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, 2015, 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         INIT_LIST_HEAD(&cli->cl_shrink_list);
 385
 386         init_waitqueue_head(&cli->cl_destroy_waitq);
 387         atomic_set(&cli->cl_destroy_in_flight, 0);
 388 #ifdef ENABLE_CHECKSUM
 389         /* Turn on checksumming by default. */
 390         cli->cl_checksum = 1;
 391         /*
 392          * The supported checksum types will be worked out at connect time
 393          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 394          * through procfs.
 395          */
 396         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 397 #endif
 398         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 399
 400         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
 401          * from OFD after connecting. */
 402         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 403
 404         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 405          * it will be updated at OSC connection time. */
 406         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 407
 408         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 409                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 410         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 411                 cli->cl_max_rpcs_in_flight = 2;
 412         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 413                 cli->cl_max_rpcs_in_flight = 3;
 414         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 415                 cli->cl_max_rpcs_in_flight = 4;
 416         } else {
 417                 if (osc_on_mdt(obddev->obd_name))
 418                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 419                 else
 420                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 421         }
 422
 423         spin_lock_init(&cli->cl_mod_rpcs_lock);
 424         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 425         cli->cl_max_mod_rpcs_in_flight = 0;
 426         cli->cl_mod_rpcs_in_flight = 0;
 427         cli->cl_close_rpcs_in_flight = 0;
 428         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 429         cli->cl_mod_tag_bitmap = NULL;
 430
 431         if (connect_op == MDS_CONNECT) {
 432                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 433                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 434                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 435                 if (cli->cl_mod_tag_bitmap == NULL)
 436                         GOTO(err, rc = -ENOMEM);
 437         }
 438
 439         rc = ldlm_get_ref();
 440         if (rc) {
 441                 CERROR("ldlm_get_ref failed: %d\n", rc);
 442                 GOTO(err, rc);
 443         }
 444
 445         ptlrpc_init_client(rq_portal, rp_portal, name,
 446                            &obddev->obd_ldlm_client);
 447
 448         imp = class_new_import(obddev);
 449         if (imp == NULL)
 450                 GOTO(err_ldlm, rc = -ENOENT);
 451         imp->imp_client = &obddev->obd_ldlm_client;
 452         imp->imp_connect_op = connect_op;
 453         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 454                LUSTRE_CFG_BUFLEN(lcfg, 1));
 455         class_import_put(imp);
 456
 457         rc = client_import_add_conn(imp, &server_uuid, 1);
 458         if (rc) {
 459                 CERROR("can't add initial connection\n");
 460                 GOTO(err_import, rc);
 461         }
 462
 463         cli->cl_import = imp;
 464         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 465         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 466
 467         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 468                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 469                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 470                                name, obddev->obd_name,
 471                                cli->cl_target_uuid.uuid);
 472                         spin_lock(&imp->imp_lock);
 473                         imp->imp_deactive = 1;
 474                         spin_unlock(&imp->imp_lock);
 475                 }
 476         }
 477
 478         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 479                                                    LDLM_NAMESPACE_CLIENT,
 480                                                    LDLM_NAMESPACE_GREEDY,
 481                                                    ns_type);
 482         if (obddev->obd_namespace == NULL) {
 483                 CERROR("Unable to create client namespace - %s\n",
 484                        obddev->obd_name);
 485                 GOTO(err_import, rc = -ENOMEM);
 486         }
 487
 488         RETURN(rc);
 489
 490 err_import:
 491         class_destroy_import(imp);
 492 err_ldlm:
 493         ldlm_put_ref();
 494 err:
 495         if (cli->cl_mod_tag_bitmap != NULL)
 496                 OBD_FREE(cli->cl_mod_tag_bitmap,
 497                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 498         cli->cl_mod_tag_bitmap = NULL;
 499         RETURN(rc);
 500
 501 }
 502 EXPORT_SYMBOL(client_obd_setup);
 503
 504 int client_obd_cleanup(struct obd_device *obddev)
 505 {
 506         struct client_obd *cli = &obddev->u.cli;
 507         ENTRY;
 508
 509         ldlm_namespace_free_post(obddev->obd_namespace);
 510         obddev->obd_namespace = NULL;
 511
 512         obd_cleanup_client_import(obddev);
 513         LASSERT(obddev->u.cli.cl_import == NULL);
 514
 515         ldlm_put_ref();
 516
 517         if (cli->cl_mod_tag_bitmap != NULL)
 518                 OBD_FREE(cli->cl_mod_tag_bitmap,
 519                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 520         cli->cl_mod_tag_bitmap = NULL;
 521
 522         RETURN(0);
 523 }
 524 EXPORT_SYMBOL(client_obd_cleanup);
 525
 526 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 527 int client_connect_import(const struct lu_env *env,
 528                           struct obd_export **exp,
 529                           struct obd_device *obd, struct obd_uuid *cluuid,
 530                           struct obd_connect_data *data, void *localdata)
 531 {
 532         struct client_obd       *cli    = &obd->u.cli;
 533         struct obd_import       *imp    = cli->cl_import;
 534         struct obd_connect_data *ocd;
 535         struct lustre_handle    conn    = { 0 };
 536         int                     rc;
 537         bool                    is_mdc = false;
 538         ENTRY;
 539
 540         *exp = NULL;
 541         down_write(&cli->cl_sem);
 542         if (cli->cl_conn_count > 0)
 543                 GOTO(out_sem, rc = -EALREADY);
 544
 545         rc = class_connect(&conn, obd, cluuid);
 546         if (rc)
 547                 GOTO(out_sem, rc);
 548
 549         cli->cl_conn_count++;
 550         *exp = class_conn2export(&conn);
 551
 552         LASSERT(obd->obd_namespace);
 553
 554         imp->imp_dlm_handle = conn;
 555         rc = ptlrpc_init_import(imp);
 556         if (rc != 0)
 557                 GOTO(out_ldlm, rc);
 558
 559         ocd = &imp->imp_connect_data;
 560         if (data) {
 561                 *ocd = *data;
 562                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
 563                                  LUSTRE_MDC_NAME, 3) == 0;
 564                 if (is_mdc)
 565                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
 566                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 567         }
 568
 569         rc = ptlrpc_connect_import(imp);
 570         if (rc != 0) {
 571                 if (data && is_mdc)
 572                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 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         struct obd_device *target;
 725         struct lustre_handle *hdl;
 726         cfs_time_t now;
 727         cfs_time_t deadline;
 728         int timeout;
 729         int rc = 0;
 730         ENTRY;
 731
 732         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 733         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 734                 conn->cookie = exp->exp_handle.h_cookie;
 735                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 736                        " cookie "LPX64"\n", cluuid->uuid, exp, conn->cookie);
 737                 RETURN(0);
 738         }
 739
 740         target = exp->exp_obd;
 741
 742         /* Might be a re-connect after a partition. */
 743         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 744                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 745                               "with handle "LPX64". Rejecting client "
 746                               "with the same UUID trying to reconnect "
 747                               "with handle "LPX64"\n", target->obd_name,
 748                               obd_uuid2str(&exp->exp_client_uuid),
 749                               obd_export_nid2str(exp),
 750                               hdl->cookie, conn->cookie);
 751                 memset(conn, 0, sizeof *conn);
 752                 /* target_handle_connect() treats EALREADY and
 753                  * -EALREADY differently.  -EALREADY is an error
 754                  * (same UUID, different handle). */
 755                 RETURN(-EALREADY);
 756         }
 757
 758         if (!target->obd_recovering) {
 759                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 760                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 761                         obd_export_nid2str(exp));
 762                 GOTO(out_already, rc);
 763         }
 764
 765         now = cfs_time_current();
 766         deadline = cfs_timer_deadline(&target->obd_recovery_timer);
 767         if (cfs_time_before(now, deadline)) {
 768                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
 769                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting,"
 770                         " waiting for %d clients in recovery for"
 771                         " %d:%.02d\n", target->obd_name,
 772                         obd_uuid2str(&exp->exp_client_uuid),
 773                         obd_export_nid2str(exp),
 774                         target->obd_max_recoverable_clients,
 775                         timeout / 60, timeout % 60);
 776         } else {
 777                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
 778                 LCONSOLE_WARN("%s: Recovery already passed deadline"
 779                         " %d:%.02d, It is most likely due to DNE"
 780                         " recovery is failed or stuck, please wait a"
 781                         " few more minutes or abort the recovery.\n",
 782                         target->obd_name, timeout / 60, timeout % 60);
 783         }
 784
 785 out_already:
 786         conn->cookie = exp->exp_handle.h_cookie;
 787         /* target_handle_connect() treats EALREADY and
 788          * -EALREADY differently.  EALREADY means we are
 789          * doing a valid reconnect from the same client. */
 790         RETURN(EALREADY);
 791 }
 792
 793 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 794                           int error)
 795 {
 796         struct obd_export *exp = cb_data;
 797
 798         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 799                obd->obd_name, exp->exp_client_uuid.uuid);
 800
 801         spin_lock(&exp->exp_lock);
 802         exp->exp_need_sync = 0;
 803         spin_unlock(&exp->exp_lock);
 804         class_export_cb_put(exp);
 805 }
 806
 807 static void
 808 check_and_start_recovery_timer(struct obd_device *obd,
 809                                struct ptlrpc_request *req, int new_client);
 810
 811 /**
 812  * update flags for import during reconnect process
 813  */
 814 static int rev_import_flags_update(struct obd_import *revimp,
 815                                    struct ptlrpc_request *req)
 816 {
 817         int rc;
 818         struct obd_connect_data *data;
 819
 820         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 821
 822         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 823                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 824         else
 825                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 826
 827         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 828
 829         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 830         if (rc) {
 831                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 832                         revimp->imp_client->cli_name,
 833                         libcfs_id2str(req->rq_peer), rc);
 834                 return rc;
 835         }
 836
 837         return 0;
 838 }
 839
 840 /**
 841  * Allocate a new reverse import for an export.
 842  *
 843  * \retval -errno in case error hit
 844  * \retval 0 if reverse import correctly init
 845  **/
 846 int rev_import_init(struct obd_export *export)
 847 {
 848         struct obd_device *obd = export->exp_obd;
 849         struct obd_import *revimp;
 850
 851         LASSERT(export->exp_imp_reverse == NULL);
 852
 853         revimp = class_new_import(obd);
 854         if (revimp == NULL)
 855                 return -ENOMEM;
 856
 857         revimp->imp_remote_handle.cookie = 0ULL;
 858         revimp->imp_client = &obd->obd_ldlm_client;
 859         revimp->imp_dlm_fake = 1;
 860
 861         /* it is safe to connect import in new state as no sends possible */
 862         spin_lock(&export->exp_lock);
 863         export->exp_imp_reverse = revimp;
 864         spin_unlock(&export->exp_lock);
 865         class_import_put(revimp);
 866
 867         return 0;
 868 }
 869 EXPORT_SYMBOL(rev_import_init);
 870
 871 /**
 872  * Handle reconnect for an export.
