lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 /**
  38  * This file deals with various client/target related logic including recovery.
  39  *
  40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  41  * should be moved.
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LDLM
  45
  46 #include <linux/kthread.h>
  47 #include <libcfs/libcfs.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_dlm.h>
  51 #include <lustre_net.h>
  52 #include <lustre_sec.h>
  53 #include "ldlm_internal.h"
  54
  55 /* @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         int rc = 0;
  64         ENTRY;
  65
  66         if (!create && !priority) {
  67                 CDEBUG(D_HA, "Nothing to do\n");
  68                 RETURN(-EINVAL);
  69         }
  70
  71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  72         if (!ptlrpc_conn) {
  73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  74                 RETURN (-ENOENT);
  75         }
  76
  77         if (create) {
  78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  79                 if (!imp_conn) {
  80                         GOTO(out_put, rc = -ENOMEM);
  81                 }
  82         }
  83
  84         spin_lock(&imp->imp_lock);
  85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  87                         if (priority) {
  88                                 list_del(&item->oic_item);
  89                                 list_add(&item->oic_item,
  90                                              &imp->imp_conn_list);
  91                                 item->oic_last_attempt = 0;
  92                         }
  93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  94                                imp, imp->imp_obd->obd_name, uuid->uuid,
  95                                (priority ? ", moved to head" : ""));
  96                         spin_unlock(&imp->imp_lock);
  97                         GOTO(out_free, rc = 0);
  98                 }
  99         }
 100         /* No existing import connection found for \a uuid. */
 101         if (create) {
 102                 imp_conn->oic_conn = ptlrpc_conn;
 103                 imp_conn->oic_uuid = *uuid;
 104                 imp_conn->oic_last_attempt = 0;
 105                 if (priority)
 106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 else
 108                         list_add_tail(&imp_conn->oic_item,
 109                                           &imp->imp_conn_list);
 110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 111                        imp, imp->imp_obd->obd_name, uuid->uuid,
 112                        (priority ? "head" : "tail"));
 113         } else {
 114                 spin_unlock(&imp->imp_lock);
 115                 GOTO(out_free, rc = -ENOENT);
 116         }
 117
 118         spin_unlock(&imp->imp_lock);
 119         RETURN(0);
 120 out_free:
 121         if (imp_conn)
 122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 123 out_put:
 124         ptlrpc_connection_put(ptlrpc_conn);
 125         RETURN(rc);
 126 }
 127
 128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 129 {
 130         return import_set_conn(imp, uuid, 1, 0);
 131 }
 132
 133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 134                            int priority)
 135 {
 136         return import_set_conn(imp, uuid, priority, 1);
 137 }
 138 EXPORT_SYMBOL(client_import_add_conn);
 139
 140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 141 {
 142         struct obd_import_conn *imp_conn;
 143         struct obd_export *dlmexp;
 144         int rc = -ENOENT;
 145         ENTRY;
 146
 147         spin_lock(&imp->imp_lock);
 148         if (list_empty(&imp->imp_conn_list)) {
 149                 LASSERT(!imp->imp_connection);
 150                 GOTO(out, rc);
 151         }
 152
 153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 155                         continue;
 156                 LASSERT(imp_conn->oic_conn);
 157
 158                 if (imp_conn == imp->imp_conn_current) {
 159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 160
 161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 162                             imp->imp_state != LUSTRE_IMP_DISCON) {
 163                                 CERROR("can't remove current connection\n");
 164                                 GOTO(out, rc = -EBUSY);
 165                         }
 166
 167                         ptlrpc_connection_put(imp->imp_connection);
 168                         imp->imp_connection = NULL;
 169
 170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 171                         if (dlmexp && dlmexp->exp_connection) {
 172                                 LASSERT(dlmexp->exp_connection ==
 173                                         imp_conn->oic_conn);
 174                                 ptlrpc_connection_put(dlmexp->exp_connection);
 175                                 dlmexp->exp_connection = NULL;
 176                         }
 177                 }
 178
 179                 list_del(&imp_conn->oic_item);
 180                 ptlrpc_connection_put(imp_conn->oic_conn);
 181                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 182                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 183                        imp, imp->imp_obd->obd_name, uuid->uuid);
 184                 rc = 0;
 185                 break;
 186         }
 187 out:
 188         spin_unlock(&imp->imp_lock);
 189         if (rc == -ENOENT)
 190                 CERROR("connection %s not found\n", uuid->uuid);
 191         RETURN(rc);
 192 }
 193 EXPORT_SYMBOL(client_import_del_conn);
 194
 195 /**
 196  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 197  * to find a conn uuid of \a imp which can reach \a peer.
 198  */
 199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 200                             struct obd_uuid *uuid)
 201 {
 202         struct obd_import_conn *conn;
 203         int rc = -ENOENT;
 204         ENTRY;
 205
 206         spin_lock(&imp->imp_lock);
 207         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 208                 /* Check if conn UUID does have this peer NID. */
 209                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 210                         *uuid = conn->oic_uuid;
 211                         rc = 0;
 212                         break;
 213                 }
 214         }
 215         spin_unlock(&imp->imp_lock);
 216         RETURN(rc);
 217 }
 218 EXPORT_SYMBOL(client_import_find_conn);
 219
 220 void client_destroy_import(struct obd_import *imp)
 221 {
 222         /* Drop security policy instance after all RPCs have finished/aborted
 223          * to let all busy contexts be released. */
 224         class_import_get(imp);
 225         class_destroy_import(imp);
 226         sptlrpc_import_sec_put(imp);
 227         class_import_put(imp);
 228 }
 229 EXPORT_SYMBOL(client_destroy_import);
 230
 231 /**
 232  * Check whether or not the OSC is on MDT.
 233  * In the config log,
 234  * osc on MDT
 235  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 236  * osc on client
 237  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 238  *
 239  **/
 240 static int osc_on_mdt(char *obdname)
 241 {
 242         char *ptr;
 243
 244         ptr = strrchr(obdname, '-');
 245         if (ptr == NULL)
 246                 return 0;
 247
 248         if (strncmp(ptr + 1, "MDT", 3) == 0)
 249                 return 1;
 250
 251         return 0;
 252 }
 253
 254 /* Configure an RPC client OBD device.
 255  *
 256  * lcfg parameters:
 257  * 1 - client UUID
 258  * 2 - server UUID
 259  * 3 - inactive-on-startup
 260  */
 261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 262 {
 263         struct client_obd *cli = &obddev->u.cli;
 264         struct obd_import *imp;
 265         struct obd_uuid server_uuid;
 266         int rq_portal, rp_portal, connect_op;
 267         char *name = obddev->obd_type->typ_name;
 268         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 269         char *cli_name = lustre_cfg_buf(lcfg, 0);
 270         int rc;
 271         ENTRY;
 272
 273         /* In a more perfect world, we would hang a ptlrpc_client off of
 274          * obd_type and just use the values from there. */
 275         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 276                 rq_portal = OST_REQUEST_PORTAL;
 277                 rp_portal = OSC_REPLY_PORTAL;
 278                 connect_op = OST_CONNECT;
 279                 cli->cl_sp_me = LUSTRE_SP_CLI;
 280                 cli->cl_sp_to = LUSTRE_SP_OST;
 281                 ns_type = LDLM_NS_TYPE_OSC;
 282         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 283                    !strcmp(name, LUSTRE_LWP_NAME)) {
 284                 rq_portal = MDS_REQUEST_PORTAL;
 285                 rp_portal = MDC_REPLY_PORTAL;
 286                 connect_op = MDS_CONNECT;
 287                 if (is_lwp_on_ost(cli_name))
 288                         cli->cl_sp_me = LUSTRE_SP_OST;
 289                 else if (is_lwp_on_mdt(cli_name))
 290                         cli->cl_sp_me = LUSTRE_SP_MDT;
 291                 else
 292                         cli->cl_sp_me = LUSTRE_SP_CLI;
 293                 cli->cl_sp_to = LUSTRE_SP_MDT;
 294                 ns_type = LDLM_NS_TYPE_MDC;
 295         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 296                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 297                         /* OSP_on_MDT for other MDTs */
 298                         connect_op = MDS_CONNECT;
 299                         cli->cl_sp_to = LUSTRE_SP_MDT;
 300                         ns_type = LDLM_NS_TYPE_MDC;
 301                         rq_portal = OUT_PORTAL;
 302                 } else {
 303                         /* OSP on MDT for OST */
 304                         connect_op = OST_CONNECT;
 305                         cli->cl_sp_to = LUSTRE_SP_OST;
 306                         ns_type = LDLM_NS_TYPE_OSC;
 307                         rq_portal = OST_REQUEST_PORTAL;
 308                 }
 309                 rp_portal = OSC_REPLY_PORTAL;
 310                 cli->cl_sp_me = LUSTRE_SP_MDT;
 311         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 312                 rq_portal = MGS_REQUEST_PORTAL;
 313                 rp_portal = MGC_REPLY_PORTAL;
 314                 connect_op = MGS_CONNECT;
 315                 cli->cl_sp_me = LUSTRE_SP_MGC;
 316                 cli->cl_sp_to = LUSTRE_SP_MGS;
 317                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 318                 ns_type = LDLM_NS_TYPE_MGC;
 319         } else {
 320                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 321                        name);
 322                 RETURN(-EINVAL);
 323         }
 324
 325         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 326                 CERROR("requires a TARGET UUID\n");
 327                 RETURN(-EINVAL);
 328         }
 329
 330         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 331                 CERROR("client UUID must be less than 38 characters\n");
 332                 RETURN(-EINVAL);
 333         }
 334
 335         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 336                 CERROR("setup requires a SERVER UUID\n");
 337                 RETURN(-EINVAL);
 338         }
 339
 340         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 341                 CERROR("target UUID must be less than 38 characters\n");
 342                 RETURN(-EINVAL);
 343         }
 344
 345         init_rwsem(&cli->cl_sem);
 346         mutex_init(&cli->cl_mgc_mutex);
 347         cli->cl_conn_count = 0;
 348         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 349                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 350                      sizeof(server_uuid)));
 351
 352         cli->cl_dirty_pages = 0;
 353         cli->cl_avail_grant = 0;
 354         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 355         /* cl_dirty_max_pages may be changed at connect time in
 356          * ptlrpc_connect_interpret(). */
 357         client_adjust_max_dirty(cli);
 358         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 359         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 360         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 361         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 362         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 363         spin_lock_init(&cli->cl_loi_list_lock);
 364         atomic_set(&cli->cl_pending_w_pages, 0);
 365         atomic_set(&cli->cl_pending_r_pages, 0);
 366         cli->cl_r_in_flight = 0;
 367         cli->cl_w_in_flight = 0;
 368
 369         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 370         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 371         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 372         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 373         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 374         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 375
 376         /* lru for osc. */
 377         INIT_LIST_HEAD(&cli->cl_lru_osc);
 378         atomic_set(&cli->cl_lru_shrinkers, 0);
 379         atomic_long_set(&cli->cl_lru_busy, 0);
 380         atomic_long_set(&cli->cl_lru_in_list, 0);
 381         INIT_LIST_HEAD(&cli->cl_lru_list);
 382         spin_lock_init(&cli->cl_lru_list_lock);
 383         atomic_long_set(&cli->cl_unstable_count, 0);
 384         INIT_LIST_HEAD(&cli->cl_shrink_list);
 385
 386         init_waitqueue_head(&cli->cl_destroy_waitq);
 387         atomic_set(&cli->cl_destroy_in_flight, 0);
 388 #ifdef ENABLE_CHECKSUM
 389         /* Turn on checksumming by default. */
 390         cli->cl_checksum = 1;
 391         /*
 392          * The supported checksum types will be worked out at connect time
 393          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 394          * through procfs.
