lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 /**
  38  * This file deals with various client/target related logic including recovery.
  39  *
  40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  41  * should be moved.
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LDLM
  45
  46 #include <linux/kthread.h>
  47 #include <libcfs/libcfs.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_dlm.h>
  51 #include <lustre_net.h>
  52 #include <lustre_sec.h>
  53 #include "ldlm_internal.h"
  54
  55 /* @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         int rc = 0;
  64         ENTRY;
  65
  66         if (!create && !priority) {
  67                 CDEBUG(D_HA, "Nothing to do\n");
  68                 RETURN(-EINVAL);
  69         }
  70
  71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  72         if (!ptlrpc_conn) {
  73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  74                 RETURN (-ENOENT);
  75         }
  76
  77         if (create) {
  78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  79                 if (!imp_conn) {
  80                         GOTO(out_put, rc = -ENOMEM);
  81                 }
  82         }
  83
  84         spin_lock(&imp->imp_lock);
  85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  87                         if (priority) {
  88                                 list_del(&item->oic_item);
  89                                 list_add(&item->oic_item,
  90                                              &imp->imp_conn_list);
  91                                 item->oic_last_attempt = 0;
  92                         }
  93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  94                                imp, imp->imp_obd->obd_name, uuid->uuid,
  95                                (priority ? ", moved to head" : ""));
  96                         spin_unlock(&imp->imp_lock);
  97                         GOTO(out_free, rc = 0);
  98                 }
  99         }
 100         /* No existing import connection found for \a uuid. */
 101         if (create) {
 102                 imp_conn->oic_conn = ptlrpc_conn;
 103                 imp_conn->oic_uuid = *uuid;
 104                 imp_conn->oic_last_attempt = 0;
 105                 if (priority)
 106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 else
 108                         list_add_tail(&imp_conn->oic_item,
 109                                           &imp->imp_conn_list);
 110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 111                        imp, imp->imp_obd->obd_name, uuid->uuid,
 112                        (priority ? "head" : "tail"));
 113         } else {
 114                 spin_unlock(&imp->imp_lock);
 115                 GOTO(out_free, rc = -ENOENT);
 116         }
 117
 118         spin_unlock(&imp->imp_lock);
 119         RETURN(0);
 120 out_free:
 121         if (imp_conn)
 122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 123 out_put:
 124         ptlrpc_connection_put(ptlrpc_conn);
 125         RETURN(rc);
 126 }
 127
 128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 129 {
 130         return import_set_conn(imp, uuid, 1, 0);
 131 }
 132
 133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 134                            int priority)
 135 {
 136         return import_set_conn(imp, uuid, priority, 1);
 137 }
 138 EXPORT_SYMBOL(client_import_add_conn);
 139
 140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 141 {
 142         struct obd_import_conn *imp_conn;
 143         struct obd_export *dlmexp;
 144         int rc = -ENOENT;
 145         ENTRY;
 146
 147         spin_lock(&imp->imp_lock);
 148         if (list_empty(&imp->imp_conn_list)) {
 149                 LASSERT(!imp->imp_connection);
 150                 GOTO(out, rc);
 151         }
 152
 153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 155                         continue;
 156                 LASSERT(imp_conn->oic_conn);
 157
 158                 if (imp_conn == imp->imp_conn_current) {
 159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 160
 161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 162                             imp->imp_state != LUSTRE_IMP_DISCON) {
 163                                 CERROR("can't remove current connection\n");
 164                                 GOTO(out, rc = -EBUSY);
 165                         }
 166
 167                         ptlrpc_connection_put(imp->imp_connection);
 168                         imp->imp_connection = NULL;
 169
 170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 171                         if (dlmexp && dlmexp->exp_connection) {
 172                                 LASSERT(dlmexp->exp_connection ==
 173                                         imp_conn->oic_conn);
 174                                 ptlrpc_connection_put(dlmexp->exp_connection);
 175                                 dlmexp->exp_connection = NULL;
 176                         }
 177                 }
 178
 179                 list_del(&imp_conn->oic_item);
 180                 ptlrpc_connection_put(imp_conn->oic_conn);
 181                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 182                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 183                        imp, imp->imp_obd->obd_name, uuid->uuid);
 184                 rc = 0;
 185                 break;
 186         }
 187 out:
 188         spin_unlock(&imp->imp_lock);
 189         if (rc == -ENOENT)
 190                 CERROR("connection %s not found\n", uuid->uuid);
 191         RETURN(rc);
 192 }
 193 EXPORT_SYMBOL(client_import_del_conn);
 194
 195 /**
 196  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 197  * to find a conn uuid of \a imp which can reach \a peer.
 198  */
 199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 200                             struct obd_uuid *uuid)
 201 {
 202         struct obd_import_conn *conn;
 203         int rc = -ENOENT;
 204         ENTRY;
 205
 206         spin_lock(&imp->imp_lock);
 207         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 208                 /* Check if conn UUID does have this peer NID. */
 209                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 210                         *uuid = conn->oic_uuid;
 211                         rc = 0;
 212                         break;
 213                 }
 214         }
 215         spin_unlock(&imp->imp_lock);
 216         RETURN(rc);
 217 }
 218 EXPORT_SYMBOL(client_import_find_conn);
 219
 220 void client_destroy_import(struct obd_import *imp)
 221 {
 222         /* Drop security policy instance after all RPCs have finished/aborted
 223          * to let all busy contexts be released. */
 224         class_import_get(imp);
 225         class_destroy_import(imp);
 226         sptlrpc_import_sec_put(imp);
 227         class_import_put(imp);
 228 }
 229 EXPORT_SYMBOL(client_destroy_import);
 230
 231 /**
 232  * Check whether or not the OSC is on MDT.
 233  * In the config log,
 234  * osc on MDT
 235  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 236  * osc on client
 237  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 238  *
 239  **/
 240 static int osc_on_mdt(char *obdname)
 241 {
 242         char *ptr;
 243
 244         ptr = strrchr(obdname, '-');
 245         if (ptr == NULL)
 246                 return 0;
 247
 248         if (strncmp(ptr + 1, "MDT", 3) == 0)
 249                 return 1;
 250
 251         return 0;
 252 }
 253
 254 /* Configure an RPC client OBD device.
 255  *
 256  * lcfg parameters:
 257  * 1 - client UUID
 258  * 2 - server UUID
 259  * 3 - inactive-on-startup
 260  */
 261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 262 {
 263         struct client_obd *cli = &obddev->u.cli;
 264         struct obd_import *imp;
 265         struct obd_uuid server_uuid;
 266         int rq_portal, rp_portal, connect_op;
 267         char *name = obddev->obd_type->typ_name;
 268         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 269         char *cli_name = lustre_cfg_buf(lcfg, 0);
 270         int rc;
 271         ENTRY;
 272
 273         /* In a more perfect world, we would hang a ptlrpc_client off of
 274          * obd_type and just use the values from there. */
 275         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 276                 rq_portal = OST_REQUEST_PORTAL;
 277                 rp_portal = OSC_REPLY_PORTAL;
 278                 connect_op = OST_CONNECT;
 279                 cli->cl_sp_me = LUSTRE_SP_CLI;
 280                 cli->cl_sp_to = LUSTRE_SP_OST;
 281                 ns_type = LDLM_NS_TYPE_OSC;
 282         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 283                    !strcmp(name, LUSTRE_LWP_NAME)) {
 284                 rq_portal = MDS_REQUEST_PORTAL;
 285                 rp_portal = MDC_REPLY_PORTAL;
 286                 connect_op = MDS_CONNECT;
 287                 if (is_lwp_on_ost(cli_name))
 288                         cli->cl_sp_me = LUSTRE_SP_OST;
 289                 else if (is_lwp_on_mdt(cli_name))
 290                         cli->cl_sp_me = LUSTRE_SP_MDT;
 291                 else
 292                         cli->cl_sp_me = LUSTRE_SP_CLI;
 293                 cli->cl_sp_to = LUSTRE_SP_MDT;
 294                 ns_type = LDLM_NS_TYPE_MDC;
 295         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 296                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 297                         /* OSP_on_MDT for other MDTs */
 298                         connect_op = MDS_CONNECT;
 299                         cli->cl_sp_to = LUSTRE_SP_MDT;
 300                         ns_type = LDLM_NS_TYPE_MDC;
 301                         rq_portal = OUT_PORTAL;
 302                 } else {
 303                         /* OSP on MDT for OST */
 304                         connect_op = OST_CONNECT;
 305                         cli->cl_sp_to = LUSTRE_SP_OST;
 306                         ns_type = LDLM_NS_TYPE_OSC;
 307                         rq_portal = OST_REQUEST_PORTAL;
 308                 }
 309                 rp_portal = OSC_REPLY_PORTAL;
 310                 cli->cl_sp_me = LUSTRE_SP_MDT;
 311         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 312                 rq_portal = MGS_REQUEST_PORTAL;
 313                 rp_portal = MGC_REPLY_PORTAL;
 314                 connect_op = MGS_CONNECT;
 315                 cli->cl_sp_me = LUSTRE_SP_MGC;
 316                 cli->cl_sp_to = LUSTRE_SP_MGS;
 317                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 318                 ns_type = LDLM_NS_TYPE_MGC;
 319         } else {
 320                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 321                        name);
 322                 RETURN(-EINVAL);
 323         }
 324
 325         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 326                 CERROR("requires a TARGET UUID\n");
 327                 RETURN(-EINVAL);
 328         }
 329
 330         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 331                 CERROR("client UUID must be less than 38 characters\n");
 332                 RETURN(-EINVAL);
 333         }
 334
 335         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 336                 CERROR("setup requires a SERVER UUID\n");
 337                 RETURN(-EINVAL);
 338         }
 339
 340         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 341                 CERROR("target UUID must be less than 38 characters\n");
 342                 RETURN(-EINVAL);
 343         }
 344
 345         init_rwsem(&cli->cl_sem);
 346         mutex_init(&cli->cl_mgc_mutex);
 347         cli->cl_conn_count = 0;
 348         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 349                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 350                      sizeof(server_uuid)));
 351
 352         cli->cl_dirty_pages = 0;
 353         cli->cl_avail_grant = 0;
 354         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 355         /* cl_dirty_max_pages may be changed at connect time in
 356          * ptlrpc_connect_interpret(). */
 357         client_adjust_max_dirty(cli);
 358         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 359         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 360         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 361         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 362         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 363         spin_lock_init(&cli->cl_loi_list_lock);
 364         atomic_set(&cli->cl_pending_w_pages, 0);
 365         atomic_set(&cli->cl_pending_r_pages, 0);
 366         cli->cl_r_in_flight = 0;
 367         cli->cl_w_in_flight = 0;
 368
 369         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 370         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 371         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 372         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 373         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 374         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 375
 376         /* lru for osc. */
 377         INIT_LIST_HEAD(&cli->cl_lru_osc);
 378         atomic_set(&cli->cl_lru_shrinkers, 0);
 379         atomic_long_set(&cli->cl_lru_busy, 0);
 380         atomic_long_set(&cli->cl_lru_in_list, 0);
 381         INIT_LIST_HEAD(&cli->cl_lru_list);
 382         spin_lock_init(&cli->cl_lru_list_lock);
 383         atomic_long_set(&cli->cl_unstable_count, 0);
 384         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         /* This value may be reduced at connect time in
 401          * ptlrpc_connect_interpret() . We initialize it to only
 402          * 1MB until we know what the performance looks like.
