lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 /**
  38  * This file deals with various client/target related logic including recovery.
  39  *
  40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  41  * should be moved.
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LDLM
  45
  46 #include <linux/kthread.h>
  47 #include <libcfs/libcfs.h>
  48 #include <obd.h>
  49 #include <obd_class.h>
  50 #include <lustre_dlm.h>
  51 #include <lustre_net.h>
  52 #include <lustre_sec.h>
  53 #include "ldlm_internal.h"
  54
  55 /* @priority: If non-zero, move the selected connection to the list head.
  56  * @create: If zero, only search in existing connections.
  57  */
  58 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  59                            int priority, int create)
  60 {
  61         struct ptlrpc_connection *ptlrpc_conn;
  62         struct obd_import_conn *imp_conn = NULL, *item;
  63         int rc = 0;
  64         ENTRY;
  65
  66         if (!create && !priority) {
  67                 CDEBUG(D_HA, "Nothing to do\n");
  68                 RETURN(-EINVAL);
  69         }
  70
  71         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  72         if (!ptlrpc_conn) {
  73                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  74                 RETURN (-ENOENT);
  75         }
  76
  77         if (create) {
  78                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  79                 if (!imp_conn) {
  80                         GOTO(out_put, rc = -ENOMEM);
  81                 }
  82         }
  83
  84         spin_lock(&imp->imp_lock);
  85         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  86                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  87                         if (priority) {
  88                                 list_del(&item->oic_item);
  89                                 list_add(&item->oic_item,
  90                                              &imp->imp_conn_list);
  91                                 item->oic_last_attempt = 0;
  92                         }
  93                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  94                                imp, imp->imp_obd->obd_name, uuid->uuid,
  95                                (priority ? ", moved to head" : ""));
  96                         spin_unlock(&imp->imp_lock);
  97                         GOTO(out_free, rc = 0);
  98                 }
  99         }
 100         /* No existing import connection found for \a uuid. */
 101         if (create) {
 102                 imp_conn->oic_conn = ptlrpc_conn;
 103                 imp_conn->oic_uuid = *uuid;
 104                 imp_conn->oic_last_attempt = 0;
 105                 if (priority)
 106                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 else
 108                         list_add_tail(&imp_conn->oic_item,
 109                                           &imp->imp_conn_list);
 110                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 111                        imp, imp->imp_obd->obd_name, uuid->uuid,
 112                        (priority ? "head" : "tail"));
 113         } else {
 114                 spin_unlock(&imp->imp_lock);
 115                 GOTO(out_free, rc = -ENOENT);
 116         }
 117
 118         spin_unlock(&imp->imp_lock);
 119         RETURN(0);
 120 out_free:
 121         if (imp_conn)
 122                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 123 out_put:
 124         ptlrpc_connection_put(ptlrpc_conn);
 125         RETURN(rc);
 126 }
 127
 128 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 129 {
 130         return import_set_conn(imp, uuid, 1, 0);
 131 }
 132
 133 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 134                            int priority)
 135 {
 136         return import_set_conn(imp, uuid, priority, 1);
 137 }
 138 EXPORT_SYMBOL(client_import_add_conn);
 139
 140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 141 {
 142         struct obd_import_conn *imp_conn;
 143         struct obd_export *dlmexp;
 144         int rc = -ENOENT;
 145         ENTRY;
 146
 147         spin_lock(&imp->imp_lock);
 148         if (list_empty(&imp->imp_conn_list)) {
 149                 LASSERT(!imp->imp_connection);
 150                 GOTO(out, rc);
 151         }
 152
 153         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 155                         continue;
 156                 LASSERT(imp_conn->oic_conn);
 157
 158                 if (imp_conn == imp->imp_conn_current) {
 159                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 160
 161                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 162                             imp->imp_state != LUSTRE_IMP_DISCON) {
 163                                 CERROR("can't remove current connection\n");
 164                                 GOTO(out, rc = -EBUSY);
 165                         }
 166
 167                         ptlrpc_connection_put(imp->imp_connection);
 168                         imp->imp_connection = NULL;
 169
 170                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 171                         if (dlmexp && dlmexp->exp_connection) {
 172                                 LASSERT(dlmexp->exp_connection ==
 173                                         imp_conn->oic_conn);
 174                                 ptlrpc_connection_put(dlmexp->exp_connection);
 175                                 dlmexp->exp_connection = NULL;
 176                         }
 177                 }
 178
 179                 list_del(&imp_conn->oic_item);
 180                 ptlrpc_connection_put(imp_conn->oic_conn);
 181                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 182                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 183                        imp, imp->imp_obd->obd_name, uuid->uuid);
 184                 rc = 0;
 185                 break;
 186         }
 187 out:
 188         spin_unlock(&imp->imp_lock);
 189         if (rc == -ENOENT)
 190                 CERROR("connection %s not found\n", uuid->uuid);
 191         RETURN(rc);
 192 }
 193 EXPORT_SYMBOL(client_import_del_conn);
 194
 195 /**
 196  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 197  * to find a conn uuid of \a imp which can reach \a peer.
 198  */
 199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 200                             struct obd_uuid *uuid)
 201 {
 202         struct obd_import_conn *conn;
 203         int rc = -ENOENT;
 204         ENTRY;
 205
 206         spin_lock(&imp->imp_lock);
 207         list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 208                 /* Check if conn UUID does have this peer NID. */
 209                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 210                         *uuid = conn->oic_uuid;
 211                         rc = 0;
 212                         break;
 213                 }
 214         }
 215         spin_unlock(&imp->imp_lock);
 216         RETURN(rc);
 217 }
 218 EXPORT_SYMBOL(client_import_find_conn);
 219
 220 void client_destroy_import(struct obd_import *imp)
 221 {
 222         /* Drop security policy instance after all RPCs have finished/aborted
 223          * to let all busy contexts be released. */
 224         class_import_get(imp);
 225         class_destroy_import(imp);
 226         sptlrpc_import_sec_put(imp);
 227         class_import_put(imp);
 228 }
 229 EXPORT_SYMBOL(client_destroy_import);
 230
 231 /**
 232  * Check whether or not the OSC is on MDT.
 233  * In the config log,
 234  * osc on MDT
 235  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 236  * osc on client
 237  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 238  *
 239  **/
 240 static int osc_on_mdt(char *obdname)
 241 {
 242         char *ptr;
 243
 244         ptr = strrchr(obdname, '-');
 245         if (ptr == NULL)
 246                 return 0;
 247
 248         if (strncmp(ptr + 1, "MDT", 3) == 0)
 249                 return 1;
 250
 251         return 0;
 252 }
 253
 254 /* Configure an RPC client OBD device.
 255  *
 256  * lcfg parameters:
 257  * 1 - client UUID
 258  * 2 - server UUID
 259  * 3 - inactive-on-startup
 260  */
 261 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 262 {
 263         struct client_obd *cli = &obddev->u.cli;
 264         struct obd_import *imp;
 265         struct obd_uuid server_uuid;
 266         int rq_portal, rp_portal, connect_op;
 267         char *name = obddev->obd_type->typ_name;
 268         enum ldlm_ns_type ns_type = LDLM_NS_TYPE_UNKNOWN;
 269         char *cli_name = lustre_cfg_buf(lcfg, 0);
 270         int rc;
 271         ENTRY;
 272
 273         /* In a more perfect world, we would hang a ptlrpc_client off of
 274          * obd_type and just use the values from there. */
 275         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 276                 rq_portal = OST_REQUEST_PORTAL;
 277                 rp_portal = OSC_REPLY_PORTAL;
 278                 connect_op = OST_CONNECT;
 279                 cli->cl_sp_me = LUSTRE_SP_CLI;
 280                 cli->cl_sp_to = LUSTRE_SP_OST;
 281                 ns_type = LDLM_NS_TYPE_OSC;
 282         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 283                    !strcmp(name, LUSTRE_LWP_NAME)) {
 284                 rq_portal = MDS_REQUEST_PORTAL;
 285                 rp_portal = MDC_REPLY_PORTAL;
 286                 connect_op = MDS_CONNECT;
 287                 if (is_lwp_on_ost(cli_name))
 288                         cli->cl_sp_me = LUSTRE_SP_OST;
 289                 else if (is_lwp_on_mdt(cli_name))
 290                         cli->cl_sp_me = LUSTRE_SP_MDT;
 291                 else
 292                         cli->cl_sp_me = LUSTRE_SP_CLI;
 293                 cli->cl_sp_to = LUSTRE_SP_MDT;
 294                 ns_type = LDLM_NS_TYPE_MDC;
 295         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 296                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 297                         /* OSP_on_MDT for other MDTs */
 298                         connect_op = MDS_CONNECT;
 299                         cli->cl_sp_to = LUSTRE_SP_MDT;
 300                         ns_type = LDLM_NS_TYPE_MDC;
 301                         rq_portal = OUT_PORTAL;
 302                 } else {
 303                         /* OSP on MDT for OST */
 304                         connect_op = OST_CONNECT;
 305                         cli->cl_sp_to = LUSTRE_SP_OST;
 306                         ns_type = LDLM_NS_TYPE_OSC;
 307                         rq_portal = OST_REQUEST_PORTAL;
 308                 }
 309                 rp_portal = OSC_REPLY_PORTAL;
 310                 cli->cl_sp_me = LUSTRE_SP_MDT;
 311         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 312                 rq_portal = MGS_REQUEST_PORTAL;
 313                 rp_portal = MGC_REPLY_PORTAL;
 314                 connect_op = MGS_CONNECT;
 315                 cli->cl_sp_me = LUSTRE_SP_MGC;
 316                 cli->cl_sp_to = LUSTRE_SP_MGS;
 317                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 318                 ns_type = LDLM_NS_TYPE_MGC;
 319         } else {
 320                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 321                        name);
 322                 RETURN(-EINVAL);
 323         }
 324
 325         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 326                 CERROR("requires a TARGET UUID\n");
 327                 RETURN(-EINVAL);
 328         }
 329
 330         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 331                 CERROR("client UUID must be less than 38 characters\n");
 332                 RETURN(-EINVAL);
 333         }
 334
 335         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 336                 CERROR("setup requires a SERVER UUID\n");
 337                 RETURN(-EINVAL);
 338         }
 339
 340         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 341                 CERROR("target UUID must be less than 38 characters\n");
 342                 RETURN(-EINVAL);
 343         }
 344
 345         init_rwsem(&cli->cl_sem);
 346         mutex_init(&cli->cl_mgc_mutex);
 347         cli->cl_conn_count = 0;
 348         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 349                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 350                      sizeof(server_uuid)));
 351
 352         cli->cl_dirty_pages = 0;
 353         cli->cl_avail_grant = 0;
 354         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 355         /* cl_dirty_max_pages may be changed at connect time in
 356          * ptlrpc_connect_interpret(). */
 357         client_adjust_max_dirty(cli);
 358         INIT_LIST_HEAD(&cli->cl_cache_waiters);
 359         INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 360         INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 361         INIT_LIST_HEAD(&cli->cl_loi_write_list);
 362         INIT_LIST_HEAD(&cli->cl_loi_read_list);
 363         spin_lock_init(&cli->cl_loi_list_lock);
 364         atomic_set(&cli->cl_pending_w_pages, 0);
 365         atomic_set(&cli->cl_pending_r_pages, 0);
 366         cli->cl_r_in_flight = 0;
 367         cli->cl_w_in_flight = 0;
 368
 369         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 370         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 371         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 372         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 373         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 374         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 375
 376         /* lru for osc. */
 377         INIT_LIST_HEAD(&cli->cl_lru_osc);
 378         atomic_set(&cli->cl_lru_shrinkers, 0);
 379         atomic_long_set(&cli->cl_lru_busy, 0);
 380         atomic_long_set(&cli->cl_lru_in_list, 0);
 381         INIT_LIST_HEAD(&cli->cl_lru_list);
 382         spin_lock_init(&cli->cl_lru_list_lock);
 383         atomic_long_set(&cli->cl_unstable_count, 0);
 384         INIT_LIST_HEAD(&cli->cl_shrink_list);
 385
 386         init_waitqueue_head(&cli->cl_destroy_waitq);
 387         atomic_set(&cli->cl_destroy_in_flight, 0);
 388 #ifdef ENABLE_CHECKSUM
 389         /* Turn on checksumming by default. */
 390         cli->cl_checksum = 1;
 391         /*
 392          * The supported checksum types will be worked out at connect time
 393          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 394          * through procfs.