 873  *
 874  * \param exp export to handle reconnect process
 875  * \param req client reconnect request
 876  *
 877  * \retval -rc in case securitfy flavor can't be changed
 878  * \retval 0 in case none problems
 879  */
 880 static int rev_import_reconnect(struct obd_export *exp,
 881                                 struct ptlrpc_request *req)
 882 {
 883         struct obd_import *revimp = exp->exp_imp_reverse;
 884         struct lustre_handle *lh;
 885         int rc;
 886
 887         /* avoid sending a request until import flags are changed */
 888         ptlrpc_import_enter_resend(revimp);
 889
 890         if (revimp->imp_connection != NULL)
 891                 ptlrpc_connection_put(revimp->imp_connection);
 892
 893         /*
 894          * client from recovery don't have a handle so we need to take from
 895          * request. it may produce situation when wrong client connected
 896          * to recovery as we trust a client uuid
 897          */
 898         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 899         revimp->imp_remote_handle = *lh;
 900
 901         /* unknown versions will be caught in
 902          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 903         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 904
 905         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 906
 907         rc = rev_import_flags_update(revimp, req);
 908         if (rc != 0) {
 909                 /* it is safe to still be in RECOVERY phase as we are not able
 910                  * to setup correct security flavor so requests are not able to
 911                  * be delivered correctly */
 912                 return rc;
 913         }
 914
 915         /* resend all rpc's via new connection */
 916         return ptlrpc_import_recovery_state_machine(revimp);
 917 }
 918
 919 int target_handle_connect(struct ptlrpc_request *req)
 920 {
 921         struct obd_device *target = NULL;
 922         struct obd_export *export = NULL;
 923         /* connect handle - filled from target_handle_reconnect in
 924          * reconnect case */
 925         struct lustre_handle conn;
 926         struct lustre_handle *tmp;
 927         struct obd_uuid tgtuuid;
 928         struct obd_uuid cluuid;
 929         char *str;
 930         int rc = 0;
 931         char *target_start;
 932         int target_len;
 933         bool     mds_conn = false, lw_client = false;
 934         bool     mds_mds_conn = false;
 935         bool     new_mds_mds_conn = false;
 936         bool     target_referenced = false;
 937         struct obd_connect_data *data, *tmpdata;
 938         int size, tmpsize;
 939         lnet_nid_t *client_nid = NULL;
 940         ENTRY;
 941
 942         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 943
 944         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 945         if (str == NULL) {
 946                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 947                 GOTO(out, rc = -EINVAL);
 948         }
 949
 950         obd_str2uuid(&tgtuuid, str);
 951         target = class_uuid2obd(&tgtuuid);
 952         if (!target)
 953                 target = class_name2obd(str);
 954
 955         if (!target) {
 956                 deuuidify(str, NULL, &target_start, &target_len);
 957                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 958                                    "from %s (no target). If you are running "
 959                                    "an HA pair check that the target is "
 960                                    "mounted on the other server.\n", str,
 961                                    libcfs_nid2str(req->rq_peer.nid));
 962                 GOTO(out, rc = -ENODEV);
 963         }
 964
 965         spin_lock(&target->obd_dev_lock);
 966         if (target->obd_stopping || !target->obd_set_up) {
 967                 spin_unlock(&target->obd_dev_lock);
 968
 969                 deuuidify(str, NULL, &target_start, &target_len);
 970                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 971                               target_len, target_start,
 972                               libcfs_nid2str(req->rq_peer.nid),
 973                               (target->obd_stopping ?
 974                                "stopping" : "not set up"));
 975                 GOTO(out, rc = -ENODEV);
 976         }
 977
 978         if (target->obd_no_conn) {
 979                 spin_unlock(&target->obd_dev_lock);
 980
 981                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 982                                "from %s\n", target->obd_name,
 983                                libcfs_nid2str(req->rq_peer.nid));
 984                 GOTO(out, rc = -EAGAIN);
 985         }
 986
 987         /* Make sure the target isn't cleaned up while we're here. Yes,
 988          * there's still a race between the above check and our incref here.
 989          * Really, class_uuid2obd should take the ref. */
 990         class_incref(target, __func__, current);
 991         target_referenced = true;
 992
 993         target->obd_conn_inprogress++;
 994         spin_unlock(&target->obd_dev_lock);
 995
 996         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 997         if (str == NULL) {
 998                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 999                 GOTO(out, rc = -EINVAL);
1000         }
1001
1002         obd_str2uuid(&cluuid, str);
1003
1004         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1005         if (tmp == NULL)
1006                 GOTO(out, rc = -EPROTO);
1007
1008         conn = *tmp;
1009
1010         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1011                                     RCL_CLIENT);
1012         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1013         if (!data)
1014                 GOTO(out, rc = -EPROTO);
1015
1016         rc = req_capsule_server_pack(&req->rq_pill);
1017         if (rc)
1018                 GOTO(out, rc);
1019
1020 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1021         /* Don't allow clients to connect that are using old 1.8 format
1022          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1023          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1024          * FULL20 flag should be set on all connections since 2.0, but no
1025          * longer affects behaviour.
1026          *
1027          * Later this check will be disabled and the flag can be retired
1028          * completely once interop with 3.0 is no longer needed.
1029          */
1030         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1031                 GOTO(out, rc = -EPROTO);
1032 #endif
1033
1034         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1035                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1036                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1037                     data->ocd_version > LUSTRE_VERSION_CODE +
1038                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1039                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1040                                   "libclient connection attempt",
1041                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1042                                   "old" : "new",
1043                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1044                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1045                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1046                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1047                         data = req_capsule_server_sized_get(&req->rq_pill,
1048                                                             &RMF_CONNECT_DATA,
1049                                     offsetof(typeof(*data), ocd_version) +
1050                                              sizeof(data->ocd_version));
1051                         if (data) {
1052                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1053                                 data->ocd_version = LUSTRE_VERSION_CODE;
1054                         }
1055                         GOTO(out, rc = -EPROTO);
1056                 }
1057         }
1058
1059         /* Note: lw_client is needed in MDS-MDS failover during update log
1060          * processing, so we needs to allow lw_client to be connected at
1061          * anytime, instead of only the initial connection */
1062         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1063
1064         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1065                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1066                 mds_mds_conn = (data->ocd_connect_flags &
1067                                 OBD_CONNECT_MDS_MDS) != 0;
1068
1069                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1070                  * for Imperative Recovery connection from MGC to MGS.
1071                  *
1072                  * Via check OBD_CONNECT_FID, we can distinguish whether
1073                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1074                  * MGC or MDT. */
1075                 if (!lw_client &&
1076                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1077                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1078                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1079                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1080                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1081                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1082
1083                         /* We do not support the MDT-MDT interoperations with
1084                          * different version MDT because of protocol changes. */
1085                         if (unlikely(major != LUSTRE_MAJOR ||
1086                                      minor != LUSTRE_MINOR ||
1087                                      abs(patch - LUSTRE_PATCH) > 3)) {
1088                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1089                                         "connection from different version MDT "
1090                                         "(%d.%d.%d.%d) %s %s\n",
1091                                         target->obd_name, LUSTRE_MAJOR,
1092                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1093                                         major, minor, patch,
1094                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1095                                         libcfs_nid2str(req->rq_peer.nid), str);
1096
1097                                 GOTO(out, rc = -EPROTO);
1098                         }
1099                 }
1100         }
1101
1102         /* lctl gets a backstage, all-access pass. */
1103         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1104                 goto dont_check_exports;
1105
1106         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1107         if (!export)
1108                 goto no_export;
1109
1110         /* We've found an export in the hash. */
1111
1112         spin_lock(&export->exp_lock);
1113
1114         if (export->exp_connecting) { /* bug 9635, et. al. */
1115                 spin_unlock(&export->exp_lock);
1116                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1117                               export->exp_obd->obd_name, export,
1118                               libcfs_nid2str(req->rq_peer.nid));
1119                 class_export_put(export);
1120                 export = NULL;
1121                 rc = -EALREADY;
1122         } else if ((mds_conn || lw_client ||
1123                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1124                    export->exp_connection != NULL) {
1125                 spin_unlock(&export->exp_lock);
1126                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1127                         /* MDS or LWP reconnected after failover. */
1128                         LCONSOLE_WARN("%s: Received %s connection from "
1129                             "%s, removing former export from %s\n",
1130                             target->obd_name, mds_conn ? "MDS" : "LWP",
1131                             libcfs_nid2str(req->rq_peer.nid),
1132                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1133                 } else {
1134                         /* New MDS connection from the same NID. */
1135                         LCONSOLE_WARN("%s: Received new %s connection from "
1136                                 "%s, removing former export from same NID\n",
1137                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1138                                 libcfs_nid2str(req->rq_peer.nid));
1139                 }
1140
1141                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1142                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1143                         /* Because exports between MDTs will always be
1144                          * kept, let's do not fail such export if they
1145                          * come from the same NID, otherwise it might
1146                          * cause eviction between MDTs, which might
1147                          * cause namespace inconsistency */
1148                         spin_lock(&export->exp_lock);
1149                         export->exp_connecting = 1;
1150                         spin_unlock(&export->exp_lock);
1151                         conn.cookie = export->exp_handle.h_cookie;
1152                         rc = EALREADY;
1153                 } else {
1154                         class_fail_export(export);
1155                         class_export_put(export);
1156                         export = NULL;
1157                         rc = 0;
1158                 }
1159         } else if (export->exp_connection != NULL &&
1160                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1161                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1162                     MSG_CONNECT_INITIAL)) {
1163                 spin_unlock(&export->exp_lock);
1164                 /* In MDS failover we have static UUID but NID can change. */
1165                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1166                               "existing nid %s is already connected\n",
1167                               target->obd_name, cluuid.uuid,
1168                               libcfs_nid2str(req->rq_peer.nid),
1169                               libcfs_nid2str(
1170                                       export->exp_connection->c_peer.nid));
1171                 rc = -EALREADY;
1172                 class_export_put(export);
1173                 export = NULL;
1174         } else {
1175                 export->exp_connecting = 1;
1176                 spin_unlock(&export->exp_lock);
1177                 LASSERT(export->exp_obd == target);
1178
1179                 rc = target_handle_reconnect(&conn, export, &cluuid);
1180         }
1181
1182         /* If we found an export, we already unlocked. */
1183         if (!export) {
1184 no_export:
1185                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1186         } else if (req->rq_export == NULL &&
1187                    atomic_read(&export->exp_rpc_count) > 0) {
1188                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1189                               "still busy with %d references\n",
1190                               target->obd_name, cluuid.uuid,
1191                               libcfs_nid2str(req->rq_peer.nid),
1192                               atomic_read(&export->exp_refcount));
1193                 GOTO(out, rc = -EBUSY);
1194         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1195                 if (!strstr(cluuid.uuid, "mdt"))
1196                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1197                                       "known client %s (at %s) because it "
1198                                       "is indicating it is a new client",
1199                                       target->obd_name, cluuid.uuid,
1200                                       libcfs_nid2str(req->rq_peer.nid));
1201                 GOTO(out, rc = -EALREADY);
1202         } else {
1203                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1204         }
1205
1206         if (rc < 0) {
1207                 GOTO(out, rc);
1208         }
1209
1210         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1211                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1212               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1213               export, (long)cfs_time_current_sec(),
1214               export ? (long)export->exp_last_request_time : 0);
1215
1216         /* If this is the first time a client connects, reset the recovery
1217          * timer. Discard lightweight connections which might be local. */
1218         if (!lw_client && rc == 0 && target->obd_recovering)
1219                 check_and_start_recovery_timer(target, req, export == NULL);
1220
1221         /* We want to handle EALREADY but *not* -EALREADY from
1222          * target_handle_reconnect(), return reconnection state in a flag. */
1223         if (rc == EALREADY) {
1224                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1225                 rc = 0;
1226         } else {
1227                 LASSERT(rc == 0);
1228         }
1229
1230         /* Tell the client if we support replayable requests. */
1231         if (target->obd_replayable)
1232                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1233         client_nid = &req->rq_peer.nid;
1234
1235         if (export == NULL) {
1236                 /* allow lightweight connections during recovery */
1237                 /* allow "new" MDT to be connected during recovery, since we
1238                  * need retrieve recovery update records from it */
1239                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1240                         cfs_time_t t;
1241                         int     c; /* connected */
1242                         int     i; /* in progress */
1243                         int     k; /* known */
1244                         int     s; /* stale/evicted */
1245
1246                         c = atomic_read(&target->obd_connected_clients);
1247                         i = atomic_read(&target->obd_lock_replay_clients);
1248                         k = target->obd_max_recoverable_clients;
1249                         s = target->obd_stale_clients;
1250                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1251                         t = cfs_time_sub(t, cfs_time_current());
1252                         t = cfs_duration_sec(t);
1253                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1254                                       "(at %s), waiting for %d known clients "
1255                                       "(%d recovered, %d in progress, and %d "
1256                                       "evicted) to recover in %d:%.02d\n",
1257                                       target->obd_name, cluuid.uuid,
1258                                       libcfs_nid2str(req->rq_peer.nid), k,
1259                                       c - i, i, s, (int)t / 60,
1260                                       (int)t % 60);
1261                         rc = -EBUSY;
1262                 } else {
1263 dont_check_exports:
1264                         rc = obd_connect(req->rq_svc_thread->t_env,
1265                                          &export, target, &cluuid, data,
1266                                          client_nid);
1267                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1268                                 lustre_msg_add_op_flags(req->rq_repmsg,
1269                                                         MSG_CONNECT_RECOVERING);
1270                         if (rc == 0) {
1271                                 conn.cookie = export->exp_handle.h_cookie;
1272                                 rc = rev_import_init(export);
1273                         }
1274
1275                         if (mds_mds_conn)
1276                                 new_mds_mds_conn = true;
1277                 }
1278         } else {
1279                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1280                                    export, target, &cluuid, data, client_nid);
1281         }
1282         if (rc)
1283                 GOTO(out, rc);
1284
1285         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1286         data->ocd_instance = target->u.obt.obt_instance;
1287
1288         /* Return only the parts of obd_connect_data that we understand, so the
1289          * client knows that we don't understand the rest. */
1290         if (data) {
1291                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1292                                                RCL_SERVER);
1293                 tmpdata = req_capsule_server_get(&req->rq_pill,
1294                                                  &RMF_CONNECT_DATA);
1295                 /* Don't use struct assignment here, because the client reply
1296                  * buffer may be smaller/larger than the local struct
1297                  * obd_connect_data. */
1298                 memcpy(tmpdata, data, min(tmpsize, size));
1299         }
1300
1301         /* If the client and the server are the same node, we will already
1302          * have an export that really points to the client's DLM export,
1303          * because we have a shared handles table.