 395          */
 396         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 397 #endif
 398         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 399
 400         /* 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 obd_device *target;
 726         struct lustre_handle *hdl;
 727         cfs_time_t now;
 728         cfs_time_t deadline;
 729         int timeout;
 730         int rc = 0;
 731         ENTRY;
 732
 733         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 734         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 735                 conn->cookie = exp->exp_handle.h_cookie;
 736                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 737                        " cookie "LPX64"\n", cluuid->uuid, exp, conn->cookie);
 738                 RETURN(0);
 739         }
 740
 741         target = exp->exp_obd;
 742
 743         /* Might be a re-connect after a partition. */
 744         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 745                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 746                               "with handle "LPX64". Rejecting client "
 747                               "with the same UUID trying to reconnect "
 748                               "with handle "LPX64"\n", target->obd_name,
 749                               obd_uuid2str(&exp->exp_client_uuid),
 750                               obd_export_nid2str(exp),
 751                               hdl->cookie, conn->cookie);
 752                 memset(conn, 0, sizeof *conn);
 753                 /* target_handle_connect() treats EALREADY and
 754                  * -EALREADY differently.  -EALREADY is an error
 755                  * (same UUID, different handle). */
 756                 RETURN(-EALREADY);
 757         }
 758
 759         if (!target->obd_recovering) {
 760                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 761                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 762                         obd_export_nid2str(exp));
 763                 GOTO(out_already, rc);
 764         }
 765
 766         now = cfs_time_current();
 767         deadline = cfs_timer_deadline(&target->obd_recovery_timer);
 768         if (cfs_time_before(now, deadline)) {
 769                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
 770                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting,"
 771                         " waiting for %d clients in recovery for"
 772                         " %d:%.02d\n", target->obd_name,
 773                         obd_uuid2str(&exp->exp_client_uuid),
 774                         obd_export_nid2str(exp),
 775                         target->obd_max_recoverable_clients,
 776                         timeout / 60, timeout % 60);
 777         } else {
 778                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
 779                 LCONSOLE_WARN("%s: Recovery already passed deadline"
 780                         " %d:%.02d, It is most likely due to DNE"
 781                         " recovery is failed or stuck, please wait a"
 782                         " few more minutes or abort the recovery.\n",
 783                         target->obd_name, timeout / 60, timeout % 60);
 784         }
 785
 786 out_already:
 787         conn->cookie = exp->exp_handle.h_cookie;
 788         /* target_handle_connect() treats EALREADY and
 789          * -EALREADY differently.  EALREADY means we are
 790          * doing a valid reconnect from the same client. */
 791         RETURN(EALREADY);
 792 }
 793
 794 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 795                           int error)
 796 {
 797         struct obd_export *exp = cb_data;
 798
 799         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 800                obd->obd_name, exp->exp_client_uuid.uuid);
 801
 802         spin_lock(&exp->exp_lock);
 803         exp->exp_need_sync = 0;
 804         spin_unlock(&exp->exp_lock);
 805         class_export_cb_put(exp);
 806 }
 807
 808 static void
 809 check_and_start_recovery_timer(struct obd_device *obd,
 810                                struct ptlrpc_request *req, int new_client);
 811
 812 /**
 813  * update flags for import during reconnect process
 814  */
 815 static int rev_import_flags_update(struct obd_import *revimp,
 816                                    struct ptlrpc_request *req)
 817 {
 818         int rc;
 819         struct obd_connect_data *data;
 820
 821         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 822
 823         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 824                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 825         else
 826                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 827
 828         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 829
 830         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 831         if (rc) {
 832                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 833                         revimp->imp_client->cli_name,
 834                         libcfs_id2str(req->rq_peer), rc);
 835                 return rc;
 836         }
 837
 838         return 0;
 839 }
 840
 841 /**
 842  * Allocate a new reverse import for an export.
 843  *
 844  * \retval -errno in case error hit
 845  * \retval 0 if reverse import correctly init
 846  **/
 847 int rev_import_init(struct obd_export *export)
 848 {
 849         struct obd_device *obd = export->exp_obd;
 850         struct obd_import *revimp;
 851
 852         LASSERT(export->exp_imp_reverse == NULL);
 853
 854         revimp = class_new_import(obd);
 855         if (revimp == NULL)
 856                 return -ENOMEM;
 857
 858         revimp->imp_remote_handle.cookie = 0ULL;
 859         revimp->imp_client = &obd->obd_ldlm_client;
 860         revimp->imp_dlm_fake = 1;
 861
 862         /* it is safe to connect import in new state as no sends possible */
 863         spin_lock(&export->exp_lock);
 864         export->exp_imp_reverse = revimp;
 865         spin_unlock(&export->exp_lock);
 866         class_import_put(revimp);
 867
 868         return 0;
 869 }
 870 EXPORT_SYMBOL(rev_import_init);
 871
 872 /**
 873  * Handle reconnect for an export.
 874  *
 875  * \param exp export to handle reconnect process
 876  * \param req client reconnect request
 877  *
 878  * \retval -rc in case securitfy flavor can't be changed
 879  * \retval 0 in case none problems
 880  */
 881 static int rev_import_reconnect(struct obd_export *exp,
 882                                 struct ptlrpc_request *req)
 883 {
 884         struct obd_import *revimp = exp->exp_imp_reverse;
 885         struct lustre_handle *lh;
 886         int rc;
 887
 888         /* avoid sending a request until import flags are changed */
 889         ptlrpc_import_enter_resend(revimp);
 890
 891         if (revimp->imp_connection != NULL)
 892                 ptlrpc_connection_put(revimp->imp_connection);
 893
 894         /*
 895          * client from recovery don't have a handle so we need to take from
 896          * request. it may produce situation when wrong client connected
 897          * to recovery as we trust a client uuid
 898          */
 899         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 900         revimp->imp_remote_handle = *lh;
 901
 902         /* unknown versions will be caught in
 903          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 904         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 905
 906         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 907
 908         rc = rev_import_flags_update(revimp, req);
 909         if (rc != 0) {
 910                 /* it is safe to still be in RECOVERY phase as we are not able
 911                  * to setup correct security flavor so requests are not able to
 912                  * be delivered correctly */
 913                 return rc;
 914         }
 915
 916         /* resend all rpc's via new connection */
 917         return ptlrpc_import_recovery_state_machine(revimp);
 918 }
 919
 920 int target_handle_connect(struct ptlrpc_request *req)
 921 {
 922         struct obd_device *target = NULL;
 923         struct obd_export *export = NULL;
 924         /* connect handle - filled from target_handle_reconnect in
 925          * reconnect case */
 926         struct lustre_handle conn;
 927         struct lustre_handle *tmp;
 928         struct obd_uuid tgtuuid;
 929         struct obd_uuid cluuid;
 930         char *str;
 931         int rc = 0;
 932         char *target_start;
 933         int target_len;
 934         bool     mds_conn = false, lw_client = false;
 935         bool     mds_mds_conn = false;
 936         bool     new_mds_mds_conn = false;
 937         bool     target_referenced = false;
 938         struct obd_connect_data *data, *tmpdata;
 939         int size, tmpsize;
 940         lnet_nid_t *client_nid = NULL;
 941         ENTRY;
 942
 943         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 944
 945         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 946         if (str == NULL) {
 947                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 948                 GOTO(out, rc = -EINVAL);
 949         }
 950
 951         obd_str2uuid(&tgtuuid, str);
 952         target = class_uuid2obd(&tgtuuid);
 953         if (!target)
 954                 target = class_name2obd(str);
 955
 956         if (!target) {
 957                 deuuidify(str, NULL, &target_start, &target_len);
 958                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 959                                    "from %s (no target). If you are running "
 960                                    "an HA pair check that the target is "
 961                                    "mounted on the other server.\n", str,
 962                                    libcfs_nid2str(req->rq_peer.nid));
 963                 GOTO(out, rc = -ENODEV);
 964         }
 965
 966         spin_lock(&target->obd_dev_lock);
 967         if (target->obd_stopping || !target->obd_set_up) {
 968                 spin_unlock(&target->obd_dev_lock);
 969
 970                 deuuidify(str, NULL, &target_start, &target_len);
 971                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 972                               target_len, target_start,
 973                               libcfs_nid2str(req->rq_peer.nid),
 974                               (target->obd_stopping ?
 975                                "stopping" : "not set up"));
 976                 GOTO(out, rc = -ENODEV);
 977         }
 978
 979         if (target->obd_no_conn) {
 980                 spin_unlock(&target->obd_dev_lock);
 981
 982                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 983                                "from %s\n", target->obd_name,
 984                                libcfs_nid2str(req->rq_peer.nid));
 985                 GOTO(out, rc = -EAGAIN);
 986         }
 987
 988         /* Make sure the target isn't cleaned up while we're here. Yes,
 989          * there's still a race between the above check and our incref here.
 990          * Really, class_uuid2obd should take the ref. */
 991         class_incref(target, __func__, current);
 992         target_referenced = true;
 993
 994         target->obd_conn_inprogress++;
 995         spin_unlock(&target->obd_dev_lock);
 996
 997         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 998         if (str == NULL) {
 999                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1000                 GOTO(out, rc = -EINVAL);
1001         }
1002
1003         obd_str2uuid(&cluuid, str);
1004
1005         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1006         if (tmp == NULL)
1007                 GOTO(out, rc = -EPROTO);
1008
1009         conn = *tmp;
1010
1011         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1012                                     RCL_CLIENT);
1013         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1014         if (!data)
1015                 GOTO(out, rc = -EPROTO);
1016
1017         rc = req_capsule_server_pack(&req->rq_pill);
1018         if (rc)
1019                 GOTO(out, rc);
1020
1021 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1022         /* Don't allow clients to connect that are using old 1.8 format
1023          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1024          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1025          * FULL20 flag should be set on all connections since 2.0, but no
1026          * longer affects behaviour.
1027          *
1028          * Later this check will be disabled and the flag can be retired
1029          * completely once interop with 3.0 is no longer needed.
1030          */
1031         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1032                 GOTO(out, rc = -EPROTO);
1033 #endif
1034
1035         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1036                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1037                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1038                     data->ocd_version > LUSTRE_VERSION_CODE +
1039                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1040                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1041                                   "libclient connection attempt",
1042                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1043                                   "old" : "new",
1044                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1045                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1046                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1047                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1048                         data = req_capsule_server_sized_get(&req->rq_pill,
1049                                                             &RMF_CONNECT_DATA,
1050                                     offsetof(typeof(*data), ocd_version) +
1051                                              sizeof(data->ocd_version));
1052                         if (data) {
1053                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1054                                 data->ocd_version = LUSTRE_VERSION_CODE;
1055                         }
1056                         GOTO(out, rc = -EPROTO);
1057                 }
1058         }
1059
1060         /* Note: lw_client is needed in MDS-MDS failover during update log
1061          * processing, so we needs to allow lw_client to be connected at
1062          * anytime, instead of only the initial connection */
1063         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1064
1065         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1066                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1067                 mds_mds_conn = (data->ocd_connect_flags &
1068                                 OBD_CONNECT_MDS_MDS) != 0;
1069
1070                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1071                  * for Imperative Recovery connection from MGC to MGS.