 395          */
 396         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 397 #endif
 398         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 399
 400         /* This value may be reduced at connect time in
 401          * ptlrpc_connect_interpret() . We initialize it to only
 402          * 1MB until we know what the performance looks like.
 403          * In the future this should likely be increased. LU-1431 */
 404         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
 405                                           LNET_MTU >> PAGE_CACHE_SHIFT);
 406
 407         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 408          * it will be updated at OSC connection time. */
 409         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 410
 411         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 412                 cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 413         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 414                 cli->cl_max_rpcs_in_flight = 2;
 415         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 416                 cli->cl_max_rpcs_in_flight = 3;
 417         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 418                 cli->cl_max_rpcs_in_flight = 4;
 419         } else {
 420                 if (osc_on_mdt(obddev->obd_name))
 421                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_MAX;
 422                 else
 423                         cli->cl_max_rpcs_in_flight = OBD_MAX_RIF_DEFAULT;
 424         }
 425
 426         spin_lock_init(&cli->cl_mod_rpcs_lock);
 427         spin_lock_init(&cli->cl_mod_rpcs_hist.oh_lock);
 428         cli->cl_max_mod_rpcs_in_flight = 0;
 429         cli->cl_mod_rpcs_in_flight = 0;
 430         cli->cl_close_rpcs_in_flight = 0;
 431         init_waitqueue_head(&cli->cl_mod_rpcs_waitq);
 432         cli->cl_mod_tag_bitmap = NULL;
 433
 434         if (connect_op == MDS_CONNECT) {
 435                 cli->cl_max_mod_rpcs_in_flight = cli->cl_max_rpcs_in_flight - 1;
 436                 OBD_ALLOC(cli->cl_mod_tag_bitmap,
 437                           BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 438                 if (cli->cl_mod_tag_bitmap == NULL)
 439                         GOTO(err, rc = -ENOMEM);
 440         }
 441
 442         rc = ldlm_get_ref();
 443         if (rc) {
 444                 CERROR("ldlm_get_ref failed: %d\n", rc);
 445                 GOTO(err, rc);
 446         }
 447
 448         ptlrpc_init_client(rq_portal, rp_portal, name,
 449                            &obddev->obd_ldlm_client);
 450
 451         imp = class_new_import(obddev);
 452         if (imp == NULL)
 453                 GOTO(err_ldlm, rc = -ENOENT);
 454         imp->imp_client = &obddev->obd_ldlm_client;
 455         imp->imp_connect_op = connect_op;
 456         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 457                LUSTRE_CFG_BUFLEN(lcfg, 1));
 458         class_import_put(imp);
 459
 460         rc = client_import_add_conn(imp, &server_uuid, 1);
 461         if (rc) {
 462                 CERROR("can't add initial connection\n");
 463                 GOTO(err_import, rc);
 464         }
 465
 466         cli->cl_import = imp;
 467         /* cli->cl_max_mds_easize updated by mdc_init_ea_size() */
 468         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 469
 470         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 471                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 472                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 473                                name, obddev->obd_name,
 474                                cli->cl_target_uuid.uuid);
 475                         spin_lock(&imp->imp_lock);
 476                         imp->imp_deactive = 1;
 477                         spin_unlock(&imp->imp_lock);
 478                 }
 479         }
 480
 481         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 482                                                    LDLM_NAMESPACE_CLIENT,
 483                                                    LDLM_NAMESPACE_GREEDY,
 484                                                    ns_type);
 485         if (obddev->obd_namespace == NULL) {
 486                 CERROR("Unable to create client namespace - %s\n",
 487                        obddev->obd_name);
 488                 GOTO(err_import, rc = -ENOMEM);
 489         }
 490
 491         RETURN(rc);
 492
 493 err_import:
 494         class_destroy_import(imp);
 495 err_ldlm:
 496         ldlm_put_ref();
 497 err:
 498         if (cli->cl_mod_tag_bitmap != NULL)
 499                 OBD_FREE(cli->cl_mod_tag_bitmap,
 500                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 501         cli->cl_mod_tag_bitmap = NULL;
 502         RETURN(rc);
 503
 504 }
 505 EXPORT_SYMBOL(client_obd_setup);
 506
 507 int client_obd_cleanup(struct obd_device *obddev)
 508 {
 509         struct client_obd *cli = &obddev->u.cli;
 510         ENTRY;
 511
 512         ldlm_namespace_free_post(obddev->obd_namespace);
 513         obddev->obd_namespace = NULL;
 514
 515         obd_cleanup_client_import(obddev);
 516         LASSERT(obddev->u.cli.cl_import == NULL);
 517
 518         ldlm_put_ref();
 519
 520         if (cli->cl_mod_tag_bitmap != NULL)
 521                 OBD_FREE(cli->cl_mod_tag_bitmap,
 522                          BITS_TO_LONGS(OBD_MAX_RIF_MAX) * sizeof(long));
 523         cli->cl_mod_tag_bitmap = NULL;
 524
 525         RETURN(0);
 526 }
 527 EXPORT_SYMBOL(client_obd_cleanup);
 528
 529 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 530 int client_connect_import(const struct lu_env *env,
 531                           struct obd_export **exp,
 532                           struct obd_device *obd, struct obd_uuid *cluuid,
 533                           struct obd_connect_data *data, void *localdata)
 534 {
 535         struct client_obd       *cli    = &obd->u.cli;
 536         struct obd_import       *imp    = cli->cl_import;
 537         struct obd_connect_data *ocd;
 538         struct lustre_handle    conn    = { 0 };
 539         int                     rc;
 540         bool                    is_mdc = false;
 541         ENTRY;
 542
 543         *exp = NULL;
 544         down_write(&cli->cl_sem);
 545         if (cli->cl_conn_count > 0)
 546                 GOTO(out_sem, rc = -EALREADY);
 547
 548         rc = class_connect(&conn, obd, cluuid);
 549         if (rc)
 550                 GOTO(out_sem, rc);
 551
 552         cli->cl_conn_count++;
 553         *exp = class_conn2export(&conn);
 554
 555         LASSERT(obd->obd_namespace);
 556
 557         imp->imp_dlm_handle = conn;
 558         rc = ptlrpc_init_import(imp);
 559         if (rc != 0)
 560                 GOTO(out_ldlm, rc);
 561
 562         ocd = &imp->imp_connect_data;
 563         if (data) {
 564                 *ocd = *data;
 565                 is_mdc = strncmp(imp->imp_obd->obd_type->typ_name,
 566                                  LUSTRE_MDC_NAME, 3) == 0;
 567                 if (is_mdc)
 568                         data->ocd_connect_flags |= OBD_CONNECT_MULTIMODRPCS;
 569                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 570         }
 571
 572         rc = ptlrpc_connect_import(imp);
 573         if (rc != 0) {
 574                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 575                 GOTO(out_ldlm, rc);
 576         }
 577         LASSERT(*exp != NULL && (*exp)->exp_connection);
 578
 579         if (data) {
 580                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 581                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 582                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 583                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 584                 /* clear the flag as it was not set and is not known
 585                  * by upper layers */
 586                 if (is_mdc)
 587                         data->ocd_connect_flags &= ~OBD_CONNECT_MULTIMODRPCS;
 588         }
 589
 590         ptlrpc_pinger_add_import(imp);
 591
 592         EXIT;
 593
 594         if (rc) {
 595 out_ldlm:
 596                 cli->cl_conn_count--;
 597                 class_disconnect(*exp);
 598                 *exp = NULL;
 599         }
 600 out_sem:
 601         up_write(&cli->cl_sem);
 602
 603         return rc;
 604 }
 605 EXPORT_SYMBOL(client_connect_import);
 606
 607 int client_disconnect_export(struct obd_export *exp)
 608 {
 609         struct obd_device *obd = class_exp2obd(exp);
 610         struct client_obd *cli;
 611         struct obd_import *imp;
 612         int rc = 0, err;
 613         ENTRY;
 614
 615         if (!obd) {
 616                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 617                        exp, exp ? exp->exp_handle.h_cookie : -1);
 618                 RETURN(-EINVAL);
 619         }
 620
 621         cli = &obd->u.cli;
 622         imp = cli->cl_import;
 623
 624         down_write(&cli->cl_sem);
 625         CDEBUG(D_INFO, "disconnect %s - %zu\n", obd->obd_name,
 626                 cli->cl_conn_count);
 627
 628         if (cli->cl_conn_count == 0) {
 629                 CERROR("disconnecting disconnected device (%s)\n",
 630                        obd->obd_name);
 631                 GOTO(out_disconnect, rc = -EINVAL);
 632         }
 633
 634         cli->cl_conn_count--;
 635         if (cli->cl_conn_count != 0)
 636                 GOTO(out_disconnect, rc = 0);
 637
 638         /* Mark import deactivated now, so we don't try to reconnect if any
 639          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 640          * fully deactivate the import, or that would drop all requests. */
 641         spin_lock(&imp->imp_lock);
 642         imp->imp_deactive = 1;
 643         spin_unlock(&imp->imp_lock);
 644
 645         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 646          * delete it regardless.  (It's safe to delete an import that was
 647          * never added.) */
 648         (void)ptlrpc_pinger_del_import(imp);
 649
 650         if (obd->obd_namespace != NULL) {
 651                 /* obd_force == local only */
 652                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 653                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 654                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 655         }
 656
 657         /* There's no need to hold sem while disconnecting an import,
 658          * and it may actually cause deadlock in GSS. */
 659         up_write(&cli->cl_sem);
 660         rc = ptlrpc_disconnect_import(imp, 0);
 661         down_write(&cli->cl_sem);
 662
 663         ptlrpc_invalidate_import(imp);
 664
 665         EXIT;
 666
 667 out_disconnect:
 668         /* Use server style - class_disconnect should be always called for
 669          * o_disconnect. */
 670         err = class_disconnect(exp);
 671         if (!rc && err)
 672                 rc = err;
 673
 674         up_write(&cli->cl_sem);
 675
 676         RETURN(rc);
 677 }
 678 EXPORT_SYMBOL(client_disconnect_export);
 679
 680 #ifdef HAVE_SERVER_SUPPORT
 681 int server_disconnect_export(struct obd_export *exp)
 682 {
 683         int rc;
 684         ENTRY;
 685
 686         /* Disconnect early so that clients can't keep using export. */
 687         rc = class_disconnect(exp);
 688         /* Close import to avoid sending any requests. */
 689         if (exp->exp_imp_reverse)
 690                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 691