1304          *
1305          * XXX this will go away when shaver stops sending the "connect" handle
1306          * in the real "remote handle" field of the request --phik 24 Apr 2003
1307          */
1308         ptlrpc_request_change_export(req, export);
1309
1310         spin_lock(&export->exp_lock);
1311         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1312                 spin_unlock(&export->exp_lock);
1313                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1314                        "or equal conn_cnt: %d >= %d\n",
1315                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1316                        export->exp_conn_cnt,
1317                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1318
1319                 GOTO(out, rc = -EALREADY);
1320         }
1321         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1322         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1323
1324         /* Don't evict liblustre clients for not pinging. */
1325         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1326                 export->exp_libclient = 1;
1327                 spin_unlock(&export->exp_lock);
1328
1329                 spin_lock(&target->obd_dev_lock);
1330                 list_del_init(&export->exp_obd_chain_timed);
1331                 spin_unlock(&target->obd_dev_lock);
1332         } else {
1333                 spin_unlock(&export->exp_lock);
1334         }
1335
1336         if (export->exp_connection != NULL) {
1337                 /* Check to see if connection came from another NID. */
1338                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1339                     !hlist_unhashed(&export->exp_nid_hash))
1340                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1341                                      &export->exp_connection->c_peer.nid,
1342                                      &export->exp_nid_hash);
1343
1344                 ptlrpc_connection_put(export->exp_connection);
1345         }
1346
1347         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1348                                                        req->rq_self,
1349                                                        &cluuid);
1350         if (hlist_unhashed(&export->exp_nid_hash))
1351                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1352                              &export->exp_connection->c_peer.nid,
1353                              &export->exp_nid_hash);
1354
1355         lustre_msg_set_handle(req->rq_repmsg, &conn);
1356
1357         rc = rev_import_reconnect(export, req);
1358         if (rc != 0)
1359                 GOTO(out, rc);
1360
1361         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1362                 int has_transno;
1363                 __u64 transno = data->ocd_transno;
1364
1365                 spin_lock(&export->exp_lock);
1366                 /* possible race with class_disconnect_stale_exports,
1367                  * export may be already in the eviction process */
1368                 if (export->exp_failed) {
1369                         spin_unlock(&export->exp_lock);
1370                         GOTO(out, rc = -ENODEV);
1371                 }
1372                 export->exp_in_recovery = 1;
1373                 export->exp_req_replay_needed = 1;
1374                 export->exp_lock_replay_needed = 1;
1375                 spin_unlock(&export->exp_lock);
1376
1377                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1378                                  MSG_CONNECT_TRANSNO);
1379                 if (has_transno && transno == 0)
1380                         CWARN("Connect with zero transno!\n");
1381
1382                 if (has_transno && transno > 0 &&
1383                     transno < target->obd_next_recovery_transno &&
1384                     transno > target->obd_last_committed) {
1385                         /* Another way is to use cmpxchg() to be lock-free. */
1386                         spin_lock(&target->obd_recovery_task_lock);
1387                         if (transno < target->obd_next_recovery_transno)
1388                                 target->obd_next_recovery_transno = transno;
1389                         spin_unlock(&target->obd_recovery_task_lock);
1390                 }
1391
1392                 atomic_inc(&target->obd_req_replay_clients);
1393                 atomic_inc(&target->obd_lock_replay_clients);
1394                 /* Note: MDS-MDS connection is allowed to be connected during
1395                  * recovery, no matter if the exports needs to be recoveried.
1396                  * Because we need retrieve updates logs from all other MDTs.
1397                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1398                  * also needs to be increased to match other recovery checking
1399                  * condition. */
1400                 if (new_mds_mds_conn)
1401                         target->obd_max_recoverable_clients++;
1402                 if (atomic_inc_return(&target->obd_connected_clients) ==
1403                     target->obd_max_recoverable_clients)
1404                         wake_up(&target->obd_next_transno_waitq);
1405         }
1406
1407         /* Tell the client we're in recovery, when client is involved in it. */
1408         if (target->obd_recovering && !lw_client)
1409                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1410
1411 out:
1412         if (export) {
1413                 spin_lock(&export->exp_lock);
1414                 export->exp_connecting = 0;
1415                 spin_unlock(&export->exp_lock);
1416
1417                 class_export_put(export);
1418         }
1419         if (target_referenced == true && target != NULL) {
1420                 spin_lock(&target->obd_dev_lock);
1421                 target->obd_conn_inprogress--;
1422                 spin_unlock(&target->obd_dev_lock);
1423
1424                 class_decref(target, __func__, current);
1425         }
1426         req->rq_status = rc;
1427         RETURN(rc);
1428 }
1429
1430 int target_handle_disconnect(struct ptlrpc_request *req)
1431 {
1432         int rc;
1433         ENTRY;
1434
1435         rc = req_capsule_server_pack(&req->rq_pill);
1436         if (rc)
1437                 RETURN(rc);
1438
1439         /* Keep the rq_export around so we can send the reply. */
1440         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1441
1442         RETURN(0);
1443 }
1444
1445 void target_destroy_export(struct obd_export *exp)
1446 {
1447         struct obd_import       *imp = NULL;
1448         /* exports created from last_rcvd data, and "fake"
1449            exports created by lctl don't have an import */
1450         spin_lock(&exp->exp_lock);
1451         if (exp->exp_imp_reverse != NULL) {
1452                 imp = exp->exp_imp_reverse;
1453                 exp->exp_imp_reverse = NULL;
1454         }
1455         spin_unlock(&exp->exp_lock);
1456         if (imp != NULL)
1457                 client_destroy_import(imp);
1458
1459         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1460         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1461         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1462         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1463 }
1464 EXPORT_SYMBOL(target_destroy_export);
1465
1466 /*
1467  * Recovery functions
1468  */
1469 static void target_request_copy_get(struct ptlrpc_request *req)
1470 {
1471         class_export_rpc_inc(req->rq_export);
1472         LASSERT(list_empty(&req->rq_list));
1473         INIT_LIST_HEAD(&req->rq_replay_list);
1474
1475         /* Increase refcount to keep request in queue. */
1476         atomic_inc(&req->rq_refcount);
1477         /* Let export know it has replays to be handled. */
1478         atomic_inc(&req->rq_export->exp_replay_count);
1479 }
1480
1481 static void target_request_copy_put(struct ptlrpc_request *req)
1482 {
1483         LASSERT(list_empty(&req->rq_replay_list));
1484         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1485
1486         atomic_dec(&req->rq_export->exp_replay_count);
1487         class_export_rpc_dec(req->rq_export);
1488         ptlrpc_server_drop_request(req);
1489 }
1490
1491 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1492 {
1493         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1494         struct obd_export     *exp = req->rq_export;
1495         struct ptlrpc_request *reqiter;
1496         struct ptlrpc_request *dup_req = NULL;
1497         int                    dup = 0;
1498
1499         LASSERT(exp);
1500
1501         spin_lock(&exp->exp_lock);
1502         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1503                                 rq_replay_list) {
1504                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1505                         dup_req = reqiter;
1506                         dup = 1;
1507                         break;
1508                 }
1509         }
1510
1511         if (dup) {
1512                 /* We expect it with RESENT and REPLAY flags. */
1513                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1514                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1515                         CERROR("invalid flags %x of resent replay\n",
1516                                lustre_msg_get_flags(req->rq_reqmsg));
1517
1518                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1519                         __u32 new_conn;
1520
1521                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1522                         if (new_conn >
1523                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1524                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1525                                                         new_conn);
1526                 }
1527         } else {
1528                 list_add_tail(&req->rq_replay_list,
1529                                   &exp->exp_req_replay_queue);
1530         }
1531
1532         spin_unlock(&exp->exp_lock);
1533         return dup;
1534 }
1535
1536 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1537 {
1538         LASSERT(!list_empty(&req->rq_replay_list));
1539         LASSERT(req->rq_export);
1540
1541         spin_lock(&req->rq_export->exp_lock);
1542         list_del_init(&req->rq_replay_list);
1543         spin_unlock(&req->rq_export->exp_lock);
1544 }
1545
1546 static void target_finish_recovery(struct lu_target *lut)
1547 {
1548         struct obd_device *obd = lut->lut_obd;
1549         ENTRY;
1550
1551         /* Only log a recovery message when recovery has occurred. */
1552         if (obd->obd_recovery_start) {
1553                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1554                                         obd->obd_recovery_start);
1555                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1556                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1557                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1558                         obd->obd_max_recoverable_clients,
1559                         atomic_read(&obd->obd_connected_clients),
1560                         obd->obd_stale_clients,
1561                         obd->obd_stale_clients == 1 ? "was" : "were");
1562         }
1563
1564         ldlm_reprocess_all_ns(obd->obd_namespace);
1565         spin_lock(&obd->obd_recovery_task_lock);
1566         if (!list_empty(&obd->obd_req_replay_queue) ||
1567             !list_empty(&obd->obd_lock_replay_queue) ||
1568             !list_empty(&obd->obd_final_req_queue)) {
1569                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1570                        obd->obd_name,
1571                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1572                        list_empty(&obd->obd_lock_replay_queue) ? \
1573                                "" : "lock ",
1574                        list_empty(&obd->obd_final_req_queue) ? \
1575                                "" : "final ");
1576                 spin_unlock(&obd->obd_recovery_task_lock);
1577                 LBUG();
1578         }
1579         spin_unlock(&obd->obd_recovery_task_lock);
1580
1581         obd->obd_recovery_end = cfs_time_current_sec();
1582
1583         /* When recovery finished, cleanup orphans on MDS and OST. */
1584         if (OBT(obd) && OBP(obd, postrecov)) {
1585                 int rc = OBP(obd, postrecov)(obd);
1586                 if (rc < 0)
1587                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1588                                       obd->obd_name, rc);
1589         }
1590         EXIT;
1591 }
1592
1593 static void abort_req_replay_queue(struct obd_device *obd)
1594 {
1595         struct ptlrpc_request *req, *n;
1596         struct list_head abort_list;
1597
1598         INIT_LIST_HEAD(&abort_list);
1599         spin_lock(&obd->obd_recovery_task_lock);
1600         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1601         spin_unlock(&obd->obd_recovery_task_lock);
1602         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1603                 DEBUG_REQ(D_WARNING, req, "aborted:");
1604                 req->rq_status = -ENOTCONN;
1605                 if (ptlrpc_error(req)) {
1606                         DEBUG_REQ(D_ERROR, req,
1607                                   "failed abort_req_reply; skipping");
1608                 }
1609                 target_exp_dequeue_req_replay(req);
1610                 target_request_copy_put(req);
1611         }
1612 }
1613
1614 static void abort_lock_replay_queue(struct obd_device *obd)
1615 {
1616         struct ptlrpc_request *req, *n;
1617         struct list_head abort_list;
1618
1619         INIT_LIST_HEAD(&abort_list);
1620         spin_lock(&obd->obd_recovery_task_lock);
1621         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1622         spin_unlock(&obd->obd_recovery_task_lock);
1623         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1624                 DEBUG_REQ(D_ERROR, req, "aborted:");
1625                 req->rq_status = -ENOTCONN;
1626                 if (ptlrpc_error(req)) {
1627                         DEBUG_REQ(D_ERROR, req,
1628                                   "failed abort_lock_reply; skipping");
1629                 }
1630                 target_request_copy_put(req);