 403          * In the future this should likely be increased. LU-1431 */
 404         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
 405                                           LNET_MTU >> PAGE_CACHE_SHIFT);
 406
 407         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 408          * it will be updated at OSC connection time. */
 409         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 410
 411         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 412                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 413         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 414                 cli->cl_max_rpcs_in_flight = 2;
 415         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 416                 cli->cl_max_rpcs_in_flight = 3;
 417         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 418                 cli->cl_max_rpcs_in_flight = 4;
 419         } else {
 420                 if (osc_on_mdt(obddev->obd_name))
 421                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 422                 else
 423                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 424         }
 425
 426         spin_lock_init(&cli->cl_mod_rpcs_lock);
 427         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 428         cli->cl_max_mod_rpcs_in_flight = 0;
 429         cli->cl_mod_rpcs_in_flight = 0;
 430         cli->cl_close_rpcs_in_flight = 0;
 431         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 432         cli->cl_mod_tag_bitmap = NULL;
 433
 434         if (connect_op == MDS_CONNECT) {
 435                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 436                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 437                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 438                 if (cli->cl_mod_tag_bitmap == NULL)
 439                         GOTO(err, rc = -ENOMEM);
 440         }
 441
 442         rc = ldlm_get_ref();
 443         if (rc) {
 444                 CERROR("ldlm_get_ref failed: %d\n", rc);
 445                 GOTO(err, rc);
 446         }
 447
 448         ptlrpc_init_client(rq_portal, rp_portal, name,
 449                            &obddev->obd_ldlm_client);
 450
 451         imp = class_new_import(obddev);
 452         if (imp == NULL)
 453                 GOTO(err_ldlm, rc = -ENOENT);
 454         imp->imp_client = &obddev->obd_ldlm_client;
 455         imp->imp_connect_op = connect_op;
 456         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 457                LUSTRE_CFG_BUFLEN(lcfg, 1));
 458         class_import_put(imp);
 459
 460         rc = client_import_add_conn(imp, &server_uuid, 1);
 461         if (rc) {
 462                 CERROR("can't add initial connection\n");
 463                 GOTO(err_import, rc);
 464         }
 465
 466         cli->cl_import = imp;
 467         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 468         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 469
 470         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 471                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 472                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 473                                name, obddev->obd_name,
 474                                cli->cl_target_uuid.uuid);
 475                         spin_lock(&imp->imp_lock);
 476                         imp->imp_deactive = 1;
 477                         spin_unlock(&imp->imp_lock);
 478                 }
 479         }
 480
 481         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 482                                                    LDLM_NAMESPACE_CLIENT,
 483                                                    LDLM_NAMESPACE_GREEDY,
 484                                                    ns_type);
 485         if (obddev->obd_namespace == NULL) {
 486                 CERROR("Unable to create client namespace - %s\n",
 487                        obddev->obd_name);
 488                 GOTO(err_import, rc = -ENOMEM);
 489         }
 490
 491         RETURN(rc);
 492
 493 err_import:
 494         class_destroy_import(imp);
 495 err_ldlm:
 496         ldlm_put_ref();
 497 err:
 498         if (cli->cl_mod_tag_bitmap != NULL)
 499                 OBD_FREE(cli->cl_mod_tag_bitmap,
 500                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 501         cli->cl_mod_tag_bitmap = NULL;
 502         RETURN(rc);
 503
 504 }
 505 EXPORT_SYMBOL(client_obd_setup);
 506
 507 int client_obd_cleanup(struct obd_device *obddev)
 508 {
 509         struct client_obd *cli = &obddev->u.cli;
 510         ENTRY;
 511
 512         ldlm_namespace_free_post(obddev->obd_namespace);
 513         obddev->obd_namespace = NULL;
 514
 515         obd_cleanup_client_import(obddev);
 516         LASSERT(obddev->u.cli.cl_import == NULL);
 517
 518         ldlm_put_ref();
 519
 520         if (cli->cl_mod_tag_bitmap != NULL)
 521                 OBD_FREE(cli->cl_mod_tag_bitmap,
 522                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 523         cli->cl_mod_tag_bitmap = NULL;
 524
 525         RETURN(0);
 526 }
 527 EXPORT_SYMBOL(client_obd_cleanup);
 528
 529 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 530 int client_connect_import(const struct lu_env *env,
 531                           struct obd_export **exp,
 532                           struct obd_device *obd, struct obd_uuid *cluuid,
 533                           struct obd_connect_data *data, void *localdata)
 534 {
 535         struct client_obd       *cli    = &obd->u.cli;
 536         struct obd_import       *imp    = cli->cl_import;
 537         struct obd_connect_data *ocd;
 538         struct lustre_handle    conn    = { 0 };
 539         int                     rc;
 540         bool                    is_mdc = false;
 541         ENTRY;
 542
 543         *exp = NULL;
 544         down_write(&cli->cl_sem);
 545         if (cli->cl_conn_count > 0)
 546                 GOTO(out_sem, rc = -EALREADY);
 547
 548         rc = class_connect(&conn, obd, cluuid);
 549         if (rc)
 550                 GOTO(out_sem, rc);
 551
 552         cli->cl_conn_count++;
 553         *exp = class_conn2export(&conn);
 554
 555         LASSERT(obd->obd_namespace);
 556
 557         imp->imp_dlm_handle = conn;
 558         rc = ptlrpc_init_import(imp);
 559         if (rc != 0)
 560                 GOTO(out_ldlm, rc);
 561
 562         ocd = &imp->imp_connect_data;
 563         if (data) {
 564                 *ocd = *data;
 565                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
 566                                  LUSTRE_MDC_NAME, 3) == 0;
 567                 if (is_mdc)
 568                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
 569                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 570         }
 571
 572         rc = ptlrpc_connect_import(imp);
 573         if (rc != 0) {
 574                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 575                 GOTO(out_ldlm, rc);
 576         }
 577         LASSERT(*exp != NULL && (*exp)->exp_connection);
 578
 579         if (data) {
 580                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 581                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 582                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 583                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 584                 /* clear the flag as it was not set and is not known
 585                  * by upper layers */
 586                 if (is_mdc)
 587                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 588         }
 589
 590         ptlrpc_pinger_add_import(imp);
 591
 592         EXIT;
 593
 594         if (rc) {
 595 out_ldlm:
 596                 cli->cl_conn_count--;
 597                 class_disconnect(*exp);
 598                 *exp = NULL;
 599         }
 600 out_sem:
 601         up_write(&cli->cl_sem);
 602
 603         return rc;
 604 }
 605 EXPORT_SYMBOL(client_connect_import);
 606
 607 int client_disconnect_export(struct obd_export *exp)
 608 {
 609         struct obd_device *obd = class_exp2obd(exp);
 610         struct client_obd *cli;
 611         struct obd_import *imp;
 612         int rc = 0, err;
 613         ENTRY;
 614
 615         if (!obd) {
 616                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 617                        exp, exp ? exp->exp_handle.h_cookie : -1);
 618                 RETURN(-EINVAL);
 619         }
 620
 621         cli = &obd->u.cli;
 622         imp = cli->cl_import;
 623
 624         down_write(&cli->cl_sem);
 625         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 626                 cli->cl_conn_count);
 627
 628         if (cli->cl_conn_count == 0) {
 629                 CERROR("disconnecting disconnected device (%s)\n",
 630                        obd->obd_name);
 631                 GOTO(out_disconnect, rc = -EINVAL);
 632         }
 633
 634         cli->cl_conn_count--;
 635         if (cli->cl_conn_count != 0)
 636                 GOTO(out_disconnect, rc = 0);
 637
 638         /* Mark import deactivated now, so we don't try to reconnect if any
 639          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 640          * fully deactivate the import, or that would drop all requests. */
 641         spin_lock(&imp->imp_lock);
 642         imp->imp_deactive = 1;
 643         spin_unlock(&imp->imp_lock);
 644
 645         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 646          * delete it regardless.  (It's safe to delete an import that was
 647          * never added.) */
 648         (void)ptlrpc_pinger_del_import(imp);
 649
 650         if (obd->obd_namespace != NULL) {
 651                 /* obd_force == local only */
 652                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 653                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 654                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 655         }
 656
 657         /* There's no need to hold sem while disconnecting an import,
 658          * and it may actually cause deadlock in GSS. */
 659         up_write(&cli->cl_sem);
 660         rc = ptlrpc_disconnect_import(imp, 0);
 661         down_write(&cli->cl_sem);
 662
 663         ptlrpc_invalidate_import(imp);
 664
 665         EXIT;
 666
 667 out_disconnect:
 668         /* Use server style - class_disconnect should be always called for
 669          * o_disconnect. */
 670         err = class_disconnect(exp);
 671         if (!rc && err)
 672                 rc = err;
 673
 674         up_write(&cli->cl_sem);
 675
 676         RETURN(rc);
 677 }
 678 EXPORT_SYMBOL(client_disconnect_export);
 679
 680 #ifdef HAVE_SERVER_SUPPORT
 681 int server_disconnect_export(struct obd_export *exp)
 682 {
 683         int rc;
 684         ENTRY;
 685
 686         /* Disconnect early so that clients can't keep using export. */
 687         rc = class_disconnect(exp);
 688         /* Close import to avoid sending any requests. */
 689         if (exp->exp_imp_reverse)
 690                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 691
 692         ldlm_bl_thread_wakeup();
 693
 694         /* complete all outstanding replies */
 695         spin_lock(&exp->exp_lock);
 696         while (!list_empty(&exp->exp_outstanding_replies)) {
 697                 struct ptlrpc_reply_state *rs =
 698                         list_entry(exp->exp_outstanding_replies.next,
 699                                        struct ptlrpc_reply_state, rs_exp_list);
 700                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 701
 702                 spin_lock(&svcpt->scp_rep_lock);
 703
 704                 list_del_init(&rs->rs_exp_list);
 705                 spin_lock(&rs->rs_lock);
 706                 ptlrpc_schedule_difficult_reply(rs);
 707                 spin_unlock(&rs->rs_lock);
 708
 709                 spin_unlock(&svcpt->scp_rep_lock);
 710         }
 711         spin_unlock(&exp->exp_lock);
 712
 713         RETURN(rc);
 714 }
 715 EXPORT_SYMBOL(server_disconnect_export);
 716
 717 /* --------------------------------------------------------------------------
 718  * from old lib/target.c
 719  * -------------------------------------------------------------------------- */
 720
 721 static int target_handle_reconnect(struct lustre_handle *conn,
 722                                    struct obd_export *exp,
 723                                    struct obd_uuid *cluuid)
 724 {
 725         struct lustre_handle *hdl;
 726         ENTRY;
 727
 728         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 729         if (exp->exp_connection && lustre_handle_is_used(hdl)) {
 730                 struct obd_device *target;
 731
 732                 target = exp->exp_obd;
 733
 734                 /* Might be a re-connect after a partition. */
 735                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 736                         if (target->obd_recovering) {
 737                                 int timeout = cfs_duration_sec(cfs_time_sub(
 738                                         cfs_timer_deadline(
 739                                         &target->obd_recovery_timer),
 740                                         cfs_time_current()));
 741
 742                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 743                                         "ing, waiting for %d clients in recov"
 744                                         "ery for %d:%.02d\n", target->obd_name,
 745                                         obd_uuid2str(&exp->exp_client_uuid),
 746                                         obd_export_nid2str(exp),
 747                                         target->obd_max_recoverable_clients,
 748                                         timeout / 60, timeout % 60);
 749                         } else {
 750                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 751                                         "reconnecting\n", target->obd_name,
 752                                         obd_uuid2str(&exp->exp_client_uuid),
 753                                         obd_export_nid2str(exp));
 754                         }
 755
 756                         conn->cookie = exp->exp_handle.h_cookie;
 757                         /* target_handle_connect() treats EALREADY and
 758                          * -EALREADY differently.  EALREADY means we are
 759                          * doing a valid reconnect from the same client. */
 760                         RETURN(EALREADY);
 761                 } else {
 762                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 763                                       "with handle "LPX64". Rejecting client "
 764                                       "with the same UUID trying to reconnect "
 765                                       "with handle "LPX64"\n", target->obd_name,
 766                                       obd_uuid2str(&exp->exp_client_uuid),
 767                                       obd_export_nid2str(exp),
 768                                       hdl->cookie, conn->cookie);
 769                         memset(conn, 0, sizeof *conn);
 770                         /* target_handle_connect() treats EALREADY and
 771                          * -EALREADY differently.  -EALREADY is an error
 772                          * (same UUID, different handle). */
 773                         RETURN(-EALREADY);
 774                 }
 775         }
 776
 777         conn->cookie = exp->exp_handle.h_cookie;
 778         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 779                cluuid->uuid, exp, conn->cookie);
 780         RETURN(0);
 781 }
 782
 783 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 784                           int error)
 785 {
 786         struct obd_export *exp = cb_data;
 787
 788         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 789                obd->obd_name, exp->exp_client_uuid.uuid);
 790
 791         spin_lock(&exp->exp_lock);
 792         exp->exp_need_sync = 0;
 793         spin_unlock(&exp->exp_lock);
 794         class_export_cb_put(exp);
 795 }
 796
 797 static void
 798 check_and_start_recovery_timer(struct obd_device *obd,
 799                                struct ptlrpc_request *req, int new_client);
 800
 801 /**
 802  * update flags for import during reconnect process
 803  */
 804 static int rev_import_flags_update(struct obd_import *revimp,
 805                                    struct ptlrpc_request *req)
 806 {
 807         int rc;
 808         struct obd_connect_data *data;
 809
 810         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 811
 812         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 813                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 814         else
 815                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 816
 817         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 818
 819         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 820         if (rc) {
 821                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 822                         revimp->imp_client->cli_name,
 823                         libcfs_id2str(req->rq_peer), rc);
 824                 return rc;
 825         }
 826
 827         return 0;
 828 }
 829
 830 /**
 831  * Allocate a new reverse import for an export.
 832  *
 833  * \retval -errno in case error hit
 834  * \retval 0 if reverse import correctly init
 835  **/
 836 int rev_import_init(struct obd_export *export)
 837 {
 838         struct obd_device *obd = export->exp_obd;
 839         struct obd_import *revimp;
 840
 841         LASSERT(export->exp_imp_reverse == NULL);
 842
 843         revimp = class_new_import(obd);
 844         if (revimp == NULL)
 845                 return -ENOMEM;
 846
 847         revimp->imp_remote_handle.cookie = 0ULL;
 848         revimp->imp_client = &obd->obd_ldlm_client;
 849         revimp->imp_dlm_fake = 1;
 850
 851         /* it is safe to connect import in new state as no sends possible */
 852         spin_lock(&export->exp_lock);
 853         export->exp_imp_reverse = revimp;
 854         spin_unlock(&export->exp_lock);
 855         class_import_put(revimp);
 856
 857         return 0;
 858 }
 859 EXPORT_SYMBOL(rev_import_init);
 860
 861 /**
 862  * Handle reconnect for an export.