 692         ldlm_bl_thread_wakeup();
 693
 694         /* complete all outstanding replies */
 695         spin_lock(&exp->exp_lock);
 696         while (!list_empty(&exp->exp_outstanding_replies)) {
 697                 struct ptlrpc_reply_state *rs =
 698                         list_entry(exp->exp_outstanding_replies.next,
 699                                        struct ptlrpc_reply_state, rs_exp_list);
 700                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 701
 702                 spin_lock(&svcpt->scp_rep_lock);
 703
 704                 list_del_init(&rs->rs_exp_list);
 705                 spin_lock(&rs->rs_lock);
 706                 ptlrpc_schedule_difficult_reply(rs);
 707                 spin_unlock(&rs->rs_lock);
 708
 709                 spin_unlock(&svcpt->scp_rep_lock);
 710         }
 711         spin_unlock(&exp->exp_lock);
 712
 713         RETURN(rc);
 714 }
 715 EXPORT_SYMBOL(server_disconnect_export);
 716
 717 /* --------------------------------------------------------------------------
 718  * from old lib/target.c
 719  * -------------------------------------------------------------------------- */
 720
 721 static int target_handle_reconnect(struct lustre_handle *conn,
 722                                    struct obd_export *exp,
 723                                    struct obd_uuid *cluuid)
 724 {
 725         struct lustre_handle *hdl;
 726         ENTRY;
 727
 728         hdl = &exp->exp_imp_reverse->imp_remote_handle;
 729         if (exp->exp_connection && lustre_handle_is_used(hdl)) {
 730                 struct obd_device *target;
 731
 732                 target = exp->exp_obd;
 733
 734                 /* Might be a re-connect after a partition. */
 735                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 736                         if (target->obd_recovering) {
 737                                 int timeout = cfs_duration_sec(cfs_time_sub(
 738                                         cfs_timer_deadline(
 739                                         &target->obd_recovery_timer),
 740                                         cfs_time_current()));
 741
 742                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 743                                         "ing, waiting for %d clients in recov"
 744                                         "ery for %d:%.02d\n", target->obd_name,
 745                                         obd_uuid2str(&exp->exp_client_uuid),
 746                                         obd_export_nid2str(exp),
 747                                         target->obd_max_recoverable_clients,
 748                                         timeout / 60, timeout % 60);
 749                         } else {
 750                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 751                                         "reconnecting\n", target->obd_name,
 752                                         obd_uuid2str(&exp->exp_client_uuid),
 753                                         obd_export_nid2str(exp));
 754                         }
 755
 756                         conn->cookie = exp->exp_handle.h_cookie;
 757                         /* target_handle_connect() treats EALREADY and
 758                          * -EALREADY differently.  EALREADY means we are
 759                          * doing a valid reconnect from the same client. */
 760                         RETURN(EALREADY);
 761                 } else {
 762                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 763                                       "with handle "LPX64". Rejecting client "
 764                                       "with the same UUID trying to reconnect "
 765                                       "with handle "LPX64"\n", target->obd_name,
 766                                       obd_uuid2str(&exp->exp_client_uuid),
 767                                       obd_export_nid2str(exp),
 768                                       hdl->cookie, conn->cookie);
 769                         memset(conn, 0, sizeof *conn);
 770                         /* target_handle_connect() treats EALREADY and
 771                          * -EALREADY differently.  -EALREADY is an error
 772                          * (same UUID, different handle). */
 773                         RETURN(-EALREADY);
 774                 }
 775         }
 776
 777         conn->cookie = exp->exp_handle.h_cookie;
 778         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 779                cluuid->uuid, exp, conn->cookie);
 780         RETURN(0);
 781 }
 782
 783 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 784                           int error)
 785 {
 786         struct obd_export *exp = cb_data;
 787
 788         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 789                obd->obd_name, exp->exp_client_uuid.uuid);
 790
 791         spin_lock(&exp->exp_lock);
 792         exp->exp_need_sync = 0;
 793         spin_unlock(&exp->exp_lock);
 794         class_export_cb_put(exp);
 795 }
 796
 797 static void
 798 check_and_start_recovery_timer(struct obd_device *obd,
 799                                struct ptlrpc_request *req, int new_client);
 800
 801 /**
 802  * update flags for import during reconnect process
 803  */
 804 static int rev_import_flags_update(struct obd_import *revimp,
 805                                    struct ptlrpc_request *req)
 806 {
 807         int rc;
 808         struct obd_connect_data *data;
 809
 810         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 811
 812         if (data->ocd_connect_flags & OBD_CONNECT_AT)
 813                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 814         else
 815                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 816
 817         revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
 818
 819         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
 820         if (rc) {
 821                 CERROR("%s: cannot get reverse import %s security: rc = %d\n",
 822                         revimp->imp_client->cli_name,
 823                         libcfs_id2str(req->rq_peer), rc);
 824                 return rc;
 825         }
 826
 827         return 0;
 828 }
 829
 830 /**
 831  * Allocate a new reverse import for an export.
 832  *
 833  * \retval -errno in case error hit
 834  * \retval 0 if reverse import correctly init
 835  **/
 836 int rev_import_init(struct obd_export *export)
 837 {
 838         struct obd_device *obd = export->exp_obd;
 839         struct obd_import *revimp;
 840
 841         LASSERT(export->exp_imp_reverse == NULL);
 842
 843         revimp = class_new_import(obd);
 844         if (revimp == NULL)
 845                 return -ENOMEM;
 846
 847         revimp->imp_remote_handle.cookie = 0ULL;
 848         revimp->imp_client = &obd->obd_ldlm_client;
 849         revimp->imp_dlm_fake = 1;
 850
 851         /* it is safe to connect import in new state as no sends possible */
 852         spin_lock(&export->exp_lock);
 853         export->exp_imp_reverse = revimp;
 854         spin_unlock(&export->exp_lock);
 855         class_import_put(revimp);
 856
 857         return 0;
 858 }
 859 EXPORT_SYMBOL(rev_import_init);
 860
 861 /**
 862  * Handle reconnect for an export.
 863  *
 864  * \param exp export to handle reconnect process
 865  * \param req client reconnect request
 866  *
 867  * \retval -rc in case securitfy flavor can't be changed
 868  * \retval 0 in case none problems
 869  */
 870 static int rev_import_reconnect(struct obd_export *exp,
 871                                 struct ptlrpc_request *req)
 872 {
 873         struct obd_import *revimp = exp->exp_imp_reverse;
 874         struct lustre_handle *lh;
 875         int rc;
 876
 877         /* avoid sending a request until import flags are changed */
 878         ptlrpc_import_enter_resend(revimp);
 879
 880         if (revimp->imp_connection != NULL)
 881                 ptlrpc_connection_put(revimp->imp_connection);
 882
 883         /*
 884          * client from recovery don't have a handle so we need to take from
 885          * request. it may produce situation when wrong client connected
 886          * to recovery as we trust a client uuid
 887          */
 888         lh = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 889         revimp->imp_remote_handle = *lh;
 890
 891         /* unknown versions will be caught in
 892          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 893         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 894
 895         revimp->imp_connection = ptlrpc_connection_addref(exp->exp_connection);
 896
 897         rc = rev_import_flags_update(revimp, req);
 898         if (rc != 0) {
 899                 /* it is safe to still be in RECOVERY phase as we are not able
 900                  * to setup correct security flavor so requests are not able to
 901                  * be delivered correctly */
 902                 return rc;
 903         }
 904
 905         /* resend all rpc's via new connection */
 906         return ptlrpc_import_recovery_state_machine(revimp);
 907 }
 908
 909 int target_handle_connect(struct ptlrpc_request *req)
 910 {
 911         struct obd_device *target = NULL;
 912         struct obd_export *export = NULL;
 913         /* connect handle - filled from target_handle_reconnect in
 914          * reconnect case */
 915         struct lustre_handle conn;
 916         struct lustre_handle *tmp;
 917         struct obd_uuid tgtuuid;
 918         struct obd_uuid cluuid;
 919         char *str;
 920         int rc = 0;
 921         char *target_start;
 922         int target_len;
 923         bool     mds_conn = false, lw_client = false;
 924         bool     mds_mds_conn = false;
 925         bool     new_mds_mds_conn = false;
 926         bool     target_referenced = false;
 927         struct obd_connect_data *data, *tmpdata;
 928         int size, tmpsize;
 929         lnet_nid_t *client_nid = NULL;
 930         ENTRY;
 931
 932         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 933
 934         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 935         if (str == NULL) {
 936                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 937                 GOTO(out, rc = -EINVAL);
 938         }
 939
 940         obd_str2uuid(&tgtuuid, str);
 941         target = class_uuid2obd(&tgtuuid);
 942         if (!target)
 943                 target = class_name2obd(str);
 944
 945         if (!target) {
 946                 deuuidify(str, NULL, &target_start, &target_len);
 947                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 948                                    "from %s (no target). If you are running "
 949                                    "an HA pair check that the target is "
 950                                    "mounted on the other server.\n", str,
 951                                    libcfs_nid2str(req->rq_peer.nid));
 952                 GOTO(out, rc = -ENODEV);
 953         }
 954
 955         spin_lock(&target->obd_dev_lock);
 956         if (target->obd_stopping || !target->obd_set_up) {
 957                 spin_unlock(&target->obd_dev_lock);
 958
 959                 deuuidify(str, NULL, &target_start, &target_len);
 960                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 961                               target_len, target_start,
 962                               libcfs_nid2str(req->rq_peer.nid),
 963                               (target->obd_stopping ?