1631         }
1632 }
1633
1634 /* Called from a cleanup function if the device is being cleaned up
1635    forcefully.  The exports should all have been disconnected already,
1636    the only thing left to do is
1637      - clear the recovery flags
1638      - cancel the timer
1639      - free queued requests and replies, but don't send replies
1640    Because the obd_stopping flag is set, no new requests should be received.
1641
1642 */
1643 void target_cleanup_recovery(struct obd_device *obd)
1644 {
1645         struct ptlrpc_request *req, *n;
1646         struct list_head clean_list;
1647
1648         INIT_LIST_HEAD(&clean_list);
1649         spin_lock(&obd->obd_dev_lock);
1650         if (!obd->obd_recovering) {
1651                 spin_unlock(&obd->obd_dev_lock);
1652                 EXIT;
1653                 return;
1654         }
1655         obd->obd_recovering = obd->obd_abort_recovery = 0;
1656         spin_unlock(&obd->obd_dev_lock);
1657
1658         spin_lock(&obd->obd_recovery_task_lock);
1659         target_cancel_recovery_timer(obd);
1660         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1661         spin_unlock(&obd->obd_recovery_task_lock);
1662
1663         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1664                 LASSERT(req->rq_reply_state == NULL);
1665                 target_exp_dequeue_req_replay(req);
1666                 target_request_copy_put(req);
1667         }
1668
1669         spin_lock(&obd->obd_recovery_task_lock);
1670         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1671         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1672         spin_unlock(&obd->obd_recovery_task_lock);
1673
1674         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1675                 LASSERT(req->rq_reply_state == NULL);
1676                 target_request_copy_put(req);
1677         }
1678
1679         EXIT;
1680 }
1681 EXPORT_SYMBOL(target_cleanup_recovery);
1682
1683 /* obd_recovery_task_lock should be held */
1684 void target_cancel_recovery_timer(struct obd_device *obd)
1685 {
1686         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1687         cfs_timer_disarm(&obd->obd_recovery_timer);
1688 }
1689
1690 static void target_start_recovery_timer(struct obd_device *obd)
1691 {
1692         if (obd->obd_recovery_start != 0)
1693                 return;
1694
1695         spin_lock(&obd->obd_dev_lock);
1696         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1697                 spin_unlock(&obd->obd_dev_lock);
1698                 return;
1699         }
1700
1701         LASSERT(obd->obd_recovery_timeout != 0);
1702
1703         if (obd->obd_recovery_start != 0) {
1704                 spin_unlock(&obd->obd_dev_lock);
1705                 return;
1706         }
1707
1708         cfs_timer_arm(&obd->obd_recovery_timer,
1709                       cfs_time_shift(obd->obd_recovery_timeout));
1710         obd->obd_recovery_start = cfs_time_current_sec();
1711         spin_unlock(&obd->obd_dev_lock);
1712
1713         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1714                       "or until %d client%s reconnect%s\n",
1715                       obd->obd_name,
1716                       obd->obd_recovery_timeout / 60,
1717                       obd->obd_recovery_timeout % 60,
1718                       obd->obd_max_recoverable_clients,
1719                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1720                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1721 }
1722
1723 /**
1724  * extend recovery window.
1725  *
1726  * if @extend is true, extend recovery window to have @drt remaining at least;
1727  * otherwise, make sure the recovery timeout value is not less than @drt.
1728  */
1729 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1730 {
1731         cfs_time_t now;
1732         cfs_time_t end;
1733         cfs_duration_t left;
1734         int to;
1735
1736         spin_lock(&obd->obd_dev_lock);
1737         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1738                 spin_unlock(&obd->obd_dev_lock);
1739                 return;
1740         }
1741         LASSERT(obd->obd_recovery_start != 0);
1742
1743         now  = cfs_time_current_sec();
1744         to   = obd->obd_recovery_timeout;
1745         end  = obd->obd_recovery_start + to;
1746         left = cfs_time_sub(end, now);
1747
1748         if (extend && (drt > left)) {
1749                 to += drt - left;
1750         } else if (!extend && (drt > to)) {
1751                 to = drt;
1752         }
1753
1754         if (to > obd->obd_recovery_time_hard)
1755                 to = obd->obd_recovery_time_hard;
1756         if (obd->obd_recovery_timeout < to) {
1757                 obd->obd_recovery_timeout = to;
1758                 end = obd->obd_recovery_start + to;
1759                 cfs_timer_arm(&obd->obd_recovery_timer,
1760                                 cfs_time_shift(end - now));
1761         }
1762         spin_unlock(&obd->obd_dev_lock);
1763
1764         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1765                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1766 }
1767
1768 /* Reset the timer with each new client connection */
1769 /*
1770  * This timer is actually reconnect_timer, which is for making sure
1771  * the total recovery window is at least as big as my reconnect
1772  * attempt timing. So the initial recovery time_out will be set to
1773  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1774  * from client is bigger than this, then the recovery time_out will
1775  * be extended to make sure the client could be reconnected, in the
1776  * process, the timeout from the new client should be ignored.
1777  */
1778
1779 static void
1780 check_and_start_recovery_timer(struct obd_device *obd,
1781                                struct ptlrpc_request *req,
1782                                int new_client)
1783 {
1784         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1785         struct obd_device_target *obt = &obd->u.obt;
1786
1787         if (!new_client && service_time)
1788                 /* Teach server about old server's estimates, as first guess
1789                  * at how long new requests will take. */
1790                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1791                             service_time);
1792
1793         target_start_recovery_timer(obd);
1794
1795         /* Convert the service time to RPC timeout,
1796          * and reuse service_time to limit stack usage. */
1797         service_time = at_est2timeout(service_time);
1798
1799         /* We expect other clients to timeout within service_time, then try
1800          * to reconnect, then try the failover server.  The max delay between
1801          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1802         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1803
1804         LASSERT(obt->obt_magic == OBT_MAGIC);
1805         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1806         if (service_time > obd->obd_recovery_timeout && !new_client)
1807                 extend_recovery_timer(obd, service_time, false);
1808 }
1809
1810 /** Health checking routines */
1811 static inline int exp_connect_healthy(struct obd_export *exp)
1812 {
1813         return (exp->exp_in_recovery);
1814 }
1815
1816 /** if export done req_replay or has replay in queue */
1817 static inline int exp_req_replay_healthy(struct obd_export *exp)
1818 {
1819         return (!exp->exp_req_replay_needed ||
1820                 atomic_read(&exp->exp_replay_count) > 0);
1821 }
1822
1823
1824 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1825 {
1826         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1827                exp_req_replay_healthy(exp);
1828 }
1829
1830 /** if export done lock_replay or has replay in queue */
1831 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1832 {
1833         return (!exp->exp_lock_replay_needed ||
1834                 atomic_read(&exp->exp_replay_count) > 0);
1835 }
1836
1837 static inline int exp_vbr_healthy(struct obd_export *exp)
1838 {
1839         return (!exp->exp_vbr_failed);
1840 }
1841
1842 static inline int exp_finished(struct obd_export *exp)
1843 {
1844         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1845 }
1846
1847 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1848 {
1849         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1850                 exp_finished(exp);
1851 }
1852
1853 static int check_for_next_transno(struct lu_target *lut)
1854 {
1855         struct ptlrpc_request *req = NULL;
1856         struct obd_device *obd = lut->lut_obd;
1857         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1858         int wake_up = 0, connected, completed, queue_len;
1859         __u64 req_transno = 0;
1860         __u64 update_transno = 0;
1861         __u64 next_transno = 0;
1862         ENTRY;
1863
1864         spin_lock(&obd->obd_recovery_task_lock);
1865         if (!list_empty(&obd->obd_req_replay_queue)) {
1866                 req = list_entry(obd->obd_req_replay_queue.next,
1867                                      struct ptlrpc_request, rq_list);
1868                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1869         }
1870
1871         if (tdtd != NULL)
1872                 update_transno = distribute_txn_get_next_transno(tdtd);
1873
1874         connected = atomic_read(&obd->obd_connected_clients);
1875         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1876         queue_len = obd->obd_requests_queued_for_recovery;
1877         next_transno = obd->obd_next_recovery_transno;
1878
1879         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1880                "req_transno: "LPU64", next_transno: "LPU64"\n",
1881                obd->obd_max_recoverable_clients, connected, completed,
1882                queue_len, req_transno, next_transno);
1883
1884         if (obd->obd_abort_recovery) {
1885                 CDEBUG(D_HA, "waking for aborted recovery\n");
1886                 wake_up = 1;
1887         } else if (obd->obd_recovery_expired) {
1888                 CDEBUG(D_HA, "waking for expired recovery\n");
1889                 wake_up = 1;
1890         } else if (tdtd != NULL && req != NULL &&
1891                    is_req_replayed_by_update(req)) {
1892                 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1893                          "next_transno"LPU64"\n", req_transno, next_transno);
1894                 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1895                        req_transno);
1896                 wake_up = 1;
1897         } else if (req_transno == next_transno ||
1898                    (update_transno != 0 && update_transno <= next_transno)) {
1899                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1900                 wake_up = 1;
1901         } else if (queue_len > 0 &&
1902                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1903                 int d_lvl = D_HA;
1904                 /** handle gaps occured due to lost reply or VBR */
1905                 LASSERTF(req_transno >= next_transno,
1906                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1907                          req_transno, next_transno);
1908                 if (req_transno > obd->obd_last_committed &&
1909                     !obd->obd_version_recov)
1910                         d_lvl = D_ERROR;
1911                 CDEBUG(d_lvl,
1912                        "%s: waking for gap in transno, VBR is %s (skip: "
1913                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1914                        ", next_update "LPD64" last_committed: "LPD64")\n",
1915                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1916                        next_transno, queue_len, completed, connected,
1917                        req_transno, update_transno, obd->obd_last_committed);
1918                 obd->obd_next_recovery_transno = req_transno;
1919                 wake_up = 1;
1920         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1921                 CDEBUG(D_HA, "waking for completed recovery\n");
1922                 wake_up = 1;
1923         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1924                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1925                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1926                 obd->obd_next_recovery_transno = req_transno;
1927                 wake_up = 1;
1928         }
1929         spin_unlock(&obd->obd_recovery_task_lock);
1930         return wake_up;
1931 }
1932
1933 static int check_for_next_lock(struct lu_target *lut)
1934 {
1935         struct obd_device *obd = lut->lut_obd;
1936         int wake_up = 0;
1937
1938         spin_lock(&obd->obd_recovery_task_lock);
1939         if (!list_empty(&obd->obd_lock_replay_queue)) {
1940                 CDEBUG(D_HA, "waking for next lock\n");
1941                 wake_up = 1;
1942         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1943                 CDEBUG(D_HA, "waking for completed lock replay\n");
1944                 wake_up = 1;
1945         } else if (obd->obd_abort_recovery) {
1946                 CDEBUG(D_HA, "waking for aborted recovery\n");
1947                 wake_up = 1;
1948         } else if (obd->obd_recovery_expired) {
1949                 CDEBUG(D_HA, "waking for expired recovery\n");
1950                 wake_up = 1;
1951         }
1952         spin_unlock(&obd->obd_recovery_task_lock);
1953
1954         return wake_up;
1955 }
1956
1957 /**
1958  * wait for recovery events,
1959  * check its status with help of check_routine
1960  * evict dead clients via health_check
