lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2010, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 /**
  34  * This file deals with various client/target related logic including recovery.
  35  *
  36  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  37  * should be moved.
  38  */
  39
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #include <linux/kthread.h>
  43 #include <libcfs/libcfs.h>
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_net.h>
  48 #include <lustre_sec.h>
  49 #include "ldlm_internal.h"
  50
  51 /* @priority: If non-zero, move the selected connection to the list head.
  52  * @create: If zero, only search in existing connections.
  53  */
  54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  55                            int priority, int create)
  56 {
  57         struct ptlrpc_connection *ptlrpc_conn;
  58         struct obd_import_conn *imp_conn = NULL, *item;
  59         int rc = 0;
  60         ENTRY;
  61
  62         if (!create && !priority) {
  63                 CDEBUG(D_HA, "Nothing to do\n");
  64                 RETURN(-EINVAL);
  65         }
  66
  67         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  68         if (!ptlrpc_conn) {
  69                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  70                 RETURN (-ENOENT);
  71         }
  72
  73         if (create) {
  74                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  75                 if (!imp_conn) {
  76                         GOTO(out_put, rc = -ENOMEM);
  77                 }
  78         }
  79
  80         spin_lock(&imp->imp_lock);
  81         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  82                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  83                         if (priority) {
  84                                 list_del(&item->oic_item);
  85                                 list_add(&item->oic_item,
  86                                              &imp->imp_conn_list);
  87                                 item->oic_last_attempt = 0;
  88                         }
  89                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  90                                imp, imp->imp_obd->obd_name, uuid->uuid,
  91                                (priority ? ", moved to head" : ""));
  92                         spin_unlock(&imp->imp_lock);
  93                         GOTO(out_free, rc = 0);
  94                 }
  95         }
  96         /* No existing import connection found for \a uuid. */
  97         if (create) {
  98                 imp_conn->oic_conn = ptlrpc_conn;
  99                 imp_conn->oic_uuid = *uuid;
 100                 imp_conn->oic_last_attempt = 0;
 101                 if (priority)
 102                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 103                 else
 104                         list_add_tail(&imp_conn->oic_item,
 105                                           &imp->imp_conn_list);
 106                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 107                        imp, imp->imp_obd->obd_name, uuid->uuid,
 108                        (priority ? "head" : "tail"));
 109         } else {
 110                 spin_unlock(&imp->imp_lock);
 111                 GOTO(out_free, rc = -ENOENT);
 112         }
 113
 114         spin_unlock(&imp->imp_lock);
 115         RETURN(0);
 116 out_free:
 117         if (imp_conn)
 118                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 119 out_put:
 120         ptlrpc_connection_put(ptlrpc_conn);
 121         RETURN(rc);
 122 }
 123
 124 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 125 {
 126         return import_set_conn(imp, uuid, 1, 0);
 127 }
 128
 129 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 130                            int priority)
 131 {
 132         return import_set_conn(imp, uuid, priority, 1);
 133 }
 134 EXPORT_SYMBOL(client_import_add_conn);
 135
 136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 137 {
 138         struct obd_import_conn *imp_conn;
 139         struct obd_export *dlmexp;
 140         int rc = -ENOENT;
 141         ENTRY;
 142
 143         spin_lock(&imp->imp_lock);
 144         if (list_empty(&imp->imp_conn_list)) {
 145                 LASSERT(!imp->imp_connection);
 146                 GOTO(out, rc);
 147         }
 148
 149         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 150                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 151                         continue;
 152                 LASSERT(imp_conn->oic_conn);
 153
 154                 if (imp_conn == imp->imp_conn_current) {
 155                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 156
 157                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 158                             imp->imp_state != LUSTRE_IMP_DISCON) {
 159                                 CERROR("can't remove current connection\n");
 160                                 GOTO(out, rc = -EBUSY);
 161                         }
 162
 163                         ptlrpc_connection_put(imp->imp_connection);
 164                         imp->imp_connection = NULL;
 165
 166                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 167                         if (dlmexp && dlmexp->exp_connection) {
 168                                 LASSERT(dlmexp->exp_connection ==
 169                                         imp_conn->oic_conn);
 170                                 ptlrpc_connection_put(dlmexp->exp_connection);
 171                                 dlmexp->exp_connection = NULL;
 172                         }
 173
 174                         if (dlmexp != NULL)
 175                                 class_export_put(dlmexp);
 176                 }
 177
 178                 list_del(&imp_conn->oic_item);
 179                 ptlrpc_connection_put(imp_conn->oic_conn);
 180                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 181                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 182                        imp, imp->imp_obd->obd_name, uuid->uuid);
 183                 rc = 0;
 184                 break;
 185         }
 186 out:
 187         spin_unlock(&imp->imp_lock);
 188         if (rc == -ENOENT)
 189                 CERROR("connection %s not found\n", uuid->uuid);
 190         RETURN(rc);
 191 }
 192 EXPORT_SYMBOL(client_import_del_conn);
 193
 194 /**
 195  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 196  * to find a conn uuid of \a imp which can reach \a peer.
 197  */
 198 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 199                             struct obd_uuid *uuid)
 200 {
 201         struct obd_import_conn *conn;
 202         int rc = -ENOENT;
 203         ENTRY;
 204
 205         spin_lock(&imp->imp_lock);
 206         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 207                 /* Check if conn UUID does have this peer NID. */
 208                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 209                         *uuid = conn->oic_uuid;
 210                         rc = 0;
 211                         break;
 212                 }
 213         }
 214         spin_unlock(&imp->imp_lock);
 215         RETURN(rc);
 216 }
 217 EXPORT_SYMBOL(client_import_find_conn);
 218
 219 void client_destroy_import(struct obd_import *imp)
 220 {
 221         /* Drop security policy instance after all RPCs have finished/aborted
 222          * to let all busy contexts be released. */
 223         class_import_get(imp);
 224         class_destroy_import(imp);
 225         sptlrpc_import_sec_put(imp);
 226         class_import_put(imp);
 227 }
 228 EXPORT_SYMBOL(client_destroy_import);
 229
 230 /**
 231  * Check whether or not the OSC is on MDT.
 232  * In the config log,
 233  * osc on MDT
 234  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 235  * osc on client
 236  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 237  *
 238  **/
 239 static int osc_on_mdt(char *obdname)
 240 {
 241         char *ptr;
 242
 243         ptr = strrchr(obdname, '-');
 244         if (ptr == NULL)
 245                 return 0;
 246
 247         if (strncmp(ptr + 1, "MDT", 3) == 0)
 248                 return 1;
 249
 250         return 0;
 251 }
 252
 253 /* Configure an RPC client OBD device.
 254  *
 255  * lcfg parameters:
 256  * 1 - client UUID
 257  * 2 - server UUID
 258  * 3 - inactive-on-startup
 259  */
 260 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 261 {
 262         struct client_obd *cli = &obddev->u.cli;
 263         struct obd_import *imp;
 264         struct obd_uuid server_uuid;
 265         int rq_portal, rp_portal, connect_op;
 266         char *name = obddev->obd_type->typ_name;
 267         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 268         char *cli_name = lustre_cfg_buf(lcfg, 0);
 269         int rc;
 270         ENTRY;
 271
 272         /* In a more perfect world, we would hang a ptlrpc_client off of
 273          * obd_type and just use the values from there. */
 274         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 275                 rq_portal = OST_REQUEST_PORTAL;
 276                 rp_portal = OSC_REPLY_PORTAL;
 277                 connect_op = OST_CONNECT;
 278                 cli->cl_sp_me = LUSTRE_SP_CLI;
 279                 cli->cl_sp_to = LUSTRE_SP_OST;
 280                 ns_type = LDLM_NS_TYPE_OSC;
 281         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 282                    !strcmp(name, LUSTRE_LWP_NAME)) {
 283                 rq_portal = MDS_REQUEST_PORTAL;
 284                 rp_portal = MDC_REPLY_PORTAL;
 285                 connect_op = MDS_CONNECT;
 286                 if (is_lwp_on_ost(cli_name))
 287                         cli->cl_sp_me = LUSTRE_SP_OST;
 288                 else if (is_lwp_on_mdt(cli_name))
 289                         cli->cl_sp_me = LUSTRE_SP_MDT;
 290                 else
 291                         cli->cl_sp_me = LUSTRE_SP_CLI;
 292                 cli->cl_sp_to = LUSTRE_SP_MDT;
 293                 ns_type = LDLM_NS_TYPE_MDC;
 294         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 295                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 296                         /* OSP_on_MDT for other MDTs */
 297                         connect_op = MDS_CONNECT;
 298                         cli->cl_sp_to = LUSTRE_SP_MDT;
 299                         ns_type = LDLM_NS_TYPE_MDC;
 300                         rq_portal = OUT_PORTAL;
 301                 } else {
 302                         /* OSP on MDT for OST */
 303                         connect_op = OST_CONNECT;
 304                         cli->cl_sp_to = LUSTRE_SP_OST;
 305                         ns_type = LDLM_NS_TYPE_OSC;
 306                         rq_portal = OST_REQUEST_PORTAL;
 307                 }
 308                 rp_portal = OSC_REPLY_PORTAL;
 309                 cli->cl_sp_me = LUSTRE_SP_MDT;
 310         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 311                 rq_portal = MGS_REQUEST_PORTAL;
 312                 rp_portal = MGC_REPLY_PORTAL;
 313                 connect_op = MGS_CONNECT;
 314                 cli->cl_sp_me = LUSTRE_SP_MGC;
 315                 cli->cl_sp_to = LUSTRE_SP_MGS;
 316                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 317                 ns_type = LDLM_NS_TYPE_MGC;
 318         } else {
 319                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 320                        name);
 321                 RETURN(-EINVAL);
 322         }
 323
 324         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 325                 CERROR("requires a TARGET UUID\n");
 326                 RETURN(-EINVAL);
 327         }
 328
 329         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 330                 CERROR("client UUID must be less than 38 characters\n");
 331                 RETURN(-EINVAL);
 332         }
 333
 334         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 335                 CERROR("setup requires a SERVER UUID\n");
 336                 RETURN(-EINVAL);
 337         }
 338
 339         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 340                 CERROR("target UUID must be less than 38 characters\n");
 341                 RETURN(-EINVAL);
 342         }
 343
 344         init_rwsem(&cli->cl_sem);
 345         mutex_init(&cli->cl_mgc_mutex);
 346         cli->cl_conn_count = 0;
 347         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 348                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 349                      sizeof(server_uuid)));
 350
 351         cli->cl_dirty_pages = 0;
 352         cli->cl_avail_grant = 0;
 353         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 354         /* cl_dirty_max_pages may be changed at connect time in
 355          * ptlrpc_connect_interpret(). */
 356         client_adjust_max_dirty(cli);
 357         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 358         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 359         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 360         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 361         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 362         spin_lock_init(&cli->cl_loi_list_lock);
 363         atomic_set(&cli->cl_pending_w_pages, 0);
 364         atomic_set(&cli->cl_pending_r_pages, 0);
 365         cli->cl_r_in_flight = 0;
 366         cli->cl_w_in_flight = 0;
 367
 368         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 369         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 370         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 371         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 372         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 373         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 374
 375         /* lru for osc. */
 376         INIT_LIST_HEAD(&cli->cl_lru_osc);
 377         atomic_set(&cli->cl_lru_shrinkers, 0);
 378         atomic_long_set(&cli->cl_lru_busy, 0);
 379         atomic_long_set(&cli->cl_lru_in_list, 0);
 380         INIT_LIST_HEAD(&cli->cl_lru_list);
 381         spin_lock_init(&cli->cl_lru_list_lock);
 382         atomic_long_set(&cli->cl_unstable_count, 0);
 383         INIT_LIST_HEAD(&cli->cl_shrink_list);
 384
 385         init_waitqueue_head(&cli->cl_destroy_waitq);
 386         atomic_set(&cli->cl_destroy_in_flight, 0);
 387 #ifdef ENABLE_CHECKSUM
 388         /* Turn on checksumming by default. */
 389         cli->cl_checksum = 1;
 390         /*
 391          * The supported checksum types will be worked out at connect time
 392          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 393          * through procfs.