 964                                "stopping" : "not set up"));
 965                 GOTO(out, rc = -ENODEV);
 966         }
 967
 968         if (target->obd_no_conn) {
 969                 spin_unlock(&target->obd_dev_lock);
 970
 971                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 972                                "from %s\n", target->obd_name,
 973                                libcfs_nid2str(req->rq_peer.nid));
 974                 GOTO(out, rc = -EAGAIN);
 975         }
 976
 977         /* Make sure the target isn't cleaned up while we're here. Yes,
 978          * there's still a race between the above check and our incref here.
 979          * Really, class_uuid2obd should take the ref. */
 980         class_incref(target, __func__, current);
 981         target_referenced = true;
 982
 983         target->obd_conn_inprogress++;
 984         spin_unlock(&target->obd_dev_lock);
 985
 986         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 987         if (str == NULL) {
 988                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 989                 GOTO(out, rc = -EINVAL);
 990         }
 991
 992         obd_str2uuid(&cluuid, str);
 993
 994         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 995         if (tmp == NULL)
 996                 GOTO(out, rc = -EPROTO);
 997
 998         conn = *tmp;
 999
1000         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1001                                     RCL_CLIENT);
1002         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
1003         if (!data)
1004                 GOTO(out, rc = -EPROTO);
1005
1006         rc = req_capsule_server_pack(&req->rq_pill);
1007         if (rc)
1008                 GOTO(out, rc);
1009
1010 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 53, 0)
1011         /* Don't allow clients to connect that are using old 1.8 format
1012          * protocol conventions (LUSTRE_MSG_MAGIC_v1, !MSGHDR_CKSUM_INCOMPAT18,
1013          * ldlm_flock_policy_wire format, MDT_ATTR_xTIME_SET, etc).  The
1014          * FULL20 flag should be set on all connections since 2.0, but no
1015          * longer affects behaviour.
1016          *
1017          * Later this check will be disabled and the flag can be retired
1018          * completely once interop with 3.0 is no longer needed.
1019          */
1020         if (!(data->ocd_connect_flags & OBD_CONNECT_FULL20))
1021                 GOTO(out, rc = -EPROTO);
1022 #endif
1023
1024         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1025                 if (data->ocd_version < LUSTRE_VERSION_CODE -
1026                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
1027                     data->ocd_version > LUSTRE_VERSION_CODE +
1028                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
1029                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
1030                                   "libclient connection attempt",
1031                                   data->ocd_version < LUSTRE_VERSION_CODE ?
1032                                   "old" : "new",
1033                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
1034                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
1035                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
1036                                   OBD_OCD_VERSION_FIX(data->ocd_version));
1037                         data = req_capsule_server_sized_get(&req->rq_pill,
1038                                                             &RMF_CONNECT_DATA,
1039                                     offsetof(typeof(*data), ocd_version) +
1040                                              sizeof(data->ocd_version));
1041                         if (data) {
1042                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
1043                                 data->ocd_version = LUSTRE_VERSION_CODE;
1044                         }
1045                         GOTO(out, rc = -EPROTO);
1046                 }
1047         }
1048
1049         /* Note: lw_client is needed in MDS-MDS failover during update log
1050          * processing, so we needs to allow lw_client to be connected at
1051          * anytime, instead of only the initial connection */
1052         lw_client = (data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0;
1053
1054         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) {
1055                 mds_conn = (data->ocd_connect_flags & OBD_CONNECT_MDS) != 0;
1056                 mds_mds_conn = (data->ocd_connect_flags &
1057                                 OBD_CONNECT_MDS_MDS) != 0;
1058
1059                 /* OBD_CONNECT_MNE_SWAB is defined as OBD_CONNECT_MDS_MDS
1060                  * for Imperative Recovery connection from MGC to MGS.
1061                  *
1062                  * Via check OBD_CONNECT_FID, we can distinguish whether
1063                  * the OBD_CONNECT_MDS_MDS/OBD_CONNECT_MNE_SWAB is from
1064                  * MGC or MDT. */
1065                 if (!lw_client &&
1066                     (data->ocd_connect_flags & OBD_CONNECT_MDS_MDS) &&
1067                     (data->ocd_connect_flags & OBD_CONNECT_FID) &&
1068                     (data->ocd_connect_flags & OBD_CONNECT_VERSION)) {
1069                         __u32 major = OBD_OCD_VERSION_MAJOR(data->ocd_version);
1070                         __u32 minor = OBD_OCD_VERSION_MINOR(data->ocd_version);
1071                         __u32 patch = OBD_OCD_VERSION_PATCH(data->ocd_version);
1072
1073                         /* We do not support the MDT-MDT interoperations with
1074                          * different version MDT because of protocol changes. */
1075                         if (unlikely(major != LUSTRE_MAJOR ||
1076                                      minor != LUSTRE_MINOR ||
1077                                      abs(patch - LUSTRE_PATCH) > 3)) {
1078                                 LCONSOLE_WARN("%s (%u.%u.%u.%u) refused the "
1079                                         "connection from different version MDT "
1080                                         "(%d.%d.%d.%d) %s %s\n",
1081                                         target->obd_name, LUSTRE_MAJOR,
1082                                         LUSTRE_MINOR, LUSTRE_PATCH, LUSTRE_FIX,
1083                                         major, minor, patch,
1084                                         OBD_OCD_VERSION_FIX(data->ocd_version),
1085                                         libcfs_nid2str(req->rq_peer.nid), str);
1086
1087                                 GOTO(out, rc = -EPROTO);
1088                         }
1089                 }
1090         }
1091
1092         /* lctl gets a backstage, all-access pass. */
1093         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
1094                 goto dont_check_exports;
1095
1096         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
1097         if (!export)
1098                 goto no_export;
1099
1100         /* We've found an export in the hash. */
1101
1102         spin_lock(&export->exp_lock);
1103
1104         if (export->exp_connecting) { /* bug 9635, et. al. */
1105                 spin_unlock(&export->exp_lock);
1106                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
1107                               export->exp_obd->obd_name, export,
1108                               libcfs_nid2str(req->rq_peer.nid));
1109                 class_export_put(export);
1110                 export = NULL;
1111                 rc = -EALREADY;
1112         } else if ((mds_conn || lw_client) && export->exp_connection != NULL) {
1113                 spin_unlock(&export->exp_lock);
1114                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
1115                         /* MDS or LWP reconnected after failover. */
1116                         LCONSOLE_WARN("%s: Received %s connection from "
1117                             "%s, removing former export from %s\n",
1118                             target->obd_name, mds_conn ? "MDS" : "LWP",
1119                             libcfs_nid2str(req->rq_peer.nid),
1120                             libcfs_nid2str(export->exp_connection->c_peer.nid));
1121                 else
1122                         /* New MDS connection from the same NID. */
1123                         LCONSOLE_WARN("%s: Received new %s connection from "
1124                                 "%s, removing former export from same NID\n",
1125                                 target->obd_name, mds_conn ? "MDS" : "LWP",
1126                                 libcfs_nid2str(req->rq_peer.nid));
1127                 class_fail_export(export);
1128                 class_export_put(export);
1129                 export = NULL;
1130                 rc = 0;
1131         } else if (export->exp_connection != NULL &&
1132                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
1133                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
1134                     MSG_CONNECT_INITIAL)) {
1135                 spin_unlock(&export->exp_lock);
1136                 /* In MDS failover we have static UUID but NID can change. */
1137                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
1138                               "existing nid %s is already connected\n",
1139                               target->obd_name, cluuid.uuid,
1140                               libcfs_nid2str(req->rq_peer.nid),
1141                               libcfs_nid2str(
1142                                       export->exp_connection->c_peer.nid));
1143                 rc = -EALREADY;
1144                 class_export_put(export);
1145                 export = NULL;
1146         } else {
1147                 export->exp_connecting = 1;
1148                 spin_unlock(&export->exp_lock);
1149                 LASSERT(export->exp_obd == target);
1150
1151                 rc = target_handle_reconnect(&conn, export, &cluuid);
1152         }
1153
1154         /* If we found an export, we already unlocked. */
1155         if (!export) {
1156 no_export:
1157                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
1158         } else if (req->rq_export == NULL &&
1159                    atomic_read(&export->exp_rpc_count) > 0) {
1160                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
1161                               "still busy with %d references\n",
1162                               target->obd_name, cluuid.uuid,
1163                               libcfs_nid2str(req->rq_peer.nid),
1164                               atomic_read(&export->exp_refcount));
1165                 GOTO(out, rc = -EBUSY);
1166         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
1167                 if (!strstr(cluuid.uuid, "mdt"))
1168                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
1169                                       "known client %s (at %s) because it "
1170                                       "is indicating it is a new client",
1171                                       target->obd_name, cluuid.uuid,
1172                                       libcfs_nid2str(req->rq_peer.nid));
1173                 GOTO(out, rc = -EALREADY);
1174         } else {
1175                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1176         }
1177
1178         if (rc < 0) {
1179                 GOTO(out, rc);
1180         }
1181
1182         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1183                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1184               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1185               export, (long)cfs_time_current_sec(),
1186               export ? (long)export->exp_last_request_time : 0);
1187
1188         /* If this is the first time a client connects, reset the recovery
1189          * timer. Discard lightweight connections which might be local. */
1190         if (!lw_client && rc == 0 && target->obd_recovering)
1191                 check_and_start_recovery_timer(target, req, export == NULL);
1192
1193         /* We want to handle EALREADY but *not* -EALREADY from
1194          * target_handle_reconnect(), return reconnection state in a flag. */
1195         if (rc == EALREADY) {
1196                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1197                 rc = 0;
1198         } else {
1199                 LASSERT(rc == 0);
1200         }
1201
1202         /* Tell the client if we support replayable requests. */
1203         if (target->obd_replayable)
1204                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1205         client_nid = &req->rq_peer.nid;
1206
1207         if (export == NULL) {
1208                 /* allow lightweight connections during recovery */
1209                 /* allow "new" MDT to be connected during recovery, since we
1210                  * need retrieve recovery update records from it */
1211                 if (target->obd_recovering && !lw_client && !mds_mds_conn) {
1212                         cfs_time_t t;
1213                         int     c; /* connected */
1214                         int     i; /* in progress */
1215                         int     k; /* known */
1216                         int     s; /* stale/evicted */
1217
1218                         c = atomic_read(&target->obd_connected_clients);
1219                         i = atomic_read(&target->obd_lock_replay_clients);
1220                         k = target->obd_max_recoverable_clients;
1221                         s = target->obd_stale_clients;
1222                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1223                         t = cfs_time_sub(t, cfs_time_current());
1224                         t = cfs_duration_sec(t);
1225                         LCONSOLE_WARN("%s: Denying connection for new client %s"
1226                                       "(at %s), waiting for %d known clients "
1227                                       "(%d recovered, %d in progress, and %d "
1228                                       "evicted) to recover in %d:%.02d\n",
1229                                       target->obd_name, cluuid.uuid,
1230                                       libcfs_nid2str(req->rq_peer.nid), k,
1231                                       c - i, i, s, (int)t / 60,
1232                                       (int)t % 60);
1233                         rc = -EBUSY;
1234                 } else {
1235 dont_check_exports:
1236                         rc = obd_connect(req->rq_svc_thread->t_env,
1237                                          &export, target, &cluuid, data,
1238                                          client_nid);
1239                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1240                                 lustre_msg_add_op_flags(req->rq_repmsg,
1241                                                         MSG_CONNECT_RECOVERING);
1242                         if (rc == 0) {
1243                                 conn.cookie = export->exp_handle.h_cookie;
1244                                 rc = rev_import_init(export);
1245                         }
1246
1247                         if (mds_mds_conn)
1248                                 new_mds_mds_conn = true;
1249                 }
1250         } else {
1251                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1252                                    export, target, &cluuid, data, client_nid);
1253         }
1254         if (rc)
1255                 GOTO(out, rc);
1256
1257         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1258         data->ocd_instance = target->u.obt.obt_instance;
1259
1260         /* Return only the parts of obd_connect_data that we understand, so the
1261          * client knows that we don't understand the rest. */
1262         if (data) {
1263                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1264                                                RCL_SERVER);
1265                 tmpdata = req_capsule_server_get(&req->rq_pill,
1266                                                  &RMF_CONNECT_DATA);
1267                 /* Don't use struct assignment here, because the client reply
1268                  * buffer may be smaller/larger than the local struct
1269                  * obd_connect_data. */
1270                 memcpy(tmpdata, data, min(tmpsize, size));
1271         }
1272
1273         /* If the client and the server are the same node, we will already
1274          * have an export that really points to the client's DLM export,
1275          * because we have a shared handles table.