1961  */
1962 static int target_recovery_overseer(struct lu_target *lut,
1963                                     int (*check_routine)(struct lu_target *),
1964                                     int (*health_check)(struct obd_export *))
1965 {
1966         struct obd_device       *obd = lut->lut_obd;
1967         struct target_distribute_txn_data *tdtd;
1968 repeat:
1969         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1970               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1971                 __u64 next_update_transno = 0;
1972
1973                 /* Only abort the recovery if there are no update recovery
1974                  * left in the queue */
1975                 spin_lock(&obd->obd_recovery_task_lock);
1976                 if (lut->lut_tdtd != NULL) {
1977                         next_update_transno =
1978                                 distribute_txn_get_next_transno(lut->lut_tdtd);
1979
1980                         tdtd = lut->lut_tdtd;
1981                         /* If next_update_transno == 0, it probably because
1982                          * updatelog retrieve threads did not get any records
1983                          * yet, let's wait those threads stopped */
1984                         if (next_update_transno == 0) {
1985                                 struct l_wait_info lwi = { 0 };
1986
1987                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1988                                        atomic_read(
1989                                        &tdtd->tdtd_recovery_threads_count) == 0,
1990                                        &lwi);
1991
1992                                 next_update_transno =
1993                                         distribute_txn_get_next_transno(
1994                                                                 lut->lut_tdtd);
1995                         }
1996                 }
1997
1998                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1999                         obd->obd_next_recovery_transno = next_update_transno;
2000                         spin_unlock(&obd->obd_recovery_task_lock);
2001                         /* Disconnect unfinished exports from clients, and
2002                          * keep connection from MDT to make sure the update
2003                          * recovery will still keep trying until some one
2004                          * manually abort the recovery */
2005                         class_disconnect_stale_exports(obd,
2006                                                 exp_finished_or_from_mdt);
2007                         /* Abort all of replay and replay lock req from
2008                          * clients */
2009                         abort_req_replay_queue(obd);
2010                         abort_lock_replay_queue(obd);
2011                         CDEBUG(D_HA, "%s: there are still update replay ("LPX64
2012                                ")in the queue.\n", obd->obd_name,
2013                                next_update_transno);
2014                 } else {
2015                         obd->obd_abort_recovery = 1;
2016                         spin_unlock(&obd->obd_recovery_task_lock);
2017                         CWARN("%s recovery is aborted by hard timeout\n",
2018                               obd->obd_name);
2019                 }
2020         }
2021
2022         while (wait_event_timeout(obd->obd_next_transno_waitq,
2023                                   check_routine(lut),
2024                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2025                 /* wait indefinitely for event, but don't trigger watchdog */;
2026
2027         if (obd->obd_abort_recovery) {
2028                 CWARN("recovery is aborted, evict exports in recovery\n");
2029                 if (lut->lut_tdtd != NULL) {
2030                         struct l_wait_info lwi = { 0 };
2031
2032                         tdtd = lut->lut_tdtd;
2033                         /* Let's wait all of the update log recovery thread
2034                          * finished */
2035                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2036                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2037                              &lwi);
2038                         /* Then abort the update recovery list */
2039                         dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2040                         dtrq_list_destroy(lut->lut_tdtd);
2041                 }
2042
2043                 /** evict exports which didn't finish recovery yet */
2044                 class_disconnect_stale_exports(obd, exp_finished);
2045                 return 1;
2046         } else if (obd->obd_recovery_expired) {
2047                 obd->obd_recovery_expired = 0;
2048                 /** If some clients died being recovered, evict them */
2049                 LCONSOLE_WARN("%s: recovery is timed out, "
2050                               "evict stale exports\n", obd->obd_name);
2051                 /** evict cexports with no replay in queue, they are stalled */
2052                 class_disconnect_stale_exports(obd, health_check);
2053
2054                 /** continue with VBR */
2055                 spin_lock(&obd->obd_dev_lock);
2056                 obd->obd_version_recov = 1;
2057                 spin_unlock(&obd->obd_dev_lock);
2058                 /**
2059                  * reset timer, recovery will proceed with versions now,
2060                  * timeout is set just to handle reconnection delays
2061                  */
2062                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2063                 /** Wait for recovery events again, after evicting bad clients */
2064                 goto repeat;
2065         }
2066         return 0;
2067 }
2068
2069 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2070 {
2071         struct obd_device       *obd = lut->lut_obd;
2072         struct ptlrpc_request *req = NULL;
2073
2074         CDEBUG(D_HA, "Waiting for lock\n");
2075         if (target_recovery_overseer(lut, check_for_next_lock,
2076                                      exp_lock_replay_healthy))
2077                 abort_lock_replay_queue(obd);
2078
2079         spin_lock(&obd->obd_recovery_task_lock);
2080         if (!list_empty(&obd->obd_lock_replay_queue)) {
2081                 req = list_entry(obd->obd_lock_replay_queue.next,
2082                                      struct ptlrpc_request, rq_list);
2083                 list_del_init(&req->rq_list);
2084                 spin_unlock(&obd->obd_recovery_task_lock);
2085         } else {
2086                 spin_unlock(&obd->obd_recovery_task_lock);
2087                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2088                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2089                 /** evict exports failed VBR */
2090                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2091         }
2092         return req;
2093 }
2094
2095 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2096 {
2097         struct ptlrpc_request *req = NULL;
2098
2099         spin_lock(&obd->obd_recovery_task_lock);
2100         if (!list_empty(&obd->obd_final_req_queue)) {
2101                 req = list_entry(obd->obd_final_req_queue.next,
2102                                      struct ptlrpc_request, rq_list);
2103                 list_del_init(&req->rq_list);
2104                 spin_unlock(&obd->obd_recovery_task_lock);
2105                 if (req->rq_export->exp_in_recovery) {
2106                         spin_lock(&req->rq_export->exp_lock);
2107                         req->rq_export->exp_in_recovery = 0;
2108                         spin_unlock(&req->rq_export->exp_lock);
2109                 }
2110         } else {
2111                 spin_unlock(&obd->obd_recovery_task_lock);
2112         }
2113         return req;
2114 }
2115
2116 static void handle_recovery_req(struct ptlrpc_thread *thread,
2117                                 struct ptlrpc_request *req,
2118                                 svc_handler_t handler)
2119 {
2120         ENTRY;
2121
2122         /**
2123          * export can be evicted during recovery, no need to handle replays for
2124          * it after that, discard such request silently
2125          */
2126         if (req->rq_export->exp_disconnected)
2127                 RETURN_EXIT;
2128
2129         req->rq_session.lc_thread = thread;
2130         req->rq_svc_thread = thread;
2131         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2132
2133         /* thread context */
2134         lu_context_enter(&thread->t_env->le_ctx);
2135         (void)handler(req);
2136         lu_context_exit(&thread->t_env->le_ctx);
2137
2138         /* don't reset timer for final stage */
2139         if (!exp_finished(req->rq_export)) {
2140                 int to = obd_timeout;
2141
2142                 /**
2143                  * Add request timeout to the recovery time so next request from
2144                  * this client may come in recovery time
2145                  */
2146                 if (!AT_OFF) {
2147                         struct ptlrpc_service_part *svcpt;
2148
2149                         svcpt = req->rq_rqbd->rqbd_svcpt;
2150                         /* If the server sent early reply for this request,
2151                          * the client will recalculate the timeout according to
2152                          * current server estimate service time, so we will
2153                          * use the maxium timeout here for waiting the client
2154                          * sending the next req */
2155                         to = max((int)at_est2timeout(
2156                                  at_get(&svcpt->scp_at_estimate)),
2157                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2158                         /* Add 2 net_latency, one for balance rq_deadline
2159                          * (see ptl_send_rpc), one for resend the req to server,
2160                          * Note: client will pack net_latency in replay req
2161                          * (see ptlrpc_replay_req) */
2162                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2163                 }
2164                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2165         }
2166         EXIT;
2167 }
2168
2169 /** Checking routines for recovery */
2170 static int check_for_recovery_ready(struct lu_target *lut)
2171 {
2172         struct obd_device *obd = lut->lut_obd;
2173         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2174
2175         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2176                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2177                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2178                obd->obd_recovery_expired);
2179
2180         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2181                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2182                 if (clnts + obd->obd_stale_clients <
2183                     obd->obd_max_recoverable_clients)
2184                         return 0;
2185         }
2186
2187         if (lut->lut_tdtd != NULL) {
2188                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2189                     !obd->obd_abort_recovery) {
2190                         /* Let's extend recovery timer, in case the recovery
2191                          * timer expired, and some clients got evicted */
2192                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2193                                               true);
2194                         CDEBUG(D_HA, "%s update recovery is not ready,"
2195                                " extend recovery %d\n", obd->obd_name,
2196                                obd->obd_recovery_timeout);
2197                         return 0;
2198                 } else {
2199                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2200                 }
2201         }
2202
2203         return 1;
2204 }
2205
2206 enum {
2207         REQUEST_RECOVERY = 1,
2208         UPDATE_RECOVERY = 2,
2209 };
2210
2211 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2212 {
2213         __u64 transno = 0;
2214
2215         if (!list_empty(&obd->obd_req_replay_queue)) {
2216                 struct ptlrpc_request *req;
2217
2218                 req = list_entry(obd->obd_req_replay_queue.next,
2219                                  struct ptlrpc_request, rq_list);
2220                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2221         }
2222
2223         return transno;
2224 }
2225
2226 static __u64 get_next_transno(struct lu_target *lut, int *type)
2227 {
2228         struct obd_device *obd = lut->lut_obd;
2229         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2230         __u64 transno = 0;
2231         __u64 update_transno;
2232         ENTRY;
2233
2234         transno = get_next_replay_req_transno(obd);
2235         if (type != NULL)
2236                 *type = REQUEST_RECOVERY;
2237
2238         if (tdtd == NULL)
2239                 RETURN(transno);
2240
2241         update_transno = distribute_txn_get_next_transno(tdtd);
2242         if (transno == 0 || (transno >= update_transno &&
2243                              update_transno != 0)) {
2244                 transno = update_transno;
2245                 if (type != NULL)
2246                         *type = UPDATE_RECOVERY;
2247         }
2248
2249         RETURN(transno);
2250 }
2251
2252 /**
2253  * drop duplicate replay request
2254  *
2255  * Because the operation has been replayed by update recovery, the request
2256  * with the same transno will be dropped and also notify the client to send
2257  * next replay request.