 394          */
 395         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 396 #endif
 397         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 398
 399         /* Set it to possible maximum size. It may be reduced by ocd_brw_size
 400          * from OFD after connecting. */
 401         cli->cl_max_pages_per_rpc = PTLRPC_MAX_BRW_PAGES;
 402
 403         /* set cl_chunkbits default value to PAGE_SHIFT,
 404          * it will be updated at OSC connection time. */
 405         cli->cl_chunkbits = PAGE_SHIFT;
 406
 407         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 408                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 409         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 128 /* MB */) {
 410                 cli->cl_max_rpcs_in_flight = 2;
 411         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 256 /* MB */) {
 412                 cli->cl_max_rpcs_in_flight = 3;
 413         } else if (totalram_pages >> (20 - PAGE_SHIFT) <= 512 /* MB */) {
 414                 cli->cl_max_rpcs_in_flight = 4;
 415         } else {
 416                 if (osc_on_mdt(obddev->obd_name))
 417                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 418                 else
 419                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 420         }
 421
 422         spin_lock_init(&cli->cl_mod_rpcs_lock);
 423         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 424         cli->cl_max_mod_rpcs_in_flight = 0;
 425         cli->cl_mod_rpcs_in_flight = 0;
 426         cli->cl_close_rpcs_in_flight = 0;
 427         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 428         cli->cl_mod_tag_bitmap = NULL;
 429
 430         if (connect_op == MDS_CONNECT) {
 431                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 432                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 433                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 434                 if (cli->cl_mod_tag_bitmap == NULL)
 435                         GOTO(err, rc = -ENOMEM);
 436         }
 437
 438         rc = ldlm_get_ref();
 439         if (rc) {
 440                 CERROR("ldlm_get_ref failed: %d\n", rc);
 441                 GOTO(err, rc);
 442         }
 443
 444         ptlrpc_init_client(rq_portal, rp_portal, name,
 445                            &obddev->obd_ldlm_client);
 446
 447         imp = class_new_import(obddev);
 448         if (imp == NULL)
 449                 GOTO(err_ldlm, rc = -ENOENT);
 450         imp->imp_client = &obddev->obd_ldlm_client;
 451         imp->imp_connect_op = connect_op;
 452         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 453                LUSTRE_CFG_BUFLEN(lcfg, 1));
 454         class_import_put(imp);
 455
 456         rc = client_import_add_conn(imp, &server_uuid, 1);
 457         if (rc) {
 458                 CERROR("can't add initial connection\n");
 459                 GOTO(err_import, rc);
 460         }
 461
 462         cli->cl_import = imp;
 463         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 464         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 465
 466         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 467                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 468                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 469                                name, obddev->obd_name,
 470                                cli->cl_target_uuid.uuid);
 471                         spin_lock(&imp->imp_lock);
 472                         imp->imp_deactive = 1;
 473                         spin_unlock(&imp->imp_lock);
 474                 }
 475         }
 476
 477         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 478                                                    LDLM_NAMESPACE_CLIENT,
 479                                                    LDLM_NAMESPACE_GREEDY,
 480                                                    ns_type);
 481         if (obddev->obd_namespace == NULL) {
 482                 CERROR("Unable to create client namespace - %s\n",
 483                        obddev->obd_name);
 484                 GOTO(err_import, rc = -ENOMEM);
 485         }
 486
 487         RETURN(rc);
 488
 489 err_import:
 490         class_destroy_import(imp);
 491 err_ldlm:
 492         ldlm_put_ref();
 493 err:
 494         if (cli->cl_mod_tag_bitmap != NULL)
 495                 OBD_FREE(cli->cl_mod_tag_bitmap,
 496                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 497         cli->cl_mod_tag_bitmap = NULL;
 498         RETURN(rc);
 499
 500 }
 501 EXPORT_SYMBOL(client_obd_setup);
 502
 503 int client_obd_cleanup(struct obd_device *obddev)
 504 {
 505         struct client_obd *cli = &obddev->u.cli;
 506         ENTRY;
 507
 508         ldlm_namespace_free_post(obddev->obd_namespace);
 509         obddev->obd_namespace = NULL;
 510
 511         obd_cleanup_client_import(obddev);
 512         LASSERT(obddev->u.cli.cl_import == NULL);
 513
 514         ldlm_put_ref();
 515
 516         if (cli->cl_mod_tag_bitmap != NULL)
 517                 OBD_FREE(cli->cl_mod_tag_bitmap,
 518                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 519         cli->cl_mod_tag_bitmap = NULL;
 520
 521         RETURN(0);
 522 }
 523 EXPORT_SYMBOL(client_obd_cleanup);
 524
 525 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 526 int client_connect_import(const struct lu_env *env,
 527                           struct obd_export **exp,
 528                           struct obd_device *obd, struct obd_uuid *cluuid,
 529                           struct obd_connect_data *data, void *localdata)
 530 {
 531         struct client_obd       *cli    = &obd->u.cli;
 532         struct obd_import       *imp    = cli->cl_import;
 533         struct obd_connect_data *ocd;
 534         struct lustre_handle    conn    = { 0 };
 535         int                     rc;
 536         bool                    is_mdc = false;
 537         ENTRY;
 538
 539         *exp = NULL;
 540         down_write(&cli->cl_sem);
 541         if (cli->cl_conn_count > 0)
 542                 GOTO(out_sem, rc = -EALREADY);
 543
 544         rc = class_connect(&conn, obd, cluuid);
 545         if (rc)
 546                 GOTO(out_sem, rc);
 547
 548         cli->cl_conn_count++;
 549         *exp = class_conn2export(&conn);
 550
 551         LASSERT(obd->obd_namespace);
 552
 553         imp->imp_dlm_handle = conn;
 554         rc = ptlrpc_init_import(imp);
 555         if (rc != 0)
 556                 GOTO(out_ldlm, rc);
 557
 558         ocd = &imp->imp_connect_data;
 559         if (data) {
 560                 *ocd = *data;
 561                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
 562                                  LUSTRE_MDC_NAME, 3) == 0;
 563                 if (is_mdc)
 564                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
 565                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 566                 imp->imp_connect_flags2_orig = data->ocd_connect_flags2;
 567         }
 568
 569         rc = ptlrpc_connect_import(imp);
 570         if (rc != 0) {
 571                 if (data && is_mdc)
 572                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 573                 LASSERT(imp->imp_state == LUSTRE_IMP_DISCON);
 574                 GOTO(out_ldlm, rc);
 575         }
 576         LASSERT(*exp != NULL && (*exp)->exp_connection);
 577
 578         if (data) {
 579                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 580                          ocd->ocd_connect_flags, "old %#llx, new %#llx\n",
 581                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 582                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 583                 /* clear the flag as it was not set and is not known
 584                  * by upper layers */
 585                 if (is_mdc)
 586                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 587         }
 588
 589         ptlrpc_pinger_add_import(imp);
 590
 591         EXIT;
 592
 593         if (rc) {
 594 out_ldlm:
 595                 cli->cl_conn_count--;
 596                 class_disconnect(*exp);
 597                 *exp = NULL;
 598         }
 599 out_sem:
 600         up_write(&cli->cl_sem);
 601
 602         return rc;
 603 }
 604 EXPORT_SYMBOL(client_connect_import);
 605
 606 int client_disconnect_export(struct obd_export *exp)
 607 {
 608         struct obd_device *obd = class_exp2obd(exp);
 609         struct client_obd *cli;
 610         struct obd_import *imp;
 611         int rc = 0, err;
 612         ENTRY;
 613
 614         if (!obd) {
 615                 CERROR("invalid export for disconnect: exp %p cookie %#llx\n",
 616                        exp, exp ? exp->exp_handle.h_cookie : -1);
 617                 RETURN(-EINVAL);
 618         }
 619
 620         cli = &obd->u.cli;
 621         imp = cli->cl_import;
 622
 623         down_write(&cli->cl_sem);
 624         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 625                 cli->cl_conn_count);
 626
 627         if (cli->cl_conn_count == 0) {
 628                 CERROR("disconnecting disconnected device (%s)\n",
 629                        obd->obd_name);
 630                 GOTO(out_disconnect, rc = -EINVAL);
 631         }
 632
 633         cli->cl_conn_count--;
 634         if (cli->cl_conn_count != 0)
 635                 GOTO(out_disconnect, rc = 0);
 636
 637         /* Mark import deactivated now, so we don't try to reconnect if any
 638          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 639          * fully deactivate the import, or that would drop all requests. */
 640         spin_lock(&imp->imp_lock);
 641         imp->imp_deactive = 1;
 642         spin_unlock(&imp->imp_lock);
 643
 644         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 645          * delete it regardless.  (It's safe to delete an import that was
 646          * never added.) */
 647         (void)ptlrpc_pinger_del_import(imp);
 648
 649         if (obd->obd_namespace != NULL) {
 650                 /* obd_force == local only */
 651                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 652                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 653                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 654         }
 655
 656         /* There's no need to hold sem while disconnecting an import,
 657          * and it may actually cause deadlock in GSS. */
 658         up_write(&cli->cl_sem);
 659         rc = ptlrpc_disconnect_import(imp, 0);
 660         down_write(&cli->cl_sem);
 661
 662         ptlrpc_invalidate_import(imp);
 663
 664         EXIT;
 665
 666 out_disconnect:
 667         /* Use server style - class_disconnect should be always called for
 668          * o_disconnect. */
 669         err = class_disconnect(exp);
 670         if (!rc && err)
 671                 rc = err;
 672
 673         up_write(&cli->cl_sem);
 674
 675         RETURN(rc);
 676 }
 677 EXPORT_SYMBOL(client_disconnect_export);
 678
 679 #ifdef HAVE_SERVER_SUPPORT
 680 int server_disconnect_export(struct obd_export *exp)
 681 {
 682         int rc;
 683         ENTRY;
 684
 685         /* Disconnect early so that clients can't keep using export. */
 686         rc = class_disconnect(exp);
 687         /* Close import to avoid sending any requests. */
 688         if (exp->exp_imp_reverse)
 689                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 690
 691         ldlm_bl_thread_wakeup();
 692
 693         /* complete all outstanding replies */
 694         spin_lock(&exp->exp_lock);
 695         while (!list_empty(&exp->exp_outstanding_replies)) {
 696                 struct ptlrpc_reply_state *rs =
 697                         list_entry(exp->exp_outstanding_replies.next,
 698                                        struct ptlrpc_reply_state, rs_exp_list);
 699                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 700
 701                 spin_lock(&svcpt->scp_rep_lock);
 702
 703                 list_del_init(&rs->rs_exp_list);
 704                 spin_lock(&rs->rs_lock);
 705                 ptlrpc_schedule_difficult_reply(rs);
 706                 spin_unlock(&rs->rs_lock);
 707
 708                 spin_unlock(&svcpt->scp_rep_lock);
 709         }
 710         spin_unlock(&exp->exp_lock);
 711
 712         RETURN(rc);
 713 }
 714 EXPORT_SYMBOL(server_disconnect_export);
 715
 716 /* --------------------------------------------------------------------------
 717  * from old lib/target.c
 718  * -------------------------------------------------------------------------- */
 719
 720 static int target_handle_reconnect(struct lustre_handle *conn,
 721                                    struct obd_export *exp,
 722                                    struct obd_uuid *cluuid)
 723 {
 724         struct obd_device *target;
 725         struct lustre_handle *hdl;
 726         cfs_time_t now;
 727         cfs_time_t deadline;
 728         int timeout;
 729         int rc = 0;
 730         ENTRY;
 731
 732         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 733         if (!exp->exp_connection || !lustre_handle_is_used(hdl)) {
 734                 conn->cookie = exp->exp_handle.h_cookie;
 735                 CDEBUG(D_HA, "connect export for UUID '%s' at %p,"
 736                        " cookie %#llx\n", cluuid->uuid, exp, conn->cookie);
 737                 RETURN(0);
 738         }
 739
 740         target = exp->exp_obd;
 741
 742         /* Might be a re-connect after a partition. */
 743         if (memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 744                 LCONSOLE_WARN("%s: already connected client %s (at %s) "
 745                               "with handle %#llx. Rejecting client "
 746                               "with the same UUID trying to reconnect "
 747                               "with handle %#llx\n", target->obd_name,
 748                               obd_uuid2str(&exp->exp_client_uuid),
 749                               obd_export_nid2str(exp),
 750                               hdl->cookie, conn->cookie);
 751                 memset(conn, 0, sizeof *conn);
 752                 /* target_handle_connect() treats EALREADY and
 753                  * -EALREADY differently.  -EALREADY is an error
 754                  * (same UUID, different handle). */
 755                 RETURN(-EALREADY);
 756         }
 757
 758         if (!target->obd_recovering) {
 759                 LCONSOLE_WARN("%s: Client %s (at %s) reconnecting\n",
 760                         target->obd_name, obd_uuid2str(&exp->exp_client_uuid),
 761                         obd_export_nid2str(exp));
 762                 GOTO(out_already, rc);
 763         }
 764
 765         now = cfs_time_current();
 766         deadline = cfs_timer_deadline(&target->obd_recovery_timer);
 767         if (cfs_time_before(now, deadline)) {
 768                 struct target_distribute_txn_data *tdtd =
 769                                         class_exp2tgt(exp)->lut_tdtd;
 770                 int size = 0;
 771                 int count = 0;
 772                 char *buf = NULL;
 773
 774                 timeout = cfs_duration_sec(cfs_time_sub(deadline, now));
 775                 if (tdtd && tdtd->tdtd_show_update_logs_retrievers)
 776                         buf = tdtd->tdtd_show_update_logs_retrievers(
 777                                 tdtd->tdtd_show_retrievers_cbdata,
 778                                 &size, &count);
 779
 780                 if (count > 0)
 781                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 782                                       "%d:%.02d. It is due to DNE recovery "
 783                                       "failed/stuck on the %d MDT(s):%s. "
 784                                       "Please wait until all MDTs recovered "
 785                                       "or abort the recovery by force.\n",
 786                                       target->obd_name, timeout / 60,
 787                                       timeout % 60, count,
 788                                       buf ? buf : "unknown (not enough RAM)");
 789                 else
 790                         LCONSOLE_WARN("%s: Recovery already passed deadline "
 791                                       "%d:%.02d. If you do not want to wait "
 792                                       "more, please abort the recovery by "
 793                                       "force.\n", target->obd_name,
 794                                       timeout / 60, timeout % 60);
 795
 796                 if (buf != NULL)
 797                         OBD_FREE(buf, size);
 798         } else {
 799                 timeout = cfs_duration_sec(cfs_time_sub(now, deadline));
 800                 LCONSOLE_WARN("%s: Recovery already passed deadline"
 801                         " %d:%.02d, It is most likely due to DNE"
 802                         " recovery is failed or stuck, please wait a"
 803                         " few more minutes or abort the recovery.\n",
 804                         target->obd_name, timeout / 60, timeout % 60);
 805         }
 806
 807 out_already:
 808         conn->cookie = exp->exp_handle.h_cookie;
 809         /* target_handle_connect() treats EALREADY and
 810          * -EALREADY differently.  EALREADY means we are
 811          * doing a valid reconnect from the same client. */
 812         RETURN(EALREADY);
 813 }
 814
 815 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 816                           int error)
 817 {
 818         struct obd_export *exp = cb_data;
 819
 820         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 821                obd->obd_name, exp->exp_client_uuid.uuid);
 822
 823         spin_lock(&exp->exp_lock);
 824         exp->exp_need_sync = 0;
 825         spin_unlock(&exp->exp_lock);
 826         class_export_cb_put(exp);
 827 }
 828
 829 static void
 830 check_and_start_recovery_timer(struct obd_device *obd,
 831                                struct ptlrpc_request *req, int new_client);
 832
 833 /**
 834  * update flags for import during reconnect process
 835  */
 836 static int rev_import_flags_update(struct obd_import *revimp,
 837                                    struct ptlrpc_request *req)
 838 {
 839         int rc;
 840         struct obd_connect_data *data;
 841
 842         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 843
 844         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 845                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 846         else
 847                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 848
 849         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 850
 851         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 852         if (rc) {
 853                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 854                         revimp->imp_client->cli_name,
 855                         libcfs_id2str(req->rq_peer), rc);
 856                 return rc;
 857         }
 858
 859         return 0;
 860 }
 861
 862 /**
 863  * Allocate a new reverse import for an export.