1276          *
1277          * XXX this will go away when shaver stops sending the "connect" handle
1278          * in the real "remote handle" field of the request --phik 24 Apr 2003
1279          */
1280         ptlrpc_request_change_export(req, export);
1281
1282         spin_lock(&export->exp_lock);
1283         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1284                 spin_unlock(&export->exp_lock);
1285                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1286                        "or equal conn_cnt: %d >= %d\n",
1287                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1288                        export->exp_conn_cnt,
1289                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1290
1291                 GOTO(out, rc = -EALREADY);
1292         }
1293         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1294         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1295
1296         /* Don't evict liblustre clients for not pinging. */
1297         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1298                 export->exp_libclient = 1;
1299                 spin_unlock(&export->exp_lock);
1300
1301                 spin_lock(&target->obd_dev_lock);
1302                 list_del_init(&export->exp_obd_chain_timed);
1303                 spin_unlock(&target->obd_dev_lock);
1304         } else {
1305                 spin_unlock(&export->exp_lock);
1306         }
1307
1308         if (export->exp_connection != NULL) {
1309                 /* Check to see if connection came from another NID. */
1310                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1311                     !hlist_unhashed(&export->exp_nid_hash))
1312                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1313                                      &export->exp_connection->c_peer.nid,
1314                                      &export->exp_nid_hash);
1315
1316                 ptlrpc_connection_put(export->exp_connection);
1317         }
1318
1319         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1320                                                        req->rq_self,
1321                                                        &cluuid);
1322         if (hlist_unhashed(&export->exp_nid_hash))
1323                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1324                              &export->exp_connection->c_peer.nid,
1325                              &export->exp_nid_hash);
1326
1327         lustre_msg_set_handle(req->rq_repmsg, &conn);
1328
1329         rc = rev_import_reconnect(export, req);
1330         if (rc != 0)
1331                 GOTO(out, rc);
1332
1333         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1334                 int has_transno;
1335                 __u64 transno = data->ocd_transno;
1336
1337                 spin_lock(&export->exp_lock);
1338                 /* possible race with class_disconnect_stale_exports,
1339                  * export may be already in the eviction process */
1340                 if (export->exp_failed) {
1341                         spin_unlock(&export->exp_lock);
1342                         GOTO(out, rc = -ENODEV);
1343                 }
1344                 export->exp_in_recovery = 1;
1345                 export->exp_req_replay_needed = 1;
1346                 export->exp_lock_replay_needed = 1;
1347                 spin_unlock(&export->exp_lock);
1348
1349                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1350                                  MSG_CONNECT_TRANSNO);
1351                 if (has_transno && transno == 0)
1352                         CWARN("Connect with zero transno!\n");
1353
1354                 if (has_transno && transno > 0 &&
1355                     transno < target->obd_next_recovery_transno &&
1356                     transno > target->obd_last_committed) {
1357                         /* Another way is to use cmpxchg() to be lock-free. */
1358                         spin_lock(&target->obd_recovery_task_lock);
1359                         if (transno < target->obd_next_recovery_transno)
1360                                 target->obd_next_recovery_transno = transno;
1361                         spin_unlock(&target->obd_recovery_task_lock);
1362                 }
1363
1364                 atomic_inc(&target->obd_req_replay_clients);
1365                 atomic_inc(&target->obd_lock_replay_clients);
1366                 /* Note: MDS-MDS connection is allowed to be connected during
1367                  * recovery, no matter if the exports needs to be recoveried.
1368                  * Because we need retrieve updates logs from all other MDTs.
1369                  * So if the MDS-MDS export is new, obd_max_recoverable_clients
1370                  * also needs to be increased to match other recovery checking
1371                  * condition. */
1372                 if (new_mds_mds_conn)
1373                         target->obd_max_recoverable_clients++;
1374                 if (atomic_inc_return(&target->obd_connected_clients) ==
1375                     target->obd_max_recoverable_clients)
1376                         wake_up(&target->obd_next_transno_waitq);
1377         }
1378
1379         /* Tell the client we're in recovery, when client is involved in it. */
1380         if (target->obd_recovering && !lw_client)
1381                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1382
1383 out:
1384         if (export) {
1385                 spin_lock(&export->exp_lock);
1386                 export->exp_connecting = 0;
1387                 spin_unlock(&export->exp_lock);
1388
1389                 class_export_put(export);
1390         }
1391         if (target_referenced == true && target != NULL) {
1392                 spin_lock(&target->obd_dev_lock);
1393                 target->obd_conn_inprogress--;
1394                 spin_unlock(&target->obd_dev_lock);
1395
1396                 class_decref(target, __func__, current);
1397         }
1398         req->rq_status = rc;
1399         RETURN(rc);
1400 }
1401
1402 int target_handle_disconnect(struct ptlrpc_request *req)
1403 {
1404         int rc;
1405         ENTRY;
1406
1407         rc = req_capsule_server_pack(&req->rq_pill);
1408         if (rc)
1409                 RETURN(rc);
1410
1411         /* Keep the rq_export around so we can send the reply. */
1412         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1413
1414         RETURN(0);
1415 }
1416
1417 void target_destroy_export(struct obd_export *exp)
1418 {
1419         struct obd_import       *imp = NULL;
1420         /* exports created from last_rcvd data, and "fake"
1421            exports created by lctl don't have an import */
1422         spin_lock(&exp->exp_lock);
1423         if (exp->exp_imp_reverse != NULL) {
1424                 imp = exp->exp_imp_reverse;
1425                 exp->exp_imp_reverse = NULL;
1426         }
1427         spin_unlock(&exp->exp_lock);
1428         if (imp != NULL)
1429                 client_destroy_import(imp);
1430
1431         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1432         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1433         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1434         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1435 }
1436 EXPORT_SYMBOL(target_destroy_export);
1437
1438 /*
1439  * Recovery functions
1440  */
1441 static void target_request_copy_get(struct ptlrpc_request *req)
1442 {
1443         class_export_rpc_inc(req->rq_export);
1444         LASSERT(list_empty(&req->rq_list));
1445         INIT_LIST_HEAD(&req->rq_replay_list);
1446
1447         /* Increase refcount to keep request in queue. */
1448         atomic_inc(&req->rq_refcount);
1449         /* Let export know it has replays to be handled. */
1450         atomic_inc(&req->rq_export->exp_replay_count);
1451 }
1452
1453 static void target_request_copy_put(struct ptlrpc_request *req)
1454 {
1455         LASSERT(list_empty(&req->rq_replay_list));
1456         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1457
1458         atomic_dec(&req->rq_export->exp_replay_count);
1459         class_export_rpc_dec(req->rq_export);
1460         ptlrpc_server_drop_request(req);
1461 }
1462
1463 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1464 {
1465         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1466         struct obd_export     *exp = req->rq_export;
1467         struct ptlrpc_request *reqiter;
1468         struct ptlrpc_request *dup_req = NULL;
1469         int                    dup = 0;
1470
1471         LASSERT(exp);
1472
1473         spin_lock(&exp->exp_lock);
1474         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1475                                 rq_replay_list) {
1476                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1477                         dup_req = reqiter;
1478                         dup = 1;
1479                         break;
1480                 }
1481         }
1482
1483         if (dup) {
1484                 /* We expect it with RESENT and REPLAY flags. */
1485                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1486                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1487                         CERROR("invalid flags %x of resent replay\n",
1488                                lustre_msg_get_flags(req->rq_reqmsg));
1489
1490                 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
1491                         __u32 new_conn;
1492
1493                         new_conn = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1494                         if (new_conn >
1495                             lustre_msg_get_conn_cnt(dup_req->rq_reqmsg))
1496                                 lustre_msg_set_conn_cnt(dup_req->rq_reqmsg,
1497                                                         new_conn);
1498                 }
1499         } else {
1500                 list_add_tail(&req->rq_replay_list,
1501                                   &exp->exp_req_replay_queue);
1502         }
1503
1504         spin_unlock(&exp->exp_lock);
1505         return dup;
1506 }
1507
1508 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1509 {
1510         LASSERT(!list_empty(&req->rq_replay_list));
