in_flight, 0);
 387 #ifdef ENABLE_CHECKSUM
 388         /* Turn on checksumming by default. */
 389         cli->cl_checksum = 1;
 390         /*
 391          * The supported checksum types will be worked out at connect time
 392          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 393          * through procfs.
 394          */
 395         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 396 #endif
 397         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 398
 399         /* This value may be reduced at connect time in
 400          * ptlrpc_connect_interpret() . We initialize it to only
 401          * 1MB until we know what the performance looks like.
 402          * In the future this should likely be increased. LU-1431 */
 403         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
 404                                           LNET_MTU >> PAGE_CACHE_SHIFT);
 405
 406         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 407          * it will be updated at OSC connection time. */
 408         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 409
 410         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 411                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 412         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 413                 cli->cl_max_rpcs_in_flight = 2;
 414         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 415                 cli->cl_max_rpcs_in_flight = 3;
 416         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 417                 cli->cl_max_rpcs_in_flight = 4;
 418         } else {
 419                 if (osc_on_mdt(obddev->obd_name))
 420                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 421                 else
 422                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 423         }
 424
 425         spin_lock_init(&cli->cl_mod_rpcs_lock);
 426         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 427         cli->cl_max_mod_rpcs_in_flight = 0;
 428         cli->cl_mod_rpcs_in_flight = 0;
 429         cli->cl_close_rpcs_in_flight = 0;
 430         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 431         cli->cl_mod_tag_bitmap = NULL;
 432
 433         if (connect_op == MDS_CONNECT) {
 434                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 435                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 436                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 437                 if (cli->cl_mod_tag_bitmap == NULL)
 438                         GOTO(err, rc = -ENOMEM);
 439         }
 440
 441         rc = ldlm_get_ref();
 442         if (rc) {
 443                 CERROR("ldlm_get_ref failed: %d\n", rc);
 444                 GOTO(err, rc);
 445         }
 446
 447         ptlrpc_init_client(rq_portal, rp_portal, name,
 448                            &obddev->obd_ldlm_client);
 449
 450         imp = class_new_import(obddev);
 451         if (imp == NULL)
 452                 GOTO(err_ldlm, rc = -ENOENT);
 453         imp->imp_client = &obddev->obd_ldlm_client;
 454         imp->imp_connect_op = connect_op;
 455         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 456                LUSTRE_CFG_BUFLEN(lcfg, 1));
 457         class_import_put(imp);
 458
 459         rc = client_import_add_conn(imp, &server_uuid, 1);
 460         if (rc) {
 461                 CERROR("can't add initial connection\n");
 462                 GOTO(err_import, rc);
 463         }
 464
 465         cli->cl_import = imp;
 466         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 467         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 468
 469         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 470                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 471                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 472                                name, obddev->obd_name,
 473                                cli->cl_target_uuid.uuid);
 474                         spin_lock(&imp->imp_lock);
 475                         imp->imp_deactive = 1;
 476                         spin_unlock(&imp->imp_lock);
 477                 }
 478         }
 479
 480         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 481                                                    LDLM_NAMESPACE_CLIENT,
 482                                                    LDLM_NAMESPACE_GREEDY,
 483                                                    ns_type);
 484         if (obddev->obd_namespace == NULL) {
 485                 CERROR("Unable to create client namespace - %s\n",
 486                        obddev->obd_name);
 487                 GOTO(err_import, rc = -ENOMEM);
 488         }
 489
 490         RETURN(rc);
 491
 492 err_import:
 493         class_destroy_import(imp);
 494 err_ldlm:
 495         ldlm_put_ref();
 496 err:
 497         if (cli->cl_mod_tag_bitmap != NULL)
 498                 OBD_FREE(cli->cl_mod_tag_bitmap,
 499                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 500         cli->cl_mod_tag_bitmap = NULL;
 501         RETURN(rc);
 502
 503 }
 504 EXPORT_SYMBOL(client_obd_setup);
 505
 506 int client_obd_cleanup(struct obd_device *obddev)
 507 {
 508         struct client_obd *cli = &obddev->u.cli;
 509         ENTRY;
 510
 511         ldlm_namespace_free_post(obddev->obd_namespace);
 512         obddev->obd_namespace = NULL;
 513
 514         obd_cleanup_client_import(obddev);
 515         LASSERT(obddev->u.cli.cl_import == NULL);
 516
 517         ldlm_put_ref();
 518
 519         if (cli->cl_mod_tag_bitmap != NULL)
 520                 OBD_FREE(cli->cl_mod_tag_bitmap,
 521                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 522         cli->cl_mod_tag_bitmap = NULL;
 523
 524         RETURN(0);
 525 }
 526 EXPORT_SYMBOL(client_obd_cleanup);
 527
 528 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 529 int client_connect_import(const struct lu_env *env,
 530                           struct obd_export **exp,
 531                           struct obd_device *obd, struct obd_uuid *cluuid,
 532                           struct obd_connect_data *data, void *localdata)
 533 {
 534         struct client_obd       *cli    = &obd->u.cli;
 535         struct obd_import       *imp    = cli->cl_import;
 536         struct obd_connect_data *ocd;
 537         struct lustre_handle    conn    = { 0 };
 538         int                     rc;
 539         bool                    is_mdc = false;
 540         ENTRY;
 541
 542         *exp = NULL;
 543         down_write(&cli->cl_sem);
 544         if (cli->cl_conn_count > 0)
 545                 GOTO(out_sem, rc = -EALREADY);
 546
 547         rc = class_connect(&conn, obd, cluuid);
 548         if (rc)
 549                 GOTO(out_sem, rc);
 550
 551         cli->cl_conn_count++;
 552         *exp = class_conn2export(&conn);
 553
 554         LASSERT(obd->obd_namespace);
 555
 556         imp->imp_dlm_handle = conn;
 557         rc = ptlrpc_init_import(imp);
 558         if (rc != 0)
 559                 GOTO(out_ldlm, rc);
 560
 561         ocd = &imp->imp_connect_data;
 562         if (data) {
 563                 *ocd = *data;
 564                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
 565                                  LUSTRE_MDC_NAME, 3) == 0;
 566                 if (is_mdc)
 567                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
 568                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 569         }
 570
 571         rc = ptlrpc_connect_import(imp);
 572         if (rc != 0) {
 573                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 574                 GOTO(out_ldlm, rc);
 575         }
 576         LASSERT(*exp != NULL && (*exp)->exp_connection);
 577
 578         if (data) {
 579                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 580                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 581                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 582                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 583                 /* clear the flag as it was not set and is not known
 584                  * by upper layers */
 585                 if (is_mdc)
 586                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 587         }
 588
 589         ptlrpc_pinger_add_import(imp);
 590
 591         EXIT;
 592
 593         if (rc) {
 594 out_ldlm:
 595                 cli->cl_conn_count--;
 596                 class_disconnect(*exp);
 597                 *exp = NULL;
 598         }
 599 out_sem:
 600         up_write(&cli->cl_sem);
 601
 602         return rc;
 603 }
 604 EXPORT_SYMBOL(client_connect_import);
 605
 606 int client_disconnect_export(struct obd_export *exp)
 607 {
 608         struct obd_device *obd = class_exp2obd(exp);
 609         struct client_obd *cli;
 610         struct obd_import *imp;
 611         int rc = 0, err;
 612         ENTRY;
 613
 614         if (!obd) {
 615                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 616                        exp, exp ? exp->exp_handle.h_cookie : -1);
 617                 RETURN(-EINVAL);
 618         }
 619
 620         cli = &obd->u.cli;
 621         imp = cli->cl_import;
 622
 623         down_write(&cli->cl_sem);
 624         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 625                 cli->cl_conn_count);
 626
 627         if (cli->cl_conn_count == 0) {
 628                 CERROR("disconnecting disconnected device (%s)\n",
 629                        obd->obd_name);
 630                 GOTO(out_disconnect, rc = -EINVAL);
 631         }
 632
 633         cli->cl_conn_count--;
 634         if (cli->cl_conn_count != 0)
 635                 GOTO(out_disconnect, rc = 0);
 636
 637         /* Mark import deactivated now, so we don't try to reconnect if any
 638          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 639          * fully deactivate the import, or that would drop all requests. */
 640         spin_lock(&imp->imp_lock);
 641         imp->imp_deactive = 1;
 642         spin_unlock(&imp->imp_lock);
 643
 644         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 645          * delete it regardless.  (It's safe to delete an import that was
 646          * never added.) */
 647         (void)ptlrpc_pinger_del_import(imp);
 648
 649         if (obd->obd_namespace != NULL) {
 650                 /* obd_force == local only */
 651                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 652                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 653                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 654         }
 655
 656         /* There's no need to hold sem while disconnecting an import,
 657          * and it may actually cause deadlock in GSS. */
 658         up_write(&cli->cl_sem);
 659         rc = ptlrpc_disconnect_import(imp, 0);
 660         down_write(&cli->cl_sem);
 661
 662         ptlrpc_invalidate_import(imp);
 663
 664         EXIT;
 665
 666 out_disconnect:
 667         /* Use server style - class_disconnect should be always called for
 668          * o_disconnect. */
 669         err = class_disconnect(exp);
 670         if (!rc && err)
 671                 rc = err;
 672
 673         up_write(&cli->cl_sem);
 674
 675         RETURN(rc);
 676 }
 677 EXPORT_SYMBOL(client_disconnect_export);
 678
 679 #ifdef HAVE_SERVER_SUPPORT
 680 int server_disconnect_export(struct obd_export *exp)
 681 {
 682         int rc;
 683         ENTRY;
 684
 685         /* Disconnect early so that clients can't keep using export. */
 686         rc = class_disconnect(exp);
 687         /* Close import to avoid sending any requests. */
 688         if (exp->exp_imp_reverse)
 689                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 690
 691         ldlm_bl_thread_wakeup();
 692
 693         /* complete all outstanding replies */
 694         spin_lock(&exp->exp_lock);
 695         while (!list_empty(&exp->exp_outstanding_replies)) {
 696                 struct ptlrpc_reply_state *rs =
 697                         list_entry(exp->exp_outstanding_replies.next,
 698                                        struct ptlrpc_reply_state, rs_exp_list);
 699                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 700
 701                 spin_lock(&svcpt->scp_rep_lock);
 702
 703                 list_del_init(&rs->rs_exp_list);
 704                 spin_lock(&rs->rs_lock);
 705                 ptlrpc_schedule_difficult_reply(rs);
 706                 spin_unlock(&rs->rs_lock);
 707
 708                 spin_unlock(&svcpt->scp_rep_lock);
 709         }
 710         spin_unlock(&exp->exp_lock);
 711
 712         RETURN(rc);
 713 }
 714 EXPORT_SYMBOL(server_disconnect_export);
 715
 716 /* --------------------------------------------------------------------------
 717  * from old lib/target.c
 718  * -------------------------------------------------------------------------- */
 719
 720 static int target_handle_reconnect(struct lustre_handle *conn,
 721                                    struct obd_export *exp,
 722                                    struct obd_uuid *cluuid)
 723 {
 724         ENTRY;
 725
 726         if (exp->exp_connection && exp->exp_imp_reverse) {
 727                 struct lustre_handle *hdl;
 728                 struct obd_device *target;
 729
 730                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 731                 target = exp->exp_obd;
 732
 733                 /* Might be a re-connect after a partition. */
 734                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 735                         if (target->obd_recovering) {
 736                                 int timeout = cfs_duration_sec(cfs_time_sub(
 737                                         cfs_timer_deadline(
 738                                         &target->obd_recovery_timer),
 739                                         cfs_time_current()));
 740
 741                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 742                                         "ing, waiting for %d clients in recov"
 743                                         "ery for %d:%.02d\n", target->obd_name,
 744                                         obd_uuid2str(&exp->exp_client_uuid),
 745                                         obd_export_nid2str(exp),
 746                                         target->obd_max_recoverable_clients,
 747                                         timeout / 60, timeout % 60);
 748                         } else {
 749                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 750                                         "reconnecting\n", target->obd_name,