 863  *
 864  * \param exp export to handle reconnect process
 865  * \param req client reconnect request
 866  *
 867  * \retval -rc in case securitfy flavor can't be changed
 868  * \retval 0 in case none problems
 869  */
 870 static int rev_import_reconnect(struct obd_export *exp,
 871                                 struct ptlrpc_request *req)
 872 {
 873         struct obd_import *revimp = exp->exp_imp_reverse;
 874         struct lustre_handle *lh;
 875         int rc;
 876
 877         /* avoid sending a request until import flags are changed */
 878         ptlrpc_import_enter_resend(revimp);
 879
 880         if (revimp->imp_connection != NULL)
 881                 ptlrpc_connection_put(revimp->imp_connection);
 882
 883         /*
 884          * client from recovery don't have a handle so we need to take from
 885          * request. it may produce situation when wrong client connected
 886          * to recovery as we trust a client uuid
 887          */
 888         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 889         revimp->imp_remote_handle = *lh;
 890
 891         /* unknown versions will be caught in
 892          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 893         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 894
 895         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 896
 897         rc = rev_import_flags_update(revimp, req);
 898         if (rc != 0) {
 899                 /* it is safe to still be in RECOVERY phase as we are not able
 900                  * to setup correct security flavor so requests are not able to
 901                  * be delivered correctly */
 902                 return rc;
 903         }
 904
 905         /* resend all rpc's via new connection */
 906         return ptlrpc_import_recovery_state_machine(revimp);
 907 }
 908
 909 int target_handle_connect(struct ptlrpc_request *req)
 910 {
 911         struct obd_device *target = NULL;
 912         struct obd_export *export = NULL;
 913         /* connect handle - filled from target_handle_reconnect in
 914          * reconnect case */
 915         struct lustre_handle conn;
 916         struct lustre_handle *tmp;
 917         struct obd_uuid tgtuuid;
 918         struct obd_uuid cluuid;
 919         char *str;
 920         int rc = 0;
 921         char *target_start;
 922         int target_len;
 923         bool     mds_conn = false, lw_client = false;
 924         bool     mds_mds_conn = false;
 925         bool     new_mds_mds_conn = false;
 926         struct obd_connect_data *data, *tmpdata;
 927         int size, tmpsize;
 928         lnet_nid_t *client_nid = NULL;
 929         ENTRY;
 930
 931         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 932
 933         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 934         if (str == NULL) {
 935                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 936                 GOTO(out, rc = -EINVAL);
 937         }
 938
 939         obd_str2uuid(&tgtuuid, str);
 940         target = class_uuid2obd(&tgtuuid);
 941         if (!target)
 942                 target = class_name2obd(str);
 943
 944         if (!target) {
 945                 deuuidify(str, NULL, &target_start, &target_len);
 946                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 947                                    "from %s (no target). If you are running "
 948                                    "an HA pair check that the target is "
 949                                    "mounted on the other server.\n", str,
 950                                    libcfs_nid2str(req->rq_peer.nid));
 951                 GOTO(out, rc = -ENODEV);
 952         }
 953
 954         spin_lock(&target->obd_dev_lock);
 955         /* Make sure the target isn't cleaned up while we're here. Yes,
 956          * there's still a race between the above check and our incref here.
 957          * Really, class_uuid2obd should take the ref. */
 958         class_incref(target, __func__, current);
 959
 960         if (target->obd_stopping || !target->obd_set_up) {
 961                 spin_unlock(&target->obd_dev_lock);
 962
 963                 deuuidify(str, NULL, &target_start, &target_len);
 964                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 965                               target_len, target_start,
 966                               libcfs_nid2str(req->rq_peer.nid),
 967                               (target->obd_stopping ?
 968                                "stopping" : "not set up"));
 969                 GOTO(out, rc = -ENODEV);
 970         }
 971
 972         if (target->obd_no_conn) {
 973                 spin_unlock(&target->obd_dev_lock);
 974
 975                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 976                                "from %s\n", target->obd_name,
 977                                libcfs_nid2str(req->rq_peer.nid));
 978                 GOTO(out, rc = -EAGAIN);
 979         }
 980
 981         target->obd_conn_inprogress++;
 982         spin_unlock(&target->obd_dev_lock);
 983
 984         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 985         if (str == NULL) {
 986                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 987                 GOTO(out, rc = -EINVAL);
 988         }
 989
 990         obd_str2uuid(&cluuid, str);
 991
 992         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 993         if (tmp == NULL)
 994                 GOTO(out, rc = -EPROTO);
 995
 996         conn = *tmp;
 997
 998         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 999                                     RCL_CLIENT);
1000         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1001         if (!data)
1002                 GOTO(out, rc = -EPROTO);
1003
1004         rc = req_capsule_server_pack(&req->rq_pill);
1005         if (rc)
1006                 GOTO(out, rc);
1007
1008 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1009         /* Don't allow clients to connect that are using old 1.8 format
1010          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1011          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1012          * FULL20 flag should be set on all connections since 2.0, but no
1013          * longer affects behaviour.
1014          *
1015          * Later this check will be disabled and the flag can be retired
1016          * completely once interop with 3.0 is no longer needed.
1017          */
1018         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1019                 GOTO(out, rc = -EPROTO);
1020 #endif
1021
1022         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1023                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1024                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1025                     data->ocd_version > LUSTRE_VERSION_CODE +
1026                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1027                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1028                                   "libclient connection attempt",
1029                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1030                                   "old" : "new",
1031                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1032                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1033                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1034                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1035                         data = req_capsule_server_sized_get(&req->rq_pill,
1036                                                             &RMF_CONNECT_DATA,
1037                                     offsetof(typeof(*data), ocd_version) +
1038                                              sizeof(data->ocd_version));
1039                         if (data) {
1040                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1041                                 data->ocd_version = LUSTRE_VERSION_CODE;
1042                         }
1043                         GOTO(out, rc = -EPROTO);
1044                 }
1045         }
1046
1047         /* Note: lw_client is needed in MDS-MDS failover during update log
1048          * processing, so we needs to allow lw_client to be connected at
1049          * anytime, instead of only the initial connection */
1050         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1051
1052         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1053                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1054                 mds_mds_conn = (data->ocd_connect_flags &
1055                                 OBD_CONNECT_MDS_MDS) != 0;
1056
1057                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1058                  * for Imperative Recovery connection from MGC to MGS.
1059                  *
1060                  * Via check OBD_CONNECT_FID, we can distinguish whether
1061                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1062                  * MGC or MDT. */
1063                 if (!lw_client &&
1064                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1065                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1066                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1067                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1068                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1069                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1070
1071                         /* We do not support the MDT-MDT interoperations with
1072                          * different version MDT because of protocol changes. */
1073                         if (unlikely(major != LUSTRE_MAJOR ||
1074                                      minor != LUSTRE_MINOR ||
1075                                      abs(patch - LUSTRE_PATCH) > 3)) {
1076                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1077                                         "connection from different version MDT "
1078                                         "(%d.%d.%d.%d) %s %s\n",
1079                                         target->obd_name, LUSTRE_MAJOR,
1080                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1081                                         major, minor, patch,
1082                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1083                                         libcfs_nid2str(req->rq_peer.nid), str);
1084
1085                                 GOTO(out, rc = -EPROTO);
1086                         }
1087                 }
1088         }
1089
1090         /* lctl gets a backstage, all-access pass. */
1091         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1092                 goto dont_check_exports;
1093
1094         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1095         if (!export)
1096                 goto no_export;
1097
1098         /* We've found an export in the hash. */
1099
1100         spin_lock(&export->exp_lock);
1101
1102         if (export->exp_connecting) { /* bug 9635, et. al. */
1103                 spin_unlock(&export->exp_lock);
1104                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1105                               export->exp_obd->obd_name, export,
1106                               libcfs_nid2str(req->rq_peer.nid));
1107                 class_export_put(export);
1108                 export = NULL;
1109                 rc = -EALREADY;
1110         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1111                 spin_unlock(&export->exp_lock);
1112                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1113                         /* MDS or LWP reconnected after failover. */
1114                         LCONSOLE_WARN("%s: Received %s connection from "
1115                             "%s, removing former export from %s\n",
1116                             target->obd_name, mds_conn ? "MDS" : "LWP",
1117                             libcfs_nid2str(req->rq_peer.nid),
1118                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1119                 else
1120                         /* New MDS connection from the same NID. */
1121                         LCONSOLE_WARN("%s: Received new %s connection from "
1122                                 "%s, removing former export from same NID\n",
1123                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1124                                 libcfs_nid2str(req->rq_peer.nid));
1125                 class_fail_export(export);
1126                 class_export_put(export);
1127                 export = NULL;
1128                 rc = 0;
1129         } else if (export->exp_connection != NULL &&
1130                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1131                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1132                     MSG_CONNECT_INITIAL)) {
1133                 spin_unlock(&export->exp_lock);
1134                 /* In MDS failover we have static UUID but NID can change. */
1135                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1136                               "existing nid %s is already connected\n",
1137                               target->obd_name, cluuid.uuid,
1138                               libcfs_nid2str(req->rq_peer.nid),
1139                               libcfs_nid2str(
1140                                       export->exp_connection->c_peer.nid));
1141                 rc = -EALREADY;
1142                 class_export_put(export);
1143                 export = NULL;
1144         } else {
1145                 export->exp_connecting = 1;
1146                 spin_unlock(&export->exp_lock);
1147                 LASSERT(export->exp_obd == target);
1148
1149                 rc = target_handle_reconnect(&conn, export, &cluuid);
1150         }
1151
1152         /* If we found an export, we already unlocked. */
1153         if (!export) {
1154 no_export:
1155                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1156         } else if (req->rq_export == NULL &&
1157                    atomic_read(&export->exp_rpc_count) > 0) {
1158                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1159                               "still busy with %d references\n",
1160                               target->obd_name, cluuid.uuid,
1161                               libcfs_nid2str(req->rq_peer.nid),
1162                               atomic_read(&export->exp_refcount));
1163                 GOTO(out, rc = -EBUSY);
1164         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1165                 if (!strstr(cluuid.uuid, "mdt"))
1166                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1167                                       "known client %s (at %s) because it "
1168                                       "is indicating it is a new client",
1169                                       target->obd_name, cluuid.uuid,
1170                                       libcfs_nid2str(req->rq_peer.nid));
1171                 GOTO(out, rc = -EALREADY);
1172         } else {
1173                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1174         }
1175
1176         if (rc < 0) {
1177                 GOTO(out, rc);
1178         }
1179
1180         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1181                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1182               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1183               export, (long)cfs_time_current_sec(),
1184               export ? (long)export->exp_last_request_time : 0);
1185
1186         /* If this is the first time a client connects, reset the recovery
1187          * timer. Discard lightweight connections which might be local. */
1188         if (!lw_client && rc == 0 && target->obd_recovering)
1189                 check_and_start_recovery_timer(target, req, export == NULL);
1190
1191         /* We want to handle EALREADY but *not* -EALREADY from
1192          * target_handle_reconnect(), return reconnection state in a flag. */
1193         if (rc == EALREADY) {
1194                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1195                 rc = 0;
1196         } else {
1197                 LASSERT(rc == 0);
1198         }
1199
1200         /* Tell the client if we support replayable requests. */
1201         if (target->obd_replayable)
1202                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1203         client_nid = &req->rq_peer.nid;
1204
1205         if (export == NULL) {
1206                 /* allow lightweight connections during recovery */
1207                 /* allow "new" MDT to be connected during recovery, since we
1208                  * need retrieve recovery update records from it */
1209                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1210                         cfs_time_t t;
1211                         int     c; /* connected */
1212                         int     i; /* in progress */
1213                         int     k; /* known */
1214                         int     s; /* stale/evicted */
1215
1216                         c = atomic_read(&target->obd_connected_clients);
1217                         i = atomic_read(&target->obd_lock_replay_clients);
1218                         k = target->obd_max_recoverable_clients;
1219                         s = target->obd_stale_clients;
1220                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1221                         t = cfs_time_sub(t, cfs_time_current());
1222                         t = cfs_duration_sec(t);
1223                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1224                                       "(at %s), waiting for %d known clients "
1225                                       "(%d recovered, %d in progress, and %d "
1226                                       "evicted) to recover in %d:%.02d\n",
1227                                       target->obd_name, cluuid.uuid,
1228                                       libcfs_nid2str(req->rq_peer.nid), k,
1229                                       c - i, i, s, (int)t / 60,
1230                                       (int)t % 60);
1231                         rc = -EBUSY;
1232                 } else {
1233 dont_check_exports:
1234                         rc = obd_connect(req->rq_svc_thread->t_env,
1235                                          &export, target, &cluuid, data,
1236                                          client_nid);
1237                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1238                                 lustre_msg_add_op_flags(req->rq_repmsg,
1239                                                         MSG_CONNECT_RECOVERING);
1240                         if (rc == 0) {
1241                                 conn.cookie = export->exp_handle.h_cookie;
1242                                 rc = rev_import_init(export);
1243                         }
1244
1245                         if (mds_mds_conn)
1246                                 new_mds_mds_conn = true;
1247                 }
1248         } else {
1249                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1250                                    export, target, &cluuid, data, client_nid);
1251         }
1252         if (rc)
1253                 GOTO(out, rc);
1254
1255         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1256         data->ocd_instance = target->u.obt.obt_instance;
1257
1258         /* Return only the parts of obd_connect_data that we understand, so the
1259          * client knows that we don't understand the rest. */
1260         if (data) {
1261                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1262                                                RCL_SERVER);
1263                 tmpdata = req_capsule_server_get(&req->rq_pill,
1264                                                  &RMF_CONNECT_DATA);
1265                 /* Don't use struct assignment here, because the client reply
1266                  * buffer may be smaller/larger than the local struct
1267                  * obd_connect_data. */
1268                 memcpy(tmpdata, data, min(tmpsize, size));
1269         }
1270
1271         /* If the client and the server are the same node, we will already
1272          * have an export that really points to the client's DLM export,
1273          * because we have a shared handles table.