1072                  *
1073                  * Via check OBD_CONNECT_FID, we can distinguish whether
1074                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1075                  * MGC or MDT. */
1076                 if (!lw_client &&
1077                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1078                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1079                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1080                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1081                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1082                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1083
1084                         /* We do not support the MDT-MDT interoperations with
1085                          * different version MDT because of protocol changes. */
1086                         if (unlikely(major != LUSTRE_MAJOR ||
1087                                      minor != LUSTRE_MINOR ||
1088                                      abs(patch - LUSTRE_PATCH) > 3)) {
1089                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1090                                         "connection from different version MDT "
1091                                         "(%d.%d.%d.%d) %s %s\n",
1092                                         target->obd_name, LUSTRE_MAJOR,
1093                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1094                                         major, minor, patch,
1095                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1096                                         libcfs_nid2str(req->rq_peer.nid), str);
1097
1098                                 GOTO(out, rc = -EPROTO);
1099                         }
1100                 }
1101         }
1102
1103         /* lctl gets a backstage, all-access pass. */
1104         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1105                 goto dont_check_exports;
1106
1107         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1108         if (!export)
1109                 goto no_export;
1110
1111         /* We've found an export in the hash. */
1112
1113         spin_lock(&export->exp_lock);
1114
1115         if (export->exp_connecting) { /* bug 9635, et. al. */
1116                 spin_unlock(&export->exp_lock);
1117                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1118                               export->exp_obd->obd_name, export,
1119                               libcfs_nid2str(req->rq_peer.nid));
1120                 class_export_put(export);
1121                 export = NULL;
1122                 rc = -EALREADY;
1123         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1124                 spin_unlock(&export->exp_lock);
1125                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1126                         /* MDS or LWP reconnected after failover. */
1127                         LCONSOLE_WARN("%s: Received %s connection from "
1128                             "%s, removing former export from %s\n",
1129                             target->obd_name, mds_conn ? "MDS" : "LWP",
1130                             libcfs_nid2str(req->rq_peer.nid),
1131                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1132                 else
1133                         /* New MDS connection from the same NID. */
1134                         LCONSOLE_WARN("%s: Received new %s connection from "
1135                                 "%s, removing former export from same NID\n",
1136                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1137                                 libcfs_nid2str(req->rq_peer.nid));
1138                 class_fail_export(export);
1139                 class_export_put(export);
1140                 export = NULL;
1141                 rc = 0;
1142         } else if (export->exp_connection != NULL &&
1143                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1144                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1145                     MSG_CONNECT_INITIAL)) {
1146                 spin_unlock(&export->exp_lock);
1147                 /* In MDS failover we have static UUID but NID can change. */
1148                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1149                               "existing nid %s is already connected\n",
1150                               target->obd_name, cluuid.uuid,
1151                               libcfs_nid2str(req->rq_peer.nid),
1152                               libcfs_nid2str(
1153                                       export->exp_connection->c_peer.nid));
1154                 rc = -EALREADY;
1155                 class_export_put(export);
1156                 export = NULL;
1157         } else {
1158                 export->exp_connecting = 1;
1159                 spin_unlock(&export->exp_lock);
1160                 LASSERT(export->exp_obd == target);
1161
1162                 rc = target_handle_reconnect(&conn, export, &cluuid);
1163         }
1164
1165         /* If we found an export, we already unlocked. */
1166         if (!export) {
1167 no_export:
1168                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1169         } else if (req->rq_export == NULL &&
1170                    atomic_read(&export->exp_rpc_count) > 0) {
1171                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1172                               "still busy with %d references\n",
1173                               target->obd_name, cluuid.uuid,
1174                               libcfs_nid2str(req->rq_peer.nid),
1175                               atomic_read(&export->exp_refcount));
1176                 GOTO(out, rc = -EBUSY);
1177         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1178                 if (!strstr(cluuid.uuid, "mdt"))
1179                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1180                                       "known client %s (at %s) because it "
1181                                       "is indicating it is a new client",
1182                                       target->obd_name, cluuid.uuid,
1183                                       libcfs_nid2str(req->rq_peer.nid));
1184                 GOTO(out, rc = -EALREADY);
1185         } else {
1186                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1187         }
1188
1189         if (rc < 0) {
1190                 GOTO(out, rc);
1191         }
1192
1193         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1194                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1195               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1196               export, (long)cfs_time_current_sec(),
1197               export ? (long)export->exp_last_request_time : 0);
1198
1199         /* If this is the first time a client connects, reset the recovery
1200          * timer. Discard lightweight connections which might be local. */
1201         if (!lw_client && rc == 0 && target->obd_recovering)
1202                 check_and_start_recovery_timer(target, req, export == NULL);
1203
1204         /* We want to handle EALREADY but *not* -EALREADY from
1205          * target_handle_reconnect(), return reconnection state in a flag. */
1206         if (rc == EALREADY) {
1207                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1208                 rc = 0;
1209         } else {
1210                 LASSERT(rc == 0);
1211         }
1212
1213         /* Tell the client if we support replayable requests. */
1214         if (target->obd_replayable)
1215                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1216         client_nid = &req->rq_peer.nid;
1217
1218         if (export == NULL) {
1219                 /* allow lightweight connections during recovery */
1220                 /* allow "new" MDT to be connected during recovery, since we
1221                  * need retrieve recovery update records from it */
1222                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1223                         cfs_time_t t;
1224                         int     c; /* connected */
1225                         int     i; /* in progress */
1226                         int     k; /* known */
1227                         int     s; /* stale/evicted */
1228
1229                         c = atomic_read(&target->obd_connected_clients);
1230                         i = atomic_read(&target->obd_lock_replay_clients);
1231                         k = target->obd_max_recoverable_clients;
1232                         s = target->obd_stale_clients;
1233                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1234                         t = cfs_time_sub(t, cfs_time_current());
1235                         t = cfs_duration_sec(t);
1236                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1237                                       "(at %s), waiting for %d known clients "
1238                                       "(%d recovered, %d in progress, and %d "
1239                                       "evicted) to recover in %d:%.02d\n",
1240                                       target->obd_name, cluuid.uuid,
1241                                       libcfs_nid2str(req->rq_peer.nid), k,
1242                                       c - i, i, s, (int)t / 60,
1243                                       (int)t % 60);
1244                         rc = -EBUSY;
1245                 } else {
1246 dont_check_exports:
1247                         rc = obd_connect(req->rq_svc_thread->t_env,
1248                                          &export, target, &cluuid, data,
1249                                          client_nid);
1250                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1251                                 lustre_msg_add_op_flags(req->rq_repmsg,
1252                                                         MSG_CONNECT_RECOVERING);
1253                         if (rc == 0) {
1254                                 conn.cookie = export->exp_handle.h_cookie;
1255                                 rc = rev_import_init(export);
1256                         }
1257
1258                         if (mds_mds_conn)
1259                                 new_mds_mds_conn = true;
1260                 }
1261         } else {
1262                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1263                                    export, target, &cluuid, data, client_nid);
1264         }
1265         if (rc)
1266                 GOTO(out, rc);
1267
1268         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1269         data->ocd_instance = target->u.obt.obt_instance;
1270
1271         /* Return only the parts of obd_connect_data that we understand, so the
1272          * client knows that we don't understand the rest. */
1273         if (data) {
1274                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1275                                                RCL_SERVER);
1276                 tmpdata = req_capsule_server_get(&req->rq_pill,
1277                                                  &RMF_CONNECT_DATA);
1278                 /* Don't use struct assignment here, because the client reply
1279                  * buffer may be smaller/larger than the local struct
1280                  * obd_connect_data. */
1281                 memcpy(tmpdata, data, min(tmpsize, size));
1282         }
1283
1284         /* If the client and the server are the same node, we will already
1285          * have an export that really points to the client's DLM export,
1286          * because we have a shared handles table.
1287          *
1288          * XXX this will go away when shaver stops sending the "connect" handle
1289          * in the real "remote handle" field of the request --phik 24 Apr 2003
1290          */
1291         ptlrpc_request_change_export(req, export);
1292
1293         spin_lock(&export->exp_lock);
1294         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1295                 spin_unlock(&export->exp_lock);
1296                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1297                        "or equal conn_cnt: %d >= %d\n",
1298                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1299                        export->exp_conn_cnt,
1300                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1301
1302                 GOTO(out, rc = -EALREADY);
1303         }
1304         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1305         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1306
1307         /* Don't evict liblustre clients for not pinging. */
1308         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1309                 export->exp_libclient = 1;
1310                 spin_unlock(&export->exp_lock);
1311
1312                 spin_lock(&target->obd_dev_lock);
1313                 list_del_init(&export->exp_obd_chain_timed);
1314                 spin_unlock(&target->obd_dev_lock);
1315         } else {
1316                 spin_unlock(&export->exp_lock);
1317         }
1318
1319         if (export->exp_connection != NULL) {
1320                 /* Check to see if connection came from another NID. */
1321                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1322                     !hlist_unhashed(&export->exp_nid_hash))
1323                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1324                                      &export->exp_connection->c_peer.nid,
1325                                      &export->exp_nid_hash);
1326
1327                 ptlrpc_connection_put(export->exp_connection);
1328         }
1329
1330         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1331                                                        req->rq_self,
1332                                                        &cluuid);
1333         if (hlist_unhashed(&export->exp_nid_hash))
1334                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1335                              &export->exp_connection->c_peer.nid,
1336                              &export->exp_nid_hash);
1337
1338         lustre_msg_set_handle(req->rq_repmsg, &conn);
1339
1340         rc = rev_import_reconnect(export, req);
1341         if (rc != 0)
1342                 GOTO(out, rc);
1343
1344         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1345                 int has_transno;
1346                 __u64 transno = data->ocd_transno;
1347
1348                 spin_lock(&export->exp_lock);
1349                 /* possible race with class_disconnect_stale_exports,
1350                  * export may be already in the eviction process */
1351                 if (export->exp_failed) {
1352                         spin_unlock(&export->exp_lock);
1353                         GOTO(out, rc = -ENODEV);
1354                 }
1355                 export->exp_in_recovery = 1;
1356                 export->exp_req_replay_needed = 1;
1357                 export->exp_lock_replay_needed = 1;
1358                 spin_unlock(&export->exp_lock);
1359
1360                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1361                                  MSG_CONNECT_TRANSNO);
1362                 if (has_transno && transno == 0)
1363                         CWARN("Connect with zero transno!\n");
1364
1365                 if (has_transno && transno > 0 &&
1366                     transno < target->obd_next_recovery_transno &&
1367                     transno > target->obd_last_committed) {
1368                         /* Another way is to use cmpxchg() to be lock-free. */
1369                         spin_lock(&target->obd_recovery_task_lock);
1370                         if (transno < target->obd_next_recovery_transno)
1371                                 target->obd_next_recovery_transno = transno;
1372                         spin_unlock(&target->obd_recovery_task_lock);
1373                 }
1374
1375                 atomic_inc(&target->obd_req_replay_clients);
1376                 atomic_inc(&target->obd_lock_replay_clients);
1377                 /* Note: MDS-MDS connection is allowed to be connected during
1378                  * recovery, no matter if the exports needs to be recoveried.
1379                  * Because we need retrieve updates logs from all other MDTs.