1511         LASSERT(req->rq_export);
1512
1513         spin_lock(&req->rq_export->exp_lock);
1514         list_del_init(&req->rq_replay_list);
1515         spin_unlock(&req->rq_export->exp_lock);
1516 }
1517
1518 static void target_finish_recovery(struct lu_target *lut)
1519 {
1520         struct obd_device *obd = lut->lut_obd;
1521         ENTRY;
1522
1523         /* Only log a recovery message when recovery has occurred. */
1524         if (obd->obd_recovery_start) {
1525                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1526                                         obd->obd_recovery_start);
1527                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1528                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1529                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1530                         obd->obd_max_recoverable_clients,
1531                         atomic_read(&obd->obd_connected_clients),
1532                         obd->obd_stale_clients,
1533                         obd->obd_stale_clients == 1 ? "was" : "were");
1534         }
1535
1536         ldlm_reprocess_all_ns(obd->obd_namespace);
1537         spin_lock(&obd->obd_recovery_task_lock);
1538         if (!list_empty(&obd->obd_req_replay_queue) ||
1539             !list_empty(&obd->obd_lock_replay_queue) ||
1540             !list_empty(&obd->obd_final_req_queue)) {
1541                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1542                        obd->obd_name,
1543                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1544                        list_empty(&obd->obd_lock_replay_queue) ? \
1545                                "" : "lock ",
1546                        list_empty(&obd->obd_final_req_queue) ? \
1547                                "" : "final ");
1548                 spin_unlock(&obd->obd_recovery_task_lock);
1549                 LBUG();
1550         }
1551         spin_unlock(&obd->obd_recovery_task_lock);
1552
1553         if (lut->lut_tdtd != NULL &&
1554             (!list_empty(&lut->lut_tdtd->tdtd_replay_list) ||
1555             !list_empty(&lut->lut_tdtd->tdtd_replay_finish_list))) {
1556                 dtrq_list_dump(lut->lut_tdtd, D_ERROR);
1557                 dtrq_list_destroy(lut->lut_tdtd);
1558         }
1559
1560         obd->obd_recovery_end = cfs_time_current_sec();
1561
1562         /* When recovery finished, cleanup orphans on MDS and OST. */
1563         if (OBT(obd) && OBP(obd, postrecov)) {
1564                 int rc = OBP(obd, postrecov)(obd);
1565                 if (rc < 0)
1566                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1567                                       obd->obd_name, rc);
1568         }
1569         EXIT;
1570 }
1571
1572 static void abort_req_replay_queue(struct obd_device *obd)
1573 {
1574         struct ptlrpc_request *req, *n;
1575         struct list_head abort_list;
1576
1577         INIT_LIST_HEAD(&abort_list);
1578         spin_lock(&obd->obd_recovery_task_lock);
1579         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1580         spin_unlock(&obd->obd_recovery_task_lock);
1581         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1582                 DEBUG_REQ(D_WARNING, req, "aborted:");
1583                 req->rq_status = -ENOTCONN;
1584                 if (ptlrpc_error(req)) {
1585                         DEBUG_REQ(D_ERROR, req,
1586                                   "failed abort_req_reply; skipping");
1587                 }
1588                 target_exp_dequeue_req_replay(req);
1589                 target_request_copy_put(req);
1590         }
1591 }
1592
1593 static void abort_lock_replay_queue(struct obd_device *obd)
1594 {
1595         struct ptlrpc_request *req, *n;
1596         struct list_head abort_list;
1597
1598         INIT_LIST_HEAD(&abort_list);
1599         spin_lock(&obd->obd_recovery_task_lock);
1600         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1601         spin_unlock(&obd->obd_recovery_task_lock);
1602         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1603                 DEBUG_REQ(D_ERROR, req, "aborted:");
1604                 req->rq_status = -ENOTCONN;
1605                 if (ptlrpc_error(req)) {
1606                         DEBUG_REQ(D_ERROR, req,
1607                                   "failed abort_lock_reply; skipping");
1608                 }
1609                 target_request_copy_put(req);
1610         }
1611 }
1612
1613 /* Called from a cleanup function if the device is being cleaned up
1614    forcefully.  The exports should all have been disconnected already,
1615    the only thing left to do is
1616      - clear the recovery flags
1617      - cancel the timer
1618      - free queued requests and replies, but don't send replies
1619    Because the obd_stopping flag is set, no new requests should be received.
1620
1621 */
1622 void target_cleanup_recovery(struct obd_device *obd)
1623 {
1624         struct ptlrpc_request *req, *n;
1625         struct list_head clean_list;
1626
1627         INIT_LIST_HEAD(&clean_list);
1628         spin_lock(&obd->obd_dev_lock);
1629         if (!obd->obd_recovering) {
1630                 spin_unlock(&obd->obd_dev_lock);
1631                 EXIT;
1632                 return;
1633         }
1634         obd->obd_recovering = obd->obd_abort_recovery = 0;
1635         obd->obd_force_abort_recovery = 0;
1636         spin_unlock(&obd->obd_dev_lock);
1637
1638         spin_lock(&obd->obd_recovery_task_lock);
1639         target_cancel_recovery_timer(obd);
1640         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1641         spin_unlock(&obd->obd_recovery_task_lock);
1642
1643         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1644                 LASSERT(req->rq_reply_state == NULL);
1645                 target_exp_dequeue_req_replay(req);
1646                 target_request_copy_put(req);
1647         }
1648
1649         spin_lock(&obd->obd_recovery_task_lock);
1650         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1651         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1652         spin_unlock(&obd->obd_recovery_task_lock);
1653
1654         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1655                 LASSERT(req->rq_reply_state == NULL);
1656                 target_request_copy_put(req);
1657         }
1658
1659         EXIT;
1660 }
1661 EXPORT_SYMBOL(target_cleanup_recovery);
1662
1663 /* obd_recovery_task_lock should be held */
1664 void target_cancel_recovery_timer(struct obd_device *obd)
1665 {
1666         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1667         cfs_timer_disarm(&obd->obd_recovery_timer);
1668 }
1669
1670 static void target_start_recovery_timer(struct obd_device *obd)
1671 {
1672         if (obd->obd_recovery_start != 0)
1673                 return;
1674
1675         spin_lock(&obd->obd_dev_lock);
1676         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1677             obd->obd_force_abort_recovery) {
1678                 spin_unlock(&obd->obd_dev_lock);
1679                 return;
1680         }
1681
1682         LASSERT(obd->obd_recovery_timeout != 0);
1683
1684         if (obd->obd_recovery_start != 0) {
1685                 spin_unlock(&obd->obd_dev_lock);
1686                 return;
1687         }
1688
1689         cfs_timer_arm(&obd->obd_recovery_timer,
1690                       cfs_time_shift(obd->obd_recovery_timeout));
1691         obd->obd_recovery_start = cfs_time_current_sec();
1692         spin_unlock(&obd->obd_dev_lock);
1693
1694         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1695                       "or until %d client%s reconnect%s\n",
1696                       obd->obd_name,
1697                       obd->obd_recovery_timeout / 60,
1698                       obd->obd_recovery_timeout % 60,
1699                       obd->obd_max_recoverable_clients,
1700                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1701                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1702 }
1703
1704 /**
1705  * extend recovery window.
1706  *
1707  * if @extend is true, extend recovery window to have @drt remaining at least;
1708  * otherwise, make sure the recovery timeout value is not less than @drt.
1709  */
1710 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1711 {
1712         cfs_time_t now;
1713         cfs_time_t end;
1714         cfs_duration_t left;
1715         int to;
1716
1717         spin_lock(&obd->obd_dev_lock);
1718         if (!obd->obd_recovering || obd->obd_abort_recovery ||
1719             obd->obd_force_abort_recovery) {
1720                 spin_unlock(&obd->obd_dev_lock);
1721                 return;
1722         }
1723         LASSERT(obd->obd_recovery_start != 0);
1724
1725         now  = cfs_time_current_sec();
1726         to   = obd->obd_recovery_timeout;
1727         end  = obd->obd_recovery_start + to;
1728         left = cfs_time_sub(end, now);
1729
1730         if (extend && (drt > left)) {
1731                 to += drt - left;
1732         } else if (!extend && (drt > to)) {
1733                 to = drt;
1734         }
1735
1736         if (to > obd->obd_recovery_time_hard)
1737                 to = obd->obd_recovery_time_hard;
1738         if (obd->obd_recovery_timeout < to) {
1739                 obd->obd_recovery_timeout = to;
1740                 end = obd->obd_recovery_start + to;
1741                 cfs_timer_arm(&obd->obd_recovery_timer,
1742                                 cfs_time_shift(end - now));
1743         }
1744         spin_unlock(&obd->obd_dev_lock);
1745
1746         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1747                 obd->obd_name, (unsigned)cfs_time_sub(end, now));
1748 }
1749
1750 /* Reset the timer with each new client connection */
1751 /*
1752  * This timer is actually reconnect_timer, which is for making sure
1753  * the total recovery window is at least as big as my reconnect
1754  * attempt timing. So the initial recovery time_out will be set to
1755  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1756  * from client is bigger than this, then the recovery time_out will
1757  * be extended to make sure the client could be reconnected, in the
1758  * process, the timeout from the new client should be ignored.