1274          *
1275          * XXX this will go away when shaver stops sending the "connect" handle
1276          * in the real "remote handle" field of the request --phik 24 Apr 2003
1277          */
1278         ptlrpc_request_change_export(req, export);
1279
1280         spin_lock(&export->exp_lock);
1281         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1282                 spin_unlock(&export->exp_lock);
1283                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1284                        "or equal conn_cnt: %d >= %d\n",
1285                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1286                        export->exp_conn_cnt,
1287                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1288
1289                 GOTO(out, rc = -EALREADY);
1290         }
1291         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1292         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1293
1294         /* Don't evict liblustre clients for not pinging. */
1295         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1296                 export->exp_libclient = 1;
1297                 spin_unlock(&export->exp_lock);
1298
1299                 spin_lock(&target->obd_dev_lock);
1300                 list_del_init(&export->exp_obd_chain_timed);
1301                 spin_unlock(&target->obd_dev_lock);
1302         } else {
1303                 spin_unlock(&export->exp_lock);
1304         }
1305
1306         if (export->exp_connection != NULL) {
1307                 /* Check to see if connection came from another NID. */
1308                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1309                     !hlist_unhashed(&export->exp_nid_hash))
1310                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1311                                      &export->exp_connection->c_peer.nid,
1312                                      &export->exp_nid_hash);
1313
1314                 ptlrpc_connection_put(export->exp_connection);
1315         }
1316
1317         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1318                                                        req->rq_self,
1319                                                        &cluuid);
1320         if (hlist_unhashed(&export->exp_nid_hash))
1321                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1322                              &export->exp_connection->c_peer.nid,
1323                              &export->exp_nid_hash);
1324
1325         lustre_msg_set_handle(req->rq_repmsg, &conn);
1326
1327         rc = rev_import_reconnect(export, req);
1328         if (rc != 0)
1329                 GOTO(out, rc);
1330
1331         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1332                 int has_transno;
1333                 __u64 transno = data->ocd_transno;
1334
1335                 spin_lock(&export->exp_lock);
1336                 /* possible race with class_disconnect_stale_exports,
1337                  * export may be already in the eviction process */
1338                 if (export->exp_failed) {
1339                         spin_unlock(&export->exp_lock);
1340                         GOTO(out, rc = -ENODEV);
1341                 }
1342                 export->exp_in_recovery = 1;
1343                 export->exp_req_replay_needed = 1;
1344                 export->exp_lock_replay_needed = 1;
1345                 spin_unlock(&export->exp_lock);
1346
1347                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1348                                  MSG_CONNECT_TRANSNO);
1349                 if (has_transno && transno == 0)
1350                         CWARN("Connect with zero transno!\n");
1351
1352                 if (has_transno && transno > 0 &&
1353                     transno < target->obd_next_recovery_transno &&
1354                     transno > target->obd_last_committed) {
1355                         /* Another way is to use cmpxchg() to be lock-free. */
1356                         spin_lock(&target->obd_recovery_task_lock);
1357                         if (transno < target->obd_next_recovery_transno)
1358                                 target->obd_next_recovery_transno = transno;
1359                         spin_unlock(&target->obd_recovery_task_lock);
1360                 }
1361
1362                 atomic_inc(&target->obd_req_replay_clients);
1363                 atomic_inc(&target->obd_lock_replay_clients);
1364                 /* Note: MDS-MDS connection is allowed to be connected during
1365                  * recovery, no matter if the exports needs to be recoveried.
1366                  * Because we need retrieve updates logs from all other MDTs.
1367                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1368                  * also needs to be increased to match other recovery checking
1369                  * condition. */
1370                 if (new_mds_mds_conn)
1371                         target->obd_max_recoverable_clients++;
1372                 if (atomic_inc_return(&target->obd_connected_clients) ==
1373                     target->obd_max_recoverable_clients)
1374                         wake_up(&target->obd_next_transno_waitq);
1375         }
1376
1377         /* Tell the client we're in recovery, when client is involved in it. */
1378         if (target->obd_recovering && !lw_client)
1379                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1380
1381 out:
1382         if (export) {
1383                 spin_lock(&export->exp_lock);
1384                 export->exp_connecting = 0;
1385                 spin_unlock(&export->exp_lock);
1386
1387                 class_export_put(export);
1388         }
1389         if (target != NULL) {
1390                 spin_lock(&target->obd_dev_lock);
1391                 target->obd_conn_inprogress--;
1392                 spin_unlock(&target->obd_dev_lock);
1393
1394                 class_decref(target, __func__, current);
1395         }
1396         req->rq_status = rc;
1397         RETURN(rc);
1398 }
1399
1400 int target_handle_disconnect(struct ptlrpc_request *req)
1401 {
1402         int rc;
1403         ENTRY;
1404
1405         rc = req_capsule_server_pack(&req->rq_pill);
1406         if (rc)
1407                 RETURN(rc);
1408
1409         /* Keep the rq_export around so we can send the reply. */
1410         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1411
1412         RETURN(0);
1413 }
1414
1415 void target_destroy_export(struct obd_export *exp)
1416 {
1417         struct obd_import       *imp = NULL;
1418         /* exports created from last_rcvd data, and "fake"
1419            exports created by lctl don't have an import */
1420         spin_lock(&exp->exp_lock);
1421         if (exp->exp_imp_reverse != NULL) {
1422                 imp = exp->exp_imp_reverse;
1423                 exp->exp_imp_reverse = NULL;
1424         }
1425         spin_unlock(&exp->exp_lock);
1426         if (imp != NULL)
1427                 client_destroy_import(imp);
1428
1429         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1430         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1431         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1432         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1433 }
1434 EXPORT_SYMBOL(target_destroy_export);
1435
1436 /*
1437  * Recovery functions
1438  */
1439 static void target_request_copy_get(struct ptlrpc_request *req)
1440 {
1441         class_export_rpc_inc(req->rq_export);
1442         LASSERT(list_empty(&req->rq_list));
1443         INIT_LIST_HEAD(&req->rq_replay_list);
1444
1445         /* Increase refcount to keep request in queue. */
1446         atomic_inc(&req->rq_refcount);
1447         /* Let export know it has replays to be handled. */
1448         atomic_inc(&req->rq_export->exp_replay_count);
1449 }
1450
1451 static void target_request_copy_put(struct ptlrpc_request *req)
1452 {
1453         LASSERT(list_empty(&req->rq_replay_list));
1454         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1455
1456         atomic_dec(&req->rq_export->exp_replay_count);
1457         class_export_rpc_dec(req->rq_export);
1458         ptlrpc_server_drop_request(req);
1459 }
1460
1461 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1462 {
1463         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1464         struct obd_export     *exp = req->rq_export;
1465         struct ptlrpc_request *reqiter;
1466         struct ptlrpc_request *dup_req = NULL;
1467         int                    dup = 0;
1468
1469         LASSERT(exp);
1470
1471         spin_lock(&exp->exp_lock);
1472         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1473                                 rq_replay_list) {
1474                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1475                         dup_req = reqiter;
1476                         dup = 1;
1477                         break;
1478                 }
1479         }
1480
1481         if (dup) {
1482                 /* We expect it with RESENT and REPLAY flags. */
1483                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1484                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1485                         CERROR("invalid flags %x of resent replay\n",
1486                                lustre_msg_get_flags(req->rq_reqmsg));
1487
1488                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1489                         __u32 new_conn;
1490
1491                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1492                         if (new_conn >
1493                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1494                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1495                                                         new_conn);
1496                 }
1497         } else {
1498                 list_add_tail(&req->rq_replay_list,
1499                                   &exp->exp_req_replay_queue);
1500         }
1501
1502         spin_unlock(&exp->exp_lock);
1503         return dup;
1504 }
1505
1506 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1507 {
1508         LASSERT(!list_empty(&req->rq_replay_list));
1509         LASSERT(req->rq_export);
1510
1511         spin_lock(&req->rq_export->exp_lock);
1512         list_del_init(&req->rq_replay_list);
1513         spin_unlock(&req->rq_export->exp_lock);
1514 }
1515
1516 static void target_finish_recovery(struct lu_target *lut)
1517 {
1518         struct obd_device *obd = lut->lut_obd;
1519         ENTRY;
1520
1521         /* Only log a recovery message when recovery has occurred. */
1522         if (obd->obd_recovery_start) {
1523                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1524                                         obd->obd_recovery_start);
1525                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1526                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1527                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1528                         obd->obd_max_recoverable_clients,
1529                         atomic_read(&obd->obd_connected_clients),
1530                         obd->obd_stale_clients,
1531                         obd->obd_stale_clients == 1 ? "was" : "were");
1532         }
1533
1534         ldlm_reprocess_all_ns(obd->obd_namespace);
1535         spin_lock(&obd->obd_recovery_task_lock);
1536         if (!list_empty(&obd->obd_req_replay_queue) ||
1537             !list_empty(&obd->obd_lock_replay_queue) ||
1538             !list_empty(&obd->obd_final_req_queue)) {
1539                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1540                        obd->obd_name,
1541                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1542                        list_empty(&obd->obd_lock_replay_queue) ? \
1543                                "" : "lock ",
1544                        list_empty(&obd->obd_final_req_queue) ? \
1545                                "" : "final ");
1546                 spin_unlock(&obd->obd_recovery_task_lock);
1547                 LBUG();
1548         }
1549         spin_unlock(&obd->obd_recovery_task_lock);
1550
1551         if (lut->lut_tdtd != NULL &&
1552             (!list_empty(&lut->lut_tdtd->tdtd_replay_list) ||
1553             !list_empty(&lut->lut_tdtd->tdtd_replay_finish_list))) {
1554                 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1555                 dtrq_list_destroy(lut->lut_tdtd);
1556         }
1557
1558         obd->obd_recovery_end = cfs_time_current_sec();
1559
1560         /* When recovery finished, cleanup orphans on MDS and OST. */
1561         if (OBT(obd) && OBP(obd, postrecov)) {
1562                 int rc = OBP(obd, postrecov)(obd);
1563                 if (rc < 0)
1564                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1565                                       obd->obd_name, rc);
1566         }
1567         EXIT;
1568 }
1569
1570 static void abort_req_replay_queue(struct obd_device *obd)
1571 {
1572         struct ptlrpc_request *req, *n;
1573         struct list_head abort_list;
1574
1575         INIT_LIST_HEAD(&abort_list);
1576         spin_lock(&obd->obd_recovery_task_lock);
1577         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1578         spin_unlock(&obd->obd_recovery_task_lock);
1579         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1580                 DEBUG_REQ(D_WARNING, req, "aborted:");
1581                 req->rq_status = -ENOTCONN;
1582                 if (ptlrpc_error(req)) {
1583                         DEBUG_REQ(D_ERROR, req,
1584                                   "failed abort_req_reply; skipping");
1585                 }
1586                 target_exp_dequeue_req_replay(req);
1587                 target_request_copy_put(req);
1588         }
1589 }
1590
1591 static void abort_lock_replay_queue(struct obd_device *obd)
1592 {
1593         struct ptlrpc_request *req, *n;
1594         struct list_head abort_list;
1595
1596         INIT_LIST_HEAD(&abort_list);
1597         spin_lock(&obd->obd_recovery_task_lock);
1598         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1599         spin_unlock(&obd->obd_recovery_task_lock);
1600         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1601                 DEBUG_REQ(D_ERROR, req, "aborted:");
1602                 req->rq_status = -ENOTCONN;
1603                 if (ptlrpc_error(req)) {
1604                         DEBUG_REQ(D_ERROR, req,
1605                                   "failed abort_lock_reply; skipping");
1606                 }
1607                 target_request_copy_put(req);