 751                                         obd_uuid2str(&exp->exp_client_uuid),
 752                                         obd_export_nid2str(exp));
 753                         }
 754
 755                         conn->cookie = exp->exp_handle.h_cookie;
 756                         /* target_handle_connect() treats EALREADY and
 757                          * -EALREADY differently.  EALREADY means we are
 758                          * doing a valid reconnect from the same client. */
 759                         RETURN(EALREADY);
 760                 } else {
 761                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 762                                       "with handle "LPX64". Rejecting client "
 763                                       "with the same UUID trying to reconnect "
 764                                       "with handle "LPX64"\n", target->obd_name,
 765                                       obd_uuid2str(&exp->exp_client_uuid),
 766                                       obd_export_nid2str(exp),
 767                                       hdl->cookie, conn->cookie);
 768                         memset(conn, 0, sizeof *conn);
 769                         /* target_handle_connect() treats EALREADY and
 770                          * -EALREADY differently.  -EALREADY is an error
 771                          * (same UUID, different handle). */
 772                         RETURN(-EALREADY);
 773                 }
 774         }
 775
 776         conn->cookie = exp->exp_handle.h_cookie;
 777         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 778                cluuid->uuid, exp, conn->cookie);
 779         RETURN(0);
 780 }
 781
 782 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 783                           int error)
 784 {
 785         struct obd_export *exp = cb_data;
 786
 787         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 788                obd->obd_name, exp->exp_client_uuid.uuid);
 789
 790         spin_lock(&exp->exp_lock);
 791         exp->exp_need_sync = 0;
 792         spin_unlock(&exp->exp_lock);
 793         class_export_cb_put(exp);
 794 }
 795
 796 static void
 797 check_and_start_recovery_timer(struct obd_device *obd,
 798                                struct ptlrpc_request *req, int new_client);
 799
 800 int target_handle_connect(struct ptlrpc_request *req)
 801 {
 802         struct obd_device *target = NULL, *targref = NULL;
 803         struct obd_export *export = NULL;
 804         struct obd_import *revimp;
 805         struct obd_import *tmp_imp = NULL;
 806         struct lustre_handle conn;
 807         struct lustre_handle *tmp;
 808         struct obd_uuid tgtuuid;
 809         struct obd_uuid cluuid;
 810         struct obd_uuid remote_uuid;
 811         char *str;
 812         int rc = 0;
 813         char *target_start;
 814         int target_len;
 815         bool     mds_conn = false, lw_client = false;
 816         bool     mds_mds_conn = false;
 817         bool     new_mds_mds_conn = false;
 818         struct obd_connect_data *data, *tmpdata;
 819         int size, tmpsize;
 820         lnet_nid_t *client_nid = NULL;
 821         ENTRY;
 822
 823         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 824
 825         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 826         if (str == NULL) {
 827                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 828                 GOTO(out, rc = -EINVAL);
 829         }
 830
 831         obd_str2uuid(&tgtuuid, str);
 832         target = class_uuid2obd(&tgtuuid);
 833         if (!target)
 834                 target = class_name2obd(str);
 835
 836         if (!target) {
 837                 deuuidify(str, NULL, &target_start, &target_len);
 838                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 839                                    "from %s (no target). If you are running "
 840                                    "an HA pair check that the target is "
 841                                    "mounted on the other server.\n", str,
 842                                    libcfs_nid2str(req->rq_peer.nid));
 843                 GOTO(out, rc = -ENODEV);
 844         }
 845
 846         spin_lock(&target->obd_dev_lock);
 847         if (target->obd_stopping || !target->obd_set_up) {
 848                 spin_unlock(&target->obd_dev_lock);
 849
 850                 deuuidify(str, NULL, &target_start, &target_len);
 851                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 852                               target_len, target_start,
 853                               libcfs_nid2str(req->rq_peer.nid),
 854                               (target->obd_stopping ?
 855                                "stopping" : "not set up"));
 856                 GOTO(out, rc = -ENODEV);
 857         }
 858
 859         if (target->obd_no_conn) {
 860                 spin_unlock(&target->obd_dev_lock);
 861
 862                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 863                                "from %s\n", target->obd_name,
 864                                libcfs_nid2str(req->rq_peer.nid));
 865                 GOTO(out, rc = -EAGAIN);
 866         }
 867
 868         /* Make sure the target isn't cleaned up while we're here. Yes,
 869          * there's still a race between the above check and our incref here.
 870          * Really, class_uuid2obd should take the ref. */
 871         targref = class_incref(target, __FUNCTION__, current);
 872
 873         target->obd_conn_inprogress++;
 874         spin_unlock(&target->obd_dev_lock);
 875
 876         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 877         if (str == NULL) {
 878                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 879                 GOTO(out, rc = -EINVAL);
 880         }
 881
 882         obd_str2uuid(&cluuid, str);
 883
 884         /* XXX Extract a nettype and format accordingly. */
 885         switch (sizeof(lnet_nid_t)) {
 886         /* NB the casts only avoid compiler warnings. */
 887         case 8:
 888                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 889                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 890                 break;
 891         case 4:
 892                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 893                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 894                 break;
 895         default:
 896                 LBUG();
 897         }
 898
 899         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 900         if (tmp == NULL)
 901                 GOTO(out, rc = -EPROTO);
 902
 903         conn = *tmp;
 904
 905         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 906                                     RCL_CLIENT);
 907         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 908         if (!data)
 909                 GOTO(out, rc = -EPROTO);
 910
 911         rc = req_capsule_server_pack(&req->rq_pill);
 912         if (rc)
 913                 GOTO(out, rc);
 914
 915 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
 916         /* Don't allow clients to connect that are using old 1.8 format
 917          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
 918          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
 919          * FULL20 flag should be set on all connections since 2.0, but no
 920          * longer affects behaviour.
 921          *
 922          * Later this check will be disabled and the flag can be retired
 923          * completely once interop with 3.0 is no longer needed.
 924          */
 925         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
 926                 GOTO(out, rc = -EPROTO);
 927 #endif
 928
 929         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 930                 if (data->ocd_version < LUSTRE_VERSION_CODE -
 931                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 932                     data->ocd_version > LUSTRE_VERSION_CODE +
 933                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 934                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 935                                   "libclient connection attempt",
 936                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 937                                   "old" : "new",
 938                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 939                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 940                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 941                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 942                         data = req_capsule_server_sized_get(&req->rq_pill,
 943                                                             &RMF_CONNECT_DATA,
 944                                     offsetof(typeof(*data), ocd_version) +
 945                                              sizeof(data->ocd_version));
 946                         if (data) {
 947                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 948                                 data->ocd_version = LUSTRE_VERSION_CODE;
 949                         }
 950                         GOTO(out, rc = -EPROTO);
 951                 }
 952         }
 953
 954         /* Note: lw_client is needed in MDS-MDS failover during update log
 955          * processing, so we needs to allow lw_client to be connected at
 956          * anytime, instead of only the initial connection */
 957         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
 958
 959         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
 960                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
 961                 mds_mds_conn = (data->ocd_connect_flags &
 962                                 OBD_CONNECT_MDS_MDS) != 0;
 963
 964                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
 965                  * for Imperative Recovery connection from MGC to MGS.
 966                  *
 967                  * Via check OBD_CONNECT_FID, we can distinguish whether
 968                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
 969                  * MGC or MDT. */
 970                 if (!lw_client &&
 971                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
 972                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
 973                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
 974                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
 975                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
 976                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
 977
 978                         /* We do not support the MDT-MDT interoperations with
 979                          * different version MDT because of protocol changes. */
 980                         if (unlikely(major != LUSTRE_MAJOR ||
 981                                      minor != LUSTRE_MINOR ||
 982                                      abs(patch - LUSTRE_PATCH) > 3)) {
 983                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
 984                                         "connection from different version MDT "
 985                                         "(%d.%d.%d.%d) %s %s\n",
 986                                         target->obd_name, LUSTRE_MAJOR,
 987                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
 988                                         major, minor, patch,
 989                                         OBD_OCD_VERSION_FIX(data->ocd_version),
 990                                         libcfs_nid2str(req->rq_peer.nid), str);
 991
 992                                 GOTO(out, rc = -EPROTO);
 993                         }
 994                 }
 995         }
 996
 997         /* lctl gets a backstage, all-access pass. */
 998         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 999                 goto dont_check_exports;
1000
1001         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1002         if (!export)
1003                 goto no_export;
1004
1005         /* We've found an export in the hash. */
1006
1007         spin_lock(&export->exp_lock);
1008
1009         if (export->exp_connecting) { /* bug 9635, et. al. */
1010                 spin_unlock(&export->exp_lock);
1011                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1012                               export->exp_obd->obd_name, export,
1013                               libcfs_nid2str(req->rq_peer.nid));
1014                 class_export_put(export);
1015                 export = NULL;
1016                 rc = -EALREADY;
1017         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1018                 spin_unlock(&export->exp_lock);
1019                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1020                         /* MDS or LWP reconnected after failover. */
1021                         LCONSOLE_WARN("%s: Received %s connection from "
1022                             "%s, removing former export from %s\n",
1023                             target->obd_name, mds_conn ? "MDS" : "LWP",
1024                             libcfs_nid2str(req->rq_peer.nid),
1025                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1026                 else
1027                         /* New MDS connection from the same NID. */
1028                         LCONSOLE_WARN("%s: Received new %s connection from "
1029                                 "%s, removing former export from same NID\n",
1030                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1031                                 libcfs_nid2str(req->rq_peer.nid));
1032                 class_fail_export(export);
1033                 class_export_put(export);
1034                 export = NULL;
1035                 rc = 0;
1036         } else if (export->exp_connection != NULL &&
1037                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1038                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1039                     MSG_CONNECT_INITIAL)) {
1040                 spin_unlock(&export->exp_lock);
1041                 /* In MDS failover we have static UUID but NID can change. */
1042                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1043                               "existing nid %s is already connected\n",
1044                               target->obd_name, cluuid.uuid,
1045                               libcfs_nid2str(req->rq_peer.nid),
1046                               libcfs_nid2str(
1047                                       export->exp_connection->c_peer.nid));