1759  */
1760
1761 static void
1762 check_and_start_recovery_timer(struct obd_device *obd,
1763                                struct ptlrpc_request *req,
1764                                int new_client)
1765 {
1766         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1767         struct obd_device_target *obt = &obd->u.obt;
1768
1769         if (!new_client && service_time)
1770                 /* Teach server about old server's estimates, as first guess
1771                  * at how long new requests will take. */
1772                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1773                             service_time);
1774
1775         target_start_recovery_timer(obd);
1776
1777         /* Convert the service time to RPC timeout,
1778          * and reuse service_time to limit stack usage. */
1779         service_time = at_est2timeout(service_time);
1780
1781         /* We expect other clients to timeout within service_time, then try
1782          * to reconnect, then try the failover server.  The max delay between
1783          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1784         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1785
1786         LASSERT(obt->obt_magic == OBT_MAGIC);
1787         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1788         if (service_time > obd->obd_recovery_timeout && !new_client)
1789                 extend_recovery_timer(obd, service_time, false);
1790 }
1791
1792 /** Health checking routines */
1793 static inline int exp_connect_healthy(struct obd_export *exp)
1794 {
1795         return (exp->exp_in_recovery);
1796 }
1797
1798 /** if export done req_replay or has replay in queue */
1799 static inline int exp_req_replay_healthy(struct obd_export *exp)
1800 {
1801         return (!exp->exp_req_replay_needed ||
1802                 atomic_read(&exp->exp_replay_count) > 0);
1803 }
1804 /** if export done lock_replay or has replay in queue */
1805 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1806 {
1807         return (!exp->exp_lock_replay_needed ||
1808                 atomic_read(&exp->exp_replay_count) > 0);
1809 }
1810
1811 static inline int exp_vbr_healthy(struct obd_export *exp)
1812 {
1813         return (!exp->exp_vbr_failed);
1814 }
1815
1816 static inline int exp_finished(struct obd_export *exp)
1817 {
1818         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1819 }
1820
1821 static int check_for_next_transno(struct lu_target *lut)
1822 {
1823         struct ptlrpc_request *req = NULL;
1824         struct obd_device *obd = lut->lut_obd;
1825         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
1826         int wake_up = 0, connected, completed, queue_len;
1827         __u64 req_transno = 0;
1828         __u64 update_transno = 0;
1829         __u64 next_transno = 0;
1830         ENTRY;
1831
1832         spin_lock(&obd->obd_recovery_task_lock);
1833         if (!list_empty(&obd->obd_req_replay_queue)) {
1834                 req = list_entry(obd->obd_req_replay_queue.next,
1835                                      struct ptlrpc_request, rq_list);
1836                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1837         }
1838
1839         if (tdtd != NULL)
1840                 update_transno = distribute_txn_get_next_transno(tdtd);
1841
1842         connected = atomic_read(&obd->obd_connected_clients);
1843         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1844         queue_len = obd->obd_requests_queued_for_recovery;
1845         next_transno = obd->obd_next_recovery_transno;
1846
1847         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1848                "req_transno: "LPU64", next_transno: "LPU64"\n",
1849                obd->obd_max_recoverable_clients, connected, completed,
1850                queue_len, req_transno, next_transno);
1851
1852         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1853                 CDEBUG(D_HA, "waking for aborted recovery\n");
1854                 wake_up = 1;
1855         } else if (obd->obd_recovery_expired) {
1856                 CDEBUG(D_HA, "waking for expired recovery\n");
1857                 wake_up = 1;
1858         } else if (tdtd != NULL && req != NULL &&
1859                    is_req_replayed_by_update(req)) {
1860                 LASSERTF(req_transno < next_transno, "req_transno "LPU64
1861                          "next_transno"LPU64"\n", req_transno, next_transno);
1862                 CDEBUG(D_HA, "waking for duplicate req ("LPU64")\n",
1863                        req_transno);
1864                 wake_up = 1;
1865         } else if (req_transno == next_transno ||
1866                    (update_transno != 0 && update_transno <= next_transno)) {
1867                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1868                 wake_up = 1;
1869         } else if (queue_len > 0 &&
1870                    queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1871                 int d_lvl = D_HA;
1872                 /** handle gaps occured due to lost reply or VBR */
1873                 LASSERTF(req_transno >= next_transno,
1874                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1875                          req_transno, next_transno);
1876                 if (req_transno > obd->obd_last_committed &&
1877                     !obd->obd_version_recov)
1878                         d_lvl = D_ERROR;
1879                 CDEBUG(d_lvl,
1880                        "%s: waking for gap in transno, VBR is %s (skip: "
1881                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1882                        ", next_update "LPD64" last_committed: "LPD64")\n",
1883                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1884                        next_transno, queue_len, completed, connected,
1885                        req_transno, update_transno, obd->obd_last_committed);
1886                 obd->obd_next_recovery_transno = req_transno;
1887                 wake_up = 1;
1888         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1889                 CDEBUG(D_HA, "waking for completed recovery\n");
1890                 wake_up = 1;
1891         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1892                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1893                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1894                 obd->obd_next_recovery_transno = req_transno;
1895                 wake_up = 1;
1896         }
1897         spin_unlock(&obd->obd_recovery_task_lock);
1898         return wake_up;
1899 }
1900
1901 static int check_for_next_lock(struct lu_target *lut)
1902 {
1903         struct obd_device *obd = lut->lut_obd;
1904         int wake_up = 0;
1905
1906         spin_lock(&obd->obd_recovery_task_lock);
1907         if (!list_empty(&obd->obd_lock_replay_queue)) {
1908                 CDEBUG(D_HA, "waking for next lock\n");
1909                 wake_up = 1;
1910         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1911                 CDEBUG(D_HA, "waking for completed lock replay\n");
1912                 wake_up = 1;
1913         } else if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1914                 CDEBUG(D_HA, "waking for aborted recovery\n");
1915                 wake_up = 1;
1916         } else if (obd->obd_recovery_expired) {
1917                 CDEBUG(D_HA, "waking for expired recovery\n");
1918                 wake_up = 1;
1919         }
1920         spin_unlock(&obd->obd_recovery_task_lock);
1921
1922         return wake_up;
1923 }
1924
1925 /**
1926  * wait for recovery events,
1927  * check its status with help of check_routine
1928  * evict dead clients via health_check
1929  */
1930 static int target_recovery_overseer(struct lu_target *lut,
1931                                     int (*check_routine)(struct lu_target *),
1932                                     int (*health_check)(struct obd_export *))
1933 {
1934         struct obd_device       *obd = lut->lut_obd;
1935 repeat:
1936         if ((obd->obd_recovery_start != 0) && (cfs_time_current_sec() >=
1937               (obd->obd_recovery_start + obd->obd_recovery_time_hard))) {
1938                 CWARN("recovery is aborted by hard timeout\n");
1939                 obd->obd_abort_recovery = 1;
1940         }
1941
1942         while (wait_event_timeout(obd->obd_next_transno_waitq,
1943                                   check_routine(lut),
1944                                   msecs_to_jiffies(60 * MSEC_PER_SEC)) == 0)
1945                 /* wait indefinitely for event, but don't trigger watchdog */;
1946
1947         if (obd->obd_abort_recovery || obd->obd_force_abort_recovery) {
1948                 CWARN("recovery is aborted, evict exports in recovery\n");
1949                 /** evict exports which didn't finish recovery yet */
1950                 class_disconnect_stale_exports(obd, exp_finished);
1951                 return 1;
1952         } else if (obd->obd_recovery_expired) {
1953                 obd->obd_recovery_expired = 0;
1954                 /** If some clients died being recovered, evict them */
1955                 LCONSOLE_WARN("%s: recovery is timed out, "
1956                               "evict stale exports\n", obd->obd_name);
1957                 /** evict cexports with no replay in queue, they are stalled */
1958                 class_disconnect_stale_exports(obd, health_check);
1959                 /** continue with VBR */
1960                 spin_lock(&obd->obd_dev_lock);
1961                 obd->obd_version_recov = 1;
1962                 spin_unlock(&obd->obd_dev_lock);
1963                 /**
1964                  * reset timer, recovery will proceed with versions now,
1965                  * timeout is set just to handle reconnection delays
1966                  */
1967                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1968                 /** Wait for recovery events again, after evicting bad clients */
1969                 goto repeat;
1970         }
1971         return 0;
1972 }
1973
1974 static struct ptlrpc_request *target_next_replay_lock(struct lu_target *lut)
1975 {
1976         struct obd_device       *obd = lut->lut_obd;
1977         struct ptlrpc_request *req = NULL;
1978
1979         CDEBUG(D_HA, "Waiting for lock\n");
1980         if (target_recovery_overseer(lut, check_for_next_lock,
1981                                      exp_lock_replay_healthy))
1982                 abort_lock_replay_queue(obd);
1983
1984         spin_lock(&obd->obd_recovery_task_lock);
1985         if (!list_empty(&obd->obd_lock_replay_queue)) {
1986                 req = list_entry(obd->obd_lock_replay_queue.next,
1987                                      struct ptlrpc_request, rq_list);
1988                 list_del_init(&req->rq_list);
1989                 spin_unlock(&obd->obd_recovery_task_lock);
1990         } else {
1991                 spin_unlock(&obd->obd_recovery_task_lock);
1992                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1993                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1994                 /** evict exports failed VBR */
1995                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1996         }
1997         return req;
1998 }
1999
2000 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
2001 {
2002         struct ptlrpc_request *req = NULL;
2003
2004         spin_lock(&obd->obd_recovery_task_lock);
2005         if (!list_empty(&obd->obd_final_req_queue)) {
2006                 req = list_entry(obd->obd_final_req_queue.next,
2007                                      struct ptlrpc_request, rq_list);
2008                 list_del_init(&req->rq_list);
2009                 spin_unlock(&obd->obd_recovery_task_lock);
2010                 if (req->rq_export->exp_in_recovery) {
2011                         spin_lock(&req->rq_export->exp_lock);
2012                         req->rq_export->exp_in_recovery = 0;
2013                         spin_unlock(&req->rq_export->exp_lock);
2014                 }
2015         } else {
2016                 spin_unlock(&obd->obd_recovery_task_lock);
2017         }
2018         return req;
2019 }
2020
2021 static void handle_recovery_req(struct ptlrpc_thread *thread,
2022                                 struct ptlrpc_request *req,
2023                                 svc_handler_t handler)
2024 {
2025         ENTRY;
2026
2027         /**
2028          * export can be evicted during recovery, no need to handle replays for
2029          * it after that, discard such request silently
2030          */
2031         if (req->rq_export->exp_disconnected)
2032                 RETURN_EXIT;
2033
2034         req->rq_session.lc_thread = thread;
2035         req->rq_svc_thread = thread;
2036         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
2037
2038         /* thread context */
2039         lu_context_enter(&thread->t_env->le_ctx);
2040         (void)handler(req);
2041         lu_context_exit(&thread->t_env->le_ctx);
2042
2043         /* don't reset timer for final stage */
2044         if (!exp_finished(req->rq_export)) {
2045                 int to = obd_timeout;
2046
2047                 /**
2048                  * Add request timeout to the recovery time so next request from
2049                  * this client may come in recovery time
2050                  */
2051                 if (!AT_OFF) {
2052                         struct ptlrpc_service_part *svcpt;
2053
2054                         svcpt = req->rq_rqbd->rqbd_svcpt;
2055                         /* If the server sent early reply for this request,
2056                          * the client will recalculate the timeout according to
2057                          * current server estimate service time, so we will
2058                          * use the maxium timeout here for waiting the client
2059                          * sending the next req */
2060                         to = max((int)at_est2timeout(
2061                                  at_get(&svcpt->scp_at_estimate)),
2062                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
2063                         /* Add 2 net_latency, one for balance rq_deadline
2064                          * (see ptl_send_rpc), one for resend the req to server,
2065                          * Note: client will pack net_latency in replay req
2066                          * (see ptlrpc_replay_req) */
2067                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
2068                 }
2069                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
2070         }
2071         EXIT;
2072 }
2073
2074 /** Checking routines for recovery */
2075 static int check_for_recovery_ready(struct lu_target *lut)
2076 {
2077         struct obd_device *obd = lut->lut_obd;
2078         unsigned int clnts = atomic_read(&obd->obd_connected_clients);
2079
2080         CDEBUG(D_HA, "connected %d stale %d max_recoverable_clients %d"
2081                " abort %d expired %d\n", clnts, obd->obd_stale_clients,
2082                obd->obd_max_recoverable_clients, obd->obd_abort_recovery,
2083                obd->obd_recovery_expired);
2084
2085         if (obd->obd_force_abort_recovery)
2086                 return 1;
2087
2088         if (!obd->obd_abort_recovery && !obd->obd_recovery_expired) {
2089                 LASSERT(clnts <= obd->obd_max_recoverable_clients);
2090                 if (clnts + obd->obd_stale_clients <
2091                     obd->obd_max_recoverable_clients)
2092                         return 0;
2093         }
2094
2095         if (lut->lut_tdtd != NULL) {
2096                 if (!lut->lut_tdtd->tdtd_replay_ready) {
2097                         /* Let's extend recovery timer, in case the recovery
2098                          * timer expired, and some clients got evicted */
2099                         extend_recovery_timer(obd, obd->obd_recovery_timeout,
2100                                               true);
2101                         return 0;
2102                 } else {
2103                         dtrq_list_dump(lut->lut_tdtd, D_HA);
2104                 }
2105         }
2106
2107         return 1;
2108 }
2109
2110 enum {
2111         REQUEST_RECOVERY = 1,
2112         UPDATE_RECOVERY = 2,
2113 };
2114
2115 static __u64 get_next_replay_req_transno(struct obd_device *obd)
2116 {
2117         __u64 transno = 0;
2118
2119         if (!list_empty(&obd->obd_req_replay_queue)) {
2120                 struct ptlrpc_request *req;
2121
2122                 req = list_entry(obd->obd_req_replay_queue.next,
2123                                  struct ptlrpc_request, rq_list);
2124                 transno = lustre_msg_get_transno(req->rq_reqmsg);
2125         }
2126
2127         return transno;
2128 }
2129 __u64 get_next_transno(struct lu_target *lut, int *type)
2130 {
2131         struct obd_device *obd = lut->lut_obd;
2132         struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2133         __u64 transno = 0;
2134         __u64 update_transno;
2135         ENTRY;
2136
2137         transno = get_next_replay_req_transno(obd);
2138         if (type != NULL)
2139                 *type = REQUEST_RECOVERY;
2140
2141         if (tdtd == NULL)
2142                 RETURN(transno);
2143
2144         update_transno = distribute_txn_get_next_transno(tdtd);
2145         if (transno == 0 || (transno >= update_transno &&
2146                              update_transno != 0)) {
2147                 transno = update_transno;
2148                 if (type != NULL)
2149                         *type = UPDATE_RECOVERY;
2150         }
2151
2152         RETURN(transno);
2153 }
2154
2155 /**
2156  * drop duplicate replay request
2157  *
2158  * Because the operation has been replayed by update recovery, the request
2159  * with the same transno will be dropped and also notify the client to send
2160  * next replay request.