1608         }
1609 }
1610
1611 /* Called from a cleanup function if the device is being cleaned up
1612    forcefully.  The exports should all have been disconnected already,
1613    the only thing left to do is
1614      - clear the recovery flags
1615      - cancel the timer
1616      - free queued requests and replies, but don't send replies
1617    Because the obd_stopping flag is set, no new requests should be received.
1618
1619 */
1620 void target_cleanup_recovery(struct obd_device *obd)
1621 {
1622         struct ptlrpc_request *req, *n;
1623         struct list_head clean_list;
1624
1625         INIT_LIST_HEAD(&clean_list);
1626         spin_lock(&obd->obd_dev_lock);
1627         if (!obd->obd_recovering) {
1628                 spin_unlock(&obd->obd_dev_lock);
1629                 EXIT;
1630                 return;
1631         }
1632         obd->obd_recovering = obd->obd_abort_recovery = 0;
1633         obd->obd_force_abort_recovery = 0;
1634         spin_unlock(&obd->obd_dev_lock);
1635
1636         spin_lock(&obd->obd_recovery_task_lock);
1637         target_cancel_recovery_timer(obd);
1638         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1639         spin_unlock(&obd->obd_recovery_task_lock);
1640
1641         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1642                 LASSERT(req->rq_reply_state == NULL);
1643                 target_exp_dequeue_req_replay(req);
1644                 target_request_copy_put(req);
1645         }
1646
1647         spin_lock(&obd->obd_recovery_task_lock);
1648         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1649         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1650         spin_unlock(&obd->obd_recovery_task_lock);
1651
1652         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1653                 LASSERT(req->rq_reply_state == NULL);
1654                 target_request_copy_put(req);
1655         }
1656
1657         EXIT;
1658 }
1659 EXPORT_SYMBOL(target_cleanup_recovery);
1660
1661 /* obd_recovery_task_lock should be held */
1662 void target_cancel_recovery_timer(struct obd_device *obd)
1663 {
1664         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1665         cfs_timer_disarm(&obd->obd_recovery_timer);
1666 }
1667
1668 static void target_start_recovery_timer(struct obd_device *obd)
1669 {
1670         if (obd->obd_recovery_start != 0)
1671                 return;
1672
1673         spin_lock(&obd->obd_dev_lock);
1674         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1675             obd->obd_force_abort_recovery) {
1676                 spin_unlock(&obd->obd_dev_lock);
1677                 return;
1678         }
1679
1680         LASSERT(obd->obd_recovery_timeout != 0);
1681
1682         if (obd->obd_recovery_start != 0) {
1683                 spin_unlock(&obd->obd_dev_lock);
1684                 return;
1685         }
1686
1687         cfs_timer_arm(&obd->obd_recovery_timer,
1688                       cfs_time_shift(obd->obd_recovery_timeout));
1689         obd->obd_recovery_start = cfs_time_current_sec();
1690         spin_unlock(&obd->obd_dev_lock);
1691
1692         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1693                       "or until %d client%s reconnect%s\n",
1694                       obd->obd_name,
1695                       obd->obd_recovery_timeout / 60,
1696                       obd->obd_recovery_timeout % 60,
1697                       obd->obd_max_recoverable_clients,
1698                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1699                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1700 }
1701
1702 /**
1703  * extend recovery window.
1704  *
1705  * if @extend is true, extend recovery window to have @drt remaining at least;
1706  * otherwise, make sure the recovery timeout value is not less than @drt.
1707  */
1708 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1709 {
1710         cfs_time_t now;
1711         cfs_time_t end;
1712         cfs_duration_t left;
1713         int to;
1714
1715         spin_lock(&obd->obd_dev_lock);
1716         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1717             obd->obd_force_abort_recovery) {
1718                 spin_unlock(&obd->obd_dev_lock);
1719                 return;
1720         }
1721         LASSERT(obd->obd_recovery_start != 0);
1722
1723         now  = cfs_time_current_sec();
1724         to   = obd->obd_recovery_timeout;
1725         end  = obd->obd_recovery_start + to;
1726         left = cfs_time_sub(end, now);
1727
1728         if (extend && (drt > left)) {
1729                 to += drt - left;
1730         } else if (!extend && (drt > to)) {
1731                 to = drt;
1732         }
1733
1734         if (to > obd->obd_recovery_time_hard)
1735                 to = obd->obd_recovery_time_hard;
1736         if (obd->obd_recovery_timeout < to) {
1737                 obd->obd_recovery_timeout = to;
1738                 end = obd->obd_recovery_start + to;
1739                 cfs_timer_arm(&obd->obd_recovery_timer,
1740                                 cfs_time_shift(end - now));
1741         }
1742         spin_unlock(&obd->obd_dev_lock);
1743
1744         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1745                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1746 }
1747
1748 /* Reset the timer with each new client connection */
1749 /*
1750  * This timer is actually reconnect_timer, which is for making sure
1751  * the total recovery window is at least as big as my reconnect
1752  * attempt timing. So the initial recovery time_out will be set to
1753  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1754  * from client is bigger than this, then the recovery time_out will
1755  * be extended to make sure the client could be reconnected, in the
1756  * process, the timeout from the new client should be ignored.
1757  */
1758
1759 static void
1760 check_and_start_recovery_timer(struct obd_device *obd,
1761                                struct ptlrpc_request *req,
1762                                int new_client)
1763 {
1764         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1765         struct obd_device_target *obt = &obd->u.obt;
1766
1767         if (!new_client && service_time)
1768                 /* Teach server about old server's estimates, as first guess
1769                  * at how long new requests will take. */
1770                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1771                             service_time);
1772
1773         target_start_recovery_timer(obd);
1774
1775         /* Convert the service time to RPC timeout,
1776          * and reuse service_time to limit stack usage. */
1777         service_time = at_est2timeout(service_time);
1778
1779         /* We expect other clients to timeout within service_time, then try
1780          * to reconnect, then try the failover server.  The max delay between
1781          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1782         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1783
1784         LASSERT(obt->obt_magic == OBT_MAGIC);
1785         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1786         if (service_time > obd->obd_recovery_timeout && !new_client)
1787                 extend_recovery_timer(obd, service_time, false);
1788 }
1789
1790 /** Health checking routines */
1791 static inline int exp_connect_healthy(struct obd_export *exp)
1792 {
1793         return (exp->exp_in_recovery);
1794 }
1795
1796 /** if export done req_replay or has replay in queue */
1797 static inline int exp_req_replay_healthy(struct obd_export *exp)
1798 {
1799         return (!exp->exp_req_replay_needed ||
1800                 atomic_read(&exp->exp_replay_count) > 0);
1801 }
1802 /** if export done lock_replay or has replay in queue */
1803 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1804 {
1805         return (!exp->exp_lock_replay_needed ||
1806                 atomic_read(&exp->exp_replay_count) > 0);
1807 }
1808
1809 static inline int exp_vbr_healthy(struct obd_export *exp)
1810 {
1811         return (!exp->exp_vbr_failed);
1812 }
1813
1814 static inline int exp_finished(struct obd_export *exp)
1815 {
1816         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1817 }
1818
1819 static int check_for_next_transno(struct lu_target *lut)
1820 {
1821         struct ptlrpc_request *req = NULL;
1822         struct obd_device *obd = lut->lut_obd;
1823         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1824         int wake_up = 0, connected, completed, queue_len;
1825         __u64 req_transno = 0;
1826         __u64 update_transno = 0;
1827         __u64 next_transno = 0;
1828         ENTRY;
1829
1830         spin_lock(&obd->obd_recovery_task_lock);
1831         if (!list_empty(&obd->obd_req_replay_queue)) {
1832                 req = list_entry(obd->obd_req_replay_queue.next,
1833                                      struct ptlrpc_request, rq_list);
1834                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1835         }
1836
1837         if (tdtd != NULL)
1838                 update_transno = distribute_txn_get_next_transno(tdtd);
1839
1840         connected = atomic_read(&obd->obd_connected_clients);
1841         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1842         queue_len = obd->obd_requests_queued_for_recovery;
1843         next_transno = obd->obd_next_recovery_transno;
1844
1845         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1846                "req_transno: "LPU64", next_transno: "LPU64"\n",
1847                obd->obd_max_recoverable_clients, connected, completed,
1848                queue_len, req_transno, next_transno);
1849
1850         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1851                 CDEBUG(D_HA, "waking for aborted recovery\n");
1852                 wake_up = 1;
1853         } else if (obd->obd_recovery_expired) {
1854                 CDEBUG(D_HA, "waking for expired recovery\n");
1855                 wake_up = 1;
1856         } else if (tdtd != NULL && req != NULL &&
1857                    is_req_replayed_by_update(req)) {
1858                 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1859                          "next_transno"LPU64"\n", req_transno, next_transno);
1860                 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1861                        req_transno);
1862                 wake_up = 1;
1863         } else if (req_transno == next_transno ||
1864                    (update_transno != 0 && update_transno <= next_transno)) {
1865                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1866                 wake_up = 1;
1867         } else if (queue_len > 0 &&
1868                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1869                 int d_lvl = D_HA;
1870                 /** handle gaps occured due to lost reply or VBR */
1871                 LASSERTF(req_transno >= next_transno,
1872                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1873                          req_transno, next_transno);
1874                 if (req_transno > obd->obd_last_committed &&
1875                     !obd->obd_version_recov)
1876                         d_lvl = D_ERROR;
1877                 CDEBUG(d_lvl,
1878                        "%s: waking for gap in transno, VBR is %s (skip: "
1879                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1880                        ", next_update "LPD64" last_committed: "LPD64")\n",
1881                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1882                        next_transno, queue_len, completed, connected,
1883                        req_transno, update_transno, obd->obd_last_committed);
1884                 obd->obd_next_recovery_transno = req_transno;
1885                 wake_up = 1;
1886         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1887                 CDEBUG(D_HA, "waking for completed recovery\n");
1888                 wake_up = 1;
1889         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1890                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1891                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1892                 obd->obd_next_recovery_transno = req_transno;
1893                 wake_up = 1;
1894         }
1895         spin_unlock(&obd->obd_recovery_task_lock);
1896         return wake_up;
1897 }
1898
1899 static int check_for_next_lock(struct lu_target *lut)
1900 {
1901         struct obd_device *obd = lut->lut_obd;
1902         int wake_up = 0;
1903
1904         spin_lock(&obd->obd_recovery_task_lock);
1905         if (!list_empty(&obd->obd_lock_replay_queue)) {
1906                 CDEBUG(D_HA, "waking for next lock\n");
1907                 wake_up = 1;
1908         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1909                 CDEBUG(D_HA, "waking for completed lock replay\n");
1910                 wake_up = 1;
1911         } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1912                 CDEBUG(D_HA, "waking for aborted recovery\n");
1913                 wake_up = 1;
1914         } else if (obd->obd_recovery_expired) {
1915                 CDEBUG(D_HA, "waking for expired recovery\n");
1916                 wake_up = 1;
1917         }
1918         spin_unlock(&obd->obd_recovery_task_lock);
1919
1920         return wake_up;
1921 }
1922
1923 /**
1924  * wait for recovery events,
1925  * check its status with help of check_routine
1926  * evict dead clients via health_check
1927  */
1928 static int target_recovery_overseer(struct lu_target *lut,
1929                                     int (*check_routine)(struct lu_target *),
1930                                     int (*health_check)(struct obd_export *))
1931 {
1932         struct obd_device       *obd = lut->lut_obd;
1933 repeat:
1934         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1935               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1936                 CWARN("recovery is aborted by hard timeout\n");
1937                 obd->obd_abort_recovery = 1;
1938         }
1939
1940         while (wait_event_timeout(obd->obd_next_transno_waitq,
1941                                   check_routine(lut),
1942                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1943                 /* wait indefinitely for event, but don't trigger watchdog */;
1944
1945         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1946                 CWARN("recovery is aborted, evict exports in recovery\n");
1947                 /** evict exports which didn't finish recovery yet */
1948                 class_disconnect_stale_exports(obd, exp_finished);
1949                 return 1;
1950         } else if (obd->obd_recovery_expired) {
1951                 obd->obd_recovery_expired = 0;
1952                 /** If some clients died being recovered, evict them */
1953                 LCONSOLE_WARN("%s: recovery is timed out, "
1954                               "evict stale exports\n", obd->obd_name);
1955                 /** evict cexports with no replay in queue, they are stalled */
1956                 class_disconnect_stale_exports(obd, health_check);
1957                 /** continue with VBR */
1958                 spin_lock(&obd->obd_dev_lock);
1959                 obd->obd_version_recov = 1;
1960                 spin_unlock(&obd->obd_dev_lock);
1961                 /**
1962                  * reset timer, recovery will proceed with versions now,
1963                  * timeout is set just to handle reconnection delays
1964                  */
1965                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1966                 /** Wait for recovery events again, after evicting bad clients */
1967                 goto repeat;
1968         }
1969         return 0;
1970 }
1971
1972 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1973 {
1974         struct obd_device       *obd = lut->lut_obd;
1975         struct ptlrpc_request *req = NULL;
1976
1977         CDEBUG(D_HA, "Waiting for lock\n");
1978         if (target_recovery_overseer(lut, check_for_next_lock,
1979                                      exp_lock_replay_healthy))
1980                 abort_lock_replay_queue(obd);
1981
1982         spin_lock(&obd->obd_recovery_task_lock);
1983         if (!list_empty(&obd->obd_lock_replay_queue)) {
1984                 req = list_entry(obd->obd_lock_replay_queue.next,