 864  *
 865  * \retval -errno in case error hit
 866  * \retval 0 if reverse import correctly init
 867  **/
 868 int rev_import_init(struct obd_export *export)
 869 {
 870         struct obd_device *obd = export->exp_obd;
 871         struct obd_import *revimp;
 872
 873         LASSERT(export->exp_imp_reverse == NULL);
 874
 875         revimp = class_new_import(obd);
 876         if (revimp == NULL)
 877                 return -ENOMEM;
 878
 879         revimp->imp_remote_handle.cookie = 0ULL;
 880         revimp->imp_client = &obd->obd_ldlm_client;
 881         revimp->imp_dlm_fake = 1;
 882
 883         /* it is safe to connect import in new state as no sends possible */
 884         spin_lock(&export->exp_lock);
 885         export->exp_imp_reverse = revimp;
 886         spin_unlock(&export->exp_lock);
 887         class_import_put(revimp);
 888
 889         return 0;
 890 }
 891 EXPORT_SYMBOL(rev_import_init);
 892
 893 /**
 894  * Handle reconnect for an export.
 895  *
 896  * \param exp export to handle reconnect process
 897  * \param req client reconnect request
 898  *
 899  * \retval -rc in case securitfy flavor can't be changed
 900  * \retval 0 in case none problems
 901  */
 902 static int rev_import_reconnect(struct obd_export *exp,
 903                                 struct ptlrpc_request *req)
 904 {
 905         struct obd_import *revimp = exp->exp_imp_reverse;
 906         struct lustre_handle *lh;
 907         int rc;
 908
 909         /* avoid sending a request until import flags are changed */
 910         ptlrpc_import_enter_resend(revimp);
 911
 912         if (revimp->imp_connection != NULL)
 913                 ptlrpc_connection_put(revimp->imp_connection);
 914
 915         /*
 916          * client from recovery don't have a handle so we need to take from
 917          * request. it may produce situation when wrong client connected
 918          * to recovery as we trust a client uuid
 919          */
 920         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 921         revimp->imp_remote_handle = *lh;
 922
 923         /* unknown versions will be caught in
 924          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 925         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 926
 927         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 928
 929         rc = rev_import_flags_update(revimp, req);
 930         if (rc != 0) {
 931                 /* it is safe to still be in RECOVERY phase as we are not able
 932                  * to setup correct security flavor so requests are not able to
 933                  * be delivered correctly */
 934                 return rc;
 935         }
 936
 937         /* resend all rpc's via new connection */
 938         return ptlrpc_import_recovery_state_machine(revimp);
 939 }
 940
 941 int target_handle_connect(struct ptlrpc_request *req)
 942 {
 943         struct obd_device *target = NULL;
 944         struct obd_export *export = NULL;
 945         /* connect handle - filled from target_handle_reconnect in
 946          * reconnect case */
 947         struct lustre_handle conn;
 948         struct lustre_handle *tmp;
 949         struct obd_uuid tgtuuid;
 950         struct obd_uuid cluuid;
 951         char *str;
 952         int rc = 0;
 953         char *target_start;
 954         int target_len;
 955         bool     mds_conn = false, lw_client = false, initial_conn = false;
 956         bool     mds_mds_conn = false;
 957         bool     new_mds_mds_conn = false;
 958         bool     target_referenced = false;
 959         struct obd_connect_data *data, *tmpdata;
 960         int size, tmpsize;
 961         lnet_nid_t *client_nid = NULL;
 962         ENTRY;
 963
 964         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 965
 966         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 967         if (str == NULL) {
 968                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 969                 GOTO(out, rc = -EINVAL);
 970         }
 971
 972         obd_str2uuid(&tgtuuid, str);
 973         target = class_uuid2obd(&tgtuuid);
 974         if (!target)
 975                 target = class_name2obd(str);
 976
 977         if (!target) {
 978                 deuuidify(str, NULL, &target_start, &target_len);
 979                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 980                                    "from %s (no target). If you are running "
 981                                    "an HA pair check that the target is "
 982                                    "mounted on the other server.\n", str,
 983                                    libcfs_nid2str(req->rq_peer.nid));
 984                 GOTO(out, rc = -ENODEV);
 985         }
 986
 987         spin_lock(&target->obd_dev_lock);
 988         if (target->obd_stopping || !target->obd_set_up) {
 989                 spin_unlock(&target->obd_dev_lock);
 990
 991                 deuuidify(str, NULL, &target_start, &target_len);
 992                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 993                               target_len, target_start,
 994                               libcfs_nid2str(req->rq_peer.nid),
 995                               (target->obd_stopping ?
 996                                "stopping" : "not set up"));
 997                 GOTO(out, rc = -ENODEV);
 998         }
 999
1000         if (target->obd_no_conn) {
1001                 spin_unlock(&target->obd_dev_lock);
1002
1003                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
1004                                "from %s\n", target->obd_name,
1005                                libcfs_nid2str(req->rq_peer.nid));
1006                 GOTO(out, rc = -EAGAIN);
1007         }
1008
1009         /* Make sure the target isn't cleaned up while we're here. Yes,
1010          * there's still a race between the above check and our incref here.
1011          * Really, class_uuid2obd should take the ref. */
1012         class_incref(target, __func__, current);
1013         target_referenced = true;
1014
1015         target->obd_conn_inprogress++;
1016         spin_unlock(&target->obd_dev_lock);
1017
1018         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
1019         if (str == NULL) {
1020                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
1021                 GOTO(out, rc = -EINVAL);
1022         }
1023
1024         obd_str2uuid(&cluuid, str);
1025
1026         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1027         if (tmp == NULL)
1028                 GOTO(out, rc = -EPROTO);
1029
1030         conn = *tmp;
1031
1032         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1033                                     RCL_CLIENT);
1034         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1035         if (!data)
1036                 GOTO(out, rc = -EPROTO);
1037
1038         rc = req_capsule_server_pack(&req->rq_pill);
1039         if (rc)
1040                 GOTO(out, rc);
1041
1042 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1043         /* Don't allow clients to connect that are using old 1.8 format
1044          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1045          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1046          * FULL20 flag should be set on all connections since 2.0, but no
1047          * longer affects behaviour.
1048          *
1049          * Later this check will be disabled and the flag can be retired
1050          * completely once interop with 3.0 is no longer needed.
1051          */
1052         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1053                 GOTO(out, rc = -EPROTO);
1054 #endif
1055
1056         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1057                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1058                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1059                     data->ocd_version > LUSTRE_VERSION_CODE +
1060                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1061                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1062                                   "libclient connection attempt",
1063                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1064                                   "old" : "new",
1065                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1066                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1067                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1068                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1069                         data = req_capsule_server_sized_get(&req->rq_pill,
1070                                                             &RMF_CONNECT_DATA,
1071                                     offsetof(typeof(*data), ocd_version) +
1072                                              sizeof(data->ocd_version));
1073                         if (data) {
1074                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1075                                 data->ocd_version = LUSTRE_VERSION_CODE;
1076                         }
1077                         GOTO(out, rc = -EPROTO);
1078                 }
1079         }
1080
1081         /* Note: lw_client is needed in MDS-MDS failover during update log
1082          * processing, so we needs to allow lw_client to be connected at
1083          * anytime, instead of only the initial connection */
1084         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1085
1086         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1087                 initial_conn = true;
1088                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1089                 mds_mds_conn = (data->ocd_connect_flags &
1090                                 OBD_CONNECT_MDS_MDS) != 0;
1091
1092                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1093                  * for Imperative Recovery connection from MGC to MGS.
1094                  *
1095                  * Via check OBD_CONNECT_FID, we can distinguish whether
1096                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1097                  * MGC or MDT. */
1098                 if (!lw_client &&
1099                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1100                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1101                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1102                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1103                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1104                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1105
1106                         /* We do not support the MDT-MDT interoperations with
1107                          * different version MDT because of protocol changes. */
1108                         if (unlikely(major != LUSTRE_MAJOR ||
1109                                      minor != LUSTRE_MINOR ||
1110                                      abs(patch - LUSTRE_PATCH) > 3)) {
1111                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1112                                         "connection from different version MDT "
1113                                         "(%d.%d.%d.%d) %s %s\n",
1114                                         target->obd_name, LUSTRE_MAJOR,
1115                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1116                                         major, minor, patch,
1117                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1118                                         libcfs_nid2str(req->rq_peer.nid), str);
1119
1120                                 GOTO(out, rc = -EPROTO);
1121                         }
1122                 }
1123         }
1124
1125         /* lctl gets a backstage, all-access pass. */
1126         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1127                 goto dont_check_exports;
1128
1129         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1130         if (!export)
1131                 goto no_export;
1132
1133         /* We've found an export in the hash. */
1134
1135         spin_lock(&export->exp_lock);
1136
1137         if (export->exp_connecting) { /* bug 9635, et. al. */
1138                 spin_unlock(&export->exp_lock);
1139                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1140                               export->exp_obd->obd_name, export,
1141                               libcfs_nid2str(req->rq_peer.nid));
1142                 class_export_put(export);
1143                 export = NULL;
1144                 rc = -EALREADY;
1145         } else if ((mds_conn || (lw_client && initial_conn) ||
1146                    data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1147                    export->exp_connection != NULL) {
1148                 spin_unlock(&export->exp_lock);
1149                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1150                         /* MDS or LWP reconnected after failover. */
1151                         LCONSOLE_WARN("%s: Received %s connection from "
1152                             "%s, removing former export from %s\n",
1153                             target->obd_name, mds_conn ? "MDS" : "LWP",
1154                             libcfs_nid2str(req->rq_peer.nid),
1155                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1156                 } else {
1157                         /* New MDS connection from the same NID. */
1158                         LCONSOLE_WARN("%s: Received new %s connection from "
1159                                 "%s, removing former export from same NID\n",
1160                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1161                                 libcfs_nid2str(req->rq_peer.nid));
1162                 }
1163
1164                 if (req->rq_peer.nid == export->exp_connection->c_peer.nid &&
1165                     data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) {
1166                         /* Because exports between MDTs will always be
1167                          * kept, let's do not fail such export if they
1168                          * come from the same NID, otherwise it might
1169                          * cause eviction between MDTs, which might
1170                          * cause namespace inconsistency */
1171                         spin_lock(&export->exp_lock);
1172                         export->exp_connecting = 1;
1173                         spin_unlock(&export->exp_lock);
1174                         conn.cookie = export->exp_handle.h_cookie;
1175                         rc = EALREADY;
1176                 } else {
1177                         class_fail_export(export);
1178                         class_export_put(export);
1179                         export = NULL;
1180                         rc = 0;
1181                 }
1182         } else if (export->exp_connection != NULL && initial_conn &&
1183                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
1184                 spin_unlock(&export->exp_lock);
1185                 /* In MDS failover we have static UUID but NID can change. */
1186                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1187                               "existing nid %s is already connected\n",
1188                               target->obd_name, cluuid.uuid,
1189                               libcfs_nid2str(req->rq_peer.nid),
1190                               libcfs_nid2str(
1191                                       export->exp_connection->c_peer.nid));
1192                 rc = -EALREADY;
1193                 class_export_put(export);
1194                 export = NULL;
1195         } else {
1196                 export->exp_connecting = 1;
1197                 spin_unlock(&export->exp_lock);
1198                 LASSERT(export->exp_obd == target);
1199
1200                 rc = target_handle_reconnect(&conn, export, &cluuid);
1201         }
1202
1203         /* If we found an export, we already unlocked. */
1204         if (!export) {
1205 no_export:
1206                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1207         } else if (req->rq_export == NULL &&
1208                    atomic_read(&export->exp_rpc_count) > 0) {
1209                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1210                               "still busy with %d references\n",
1211                               target->obd_name, cluuid.uuid,
1212                               libcfs_nid2str(req->rq_peer.nid),
1213                               atomic_read(&export->exp_refcount));
1214                 GOTO(out, rc = -EBUSY);
1215         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1216                 if (!strstr(cluuid.uuid, "mdt"))
1217                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1218                                       "known client %s (at %s) because it "
1219                                       "is indicating it is a new client",
1220                                       target->obd_name, cluuid.uuid,
1221                                       libcfs_nid2str(req->rq_peer.nid));
1222                 GOTO(out, rc = -EALREADY);
1223         } else {
1224                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1225         }
1226
1227         if (rc < 0) {
1228                 GOTO(out, rc);
1229         }
1230
1231         CDEBUG(D_HA, "%s: connection from %s@%s %st%llu exp %p cur %ld last %ld\n",
1232                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1233               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1234               export, (long)cfs_time_current_sec(),
1235               export ? (long)export->exp_last_request_time : 0);
1236
1237         /* If this is the first time a client connects, reset the recovery
1238          * timer. Discard lightweight connections which might be local. */
1239         if (!lw_client && rc == 0 && target->obd_recovering)
1240                 check_and_start_recovery_timer(target, req, export == NULL);
1241
1242         /* We want to handle EALREADY but *not* -EALREADY from
1243          * target_handle_reconnect(), return reconnection state in a flag. */
1244         if (rc == EALREADY) {
1245                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1246                 rc = 0;
1247         } else {
1248                 LASSERT(rc == 0);
1249         }
1250
1251         /* Tell the client if we support replayable requests. */
1252         if (target->obd_replayable)
1253                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1254         client_nid = &req->rq_peer.nid;
1255
1256         if (export == NULL) {
1257                 /* allow lightweight connections during recovery */
1258                 /* allow "new" MDT to be connected during recovery, since we
1259                  * need retrieve recovery update records from it */
1260                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1261                         cfs_time_t t;
1262                         int     c; /* connected */
1263                         int     i; /* in progress */
1264                         int     k; /* known */
1265                         int     s; /* stale/evicted */
1266
1267                         c = atomic_read(&target->obd_connected_clients);
1268                         i = atomic_read(&target->obd_lock_replay_clients);
1269                         k = target->obd_max_recoverable_clients;
1270                         s = target->obd_stale_clients;
1271                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1272                         t = cfs_time_sub(t, cfs_time_current());
1273                         t = cfs_duration_sec(t);
1274                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1275                                       "(at %s), waiting for %d known clients "
1276                                       "(%d recovered, %d in progress, and %d "
1277                                       "evicted) to recover in %d:%.02d\n",
1278                                       target->obd_name, cluuid.uuid,
1279                                       libcfs_nid2str(req->rq_peer.nid), k,
1280                                       c - i, i, s, (int)t / 60,
1281                                       (int)t % 60);
1282                         rc = -EBUSY;
1283                 } else {
1284 dont_check_exports:
1285                         rc = obd_connect(req->rq_svc_thread->t_env,
1286                                          &export, target, &cluuid, data,
1287                                          client_nid);
1288                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1289                                 lustre_msg_add_op_flags(req->rq_repmsg,
1290                                                         MSG_CONNECT_RECOVERING);
1291                         if (rc == 0) {
1292                                 conn.cookie = export->exp_handle.h_cookie;
1293                                 rc = rev_import_init(export);
1294                         }
1295
1296                         if (mds_mds_conn)
1297                                 new_mds_mds_conn = true;
1298                 }
1299         } else {
1300                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1301                                    export, target, &cluuid, data, client_nid);
1302         }
1303         if (rc)
1304                 GOTO(out, rc);
1305
1306         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1307         data->ocd_instance = target->u.obt.obt_instance;
1308
1309         /* Return only the parts of obd_connect_data that we understand, so the
1310          * client knows that we don't understand the rest. */
1311         if (data) {
1312                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1313                                                RCL_SERVER);
1314                 tmpdata = req_capsule_server_get(&req->rq_pill,
1315                                                  &RMF_CONNECT_DATA);
1316                 /* Don't use struct assignment here, because the client reply
1317                  * buffer may be smaller/larger than the local struct
1318                  * obd_connect_data. */
1319                 memcpy(tmpdata, data, min(tmpsize, size));
1320         }
1321
1322         /* If the client and the server are the same node, we will already
1323          * have an export that really points to the client's DLM export,
1324          * because we have a shared handles table.