2258  *
2259  * \param[in] env       execution environment
2260  * \param[in] obd       failover obd device
2261  * \param[in] req       request to be dropped
2262  */
2263 static void drop_duplicate_replay_req(struct lu_env *env,
2264                                       struct obd_device *obd,
2265                                       struct ptlrpc_request *req)
2266 {
2267         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2268                   " update records are found.\n",
2269                   lustre_msg_get_transno(req->rq_reqmsg),
2270                   libcfs_nid2str(req->rq_peer.nid));
2271
2272         /* Right now, only for MDS reint operation update replay and
2273          * normal request replay can have the same transno */
2274         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2275                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2276                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2277                 if (likely(req->rq_export))
2278                         target_committed_to_req(req);
2279                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2280                 target_send_reply(req, req->rq_status, 0);
2281         } else {
2282                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2283                 libcfs_nid2str(req->rq_peer.nid));
2284         }
2285         target_exp_dequeue_req_replay(req);
2286         target_request_copy_put(req);
2287         obd->obd_replayed_requests++;
2288 }
2289
2290 static void replay_request_or_update(struct lu_env *env,
2291                                      struct lu_target *lut,
2292                                      struct target_recovery_data *trd,
2293                                      struct ptlrpc_thread *thread)
2294 {
2295         struct obd_device *obd = lut->lut_obd;
2296         struct ptlrpc_request *req = NULL;
2297         int                     type;
2298         __u64                   transno;
2299         ENTRY;
2300
2301         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2302                obd->obd_next_recovery_transno);
2303
2304         /* Replay all of request and update by transno */
2305         do {
2306                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2307
2308                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2309
2310                 /** It is needed to extend recovery window above
2311                  *  recovery_time_soft. Extending is possible only in the
2312                  *  end of recovery window (see more details in
2313                  *  handle_recovery_req()).
2314                  */
2315                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2316
2317                 if (target_recovery_overseer(lut, check_for_next_transno,
2318                                         exp_req_replay_healthy_or_from_mdt)) {
2319                         abort_req_replay_queue(obd);
2320                         abort_lock_replay_queue(obd);
2321                         goto abort;
2322                 }
2323
2324                 spin_lock(&obd->obd_recovery_task_lock);
2325                 transno = get_next_transno(lut, &type);
2326                 if (type == REQUEST_RECOVERY && transno != 0) {
2327                         /* Drop replay request from client side, if the
2328                          * replay has been executed by update with the
2329                          * same transno */
2330                         req = list_entry(obd->obd_req_replay_queue.next,
2331                                         struct ptlrpc_request, rq_list);
2332
2333                         list_del_init(&req->rq_list);
2334                         obd->obd_requests_queued_for_recovery--;
2335                         spin_unlock(&obd->obd_recovery_task_lock);
2336
2337                         /* Let's check if the request has been redone by
2338                          * update replay */
2339                         if (is_req_replayed_by_update(req)) {
2340                                 struct distribute_txn_replay_req *dtrq;
2341
2342                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2343                                                                    req->rq_xid);
2344                                 LASSERT(dtrq != NULL);
2345                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2346                                 list_del_init(&dtrq->dtrq_list);
2347                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2348                                 dtrq_destroy(dtrq);
2349
2350                                 drop_duplicate_replay_req(env, obd, req);
2351
2352                                 continue;
2353                         }
2354
2355                         LASSERT(trd->trd_processing_task == current_pid());
2356                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2357                                   lustre_msg_get_transno(req->rq_reqmsg),
2358                                   libcfs_nid2str(req->rq_peer.nid));
2359
2360                         handle_recovery_req(thread, req,
2361                                             trd->trd_recovery_handler);
2362                         /**
2363                          * bz18031: increase next_recovery_transno before
2364                          * target_request_copy_put() will drop exp_rpc reference
2365                          */
2366                         spin_lock(&obd->obd_recovery_task_lock);
2367                         obd->obd_next_recovery_transno++;
2368                         spin_unlock(&obd->obd_recovery_task_lock);
2369                         target_exp_dequeue_req_replay(req);
2370                         target_request_copy_put(req);
2371                         obd->obd_replayed_requests++;
2372                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2373                         struct distribute_txn_replay_req *dtrq;
2374                         int rc;
2375
2376                         spin_unlock(&obd->obd_recovery_task_lock);
2377
2378                         LASSERT(tdtd != NULL);
2379                         dtrq = distribute_txn_get_next_req(tdtd);
2380                         lu_context_enter(&thread->t_env->le_ctx);
2381                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2382                         lu_context_exit(&thread->t_env->le_ctx);
2383                         extend_recovery_timer(obd, obd_timeout, true);
2384
2385                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2386                                 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2387                                        " to finish list\n", dtrq->dtrq_xid,
2388                                        dtrq->dtrq_master_transno);
2389
2390                                 /* Add it to the replay finish list */
2391                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2392                                 list_add(&dtrq->dtrq_list,
2393                                          &tdtd->tdtd_replay_finish_list);
2394                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2395
2396                                 spin_lock(&obd->obd_recovery_task_lock);
2397                                 if (transno == obd->obd_next_recovery_transno)
2398                                         obd->obd_next_recovery_transno++;
2399                                 else if (transno >
2400                                          obd->obd_next_recovery_transno)
2401                                         obd->obd_next_recovery_transno =
2402                                                                 transno + 1;
2403                                 spin_unlock(&obd->obd_recovery_task_lock);
2404                         } else {
2405                                 dtrq_destroy(dtrq);
2406                         }
2407                 } else {
2408                         spin_unlock(&obd->obd_recovery_task_lock);
2409 abort:
2410                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2411                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2412                         /** evict exports failed VBR */
2413                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2414                         break;
2415                 }
2416         } while (1);
2417 }
2418
2419 static int target_recovery_thread(void *arg)
2420 {
2421         struct lu_target *lut = arg;
2422         struct obd_device *obd = lut->lut_obd;
2423         struct ptlrpc_request *req;
2424         struct target_recovery_data *trd = &obd->obd_recovery_data;
2425         unsigned long delta;
2426         struct lu_env *env;
2427         struct ptlrpc_thread *thread = NULL;
2428         int rc = 0;
2429         ENTRY;
2430
2431         unshare_fs_struct();
2432         OBD_ALLOC_PTR(thread);
2433         if (thread == NULL)
2434                 RETURN(-ENOMEM);
2435
2436         OBD_ALLOC_PTR(env);
2437         if (env == NULL) {
2438                 OBD_FREE_PTR(thread);
2439                 RETURN(-ENOMEM);
2440         }
2441
2442         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2443         if (rc) {
2444                 OBD_FREE_PTR(thread);
2445                 OBD_FREE_PTR(env);
2446                 RETURN(rc);
2447         }
2448
2449         thread->t_env = env;
2450         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2451         env->le_ctx.lc_thread = thread;
2452         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2453         thread->t_watchdog = NULL;
2454
2455         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2456                current_pid());
2457         trd->trd_processing_task = current_pid();
2458
2459         spin_lock(&obd->obd_dev_lock);
2460         obd->obd_recovering = 1;
2461         spin_unlock(&obd->obd_dev_lock);
2462         complete(&trd->trd_starting);
2463
2464         /* first of all, we have to know the first transno to replay */
2465         if (target_recovery_overseer(lut, check_for_recovery_ready,
2466                                      exp_connect_healthy)) {
2467                 abort_req_replay_queue(obd);
2468                 abort_lock_replay_queue(obd);
2469                 if (lut->lut_tdtd != NULL)
2470                         dtrq_list_destroy(lut->lut_tdtd);
2471         }
2472
2473         /* next stage: replay requests or update */
2474         delta = jiffies;
2475         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2476                atomic_read(&obd->obd_req_replay_clients),
2477                obd->obd_next_recovery_transno);
2478         replay_request_or_update(env, lut, trd, thread);
2479
2480         /**
2481          * The second stage: replay locks
2482          */
2483         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2484                atomic_read(&obd->obd_lock_replay_clients));
2485         while ((req = target_next_replay_lock(lut))) {
2486                 LASSERT(trd->trd_processing_task == current_pid());
2487                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2488                           libcfs_nid2str(req->rq_peer.nid));
2489                 handle_recovery_req(thread, req,
2490                                     trd->trd_recovery_handler);
2491                 target_request_copy_put(req);
2492                 obd->obd_replayed_locks++;
2493         }
2494
2495         /**
2496          * The third stage: reply on final pings, at this moment all clients
2497          * must have request in final queue
2498          */
2499         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2500         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2501         /** Update server last boot epoch */
2502         tgt_boot_epoch_update(lut);
2503         /* We drop recoverying flag to forward all new requests
2504          * to regular mds_handle() since now */
2505         spin_lock(&obd->obd_dev_lock);
2506         obd->obd_recovering = obd->obd_abort_recovery = 0;
2507         spin_unlock(&obd->obd_dev_lock);
2508         spin_lock(&obd->obd_recovery_task_lock);
2509         target_cancel_recovery_timer(obd);
2510         spin_unlock(&obd->obd_recovery_task_lock);
2511         while ((req = target_next_final_ping(obd))) {
2512                 LASSERT(trd->trd_processing_task == current_pid());
2513                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2514                           libcfs_nid2str(req->rq_peer.nid));
2515                 handle_recovery_req(thread, req,
2516                                     trd->trd_recovery_handler);
2517                 /* Because the waiting client can not send ping to server,
2518                  * so we need refresh the last_request_time, to avoid the
2519                  * export is being evicted */
2520                 ptlrpc_update_export_timer(req->rq_export, 0);
2521                 target_request_copy_put(req);
2522         }
2523