1985                                      struct ptlrpc_request, rq_list);
1986                 list_del_init(&req->rq_list);
1987                 spin_unlock(&obd->obd_recovery_task_lock);
1988         } else {
1989                 spin_unlock(&obd->obd_recovery_task_lock);
1990                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1991                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1992                 /** evict exports failed VBR */
1993                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1994         }
1995         return req;
1996 }
1997
1998 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1999 {
2000         struct ptlrpc_request *req = NULL;
2001
2002         spin_lock(&obd->obd_recovery_task_lock);
2003         if (!list_empty(&obd->obd_final_req_queue)) {
2004                 req = list_entry(obd->obd_final_req_queue.next,
2005                                      struct ptlrpc_request, rq_list);
2006                 list_del_init(&req->rq_list);
2007                 spin_unlock(&obd->obd_recovery_task_lock);
2008                 if (req->rq_export->exp_in_recovery) {
2009                         spin_lock(&req->rq_export->exp_lock);
2010                         req->rq_export->exp_in_recovery = 0;
2011                         spin_unlock(&req->rq_export->exp_lock);
2012                 }
2013         } else {
2014                 spin_unlock(&obd->obd_recovery_task_lock);
2015         }
2016         return req;
2017 }
2018
2019 static void handle_recovery_req(struct ptlrpc_thread *thread,
2020                                 struct ptlrpc_request *req,
2021                                 svc_handler_t handler)
2022 {
2023         ENTRY;
2024
2025         /**
2026          * export can be evicted during recovery, no need to handle replays for
2027          * it after that, discard such request silently
2028          */
2029         if (req->rq_export->exp_disconnected)
2030                 RETURN_EXIT;
2031
2032         req->rq_session.lc_thread = thread;
2033         req->rq_svc_thread = thread;
2034         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2035
2036         /* thread context */
2037         lu_context_enter(&thread->t_env->le_ctx);
2038         (void)handler(req);
2039         lu_context_exit(&thread->t_env->le_ctx);
2040
2041         /* don't reset timer for final stage */
2042         if (!exp_finished(req->rq_export)) {
2043                 int to = obd_timeout;
2044
2045                 /**
2046                  * Add request timeout to the recovery time so next request from
2047                  * this client may come in recovery time
2048                  */
2049                 if (!AT_OFF) {
2050                         struct ptlrpc_service_part *svcpt;
2051
2052                         svcpt = req->rq_rqbd->rqbd_svcpt;
2053                         /* If the server sent early reply for this request,
2054                          * the client will recalculate the timeout according to
2055                          * current server estimate service time, so we will
2056                          * use the maxium timeout here for waiting the client
2057                          * sending the next req */
2058                         to = max((int)at_est2timeout(
2059                                  at_get(&svcpt->scp_at_estimate)),
2060                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2061                         /* Add 2 net_latency, one for balance rq_deadline
2062                          * (see ptl_send_rpc), one for resend the req to server,
2063                          * Note: client will pack net_latency in replay req
2064                          * (see ptlrpc_replay_req) */
2065                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2066                 }
2067                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2068         }
2069         EXIT;
2070 }
2071
2072 /** Checking routines for recovery */
2073 static int check_for_recovery_ready(struct lu_target *lut)
2074 {
2075         struct obd_device *obd = lut->lut_obd;
2076         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2077
2078         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2079                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2080                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2081                obd->obd_recovery_expired);
2082
2083         if (obd->obd_force_abort_recovery)
2084                 return 1;
2085
2086         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2087                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2088                 if (clnts + obd->obd_stale_clients <
2089                     obd->obd_max_recoverable_clients)
2090                         return 0;
2091         }
2092
2093         if (lut->lut_tdtd != NULL) {
2094                 if (!lut->lut_tdtd->tdtd_replay_ready) {
2095                         /* Let's extend recovery timer, in case the recovery
2096                          * timer expired, and some clients got evicted */
2097                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2098                                               true);
2099                         return 0;
2100                 } else {
2101                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2102                 }
2103         }
2104
2105         return 1;
2106 }
2107
2108 enum {
2109         REQUEST_RECOVERY = 1,
2110         UPDATE_RECOVERY = 2,
2111 };
2112
2113 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2114 {
2115         __u64 transno = 0;
2116
2117         if (!list_empty(&obd->obd_req_replay_queue)) {
2118                 struct ptlrpc_request *req;
2119
2120                 req = list_entry(obd->obd_req_replay_queue.next,
2121                                  struct ptlrpc_request, rq_list);
2122                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2123         }
2124
2125         return transno;
2126 }
2127 __u64 get_next_transno(struct lu_target *lut, int *type)
2128 {
2129         struct obd_device *obd = lut->lut_obd;
2130         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2131         __u64 transno = 0;
2132         __u64 update_transno;
2133         ENTRY;
2134
2135         transno = get_next_replay_req_transno(obd);
2136         if (type != NULL)
2137                 *type = REQUEST_RECOVERY;
2138
2139         if (tdtd == NULL)
2140                 RETURN(transno);
2141
2142         update_transno = distribute_txn_get_next_transno(tdtd);
2143         if (transno == 0 || (transno >= update_transno &&
2144                              update_transno != 0)) {
2145                 transno = update_transno;
2146                 if (type != NULL)
2147                         *type = UPDATE_RECOVERY;
2148         }
2149
2150         RETURN(transno);
2151 }
2152
2153 /**
2154  * drop duplicate replay request
2155  *
2156  * Because the operation has been replayed by update recovery, the request
2157  * with the same transno will be dropped and also notify the client to send
2158  * next replay request.
2159  *
2160  * \param[in] env       execution environment
2161  * \param[in] obd       failover obd device
2162  * \param[in] req       request to be dropped
2163  */
2164 static void drop_duplicate_replay_req(struct lu_env *env,
2165                                       struct obd_device *obd,
2166                                       struct ptlrpc_request *req)
2167 {
2168         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2169                   " update records are found.\n",
2170                   lustre_msg_get_transno(req->rq_reqmsg),
2171                   libcfs_nid2str(req->rq_peer.nid));
2172
2173         /* Right now, only for MDS reint operation update replay and
2174          * normal request replay can have the same transno */
2175         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2176                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2177                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2178                 if (likely(req->rq_export))
2179                         target_committed_to_req(req);
2180                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2181                 target_send_reply(req, req->rq_status, 0);
2182         } else {
2183                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2184                 libcfs_nid2str(req->rq_peer.nid));
2185         }
2186         target_exp_dequeue_req_replay(req);
2187         target_request_copy_put(req);
2188         obd->obd_replayed_requests++;
2189 }
2190
2191 static void replay_request_or_update(struct lu_env *env,
2192                                      struct lu_target *lut,
2193                                      struct target_recovery_data *trd,
2194                                      struct ptlrpc_thread *thread)
2195 {
2196         struct obd_device *obd = lut->lut_obd;
2197         struct ptlrpc_request *req = NULL;
2198         int                     type;
2199         __u64                   transno;
2200         ENTRY;
2201
2202         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2203                obd->obd_next_recovery_transno);
2204
2205         /* Replay all of request and update by transno */
2206         do {
2207                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2208
2209                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2210
2211                 /** It is needed to extend recovery window above
2212                  *  recovery_time_soft. Extending is possible only in the
2213                  *  end of recovery window (see more details in
2214                  *  handle_recovery_req()).
2215                  */
2216                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2217
2218                 if (target_recovery_overseer(lut, check_for_next_transno,
2219                                              exp_req_replay_healthy)) {
2220                         abort_req_replay_queue(obd);
2221                         abort_lock_replay_queue(obd);
2222                 }
2223
2224                 spin_lock(&obd->obd_recovery_task_lock);
2225                 transno = get_next_transno(lut, &type);
2226                 if (type == REQUEST_RECOVERY && transno != 0) {
2227                         /* Drop replay request from client side, if the
2228                          * replay has been executed by update with the
2229                          * same transno */
2230                         req = list_entry(obd->obd_req_replay_queue.next,
2231                                         struct ptlrpc_request, rq_list);
2232
2233                         list_del_init(&req->rq_list);
2234                         obd->obd_requests_queued_for_recovery--;
2235                         spin_unlock(&obd->obd_recovery_task_lock);
2236
2237                         /* Let's check if the request has been redone by
2238                          * update replay */
2239                         if (is_req_replayed_by_update(req)) {
2240                                 struct distribute_txn_replay_req *dtrq;
2241
2242                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2243                                                                    req->rq_xid);
2244                                 LASSERT(dtrq != NULL);
2245                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2246                                 list_del_init(&dtrq->dtrq_list);
2247                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2248                                 dtrq_destroy(dtrq);
2249
2250                                 drop_duplicate_replay_req(env, obd, req);
2251
2252                                 continue;
2253                         }
2254
2255                         LASSERT(trd->trd_processing_task == current_pid());
2256                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2257                                   lustre_msg_get_transno(req->rq_reqmsg),
2258                                   libcfs_nid2str(req->rq_peer.nid));
2259
2260                         handle_recovery_req(thread, req,
2261                                             trd->trd_recovery_handler);
2262                         /**
2263                          * bz18031: increase next_recovery_transno before
2264                          * target_request_copy_put() will drop exp_rpc reference
2265                          */
2266                         spin_lock(&obd->obd_recovery_task_lock);
2267                         obd->obd_next_recovery_transno++;
2268                         spin_unlock(&obd->obd_recovery_task_lock);
2269                         target_exp_dequeue_req_replay(req);
2270                         target_request_copy_put(req);
2271                         obd->obd_replayed_requests++;
2272                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2273                         struct distribute_txn_replay_req *dtrq;
2274
2275                         spin_unlock(&obd->obd_recovery_task_lock);
2276
2277                         LASSERT(tdtd != NULL);
2278                         dtrq = distribute_txn_get_next_req(tdtd);
2279                         lu_context_enter(&thread->t_env->le_ctx);
2280                         tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2281                         lu_context_exit(&thread->t_env->le_ctx);
2282                         extend_recovery_timer(obd, obd_timeout, true);
2283
2284                         /* Add it to the replay finish list */
2285                         spin_lock(&tdtd->tdtd_replay_list_lock);
2286                         if (dtrq->dtrq_xid != 0) {
2287                                 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2288                                        " to finish list\n", dtrq->dtrq_xid,
2289                                        dtrq->dtrq_master_transno);
2290                                 list_add(&dtrq->dtrq_list,
2291                                          &tdtd->tdtd_replay_finish_list);
2292                         } else {
2293                                 dtrq_destroy(dtrq);
2294                         }
2295                         spin_unlock(&tdtd->tdtd_replay_list_lock);
2296
2297                         spin_lock(&obd->obd_recovery_task_lock);
2298                         if (transno == obd->obd_next_recovery_transno)
2299                                 obd->obd_next_recovery_transno++;
2300                         else if (transno > obd->obd_next_recovery_transno)
2301                                 obd->obd_next_recovery_transno = transno + 1;
2302                         spin_unlock(&obd->obd_recovery_task_lock);
2303
2304                 } else {
2305                         spin_unlock(&obd->obd_recovery_task_lock);
2306                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2307                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2308                         /** evict exports failed VBR */
2309                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2310                         break;
2311                 }
2312         } while (1);
2313 }
2314
2315 static int target_recovery_thread(void *arg)
2316 {
2317         struct lu_target *lut = arg;
2318         struct obd_device *obd = lut->lut_obd;
2319         struct ptlrpc_request *req;
2320         struct target_recovery_data *trd = &obd->obd_recovery_data;
2321         unsigned long delta;
2322         struct lu_env *env;
2323         struct ptlrpc_thread *thread = NULL;
2324         int rc = 0;
2325         ENTRY;
2326
2327         unshare_fs_struct();
2328         OBD_ALLOC_PTR(thread);
2329         if (thread == NULL)
2330                 RETURN(-ENOMEM);
2331
2332         OBD_ALLOC_PTR(env);
2333         if (env == NULL) {
2334                 OBD_FREE_PTR(thread);
2335                 RETURN(-ENOMEM);
2336         }
2337
2338         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2339         if (rc) {
2340                 OBD_FREE_PTR(thread);
2341                 OBD_FREE_PTR(env);
2342                 RETURN(rc);
2343         }
2344
2345         thread->t_env = env;
2346         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2347         env->le_ctx.lc_thread = thread;