1325          *
1326          * XXX this will go away when shaver stops sending the "connect" handle
1327          * in the real "remote handle" field of the request --phik 24 Apr 2003
1328          */
1329         ptlrpc_request_change_export(req, export);
1330
1331         spin_lock(&export->exp_lock);
1332         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1333                 spin_unlock(&export->exp_lock);
1334                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1335                        "or equal conn_cnt: %d >= %d\n",
1336                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1337                        export->exp_conn_cnt,
1338                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1339
1340                 GOTO(out, rc = -EALREADY);
1341         }
1342         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1343         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1344
1345         /* Don't evict liblustre clients for not pinging. */
1346         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1347                 export->exp_libclient = 1;
1348                 spin_unlock(&export->exp_lock);
1349
1350                 spin_lock(&target->obd_dev_lock);
1351                 list_del_init(&export->exp_obd_chain_timed);
1352                 spin_unlock(&target->obd_dev_lock);
1353         } else {
1354                 spin_unlock(&export->exp_lock);
1355         }
1356
1357         if (export->exp_connection != NULL) {
1358                 /* Check to see if connection came from another NID. */
1359                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1360                     !hlist_unhashed(&export->exp_nid_hash))
1361                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1362                                      &export->exp_connection->c_peer.nid,
1363                                      &export->exp_nid_hash);
1364
1365                 ptlrpc_connection_put(export->exp_connection);
1366         }
1367
1368         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1369                                                        req->rq_self,
1370                                                        &cluuid);
1371         if (hlist_unhashed(&export->exp_nid_hash))
1372                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1373                              &export->exp_connection->c_peer.nid,
1374                              &export->exp_nid_hash);
1375
1376         lustre_msg_set_handle(req->rq_repmsg, &conn);
1377
1378         rc = rev_import_reconnect(export, req);
1379         if (rc != 0)
1380                 GOTO(out, rc);
1381
1382         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1383                 int has_transno;
1384                 __u64 transno = data->ocd_transno;
1385
1386                 spin_lock(&export->exp_lock);
1387                 /* possible race with class_disconnect_stale_exports,
1388                  * export may be already in the eviction process */
1389                 if (export->exp_failed) {
1390                         spin_unlock(&export->exp_lock);
1391                         GOTO(out, rc = -ENODEV);
1392                 }
1393                 export->exp_in_recovery = 1;
1394                 export->exp_req_replay_needed = 1;
1395                 export->exp_lock_replay_needed = 1;
1396                 spin_unlock(&export->exp_lock);
1397
1398                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1399                                  MSG_CONNECT_TRANSNO);
1400                 if (has_transno && transno == 0)
1401                         CWARN("Connect with zero transno!\n");
1402
1403                 if (has_transno && transno > 0 &&
1404                     transno < target->obd_next_recovery_transno &&
1405                     transno > target->obd_last_committed) {
1406                         /* Another way is to use cmpxchg() to be lock-free. */
1407                         spin_lock(&target->obd_recovery_task_lock);
1408                         if (transno < target->obd_next_recovery_transno)
1409                                 target->obd_next_recovery_transno = transno;
1410                         spin_unlock(&target->obd_recovery_task_lock);
1411                 }
1412
1413                 atomic_inc(&target->obd_req_replay_clients);
1414                 atomic_inc(&target->obd_lock_replay_clients);
1415                 /* Note: MDS-MDS connection is allowed to be connected during
1416                  * recovery, no matter if the exports needs to be recoveried.
1417                  * Because we need retrieve updates logs from all other MDTs.
1418                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1419                  * also needs to be increased to match other recovery checking
1420                  * condition. */
1421                 if (new_mds_mds_conn)
1422                         target->obd_max_recoverable_clients++;
1423                 if (atomic_inc_return(&target->obd_connected_clients) ==
1424                     target->obd_max_recoverable_clients)
1425                         wake_up(&target->obd_next_transno_waitq);
1426         }
1427
1428         /* Tell the client we're in recovery, when client is involved in it. */
1429         if (target->obd_recovering && !lw_client)
1430                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1431
1432 out:
1433         if (export) {
1434                 spin_lock(&export->exp_lock);
1435                 export->exp_connecting = 0;
1436                 spin_unlock(&export->exp_lock);
1437
1438                 class_export_put(export);
1439         }
1440         if (target_referenced == true && target != NULL) {
1441                 spin_lock(&target->obd_dev_lock);
1442                 target->obd_conn_inprogress--;
1443                 spin_unlock(&target->obd_dev_lock);
1444
1445                 class_decref(target, __func__, current);
1446         }
1447         req->rq_status = rc;
1448         RETURN(rc);
1449 }
1450
1451 int target_handle_disconnect(struct ptlrpc_request *req)
1452 {
1453         int rc;
1454         ENTRY;
1455
1456         rc = req_capsule_server_pack(&req->rq_pill);
1457         if (rc)
1458                 RETURN(rc);
1459
1460         /* Keep the rq_export around so we can send the reply. */
1461         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1462
1463         RETURN(0);
1464 }
1465
1466 void target_destroy_export(struct obd_export *exp)
1467 {
1468         struct obd_import       *imp = NULL;
1469         /* exports created from last_rcvd data, and "fake"
1470            exports created by lctl don't have an import */
1471         spin_lock(&exp->exp_lock);
1472         if (exp->exp_imp_reverse != NULL) {
1473                 imp = exp->exp_imp_reverse;
1474                 exp->exp_imp_reverse = NULL;
1475         }
1476         spin_unlock(&exp->exp_lock);
1477         if (imp != NULL)
1478                 client_destroy_import(imp);
1479
1480         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1481         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1482         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1483         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1484 }
1485 EXPORT_SYMBOL(target_destroy_export);
1486
1487 /*
1488  * Recovery functions
1489  */
1490 static void target_request_copy_get(struct ptlrpc_request *req)
1491 {
1492         class_export_rpc_inc(req->rq_export);
1493         LASSERT(list_empty(&req->rq_list));
1494         INIT_LIST_HEAD(&req->rq_replay_list);
1495
1496         /* Increase refcount to keep request in queue. */
1497         atomic_inc(&req->rq_refcount);
1498         /* Let export know it has replays to be handled. */
1499         atomic_inc(&req->rq_export->exp_replay_count);
1500 }
1501
1502 static void target_request_copy_put(struct ptlrpc_request *req)
1503 {
1504         LASSERT(list_empty(&req->rq_replay_list));
1505         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1506
1507         atomic_dec(&req->rq_export->exp_replay_count);
1508         class_export_rpc_dec(req->rq_export);
1509         ptlrpc_server_drop_request(req);
1510 }
1511
1512 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1513 {
1514         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1515         struct obd_export     *exp = req->rq_export;
1516         struct ptlrpc_request *reqiter;
1517         struct ptlrpc_request *dup_req = NULL;
1518         int                    dup = 0;
1519
1520         LASSERT(exp);
1521
1522         spin_lock(&exp->exp_lock);
1523         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1524                                 rq_replay_list) {
1525                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1526                         dup_req = reqiter;
1527                         dup = 1;
1528                         break;
1529                 }
1530         }
1531
1532         if (dup) {
1533                 /* We expect it with RESENT and REPLAY flags. */
1534                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1535                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1536                         CERROR("invalid flags %x of resent replay\n",
1537                                lustre_msg_get_flags(req->rq_reqmsg));
1538
1539                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1540                         __u32 new_conn;
1541
1542                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1543                         if (new_conn >
1544                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1545                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1546                                                         new_conn);
1547                 }
1548         } else {
1549                 list_add_tail(&req->rq_replay_list,
1550                                   &exp->exp_req_replay_queue);
1551         }
1552
1553         spin_unlock(&exp->exp_lock);
1554         return dup;
1555 }
1556
1557 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1558 {
1559         LASSERT(!list_empty(&req->rq_replay_list));
1560         LASSERT(req->rq_export);
1561
1562         spin_lock(&req->rq_export->exp_lock);
1563         list_del_init(&req->rq_replay_list);
1564         spin_unlock(&req->rq_export->exp_lock);
1565 }
1566
1567 static void target_finish_recovery(struct lu_target *lut)
1568 {
1569         struct obd_device *obd = lut->lut_obd;
1570         ENTRY;
1571
1572         /* Only log a recovery message when recovery has occurred. */
1573         if (obd->obd_recovery_start) {
1574                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1575                                         obd->obd_recovery_start);
1576                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1577                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1578                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1579                         obd->obd_max_recoverable_clients,
1580                         atomic_read(&obd->obd_connected_clients),
1581                         obd->obd_stale_clients,
1582                         obd->obd_stale_clients == 1 ? "was" : "were");
1583         }
1584
1585         ldlm_reprocess_all_ns(obd->obd_namespace);
1586         spin_lock(&obd->obd_recovery_task_lock);
1587         if (!list_empty(&obd->obd_req_replay_queue) ||
1588             !list_empty(&obd->obd_lock_replay_queue) ||
1589             !list_empty(&obd->obd_final_req_queue)) {
1590                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1591                        obd->obd_name,
1592                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1593                        list_empty(&obd->obd_lock_replay_queue) ? \
1594                                "" : "lock ",
1595                        list_empty(&obd->obd_final_req_queue) ? \
1596                                "" : "final ");
1597                 spin_unlock(&obd->obd_recovery_task_lock);
1598                 LBUG();
1599         }
1600         spin_unlock(&obd->obd_recovery_task_lock);
1601
1602         obd->obd_recovery_end = cfs_time_current_sec();
1603
1604         /* When recovery finished, cleanup orphans on MDS and OST. */
1605         if (OBT(obd) && OBP(obd, postrecov)) {
1606                 int rc = OBP(obd, postrecov)(obd);
1607                 if (rc < 0)
1608                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1609                                       obd->obd_name, rc);
1610         }
1611         EXIT;
1612 }
1613
1614 static void abort_req_replay_queue(struct obd_device *obd)
1615 {
1616         struct ptlrpc_request *req, *n;
1617         struct list_head abort_list;
1618
1619         INIT_LIST_HEAD(&abort_list);
1620         spin_lock(&obd->obd_recovery_task_lock);
1621         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1622         spin_unlock(&obd->obd_recovery_task_lock);
1623         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1624                 DEBUG_REQ(D_WARNING, req, "aborted:");
1625                 req->rq_status = -ENOTCONN;
1626                 if (ptlrpc_error(req)) {
1627                         DEBUG_REQ(D_ERROR, req,
1628                                   "failed abort_req_reply; skipping");
1629                 }
1630                 target_exp_dequeue_req_replay(req);
1631                 target_request_copy_put(req);
1632         }
1633 }
1634
1635 static void abort_lock_replay_queue(struct obd_device *obd)
1636 {
1637         struct ptlrpc_request *req, *n;
1638         struct list_head abort_list;
1639
1640         INIT_LIST_HEAD(&abort_list);
1641         spin_lock(&obd->obd_recovery_task_lock);
1642         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1643         spin_unlock(&obd->obd_recovery_task_lock);
1644         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1645                 DEBUG_REQ(D_ERROR, req, "aborted:");
1646                 req->rq_status = -ENOTCONN;
1647                 if (ptlrpc_error(req)) {
1648                         DEBUG_REQ(D_ERROR, req,
1649                                   "failed abort_lock_reply; skipping");
1650                 }
1651                 target_request_copy_put(req);