1048                 rc = -EALREADY;
1049                 class_export_put(export);
1050                 export = NULL;
1051         } else {
1052                 export->exp_connecting = 1;
1053                 spin_unlock(&export->exp_lock);
1054                 LASSERT(export->exp_obd == target);
1055
1056                 rc = target_handle_reconnect(&conn, export, &cluuid);
1057         }
1058
1059         /* If we found an export, we already unlocked. */
1060         if (!export) {
1061 no_export:
1062                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1063         } else if (req->rq_export == NULL &&
1064                    atomic_read(&export->exp_rpc_count) > 0) {
1065                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1066                               "still busy with %d references\n",
1067                               target->obd_name, cluuid.uuid,
1068                               libcfs_nid2str(req->rq_peer.nid),
1069                               atomic_read(&export->exp_refcount));
1070                 GOTO(out, rc = -EBUSY);
1071         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1072                 if (!strstr(cluuid.uuid, "mdt"))
1073                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1074                                       "known client %s (at %s) because it "
1075                                       "is indicating it is a new client",
1076                                       target->obd_name, cluuid.uuid,
1077                                       libcfs_nid2str(req->rq_peer.nid));
1078                 GOTO(out, rc = -EALREADY);
1079         } else {
1080                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1081         }
1082
1083         if (rc < 0) {
1084                 GOTO(out, rc);
1085         }
1086
1087         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1088                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1089               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1090               export, (long)cfs_time_current_sec(),
1091               export ? (long)export->exp_last_request_time : 0);
1092
1093         /* If this is the first time a client connects, reset the recovery
1094          * timer. Discard lightweight connections which might be local. */
1095         if (!lw_client && rc == 0 && target->obd_recovering)
1096                 check_and_start_recovery_timer(target, req, export == NULL);
1097
1098         /* We want to handle EALREADY but *not* -EALREADY from
1099          * target_handle_reconnect(), return reconnection state in a flag. */
1100         if (rc == EALREADY) {
1101                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1102                 rc = 0;
1103         } else {
1104                 LASSERT(rc == 0);
1105         }
1106
1107         /* Tell the client if we support replayable requests. */
1108         if (target->obd_replayable)
1109                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1110         client_nid = &req->rq_peer.nid;
1111
1112         if (export == NULL) {
1113                 /* allow lightweight connections during recovery */
1114                 /* allow "new" MDT to be connected during recovery, since we
1115                  * need retrieve recovery update records from it */
1116                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1117                         cfs_time_t t;
1118                         int     c; /* connected */
1119                         int     i; /* in progress */
1120                         int     k; /* known */
1121                         int     s; /* stale/evicted */
1122
1123                         c = atomic_read(&target->obd_connected_clients);
1124                         i = atomic_read(&target->obd_lock_replay_clients);
1125                         k = target->obd_max_recoverable_clients;
1126                         s = target->obd_stale_clients;
1127                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1128                         t = cfs_time_sub(t, cfs_time_current());
1129                         t = cfs_duration_sec(t);
1130                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1131                                       "(at %s), waiting for %d known clients "
1132                                       "(%d recovered, %d in progress, and %d "
1133                                       "evicted) to recover in %d:%.02d\n",
1134                                       target->obd_name, cluuid.uuid,
1135                                       libcfs_nid2str(req->rq_peer.nid), k,
1136                                       c - i, i, s, (int)t / 60,
1137                                       (int)t % 60);
1138                         rc = -EBUSY;
1139                 } else {
1140 dont_check_exports:
1141                         rc = obd_connect(req->rq_svc_thread->t_env,
1142                                          &export, target, &cluuid, data,
1143                                          client_nid);
1144                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1145                                 lustre_msg_add_op_flags(req->rq_repmsg,
1146                                                         MSG_CONNECT_RECOVERING);
1147                         if (rc == 0)
1148                                 conn.cookie = export->exp_handle.h_cookie;
1149
1150                         if (mds_mds_conn)
1151                                 new_mds_mds_conn = true;
1152                 }
1153         } else {
1154                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1155                                    export, target, &cluuid, data, client_nid);
1156         }
1157         if (rc)
1158                 GOTO(out, rc);
1159
1160         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1161         data->ocd_instance = target->u.obt.obt_instance;
1162
1163         /* Return only the parts of obd_connect_data that we understand, so the
1164          * client knows that we don't understand the rest. */
1165         if (data) {
1166                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1167                                                RCL_SERVER);
1168                 tmpdata = req_capsule_server_get(&req->rq_pill,
1169                                                  &RMF_CONNECT_DATA);
1170                 /* Don't use struct assignment here, because the client reply
1171                  * buffer may be smaller/larger than the local struct
1172                  * obd_connect_data. */
1173                 memcpy(tmpdata, data, min(tmpsize, size));
1174         }
1175
1176         /* If all else goes well, this is our RPC return code. */
1177         req->rq_status = 0;
1178
1179         lustre_msg_set_handle(req->rq_repmsg, &conn);
1180
1181         /* If the client and the server are the same node, we will already
1182          * have an export that really points to the client's DLM export,
1183          * because we have a shared handles table.
1184          *
1185          * XXX this will go away when shaver stops sending the "connect" handle
1186          * in the real "remote handle" field of the request --phik 24 Apr 2003
1187          */
1188         ptlrpc_request_change_export(req, export);
1189
1190         spin_lock(&export->exp_lock);
1191         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1192                 spin_unlock(&export->exp_lock);
1193                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1194                        "or equal conn_cnt: %d >= %d\n",
1195                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1196                        export->exp_conn_cnt,
1197                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1198
1199                 GOTO(out, rc = -EALREADY);
1200         }
1201         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1202         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1203
1204         /* Don't evict liblustre clients for not pinging. */
1205         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1206                 export->exp_libclient = 1;
1207                 spin_unlock(&export->exp_lock);
1208
1209                 spin_lock(&target->obd_dev_lock);
1210                 list_del_init(&export->exp_obd_chain_timed);
1211                 spin_unlock(&target->obd_dev_lock);
1212         } else {
1213                 spin_unlock(&export->exp_lock);
1214         }
1215
1216         if (export->exp_connection != NULL) {
1217                 /* Check to see if connection came from another NID. */
1218                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1219                     !hlist_unhashed(&export->exp_nid_hash))
1220                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1221                                      &export->exp_connection->c_peer.nid,
1222                                      &export->exp_nid_hash);
1223
1224                 ptlrpc_connection_put(export->exp_connection);
1225         }
1226
1227         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1228                                                        req->rq_self,
1229                                                        &remote_uuid);
1230         if (hlist_unhashed(&export->exp_nid_hash)) {
1231                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1232                              &export->exp_connection->c_peer.nid,
1233                              &export->exp_nid_hash);
1234         }
1235
1236         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1237                 int has_transno;
1238                 __u64 transno = data->ocd_transno;
1239
1240                 spin_lock(&export->exp_lock);
1241                 /* possible race with class_disconnect_stale_exports,
1242                  * export may be already in the eviction process */
1243                 if (export->exp_failed) {
1244                         spin_unlock(&export->exp_lock);
1245                         GOTO(out, rc = -ENODEV);
1246                 }
1247                 export->exp_in_recovery = 1;
1248                 export->exp_req_replay_needed = 1;
1249                 export->exp_lock_replay_needed = 1;
1250                 spin_unlock(&export->exp_lock);
1251
1252                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1253                                  MSG_CONNECT_TRANSNO);
1254                 if (has_transno && transno == 0)
1255                         CWARN("Connect with zero transno!\n");
1256
1257                 if (has_transno && transno > 0 &&
1258                     transno < target->obd_next_recovery_transno &&
1259                     transno > target->obd_last_committed) {
1260                         /* Another way is to use cmpxchg() to be lock-free. */
1261                         spin_lock(&target->obd_recovery_task_lock);
1262                         if (transno < target->obd_next_recovery_transno)
1263                                 target->obd_next_recovery_transno = transno;
1264                         spin_unlock(&target->obd_recovery_task_lock);
1265                 }
1266
1267                 atomic_inc(&target->obd_req_replay_clients);
1268                 atomic_inc(&target->obd_lock_replay_clients);
1269                 /* Note: MDS-MDS connection is allowed to be connected during
1270                  * recovery, no matter if the exports needs to be recoveried.
1271                  * Because we need retrieve updates logs from all other MDTs.
1272                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1273                  * also needs to be increased to match other recovery checking
1274                  * condition. */
1275                 if (new_mds_mds_conn)
1276                         target->obd_max_recoverable_clients++;
1277                 if (atomic_inc_return(&target->obd_connected_clients) ==
1278                     target->obd_max_recoverable_clients)
1279                         wake_up(&target->obd_next_transno_waitq);
1280         }
1281
1282         /* Tell the client we're in recovery, when client is involved in it. */
1283         if (target->obd_recovering && !lw_client)
1284                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1285
1286         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1287         conn = *tmp;
1288
1289         /* Return -ENOTCONN in case of errors to let client reconnect. */
1290         revimp = class_new_import(target);
1291         if (revimp == NULL) {
1292                 CERROR("fail to alloc new reverse import.\n");
1293                 GOTO(out, rc = -ENOTCONN);
1294         }
1295
1296         spin_lock(&export->exp_lock);
1297         if (export->exp_imp_reverse != NULL)
1298                 /* destroyed import can be still referenced in ctxt */
1299                 tmp_imp = export->exp_imp_reverse;
1300         export->exp_imp_reverse = revimp;
1301         spin_unlock(&export->exp_lock);
1302
1303         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1304         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1305         revimp->imp_remote_handle = conn;
1306         revimp->imp_dlm_fake = 1;
1307         revimp->imp_state = LUSTRE_IMP_FULL;
1308
1309         /* Unknown versions will be caught in
1310          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1311         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1312
1313         if (data->ocd_connect_flags & OBD_CONNECT_AT)
1314                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1315         else
1316                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1317
1318         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1319
1320         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1321         if (rc) {
1322                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1323                 spin_lock(&export->exp_lock);
1324                 export->exp_imp_reverse = NULL;
1325                 spin_unlock(&export->exp_lock);
1326                 class_destroy_import(revimp);
1327         }
1328
1329         class_import_put(revimp);
1330
1331 out:
1332         if (tmp_imp != NULL)
1333                 client_destroy_import(tmp_imp);
1334         if (export) {
1335                 spin_lock(&export->exp_lock);
1336                 export->exp_connecting = 0;
1337                 spin_unlock(&export->exp_lock);
1338
1339                 class_export_put(export);
1340         }
1341         if (targref) {
1342                 spin_lock(&target->obd_dev_lock);
1343                 target->obd_conn_inprogress--;