2524         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2525         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2526               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2527         if (delta > OBD_RECOVERY_TIME_SOFT) {
2528                 CWARN("too long recovery - read logs\n");
2529                 libcfs_debug_dumplog();
2530         }
2531
2532         target_finish_recovery(lut);
2533
2534         lu_context_fini(&env->le_ctx);
2535         trd->trd_processing_task = 0;
2536         complete(&trd->trd_finishing);
2537
2538         tgt_io_thread_done(thread);
2539         OBD_FREE_PTR(thread);
2540         OBD_FREE_PTR(env);
2541         RETURN(rc);
2542 }
2543
2544 static int target_start_recovery_thread(struct lu_target *lut,
2545                                         svc_handler_t handler)
2546 {
2547         struct obd_device *obd = lut->lut_obd;
2548         int rc = 0;
2549         struct target_recovery_data *trd = &obd->obd_recovery_data;
2550         int index;
2551
2552         memset(trd, 0, sizeof(*trd));
2553         init_completion(&trd->trd_starting);
2554         init_completion(&trd->trd_finishing);
2555         trd->trd_recovery_handler = handler;
2556
2557         rc = server_name2index(obd->obd_name, &index, NULL);
2558         if (rc < 0)
2559                 return rc;
2560
2561         if (!IS_ERR(kthread_run(target_recovery_thread,
2562                                 lut, "tgt_recover_%d", index))) {
2563                 wait_for_completion(&trd->trd_starting);
2564                 LASSERT(obd->obd_recovering != 0);
2565         } else {
2566                 rc = -ECHILD;
2567         }
2568
2569         return rc;
2570 }
2571
2572 void target_stop_recovery_thread(struct obd_device *obd)
2573 {
2574         if (obd->obd_recovery_data.trd_processing_task > 0) {
2575                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2576                 /** recovery can be done but postrecovery is not yet */
2577                 spin_lock(&obd->obd_dev_lock);
2578                 if (obd->obd_recovering) {
2579                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2580                         obd->obd_abort_recovery = 1;
2581                         wake_up(&obd->obd_next_transno_waitq);
2582                 }
2583                 spin_unlock(&obd->obd_dev_lock);
2584                 wait_for_completion(&trd->trd_finishing);
2585         }
2586 }
2587 EXPORT_SYMBOL(target_stop_recovery_thread);
2588
2589 void target_recovery_fini(struct obd_device *obd)
2590 {
2591         class_disconnect_exports(obd);
2592         target_stop_recovery_thread(obd);
2593         target_cleanup_recovery(obd);
2594 }
2595 EXPORT_SYMBOL(target_recovery_fini);
2596
2597 static void target_recovery_expired(unsigned long castmeharder)
2598 {
2599         struct obd_device *obd = (struct obd_device *)castmeharder;
2600         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2601                " after %lds (%d clients connected)\n",
2602                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2603                cfs_time_current_sec()- obd->obd_recovery_start,
2604                atomic_read(&obd->obd_connected_clients));
2605
2606         obd->obd_recovery_expired = 1;
2607         wake_up(&obd->obd_next_transno_waitq);
2608 }
2609
2610 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2611 {
2612         struct obd_device *obd = lut->lut_obd;
2613
2614         if (obd->obd_max_recoverable_clients == 0) {
2615                 /** Update server last boot epoch */
2616                 tgt_boot_epoch_update(lut);
2617                 return;
2618         }
2619
2620         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2621                "last_transno "LPU64"\n", obd->obd_name,
2622                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2623         LASSERT(obd->obd_stopping == 0);
2624         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2625         obd->obd_recovery_start = 0;
2626         obd->obd_recovery_end = 0;
2627
2628         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2629         target_start_recovery_thread(lut, handler);
2630 }
2631 EXPORT_SYMBOL(target_recovery_init);
2632
2633 static int target_process_req_flags(struct obd_device *obd,
2634                                     struct ptlrpc_request *req)
2635 {
2636         struct obd_export *exp = req->rq_export;
2637         LASSERT(exp != NULL);
2638         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2639                 /* client declares he's ready to replay locks */
2640                 spin_lock(&exp->exp_lock);
2641                 if (exp->exp_req_replay_needed) {
2642                         exp->exp_req_replay_needed = 0;
2643                         spin_unlock(&exp->exp_lock);
2644
2645                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2646                         atomic_dec(&obd->obd_req_replay_clients);
2647                 } else {
2648                         spin_unlock(&exp->exp_lock);
2649                 }
2650         }
2651         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2652                 /* client declares he's ready to complete recovery
2653                  * so, we put the request on th final queue */
2654                 spin_lock(&exp->exp_lock);
2655                 if (exp->exp_lock_replay_needed) {
2656                         exp->exp_lock_replay_needed = 0;
2657                         spin_unlock(&exp->exp_lock);
2658
2659                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2660                         atomic_dec(&obd->obd_lock_replay_clients);
2661                 } else {
2662                         spin_unlock(&exp->exp_lock);
2663                 }
2664         }
2665         return 0;
2666 }
2667
2668 int target_queue_recovery_request(struct ptlrpc_request *req,
2669                                   struct obd_device *obd)
2670 {
2671         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2672         struct ptlrpc_request *reqiter;
2673         int inserted = 0;
2674         ENTRY;
2675
2676         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2677                 /* Processing the queue right now, don't re-add. */
2678                 RETURN(1);
2679         }
2680
2681         target_process_req_flags(obd, req);
2682
2683         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2684                 /* client declares he's ready to complete recovery
2685                  * so, we put the request on th final queue */
2686                 target_request_copy_get(req);
2687                 DEBUG_REQ(D_HA, req, "queue final req");
2688                 wake_up(&obd->obd_next_transno_waitq);
2689                 spin_lock(&obd->obd_recovery_task_lock);
2690                 if (obd->obd_recovering) {
2691                         list_add_tail(&req->rq_list,
2692                                           &obd->obd_final_req_queue);
2693                 } else {
2694                         spin_unlock(&obd->obd_recovery_task_lock);
2695                         target_request_copy_put(req);
2696                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2697                 }
2698                 spin_unlock(&obd->obd_recovery_task_lock);
2699                 RETURN(0);
2700         }
2701         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2702                 /* client declares he's ready to replay locks */
2703                 target_request_copy_get(req);
2704                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2705                 wake_up(&obd->obd_next_transno_waitq);
2706                 spin_lock(&obd->obd_recovery_task_lock);
2707                 LASSERT(obd->obd_recovering);
2708                 /* usually due to recovery abort */
2709                 if (!req->rq_export->exp_in_recovery) {
2710                         spin_unlock(&obd->obd_recovery_task_lock);
2711                         target_request_copy_put(req);
2712                         RETURN(-ENOTCONN);
2713                 }
2714                 LASSERT(req->rq_export->exp_lock_replay_needed);
2715                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2716                 spin_unlock(&obd->obd_recovery_task_lock);
2717                 RETURN(0);
2718         }
2719
2720         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2721          * (i.e. buflens etc are in my own byte order), but type-dependent
2722          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2723
2724         if (!transno) {
2725                 INIT_LIST_HEAD(&req->rq_list);
2726                 DEBUG_REQ(D_HA, req, "not queueing");
2727                 RETURN(1);
2728         }
2729
2730         /* If we're processing the queue, we want don't want to queue this
2731          * message.
2732          *
2733          * Also, if this request has a transno less than the one we're waiting
2734          * for, we should process it now.  It could (and currently always will)
2735          * be an open request for a descriptor that was opened some time ago.
2736          *
2737          * Also, a resent, replayed request that has already been
2738          * handled will pass through here and be processed immediately.
2739          */
2740         CDEBUG(D_HA, "Next recovery transno: "LPU64
2741                ", current: "LPU64", replaying\n",
2742                obd->obd_next_recovery_transno, transno);
2743
2744         /* If the request has been replayed by update replay, then sends this
2745          * request to the recovery thread (replay_request_or_update()), where
2746          * it will be handled */
2747         spin_lock(&obd->obd_recovery_task_lock);
2748         if (transno < obd->obd_next_recovery_transno &&
2749             !is_req_replayed_by_update(req)) {
2750                 /* Processing the queue right now, don't re-add. */
2751                 LASSERT(list_empty(&req->rq_list));
2752                 spin_unlock(&obd->obd_recovery_task_lock);
2753                 RETURN(1);
2754         }
2755         spin_unlock(&obd->obd_recovery_task_lock);
2756
2757         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2758                 RETURN(0);
2759
2760         target_request_copy_get(req);
2761         if (!req->rq_export->exp_in_recovery) {
2762                 target_request_copy_put(req);
2763                 RETURN(-ENOTCONN);
2764         }
2765         LASSERT(req->rq_export->exp_req_replay_needed);
2766
2767         if (target_exp_enqueue_req_replay(req)) {
2768                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2769                 target_request_copy_put(req);
2770                 RETURN(0);
2771         }
2772
2773         /* XXX O(n^2) */
2774         spin_lock(&obd->obd_recovery_task_lock);
2775         LASSERT(obd->obd_recovering);
2776         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2777                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2778                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2779                         inserted = 1;
2780                         goto added;
2781                 }
2782
2783                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2784                              transno)) {
2785                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2786                                   "has been claimed by another client");
2787                         spin_unlock(&obd->obd_recovery_task_lock);
2788                         target_exp_dequeue_req_replay(req);
2789                         target_request_copy_put(req);
2790                         RETURN(0);
2791                 }
2792         }
2793 added:
2794         if (!inserted)
2795                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2796
2797         obd->obd_requests_queued_for_recovery++;
2798         spin_unlock(&obd->obd_recovery_task_lock);
2799         wake_up(&obd->obd_next_transno_waitq);
2800         RETURN(0);
2801 }
2802
2803 int target_handle_ping(struct ptlrpc_request *req)
2804 {
2805         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2806         return req_capsule_server_pack(&req->rq_pill);
2807 }
2808
2809 void target_committed_to_req(struct ptlrpc_request *req)
2810 {
2811         struct obd_export *exp = req->rq_export;
2812
2813         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2814                 lustre_msg_set_last_committed(req->rq_repmsg,
2815                                               exp->exp_last_committed);
2816         else
2817                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2818                           "%d)", exp->exp_obd->obd_no_transno,
2819                           req->rq_repmsg == NULL);
2820
2821         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2822                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2823 }
2824
2825 #endif /* HAVE_SERVER_SUPPORT */
2826
2827 /**
2828  * Packs current SLV and Limit into \a req.