1652         }
1653 }
1654
1655 /* Called from a cleanup function if the device is being cleaned up
1656    forcefully.  The exports should all have been disconnected already,
1657    the only thing left to do is
1658      - clear the recovery flags
1659      - cancel the timer
1660      - free queued requests and replies, but don't send replies
1661    Because the obd_stopping flag is set, no new requests should be received.
1662
1663 */
1664 void target_cleanup_recovery(struct obd_device *obd)
1665 {
1666         struct ptlrpc_request *req, *n;
1667         struct list_head clean_list;
1668
1669         INIT_LIST_HEAD(&clean_list);
1670         spin_lock(&obd->obd_dev_lock);
1671         if (!obd->obd_recovering) {
1672                 spin_unlock(&obd->obd_dev_lock);
1673                 EXIT;
1674                 return;
1675         }
1676         obd->obd_recovering = obd->obd_abort_recovery = 0;
1677         spin_unlock(&obd->obd_dev_lock);
1678
1679         spin_lock(&obd->obd_recovery_task_lock);
1680         target_cancel_recovery_timer(obd);
1681         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1682         spin_unlock(&obd->obd_recovery_task_lock);
1683
1684         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1685                 LASSERT(req->rq_reply_state == NULL);
1686                 target_exp_dequeue_req_replay(req);
1687                 target_request_copy_put(req);
1688         }
1689
1690         spin_lock(&obd->obd_recovery_task_lock);
1691         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1692         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1693         spin_unlock(&obd->obd_recovery_task_lock);
1694
1695         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1696                 LASSERT(req->rq_reply_state == NULL);
1697                 target_request_copy_put(req);
1698         }
1699
1700         EXIT;
1701 }
1702 EXPORT_SYMBOL(target_cleanup_recovery);
1703
1704 /* obd_recovery_task_lock should be held */
1705 void target_cancel_recovery_timer(struct obd_device *obd)
1706 {
1707         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1708         cfs_timer_disarm(&obd->obd_recovery_timer);
1709 }
1710
1711 static void target_start_recovery_timer(struct obd_device *obd)
1712 {
1713         if (obd->obd_recovery_start != 0)
1714                 return;
1715
1716         spin_lock(&obd->obd_dev_lock);
1717         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1718                 spin_unlock(&obd->obd_dev_lock);
1719                 return;
1720         }
1721
1722         LASSERT(obd->obd_recovery_timeout != 0);
1723
1724         if (obd->obd_recovery_start != 0) {
1725                 spin_unlock(&obd->obd_dev_lock);
1726                 return;
1727         }
1728
1729         cfs_timer_arm(&obd->obd_recovery_timer,
1730                       cfs_time_shift(obd->obd_recovery_timeout));
1731         obd->obd_recovery_start = cfs_time_current_sec();
1732         spin_unlock(&obd->obd_dev_lock);
1733
1734         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1735                       "or until %d client%s reconnect%s\n",
1736                       obd->obd_name,
1737                       obd->obd_recovery_timeout / 60,
1738                       obd->obd_recovery_timeout % 60,
1739                       obd->obd_max_recoverable_clients,
1740                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1741                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1742 }
1743
1744 /**
1745  * extend recovery window.
1746  *
1747  * if @extend is true, extend recovery window to have @drt remaining at least;
1748  * otherwise, make sure the recovery timeout value is not less than @drt.
1749  */
1750 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1751 {
1752         cfs_time_t now;
1753         cfs_time_t end;
1754         cfs_duration_t left;
1755         int to;
1756
1757         spin_lock(&obd->obd_dev_lock);
1758         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1759                 spin_unlock(&obd->obd_dev_lock);
1760                 return;
1761         }
1762         LASSERT(obd->obd_recovery_start != 0);
1763
1764         now  = cfs_time_current_sec();
1765         to   = obd->obd_recovery_timeout;
1766         end  = obd->obd_recovery_start + to;
1767         left = cfs_time_sub(end, now);
1768
1769         if (extend && (drt > left)) {
1770                 to += drt - left;
1771         } else if (!extend && (drt > to)) {
1772                 to = drt;
1773         }
1774
1775         if (to > obd->obd_recovery_time_hard)
1776                 to = obd->obd_recovery_time_hard;
1777         if (obd->obd_recovery_timeout < to) {
1778                 obd->obd_recovery_timeout = to;
1779                 end = obd->obd_recovery_start + to;
1780                 cfs_timer_arm(&obd->obd_recovery_timer,
1781                                 cfs_time_shift(end - now));
1782         }
1783         spin_unlock(&obd->obd_dev_lock);
1784
1785         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1786                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1787 }
1788
1789 /* Reset the timer with each new client connection */
1790 /*
1791  * This timer is actually reconnect_timer, which is for making sure
1792  * the total recovery window is at least as big as my reconnect
1793  * attempt timing. So the initial recovery time_out will be set to
1794  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1795  * from client is bigger than this, then the recovery time_out will
1796  * be extended to make sure the client could be reconnected, in the
1797  * process, the timeout from the new client should be ignored.
1798  */
1799
1800 static void
1801 check_and_start_recovery_timer(struct obd_device *obd,
1802                                struct ptlrpc_request *req,
1803                                int new_client)
1804 {
1805         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1806         struct obd_device_target *obt = &obd->u.obt;
1807
1808         if (!new_client && service_time)
1809                 /* Teach server about old server's estimates, as first guess
1810                  * at how long new requests will take. */
1811                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1812                             service_time);
1813
1814         target_start_recovery_timer(obd);
1815
1816         /* Convert the service time to RPC timeout,
1817          * and reuse service_time to limit stack usage. */
1818         service_time = at_est2timeout(service_time);
1819
1820         /* We expect other clients to timeout within service_time, then try
1821          * to reconnect, then try the failover server.  The max delay between
1822          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1823         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1824
1825         LASSERT(obt->obt_magic == OBT_MAGIC);
1826         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1827         if (service_time > obd->obd_recovery_timeout && !new_client)
1828                 extend_recovery_timer(obd, service_time, false);
1829 }
1830
1831 /** Health checking routines */
1832 static inline int exp_connect_healthy(struct obd_export *exp)
1833 {
1834         return (exp->exp_in_recovery);
1835 }
1836
1837 /** if export done req_replay or has replay in queue */
1838 static inline int exp_req_replay_healthy(struct obd_export *exp)
1839 {
1840         return (!exp->exp_req_replay_needed ||
1841                 atomic_read(&exp->exp_replay_count) > 0);
1842 }
1843
1844
1845 static inline int exp_req_replay_healthy_or_from_mdt(struct obd_export *exp)
1846 {
1847         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1848                exp_req_replay_healthy(exp);
1849 }
1850
1851 /** if export done lock_replay or has replay in queue */
1852 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1853 {
1854         return (!exp->exp_lock_replay_needed ||
1855                 atomic_read(&exp->exp_replay_count) > 0);
1856 }
1857
1858 static inline int exp_vbr_healthy(struct obd_export *exp)
1859 {
1860         return (!exp->exp_vbr_failed);
1861 }
1862
1863 static inline int exp_finished(struct obd_export *exp)
1864 {
1865         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1866 }
1867
1868 static inline int exp_finished_or_from_mdt(struct obd_export *exp)
1869 {
1870         return (exp_connect_flags(exp) & OBD_CONNECT_MDS_MDS) ||
1871                 exp_finished(exp);
1872 }
1873
1874 static int check_for_next_transno(struct lu_target *lut)
1875 {
1876         struct ptlrpc_request *req = NULL;
1877         struct obd_device *obd = lut->lut_obd;
1878         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1879         int wake_up = 0, connected, completed, queue_len;
1880         __u64 req_transno = 0;
1881         __u64 update_transno = 0;
1882         __u64 next_transno = 0;
1883         ENTRY;
1884
1885         spin_lock(&obd->obd_recovery_task_lock);
1886         if (!list_empty(&obd->obd_req_replay_queue)) {
1887                 req = list_entry(obd->obd_req_replay_queue.next,
1888                                      struct ptlrpc_request, rq_list);
1889                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1890         }
1891
1892         if (tdtd != NULL)
1893                 update_transno = distribute_txn_get_next_transno(tdtd);
1894
1895         connected = atomic_read(&obd->obd_connected_clients);
1896         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1897         queue_len = obd->obd_requests_queued_for_recovery;
1898         next_transno = obd->obd_next_recovery_transno;
1899
1900         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1901                "req_transno: %llu, next_transno: %llu\n",
1902                obd->obd_max_recoverable_clients, connected, completed,
1903                queue_len, req_transno, next_transno);
1904
1905         if (obd->obd_abort_recovery) {
1906                 CDEBUG(D_HA, "waking for aborted recovery\n");
1907                 wake_up = 1;
1908         } else if (obd->obd_recovery_expired) {
1909                 CDEBUG(D_HA, "waking for expired recovery\n");
1910                 wake_up = 1;
1911         } else if (tdtd != NULL && req != NULL &&
1912                    is_req_replayed_by_update(req)) {
1913                 LASSERTF(req_transno < next_transno, "req_transno %llu"
1914                          "next_transno%llu\n", req_transno, next_transno);
1915                 CDEBUG(D_HA, "waking for duplicate req (%llu)\n",
1916                        req_transno);
1917                 wake_up = 1;
1918         } else if (req_transno == next_transno ||
1919                    (update_transno != 0 && update_transno <= next_transno)) {
1920                 CDEBUG(D_HA, "waking for next (%lld)\n", next_transno);
1921                 wake_up = 1;
1922         } else if (queue_len > 0 &&
1923                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1924                 /** handle gaps occured due to lost reply or VBR */
1925                 LASSERTF(req_transno >= next_transno,
1926                          "req_transno: %llu, next_transno: %llu\n",
1927                          req_transno, next_transno);
1928                 CDEBUG(D_HA,
1929                        "%s: waking for gap in transno, VBR is %s (skip: "
1930                        "%lld, ql: %d, comp: %d, conn: %d, next: %lld"
1931                        ", next_update %lld last_committed: %lld)\n",
1932                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1933                        next_transno, queue_len, completed, connected,
1934                        req_transno, update_transno, obd->obd_last_committed);
1935                 obd->obd_next_recovery_transno = req_transno;
1936                 wake_up = 1;
1937         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1938                 CDEBUG(D_HA, "waking for completed recovery\n");
1939                 wake_up = 1;
1940         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1941                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1942                        " by fail_lock, waking up (%lld)\n", next_transno);
1943                 obd->obd_next_recovery_transno = req_transno;
1944                 wake_up = 1;
1945         }
1946         spin_unlock(&obd->obd_recovery_task_lock);
1947         return wake_up;
1948 }
1949
1950 static int check_for_next_lock(struct lu_target *lut)
1951 {
1952         struct obd_device *obd = lut->lut_obd;
1953         int wake_up = 0;
1954
1955         spin_lock(&obd->obd_recovery_task_lock);
1956         if (!list_empty(&obd->obd_lock_replay_queue)) {
1957                 CDEBUG(D_HA, "waking for next lock\n");
1958                 wake_up = 1;
1959         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1960                 CDEBUG(D_HA, "waking for completed lock replay\n");
1961                 wake_up = 1;
1962         } else if (obd->obd_abort_recovery) {
1963                 CDEBUG(D_HA, "waking for aborted recovery\n");
1964                 wake_up = 1;
1965         } else if (obd->obd_recovery_expired) {
1966                 CDEBUG(D_HA, "waking for expired recovery\n");
1967                 wake_up = 1;
1968         }
1969         spin_unlock(&obd->obd_recovery_task_lock);
1970
1971         return wake_up;
1972 }
1973
1974 /**
1975  * wait for recovery events,
1976  * check its status with help of check_routine
1977  * evict dead clients via health_check
1978  */
1979 static int target_recovery_overseer(struct lu_target *lut,
1980                                     int (*check_routine)(struct lu_target *),
1981                                     int (*health_check)(struct obd_export *))
1982 {
1983         struct obd_device       *obd = lut->lut_obd;
1984         struct target_distribute_txn_data *tdtd;
1985 repeat:
1986         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1987               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1988                 __u64 next_update_transno = 0;
1989
1990                 /* Only abort the recovery if there are no update recovery
1991                  * left in the queue */
1992                 spin_lock(&obd->obd_recovery_task_lock);
1993                 if (lut->lut_tdtd != NULL) {
1994                         next_update_transno =
1995                                 distribute_txn_get_next_transno(lut->lut_tdtd);
1996
1997                         tdtd = lut->lut_tdtd;
1998                         /* If next_update_transno == 0, it probably because
1999                          * updatelog retrieve threads did not get any records
2000                          * yet, let's wait those threads stopped */
2001                         if (next_update_transno == 0) {
2002                                 struct l_wait_info lwi = { 0 };
2003
2004                                 l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2005                                        atomic_read(
2006                                        &tdtd->tdtd_recovery_threads_count) == 0,
2007                                        &lwi);
2008
2009                                 next_update_transno =
2010                                         distribute_txn_get_next_transno(
2011                                                                 lut->lut_tdtd);
2012                         }
2013                 }
2014
2015                 if (next_update_transno != 0 && !obd->obd_abort_recovery) {
2016                         obd->obd_next_recovery_transno = next_update_transno;
2017                         spin_unlock(&obd->obd_recovery_task_lock);
2018                         /* Disconnect unfinished exports from clients, and
2019                          * keep connection from MDT to make sure the update
2020                          * recovery will still keep trying until some one
2021                          * manually abort the recovery */
2022                         class_disconnect_stale_exports(obd,
2023                                                 exp_finished_or_from_mdt);
2024                         /* Abort all of replay and replay lock req from
2025                          * clients */
2026                         abort_req_replay_queue(obd);
2027                         abort_lock_replay_queue(obd);
2028                         CDEBUG(D_HA, "%s: there are still update replay (%#llx"
2029                                ")in the queue.\n", obd->obd_name,
2030                                next_update_transno);
2031                 } else {
2032                         obd->obd_abort_recovery = 1;
2033                         spin_unlock(&obd->obd_recovery_task_lock);
2034                         CWARN("%s recovery is aborted by hard timeout\n",
2035                               obd->obd_name);
2036                 }
2037         }
2038
2039         while (wait_event_timeout(obd->obd_next_transno_waitq,
2040                                   check_routine(lut),
2041                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
2042                 /* wait indefinitely for event, but don't trigger watchdog */;
2043
2044         if (obd->obd_abort_recovery) {
2045                 CWARN("recovery is aborted, evict exports in recovery\n");
2046                 if (lut->lut_tdtd != NULL) {
2047                         struct l_wait_info lwi = { 0 };
2048
2049                         tdtd = lut->lut_tdtd;
2050                         /* Let's wait all of the update log recovery thread
2051                          * finished */
2052                         l_wait_event(tdtd->tdtd_recovery_threads_waitq,
2053                          atomic_read(&tdtd->tdtd_recovery_threads_count) == 0,
2054                              &lwi);
2055                         /* Then abort the update recovery list */
2056                         dtrq_list_dump(lut->lut_tdtd, D_ERROR);
2057                         dtrq_list_destroy(lut->lut_tdtd);
2058                 }
2059
2060                 /** evict exports which didn't finish recovery yet */
2061                 class_disconnect_stale_exports(obd, exp_finished);
2062                 return 1;
2063         } else if (obd->obd_recovery_expired) {
2064                 obd->obd_recovery_expired = 0;