1344                 spin_unlock(&target->obd_dev_lock);
1345
1346                 class_decref(targref, __func__, current);
1347         }
1348         if (rc)
1349                 req->rq_status = rc;
1350         RETURN(rc);
1351 }
1352
1353 int target_handle_disconnect(struct ptlrpc_request *req)
1354 {
1355         int rc;
1356         ENTRY;
1357
1358         rc = req_capsule_server_pack(&req->rq_pill);
1359         if (rc)
1360                 RETURN(rc);
1361
1362         /* Keep the rq_export around so we can send the reply. */
1363         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1364
1365         RETURN(0);
1366 }
1367
1368 void target_destroy_export(struct obd_export *exp)
1369 {
1370         struct obd_import       *imp = NULL;
1371         /* exports created from last_rcvd data, and "fake"
1372            exports created by lctl don't have an import */
1373         spin_lock(&exp->exp_lock);
1374         if (exp->exp_imp_reverse != NULL) {
1375                 imp = exp->exp_imp_reverse;
1376                 exp->exp_imp_reverse = NULL;
1377         }
1378         spin_unlock(&exp->exp_lock);
1379         if (imp != NULL)
1380                 client_destroy_import(imp);
1381
1382         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1383         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1384         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1385         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1386 }
1387 EXPORT_SYMBOL(target_destroy_export);
1388
1389 /*
1390  * Recovery functions
1391  */
1392 static void target_request_copy_get(struct ptlrpc_request *req)
1393 {
1394         class_export_rpc_inc(req->rq_export);
1395         LASSERT(list_empty(&req->rq_list));
1396         INIT_LIST_HEAD(&req->rq_replay_list);
1397
1398         /* Increase refcount to keep request in queue. */
1399         atomic_inc(&req->rq_refcount);
1400         /* Let export know it has replays to be handled. */
1401         atomic_inc(&req->rq_export->exp_replay_count);
1402 }
1403
1404 static void target_request_copy_put(struct ptlrpc_request *req)
1405 {
1406         LASSERT(list_empty(&req->rq_replay_list));
1407         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1408
1409         atomic_dec(&req->rq_export->exp_replay_count);
1410         class_export_rpc_dec(req->rq_export);
1411         ptlrpc_server_drop_request(req);
1412 }
1413
1414 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1415 {
1416         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1417         struct obd_export     *exp = req->rq_export;
1418         struct ptlrpc_request *reqiter;
1419         struct ptlrpc_request *dup_req = NULL;
1420         int                    dup = 0;
1421
1422         LASSERT(exp);
1423
1424         spin_lock(&exp->exp_lock);
1425         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1426                                 rq_replay_list) {
1427                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1428                         dup_req = reqiter;
1429                         dup = 1;
1430                         break;
1431                 }
1432         }
1433
1434         if (dup) {
1435                 /* We expect it with RESENT and REPLAY flags. */
1436                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1437                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1438                         CERROR("invalid flags %x of resent replay\n",
1439                                lustre_msg_get_flags(req->rq_reqmsg));
1440
1441                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1442                         __u32 new_conn;
1443
1444                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1445                         if (new_conn >
1446                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1447                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1448                                                         new_conn);
1449                 }
1450         } else {
1451                 list_add_tail(&req->rq_replay_list,
1452                                   &exp->exp_req_replay_queue);
1453         }
1454
1455         spin_unlock(&exp->exp_lock);
1456         return dup;
1457 }
1458
1459 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1460 {
1461         LASSERT(!list_empty(&req->rq_replay_list));
1462         LASSERT(req->rq_export);
1463
1464         spin_lock(&req->rq_export->exp_lock);
1465         list_del_init(&req->rq_replay_list);
1466         spin_unlock(&req->rq_export->exp_lock);
1467 }
1468
1469 static void target_finish_recovery(struct lu_target *lut)
1470 {
1471         struct obd_device *obd = lut->lut_obd;
1472         ENTRY;
1473
1474         /* Only log a recovery message when recovery has occurred. */
1475         if (obd->obd_recovery_start) {
1476                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1477                                         obd->obd_recovery_start);
1478                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1479                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1480                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1481                         obd->obd_max_recoverable_clients,
1482                         atomic_read(&obd->obd_connected_clients),
1483                         obd->obd_stale_clients,
1484                         obd->obd_stale_clients == 1 ? "was" : "were");
1485         }
1486
1487         ldlm_reprocess_all_ns(obd->obd_namespace);
1488         spin_lock(&obd->obd_recovery_task_lock);
1489         if (!list_empty(&obd->obd_req_replay_queue) ||
1490             !list_empty(&obd->obd_lock_replay_queue) ||
1491             !list_empty(&obd->obd_final_req_queue)) {
1492                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1493                        obd->obd_name,
1494                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1495                        list_empty(&obd->obd_lock_replay_queue) ? \
1496                                "" : "lock ",
1497                        list_empty(&obd->obd_final_req_queue) ? \
1498                                "" : "final ");
1499                 spin_unlock(&obd->obd_recovery_task_lock);
1500                 LBUG();
1501         }
1502         spin_unlock(&obd->obd_recovery_task_lock);
1503
1504         if (lut->lut_tdtd != NULL &&
1505             (!list_empty(&lut->lut_tdtd->tdtd_replay_list) ||
1506             !list_empty(&lut->lut_tdtd->tdtd_replay_finish_list))) {
1507                 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1508                 dtrq_list_destroy(lut->lut_tdtd);
1509         }
1510
1511         obd->obd_recovery_end = cfs_time_current_sec();
1512
1513         /* When recovery finished, cleanup orphans on MDS and OST. */
1514         if (OBT(obd) && OBP(obd, postrecov)) {
1515                 int rc = OBP(obd, postrecov)(obd);
1516                 if (rc < 0)
1517                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1518                                       obd->obd_name, rc);
1519         }
1520         EXIT;
1521 }
1522
1523 static void abort_req_replay_queue(struct obd_device *obd)
1524 {
1525         struct ptlrpc_request *req, *n;
1526         struct list_head abort_list;
1527
1528         INIT_LIST_HEAD(&abort_list);
1529         spin_lock(&obd->obd_recovery_task_lock);
1530         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1531         spin_unlock(&obd->obd_recovery_task_lock);
1532         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1533                 DEBUG_REQ(D_WARNING, req, "aborted:");
1534                 req->rq_status = -ENOTCONN;
1535                 if (ptlrpc_error(req)) {
1536                         DEBUG_REQ(D_ERROR, req,
1537                                   "failed abort_req_reply; skipping");
1538                 }
1539                 target_exp_dequeue_req_replay(req);
1540                 target_request_copy_put(req);
1541         }
1542 }
1543
1544 static void abort_lock_replay_queue(struct obd_device *obd)
1545 {
1546         struct ptlrpc_request *req, *n;
1547         struct list_head abort_list;
1548
1549         INIT_LIST_HEAD(&abort_list);
1550         spin_lock(&obd->obd_recovery_task_lock);
1551         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1552         spin_unlock(&obd->obd_recovery_task_lock);
1553         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1554                 DEBUG_REQ(D_ERROR, req, "aborted:");
1555                 req->rq_status = -ENOTCONN;
1556                 if (ptlrpc_error(req)) {
1557                         DEBUG_REQ(D_ERROR, req,
1558                                   "failed abort_lock_reply; skipping");
1559                 }
1560                 target_request_copy_put(req);
1561         }
1562 }
1563
1564 /* Called from a cleanup function if the device is being cleaned up
1565    forcefully.  The exports should all have been disconnected already,
1566    the only thing left to do is
1567      - clear the recovery flags
1568      - cancel the timer
1569      - free queued requests and replies, but don't send replies
1570    Because the obd_stopping flag is set, no new requests should be received.
1571
1572 */
1573 void target_cleanup_recovery(struct obd_device *obd)
1574 {
1575         struct ptlrpc_request *req, *n;
1576         struct list_head clean_list;
1577
1578         INIT_LIST_HEAD(&clean_list);
1579         spin_lock(&obd->obd_dev_lock);
1580         if (!obd->obd_recovering) {
1581                 spin_unlock(&obd->obd_dev_lock);
1582                 EXIT;
1583                 return;
1584         }
1585         obd->obd_recovering = obd->obd_abort_recovery = 0;
1586         obd->obd_force_abort_recovery = 0;
1587         spin_unlock(&obd->obd_dev_lock);
1588
1589         spin_lock(&obd->obd_recovery_task_lock);
1590         target_cancel_recovery_timer(obd);
1591         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1592         spin_unlock(&obd->obd_recovery_task_lock);
1593
1594         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1595                 LASSERT(req->rq_reply_state == NULL);
1596                 target_exp_dequeue_req_replay(req);
1597                 target_request_copy_put(req);
1598         }
1599
1600         spin_lock(&obd->obd_recovery_task_lock);
1601         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1602         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1603         spin_unlock(&obd->obd_recovery_task_lock);
1604
1605         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1606                 LASSERT(req->rq_reply_state == NULL);
1607                 target_request_copy_put(req);
1608         }
1609
1610         EXIT;
1611 }
1612 EXPORT_SYMBOL(target_cleanup_recovery);
1613
1614 /* obd_recovery_task_lock should be held */
1615 void target_cancel_recovery_timer(struct obd_device *obd)
1616 {
1617         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1618         cfs_timer_disarm(&obd->obd_recovery_timer);
1619 }
1620
1621 static void target_start_recovery_timer(struct obd_device *obd)
1622 {
1623         if (obd->obd_recovery_start != 0)
1624                 return;
1625
1626         spin_lock(&obd->obd_dev_lock);
1627         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1628             obd->obd_force_abort_recovery) {
1629                 spin_unlock(&obd->obd_dev_lock);
1630                 return;
1631         }
1632
1633         LASSERT(obd->obd_recovery_timeout != 0);
1634
1635         if (obd->obd_recovery_start != 0) {
1636                 spin_unlock(&obd->obd_dev_lock);
1637                 return;
1638         }
1639
1640         cfs_timer_arm(&obd->obd_recovery_timer,
1641                       cfs_time_shift(obd->obd_recovery_timeout));
1642         obd->obd_recovery_start = cfs_time_current_sec();
1643         spin_unlock(&obd->obd_dev_lock);
1644
1645         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1646                       "or until %d client%s reconnect%s\n",
1647                       obd->obd_name,
1648                       obd->obd_recovery_timeout / 60,
1649                       obd->obd_recovery_timeout % 60,
1650                       obd->obd_max_recoverable_clients,
1651                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1652                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1653 }
1654
1655 /**
1656  * extend recovery window.
1657  *
1658  * if @extend is true, extend recovery window to have @drt remaining at least;
1659  * otherwise, make sure the recovery timeout value is not less than @drt.
1660  */
1661 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1662 {
1663         cfs_time_t now;
1664         cfs_time_t end;
1665         cfs_duration_t left;
1666         int to;
1667
1668         spin_lock(&obd->obd_dev_lock);
1669         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1670             obd->obd_force_abort_recovery) {
1671                 spin_unlock(&obd->obd_dev_lock);
1672                 return;
1673         }
1674         LASSERT(obd->obd_recovery_start != 0);
1675
1676         now  = cfs_time_current_sec();
1677         to   = obd->obd_recovery_timeout;
1678         end  = obd->obd_recovery_start + to;
1679         left = cfs_time_sub(end, now);
1680
1681         if (extend && (drt > left)) {
1682                 to += drt - left;
1683         } else if (!extend && (drt > to)) {
1684                 to = drt;
1685         }
1686
1687         if (to > obd->obd_recovery_time_hard)
1688                 to = obd->obd_recovery_time_hard;
1689         if (obd->obd_recovery_timeout < to) {
1690                 obd->obd_recovery_timeout = to;
1691                 end = obd->obd_recovery_start + to;
1692                 cfs_timer_arm(&obd->obd_recovery_timer,
1693                                 cfs_time_shift(end - now));
1694         }
1695         spin_unlock(&obd->obd_dev_lock);
1696
1697         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1698                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1699 }
1700
1701 /* Reset the timer with each new client connection */
1702 /*
1703  * This timer is actually reconnect_timer, which is for making sure
1704  * the total recovery window is at least as big as my reconnect
1705  * attempt timing. So the initial recovery time_out will be set to
1706  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1707  * from client is bigger than this, then the recovery time_out will
1708  * be extended to make sure the client could be reconnected, in the
1709  * process, the timeout from the new client should be ignored.