1380                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1381                  * also needs to be increased to match other recovery checking
1382                  * condition. */
1383                 if (new_mds_mds_conn)
1384                         target->obd_max_recoverable_clients++;
1385                 if (atomic_inc_return(&target->obd_connected_clients) ==
1386                     target->obd_max_recoverable_clients)
1387                         wake_up(&target->obd_next_transno_waitq);
1388         }
1389
1390         /* Tell the client we're in recovery, when client is involved in it. */
1391         if (target->obd_recovering && !lw_client)
1392                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1393
1394 out:
1395         if (export) {
1396                 spin_lock(&export->exp_lock);
1397                 export->exp_connecting = 0;
1398                 spin_unlock(&export->exp_lock);
1399
1400                 class_export_put(export);
1401         }
1402         if (target_referenced == true && target != NULL) {
1403                 spin_lock(&target->obd_dev_lock);
1404                 target->obd_conn_inprogress--;
1405                 spin_unlock(&target->obd_dev_lock);
1406
1407                 class_decref(target, __func__, current);
1408         }
1409         req->rq_status = rc;
1410         RETURN(rc);
1411 }
1412
1413 int target_handle_disconnect(struct ptlrpc_request *req)
1414 {
1415         int rc;
1416         ENTRY;
1417
1418         rc = req_capsule_server_pack(&req->rq_pill);
1419         if (rc)
1420                 RETURN(rc);
1421
1422         /* Keep the rq_export around so we can send the reply. */
1423         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1424
1425         RETURN(0);
1426 }
1427
1428 void target_destroy_export(struct obd_export *exp)
1429 {
1430         struct obd_import       *imp = NULL;
1431         /* exports created from last_rcvd data, and "fake"
1432            exports created by lctl don't have an import */
1433         spin_lock(&exp->exp_lock);
1434         if (exp->exp_imp_reverse != NULL) {
1435                 imp = exp->exp_imp_reverse;
1436                 exp->exp_imp_reverse = NULL;
1437         }
1438         spin_unlock(&exp->exp_lock);
1439         if (imp != NULL)
1440                 client_destroy_import(imp);
1441
1442         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1443         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1444         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1445         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1446 }
1447 EXPORT_SYMBOL(target_destroy_export);
1448
1449 /*
1450  * Recovery functions
1451  */
1452 static void target_request_copy_get(struct ptlrpc_request *req)
1453 {
1454         class_export_rpc_inc(req->rq_export);
1455         LASSERT(list_empty(&req->rq_list));
1456         INIT_LIST_HEAD(&req->rq_replay_list);
1457
1458         /* Increase refcount to keep request in queue. */
1459         atomic_inc(&req->rq_refcount);
1460         /* Let export know it has replays to be handled. */
1461         atomic_inc(&req->rq_export->exp_replay_count);
1462 }
1463
1464 static void target_request_copy_put(struct ptlrpc_request *req)
1465 {
1466         LASSERT(list_empty(&req->rq_replay_list));
1467         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1468
1469         atomic_dec(&req->rq_export->exp_replay_count);
1470         class_export_rpc_dec(req->rq_export);
1471         ptlrpc_server_drop_request(req);
1472 }
1473
1474 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1475 {
1476         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1477         struct obd_export     *exp = req->rq_export;
1478         struct ptlrpc_request *reqiter;
1479         struct ptlrpc_request *dup_req = NULL;
1480         int                    dup = 0;
1481
1482         LASSERT(exp);
1483
1484         spin_lock(&exp->exp_lock);
1485         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1486                                 rq_replay_list) {
1487                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1488                         dup_req = reqiter;
1489                         dup = 1;
1490                         break;
1491                 }
1492         }
1493
1494         if (dup) {
1495                 /* We expect it with RESENT and REPLAY flags. */
1496                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1497                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1498                         CERROR("invalid flags %x of resent replay\n",
1499                                lustre_msg_get_flags(req->rq_reqmsg));
1500
1501                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1502                         __u32 new_conn;
1503
1504                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1505                         if (new_conn >
1506                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1507                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1508                                                         new_conn);
1509                 }
1510         } else {
1511                 list_add_tail(&req->rq_replay_list,
1512                                   &exp->exp_req_replay_queue);
1513         }
1514
1515         spin_unlock(&exp->exp_lock);
1516         return dup;
1517 }
1518
1519 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1520 {
1521         LASSERT(!list_empty(&req->rq_replay_list));
1522         LASSERT(req->rq_export);
1523
1524         spin_lock(&req->rq_export->exp_lock);
1525         list_del_init(&req->rq_replay_list);
1526         spin_unlock(&req->rq_export->exp_lock);
1527 }
1528
1529 static void target_finish_recovery(struct lu_target *lut)
1530 {
1531         struct obd_device *obd = lut->lut_obd;
1532         ENTRY;
1533
1534         /* Only log a recovery message when recovery has occurred. */
1535         if (obd->obd_recovery_start) {
1536                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1537                                         obd->obd_recovery_start);
1538                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1539                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1540                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1541                         obd->obd_max_recoverable_clients,
1542                         atomic_read(&obd->obd_connected_clients),
1543                         obd->obd_stale_clients,
1544                         obd->obd_stale_clients == 1 ? "was" : "were");
1545         }
1546
1547         ldlm_reprocess_all_ns(obd->obd_namespace);
1548         spin_lock(&obd->obd_recovery_task_lock);
1549         if (!list_empty(&obd->obd_req_replay_queue) ||
1550             !list_empty(&obd->obd_lock_replay_queue) ||
1551             !list_empty(&obd->obd_final_req_queue)) {
1552                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1553                        obd->obd_name,
1554                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1555                        list_empty(&obd->obd_lock_replay_queue) ? \
1556                                "" : "lock ",
1557                        list_empty(&obd->obd_final_req_queue) ? \
1558                                "" : "final ");
1559                 spin_unlock(&obd->obd_recovery_task_lock);
1560                 LBUG();
1561         }
1562         spin_unlock(&obd->obd_recovery_task_lock);
1563
1564         obd->obd_recovery_end = cfs_time_current_sec();
1565
1566         /* When recovery finished, cleanup orphans on MDS and OST. */
1567         if (OBT(obd) && OBP(obd, postrecov)) {
1568                 int rc = OBP(obd, postrecov)(obd);
1569                 if (rc < 0)
1570                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1571                                       obd->obd_name, rc);
1572         }
1573         EXIT;
1574 }
1575
1576 static void abort_req_replay_queue(struct obd_device *obd)
1577 {
1578         struct ptlrpc_request *req, *n;
1579         struct list_head abort_list;
1580
1581         INIT_LIST_HEAD(&abort_list);
1582         spin_lock(&obd->obd_recovery_task_lock);
1583         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1584         spin_unlock(&obd->obd_recovery_task_lock);
1585         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1586                 DEBUG_REQ(D_WARNING, req, "aborted:");
1587                 req->rq_status = -ENOTCONN;
1588                 if (ptlrpc_error(req)) {
1589                         DEBUG_REQ(D_ERROR, req,
1590                                   "failed abort_req_reply; skipping");
1591                 }
1592                 target_exp_dequeue_req_replay(req);
1593                 target_request_copy_put(req);
1594         }
1595 }
1596
1597 static void abort_lock_replay_queue(struct obd_device *obd)
1598 {
1599         struct ptlrpc_request *req, *n;
1600         struct list_head abort_list;
1601
1602         INIT_LIST_HEAD(&abort_list);
1603         spin_lock(&obd->obd_recovery_task_lock);
1604         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1605         spin_unlock(&obd->obd_recovery_task_lock);
1606         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1607                 DEBUG_REQ(D_ERROR, req, "aborted:");
1608                 req->rq_status = -ENOTCONN;
1609                 if (ptlrpc_error(req)) {
1610                         DEBUG_REQ(D_ERROR, req,
1611                                   "failed abort_lock_reply; skipping");
1612                 }
1613                 target_request_copy_put(req);
1614         }
1615 }
1616
1617 /* Called from a cleanup function if the device is being cleaned up
1618    forcefully.  The exports should all have been disconnected already,
1619    the only thing left to do is
1620      - clear the recovery flags
1621      - cancel the timer
1622      - free queued requests and replies, but don't send replies
1623    Because the obd_stopping flag is set, no new requests should be received.
1624
1625 */
1626 void target_cleanup_recovery(struct obd_device *obd)
1627 {
1628         struct ptlrpc_request *req, *n;
1629         struct list_head clean_list;
1630
1631         INIT_LIST_HEAD(&clean_list);
1632         spin_lock(&obd->obd_dev_lock);
1633         if (!obd->obd_recovering) {
1634                 spin_unlock(&obd->obd_dev_lock);
1635                 EXIT;
1636                 return;
1637         }
1638         obd->obd_recovering = obd->obd_abort_recovery = 0;
1639         spin_unlock(&obd->obd_dev_lock);
1640
1641         spin_lock(&obd->obd_recovery_task_lock);
1642         target_cancel_recovery_timer(obd);
1643         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1644         spin_unlock(&obd->obd_recovery_task_lock);
1645
1646         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1647                 LASSERT(req->rq_reply_state == NULL);
1648                 target_exp_dequeue_req_replay(req);
1649                 target_request_copy_put(req);
1650         }
1651
1652         spin_lock(&obd->obd_recovery_task_lock);
1653         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1654         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1655         spin_unlock(&obd->obd_recovery_task_lock);
1656
1657         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1658                 LASSERT(req->rq_reply_state == NULL);
1659                 target_request_copy_put(req);
1660         }
1661
1662         EXIT;
1663 }
1664 EXPORT_SYMBOL(target_cleanup_recovery);
1665
1666 /* obd_recovery_task_lock should be held */
1667 void target_cancel_recovery_timer(struct obd_device *obd)
1668 {
1669         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1670         cfs_timer_disarm(&obd->obd_recovery_timer);
1671 }
1672
1673 static void target_start_recovery_timer(struct obd_device *obd)
1674 {
1675         if (obd->obd_recovery_start != 0)
1676                 return;
1677
1678         spin_lock(&obd->obd_dev_lock);
1679         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1680                 spin_unlock(&obd->obd_dev_lock);
1681                 return;
1682         }
1683
1684         LASSERT(obd->obd_recovery_timeout != 0);
1685
1686         if (obd->obd_recovery_start != 0) {
1687                 spin_unlock(&obd->obd_dev_lock);
1688                 return;
1689         }
1690
1691         cfs_timer_arm(&obd->obd_recovery_timer,
1692                       cfs_time_shift(obd->obd_recovery_timeout));
1693         obd->obd_recovery_start = cfs_time_current_sec();
1694         spin_unlock(&obd->obd_dev_lock);
1695
1696         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1697                       "or until %d client%s reconnect%s\n",
1698                       obd->obd_name,
1699                       obd->obd_recovery_timeout / 60,
1700                       obd->obd_recovery_timeout % 60,
1701                       obd->obd_max_recoverable_clients,
1702                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1703                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1704 }
1705
1706 /**
1707  * extend recovery window.
1708  *
1709  * if @extend is true, extend recovery window to have @drt remaining at least;
1710  * otherwise, make sure the recovery timeout value is not less than @drt.
1711  */
1712 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1713 {
1714         cfs_time_t now;
1715         cfs_time_t end;
1716         cfs_duration_t left;
1717         int to;
1718
1719         spin_lock(&obd->obd_dev_lock);
1720         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1721                 spin_unlock(&obd->obd_dev_lock);
1722                 return;
1723         }
1724         LASSERT(obd->obd_recovery_start != 0);
1725
1726         now  = cfs_time_current_sec();
1727         to   = obd->obd_recovery_timeout;
1728         end  = obd->obd_recovery_start + to;
1729         left = cfs_time_sub(end, now);
1730
1731         if (extend && (drt > left)) {
1732                 to += drt - left;
1733         } else if (!extend && (drt > to)) {
1734                 to = drt;
1735         }
1736
1737         if (to > obd->obd_recovery_time_hard)
1738                 to = obd->obd_recovery_time_hard;
1739         if (obd->obd_recovery_timeout < to) {
1740                 obd->obd_recovery_timeout = to;
1741                 end = obd->obd_recovery_start + to;
1742                 cfs_timer_arm(&obd->obd_recovery_timer,
1743                                 cfs_time_shift(end - now));
1744         }
1745         spin_unlock(&obd->obd_dev_lock);
1746
1747         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1748                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1749 }
1750
1751 /* Reset the timer with each new client connection */
1752 /*
1753  * This timer is actually reconnect_timer, which is for making sure
1754  * the total recovery window is at least as big as my reconnect
1755  * attempt timing. So the initial recovery time_out will be set to
1756  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1757  * from client is bigger than this, then the recovery time_out will
1758  * be extended to make sure the client could be reconnected, in the
1759  * process, the timeout from the new client should be ignored.