2065                 /** If some clients died being recovered, evict them */
2066                 LCONSOLE_WARN("%s: recovery is timed out, "
2067                               "evict stale exports\n", obd->obd_name);
2068                 /** evict cexports with no replay in queue, they are stalled */
2069                 class_disconnect_stale_exports(obd, health_check);
2070
2071                 /** continue with VBR */
2072                 spin_lock(&obd->obd_dev_lock);
2073                 obd->obd_version_recov = 1;
2074                 spin_unlock(&obd->obd_dev_lock);
2075                 /**
2076                  * reset timer, recovery will proceed with versions now,
2077                  * timeout is set just to handle reconnection delays
2078                  */
2079                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
2080                 /** Wait for recovery events again, after evicting bad clients */
2081                 goto repeat;
2082         }
2083         return 0;
2084 }
2085
2086 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
2087 {
2088         struct obd_device       *obd = lut->lut_obd;
2089         struct ptlrpc_request *req = NULL;
2090
2091         CDEBUG(D_HA, "Waiting for lock\n");
2092         if (target_recovery_overseer(lut, check_for_next_lock,
2093                                      exp_lock_replay_healthy))
2094                 abort_lock_replay_queue(obd);
2095
2096         spin_lock(&obd->obd_recovery_task_lock);
2097         if (!list_empty(&obd->obd_lock_replay_queue)) {
2098                 req = list_entry(obd->obd_lock_replay_queue.next,
2099                                      struct ptlrpc_request, rq_list);
2100                 list_del_init(&req->rq_list);
2101                 spin_unlock(&obd->obd_recovery_task_lock);
2102         } else {
2103                 spin_unlock(&obd->obd_recovery_task_lock);
2104                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
2105                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
2106                 /** evict exports failed VBR */
2107                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
2108         }
2109         return req;
2110 }
2111
2112 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2113 {
2114         struct ptlrpc_request *req = NULL;
2115
2116         spin_lock(&obd->obd_recovery_task_lock);
2117         if (!list_empty(&obd->obd_final_req_queue)) {
2118                 req = list_entry(obd->obd_final_req_queue.next,
2119                                      struct ptlrpc_request, rq_list);
2120                 list_del_init(&req->rq_list);
2121                 spin_unlock(&obd->obd_recovery_task_lock);
2122                 if (req->rq_export->exp_in_recovery) {
2123                         spin_lock(&req->rq_export->exp_lock);
2124                         req->rq_export->exp_in_recovery = 0;
2125                         spin_unlock(&req->rq_export->exp_lock);
2126                 }
2127         } else {
2128                 spin_unlock(&obd->obd_recovery_task_lock);
2129         }
2130         return req;
2131 }
2132
2133 static void handle_recovery_req(struct ptlrpc_thread *thread,
2134                                 struct ptlrpc_request *req,
2135                                 svc_handler_t handler)
2136 {
2137         ENTRY;
2138
2139         /**
2140          * export can be evicted during recovery, no need to handle replays for
2141          * it after that, discard such request silently
2142          */
2143         if (req->rq_export->exp_disconnected)
2144                 RETURN_EXIT;
2145
2146         req->rq_session.lc_thread = thread;
2147         req->rq_svc_thread = thread;
2148         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2149
2150         /* thread context */
2151         lu_context_enter(&thread->t_env->le_ctx);
2152         (void)handler(req);
2153         lu_context_exit(&thread->t_env->le_ctx);
2154
2155         /* don't reset timer for final stage */
2156         if (!exp_finished(req->rq_export)) {
2157                 int to = obd_timeout;
2158
2159                 /**
2160                  * Add request timeout to the recovery time so next request from
2161                  * this client may come in recovery time
2162                  */
2163                 if (!AT_OFF) {
2164                         struct ptlrpc_service_part *svcpt;
2165
2166                         svcpt = req->rq_rqbd->rqbd_svcpt;
2167                         /* If the server sent early reply for this request,
2168                          * the client will recalculate the timeout according to
2169                          * current server estimate service time, so we will
2170                          * use the maxium timeout here for waiting the client
2171                          * sending the next req */
2172                         to = max((int)at_est2timeout(
2173                                  at_get(&svcpt->scp_at_estimate)),
2174                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2175                         /* Add 2 net_latency, one for balance rq_deadline
2176                          * (see ptl_send_rpc), one for resend the req to server,
2177                          * Note: client will pack net_latency in replay req
2178                          * (see ptlrpc_replay_req) */
2179                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2180                 }
2181                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2182         }
2183         EXIT;
2184 }
2185
2186 /** Checking routines for recovery */
2187 static int check_for_recovery_ready(struct lu_target *lut)
2188 {
2189         struct obd_device *obd = lut->lut_obd;
2190         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2191
2192         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2193                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2194                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2195                obd->obd_recovery_expired);
2196
2197         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2198                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2199                 if (clnts + obd->obd_stale_clients <
2200                     obd->obd_max_recoverable_clients)
2201                         return 0;
2202         }
2203
2204         if (lut->lut_tdtd != NULL) {
2205                 if (!lut->lut_tdtd->tdtd_replay_ready &&
2206                     !obd->obd_abort_recovery) {
2207                         /* Let's extend recovery timer, in case the recovery
2208                          * timer expired, and some clients got evicted */
2209                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2210                                               true);
2211                         CDEBUG(D_HA, "%s update recovery is not ready,"
2212                                " extend recovery %d\n", obd->obd_name,
2213                                obd->obd_recovery_timeout);
2214                         return 0;
2215                 } else {
2216                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2217                 }
2218         }
2219
2220         return 1;
2221 }
2222
2223 enum {
2224         REQUEST_RECOVERY = 1,
2225         UPDATE_RECOVERY = 2,
2226 };
2227
2228 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2229 {
2230         __u64 transno = 0;
2231
2232         if (!list_empty(&obd->obd_req_replay_queue)) {
2233                 struct ptlrpc_request *req;
2234
2235                 req = list_entry(obd->obd_req_replay_queue.next,
2236                                  struct ptlrpc_request, rq_list);
2237                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2238         }
2239
2240         return transno;
2241 }
2242
2243 static __u64 get_next_transno(struct lu_target *lut, int *type)
2244 {
2245         struct obd_device *obd = lut->lut_obd;
2246         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2247         __u64 transno = 0;
2248         __u64 update_transno;
2249         ENTRY;
2250
2251         transno = get_next_replay_req_transno(obd);
2252         if (type != NULL)
2253                 *type = REQUEST_RECOVERY;
2254
2255         if (tdtd == NULL)
2256                 RETURN(transno);
2257
2258         update_transno = distribute_txn_get_next_transno(tdtd);
2259         if (transno == 0 || (transno >= update_transno &&
2260                              update_transno != 0)) {
2261                 transno = update_transno;
2262                 if (type != NULL)
2263                         *type = UPDATE_RECOVERY;
2264         }
2265
2266         RETURN(transno);
2267 }
2268
2269 /**
2270  * drop duplicate replay request
2271  *
2272  * Because the operation has been replayed by update recovery, the request
2273  * with the same transno will be dropped and also notify the client to send
2274  * next replay request.
2275  *
2276  * \param[in] env       execution environment
2277  * \param[in] obd       failover obd device
2278  * \param[in] req       request to be dropped
2279  */
2280 static void drop_duplicate_replay_req(struct lu_env *env,
2281                                       struct obd_device *obd,
2282                                       struct ptlrpc_request *req)
2283 {
2284         DEBUG_REQ(D_HA, req, "remove t%lld from %s because of duplicate"
2285                   " update records are found.\n",
2286                   lustre_msg_get_transno(req->rq_reqmsg),
2287                   libcfs_nid2str(req->rq_peer.nid));
2288
2289         /* Right now, only for MDS reint operation update replay and
2290          * normal request replay can have the same transno */
2291         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2292                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2293                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2294                 if (likely(req->rq_export))
2295                         target_committed_to_req(req);
2296                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2297                 target_send_reply(req, req->rq_status, 0);
2298         } else {
2299                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2300                 libcfs_nid2str(req->rq_peer.nid));
2301         }
2302         target_exp_dequeue_req_replay(req);
2303         target_request_copy_put(req);
2304         obd->obd_replayed_requests++;
2305 }
2306
2307 static void replay_request_or_update(struct lu_env *env,
2308                                      struct lu_target *lut,
2309                                      struct target_recovery_data *trd,
2310                                      struct ptlrpc_thread *thread)
2311 {
2312         struct obd_device *obd = lut->lut_obd;
2313         struct ptlrpc_request *req = NULL;
2314         int                     type;
2315         __u64                   transno;
2316         ENTRY;
2317
2318         CDEBUG(D_HA, "Waiting for transno %lld\n",
2319                obd->obd_next_recovery_transno);
2320
2321         /* Replay all of request and update by transno */
2322         do {
2323                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2324
2325                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2326
2327                 /** It is needed to extend recovery window above
2328                  *  recovery_time_soft. Extending is possible only in the
2329                  *  end of recovery window (see more details in
2330                  *  handle_recovery_req()).
2331                  */
2332                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2333
2334                 if (target_recovery_overseer(lut, check_for_next_transno,
2335                                         exp_req_replay_healthy_or_from_mdt)) {
2336                         abort_req_replay_queue(obd);
2337                         abort_lock_replay_queue(obd);
2338                         goto abort;
2339                 }
2340
2341                 spin_lock(&obd->obd_recovery_task_lock);
2342                 transno = get_next_transno(lut, &type);
2343                 if (type == REQUEST_RECOVERY && transno != 0) {
2344                         /* Drop replay request from client side, if the
2345                          * replay has been executed by update with the
2346                          * same transno */
2347                         req = list_entry(obd->obd_req_replay_queue.next,
2348                                         struct ptlrpc_request, rq_list);
2349
2350                         list_del_init(&req->rq_list);
2351                         obd->obd_requests_queued_for_recovery--;
2352                         spin_unlock(&obd->obd_recovery_task_lock);
2353
2354                         /* Let's check if the request has been redone by
2355                          * update replay */
2356                         if (is_req_replayed_by_update(req)) {
2357                                 struct distribute_txn_replay_req *dtrq;
2358
2359                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2360                                                                    req->rq_xid);
2361                                 LASSERT(dtrq != NULL);
2362                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2363                                 list_del_init(&dtrq->dtrq_list);
2364                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2365                                 dtrq_destroy(dtrq);
2366
2367                                 drop_duplicate_replay_req(env, obd, req);
2368
2369                                 continue;
2370                         }
2371
2372                         LASSERT(trd->trd_processing_task == current_pid());
2373                         DEBUG_REQ(D_HA, req, "processing t%lld from %s",
2374                                   lustre_msg_get_transno(req->rq_reqmsg),
2375                                   libcfs_nid2str(req->rq_peer.nid));
2376
2377                         handle_recovery_req(thread, req,
2378                                             trd->trd_recovery_handler);
2379                         /**
2380                          * bz18031: increase next_recovery_transno before
2381                          * target_request_copy_put() will drop exp_rpc reference
2382                          */
2383                         spin_lock(&obd->obd_recovery_task_lock);
2384                         obd->obd_next_recovery_transno++;
2385                         spin_unlock(&obd->obd_recovery_task_lock);
2386                         target_exp_dequeue_req_replay(req);
2387                         target_request_copy_put(req);
2388                         obd->obd_replayed_requests++;
2389                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2390                         struct distribute_txn_replay_req *dtrq;
2391                         int rc;
2392
2393                         spin_unlock(&obd->obd_recovery_task_lock);
2394
2395                         LASSERT(tdtd != NULL);
2396                         dtrq = distribute_txn_get_next_req(tdtd);
2397                         lu_context_enter(&thread->t_env->le_ctx);
2398                         rc = tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2399                         lu_context_exit(&thread->t_env->le_ctx);
2400                         extend_recovery_timer(obd, obd_timeout, true);
2401
2402                         if (rc == 0 && dtrq->dtrq_xid != 0) {
2403                                 CDEBUG(D_HA, "Move x%llu t%llu"
2404                                        " to finish list\n", dtrq->dtrq_xid,
2405                                        dtrq->dtrq_master_transno);
2406
2407                                 /* Add it to the replay finish list */
2408                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2409                                 list_add(&dtrq->dtrq_list,
2410                                          &tdtd->tdtd_replay_finish_list);
2411                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2412
2413                                 spin_lock(&obd->obd_recovery_task_lock);
2414                                 if (transno == obd->obd_next_recovery_transno)
2415                                         obd->obd_next_recovery_transno++;
2416                                 else if (transno >
2417                                          obd->obd_next_recovery_transno)
2418                                         obd->obd_next_recovery_transno =
2419                                                                 transno + 1;
2420                                 spin_unlock(&obd->obd_recovery_task_lock);
2421                         } else {
2422                                 dtrq_destroy(dtrq);
2423                         }
2424                 } else {
2425                         spin_unlock(&obd->obd_recovery_task_lock);
2426 abort:
2427                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2428                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2429                         /** evict exports failed VBR */
2430                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2431                         break;
2432                 }
2433         } while (1);
2434 }
2435
2436 static int target_recovery_thread(void *arg)
2437 {
2438         struct lu_target *lut = arg;
2439         struct obd_device *obd = lut->lut_obd;
2440         struct ptlrpc_request *req;
2441         struct target_recovery_data *trd = &obd->obd_recovery_data;
2442         unsigned long delta;
2443         struct lu_env *env;
2444         struct ptlrpc_thread *thread = NULL;
2445         int rc = 0;
2446         ENTRY;
2447
2448         unshare_fs_struct();
2449         OBD_ALLOC_PTR(thread);
2450         if (thread == NULL)
2451                 RETURN(-ENOMEM);
2452
2453         OBD_ALLOC_PTR(env);
2454         if (env == NULL) {
2455                 OBD_FREE_PTR(thread);
2456                 RETURN(-ENOMEM);
2457         }
2458
2459         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2460         if (rc) {
2461                 OBD_FREE_PTR(thread);
2462                 OBD_FREE_PTR(env);
2463                 RETURN(rc);
2464         }
2465
2466         thread->t_env = env;