1710  */
1711
1712 static void
1713 check_and_start_recovery_timer(struct obd_device *obd,
1714                                struct ptlrpc_request *req,
1715                                int new_client)
1716 {
1717         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1718         struct obd_device_target *obt = &obd->u.obt;
1719
1720         if (!new_client && service_time)
1721                 /* Teach server about old server's estimates, as first guess
1722                  * at how long new requests will take. */
1723                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1724                             service_time);
1725
1726         target_start_recovery_timer(obd);
1727
1728         /* Convert the service time to RPC timeout,
1729          * and reuse service_time to limit stack usage. */
1730         service_time = at_est2timeout(service_time);
1731
1732         /* We expect other clients to timeout within service_time, then try
1733          * to reconnect, then try the failover server.  The max delay between
1734          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1735         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1736
1737         LASSERT(obt->obt_magic == OBT_MAGIC);
1738         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1739         if (service_time > obd->obd_recovery_timeout && !new_client)
1740                 extend_recovery_timer(obd, service_time, false);
1741 }
1742
1743 /** Health checking routines */
1744 static inline int exp_connect_healthy(struct obd_export *exp)
1745 {
1746         return (exp->exp_in_recovery);
1747 }
1748
1749 /** if export done req_replay or has replay in queue */
1750 static inline int exp_req_replay_healthy(struct obd_export *exp)
1751 {
1752         return (!exp->exp_req_replay_needed ||
1753                 atomic_read(&exp->exp_replay_count) > 0);
1754 }
1755 /** if export done lock_replay or has replay in queue */
1756 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1757 {
1758         return (!exp->exp_lock_replay_needed ||
1759                 atomic_read(&exp->exp_replay_count) > 0);
1760 }
1761
1762 static inline int exp_vbr_healthy(struct obd_export *exp)
1763 {
1764         return (!exp->exp_vbr_failed);
1765 }
1766
1767 static inline int exp_finished(struct obd_export *exp)
1768 {
1769         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1770 }
1771
1772 static int check_for_next_transno(struct lu_target *lut)
1773 {
1774         struct ptlrpc_request *req = NULL;
1775         struct obd_device *obd = lut->lut_obd;
1776         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1777         int wake_up = 0, connected, completed, queue_len;
1778         __u64 req_transno = 0;
1779         __u64 update_transno = 0;
1780         __u64 next_transno = 0;
1781         ENTRY;
1782
1783         spin_lock(&obd->obd_recovery_task_lock);
1784         if (!list_empty(&obd->obd_req_replay_queue)) {
1785                 req = list_entry(obd->obd_req_replay_queue.next,
1786                                      struct ptlrpc_request, rq_list);
1787                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1788         }
1789
1790         if (tdtd != NULL)
1791                 update_transno = distribute_txn_get_next_transno(tdtd);
1792
1793         connected = atomic_read(&obd->obd_connected_clients);
1794         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1795         queue_len = obd->obd_requests_queued_for_recovery;
1796         next_transno = obd->obd_next_recovery_transno;
1797
1798         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1799                "req_transno: "LPU64", next_transno: "LPU64"\n",
1800                obd->obd_max_recoverable_clients, connected, completed,
1801                queue_len, req_transno, next_transno);
1802
1803         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1804                 CDEBUG(D_HA, "waking for aborted recovery\n");
1805                 wake_up = 1;
1806         } else if (obd->obd_recovery_expired) {
1807                 CDEBUG(D_HA, "waking for expired recovery\n");
1808                 wake_up = 1;
1809         } else if (tdtd != NULL && req != NULL &&
1810                    is_req_replayed_by_update(req)) {
1811                 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1812                          "next_transno"LPU64"\n", req_transno, next_transno);
1813                 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1814                        req_transno);
1815                 wake_up = 1;
1816         } else if (req_transno == next_transno ||
1817                    (update_transno != 0 && update_transno <= next_transno)) {
1818                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1819                 wake_up = 1;
1820         } else if (queue_len > 0 &&
1821                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1822                 int d_lvl = D_HA;
1823                 /** handle gaps occured due to lost reply or VBR */
1824                 LASSERTF(req_transno >= next_transno,
1825                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1826                          req_transno, next_transno);
1827                 if (req_transno > obd->obd_last_committed &&
1828                     !obd->obd_version_recov)
1829                         d_lvl = D_ERROR;
1830                 CDEBUG(d_lvl,
1831                        "%s: waking for gap in transno, VBR is %s (skip: "
1832                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1833                        ", next_update "LPD64" last_committed: "LPD64")\n",
1834                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1835                        next_transno, queue_len, completed, connected,
1836                        req_transno, update_transno, obd->obd_last_committed);
1837                 obd->obd_next_recovery_transno = req_transno;
1838                 wake_up = 1;
1839         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1840                 CDEBUG(D_HA, "waking for completed recovery\n");
1841                 wake_up = 1;
1842         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1843                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1844                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1845                 obd->obd_next_recovery_transno = req_transno;
1846                 wake_up = 1;
1847         }
1848         spin_unlock(&obd->obd_recovery_task_lock);
1849         return wake_up;
1850 }
1851
1852 static int check_for_next_lock(struct lu_target *lut)
1853 {
1854         struct obd_device *obd = lut->lut_obd;
1855         int wake_up = 0;
1856
1857         spin_lock(&obd->obd_recovery_task_lock);
1858         if (!list_empty(&obd->obd_lock_replay_queue)) {
1859                 CDEBUG(D_HA, "waking for next lock\n");
1860                 wake_up = 1;
1861         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1862                 CDEBUG(D_HA, "waking for completed lock replay\n");
1863                 wake_up = 1;
1864         } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1865                 CDEBUG(D_HA, "waking for aborted recovery\n");
1866                 wake_up = 1;
1867         } else if (obd->obd_recovery_expired) {
1868                 CDEBUG(D_HA, "waking for expired recovery\n");
1869                 wake_up = 1;
1870         }
1871         spin_unlock(&obd->obd_recovery_task_lock);
1872
1873         return wake_up;
1874 }
1875
1876 /**
1877  * wait for recovery events,
1878  * check its status with help of check_routine
1879  * evict dead clients via health_check
1880  */
1881 static int target_recovery_overseer(struct lu_target *lut,
1882                                     int (*check_routine)(struct lu_target *),
1883                                     int (*health_check)(struct obd_export *))
1884 {
1885         struct obd_device       *obd = lut->lut_obd;
1886 repeat:
1887         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1888               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1889                 CWARN("recovery is aborted by hard timeout\n");
1890                 obd->obd_abort_recovery = 1;
1891         }
1892
1893         while (wait_event_timeout(obd->obd_next_transno_waitq,
1894                                   check_routine(lut),
1895                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1896                 /* wait indefinitely for event, but don't trigger watchdog */;
1897
1898         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1899                 CWARN("recovery is aborted, evict exports in recovery\n");
1900                 /** evict exports which didn't finish recovery yet */
1901                 class_disconnect_stale_exports(obd, exp_finished);
1902                 return 1;
1903         } else if (obd->obd_recovery_expired) {
1904                 obd->obd_recovery_expired = 0;
1905                 /** If some clients died being recovered, evict them */
1906                 LCONSOLE_WARN("%s: recovery is timed out, "
1907                               "evict stale exports\n", obd->obd_name);
1908                 /** evict cexports with no replay in queue, they are stalled */
1909                 class_disconnect_stale_exports(obd, health_check);
1910                 /** continue with VBR */
1911                 spin_lock(&obd->obd_dev_lock);
1912                 obd->obd_version_recov = 1;
1913                 spin_unlock(&obd->obd_dev_lock);
1914                 /**
1915                  * reset timer, recovery will proceed with versions now,
1916                  * timeout is set just to handle reconnection delays
1917                  */
1918                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1919                 /** Wait for recovery events again, after evicting bad clients */
1920                 goto repeat;
1921         }
1922         return 0;
1923 }
1924
1925 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1926 {
1927         struct obd_device       *obd = lut->lut_obd;
1928         struct ptlrpc_request *req = NULL;
1929
1930         CDEBUG(D_HA, "Waiting for lock\n");
1931         if (target_recovery_overseer(lut, check_for_next_lock,
1932                                      exp_lock_replay_healthy))
1933                 abort_lock_replay_queue(obd);
1934
1935         spin_lock(&obd->obd_recovery_task_lock);
1936         if (!list_empty(&obd->obd_lock_replay_queue)) {
1937                 req = list_entry(obd->obd_lock_replay_queue.next,
1938                                      struct ptlrpc_request, rq_list);
1939                 list_del_init(&req->rq_list);
1940                 spin_unlock(&obd->obd_recovery_task_lock);
1941         } else {
1942                 spin_unlock(&obd->obd_recovery_task_lock);
1943                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1944                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1945                 /** evict exports failed VBR */
1946                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1947         }
1948         return req;
1949 }
1950
1951 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1952 {
1953         struct ptlrpc_request *req = NULL;
1954
1955         spin_lock(&obd->obd_recovery_task_lock);
1956         if (!list_empty(&obd->obd_final_req_queue)) {
1957                 req = list_entry(obd->obd_final_req_queue.next,
1958                                      struct ptlrpc_request, rq_list);
1959                 list_del_init(&req->rq_list);
1960                 spin_unlock(&obd->obd_recovery_task_lock);
1961                 if (req->rq_export->exp_in_recovery) {
1962                         spin_lock(&req->rq_export->exp_lock);
1963                         req->rq_export->exp_in_recovery = 0;
1964                         spin_unlock(&req->rq_export->exp_lock);
1965                 }
1966         } else {
1967                 spin_unlock(&obd->obd_recovery_task_lock);
1968         }
1969         return req;
1970 }
1971
1972 static void handle_recovery_req(struct ptlrpc_thread *thread,
1973                                 struct ptlrpc_request *req,
1974                                 svc_handler_t handler)
1975 {
1976         ENTRY;
1977
1978         /**
1979          * export can be evicted during recovery, no need to handle replays for
1980          * it after that, discard such request silently
1981          */
1982         if (req->rq_export->exp_disconnected)
1983                 RETURN_EXIT;
1984
1985         req->rq_session.lc_thread = thread;
1986         req->rq_svc_thread = thread;
1987         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1988
1989         /* thread context */
1990         lu_context_enter(&thread->t_env->le_ctx);
1991         (void)handler(req);
1992         lu_context_exit(&thread->t_env->le_ctx);
1993
1994         /* don't reset timer for final stage */
1995         if (!exp_finished(req->rq_export)) {
1996                 int to = obd_timeout;
1997
1998                 /**
1999                  * Add request timeout to the recovery time so next request from
2000                  * this client may come in recovery time
2001                  */
2002                 if (!AT_OFF) {
2003                         struct ptlrpc_service_part *svcpt;
2004
2005                         svcpt = req->rq_rqbd->rqbd_svcpt;
2006                         /* If the server sent early reply for this request,
2007                          * the client will recalculate the timeout according to
2008                          * current server estimate service time, so we will
2009                          * use the maxium timeout here for waiting the client
2010                          * sending the next req */
2011                         to = max((int)at_est2timeout(
2012                                  at_get(&svcpt->scp_at_estimate)),
2013                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2014                         /* Add 2 net_latency, one for balance rq_deadline
2015                          * (see ptl_send_rpc), one for resend the req to server,
2016                          * Note: client will pack net_latency in replay req
2017                          * (see ptlrpc_replay_req) */
2018                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2019                 }
2020                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2021         }
2022         EXIT;
2023 }
2024
2025 /** Checking routines for recovery */
2026 static int check_for_recovery_ready(struct lu_target *lut)
2027 {
2028         struct obd_device *obd = lut->lut_obd;
2029         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2030
2031         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2032                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2033                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2034                obd->obd_recovery_expired);
2035
2036         if (obd->obd_force_abort_recovery)
2037                 return 1;
2038
2039         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2040                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2041                 if (clnts + obd->obd_stale_clients <
2042                     obd->obd_max_recoverable_clients)
2043                         return 0;
2044         }
2045
2046         if (lut->lut_tdtd != NULL) {
2047                 if (!lut->lut_tdtd->tdtd_replay_ready) {
2048                         /* Let's extend recovery timer, in case the recovery
2049                          * timer expired, and some clients got evicted */
2050                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2051                                               true);
2052                         return 0;
2053                 } else {
2054                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2055                 }
2056         }
2057
2058         return 1;
2059 }
2060
2061 enum {
2062         REQUEST_RECOVERY = 1,
2063         UPDATE_RECOVERY = 2,
2064 };
2065
2066 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2067 {
2068         __u64 transno = 0;
2069
2070         if (!list_empty(&obd->obd_req_replay_queue)) {
2071                 struct ptlrpc_request *req;
2072
2073                 req = list_entry(obd->obd_req_replay_queue.next,
2074                                  struct ptlrpc_request, rq_list);
2075                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2076         }
2077
2078         return transno;
2079 }
2080 __u64 get_next_transno(struct lu_target *lut, int *type)
2081 {
2082         struct obd_device *obd = lut->lut_obd;
2083         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2084         __u64 transno = 0;
2085         __u64 update_transno;
2086         ENTRY;
2087
2088         transno = get_next_replay_req_transno(obd);
2089         if (type != NULL)
2090                 *type = REQUEST_RECOVERY;
2091
2092         if (tdtd == NULL)
2093                 RETURN(transno);
2094
2095         update_transno = distribute_txn_get_next_transno(tdtd);
2096         if (transno == 0 || (transno >= update_transno &&
2097                              update_transno != 0)) {
2098                 transno = update_transno;
2099                 if (type != NULL)
2100                         *type = UPDATE_RECOVERY;
2101         }
2102
2103         RETURN(transno);
2104 }
2105
2106 /**
2107  * drop duplicate replay request
2108  *
2109  * Because the operation has been replayed by update recovery, the request
2110  * with the same transno will be dropped and also notify the client to send
2111  * next replay request.