2348         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2349         thread->t_watchdog = NULL;
2350
2351         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2352                current_pid());
2353         trd->trd_processing_task = current_pid();
2354
2355         spin_lock(&obd->obd_dev_lock);
2356         obd->obd_recovering = 1;
2357         spin_unlock(&obd->obd_dev_lock);
2358         complete(&trd->trd_starting);
2359
2360         /* first of all, we have to know the first transno to replay */
2361         if (target_recovery_overseer(lut, check_for_recovery_ready,
2362                                      exp_connect_healthy)) {
2363                 abort_req_replay_queue(obd);
2364                 abort_lock_replay_queue(obd);
2365                 if (lut->lut_tdtd != NULL)
2366                         dtrq_list_destroy(lut->lut_tdtd);
2367         }
2368
2369         /* next stage: replay requests or update */
2370         delta = jiffies;
2371         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2372                atomic_read(&obd->obd_req_replay_clients),
2373                obd->obd_next_recovery_transno);
2374         replay_request_or_update(env, lut, trd, thread);
2375
2376         /**
2377          * The second stage: replay locks
2378          */
2379         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2380                atomic_read(&obd->obd_lock_replay_clients));
2381         while ((req = target_next_replay_lock(lut))) {
2382                 LASSERT(trd->trd_processing_task == current_pid());
2383                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2384                           libcfs_nid2str(req->rq_peer.nid));
2385                 handle_recovery_req(thread, req,
2386                                     trd->trd_recovery_handler);
2387                 target_request_copy_put(req);
2388                 obd->obd_replayed_locks++;
2389         }
2390
2391         /**
2392          * The third stage: reply on final pings, at this moment all clients
2393          * must have request in final queue
2394          */
2395         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2396         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2397         /** Update server last boot epoch */
2398         tgt_boot_epoch_update(lut);
2399         /* We drop recoverying flag to forward all new requests
2400          * to regular mds_handle() since now */
2401         spin_lock(&obd->obd_dev_lock);
2402         obd->obd_recovering = obd->obd_abort_recovery = 0;
2403         spin_unlock(&obd->obd_dev_lock);
2404         spin_lock(&obd->obd_recovery_task_lock);
2405         target_cancel_recovery_timer(obd);
2406         spin_unlock(&obd->obd_recovery_task_lock);
2407         while ((req = target_next_final_ping(obd))) {
2408                 LASSERT(trd->trd_processing_task == current_pid());
2409                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2410                           libcfs_nid2str(req->rq_peer.nid));
2411                 handle_recovery_req(thread, req,
2412                                     trd->trd_recovery_handler);
2413                 /* Because the waiting client can not send ping to server,
2414                  * so we need refresh the last_request_time, to avoid the
2415                  * export is being evicted */
2416                 ptlrpc_update_export_timer(req->rq_export, 0);
2417                 target_request_copy_put(req);
2418         }
2419
2420         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2421         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2422               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2423         if (delta > OBD_RECOVERY_TIME_SOFT) {
2424                 CWARN("too long recovery - read logs\n");
2425                 libcfs_debug_dumplog();
2426         }
2427
2428         target_finish_recovery(lut);
2429
2430         lu_context_fini(&env->le_ctx);
2431         trd->trd_processing_task = 0;
2432         complete(&trd->trd_finishing);
2433
2434         tgt_io_thread_done(thread);
2435         OBD_FREE_PTR(thread);
2436         OBD_FREE_PTR(env);
2437         RETURN(rc);
2438 }
2439
2440 static int target_start_recovery_thread(struct lu_target *lut,
2441                                         svc_handler_t handler)
2442 {
2443         struct obd_device *obd = lut->lut_obd;
2444         int rc = 0;
2445         struct target_recovery_data *trd = &obd->obd_recovery_data;
2446
2447         memset(trd, 0, sizeof(*trd));
2448         init_completion(&trd->trd_starting);
2449         init_completion(&trd->trd_finishing);
2450         trd->trd_recovery_handler = handler;
2451
2452         if (!IS_ERR(kthread_run(target_recovery_thread,
2453                                 lut, "tgt_recov"))) {
2454                 wait_for_completion(&trd->trd_starting);
2455                 LASSERT(obd->obd_recovering != 0);
2456         } else {
2457                 rc = -ECHILD;
2458         }
2459
2460         return rc;
2461 }
2462
2463 void target_stop_recovery_thread(struct obd_device *obd)
2464 {
2465         if (obd->obd_recovery_data.trd_processing_task > 0) {
2466                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2467                 /** recovery can be done but postrecovery is not yet */
2468                 spin_lock(&obd->obd_dev_lock);
2469                 if (obd->obd_recovering) {
2470                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2471                         obd->obd_abort_recovery = 1;
2472                         wake_up(&obd->obd_next_transno_waitq);
2473                 }
2474                 spin_unlock(&obd->obd_dev_lock);
2475                 wait_for_completion(&trd->trd_finishing);
2476         }
2477 }
2478 EXPORT_SYMBOL(target_stop_recovery_thread);
2479
2480 void target_recovery_fini(struct obd_device *obd)
2481 {
2482         class_disconnect_exports(obd);
2483         target_stop_recovery_thread(obd);
2484         target_cleanup_recovery(obd);
2485 }
2486 EXPORT_SYMBOL(target_recovery_fini);
2487
2488 static void target_recovery_expired(unsigned long castmeharder)
2489 {
2490         struct obd_device *obd = (struct obd_device *)castmeharder;
2491         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2492                " after %lds (%d clients connected)\n",
2493                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2494                cfs_time_current_sec()- obd->obd_recovery_start,
2495                atomic_read(&obd->obd_connected_clients));
2496
2497         obd->obd_recovery_expired = 1;
2498         wake_up(&obd->obd_next_transno_waitq);
2499 }
2500
2501 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2502 {
2503         struct obd_device *obd = lut->lut_obd;
2504
2505         if (obd->obd_max_recoverable_clients == 0) {
2506                 /** Update server last boot epoch */
2507                 tgt_boot_epoch_update(lut);
2508                 return;
2509         }
2510
2511         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2512                "last_transno "LPU64"\n", obd->obd_name,
2513                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2514         LASSERT(obd->obd_stopping == 0);
2515         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2516         obd->obd_recovery_start = 0;
2517         obd->obd_recovery_end = 0;
2518
2519         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2520         target_start_recovery_thread(lut, handler);
2521 }
2522 EXPORT_SYMBOL(target_recovery_init);
2523
2524 static int target_process_req_flags(struct obd_device *obd,
2525                                     struct ptlrpc_request *req)
2526 {
2527         struct obd_export *exp = req->rq_export;
2528         LASSERT(exp != NULL);
2529         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2530                 /* client declares he's ready to replay locks */
2531                 spin_lock(&exp->exp_lock);
2532                 if (exp->exp_req_replay_needed) {
2533                         exp->exp_req_replay_needed = 0;
2534                         spin_unlock(&exp->exp_lock);
2535
2536                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2537                         atomic_dec(&obd->obd_req_replay_clients);
2538                 } else {
2539                         spin_unlock(&exp->exp_lock);
2540                 }
2541         }
2542         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2543                 /* client declares he's ready to complete recovery
2544                  * so, we put the request on th final queue */
2545                 spin_lock(&exp->exp_lock);
2546                 if (exp->exp_lock_replay_needed) {
2547                         exp->exp_lock_replay_needed = 0;
2548                         spin_unlock(&exp->exp_lock);
2549
2550                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2551                         atomic_dec(&obd->obd_lock_replay_clients);
2552                 } else {
2553                         spin_unlock(&exp->exp_lock);
2554                 }
2555         }
2556         return 0;
2557 }
2558
2559 int target_queue_recovery_request(struct ptlrpc_request *req,
2560                                   struct obd_device *obd)
2561 {
2562         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2563         struct ptlrpc_request *reqiter;
2564         int inserted = 0;
2565         ENTRY;
2566
2567         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2568                 /* Processing the queue right now, don't re-add. */
2569                 RETURN(1);
2570         }
2571
2572         target_process_req_flags(obd, req);
2573
2574         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2575                 /* client declares he's ready to complete recovery
2576                  * so, we put the request on th final queue */
2577                 target_request_copy_get(req);
2578                 DEBUG_REQ(D_HA, req, "queue final req");
2579                 wake_up(&obd->obd_next_transno_waitq);
2580                 spin_lock(&obd->obd_recovery_task_lock);
2581                 if (obd->obd_recovering) {
2582                         list_add_tail(&req->rq_list,
2583                                           &obd->obd_final_req_queue);
2584                 } else {
2585                         spin_unlock(&obd->obd_recovery_task_lock);
2586                         target_request_copy_put(req);
2587                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2588                 }
2589                 spin_unlock(&obd->obd_recovery_task_lock);
2590                 RETURN(0);
2591         }
2592         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2593                 /* client declares he's ready to replay locks */
2594                 target_request_copy_get(req);
2595                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2596                 wake_up(&obd->obd_next_transno_waitq);
2597                 spin_lock(&obd->obd_recovery_task_lock);
2598                 LASSERT(obd->obd_recovering);
2599                 /* usually due to recovery abort */
2600                 if (!req->rq_export->exp_in_recovery) {
2601                         spin_unlock(&obd->obd_recovery_task_lock);
2602                         target_request_copy_put(req);
2603                         RETURN(-ENOTCONN);
2604                 }
2605                 LASSERT(req->rq_export->exp_lock_replay_needed);
2606                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2607                 spin_unlock(&obd->obd_recovery_task_lock);
2608                 RETURN(0);
2609         }
2610
2611         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2612          * (i.e. buflens etc are in my own byte order), but type-dependent
2613          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2614
2615         if (!transno) {
2616                 INIT_LIST_HEAD(&req->rq_list);
2617                 DEBUG_REQ(D_HA, req, "not queueing");
2618                 RETURN(1);
2619         }
2620
2621         /* If we're processing the queue, we want don't want to queue this
2622          * message.
2623          *
2624          * Also, if this request has a transno less than the one we're waiting
2625          * for, we should process it now.  It could (and currently always will)
2626          * be an open request for a descriptor that was opened some time ago.
2627          *
2628          * Also, a resent, replayed request that has already been
2629          * handled will pass through here and be processed immediately.
2630          */
2631         CDEBUG(D_HA, "Next recovery transno: "LPU64
2632                ", current: "LPU64", replaying\n",
2633                obd->obd_next_recovery_transno, transno);
2634
2635         /* If the request has been replayed by update replay, then sends this
2636          * request to the recovery thread (replay_request_or_update()), where
2637          * it will be handled */
2638         spin_lock(&obd->obd_recovery_task_lock);
2639         if (transno < obd->obd_next_recovery_transno &&
2640             !is_req_replayed_by_update(req)) {
2641                 /* Processing the queue right now, don't re-add. */
2642                 LASSERT(list_empty(&req->rq_list));
2643                 spin_unlock(&obd->obd_recovery_task_lock);
2644                 RETURN(1);
2645         }
2646         spin_unlock(&obd->obd_recovery_task_lock);
2647
2648         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2649                 RETURN(0);
2650
2651         target_request_copy_get(req);
2652         if (!req->rq_export->exp_in_recovery) {
2653                 target_request_copy_put(req);
2654                 RETURN(-ENOTCONN);
2655         }
2656         LASSERT(req->rq_export->exp_req_replay_needed);
2657
2658         if (target_exp_enqueue_req_replay(req)) {
2659                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2660                 target_request_copy_put(req);
2661                 RETURN(0);
2662         }
2663
2664         /* XXX O(n^2) */
2665         spin_lock(&obd->obd_recovery_task_lock);
2666         LASSERT(obd->obd_recovering);
2667         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2668                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2669                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2670                         inserted = 1;
2671                         goto added;
2672                 }
2673
2674                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2675                              transno)) {
2676                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2677                                   "has been claimed by another client");
2678                         spin_unlock(&obd->obd_recovery_task_lock);
2679                         target_exp_dequeue_req_replay(req);
2680                         target_request_copy_put(req);
2681                         RETURN(0);
2682                 }
2683         }
2684 added:
2685         if (!inserted)
2686                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2687
2688         obd->obd_requests_queued_for_recovery++;
2689         spin_unlock(&obd->obd_recovery_task_lock);
2690         wake_up(&obd->obd_next_transno_waitq);
2691         RETURN(0);
2692 }
2693
2694 int target_handle_ping(struct ptlrpc_request *req)
2695 {
2696         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2697         return req_capsule_server_pack(&req->rq_pill);
2698 }
2699
2700 void target_committed_to_req(struct ptlrpc_request *req)
2701 {
2702         struct obd_export *exp = req->rq_export;
2703
2704         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2705                 lustre_msg_set_last_committed(req->rq_repmsg,
2706                                               exp->exp_last_committed);
2707         else
2708                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2709                           "%d)", exp->exp_obd->obd_no_transno,
2710                           req->rq_repmsg == NULL);
2711
2712         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2713                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2714 }
2715
2716 #endif /* HAVE_SERVER_SUPPORT */
2717
2718 /**
2719  * Packs current SLV and Limit into \a req.