1760  */
1761
1762 static void
1763 check_and_start_recovery_timer(struct obd_device *obd,
1764                                struct ptlrpc_request *req,
1765                                int new_client)
1766 {
1767         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1768         struct obd_device_target *obt = &obd->u.obt;
1769
1770         if (!new_client && service_time)
1771                 /* Teach server about old server's estimates, as first guess
1772                  * at how long new requests will take. */
1773                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1774                             service_time);
1775
1776         target_start_recovery_timer(obd);
1777
1778         /* Convert the service time to RPC timeout,
1779          * and reuse service_time to limit stack usage. */
1780         service_time = at_est2timeout(service_time);
1781
1782         /* We expect other clients to timeout within service_time, then try
1783          * to reconnect, then try the failover server.  The max delay between
1784          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1785         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1786
1787         LASSERT(obt->obt_magic == OBT_MAGIC);
1788         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1789         if (service_time > obd->obd_recovery_timeout && !new_client)
1790                 extend_recovery_timer(obd, service_time, false);
1791 }
1792
1793 /** Health checking routines */
1794 static inline int exp_connect_healthy(struct obd_export *exp)
1795 {
1796         return (exp->exp_in_recovery);
1797 }
1798
1799 /** if export done req_replay or has replay in queue */
1800 static inline int exp_req_replay_healthy(struct obd_export *exp)
1801 {
1802         return (!exp->exp_req_replay_needed ||
1803                 atomic_read(&exp->exp_replay_count) > 0);
1804 }
1805
1806
1807 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1808 {
1809         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1810                exp_req_replay_healthy(exp);
1811 }
1812
1813 /** if export done lock_replay or has replay in queue */
1814 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1815 {
1816         return (!exp->exp_lock_replay_needed ||
1817                 atomic_read(&exp->exp_replay_count) > 0);
1818 }
1819
1820 static inline int exp_vbr_healthy(struct obd_export *exp)
1821 {
1822         return (!exp->exp_vbr_failed);
1823 }
1824
1825 static inline int exp_finished(struct obd_export *exp)
1826 {
1827         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1828 }
1829
1830 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1831 {
1832         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1833                 exp_finished(exp);
1834 }
1835
1836 static int check_for_next_transno(struct lu_target *lut)
1837 {
1838         struct ptlrpc_request *req = NULL;
1839         struct obd_device *obd = lut->lut_obd;
1840         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1841         int wake_up = 0, connected, completed, queue_len;
1842         __u64 req_transno = 0;
1843         __u64 update_transno = 0;
1844         __u64 next_transno = 0;
1845         ENTRY;
1846
1847         spin_lock(&obd->obd_recovery_task_lock);
1848         if (!list_empty(&obd->obd_req_replay_queue)) {
1849                 req = list_entry(obd->obd_req_replay_queue.next,
1850                                      struct ptlrpc_request, rq_list);
1851                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1852         }
1853
1854         if (tdtd != NULL)
1855                 update_transno = distribute_txn_get_next_transno(tdtd);
1856
1857         connected = atomic_read(&obd->obd_connected_clients);
1858         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1859         queue_len = obd->obd_requests_queued_for_recovery;
1860         next_transno = obd->obd_next_recovery_transno;
1861
1862         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1863                "req_transno: "LPU64", next_transno: "LPU64"\n",
1864                obd->obd_max_recoverable_clients, connected, completed,
1865                queue_len, req_transno, next_transno);
1866
1867         if (obd->obd_abort_recovery) {
1868                 CDEBUG(D_HA, "waking for aborted recovery\n");
1869                 wake_up = 1;
1870         } else if (obd->obd_recovery_expired) {
1871                 CDEBUG(D_HA, "waking for expired recovery\n");
1872                 wake_up = 1;
1873         } else if (tdtd != NULL && req != NULL &&
1874                    is_req_replayed_by_update(req)) {
1875                 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1876                          "next_transno"LPU64"\n", req_transno, next_transno);
1877                 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1878                        req_transno);
1879                 wake_up = 1;
1880         } else if (req_transno == next_transno ||
1881                    (update_transno != 0 && update_transno <= next_transno)) {
1882                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1883                 wake_up = 1;
1884         } else if (queue_len > 0 &&
1885                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1886                 int d_lvl = D_HA;
1887                 /** handle gaps occured due to lost reply or VBR */
1888                 LASSERTF(req_transno >= next_transno,
1889                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1890                          req_transno, next_transno);
1891                 if (req_transno > obd->obd_last_committed &&
1892                     !obd->obd_version_recov)
1893                         d_lvl = D_ERROR;
1894                 CDEBUG(d_lvl,
1895                        "%s: waking for gap in transno, VBR is %s (skip: "
1896                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1897                        ", next_update "LPD64" last_committed: "LPD64")\n",
1898                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1899                        next_transno, queue_len, completed, connected,
1900                        req_transno, update_transno, obd->obd_last_committed);
1901                 obd->obd_next_recovery_transno = req_transno;
1902                 wake_up = 1;
1903         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1904                 CDEBUG(D_HA, "waking for completed recovery\n");
1905                 wake_up = 1;
1906         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1907                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1908                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1909                 obd->obd_next_recovery_transno = req_transno;
1910                 wake_up = 1;
1911         }
1912         spin_unlock(&obd->obd_recovery_task_lock);
1913         return wake_up;
1914 }
1915
1916 static int check_for_next_lock(struct lu_target *lut)
1917 {
1918         struct obd_device *obd = lut->lut_obd;
1919         int wake_up = 0;
1920
1921         spin_lock(&obd->obd_recovery_task_lock);
1922         if (!list_empty(&obd->obd_lock_replay_queue)) {
1923                 CDEBUG(D_HA, "waking for next lock\n");
1924                 wake_up = 1;
1925         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1926                 CDEBUG(D_HA, "waking for completed lock replay\n");
1927                 wake_up = 1;
1928         } else if (obd->obd_abort_recovery) {
1929                 CDEBUG(D_HA, "waking for aborted recovery\n");
1930                 wake_up = 1;
1931         } else if (obd->obd_recovery_expired) {
1932                 CDEBUG(D_HA, "waking for expired recovery\n");
1933                 wake_up = 1;
1934         }
1935         spin_unlock(&obd->obd_recovery_task_lock);
1936
1937         return wake_up;
1938 }
1939
1940 /**
1941  * wait for recovery events,
1942  * check its status with help of check_routine
1943  * evict dead clients via health_check
1944  */
1945 static int target_recovery_overseer(struct lu_target *lut,
1946                                     int (*check_routine)(struct lu_target *),
1947                                     int (*health_check)(struct obd_export *))
1948 {
1949         struct obd_device       *obd = lut->lut_obd;
1950         struct target_distribute_txn_data *tdtd;
1951 repeat:
1952         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1953               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1954                 __u64 next_update_transno = 0;
1955
1956                 /* Only abort the recovery if there are no update recovery
1957                  * left in the queue */
1958                 spin_lock(&obd->obd_recovery_task_lock);
1959                 if (lut->lut_tdtd != NULL) {
1960                         next_update_transno =
1961                                 distribute_txn_get_next_transno(lut->lut_tdtd);
1962
1963                         tdtd = lut->lut_tdtd;
1964                         /* If next_update_transno == 0, it probably because
1965                          * updatelog retrieve threads did not get any records
1966                          * yet, let's wait those threads stopped */
1967                         if (next_update_transno == 0) {
1968                                 struct l_wait_info lwi = { 0 };
1969
1970                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
1971                                        atomic_read(
1972                                        &tdtd->tdtd_recovery_threads_count) == 0,
1973                                        &lwi);
1974
1975                                 next_update_transno =
1976                                         distribute_txn_get_next_transno(
1977                                                                 lut->lut_tdtd);
1978                         }
1979                 }
1980
1981                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
1982                         obd->obd_next_recovery_transno = next_update_transno;
1983                         spin_unlock(&obd->obd_recovery_task_lock);
1984                         /* Disconnect unfinished exports from clients, and
1985                          * keep connection from MDT to make sure the update
1986                          * recovery will still keep trying until some one
1987                          * manually abort the recovery */
1988                         class_disconnect_stale_exports(obd,
1989                                                 exp_finished_or_from_mdt);
1990                         /* Abort all of replay and replay lock req from
1991                          * clients */
1992                         abort_req_replay_queue(obd);
1993                         abort_lock_replay_queue(obd);
1994                         CDEBUG(D_HA, "%s: there are still update replay ("LPX64
1995                                ")in the queue.\n", obd->obd_name,
1996                                next_update_transno);
1997                 } else {
1998                         obd->obd_abort_recovery = 1;
1999                         spin_unlock(&obd->obd_recovery_task_lock);
2000                         CWARN("%s recovery is aborted by hard timeout\n",
2001                               obd->obd_name);
2002                 }
2003         }
2004
2005         while (wait_event_timeout(obd->obd_next_transno_waitq,
2006                                   check_routine(lut),
2007                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2008                 /* wait indefinitely for event, but don't trigger watchdog */;
2009
2010         if (obd->obd_abort_recovery) {
2011                 CWARN("recovery is aborted, evict exports in recovery\n");
2012                 if (lut->lut_tdtd != NULL) {
2013                         struct l_wait_info lwi = { 0 };
2014
2015                         tdtd = lut->lut_tdtd;
2016                         /* Let's wait all of the update log recovery thread
2017                          * finished */
2018                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2019                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2020                              &lwi);
2021                         /* Then abort the update recovery list */
2022                         dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2023                         dtrq_list_destroy(lut->lut_tdtd);
2024                 }
2025
2026                 /** evict exports which didn't finish recovery yet */
2027                 class_disconnect_stale_exports(obd, exp_finished);
2028                 return 1;
2029         } else if (obd->obd_recovery_expired) {
2030                 obd->obd_recovery_expired = 0;
2031                 /** If some clients died being recovered, evict them */
2032                 LCONSOLE_WARN("%s: recovery is timed out, "
2033                               "evict stale exports\n", obd->obd_name);
2034                 /** evict cexports with no replay in queue, they are stalled */
2035                 class_disconnect_stale_exports(obd, health_check);
2036
2037                 /** continue with VBR */
2038                 spin_lock(&obd->obd_dev_lock);
2039                 obd->obd_version_recov = 1;
2040                 spin_unlock(&obd->obd_dev_lock);
2041                 /**
2042                  * reset timer, recovery will proceed with versions now,
2043                  * timeout is set just to handle reconnection delays
2044                  */
2045                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2046                 /** Wait for recovery events again, after evicting bad clients */
2047                 goto repeat;
2048         }
2049         return 0;
2050 }
2051
2052 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2053 {
2054         struct obd_device       *obd = lut->lut_obd;
2055         struct ptlrpc_request *req = NULL;
2056
2057         CDEBUG(D_HA, "Waiting for lock\n");
2058         if (target_recovery_overseer(lut, check_for_next_lock,
2059                                      exp_lock_replay_healthy))
2060                 abort_lock_replay_queue(obd);
2061
2062         spin_lock(&obd->obd_recovery_task_lock);
2063         if (!list_empty(&obd->obd_lock_replay_queue)) {
2064                 req = list_entry(obd->obd_lock_replay_queue.next,
2065                                      struct ptlrpc_request, rq_list);
2066                 list_del_init(&req->rq_list);
2067                 spin_unlock(&obd->obd_recovery_task_lock);
2068         } else {
2069                 spin_unlock(&obd->obd_recovery_task_lock);
2070                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2071                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2072                 /** evict exports failed VBR */
2073                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2074         }
2075         return req;
2076 }
2077
2078 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2079 {
2080         struct ptlrpc_request *req = NULL;
2081
2082         spin_lock(&obd->obd_recovery_task_lock);
2083         if (!list_empty(&obd->obd_final_req_queue)) {
2084                 req = list_entry(obd->obd_final_req_queue.next,
2085                                      struct ptlrpc_request, rq_list);
2086                 list_del_init(&req->rq_list);
2087                 spin_unlock(&obd->obd_recovery_task_lock);
2088                 if (req->rq_export->exp_in_recovery) {
2089                         spin_lock(&req->rq_export->exp_lock);
2090                         req->rq_export->exp_in_recovery = 0;
2091                         spin_unlock(&req->rq_export->exp_lock);
2092                 }
2093         } else {
2094                 spin_unlock(&obd->obd_recovery_task_lock);
2095         }
2096         return req;
2097 }
2098
2099 static void handle_recovery_req(struct ptlrpc_thread *thread,
2100                                 struct ptlrpc_request *req,
2101                                 svc_handler_t handler)
2102 {
2103         ENTRY;
2104
2105         /**
2106          * export can be evicted during recovery, no need to handle replays for
2107          * it after that, discard such request silently
2108          */
2109         if (req->rq_export->exp_disconnected)
2110                 RETURN_EXIT;
2111
2112         req->rq_session.lc_thread = thread;
2113         req->rq_svc_thread = thread;
2114         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2115
2116         /* thread context */
2117         lu_context_enter(&thread->t_env->le_ctx);
2118         (void)handler(req);
2119         lu_context_exit(&thread->t_env->le_ctx);
2120
2121         /* don't reset timer for final stage */
2122         if (!exp_finished(req->rq_export)) {
2123                 int to = obd_timeout;
2124
2125                 /**
2126                  * Add request timeout to the recovery time so next request from
2127                  * this client may come in recovery time
2128                  */
2129                 if (!AT_OFF) {
2130                         struct ptlrpc_service_part *svcpt;
2131
2132                         svcpt = req->rq_rqbd->rqbd_svcpt;
2133                         /* If the server sent early reply for this request,
2134                          * the client will recalculate the timeout according to
2135                          * current server estimate service time, so we will
2136                          * use the maxium timeout here for waiting the client
2137                          * sending the next req */
2138                         to = max((int)at_est2timeout(
2139                                  at_get(&svcpt->scp_at_estimate)),
2140                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2141                         /* Add 2 net_latency, one for balance rq_deadline
2142                          * (see ptl_send_rpc), one for resend the req to server,
2143                          * Note: client will pack net_latency in replay req
2144                          * (see ptlrpc_replay_req) */
2145                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2146                 }
2147                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2148         }
2149         EXIT;
2150 }
2151
2152 /** Checking routines for recovery */
2153 static int check_for_recovery_ready(struct lu_target *lut)
2154 {
2155         struct obd_device *obd = lut->lut_obd;
2156         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2157
2158         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2159                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2160                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2161                obd->obd_recovery_expired);
2162
2163         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2164                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2165                 if (clnts + obd->obd_stale_clients <
2166                     obd->obd_max_recoverable_clients)
2167                         return 0;
2168         }
2169
2170         if (lut->lut_tdtd != NULL) {
2171                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2172                     !obd->obd_abort_recovery) {
2173                         /* Let's extend recovery timer, in case the recovery
2174                          * timer expired, and some clients got evicted */
2175                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2176                                               true);
2177                         CDEBUG(D_HA, "%s update recovery is not ready,"
2178                                " extend recovery %d\n", obd->obd_name,
2179                                obd->obd_recovery_timeout);
2180                         return 0;
2181                 } else {
2182                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2183                 }
2184         }
2185
2186         return 1;
2187 }
2188
2189 enum {
2190         REQUEST_RECOVERY = 1,
2191         UPDATE_RECOVERY = 2,
2192 };
2193
2194 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2195 {
2196         __u64 transno = 0;
2197
2198         if (!list_empty(&obd->obd_req_replay_queue)) {
2199                 struct ptlrpc_request *req;
2200
2201                 req = list_entry(obd->obd_req_replay_queue.next,
2202                                  struct ptlrpc_request, rq_list);
2203                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2204         }
2205
2206         return transno;
2207 }
2208
2209 static __u64 get_next_transno(struct lu_target *lut, int *type)
2210 {
2211         struct obd_device *obd = lut->lut_obd;
2212         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2213         __u64 transno = 0;
2214         __u64 update_transno;
2215         ENTRY;
2216
2217         transno = get_next_replay_req_transno(obd);
2218         if (type != NULL)
2219                 *type = REQUEST_RECOVERY;
2220
2221         if (tdtd == NULL)
2222                 RETURN(transno);
2223
2224         update_transno = distribute_txn_get_next_transno(tdtd);
2225         if (transno == 0 || (transno >= update_transno &&
2226                              update_transno != 0)) {
2227                 transno = update_transno;
2228                 if (type != NULL)
2229                         *type = UPDATE_RECOVERY;
2230         }
2231
2232         RETURN(transno);
2233 }
2234
2235 /**
2236  * drop duplicate replay request
2237  *
2238  * Because the operation has been replayed by update recovery, the request
2239  * with the same transno will be dropped and also notify the client to send
2240  * next replay request.