2467         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2468         env->le_ctx.lc_thread = thread;
2469         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2470         thread->t_watchdog = NULL;
2471
2472         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2473                current_pid());
2474         trd->trd_processing_task = current_pid();
2475
2476         spin_lock(&obd->obd_dev_lock);
2477         obd->obd_recovering = 1;
2478         spin_unlock(&obd->obd_dev_lock);
2479         complete(&trd->trd_starting);
2480
2481         /* first of all, we have to know the first transno to replay */
2482         if (target_recovery_overseer(lut, check_for_recovery_ready,
2483                                      exp_connect_healthy)) {
2484                 abort_req_replay_queue(obd);
2485                 abort_lock_replay_queue(obd);
2486                 if (lut->lut_tdtd != NULL)
2487                         dtrq_list_destroy(lut->lut_tdtd);
2488         }
2489
2490         /* next stage: replay requests or update */
2491         delta = jiffies;
2492         CDEBUG(D_INFO, "1: request replay stage - %d clients from t%llu\n",
2493                atomic_read(&obd->obd_req_replay_clients),
2494                obd->obd_next_recovery_transno);
2495         replay_request_or_update(env, lut, trd, thread);
2496
2497         /**
2498          * The second stage: replay locks
2499          */
2500         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2501                atomic_read(&obd->obd_lock_replay_clients));
2502         while ((req = target_next_replay_lock(lut))) {
2503                 LASSERT(trd->trd_processing_task == current_pid());
2504                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2505                           libcfs_nid2str(req->rq_peer.nid));
2506                 handle_recovery_req(thread, req,
2507                                     trd->trd_recovery_handler);
2508                 target_request_copy_put(req);
2509                 obd->obd_replayed_locks++;
2510         }
2511
2512         /**
2513          * The third stage: reply on final pings, at this moment all clients
2514          * must have request in final queue
2515          */
2516         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2517         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2518         /** Update server last boot epoch */
2519         tgt_boot_epoch_update(lut);
2520         /* We drop recoverying flag to forward all new requests
2521          * to regular mds_handle() since now */
2522         spin_lock(&obd->obd_dev_lock);
2523         obd->obd_recovering = obd->obd_abort_recovery = 0;
2524         spin_unlock(&obd->obd_dev_lock);
2525         spin_lock(&obd->obd_recovery_task_lock);
2526         target_cancel_recovery_timer(obd);
2527         spin_unlock(&obd->obd_recovery_task_lock);
2528         while ((req = target_next_final_ping(obd))) {
2529                 LASSERT(trd->trd_processing_task == current_pid());
2530                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2531                           libcfs_nid2str(req->rq_peer.nid));
2532                 handle_recovery_req(thread, req,
2533                                     trd->trd_recovery_handler);
2534                 /* Because the waiting client can not send ping to server,
2535                  * so we need refresh the last_request_time, to avoid the
2536                  * export is being evicted */
2537                 ptlrpc_update_export_timer(req->rq_export, 0);
2538                 target_request_copy_put(req);
2539         }
2540
2541         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2542         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2543               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2544         if (delta > OBD_RECOVERY_TIME_SOFT) {
2545                 CWARN("too long recovery - read logs\n");
2546                 libcfs_debug_dumplog();
2547         }
2548
2549         target_finish_recovery(lut);
2550
2551         lu_context_fini(&env->le_ctx);
2552         trd->trd_processing_task = 0;
2553         complete(&trd->trd_finishing);
2554
2555         tgt_io_thread_done(thread);
2556         OBD_FREE_PTR(thread);
2557         OBD_FREE_PTR(env);
2558         RETURN(rc);
2559 }
2560
2561 static int target_start_recovery_thread(struct lu_target *lut,
2562                                         svc_handler_t handler)
2563 {
2564         struct obd_device *obd = lut->lut_obd;
2565         int rc = 0;
2566         struct target_recovery_data *trd = &obd->obd_recovery_data;
2567         int index;
2568
2569         memset(trd, 0, sizeof(*trd));
2570         init_completion(&trd->trd_starting);
2571         init_completion(&trd->trd_finishing);
2572         trd->trd_recovery_handler = handler;
2573
2574         rc = server_name2index(obd->obd_name, &index, NULL);
2575         if (rc < 0)
2576                 return rc;
2577
2578         if (!IS_ERR(kthread_run(target_recovery_thread,
2579                                 lut, "tgt_recover_%d", index))) {
2580                 wait_for_completion(&trd->trd_starting);
2581                 LASSERT(obd->obd_recovering != 0);
2582         } else {
2583                 rc = -ECHILD;
2584         }
2585
2586         return rc;
2587 }
2588
2589 void target_stop_recovery_thread(struct obd_device *obd)
2590 {
2591         if (obd->obd_recovery_data.trd_processing_task > 0) {
2592                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2593                 /** recovery can be done but postrecovery is not yet */
2594                 spin_lock(&obd->obd_dev_lock);
2595                 if (obd->obd_recovering) {
2596                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2597                         obd->obd_abort_recovery = 1;
2598                         wake_up(&obd->obd_next_transno_waitq);
2599                 }
2600                 spin_unlock(&obd->obd_dev_lock);
2601                 wait_for_completion(&trd->trd_finishing);
2602         }
2603 }
2604 EXPORT_SYMBOL(target_stop_recovery_thread);
2605
2606 void target_recovery_fini(struct obd_device *obd)
2607 {
2608         class_disconnect_exports(obd);
2609         target_stop_recovery_thread(obd);
2610         target_cleanup_recovery(obd);
2611 }
2612 EXPORT_SYMBOL(target_recovery_fini);
2613
2614 static void target_recovery_expired(unsigned long castmeharder)
2615 {
2616         struct obd_device *obd = (struct obd_device *)castmeharder;
2617         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2618                " after %lds (%d clients connected)\n",
2619                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2620                cfs_time_current_sec()- obd->obd_recovery_start,
2621                atomic_read(&obd->obd_connected_clients));
2622
2623         obd->obd_recovery_expired = 1;
2624         wake_up(&obd->obd_next_transno_waitq);
2625 }
2626
2627 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2628 {
2629         struct obd_device *obd = lut->lut_obd;
2630
2631         if (obd->obd_max_recoverable_clients == 0) {
2632                 /** Update server last boot epoch */
2633                 tgt_boot_epoch_update(lut);
2634                 return;
2635         }
2636
2637         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2638                "last_transno %llu\n", obd->obd_name,
2639                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2640         LASSERT(obd->obd_stopping == 0);
2641         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2642         obd->obd_recovery_start = 0;
2643         obd->obd_recovery_end = 0;
2644
2645         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2646         target_start_recovery_thread(lut, handler);
2647 }
2648 EXPORT_SYMBOL(target_recovery_init);
2649
2650 static int target_process_req_flags(struct obd_device *obd,
2651                                     struct ptlrpc_request *req)
2652 {
2653         struct obd_export *exp = req->rq_export;
2654         LASSERT(exp != NULL);
2655         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2656                 /* client declares he's ready to replay locks */
2657                 spin_lock(&exp->exp_lock);
2658                 if (exp->exp_req_replay_needed) {
2659                         exp->exp_req_replay_needed = 0;
2660                         spin_unlock(&exp->exp_lock);
2661
2662                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2663                         atomic_dec(&obd->obd_req_replay_clients);
2664                 } else {
2665                         spin_unlock(&exp->exp_lock);
2666                 }
2667         }
2668         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2669                 /* client declares he's ready to complete recovery
2670                  * so, we put the request on th final queue */
2671                 spin_lock(&exp->exp_lock);
2672                 if (exp->exp_lock_replay_needed) {
2673                         exp->exp_lock_replay_needed = 0;
2674                         spin_unlock(&exp->exp_lock);
2675
2676                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2677                         atomic_dec(&obd->obd_lock_replay_clients);
2678                 } else {
2679                         spin_unlock(&exp->exp_lock);
2680                 }
2681         }
2682         return 0;
2683 }
2684
2685 int target_queue_recovery_request(struct ptlrpc_request *req,
2686                                   struct obd_device *obd)
2687 {
2688         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2689         struct ptlrpc_request *reqiter;
2690         int inserted = 0;
2691         ENTRY;
2692
2693         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2694                 /* Processing the queue right now, don't re-add. */
2695                 RETURN(1);
2696         }
2697
2698         target_process_req_flags(obd, req);
2699
2700         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2701                 /* client declares he's ready to complete recovery
2702                  * so, we put the request on th final queue */
2703                 target_request_copy_get(req);
2704                 DEBUG_REQ(D_HA, req, "queue final req");
2705                 wake_up(&obd->obd_next_transno_waitq);
2706                 spin_lock(&obd->obd_recovery_task_lock);
2707                 if (obd->obd_recovering) {
2708                         struct ptlrpc_request *tmp;
2709                         struct ptlrpc_request *duplicate = NULL;
2710
2711                         if (likely(!req->rq_export->exp_replay_done)) {
2712                                 req->rq_export->exp_replay_done = 1;
2713                                 list_add_tail(&req->rq_list,
2714                                               &obd->obd_final_req_queue);
2715                                 spin_unlock(&obd->obd_recovery_task_lock);
2716                                 RETURN(0);
2717                         }
2718
2719                         /* XXX O(n), but only happens if final ping is
2720                          * timed out, probably reorganize the list as
2721                          * a hash list later */
2722                         list_for_each_entry_safe(reqiter, tmp,
2723                                                  &obd->obd_final_req_queue,
2724                                                  rq_list) {
2725                                 if (reqiter->rq_export == req->rq_export) {
2726                                         list_del_init(&reqiter->rq_list);
2727                                         duplicate = reqiter;
2728                                         break;
2729                                 }
2730                         }
2731
2732                         list_add_tail(&req->rq_list,
2733                                       &obd->obd_final_req_queue);
2734                         req->rq_export->exp_replay_done = 1;
2735                         spin_unlock(&obd->obd_recovery_task_lock);
2736
2737                         if (duplicate != NULL) {
2738                                 DEBUG_REQ(D_HA, duplicate,
2739                                           "put prev final req\n");
2740                                 target_request_copy_put(duplicate);
2741                         }
2742                         RETURN(0);
2743                 } else {
2744                         spin_unlock(&obd->obd_recovery_task_lock);
2745                         target_request_copy_put(req);
2746                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2747                 }
2748         }
2749         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2750                 /* client declares he's ready to replay locks */
2751                 target_request_copy_get(req);
2752                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2753                 wake_up(&obd->obd_next_transno_waitq);
2754                 spin_lock(&obd->obd_recovery_task_lock);
2755                 LASSERT(obd->obd_recovering);
2756                 /* usually due to recovery abort */
2757                 if (!req->rq_export->exp_in_recovery) {
2758                         spin_unlock(&obd->obd_recovery_task_lock);
2759                         target_request_copy_put(req);
2760                         RETURN(-ENOTCONN);
2761                 }
2762                 LASSERT(req->rq_export->exp_lock_replay_needed);
2763                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2764                 spin_unlock(&obd->obd_recovery_task_lock);
2765                 RETURN(0);
2766         }
2767
2768         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2769          * (i.e. buflens etc are in my own byte order), but type-dependent
2770          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2771
2772         if (!transno) {
2773                 INIT_LIST_HEAD(&req->rq_list);
2774                 DEBUG_REQ(D_HA, req, "not queueing");
2775                 RETURN(1);
2776         }
2777
2778         /* If we're processing the queue, we want don't want to queue this
2779          * message.
2780          *
2781          * Also, if this request has a transno less than the one we're waiting
2782          * for, we should process it now.  It could (and currently always will)
2783          * be an open request for a descriptor that was opened some time ago.
2784          *
2785          * Also, a resent, replayed request that has already been
2786          * handled will pass through here and be processed immediately.
2787          */
2788         CDEBUG(D_HA, "Next recovery transno: %llu"
2789                ", current: %llu, replaying\n",
2790                obd->obd_next_recovery_transno, transno);
2791
2792         /* If the request has been replayed by update replay, then sends this
2793          * request to the recovery thread (replay_request_or_update()), where
2794          * it will be handled */
2795         spin_lock(&obd->obd_recovery_task_lock);
2796         if (transno < obd->obd_next_recovery_transno &&
2797             !is_req_replayed_by_update(req)) {
2798                 /* Processing the queue right now, don't re-add. */
2799                 LASSERT(list_empty(&req->rq_list));
2800                 spin_unlock(&obd->obd_recovery_task_lock);
2801                 RETURN(1);
2802         }
2803         spin_unlock(&obd->obd_recovery_task_lock);
2804
2805         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2806                 RETURN(0);
2807
2808         target_request_copy_get(req);
2809         if (!req->rq_export->exp_in_recovery) {
2810                 target_request_copy_put(req);
2811                 RETURN(-ENOTCONN);
2812         }
2813         LASSERT(req->rq_export->exp_req_replay_needed);
2814
2815         if (target_exp_enqueue_req_replay(req)) {
2816                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2817                 target_request_copy_put(req);
2818                 RETURN(0);
2819         }
2820
2821         /* XXX O(n^2) */
2822         spin_lock(&obd->obd_recovery_task_lock);
2823         LASSERT(obd->obd_recovering);
2824         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2825                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2826                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2827                         inserted = 1;
2828                         goto added;
2829                 }
2830
2831                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2832                              transno)) {
2833                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2834                                   "has been claimed by another client");
2835                         spin_unlock(&obd->obd_recovery_task_lock);
2836                         target_exp_dequeue_req_replay(req);
2837                         target_request_copy_put(req);
2838                         RETURN(0);
2839                 }
2840         }
2841 added:
2842         if (!inserted)
2843                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2844
2845         obd->obd_requests_queued_for_recovery++;
2846         spin_unlock(&obd->obd_recovery_task_lock);
2847         wake_up(&obd->obd_next_transno_waitq);
2848         RETURN(0);
2849 }
2850
2851 int target_handle_ping(struct ptlrpc_request *req)
2852 {
2853         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2854         return req_capsule_server_pack(&req->rq_pill);
2855 }
2856
2857 void target_committed_to_req(struct ptlrpc_request *req)
2858 {
2859         struct obd_export *exp = req->rq_export;
2860
2861         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2862                 lustre_msg_set_last_committed(req->rq_repmsg,
2863                                               exp->exp_last_committed);
2864         else
2865                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2866                           "%d)", exp->exp_obd->obd_no_transno,
2867                           req->rq_repmsg == NULL);
2868
2869         CDEBUG(D_INFO, "last_committed %llu, transno %llu, xid %llu\n",
2870                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2871 }
2872
2873 #endif /* HAVE_SERVER_SUPPORT */
2874
2875 /**
2876  * Packs current SLV and Limit into \a req.