2720  */
2721 int target_pack_pool_reply(struct ptlrpc_request *req)
2722 {
2723         struct obd_device *obd;
2724         ENTRY;
2725
2726         /* Check that we still have all structures alive as this may
2727          * be some late RPC at shutdown time. */
2728         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2729                      !exp_connect_lru_resize(req->rq_export))) {
2730                 lustre_msg_set_slv(req->rq_repmsg, 0);
2731                 lustre_msg_set_limit(req->rq_repmsg, 0);
2732                 RETURN(0);
2733         }
2734
2735         /* OBD is alive here as export is alive, which we checked above. */
2736         obd = req->rq_export->exp_obd;
2737
2738         read_lock(&obd->obd_pool_lock);
2739         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2740         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2741         read_unlock(&obd->obd_pool_lock);
2742
2743         RETURN(0);
2744 }
2745
2746 static int target_send_reply_msg(struct ptlrpc_request *req,
2747                                  int rc, int fail_id)
2748 {
2749         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2750                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2751                 return -ECOMM;
2752         }
2753         if (unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2754                      OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2755                 return -ECOMM;
2756
2757         if (unlikely(rc)) {
2758                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2759                 req->rq_status = rc;
2760                 return ptlrpc_send_error(req, 1);
2761         } else {
2762                 DEBUG_REQ(D_NET, req, "sending reply");
2763         }
2764
2765         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2766 }
2767
2768 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2769 {
2770         struct ptlrpc_service_part *svcpt;
2771         int                        netrc;
2772         struct ptlrpc_reply_state *rs;
2773         struct obd_export         *exp;
2774         ENTRY;
2775
2776         if (req->rq_no_reply) {
2777                 EXIT;
2778                 return;
2779         }
2780
2781         svcpt = req->rq_rqbd->rqbd_svcpt;
2782         rs = req->rq_reply_state;
2783         if (rs == NULL || !rs->rs_difficult) {
2784                 /* no notifiers */
2785                 target_send_reply_msg (req, rc, fail_id);
2786                 EXIT;
2787                 return;
2788         }
2789
2790         /* must be an export if locks saved */
2791         LASSERT(req->rq_export != NULL);
2792         /* req/reply consistent */
2793         LASSERT(rs->rs_svcpt == svcpt);
2794
2795         /* "fresh" reply */
2796         LASSERT(!rs->rs_scheduled);
2797         LASSERT(!rs->rs_scheduled_ever);
2798         LASSERT(!rs->rs_handled);
2799         LASSERT(!rs->rs_on_net);
2800         LASSERT(rs->rs_export == NULL);
2801         LASSERT(list_empty(&rs->rs_obd_list));
2802         LASSERT(list_empty(&rs->rs_exp_list));
2803
2804         exp = class_export_get(req->rq_export);
2805
2806         /* disable reply scheduling while I'm setting up */
2807         rs->rs_scheduled = 1;
2808         rs->rs_on_net    = 1;
2809         rs->rs_xid       = req->rq_xid;
2810         rs->rs_transno   = req->rq_transno;
2811         rs->rs_export    = exp;
2812         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2813
2814         spin_lock(&exp->exp_uncommitted_replies_lock);
2815         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2816                rs->rs_transno, exp->exp_last_committed);
2817         if (rs->rs_transno > exp->exp_last_committed) {
2818                 /* not committed already */
2819                 list_add_tail(&rs->rs_obd_list,
2820                                   &exp->exp_uncommitted_replies);
2821         }
2822         spin_unlock(&exp->exp_uncommitted_replies_lock);
2823
2824         spin_lock(&exp->exp_lock);
2825         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2826         spin_unlock(&exp->exp_lock);
2827
2828         netrc = target_send_reply_msg(req, rc, fail_id);
2829
2830         spin_lock(&svcpt->scp_rep_lock);
2831
2832         atomic_inc(&svcpt->scp_nreps_difficult);
2833
2834         if (netrc != 0) {
2835                 /* error sending: reply is off the net.  Also we need +1
2836                  * reply ref until ptlrpc_handle_rs() is done
2837                  * with the reply state (if the send was successful, there
2838                  * would have been +1 ref for the net, which
2839                  * reply_out_callback leaves alone) */
2840                 rs->rs_on_net = 0;
2841                 ptlrpc_rs_addref(rs);
2842         }
2843
2844         spin_lock(&rs->rs_lock);
2845         if (rs->rs_transno <= exp->exp_last_committed ||
2846             (!rs->rs_on_net && !rs->rs_no_ack) ||
2847             list_empty(&rs->rs_exp_list) ||     /* completed already */
2848             list_empty(&rs->rs_obd_list)) {
2849                 CDEBUG(D_HA, "Schedule reply immediately\n");
2850                 ptlrpc_dispatch_difficult_reply(rs);
2851         } else {
2852                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2853                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2854         }
2855         spin_unlock(&rs->rs_lock);
2856         spin_unlock(&svcpt->scp_rep_lock);
2857         EXIT;
2858 }
2859
2860 enum ldlm_mode lck_compat_array[] = {
2861         [LCK_EX]    = LCK_COMPAT_EX,
2862         [LCK_PW]    = LCK_COMPAT_PW,
2863         [LCK_PR]    = LCK_COMPAT_PR,
2864         [LCK_CW]    = LCK_COMPAT_CW,
2865         [LCK_CR]    = LCK_COMPAT_CR,
2866         [LCK_NL]    = LCK_COMPAT_NL,
2867         [LCK_GROUP] = LCK_COMPAT_GROUP,
2868         [LCK_COS]   = LCK_COMPAT_COS,
2869 };
2870
2871 /**
2872  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2873  * not escape to the user level.
2874  */
2875 int ldlm_error2errno(enum ldlm_error error)
2876 {
2877         int result;
2878
2879         switch (error) {
2880         case ELDLM_OK:
2881         case ELDLM_LOCK_MATCHED:
2882                 result = 0;
2883                 break;
2884         case ELDLM_LOCK_CHANGED:
2885                 result = -ESTALE;
2886                 break;
2887         case ELDLM_LOCK_ABORTED:
2888                 result = -ENAVAIL;
2889                 break;
2890         case ELDLM_LOCK_REPLACED:
2891                 result = -ESRCH;
2892                 break;
2893         case ELDLM_NO_LOCK_DATA:
2894                 result = -ENOENT;
2895                 break;
2896         case ELDLM_NAMESPACE_EXISTS:
2897                 result = -EEXIST;
2898                 break;
2899         case ELDLM_BAD_NAMESPACE:
2900                 result = -EBADF;
2901                 break;
2902         default:
2903                 if (((int)error) < 0) { /* cast to signed type */
2904                         result = error; /* as ldlm_error can be unsigned */
2905                 } else {
2906                         CERROR("Invalid DLM result code: %d\n", error);
2907                         result = -EPROTO;
2908                 }
2909         }
2910         return result;
2911 }
2912 EXPORT_SYMBOL(ldlm_error2errno);
2913
2914 /**
2915  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
2916  */
2917 enum ldlm_error ldlm_errno2error(int err_no)
2918 {
2919         int error;
2920
2921         switch (err_no) {
2922         case 0:
2923                 error = ELDLM_OK;
2924                 break;
2925         case -ESTALE:
2926                 error = ELDLM_LOCK_CHANGED;
2927                 break;
2928         case -ENAVAIL:
2929                 error = ELDLM_LOCK_ABORTED;
2930                 break;
2931         case -ESRCH:
2932                 error = ELDLM_LOCK_REPLACED;
2933                 break;
2934         case -ENOENT:
2935                 error = ELDLM_NO_LOCK_DATA;
2936                 break;
2937         case -EEXIST:
2938                 error = ELDLM_NAMESPACE_EXISTS;
2939                 break;
2940         case -EBADF:
2941                 error = ELDLM_BAD_NAMESPACE;
2942                 break;
2943         default:
2944                 error = err_no;
2945         }
2946         return error;
2947 }
2948
2949 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2950 void ldlm_dump_export_locks(struct obd_export *exp)
2951 {
2952         spin_lock(&exp->exp_locks_list_guard);
2953         if (!list_empty(&exp->exp_locks_list)) {
2954                 struct ldlm_lock *lock;
2955
2956                 CERROR("dumping locks for export %p,"
2957                        "ignore if the unmount doesn't hang\n", exp);
2958                 list_for_each_entry(lock, &exp->exp_locks_list,
2959                                         l_exp_refs_link)
2960                         LDLM_ERROR(lock, "lock:");
2961         }
2962         spin_unlock(&exp->exp_locks_list_guard);
2963 }
2964 #endif
2965
2966 #ifdef HAVE_SERVER_SUPPORT
2967 static int target_bulk_timeout(void *data)
2968 {
2969         ENTRY;
2970         /* We don't fail the connection here, because having the export
2971          * killed makes the (vital) call to commitrw very sad.
2972          */
2973         RETURN(1);
2974 }
2975
2976 static inline const char *bulk2type(struct ptlrpc_request *req)
2977 {
2978         if (req->rq_bulk_read)
2979                 return "READ";
2980         if (req->rq_bulk_write)
2981                 return "WRITE";
2982         return "UNKNOWN";
2983 }
2984
2985 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2986                    struct l_wait_info *lwi)
2987 {
2988         struct ptlrpc_request   *req = desc->bd_req;
2989         time_t                   start = cfs_time_current_sec();
2990         time_t                   deadline;
2991         int                      rc = 0;
2992
2993         ENTRY;
2994
2995         /* If there is eviction in progress, wait for it to finish. */
2996         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2997                 *lwi = LWI_INTR(NULL, NULL);
2998                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2999                                   !atomic_read(&exp->exp_obd->
3000                                                    obd_evict_inprogress),
3001                                   lwi);
3002         }
3003
3004         /* Check if client was evicted or reconnected already. */
3005         if (exp->exp_failed ||
3006             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3007                 rc = -ENOTCONN;
3008         } else {
3009                 if (req->rq_bulk_read)
3010                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3011
3012                 if ((exp->exp_connect_data.ocd_connect_flags &
3013                      OBD_CONNECT_BULK_MBITS) != 0)
3014                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3015                 else /* old version, bulk matchbits is rq_xid */
3016                         req->rq_mbits = req->rq_xid;
3017
3018                 if (rc == 0)
3019                         rc = ptlrpc_start_bulk_transfer(desc);
3020         }
3021
3022         if (rc < 0) {
3023                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3024                           bulk2type(req), rc);
3025                 RETURN(rc);
3026         }
3027
3028         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3029                 ptlrpc_abort_bulk(desc);
3030                 RETURN(0);
3031         }
3032
3033         /* limit actual bulk transfer to bulk_timeout seconds */
3034         deadline = start + bulk_timeout;
3035         if (deadline > req->rq_deadline)
3036                 deadline = req->rq_deadline;
3037
3038         do {
3039                 long timeoutl = deadline - cfs_time_current_sec();
3040                 cfs_duration_t timeout = timeoutl <= 0 ?
3041                                          CFS_TICK : cfs_time_seconds(timeoutl);
3042                 time_t  rq_deadline;
3043
3044                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3045                                             target_bulk_timeout, desc);
3046                 rc = l_wait_event(desc->bd_waitq,
3047                                   !ptlrpc_server_bulk_active(desc) ||
3048                                   exp->exp_failed ||
3049                                   exp->exp_conn_cnt >
3050                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3051                                   lwi);
3052                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3053                 /* Wait again if we changed rq_deadline. */
3054                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3055                 deadline = start + bulk_timeout;
3056                 if (deadline > rq_deadline)
3057                         deadline = rq_deadline;
3058         } while ((rc == -ETIMEDOUT) &&
3059                  (deadline > cfs_time_current_sec()));
3060
3061         if (rc == -ETIMEDOUT) {
3062                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3063                           bulk2type(req), deadline - start,
3064                           cfs_time_current_sec() - deadline);
3065                 ptlrpc_abort_bulk(desc);
3066         } else if (exp->exp_failed) {
3067                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3068                           bulk2type(req));
3069                 rc = -ENOTCONN;
3070                 ptlrpc_abort_bulk(desc);
3071         } else if (exp->exp_conn_cnt >
3072                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3073                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3074                           bulk2type(req));
3075                 /* We don't reply anyway. */
3076                 rc = -ETIMEDOUT;
3077                 ptlrpc_abort_bulk(desc);
3078         } else if (desc->bd_failure) {
3079                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3080                           bulk2type(req));
3081                 /* XXX should this be a different errno? */
3082                 rc = -ETIMEDOUT;
3083         } else {
3084                 if (req->rq_bulk_write)
3085                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3086                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3087                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3088                                   bulk2type(req), desc->bd_nob_transferred,
3089                                   desc->bd_nob);
3090                         /* XXX should this be a different errno? */
3091                         rc = -ETIMEDOUT;
3092                 }
3093         }
3094
3095         RETURN(rc);
3096 }
3097 EXPORT_SYMBOL(target_bulk_io);
3098
3099 #endif /* HAVE_SERVER_SUPPORT */