2241  *
2242  * \param[in] env       execution environment
2243  * \param[in] obd       failover obd device
2244  * \param[in] req       request to be dropped
2245  */
2246 static void drop_duplicate_replay_req(struct lu_env *env,
2247                                       struct obd_device *obd,
2248                                       struct ptlrpc_request *req)
2249 {
2250         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2251                   " update records are found.\n",
2252                   lustre_msg_get_transno(req->rq_reqmsg),
2253                   libcfs_nid2str(req->rq_peer.nid));
2254
2255         /* Right now, only for MDS reint operation update replay and
2256          * normal request replay can have the same transno */
2257         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2258                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2259                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2260                 if (likely(req->rq_export))
2261                         target_committed_to_req(req);
2262                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2263                 target_send_reply(req, req->rq_status, 0);
2264         } else {
2265                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2266                 libcfs_nid2str(req->rq_peer.nid));
2267         }
2268         target_exp_dequeue_req_replay(req);
2269         target_request_copy_put(req);
2270         obd->obd_replayed_requests++;
2271 }
2272
2273 static void replay_request_or_update(struct lu_env *env,
2274                                      struct lu_target *lut,
2275                                      struct target_recovery_data *trd,
2276                                      struct ptlrpc_thread *thread)
2277 {
2278         struct obd_device *obd = lut->lut_obd;
2279         struct ptlrpc_request *req = NULL;
2280         int                     type;
2281         __u64                   transno;
2282         ENTRY;
2283
2284         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2285                obd->obd_next_recovery_transno);
2286
2287         /* Replay all of request and update by transno */
2288         do {
2289                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2290
2291                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2292
2293                 /** It is needed to extend recovery window above
2294                  *  recovery_time_soft. Extending is possible only in the
2295                  *  end of recovery window (see more details in
2296                  *  handle_recovery_req()).
2297                  */
2298                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2299
2300                 if (target_recovery_overseer(lut, check_for_next_transno,
2301                                         exp_req_replay_healthy_or_from_mdt)) {
2302                         abort_req_replay_queue(obd);
2303                         abort_lock_replay_queue(obd);
2304                         goto abort;
2305                 }
2306
2307                 spin_lock(&obd->obd_recovery_task_lock);
2308                 transno = get_next_transno(lut, &type);
2309                 if (type == REQUEST_RECOVERY && transno != 0) {
2310                         /* Drop replay request from client side, if the
2311                          * replay has been executed by update with the
2312                          * same transno */
2313                         req = list_entry(obd->obd_req_replay_queue.next,
2314                                         struct ptlrpc_request, rq_list);
2315
2316                         list_del_init(&req->rq_list);
2317                         obd->obd_requests_queued_for_recovery--;
2318                         spin_unlock(&obd->obd_recovery_task_lock);
2319
2320                         /* Let's check if the request has been redone by
2321                          * update replay */
2322                         if (is_req_replayed_by_update(req)) {
2323                                 struct distribute_txn_replay_req *dtrq;
2324
2325                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2326                                                                    req->rq_xid);
2327                                 LASSERT(dtrq != NULL);
2328                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2329                                 list_del_init(&dtrq->dtrq_list);
2330                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2331                                 dtrq_destroy(dtrq);
2332
2333                                 drop_duplicate_replay_req(env, obd, req);
2334
2335                                 continue;
2336                         }
2337
2338                         LASSERT(trd->trd_processing_task == current_pid());
2339                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2340                                   lustre_msg_get_transno(req->rq_reqmsg),
2341                                   libcfs_nid2str(req->rq_peer.nid));
2342
2343                         handle_recovery_req(thread, req,
2344                                             trd->trd_recovery_handler);
2345                         /**
2346                          * bz18031: increase next_recovery_transno before
2347                          * target_request_copy_put() will drop exp_rpc reference
2348                          */
2349                         spin_lock(&obd->obd_recovery_task_lock);
2350                         obd->obd_next_recovery_transno++;
2351                         spin_unlock(&obd->obd_recovery_task_lock);
2352                         target_exp_dequeue_req_replay(req);
2353                         target_request_copy_put(req);
2354                         obd->obd_replayed_requests++;
2355                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2356                         struct distribute_txn_replay_req *dtrq;
2357                         int rc;
2358
2359                         spin_unlock(&obd->obd_recovery_task_lock);
2360
2361                         LASSERT(tdtd != NULL);
2362                         dtrq = distribute_txn_get_next_req(tdtd);
2363                         lu_context_enter(&thread->t_env->le_ctx);
2364                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2365                         lu_context_exit(&thread->t_env->le_ctx);
2366                         extend_recovery_timer(obd, obd_timeout, true);
2367
2368                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2369                                 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2370                                        " to finish list\n", dtrq->dtrq_xid,
2371                                        dtrq->dtrq_master_transno);
2372
2373                                 /* Add it to the replay finish list */
2374                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2375                                 list_add(&dtrq->dtrq_list,
2376                                          &tdtd->tdtd_replay_finish_list);
2377                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2378
2379                                 spin_lock(&obd->obd_recovery_task_lock);
2380                                 if (transno == obd->obd_next_recovery_transno)
2381                                         obd->obd_next_recovery_transno++;
2382                                 else if (transno >
2383                                          obd->obd_next_recovery_transno)
2384                                         obd->obd_next_recovery_transno =
2385                                                                 transno + 1;
2386                                 spin_unlock(&obd->obd_recovery_task_lock);
2387                         } else {
2388                                 dtrq_destroy(dtrq);
2389                         }
2390                 } else {
2391                         spin_unlock(&obd->obd_recovery_task_lock);
2392 abort:
2393                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2394                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2395                         /** evict exports failed VBR */
2396                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2397                         break;
2398                 }
2399         } while (1);
2400 }
2401
2402 static int target_recovery_thread(void *arg)
2403 {
2404         struct lu_target *lut = arg;
2405         struct obd_device *obd = lut->lut_obd;
2406         struct ptlrpc_request *req;
2407         struct target_recovery_data *trd = &obd->obd_recovery_data;
2408         unsigned long delta;
2409         struct lu_env *env;
2410         struct ptlrpc_thread *thread = NULL;
2411         int rc = 0;
2412         ENTRY;
2413
2414         unshare_fs_struct();
2415         OBD_ALLOC_PTR(thread);
2416         if (thread == NULL)
2417                 RETURN(-ENOMEM);
2418
2419         OBD_ALLOC_PTR(env);
2420         if (env == NULL) {
2421                 OBD_FREE_PTR(thread);
2422                 RETURN(-ENOMEM);
2423         }
2424
2425         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2426         if (rc) {
2427                 OBD_FREE_PTR(thread);
2428                 OBD_FREE_PTR(env);
2429                 RETURN(rc);
2430         }
2431
2432         thread->t_env = env;
2433         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2434         env->le_ctx.lc_thread = thread;
2435         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2436         thread->t_watchdog = NULL;
2437
2438         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2439                current_pid());
2440         trd->trd_processing_task = current_pid();
2441
2442         spin_lock(&obd->obd_dev_lock);
2443         obd->obd_recovering = 1;
2444         spin_unlock(&obd->obd_dev_lock);
2445         complete(&trd->trd_starting);
2446
2447         /* first of all, we have to know the first transno to replay */
2448         if (target_recovery_overseer(lut, check_for_recovery_ready,
2449                                      exp_connect_healthy)) {
2450                 abort_req_replay_queue(obd);
2451                 abort_lock_replay_queue(obd);
2452                 if (lut->lut_tdtd != NULL)
2453                         dtrq_list_destroy(lut->lut_tdtd);
2454         }
2455
2456         /* next stage: replay requests or update */
2457         delta = jiffies;
2458         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2459                atomic_read(&obd->obd_req_replay_clients),
2460                obd->obd_next_recovery_transno);
2461         replay_request_or_update(env, lut, trd, thread);
2462
2463         /**
2464          * The second stage: replay locks
2465          */
2466         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2467                atomic_read(&obd->obd_lock_replay_clients));
2468         while ((req = target_next_replay_lock(lut))) {
2469                 LASSERT(trd->trd_processing_task == current_pid());
2470                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2471                           libcfs_nid2str(req->rq_peer.nid));
2472                 handle_recovery_req(thread, req,
2473                                     trd->trd_recovery_handler);
2474                 target_request_copy_put(req);
2475                 obd->obd_replayed_locks++;
2476         }
2477
2478         /**
2479          * The third stage: reply on final pings, at this moment all clients
2480          * must have request in final queue
2481          */
2482         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2483         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2484         /** Update server last boot epoch */
2485         tgt_boot_epoch_update(lut);
2486         /* We drop recoverying flag to forward all new requests
2487          * to regular mds_handle() since now */
2488         spin_lock(&obd->obd_dev_lock);
2489         obd->obd_recovering = obd->obd_abort_recovery = 0;
2490         spin_unlock(&obd->obd_dev_lock);
2491         spin_lock(&obd->obd_recovery_task_lock);
2492         target_cancel_recovery_timer(obd);
2493         spin_unlock(&obd->obd_recovery_task_lock);
2494         while ((req = target_next_final_ping(obd))) {
2495                 LASSERT(trd->trd_processing_task == current_pid());
2496                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2497                           libcfs_nid2str(req->rq_peer.nid));
2498                 handle_recovery_req(thread, req,
2499                                     trd->trd_recovery_handler);
2500                 /* Because the waiting client can not send ping to server,
2501                  * so we need refresh the last_request_time, to avoid the
2502                  * export is being evicted */
2503                 ptlrpc_update_export_timer(req->rq_export, 0);
2504                 target_request_copy_put(req);
2505         }
2506