2877  */
2878 int target_pack_pool_reply(struct ptlrpc_request *req)
2879 {
2880         struct obd_device *obd;
2881         ENTRY;
2882
2883         /* Check that we still have all structures alive as this may
2884          * be some late RPC at shutdown time. */
2885         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2886                      !exp_connect_lru_resize(req->rq_export))) {
2887                 lustre_msg_set_slv(req->rq_repmsg, 0);
2888                 lustre_msg_set_limit(req->rq_repmsg, 0);
2889                 RETURN(0);
2890         }
2891
2892         /* OBD is alive here as export is alive, which we checked above. */
2893         obd = req->rq_export->exp_obd;
2894
2895         read_lock(&obd->obd_pool_lock);
2896         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2897         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2898         read_unlock(&obd->obd_pool_lock);
2899
2900         RETURN(0);
2901 }
2902
2903 static int target_send_reply_msg(struct ptlrpc_request *req,
2904                                  int rc, int fail_id)
2905 {
2906         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2907                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2908                 return -ECOMM;
2909         }
2910         /* We can have a null rq_reqmsg in the event of bad signature or
2911          * no context when unwrapping */
2912         if (req->rq_reqmsg &&
2913             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2914             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2915                 return -ECOMM;
2916
2917         if (unlikely(rc)) {
2918                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2919                 req->rq_status = rc;
2920                 return ptlrpc_send_error(req, 1);
2921         } else {
2922                 DEBUG_REQ(D_NET, req, "sending reply");
2923         }
2924
2925         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2926 }
2927
2928 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2929 {
2930         struct ptlrpc_service_part *svcpt;
2931         int                        netrc;
2932         struct ptlrpc_reply_state *rs;
2933         struct obd_export         *exp;
2934         ENTRY;
2935
2936         if (req->rq_no_reply) {
2937                 EXIT;
2938                 return;
2939         }
2940
2941         svcpt = req->rq_rqbd->rqbd_svcpt;
2942         rs = req->rq_reply_state;
2943         if (rs == NULL || !rs->rs_difficult) {
2944                 /* no notifiers */
2945                 target_send_reply_msg (req, rc, fail_id);
2946                 EXIT;
2947                 return;
2948         }
2949
2950         /* must be an export if locks saved */
2951         LASSERT(req->rq_export != NULL);
2952         /* req/reply consistent */
2953         LASSERT(rs->rs_svcpt == svcpt);
2954
2955         /* "fresh" reply */
2956         LASSERT(!rs->rs_scheduled);
2957         LASSERT(!rs->rs_scheduled_ever);
2958         LASSERT(!rs->rs_handled);
2959         LASSERT(!rs->rs_on_net);
2960         LASSERT(rs->rs_export == NULL);
2961         LASSERT(list_empty(&rs->rs_obd_list));
2962         LASSERT(list_empty(&rs->rs_exp_list));
2963
2964         exp = class_export_get(req->rq_export);
2965
2966         /* disable reply scheduling while I'm setting up */
2967         rs->rs_scheduled = 1;
2968         rs->rs_on_net    = 1;
2969         rs->rs_xid       = req->rq_xid;
2970         rs->rs_transno   = req->rq_transno;
2971         rs->rs_export    = exp;
2972         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2973
2974         spin_lock(&exp->exp_uncommitted_replies_lock);
2975         CDEBUG(D_NET, "rs transno = %llu, last committed = %llu\n",
2976                rs->rs_transno, exp->exp_last_committed);
2977         if (rs->rs_transno > exp->exp_last_committed) {
2978                 /* not committed already */
2979                 list_add_tail(&rs->rs_obd_list,
2980                                   &exp->exp_uncommitted_replies);
2981         }
2982         spin_unlock(&exp->exp_uncommitted_replies_lock);
2983
2984         spin_lock(&exp->exp_lock);
2985         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2986         spin_unlock(&exp->exp_lock);
2987
2988         netrc = target_send_reply_msg(req, rc, fail_id);
2989
2990         spin_lock(&svcpt->scp_rep_lock);
2991
2992         atomic_inc(&svcpt->scp_nreps_difficult);
2993
2994         if (netrc != 0) {
2995                 /* error sending: reply is off the net.  Also we need +1
2996                  * reply ref until ptlrpc_handle_rs() is done
2997                  * with the reply state (if the send was successful, there
2998                  * would have been +1 ref for the net, which
2999                  * reply_out_callback leaves alone) */
3000                 rs->rs_on_net = 0;
3001                 ptlrpc_rs_addref(rs);
3002         }
3003
3004         spin_lock(&rs->rs_lock);
3005         if (rs->rs_transno <= exp->exp_last_committed ||
3006             (!rs->rs_on_net && !rs->rs_no_ack) ||
3007             list_empty(&rs->rs_exp_list) ||     /* completed already */
3008             list_empty(&rs->rs_obd_list)) {
3009                 CDEBUG(D_HA, "Schedule reply immediately\n");
3010                 ptlrpc_dispatch_difficult_reply(rs);
3011         } else {
3012                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
3013                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
3014         }
3015         spin_unlock(&rs->rs_lock);
3016         spin_unlock(&svcpt->scp_rep_lock);
3017         EXIT;
3018 }
3019
3020 enum ldlm_mode lck_compat_array[] = {
3021         [LCK_EX]    = LCK_COMPAT_EX,
3022         [LCK_PW]    = LCK_COMPAT_PW,
3023         [LCK_PR]    = LCK_COMPAT_PR,
3024         [LCK_CW]    = LCK_COMPAT_CW,
3025         [LCK_CR]    = LCK_COMPAT_CR,
3026         [LCK_NL]    = LCK_COMPAT_NL,
3027         [LCK_GROUP] = LCK_COMPAT_GROUP,
3028         [LCK_COS]   = LCK_COMPAT_COS,
3029 };
3030
3031 /**
3032  * Rather arbitrary mapping from LDLM error codes to errno values. This should
3033  * not escape to the user level.
3034  */
3035 int ldlm_error2errno(enum ldlm_error error)
3036 {
3037         int result;
3038
3039         switch (error) {
3040         case ELDLM_OK:
3041         case ELDLM_LOCK_MATCHED:
3042                 result = 0;
3043                 break;
3044         case ELDLM_LOCK_CHANGED:
3045                 result = -ESTALE;
3046                 break;
3047         case ELDLM_LOCK_ABORTED:
3048                 result = -ENAVAIL;
3049                 break;
3050         case ELDLM_LOCK_REPLACED:
3051                 result = -ESRCH;
3052                 break;
3053         case ELDLM_NO_LOCK_DATA:
3054                 result = -ENOENT;
3055                 break;
3056         case ELDLM_NAMESPACE_EXISTS:
3057                 result = -EEXIST;
3058                 break;
3059         case ELDLM_BAD_NAMESPACE:
3060                 result = -EBADF;
3061                 break;
3062         default:
3063                 if (((int)error) < 0) { /* cast to signed type */
3064                         result = error; /* as ldlm_error can be unsigned */
3065                 } else {
3066                         CERROR("Invalid DLM result code: %d\n", error);
3067                         result = -EPROTO;
3068                 }
3069         }
3070         return result;
3071 }
3072 EXPORT_SYMBOL(ldlm_error2errno);
3073
3074 /**
3075  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
3076  */
3077 enum ldlm_error ldlm_errno2error(int err_no)
3078 {
3079         int error;
3080
3081         switch (err_no) {
3082         case 0:
3083                 error = ELDLM_OK;
3084                 break;
3085         case -ESTALE:
3086                 error = ELDLM_LOCK_CHANGED;
3087                 break;
3088         case -ENAVAIL:
3089                 error = ELDLM_LOCK_ABORTED;
3090                 break;
3091         case -ESRCH:
3092                 error = ELDLM_LOCK_REPLACED;
3093                 break;
3094         case -ENOENT:
3095                 error = ELDLM_NO_LOCK_DATA;
3096                 break;
3097         case -EEXIST:
3098                 error = ELDLM_NAMESPACE_EXISTS;
3099                 break;
3100         case -EBADF:
3101                 error = ELDLM_BAD_NAMESPACE;
3102                 break;
3103         default:
3104                 error = err_no;
3105         }
3106         return error;
3107 }
3108
3109 #if LUSTRE_TRACKS_LOCK_EXP_REFS
3110 void ldlm_dump_export_locks(struct obd_export *exp)
3111 {
3112         spin_lock(&exp->exp_locks_list_guard);
3113         if (!list_empty(&exp->exp_locks_list)) {
3114                 struct ldlm_lock *lock;
3115
3116                 CERROR("dumping locks for export %p,"
3117                        "ignore if the unmount doesn't hang\n", exp);
3118                 list_for_each_entry(lock, &exp->exp_locks_list,
3119                                         l_exp_refs_link)
3120                         LDLM_ERROR(lock, "lock:");
3121         }
3122         spin_unlock(&exp->exp_locks_list_guard);
3123 }
3124 #endif
3125
3126 #ifdef HAVE_SERVER_SUPPORT
3127 static int target_bulk_timeout(void *data)
3128 {
3129         ENTRY;
3130         /* We don't fail the connection here, because having the export
3131          * killed makes the (vital) call to commitrw very sad.
3132          */
3133         RETURN(1);
3134 }
3135
3136 static inline const char *bulk2type(struct ptlrpc_request *req)
3137 {
3138         if (req->rq_bulk_read)
3139                 return "READ";
3140         if (req->rq_bulk_write)
3141                 return "WRITE";
3142         return "UNKNOWN";
3143 }
3144
3145 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
3146                    struct l_wait_info *lwi)
3147 {
3148         struct ptlrpc_request   *req = desc->bd_req;
3149         time_t                   start = cfs_time_current_sec();
3150         time_t                   deadline;
3151         int                      rc = 0;
3152
3153         ENTRY;
3154
3155         /* If there is eviction in progress, wait for it to finish. */
3156         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3157                 *lwi = LWI_INTR(NULL, NULL);
3158                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3159                                   !atomic_read(&exp->exp_obd->
3160                                                    obd_evict_inprogress),
3161                                   lwi);
3162         }
3163
3164         /* Check if client was evicted or reconnected already. */
3165         if (exp->exp_failed ||
3166             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3167                 rc = -ENOTCONN;
3168         } else {
3169                 if (req->rq_bulk_read)
3170                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3171
3172                 if (OCD_HAS_FLAG(&exp->exp_connect_data, BULK_MBITS))
3173                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3174                 else /* old version, bulk matchbits is rq_xid */
3175                         req->rq_mbits = req->rq_xid;
3176
3177                 if (rc == 0)
3178                         rc = ptlrpc_start_bulk_transfer(desc);
3179         }
3180
3181         if (rc < 0) {
3182                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3183                           bulk2type(req), rc);
3184                 RETURN(rc);
3185         }
3186
3187         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3188                 ptlrpc_abort_bulk(desc);
3189                 RETURN(0);
3190         }
3191
3192         /* limit actual bulk transfer to bulk_timeout seconds */
3193         deadline = start + bulk_timeout;
3194         if (deadline > req->rq_deadline)
3195                 deadline = req->rq_deadline;
3196
3197         do {
3198                 long timeoutl = deadline - cfs_time_current_sec();
3199                 cfs_duration_t timeout = timeoutl <= 0 ?
3200                                          CFS_TICK : cfs_time_seconds(timeoutl);
3201                 time_t  rq_deadline;
3202
3203                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3204                                             target_bulk_timeout, desc);
3205                 rc = l_wait_event(desc->bd_waitq,
3206                                   !ptlrpc_server_bulk_active(desc) ||
3207                                   exp->exp_failed ||
3208                                   exp->exp_conn_cnt >
3209                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3210                                   lwi);
3211                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3212                 /* Wait again if we changed rq_deadline. */
3213                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3214                 deadline = start + bulk_timeout;
3215                 if (deadline > rq_deadline)
3216                         deadline = rq_deadline;
3217         } while ((rc == -ETIMEDOUT) &&
3218                  (deadline > cfs_time_current_sec()));
3219
3220         if (rc == -ETIMEDOUT) {
3221                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3222                           bulk2type(req), deadline - start,
3223                           cfs_time_current_sec() - deadline);
3224                 ptlrpc_abort_bulk(desc);
3225         } else if (exp->exp_failed) {
3226                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3227                           bulk2type(req));
3228                 rc = -ENOTCONN;
3229                 ptlrpc_abort_bulk(desc);
3230         } else if (exp->exp_conn_cnt >
3231                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3232                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3233                           bulk2type(req));
3234                 /* We don't reply anyway. */
3235                 rc = -ETIMEDOUT;
3236                 ptlrpc_abort_bulk(desc);
3237         } else if (desc->bd_failure) {
3238                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3239                           bulk2type(req));
3240                 /* XXX should this be a different errno? */
3241                 rc = -ETIMEDOUT;
3242         } else {
3243                 if (req->rq_bulk_write)
3244                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3245                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3246                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3247                                   bulk2type(req), desc->bd_nob_transferred,
3248                                   desc->bd_nob);
3249                         /* XXX should this be a different errno? */
3250                         rc = -ETIMEDOUT;
3251                 }
3252         }
3253
3254         RETURN(rc);
3255 }
3256 EXPORT_SYMBOL(target_bulk_io);
3257
3258 #endif /* HAVE_SERVER_SUPPORT */