2112  *
2113  * \param[in] env       execution environment
2114  * \param[in] obd       failover obd device
2115  * \param[in] req       request to be dropped
2116  */
2117 static void drop_duplicate_replay_req(struct lu_env *env,
2118                                       struct obd_device *obd,
2119                                       struct ptlrpc_request *req)
2120 {
2121         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2122                   " update records are found.\n",
2123                   lustre_msg_get_transno(req->rq_reqmsg),
2124                   libcfs_nid2str(req->rq_peer.nid));
2125
2126         /* Right now, only for MDS reint operation update replay and
2127          * normal request replay can have the same transno */
2128         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2129                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2130                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2131                 if (likely(req->rq_export))
2132                         target_committed_to_req(req);
2133                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2134                 target_send_reply(req, req->rq_status, 0);
2135         } else {
2136                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2137                 libcfs_nid2str(req->rq_peer.nid));
2138         }
2139         target_exp_dequeue_req_replay(req);
2140         target_request_copy_put(req);
2141         obd->obd_replayed_requests++;
2142 }
2143
2144 static void replay_request_or_update(struct lu_env *env,
2145                                      struct lu_target *lut,
2146                                      struct target_recovery_data *trd,
2147                                      struct ptlrpc_thread *thread)
2148 {
2149         struct obd_device *obd = lut->lut_obd;
2150         struct ptlrpc_request *req = NULL;
2151         int                     type;
2152         __u64                   transno;
2153         ENTRY;
2154
2155         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2156                obd->obd_next_recovery_transno);
2157
2158         /* Replay all of request and update by transno */
2159         do {
2160                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2161
2162                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2163
2164                 /** It is needed to extend recovery window above
2165                  *  recovery_time_soft. Extending is possible only in the
2166                  *  end of recovery window (see more details in
2167                  *  handle_recovery_req()).
2168                  */
2169                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2170
2171                 if (target_recovery_overseer(lut, check_for_next_transno,
2172                                              exp_req_replay_healthy)) {
2173                         abort_req_replay_queue(obd);
2174                         abort_lock_replay_queue(obd);
2175                 }
2176
2177                 spin_lock(&obd->obd_recovery_task_lock);
2178                 transno = get_next_transno(lut, &type);
2179                 if (type == REQUEST_RECOVERY && transno != 0) {
2180                         /* Drop replay request from client side, if the
2181                          * replay has been executed by update with the
2182                          * same transno */
2183                         req = list_entry(obd->obd_req_replay_queue.next,
2184                                         struct ptlrpc_request, rq_list);
2185
2186                         list_del_init(&req->rq_list);
2187                         obd->obd_requests_queued_for_recovery--;
2188                         spin_unlock(&obd->obd_recovery_task_lock);
2189
2190                         /* Let's check if the request has been redone by
2191                          * update replay */
2192                         if (is_req_replayed_by_update(req)) {
2193                                 struct distribute_txn_replay_req *dtrq;
2194
2195                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2196                                                                    req->rq_xid);
2197                                 LASSERT(dtrq != NULL);
2198                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2199                                 list_del_init(&dtrq->dtrq_list);
2200                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2201                                 dtrq_destroy(dtrq);
2202
2203                                 drop_duplicate_replay_req(env, obd, req);
2204
2205                                 continue;
2206                         }
2207
2208                         LASSERT(trd->trd_processing_task == current_pid());
2209                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2210                                   lustre_msg_get_transno(req->rq_reqmsg),
2211                                   libcfs_nid2str(req->rq_peer.nid));
2212
2213                         handle_recovery_req(thread, req,
2214                                             trd->trd_recovery_handler);
2215                         /**
2216                          * bz18031: increase next_recovery_transno before
2217                          * target_request_copy_put() will drop exp_rpc reference
2218                          */
2219                         spin_lock(&obd->obd_recovery_task_lock);
2220                         obd->obd_next_recovery_transno++;
2221                         spin_unlock(&obd->obd_recovery_task_lock);
2222                         target_exp_dequeue_req_replay(req);
2223                         target_request_copy_put(req);
2224                         obd->obd_replayed_requests++;
2225                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2226                         struct distribute_txn_replay_req *dtrq;
2227
2228                         spin_unlock(&obd->obd_recovery_task_lock);
2229
2230                         LASSERT(tdtd != NULL);
2231                         dtrq = distribute_txn_get_next_req(tdtd);
2232                         lu_context_enter(&thread->t_env->le_ctx);
2233                         tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2234                         lu_context_exit(&thread->t_env->le_ctx);
2235                         extend_recovery_timer(obd, obd_timeout, true);
2236
2237                         /* Add it to the replay finish list */
2238                         spin_lock(&tdtd->tdtd_replay_list_lock);
2239                         if (dtrq->dtrq_xid != 0) {
2240                                 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2241                                        " to finish list\n", dtrq->dtrq_xid,
2242                                        dtrq->dtrq_master_transno);
2243                                 list_add(&dtrq->dtrq_list,
2244                                          &tdtd->tdtd_replay_finish_list);
2245                         } else {
2246                                 dtrq_destroy(dtrq);
2247                         }
2248                         spin_unlock(&tdtd->tdtd_replay_list_lock);
2249
2250                         spin_lock(&obd->obd_recovery_task_lock);
2251                         if (transno == obd->obd_next_recovery_transno)
2252                                 obd->obd_next_recovery_transno++;
2253                         else if (transno > obd->obd_next_recovery_transno)
2254                                 obd->obd_next_recovery_transno = transno + 1;
2255                         spin_unlock(&obd->obd_recovery_task_lock);
2256
2257                 } else {
2258                         spin_unlock(&obd->obd_recovery_task_lock);
2259                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2260                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2261                         /** evict exports failed VBR */
2262                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2263                         break;
2264                 }
2265         } while (1);
2266 }
2267
2268 static int target_recovery_thread(void *arg)
2269 {
2270         struct lu_target *lut = arg;
2271         struct obd_device *obd = lut->lut_obd;
2272         struct ptlrpc_request *req;
2273         struct target_recovery_data *trd = &obd->obd_recovery_data;
2274         unsigned long delta;
2275         struct lu_env *env;
2276         struct ptlrpc_thread *thread = NULL;
2277         int rc = 0;
2278         ENTRY;
2279
2280         unshare_fs_struct();
2281         OBD_ALLOC_PTR(thread);
2282         if (thread == NULL)
2283                 RETURN(-ENOMEM);
2284
2285         OBD_ALLOC_PTR(env);
2286         if (env == NULL) {
2287                 OBD_FREE_PTR(thread);
2288                 RETURN(-ENOMEM);
2289         }
2290
2291         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2292         if (rc) {
2293                 OBD_FREE_PTR(thread);
2294                 OBD_FREE_PTR(env);
2295                 RETURN(rc);
2296         }
2297
2298         thread->t_env = env;
2299         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2300         env->le_ctx.lc_thread = thread;
2301         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2302         thread->t_watchdog = NULL;
2303
2304         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2305                current_pid());
2306         trd->trd_processing_task = current_pid();
2307
2308         spin_lock(&obd->obd_dev_lock);
2309         obd->obd_recovering = 1;
2310         spin_unlock(&obd->obd_dev_lock);
2311         complete(&trd->trd_starting);
2312
2313         /* first of all, we have to know the first transno to replay */
2314         if (target_recovery_overseer(lut, check_for_recovery_ready,
2315                                      exp_connect_healthy)) {
2316                 abort_req_replay_queue(obd);
2317                 abort_lock_replay_queue(obd);
2318                 if (lut->lut_tdtd != NULL)
2319                         dtrq_list_destroy(lut->lut_tdtd);
2320         }
2321
2322         /* next stage: replay requests or update */
2323         delta = jiffies;
2324         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2325                atomic_read(&obd->obd_req_replay_clients),
2326                obd->obd_next_recovery_transno);
2327         replay_request_or_update(env, lut, trd, thread);
2328
2329         /**
2330          * The second stage: replay locks
2331          */
2332         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2333                atomic_read(&obd->obd_lock_replay_clients));
2334         while ((req = target_next_replay_lock(lut))) {
2335                 LASSERT(trd->trd_processing_task == current_pid());
2336                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2337                           libcfs_nid2str(req->rq_peer.nid));
2338                 handle_recovery_req(thread, req,
2339                                     trd->trd_recovery_handler);
2340                 target_request_copy_put(req);
2341                 obd->obd_replayed_locks++;
2342         }
2343
2344         /**
2345          * The third stage: reply on final pings, at this moment all clients
2346          * must have request in final queue
2347          */
2348         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2349         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2350         /** Update server last boot epoch */
2351         tgt_boot_epoch_update(lut);
2352         /* We drop recoverying flag to forward all new requests
2353          * to regular mds_handle() since now */
2354         spin_lock(&obd->obd_dev_lock);
2355         obd->obd_recovering = obd->obd_abort_recovery = 0;
2356         spin_unlock(&obd->obd_dev_lock);
2357         spin_lock(&obd->obd_recovery_task_lock);
2358         target_cancel_recovery_timer(obd);
2359         spin_unlock(&obd->obd_recovery_task_lock);
2360         while ((req = target_next_final_ping(obd))) {
2361                 LASSERT(trd->trd_processing_task == current_pid());
2362                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2363                           libcfs_nid2str(req->rq_peer.nid));
2364                 handle_recovery_req(thread, req,
2365                                     trd->trd_recovery_handler);
2366                 /* Because the waiting client can not send ping to server,
2367                  * so we need refresh the last_request_time, to avoid the
2368                  * export is being evicted */
2369                 ptlrpc_update_export_timer(req->rq_export, 0);
2370                 target_request_copy_put(req);
2371         }
2372