2829  */
2830 int target_pack_pool_reply(struct ptlrpc_request *req)
2831 {
2832         struct obd_device *obd;
2833         ENTRY;
2834
2835         /* Check that we still have all structures alive as this may
2836          * be some late RPC at shutdown time. */
2837         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2838                      !exp_connect_lru_resize(req->rq_export))) {
2839                 lustre_msg_set_slv(req->rq_repmsg, 0);
2840                 lustre_msg_set_limit(req->rq_repmsg, 0);
2841                 RETURN(0);
2842         }
2843
2844         /* OBD is alive here as export is alive, which we checked above. */
2845         obd = req->rq_export->exp_obd;
2846
2847         read_lock(&obd->obd_pool_lock);
2848         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2849         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2850         read_unlock(&obd->obd_pool_lock);
2851
2852         RETURN(0);
2853 }
2854
2855 static int target_send_reply_msg(struct ptlrpc_request *req,
2856                                  int rc, int fail_id)
2857 {
2858         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2859                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2860                 return -ECOMM;
2861         }
2862         /* We can have a null rq_reqmsg in the event of bad signature or
2863          * no context when unwrapping */
2864         if (req->rq_reqmsg &&
2865             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2866             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2867                 return -ECOMM;
2868
2869         if (unlikely(rc)) {
2870                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2871                 req->rq_status = rc;
2872                 return ptlrpc_send_error(req, 1);
2873         } else {
2874                 DEBUG_REQ(D_NET, req, "sending reply");
2875         }
2876
2877         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2878 }
2879
2880 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2881 {
2882         struct ptlrpc_service_part *svcpt;
2883         int                        netrc;
2884         struct ptlrpc_reply_state *rs;
2885         struct obd_export         *exp;
2886         ENTRY;
2887
2888         if (req->rq_no_reply) {
2889                 EXIT;
2890                 return;
2891         }
2892
2893         svcpt = req->rq_rqbd->rqbd_svcpt;
2894         rs = req->rq_reply_state;
2895         if (rs == NULL || !rs->rs_difficult) {
2896                 /* no notifiers */
2897                 target_send_reply_msg (req, rc, fail_id);
2898                 EXIT;
2899                 return;
2900         }
2901
2902         /* must be an export if locks saved */
2903         LASSERT(req->rq_export != NULL);
2904         /* req/reply consistent */
2905         LASSERT(rs->rs_svcpt == svcpt);
2906
2907         /* "fresh" reply */
2908         LASSERT(!rs->rs_scheduled);
2909         LASSERT(!rs->rs_scheduled_ever);
2910         LASSERT(!rs->rs_handled);
2911         LASSERT(!rs->rs_on_net);
2912         LASSERT(rs->rs_export == NULL);
2913         LASSERT(list_empty(&rs->rs_obd_list));
2914         LASSERT(list_empty(&rs->rs_exp_list));
2915
2916         exp = class_export_get(req->rq_export);
2917
2918         /* disable reply scheduling while I'm setting up */
2919         rs->rs_scheduled = 1;
2920         rs->rs_on_net    = 1;
2921         rs->rs_xid       = req->rq_xid;
2922         rs->rs_transno   = req->rq_transno;
2923         rs->rs_export    = exp;
2924         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2925
2926         spin_lock(&exp->exp_uncommitted_replies_lock);
2927         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2928                rs->rs_transno, exp->exp_last_committed);
2929         if (rs->rs_transno > exp->exp_last_committed) {
2930                 /* not committed already */
2931                 list_add_tail(&rs->rs_obd_list,
2932                                   &exp->exp_uncommitted_replies);
2933         }
2934         spin_unlock(&exp->exp_uncommitted_replies_lock);
2935
2936         spin_lock(&exp->exp_lock);
2937         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2938         spin_unlock(&exp->exp_lock);
2939
2940         netrc = target_send_reply_msg(req, rc, fail_id);
2941
2942         spin_lock(&svcpt->scp_rep_lock);
2943
2944         atomic_inc(&svcpt->scp_nreps_difficult);
2945
2946         if (netrc != 0) {
2947                 /* error sending: reply is off the net.  Also we need +1
2948                  * reply ref until ptlrpc_handle_rs() is done
2949                  * with the reply state (if the send was successful, there
2950                  * would have been +1 ref for the net, which
2951                  * reply_out_callback leaves alone) */
2952                 rs->rs_on_net = 0;
2953                 ptlrpc_rs_addref(rs);
2954         }
2955
2956         spin_lock(&rs->rs_lock);
2957         if (rs->rs_transno <= exp->exp_last_committed ||
2958             (!rs->rs_on_net && !rs->rs_no_ack) ||
2959             list_empty(&rs->rs_exp_list) ||     /* completed already */
2960             list_empty(&rs->rs_obd_list)) {
2961                 CDEBUG(D_HA, "Schedule reply immediately\n");
2962                 ptlrpc_dispatch_difficult_reply(rs);
2963         } else {
2964                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2965                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2966         }
2967         spin_unlock(&rs->rs_lock);
2968         spin_unlock(&svcpt->scp_rep_lock);
2969         EXIT;
2970 }
2971
2972 enum ldlm_mode lck_compat_array[] = {
2973         [LCK_EX]    = LCK_COMPAT_EX,
2974         [LCK_PW]    = LCK_COMPAT_PW,
2975         [LCK_PR]    = LCK_COMPAT_PR,
2976         [LCK_CW]    = LCK_COMPAT_CW,
2977         [LCK_CR]    = LCK_COMPAT_CR,
2978         [LCK_NL]    = LCK_COMPAT_NL,
2979         [LCK_GROUP] = LCK_COMPAT_GROUP,
2980         [LCK_COS]   = LCK_COMPAT_COS,
2981 };
2982
2983 /**
2984  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2985  * not escape to the user level.
2986  */
2987 int ldlm_error2errno(enum ldlm_error error)
2988 {
2989         int result;
2990
2991         switch (error) {
2992         case ELDLM_OK:
2993         case ELDLM_LOCK_MATCHED:
2994                 result = 0;
2995                 break;
2996         case ELDLM_LOCK_CHANGED:
2997                 result = -ESTALE;
2998                 break;
2999         case ELDLM_LOCK_ABORTED:
3000                 result = -ENAVAIL;
3001                 break;
3002         case ELDLM_LOCK_REPLACED:
3003                 result = -ESRCH;
3004                 break;
3005         case ELDLM_NO_LOCK_DATA:
3006                 result = -ENOENT;
3007                 break;
3008         case ELDLM_NAMESPACE_EXISTS:
3009                 result = -EEXIST;
3010                 break;
3011         case ELDLM_BAD_NAMESPACE:
3012                 result = -EBADF;
3013                 break;
3014         default:
3015                 if (((int)error) < 0) { /* cast to signed type */
3016                         result = error; /* as ldlm_error can be unsigned */
3017                 } else {
3018                         CERROR("Invalid DLM result code: %d\n", error);
3019                         result = -EPROTO;
3020                 }
3021         }
3022         return result;
3023 }
3024 EXPORT_SYMBOL(ldlm_error2errno);
3025
3026 /**
3027  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3028  */
3029 enum ldlm_error ldlm_errno2error(int err_no)
3030 {
3031         int error;
3032
3033         switch (err_no) {
3034         case 0:
3035                 error = ELDLM_OK;
3036                 break;
3037         case -ESTALE:
3038                 error = ELDLM_LOCK_CHANGED;
3039                 break;
3040         case -ENAVAIL:
3041                 error = ELDLM_LOCK_ABORTED;
3042                 break;
3043         case -ESRCH:
3044                 error = ELDLM_LOCK_REPLACED;
3045                 break;
3046         case -ENOENT:
3047                 error = ELDLM_NO_LOCK_DATA;
3048                 break;
3049         case -EEXIST:
3050                 error = ELDLM_NAMESPACE_EXISTS;
3051                 break;
3052         case -EBADF:
3053                 error = ELDLM_BAD_NAMESPACE;
3054                 break;
3055         default:
3056                 error = err_no;
3057         }
3058         return error;
3059 }
3060
3061 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3062 void ldlm_dump_export_locks(struct obd_export *exp)
3063 {
3064         spin_lock(&exp->exp_locks_list_guard);
3065         if (!list_empty(&exp->exp_locks_list)) {
3066                 struct ldlm_lock *lock;
3067
3068                 CERROR("dumping locks for export %p,"
3069                        "ignore if the unmount doesn't hang\n", exp);
3070                 list_for_each_entry(lock, &exp->exp_locks_list,
3071                                         l_exp_refs_link)
3072                         LDLM_ERROR(lock, "lock:");
3073         }
3074         spin_unlock(&exp->exp_locks_list_guard);
3075 }
3076 #endif
3077
3078 #ifdef HAVE_SERVER_SUPPORT
3079 static int target_bulk_timeout(void *data)
3080 {
3081         ENTRY;
3082         /* We don't fail the connection here, because having the export
3083          * killed makes the (vital) call to commitrw very sad.
3084          */
3085         RETURN(1);
3086 }
3087
3088 static inline const char *bulk2type(struct ptlrpc_request *req)
3089 {
3090         if (req->rq_bulk_read)
3091                 return "READ";
3092         if (req->rq_bulk_write)
3093                 return "WRITE";
3094         return "UNKNOWN";
3095 }
3096
3097 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3098                    struct l_wait_info *lwi)
3099 {
3100         struct ptlrpc_request   *req = desc->bd_req;
3101         time_t                   start = cfs_time_current_sec();
3102         time_t                   deadline;
3103         int                      rc = 0;
3104
3105         ENTRY;
3106
3107         /* If there is eviction in progress, wait for it to finish. */
3108         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3109                 *lwi = LWI_INTR(NULL, NULL);
3110                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3111                                   !atomic_read(&exp->exp_obd->
3112                                                    obd_evict_inprogress),
3113                                   lwi);
3114         }
3115
3116         /* Check if client was evicted or reconnected already. */
3117         if (exp->exp_failed ||
3118             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3119                 rc = -ENOTCONN;
3120         } else {
3121                 if (req->rq_bulk_read)
3122                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3123
3124                 if ((exp->exp_connect_data.ocd_connect_flags &
3125                      OBD_CONNECT_BULK_MBITS) != 0)
3126                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3127                 else /* old version, bulk matchbits is rq_xid */
3128                         req->rq_mbits = req->rq_xid;
3129
3130                 if (rc == 0)
3131                         rc = ptlrpc_start_bulk_transfer(desc);
3132         }
3133
3134         if (rc < 0) {
3135                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3136                           bulk2type(req), rc);
3137                 RETURN(rc);
3138         }
3139
3140         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3141                 ptlrpc_abort_bulk(desc);
3142                 RETURN(0);
3143         }
3144
3145         /* limit actual bulk transfer to bulk_timeout seconds */
3146         deadline = start + bulk_timeout;
3147         if (deadline > req->rq_deadline)
3148                 deadline = req->rq_deadline;
3149
3150         do {
3151                 long timeoutl = deadline - cfs_time_current_sec();
3152                 cfs_duration_t timeout = timeoutl <= 0 ?
3153                                          CFS_TICK : cfs_time_seconds(timeoutl);
3154                 time_t  rq_deadline;
3155
3156                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3157                                             target_bulk_timeout, desc);
3158                 rc = l_wait_event(desc->bd_waitq,
3159                                   !ptlrpc_server_bulk_active(desc) ||
3160                                   exp->exp_failed ||
3161                                   exp->exp_conn_cnt >
3162                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3163                                   lwi);
3164                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3165                 /* Wait again if we changed rq_deadline. */
3166                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3167                 deadline = start + bulk_timeout;
3168                 if (deadline > rq_deadline)
3169                         deadline = rq_deadline;
3170         } while ((rc == -ETIMEDOUT) &&
3171                  (deadline > cfs_time_current_sec()));
3172
3173         if (rc == -ETIMEDOUT) {
3174                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3175                           bulk2type(req), deadline - start,
3176                           cfs_time_current_sec() - deadline);
3177                 ptlrpc_abort_bulk(desc);
3178         } else if (exp->exp_failed) {
3179                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3180                           bulk2type(req));
3181                 rc = -ENOTCONN;
3182                 ptlrpc_abort_bulk(desc);
3183         } else if (exp->exp_conn_cnt >
3184                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3185                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3186                           bulk2type(req));
3187                 /* We don't reply anyway. */
3188                 rc = -ETIMEDOUT;
3189                 ptlrpc_abort_bulk(desc);
3190         } else if (desc->bd_failure) {
3191                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3192                           bulk2type(req));
3193                 /* XXX should this be a different errno? */
3194                 rc = -ETIMEDOUT;
3195         } else {
3196                 if (req->rq_bulk_write)
3197                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3198                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3199                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3200                                   bulk2type(req), desc->bd_nob_transferred,
3201                                   desc->bd_nob);
3202                         /* XXX should this be a different errno? */
3203                         rc = -ETIMEDOUT;
3204                 }
3205         }
3206
3207         RETURN(rc);
3208 }
3209 EXPORT_SYMBOL(target_bulk_io);
3210
3211 #endif /* HAVE_SERVER_SUPPORT */