2507         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2508         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2509               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2510         if (delta > OBD_RECOVERY_TIME_SOFT) {
2511                 CWARN("too long recovery - read logs\n");
2512                 libcfs_debug_dumplog();
2513         }
2514
2515         target_finish_recovery(lut);
2516
2517         lu_context_fini(&env->le_ctx);
2518         trd->trd_processing_task = 0;
2519         complete(&trd->trd_finishing);
2520
2521         tgt_io_thread_done(thread);
2522         OBD_FREE_PTR(thread);
2523         OBD_FREE_PTR(env);
2524         RETURN(rc);
2525 }
2526
2527 static int target_start_recovery_thread(struct lu_target *lut,
2528                                         svc_handler_t handler)
2529 {
2530         struct obd_device *obd = lut->lut_obd;
2531         int rc = 0;
2532         struct target_recovery_data *trd = &obd->obd_recovery_data;
2533         int index;
2534
2535         memset(trd, 0, sizeof(*trd));
2536         init_completion(&trd->trd_starting);
2537         init_completion(&trd->trd_finishing);
2538         trd->trd_recovery_handler = handler;
2539
2540         rc = server_name2index(obd->obd_name, &index, NULL);
2541         if (rc < 0)
2542                 return rc;
2543
2544         if (!IS_ERR(kthread_run(target_recovery_thread,
2545                                 lut, "tgt_recover_%d", index))) {
2546                 wait_for_completion(&trd->trd_starting);
2547                 LASSERT(obd->obd_recovering != 0);
2548         } else {
2549                 rc = -ECHILD;
2550         }
2551
2552         return rc;
2553 }
2554
2555 void target_stop_recovery_thread(struct obd_device *obd)
2556 {
2557         if (obd->obd_recovery_data.trd_processing_task > 0) {
2558                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2559                 /** recovery can be done but postrecovery is not yet */
2560                 spin_lock(&obd->obd_dev_lock);
2561                 if (obd->obd_recovering) {
2562                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2563                         obd->obd_abort_recovery = 1;
2564                         wake_up(&obd->obd_next_transno_waitq);
2565                 }
2566                 spin_unlock(&obd->obd_dev_lock);
2567                 wait_for_completion(&trd->trd_finishing);
2568         }
2569 }
2570 EXPORT_SYMBOL(target_stop_recovery_thread);
2571
2572 void target_recovery_fini(struct obd_device *obd)
2573 {
2574         class_disconnect_exports(obd);
2575         target_stop_recovery_thread(obd);
2576         target_cleanup_recovery(obd);
2577 }
2578 EXPORT_SYMBOL(target_recovery_fini);
2579
2580 static void target_recovery_expired(unsigned long castmeharder)
2581 {
2582         struct obd_device *obd = (struct obd_device *)castmeharder;
2583         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2584                " after %lds (%d clients connected)\n",
2585                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2586                cfs_time_current_sec()- obd->obd_recovery_start,
2587                atomic_read(&obd->obd_connected_clients));
2588
2589         obd->obd_recovery_expired = 1;
2590         wake_up(&obd->obd_next_transno_waitq);
2591 }
2592
2593 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2594 {
2595         struct obd_device *obd = lut->lut_obd;
2596
2597         if (obd->obd_max_recoverable_clients == 0) {
2598                 /** Update server last boot epoch */
2599                 tgt_boot_epoch_update(lut);
2600                 return;
2601         }
2602
2603         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2604                "last_transno "LPU64"\n", obd->obd_name,
2605                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2606         LASSERT(obd->obd_stopping == 0);
2607         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2608         obd->obd_recovery_start = 0;
2609         obd->obd_recovery_end = 0;
2610
2611         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2612         target_start_recovery_thread(lut, handler);
2613 }
2614 EXPORT_SYMBOL(target_recovery_init);
2615
2616 static int target_process_req_flags(struct obd_device *obd,
2617                                     struct ptlrpc_request *req)
2618 {
2619         struct obd_export *exp = req->rq_export;
2620         LASSERT(exp != NULL);
2621         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2622                 /* client declares he's ready to replay locks */
2623                 spin_lock(&exp->exp_lock);
2624                 if (exp->exp_req_replay_needed) {
2625                         exp->exp_req_replay_needed = 0;
2626                         spin_unlock(&exp->exp_lock);
2627
2628                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2629                         atomic_dec(&obd->obd_req_replay_clients);
2630                 } else {
2631                         spin_unlock(&exp->exp_lock);
2632                 }
2633         }
2634         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2635                 /* client declares he's ready to complete recovery
2636                  * so, we put the request on th final queue */
2637                 spin_lock(&exp->exp_lock);
2638                 if (exp->exp_lock_replay_needed) {
2639                         exp->exp_lock_replay_needed = 0;
2640                         spin_unlock(&exp->exp_lock);
2641
2642                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2643                         atomic_dec(&obd->obd_lock_replay_clients);
2644                 } else {
2645                         spin_unlock(&exp->exp_lock);
2646                 }
2647         }
2648         return 0;
2649 }
2650
2651 int target_queue_recovery_request(struct ptlrpc_request *req,
2652                                   struct obd_device *obd)
2653 {
2654         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2655         struct ptlrpc_request *reqiter;
2656         int inserted = 0;
2657         ENTRY;
2658
2659         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2660                 /* Processing the queue right now, don't re-add. */
2661                 RETURN(1);
2662         }
2663
2664         target_process_req_flags(obd, req);
2665
2666         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2667                 /* client declares he's ready to complete recovery
2668                  * so, we put the request on th final queue */
2669                 target_request_copy_get(req);
2670                 DEBUG_REQ(D_HA, req, "queue final req");
2671                 wake_up(&obd->obd_next_transno_waitq);
2672                 spin_lock(&obd->obd_recovery_task_lock);
2673                 if (obd->obd_recovering) {
2674                         list_add_tail(&req->rq_list,
2675                                           &obd->obd_final_req_queue);
2676                 } else {
2677                         spin_unlock(&obd->obd_recovery_task_lock);
2678                         target_request_copy_put(req);
2679                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2680                 }
2681                 spin_unlock(&obd->obd_recovery_task_lock);
2682                 RETURN(0);
2683         }
2684         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2685                 /* client declares he's ready to replay locks */
2686                 target_request_copy_get(req);
2687                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2688                 wake_up(&obd->obd_next_transno_waitq);
2689                 spin_lock(&obd->obd_recovery_task_lock);
2690                 LASSERT(obd->obd_recovering);
2691                 /* usually due to recovery abort */
2692                 if (!req->rq_export->exp_in_recovery) {
2693                         spin_unlock(&obd->obd_recovery_task_lock);
2694                         target_request_copy_put(req);
2695                         RETURN(-ENOTCONN);
2696                 }
2697                 LASSERT(req->rq_export->exp_lock_replay_needed);
2698                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2699                 spin_unlock(&obd->obd_recovery_task_lock);
2700                 RETURN(0);
2701         }
2702
2703         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2704          * (i.e. buflens etc are in my own byte order), but type-dependent
2705          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2706
2707         if (!transno) {
2708                 INIT_LIST_HEAD(&req->rq_list);
2709                 DEBUG_REQ(D_HA, req, "not queueing");
2710                 RETURN(1);
2711         }
2712
2713         /* If we're processing the queue, we want don't want to queue this
2714          * message.
2715          *
2716          * Also, if this request has a transno less than the one we're waiting
2717          * for, we should process it now.  It could (and currently always will)
2718          * be an open request for a descriptor that was opened some time ago.
2719          *
2720          * Also, a resent, replayed request that has already been
2721          * handled will pass through here and be processed immediately.
2722          */
2723         CDEBUG(D_HA, "Next recovery transno: "LPU64
2724                ", current: "LPU64", replaying\n",
2725                obd->obd_next_recovery_transno, transno);
2726
2727         /* If the request has been replayed by update replay, then sends this
2728          * request to the recovery thread (replay_request_or_update()), where
2729          * it will be handled */
2730         spin_lock(&obd->obd_recovery_task_lock);
2731         if (transno < obd->obd_next_recovery_transno &&
2732             !is_req_replayed_by_update(req)) {
2733                 /* Processing the queue right now, don't re-add. */
2734                 LASSERT(list_empty(&req->rq_list));
2735                 spin_unlock(&obd->obd_recovery_task_lock);
2736                 RETURN(1);
2737         }
2738         spin_unlock(&obd->obd_recovery_task_lock);
2739
2740         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2741                 RETURN(0);
2742
2743         target_request_copy_get(req);
2744         if (!req->rq_export->exp_in_recovery) {
2745                 target_request_copy_put(req);
2746                 RETURN(-ENOTCONN);
2747         }
2748         LASSERT(req->rq_export->exp_req_replay_needed);
2749
2750         if (target_exp_enqueue_req_replay(req)) {
2751                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2752                 target_request_copy_put(req);
2753                 RETURN(0);
2754         }
2755
2756         /* XXX O(n^2) */
2757         spin_lock(&obd->obd_recovery_task_lock);
2758         LASSERT(obd->obd_recovering);
2759         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2760                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2761                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2762                         inserted = 1;
2763                         goto added;
2764                 }
2765
2766                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2767                              transno)) {
2768                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2769                                   "has been claimed by another client");
2770                         spin_unlock(&obd->obd_recovery_task_lock);
2771                         target_exp_dequeue_req_replay(req);
2772                         target_request_copy_put(req);
2773                         RETURN(0);
2774                 }
2775         }
2776 added:
2777         if (!inserted)
2778                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2779
2780         obd->obd_requests_queued_for_recovery++;
2781         spin_unlock(&obd->obd_recovery_task_lock);
2782         wake_up(&obd->obd_next_transno_waitq);
2783         RETURN(0);
2784 }
2785
2786 int target_handle_ping(struct ptlrpc_request *req)
2787 {
2788         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2789         return req_capsule_server_pack(&req->rq_pill);
2790 }
2791
2792 void target_committed_to_req(struct ptlrpc_request *req)
2793 {
2794         struct obd_export *exp = req->rq_export;
2795
2796         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2797                 lustre_msg_set_last_committed(req->rq_repmsg,
2798                                               exp->exp_last_committed);
2799         else
2800                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2801                           "%d)", exp->exp_obd->obd_no_transno,
2802                           req->rq_repmsg == NULL);
2803
2804         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2805                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2806 }
2807
2808 #endif /* HAVE_SERVER_SUPPORT */
2809
2810 /**
2811  * Packs current SLV and Limit into \a req.