2373         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2374         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2375               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2376         if (delta > OBD_RECOVERY_TIME_SOFT) {
2377                 CWARN("too long recovery - read logs\n");
2378                 libcfs_debug_dumplog();
2379         }
2380
2381         target_finish_recovery(lut);
2382
2383         lu_context_fini(&env->le_ctx);
2384         trd->trd_processing_task = 0;
2385         complete(&trd->trd_finishing);
2386
2387         tgt_io_thread_done(thread);
2388         OBD_FREE_PTR(thread);
2389         OBD_FREE_PTR(env);
2390         RETURN(rc);
2391 }
2392
2393 static int target_start_recovery_thread(struct lu_target *lut,
2394                                         svc_handler_t handler)
2395 {
2396         struct obd_device *obd = lut->lut_obd;
2397         int rc = 0;
2398         struct target_recovery_data *trd = &obd->obd_recovery_data;
2399
2400         memset(trd, 0, sizeof(*trd));
2401         init_completion(&trd->trd_starting);
2402         init_completion(&trd->trd_finishing);
2403         trd->trd_recovery_handler = handler;
2404
2405         if (!IS_ERR(kthread_run(target_recovery_thread,
2406                                 lut, "tgt_recov"))) {
2407                 wait_for_completion(&trd->trd_starting);
2408                 LASSERT(obd->obd_recovering != 0);
2409         } else {
2410                 rc = -ECHILD;
2411         }
2412
2413         return rc;
2414 }
2415
2416 void target_stop_recovery_thread(struct obd_device *obd)
2417 {
2418         if (obd->obd_recovery_data.trd_processing_task > 0) {
2419                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2420                 /** recovery can be done but postrecovery is not yet */
2421                 spin_lock(&obd->obd_dev_lock);
2422                 if (obd->obd_recovering) {
2423                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2424                         obd->obd_abort_recovery = 1;
2425                         wake_up(&obd->obd_next_transno_waitq);
2426                 }
2427                 spin_unlock(&obd->obd_dev_lock);
2428                 wait_for_completion(&trd->trd_finishing);
2429         }
2430 }
2431 EXPORT_SYMBOL(target_stop_recovery_thread);
2432
2433 void target_recovery_fini(struct obd_device *obd)
2434 {
2435         class_disconnect_exports(obd);
2436         target_stop_recovery_thread(obd);
2437         target_cleanup_recovery(obd);
2438 }
2439 EXPORT_SYMBOL(target_recovery_fini);
2440
2441 static void target_recovery_expired(unsigned long castmeharder)
2442 {
2443         struct obd_device *obd = (struct obd_device *)castmeharder;
2444         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2445                " after %lds (%d clients connected)\n",
2446                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2447                cfs_time_current_sec()- obd->obd_recovery_start,
2448                atomic_read(&obd->obd_connected_clients));
2449
2450         obd->obd_recovery_expired = 1;
2451         wake_up(&obd->obd_next_transno_waitq);
2452 }
2453
2454 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2455 {
2456         struct obd_device *obd = lut->lut_obd;
2457
2458         if (obd->obd_max_recoverable_clients == 0) {
2459                 /** Update server last boot epoch */
2460                 tgt_boot_epoch_update(lut);
2461                 return;
2462         }
2463
2464         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2465                "last_transno "LPU64"\n", obd->obd_name,
2466                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2467         LASSERT(obd->obd_stopping == 0);
2468         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2469         obd->obd_recovery_start = 0;
2470         obd->obd_recovery_end = 0;
2471
2472         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2473         target_start_recovery_thread(lut, handler);
2474 }
2475 EXPORT_SYMBOL(target_recovery_init);
2476
2477 static int target_process_req_flags(struct obd_device *obd,
2478                                     struct ptlrpc_request *req)
2479 {
2480         struct obd_export *exp = req->rq_export;
2481         LASSERT(exp != NULL);
2482         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2483                 /* client declares he's ready to replay locks */
2484                 spin_lock(&exp->exp_lock);
2485                 if (exp->exp_req_replay_needed) {
2486                         exp->exp_req_replay_needed = 0;
2487                         spin_unlock(&exp->exp_lock);
2488
2489                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2490                         atomic_dec(&obd->obd_req_replay_clients);
2491                 } else {
2492                         spin_unlock(&exp->exp_lock);
2493                 }
2494         }
2495         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2496                 /* client declares he's ready to complete recovery
2497                  * so, we put the request on th final queue */
2498                 spin_lock(&exp->exp_lock);
2499                 if (exp->exp_lock_replay_needed) {
2500                         exp->exp_lock_replay_needed = 0;
2501                         spin_unlock(&exp->exp_lock);
2502
2503                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2504                         atomic_dec(&obd->obd_lock_replay_clients);
2505                 } else {
2506                         spin_unlock(&exp->exp_lock);
2507                 }
2508         }
2509         return 0;
2510 }
2511
2512 int target_queue_recovery_request(struct ptlrpc_request *req,
2513                                   struct obd_device *obd)
2514 {
2515         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2516         struct ptlrpc_request *reqiter;
2517         int inserted = 0;
2518         ENTRY;
2519
2520         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2521                 /* Processing the queue right now, don't re-add. */
2522                 RETURN(1);
2523         }
2524
2525         target_process_req_flags(obd, req);
2526
2527         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2528                 /* client declares he's ready to complete recovery
2529                  * so, we put the request on th final queue */
2530                 target_request_copy_get(req);
2531                 DEBUG_REQ(D_HA, req, "queue final req");
2532                 wake_up(&obd->obd_next_transno_waitq);
2533                 spin_lock(&obd->obd_recovery_task_lock);
2534                 if (obd->obd_recovering) {
2535                         list_add_tail(&req->rq_list,
2536                                           &obd->obd_final_req_queue);
2537                 } else {
2538                         spin_unlock(&obd->obd_recovery_task_lock);
2539                         target_request_copy_put(req);
2540                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2541                 }
2542                 spin_unlock(&obd->obd_recovery_task_lock);
2543                 RETURN(0);
2544         }
2545         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2546                 /* client declares he's ready to replay locks */
2547                 target_request_copy_get(req);
2548                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2549                 wake_up(&obd->obd_next_transno_waitq);
2550                 spin_lock(&obd->obd_recovery_task_lock);
2551                 LASSERT(obd->obd_recovering);
2552                 /* usually due to recovery abort */
2553                 if (!req->rq_export->exp_in_recovery) {
2554                         spin_unlock(&obd->obd_recovery_task_lock);
2555                         target_request_copy_put(req);
2556                         RETURN(-ENOTCONN);
2557                 }
2558                 LASSERT(req->rq_export->exp_lock_replay_needed);
2559                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2560                 spin_unlock(&obd->obd_recovery_task_lock);
2561                 RETURN(0);
2562         }
2563
2564         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2565          * (i.e. buflens etc are in my own byte order), but type-dependent
2566          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2567
2568         if (!transno) {
2569                 INIT_LIST_HEAD(&req->rq_list);
2570                 DEBUG_REQ(D_HA, req, "not queueing");
2571                 RETURN(1);
2572         }
2573
2574         /* If we're processing the queue, we want don't want to queue this
2575          * message.
2576          *
2577          * Also, if this request has a transno less than the one we're waiting
2578          * for, we should process it now.  It could (and currently always will)
2579          * be an open request for a descriptor that was opened some time ago.
2580          *
2581          * Also, a resent, replayed request that has already been
2582          * handled will pass through here and be processed immediately.
2583          */
2584         CDEBUG(D_HA, "Next recovery transno: "LPU64
2585                ", current: "LPU64", replaying\n",
2586                obd->obd_next_recovery_transno, transno);
2587
2588         /* If the request has been replayed by update replay, then sends this
2589          * request to the recovery thread (replay_request_or_update()), where
2590          * it will be handled */
2591         spin_lock(&obd->obd_recovery_task_lock);
2592         if (transno < obd->obd_next_recovery_transno &&
2593             !is_req_replayed_by_update(req)) {
2594                 /* Processing the queue right now, don't re-add. */
2595                 LASSERT(list_empty(&req->rq_list));
2596                 spin_unlock(&obd->obd_recovery_task_lock);
2597                 RETURN(1);
2598         }
2599         spin_unlock(&obd->obd_recovery_task_lock);
2600
2601         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2602                 RETURN(0);
2603
2604         target_request_copy_get(req);
2605         if (!req->rq_export->exp_in_recovery) {
2606                 target_request_copy_put(req);
2607                 RETURN(-ENOTCONN);
2608         }
2609         LASSERT(req->rq_export->exp_req_replay_needed);
2610
2611         if (target_exp_enqueue_req_replay(req)) {
2612                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2613                 target_request_copy_put(req);
2614                 RETURN(0);
2615         }
2616
2617         /* XXX O(n^2) */
2618         spin_lock(&obd->obd_recovery_task_lock);
2619         LASSERT(obd->obd_recovering);
2620         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2621                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2622                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2623                         inserted = 1;
2624                         goto added;
2625                 }
2626
2627                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2628                              transno)) {
2629                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2630                                   "has been claimed by another client");
2631                         spin_unlock(&obd->obd_recovery_task_lock);
2632                         target_exp_dequeue_req_replay(req);
2633                         target_request_copy_put(req);
2634                         RETURN(0);
2635                 }
2636         }
2637 added:
2638         if (!inserted)
2639                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2640
2641         obd->obd_requests_queued_for_recovery++;
2642         spin_unlock(&obd->obd_recovery_task_lock);
2643         wake_up(&obd->obd_next_transno_waitq);
2644         RETURN(0);
2645 }
2646
2647 int target_handle_ping(struct ptlrpc_request *req)
2648 {
2649         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2650         return req_capsule_server_pack(&req->rq_pill);
2651 }
2652
2653 void target_committed_to_req(struct ptlrpc_request *req)
2654 {
2655         struct obd_export *exp = req->rq_export;
2656
2657         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2658                 lustre_msg_set_last_committed(req->rq_repmsg,
2659                                               exp->exp_last_committed);
2660         else
2661                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2662                           "%d)", exp->exp_obd->obd_no_transno,
2663                           req->rq_repmsg == NULL);
2664
2665         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2666                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2667 }
2668
2669 #endif /* HAVE_SERVER_SUPPORT */
2670
2671 /**
2672  * Packs current SLV and Limit into \a req.
2673  */
2674 int target_pack_pool_reply(struct ptlrpc_request *req)
2675 {
2676         struct obd_device *obd;
2677         ENTRY;
2678
2679         /* Check that we still have all structures alive as this may
2680          * be some late RPC at shutdown time. */
2681         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2682                      !exp_connect_lru_resize(req->rq_export))) {
2683                 lustre_msg_set_slv(req->rq_repmsg, 0);
2684                 lustre_msg_set_limit(req->rq_repmsg, 0);
2685                 RETURN(0);
2686         }
2687
2688         /* OBD is alive here as export is alive, which we checked above. */
2689         obd = req->rq_export->exp_obd;
2690
2691         read_lock(&obd->obd_pool_lock);
2692         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2693         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2694         read_unlock(&obd->obd_pool_lock);
2695
2696         RETURN(0);
2697 }
2698
2699 static int target_send_reply_msg(struct ptlrpc_request *req,
2700                                  int rc, int fail_id)
2701 {
2702         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2703                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2704                 return -ECOMM;
2705         }
2706         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2707                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2708                 return -ECOMM;
2709
2710         if (unlikely(rc)) {
2711                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2712                 req->rq_status = rc;
2713                 return ptlrpc_send_error(req, 1);
2714         } else {
2715                 DEBUG_REQ(D_NET, req, "sending reply");
2716         }
2717
2718         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2719 }
2720
2721 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2722 {
2723         struct ptlrpc_service_part *svcpt;
2724         int                        netrc;
2725         struct ptlrpc_reply_state *rs;
2726         struct obd_export         *exp;
2727         ENTRY;
2728
2729         if (req->rq_no_reply) {
2730                 EXIT;
2731                 return;
2732         }
2733
2734         svcpt = req->rq_rqbd->rqbd_svcpt;
2735         rs = req->rq_reply_state;
2736         if (rs == NULL || !rs->rs_difficult) {
2737                 /* no notifiers */
2738                 target_send_reply_msg (req, rc, fail_id);
2739                 EXIT;
2740                 return;
2741         }
2742
2743         /* must be an export if locks saved */
2744         LASSERT(req->rq_export != NULL);
2745         /* req/reply consistent */
2746         LASSERT(rs->rs_svcpt == svcpt);
2747
2748         /* "fresh" reply */
2749         LASSERT(!rs->rs_scheduled);
2750         LASSERT(!rs->rs_scheduled_ever);
2751         LASSERT(!rs->rs_handled);
2752         LASSERT(!rs->rs_on_net);
2753         LASSERT(rs->rs_export == NULL);
2754         LASSERT(list_empty(&rs->rs_obd_list));
2755         LASSERT(list_empty(&rs->rs_exp_list));
2756
2757         exp = class_export_get(req->rq_export);
2758
2759         /* disable reply scheduling while I'm setting up */
2760         rs->rs_scheduled = 1;
2761         rs->rs_on_net    = 1;
2762         rs->rs_xid       = req->rq_xid;
2763         rs->rs_transno   = req->rq_transno;
2764         rs->rs_export    = exp;
2765         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2766
2767         spin_lock(&exp->exp_uncommitted_replies_lock);
2768         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2769                rs->rs_transno, exp->exp_last_committed);
2770         if (rs->rs_transno > exp->exp_last_committed) {
2771                 /* not committed already */
2772                 list_add_tail(&rs->rs_obd_list,
2773                                   &exp->exp_uncommitted_replies);
2774         }
2775         spin_unlock(&exp->exp_uncommitted_replies_lock);
2776
2777         spin_lock(&exp->exp_lock);
2778         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2779         spin_unlock(&exp->exp_lock);
2780
2781         netrc = target_send_reply_msg(req, rc, fail_id);
2782
2783         spin_lock(&svcpt->scp_rep_lock);
2784
2785         atomic_inc(&svcpt->scp_nreps_difficult);
2786
2787         if (netrc != 0) {
2788                 /* error sending: reply is off the net.  Also we need +1
2789                  * reply ref until ptlrpc_handle_rs() is done
2790                  * with the reply state (if the send was successful, there
2791                  * would have been +1 ref for the net, which
2792                  * reply_out_callback leaves alone) */
2793                 rs->rs_on_net = 0;
2794                 ptlrpc_rs_addref(rs);
2795         }
2796
2797         spin_lock(&rs->rs_lock);
2798         if (rs->rs_transno <= exp->exp_last_committed ||
2799             (!rs->rs_on_net && !rs->rs_no_ack) ||
2800             list_empty(&rs->rs_exp_list) ||     /* completed already */
2801             list_empty(&rs->rs_obd_list)) {
2802                 CDEBUG(D_HA, "Schedule reply immediately\n");
2803                 ptlrpc_dispatch_difficult_reply(rs);
2804         } else {
2805                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2806                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2807         }
2808         spin_unlock(&rs->rs_lock);
2809         spin_unlock(&svcpt->scp_rep_lock);
2810         EXIT;
2811 }
2812
2813 enum ldlm_mode lck_compat_array[] = {
2814         [LCK_EX]    = LCK_COMPAT_EX,
2815         [LCK_PW]    = LCK_COMPAT_PW,
2816         [LCK_PR]    = LCK_COMPAT_PR,
2817         [LCK_CW]    = LCK_COMPAT_CW,
2818         [LCK_CR]    = LCK_COMPAT_CR,
2819         [LCK_NL]    = LCK_COMPAT_NL,
2820         [LCK_GROUP] = LCK_COMPAT_GROUP,
2821         [LCK_COS]   = LCK_COMPAT_COS,
2822 };
2823
2824 /**
2825  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2826  * not escape to the user level.
2827  */
2828 int ldlm_error2errno(enum ldlm_error error)
2829 {
2830         int result;
2831
2832         switch (error) {
2833         case ELDLM_OK:
2834         case ELDLM_LOCK_MATCHED:
2835                 result = 0;
2836                 break;
2837         case ELDLM_LOCK_CHANGED:
2838                 result = -ESTALE;
2839                 break;
2840         case ELDLM_LOCK_ABORTED:
2841                 result = -ENAVAIL;
2842                 break;
2843         case ELDLM_LOCK_REPLACED:
2844                 result = -ESRCH;
2845                 break;
2846         case ELDLM_NO_LOCK_DATA:
2847                 result = -ENOENT;
2848                 break;
2849         case ELDLM_NAMESPACE_EXISTS:
2850                 result = -EEXIST;
2851                 break;
2852         case ELDLM_BAD_NAMESPACE:
2853                 result = -EBADF;
2854                 break;
2855         default:
2856                 if (((int)error) < 0) { /* cast to signed type */
2857                         result = error; /* as ldlm_error can be unsigned */
2858                 } else {
2859                         CERROR("Invalid DLM result code: %d\n", error);
2860                         result = -EPROTO;
2861                 }
2862         }
2863         return result;
2864 }
2865 EXPORT_SYMBOL(ldlm_error2errno);
2866
2867 /**
2868  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
2869  */
2870 enum ldlm_error ldlm_errno2error(int err_no)
2871 {
2872         int error;
2873
2874         switch (err_no) {
2875         case 0:
2876                 error = ELDLM_OK;
2877                 break;
2878         case -ESTALE:
2879                 error = ELDLM_LOCK_CHANGED;
2880                 break;
2881         case -ENAVAIL:
2882                 error = ELDLM_LOCK_ABORTED;
2883                 break;
2884         case -ESRCH:
2885                 error = ELDLM_LOCK_REPLACED;
2886                 break;
2887         case -ENOENT:
2888                 error = ELDLM_NO_LOCK_DATA;
2889                 break;
2890         case -EEXIST:
2891                 error = ELDLM_NAMESPACE_EXISTS;
2892                 break;
2893         case -EBADF:
2894                 error = ELDLM_BAD_NAMESPACE;
2895                 break;
2896         default:
2897                 error = err_no;
2898         }
2899         return error;
2900 }
2901
2902 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2903 void ldlm_dump_export_locks(struct obd_export *exp)
2904 {
2905         spin_lock(&exp->exp_locks_list_guard);
2906         if (!list_empty(&exp->exp_locks_list)) {
2907                 struct ldlm_lock *lock;
2908
2909                 CERROR("dumping locks for export %p,"
2910                        "ignore if the unmount doesn't hang\n", exp);
2911                 list_for_each_entry(lock, &exp->exp_locks_list,
2912                                         l_exp_refs_link)
2913                         LDLM_ERROR(lock, "lock:");
2914         }
2915         spin_unlock(&exp->exp_locks_list_guard);
2916 }
2917 #endif
2918
2919 #ifdef HAVE_SERVER_SUPPORT
2920 static int target_bulk_timeout(void *data)
2921 {
2922         ENTRY;
2923         /* We don't fail the connection here, because having the export
2924          * killed makes the (vital) call to commitrw very sad.
2925          */
2926         RETURN(1);
2927 }
2928
2929 static inline const char *bulk2type(struct ptlrpc_request *req)
2930 {
2931         if (req->rq_bulk_read)
2932                 return "READ";
2933         if (req->rq_bulk_write)
2934                 return "WRITE";
2935         return "UNKNOWN";
2936 }
2937
2938 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2939                    struct l_wait_info *lwi)
2940 {
2941         struct ptlrpc_request   *req = desc->bd_req;
2942         time_t                   start = cfs_time_current_sec();
2943         time_t                   deadline;
2944         int                      rc = 0;
2945
2946         ENTRY;
2947
2948         /* If there is eviction in progress, wait for it to finish. */
2949         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2950                 *lwi = LWI_INTR(NULL, NULL);
2951                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2952                                   !atomic_read(&exp->exp_obd->
2953                                                    obd_evict_inprogress),
2954                                   lwi);
2955         }
2956
2957         /* Check if client was evicted or reconnected already. */
2958         if (exp->exp_failed ||
2959             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2960                 rc = -ENOTCONN;
2961         } else {
2962                 if (req->rq_bulk_read)
2963                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2964
2965                 if ((exp->exp_connect_data.ocd_connect_flags &
2966                      OBD_CONNECT_BULK_MBITS) != 0)
2967                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
2968                 else /* old version, bulk matchbits is rq_xid */
2969                         req->rq_mbits = req->rq_xid;
2970
2971                 if (rc == 0)
2972                         rc = ptlrpc_start_bulk_transfer(desc);
2973         }
2974
2975         if (rc < 0) {
2976                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2977                           bulk2type(req), rc);
2978                 RETURN(rc);
2979         }
2980
2981         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2982                 ptlrpc_abort_bulk(desc);
2983                 RETURN(0);
2984         }
2985
2986         /* limit actual bulk transfer to bulk_timeout seconds */
2987         deadline = start + bulk_timeout;
2988         if (deadline > req->rq_deadline)
2989                 deadline = req->rq_deadline;
2990
2991         do {
2992                 long timeoutl = deadline - cfs_time_current_sec();
2993                 cfs_duration_t timeout = timeoutl <= 0 ?
2994                                          CFS_TICK : cfs_time_seconds(timeoutl);
2995                 time_t  rq_deadline;
2996
2997                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2998                                             target_bulk_timeout, desc);
2999                 rc = l_wait_event(desc->bd_waitq,
3000                                   !ptlrpc_server_bulk_active(desc) ||
3001                                   exp->exp_failed ||
3002                                   exp->exp_conn_cnt >
3003                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3004                                   lwi);
3005                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3006                 /* Wait again if we changed rq_deadline. */
3007                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3008                 deadline = start + bulk_timeout;
3009                 if (deadline > rq_deadline)
3010                         deadline = rq_deadline;
3011         } while ((rc == -ETIMEDOUT) &&
3012                  (deadline > cfs_time_current_sec()));
3013
3014         if (rc == -ETIMEDOUT) {
3015                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3016                           bulk2type(req), deadline - start,
3017                           cfs_time_current_sec() - deadline);
3018                 ptlrpc_abort_bulk(desc);
3019         } else if (exp->exp_failed) {
3020                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3021                           bulk2type(req));
3022                 rc = -ENOTCONN;
3023                 ptlrpc_abort_bulk(desc);
3024         } else if (exp->exp_conn_cnt >
3025                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3026                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3027                           bulk2type(req));
3028                 /* We don't reply anyway. */
3029                 rc = -ETIMEDOUT;
3030                 ptlrpc_abort_bulk(desc);
3031         } else if (desc->bd_failure) {
3032                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3033                           bulk2type(req));
3034                 /* XXX should this be a different errno? */
3035                 rc = -ETIMEDOUT;
3036         } else {
3037                 if (req->rq_bulk_write)
3038                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3039                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3040                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3041                                   bulk2type(req), desc->bd_nob_transferred,
3042                                   desc->bd_nob);
3043                         /* XXX should this be a different errno? */
3044                         rc = -ETIMEDOUT;
3045                 }
3046         }
3047
3048         RETURN(rc);
3049 }
3050 EXPORT_SYMBOL(target_bulk_io);
3051
3052 #endif /* HAVE_SERVER_SUPPORT */