2812  */
2813 int target_pack_pool_reply(struct ptlrpc_request *req)
2814 {
2815         struct obd_device *obd;
2816         ENTRY;
2817
2818         /* Check that we still have all structures alive as this may
2819          * be some late RPC at shutdown time. */
2820         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2821                      !exp_connect_lru_resize(req->rq_export))) {
2822                 lustre_msg_set_slv(req->rq_repmsg, 0);
2823                 lustre_msg_set_limit(req->rq_repmsg, 0);
2824                 RETURN(0);
2825         }
2826
2827         /* OBD is alive here as export is alive, which we checked above. */
2828         obd = req->rq_export->exp_obd;
2829
2830         read_lock(&obd->obd_pool_lock);
2831         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2832         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2833         read_unlock(&obd->obd_pool_lock);
2834
2835         RETURN(0);
2836 }
2837
2838 static int target_send_reply_msg(struct ptlrpc_request *req,
2839                                  int rc, int fail_id)
2840 {
2841         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2842                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2843                 return -ECOMM;
2844         }
2845         /* We can have a null rq_reqmsg in the event of bad signature or
2846          * no context when unwrapping */
2847         if (req->rq_reqmsg &&
2848             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2849             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2850                 return -ECOMM;
2851
2852         if (unlikely(rc)) {
2853                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2854                 req->rq_status = rc;
2855                 return ptlrpc_send_error(req, 1);
2856         } else {
2857                 DEBUG_REQ(D_NET, req, "sending reply");
2858         }
2859
2860         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2861 }
2862
2863 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2864 {
2865         struct ptlrpc_service_part *svcpt;
2866         int                        netrc;
2867         struct ptlrpc_reply_state *rs;
2868         struct obd_export         *exp;
2869         ENTRY;
2870
2871         if (req->rq_no_reply) {
2872                 EXIT;
2873                 return;
2874         }
2875
2876         svcpt = req->rq_rqbd->rqbd_svcpt;
2877         rs = req->rq_reply_state;
2878         if (rs == NULL || !rs->rs_difficult) {
2879                 /* no notifiers */
2880                 target_send_reply_msg (req, rc, fail_id);
2881                 EXIT;
2882                 return;
2883         }
2884
2885         /* must be an export if locks saved */
2886         LASSERT(req->rq_export != NULL);
2887         /* req/reply consistent */
2888         LASSERT(rs->rs_svcpt == svcpt);
2889
2890         /* "fresh" reply */
2891         LASSERT(!rs->rs_scheduled);
2892         LASSERT(!rs->rs_scheduled_ever);
2893         LASSERT(!rs->rs_handled);
2894         LASSERT(!rs->rs_on_net);
2895         LASSERT(rs->rs_export == NULL);
2896         LASSERT(list_empty(&rs->rs_obd_list));
2897         LASSERT(list_empty(&rs->rs_exp_list));
2898
2899         exp = class_export_get(req->rq_export);
2900
2901         /* disable reply scheduling while I'm setting up */
2902         rs->rs_scheduled = 1;
2903         rs->rs_on_net    = 1;
2904         rs->rs_xid       = req->rq_xid;
2905         rs->rs_transno   = req->rq_transno;
2906         rs->rs_export    = exp;
2907         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2908
2909         spin_lock(&exp->exp_uncommitted_replies_lock);
2910         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2911                rs->rs_transno, exp->exp_last_committed);
2912         if (rs->rs_transno > exp->exp_last_committed) {
2913                 /* not committed already */
2914                 list_add_tail(&rs->rs_obd_list,
2915                                   &exp->exp_uncommitted_replies);
2916         }
2917         spin_unlock(&exp->exp_uncommitted_replies_lock);
2918
2919         spin_lock(&exp->exp_lock);
2920         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2921         spin_unlock(&exp->exp_lock);
2922
2923         netrc = target_send_reply_msg(req, rc, fail_id);
2924
2925         spin_lock(&svcpt->scp_rep_lock);
2926
2927         atomic_inc(&svcpt->scp_nreps_difficult);
2928
2929         if (netrc != 0) {
2930                 /* error sending: reply is off the net.  Also we need +1
2931                  * reply ref until ptlrpc_handle_rs() is done
2932                  * with the reply state (if the send was successful, there
2933                  * would have been +1 ref for the net, which
2934                  * reply_out_callback leaves alone) */
2935                 rs->rs_on_net = 0;
2936                 ptlrpc_rs_addref(rs);
2937         }
2938
2939         spin_lock(&rs->rs_lock);
2940         if (rs->rs_transno <= exp->exp_last_committed ||
2941             (!rs->rs_on_net && !rs->rs_no_ack) ||
2942             list_empty(&rs->rs_exp_list) ||     /* completed already */
2943             list_empty(&rs->rs_obd_list)) {
2944                 CDEBUG(D_HA, "Schedule reply immediately\n");
2945                 ptlrpc_dispatch_difficult_reply(rs);
2946         } else {
2947                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2948                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2949         }
2950         spin_unlock(&rs->rs_lock);
2951         spin_unlock(&svcpt->scp_rep_lock);
2952         EXIT;
2953 }
2954
2955 enum ldlm_mode lck_compat_array[] = {
2956         [LCK_EX]    = LCK_COMPAT_EX,
2957         [LCK_PW]    = LCK_COMPAT_PW,
2958         [LCK_PR]    = LCK_COMPAT_PR,
2959         [LCK_CW]    = LCK_COMPAT_CW,
2960         [LCK_CR]    = LCK_COMPAT_CR,
2961         [LCK_NL]    = LCK_COMPAT_NL,
2962         [LCK_GROUP] = LCK_COMPAT_GROUP,
2963         [LCK_COS]   = LCK_COMPAT_COS,
2964 };
2965
2966 /**
2967  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2968  * not escape to the user level.
2969  */
2970 int ldlm_error2errno(enum ldlm_error error)
2971 {
2972         int result;
2973
2974         switch (error) {
2975         case ELDLM_OK:
2976         case ELDLM_LOCK_MATCHED:
2977                 result = 0;
2978                 break;
2979         case ELDLM_LOCK_CHANGED:
2980                 result = -ESTALE;
2981                 break;
2982         case ELDLM_LOCK_ABORTED:
2983                 result = -ENAVAIL;
2984                 break;
2985         case ELDLM_LOCK_REPLACED:
2986                 result = -ESRCH;
2987                 break;
2988         case ELDLM_NO_LOCK_DATA:
2989                 result = -ENOENT;
2990                 break;
2991         case ELDLM_NAMESPACE_EXISTS:
2992                 result = -EEXIST;
2993                 break;
2994         case ELDLM_BAD_NAMESPACE:
2995                 result = -EBADF;
2996                 break;
2997         default:
2998                 if (((int)error) < 0) { /* cast to signed type */
2999                         result = error; /* as ldlm_error can be unsigned */
3000                 } else {
3001                         CERROR("Invalid DLM result code: %d\n", error);
3002                         result = -EPROTO;
3003                 }
3004         }
3005         return result;
3006 }
3007 EXPORT_SYMBOL(ldlm_error2errno);
3008
3009 /**
3010  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3011  */
3012 enum ldlm_error ldlm_errno2error(int err_no)
3013 {
3014         int error;
3015
3016         switch (err_no) {
3017         case 0:
3018                 error = ELDLM_OK;
3019                 break;
3020         case -ESTALE:
3021                 error = ELDLM_LOCK_CHANGED;
3022                 break;
3023         case -ENAVAIL:
3024                 error = ELDLM_LOCK_ABORTED;
3025                 break;
3026         case -ESRCH:
3027                 error = ELDLM_LOCK_REPLACED;
3028                 break;
3029         case -ENOENT:
3030                 error = ELDLM_NO_LOCK_DATA;
3031                 break;
3032         case -EEXIST:
3033                 error = ELDLM_NAMESPACE_EXISTS;
3034                 break;
3035         case -EBADF:
3036                 error = ELDLM_BAD_NAMESPACE;
3037                 break;
3038         default:
3039                 error = err_no;
3040         }
3041         return error;
3042 }
3043
3044 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3045 void ldlm_dump_export_locks(struct obd_export *exp)
3046 {
3047         spin_lock(&exp->exp_locks_list_guard);
3048         if (!list_empty(&exp->exp_locks_list)) {
3049                 struct ldlm_lock *lock;
3050
3051                 CERROR("dumping locks for export %p,"
3052                        "ignore if the unmount doesn't hang\n", exp);
3053                 list_for_each_entry(lock, &exp->exp_locks_list,
3054                                         l_exp_refs_link)
3055                         LDLM_ERROR(lock, "lock:");
3056         }
3057         spin_unlock(&exp->exp_locks_list_guard);
3058 }
3059 #endif
3060
3061 #ifdef HAVE_SERVER_SUPPORT
3062 static int target_bulk_timeout(void *data)
3063 {
3064         ENTRY;
3065         /* We don't fail the connection here, because having the export
3066          * killed makes the (vital) call to commitrw very sad.
3067          */
3068         RETURN(1);
3069 }
3070
3071 static inline const char *bulk2type(struct ptlrpc_request *req)
3072 {
3073         if (req->rq_bulk_read)
3074                 return "READ";
3075         if (req->rq_bulk_write)
3076                 return "WRITE";
3077         return "UNKNOWN";
3078 }
3079
3080 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3081                    struct l_wait_info *lwi)
3082 {
3083         struct ptlrpc_request   *req = desc->bd_req;
3084         time_t                   start = cfs_time_current_sec();
3085         time_t                   deadline;
3086         int                      rc = 0;
3087
3088         ENTRY;
3089
3090         /* If there is eviction in progress, wait for it to finish. */
3091         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3092                 *lwi = LWI_INTR(NULL, NULL);
3093                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3094                                   !atomic_read(&exp->exp_obd->
3095                                                    obd_evict_inprogress),
3096                                   lwi);
3097         }
3098
3099         /* Check if client was evicted or reconnected already. */
3100         if (exp->exp_failed ||
3101             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3102                 rc = -ENOTCONN;
3103         } else {
3104                 if (req->rq_bulk_read)
3105                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3106
3107                 if ((exp->exp_connect_data.ocd_connect_flags &
3108                      OBD_CONNECT_BULK_MBITS) != 0)
3109                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3110                 else /* old version, bulk matchbits is rq_xid */
3111                         req->rq_mbits = req->rq_xid;
3112
3113                 if (rc == 0)
3114                         rc = ptlrpc_start_bulk_transfer(desc);
3115         }
3116
3117         if (rc < 0) {
3118                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3119                           bulk2type(req), rc);
3120                 RETURN(rc);
3121         }
3122
3123         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3124                 ptlrpc_abort_bulk(desc);
3125                 RETURN(0);
3126         }
3127
3128         /* limit actual bulk transfer to bulk_timeout seconds */
3129         deadline = start + bulk_timeout;
3130         if (deadline > req->rq_deadline)
3131                 deadline = req->rq_deadline;
3132
3133         do {
3134                 long timeoutl = deadline - cfs_time_current_sec();
3135                 cfs_duration_t timeout = timeoutl <= 0 ?
3136                                          CFS_TICK : cfs_time_seconds(timeoutl);
3137                 time_t  rq_deadline;
3138
3139                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3140                                             target_bulk_timeout, desc);
3141                 rc = l_wait_event(desc->bd_waitq,
3142                                   !ptlrpc_server_bulk_active(desc) ||
3143                                   exp->exp_failed ||
3144                                   exp->exp_conn_cnt >
3145                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3146                                   lwi);
3147                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3148                 /* Wait again if we changed rq_deadline. */
3149                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3150                 deadline = start + bulk_timeout;
3151                 if (deadline > rq_deadline)
3152                         deadline = rq_deadline;
3153         } while ((rc == -ETIMEDOUT) &&
3154                  (deadline > cfs_time_current_sec()));
3155
3156         if (rc == -ETIMEDOUT) {
3157                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3158                           bulk2type(req), deadline - start,
3159                           cfs_time_current_sec() - deadline);
3160                 ptlrpc_abort_bulk(desc);
3161         } else if (exp->exp_failed) {
3162                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3163                           bulk2type(req));
3164                 rc = -ENOTCONN;
3165                 ptlrpc_abort_bulk(desc);
3166         } else if (exp->exp_conn_cnt >
3167                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3168                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3169                           bulk2type(req));
3170                 /* We don't reply anyway. */
3171                 rc = -ETIMEDOUT;
3172                 ptlrpc_abort_bulk(desc);
3173         } else if (desc->bd_failure) {
3174                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3175                           bulk2type(req));
3176                 /* XXX should this be a different errno? */
3177                 rc = -ETIMEDOUT;
3178         } else {
3179                 if (req->rq_bulk_write)
3180                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3181                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3182                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3183                                   bulk2type(req), desc->bd_nob_transferred,
3184                                   desc->bd_nob);
3185                         /* XXX should this be a different errno? */
3186                         rc = -ETIMEDOUT;
3187                 }
3188         }
3189
3190         RETURN(rc);
3191 }
3192 EXPORT_SYMBOL(target_bulk_io);
3193
3194 #endif /* HAVE_SERVER_SUPPORT */