2161  *
2162  * \param[in] env       execution environment
2163  * \param[in] obd       failover obd device
2164  * \param[in] req       request to be dropped
2165  */
2166 static void drop_duplicate_replay_req(struct lu_env *env,
2167                                       struct obd_device *obd,
2168                                       struct ptlrpc_request *req)
2169 {
2170         DEBUG_REQ(D_HA, req, "remove t"LPD64" from %s because of duplicate"
2171                   " update records are found.\n",
2172                   lustre_msg_get_transno(req->rq_reqmsg),
2173                   libcfs_nid2str(req->rq_peer.nid));
2174
2175         /* Right now, only for MDS reint operation update replay and
2176          * normal request replay can have the same transno */
2177         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT) {
2178                 req_capsule_set(&req->rq_pill, &RQF_MDS_REINT);
2179                 req->rq_status = req_capsule_server_pack(&req->rq_pill);
2180                 if (likely(req->rq_export))
2181                         target_committed_to_req(req);
2182                 lustre_msg_set_transno(req->rq_repmsg, req->rq_transno);
2183                 target_send_reply(req, req->rq_status, 0);
2184         } else {
2185                 DEBUG_REQ(D_ERROR, req, "wrong opc" "from %s\n",
2186                 libcfs_nid2str(req->rq_peer.nid));
2187         }
2188         target_exp_dequeue_req_replay(req);
2189         target_request_copy_put(req);
2190         obd->obd_replayed_requests++;
2191 }
2192
2193 static void replay_request_or_update(struct lu_env *env,
2194                                      struct lu_target *lut,
2195                                      struct target_recovery_data *trd,
2196                                      struct ptlrpc_thread *thread)
2197 {
2198         struct obd_device *obd = lut->lut_obd;
2199         struct ptlrpc_request *req = NULL;
2200         int                     type;
2201         __u64                   transno;
2202         ENTRY;
2203
2204         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
2205                obd->obd_next_recovery_transno);
2206
2207         /* Replay all of request and update by transno */
2208         do {
2209                 struct target_distribute_txn_data *tdtd = lut->lut_tdtd;
2210
2211                 CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
2212
2213                 /** It is needed to extend recovery window above
2214                  *  recovery_time_soft. Extending is possible only in the
2215                  *  end of recovery window (see more details in
2216                  *  handle_recovery_req()).
2217                  */
2218                 CFS_FAIL_TIMEOUT_MS(OBD_FAIL_TGT_REPLAY_DELAY, 300);
2219
2220                 if (target_recovery_overseer(lut, check_for_next_transno,
2221                                              exp_req_replay_healthy)) {
2222                         abort_req_replay_queue(obd);
2223                         abort_lock_replay_queue(obd);
2224                 }
2225
2226                 spin_lock(&obd->obd_recovery_task_lock);
2227                 transno = get_next_transno(lut, &type);
2228                 if (type == REQUEST_RECOVERY && transno != 0) {
2229                         /* Drop replay request from client side, if the
2230                          * replay has been executed by update with the
2231                          * same transno */
2232                         req = list_entry(obd->obd_req_replay_queue.next,
2233                                         struct ptlrpc_request, rq_list);
2234
2235                         list_del_init(&req->rq_list);
2236                         obd->obd_requests_queued_for_recovery--;
2237                         spin_unlock(&obd->obd_recovery_task_lock);
2238
2239                         /* Let's check if the request has been redone by
2240                          * update replay */
2241                         if (is_req_replayed_by_update(req)) {
2242                                 struct distribute_txn_replay_req *dtrq;
2243
2244                                 dtrq = distribute_txn_lookup_finish_list(tdtd,
2245                                                                    req->rq_xid);
2246                                 LASSERT(dtrq != NULL);
2247                                 spin_lock(&tdtd->tdtd_replay_list_lock);
2248                                 list_del_init(&dtrq->dtrq_list);
2249                                 spin_unlock(&tdtd->tdtd_replay_list_lock);
2250                                 dtrq_destroy(dtrq);
2251
2252                                 drop_duplicate_replay_req(env, obd, req);
2253
2254                                 continue;
2255                         }
2256
2257                         LASSERT(trd->trd_processing_task == current_pid());
2258                         DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2259                                   lustre_msg_get_transno(req->rq_reqmsg),
2260                                   libcfs_nid2str(req->rq_peer.nid));
2261
2262                         handle_recovery_req(thread, req,
2263                                             trd->trd_recovery_handler);
2264                         /**
2265                          * bz18031: increase next_recovery_transno before
2266                          * target_request_copy_put() will drop exp_rpc reference
2267                          */
2268                         spin_lock(&obd->obd_recovery_task_lock);
2269                         obd->obd_next_recovery_transno++;
2270                         spin_unlock(&obd->obd_recovery_task_lock);
2271                         target_exp_dequeue_req_replay(req);
2272                         target_request_copy_put(req);
2273                         obd->obd_replayed_requests++;
2274                 } else if (type == UPDATE_RECOVERY && transno != 0) {
2275                         struct distribute_txn_replay_req *dtrq;
2276                         bool update_transno = false;
2277
2278                         spin_unlock(&obd->obd_recovery_task_lock);
2279
2280                         LASSERT(tdtd != NULL);
2281                         dtrq = distribute_txn_get_next_req(tdtd);
2282                         lu_context_enter(&thread->t_env->le_ctx);
2283                         tdtd->tdtd_replay_handler(env, tdtd, dtrq);
2284                         lu_context_exit(&thread->t_env->le_ctx);
2285                         extend_recovery_timer(obd, obd_timeout, true);
2286
2287                         /* Add it to the replay finish list */
2288                         spin_lock(&tdtd->tdtd_replay_list_lock);
2289                         if (dtrq->dtrq_xid != 0) {
2290                                 CDEBUG(D_HA, "Move x"LPU64" t"LPU64
2291                                        " to finish list\n", dtrq->dtrq_xid,
2292                                        dtrq->dtrq_master_transno);
2293                                 list_add(&dtrq->dtrq_list,
2294                                          &tdtd->tdtd_replay_finish_list);
2295                                 update_transno = true;
2296                         } else {
2297                                 dtrq_destroy(dtrq);
2298                         }
2299                         spin_unlock(&tdtd->tdtd_replay_list_lock);
2300
2301                         if (update_transno) {
2302                                 spin_lock(&obd->obd_recovery_task_lock);
2303                                 if (transno == obd->obd_next_recovery_transno)
2304                                         obd->obd_next_recovery_transno++;
2305                                 else if (transno >
2306                                          obd->obd_next_recovery_transno)
2307                                         obd->obd_next_recovery_transno =
2308                                                                 transno + 1;
2309                                 spin_unlock(&obd->obd_recovery_task_lock);
2310                         }
2311                 } else {
2312                         spin_unlock(&obd->obd_recovery_task_lock);
2313                         LASSERT(list_empty(&obd->obd_req_replay_queue));
2314                         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
2315                         /** evict exports failed VBR */
2316                         class_disconnect_stale_exports(obd, exp_vbr_healthy);
2317                         break;
2318                 }
2319         } while (1);
2320 }
2321
2322 static int target_recovery_thread(void *arg)
2323 {
2324         struct lu_target *lut = arg;
2325         struct obd_device *obd = lut->lut_obd;
2326         struct ptlrpc_request *req;
2327         struct target_recovery_data *trd = &obd->obd_recovery_data;
2328         unsigned long delta;
2329         struct lu_env *env;
2330         struct ptlrpc_thread *thread = NULL;
2331         int rc = 0;
2332         ENTRY;
2333
2334         unshare_fs_struct();
2335         OBD_ALLOC_PTR(thread);
2336         if (thread == NULL)
2337                 RETURN(-ENOMEM);
2338
2339         OBD_ALLOC_PTR(env);
2340         if (env == NULL) {
2341                 OBD_FREE_PTR(thread);
2342                 RETURN(-ENOMEM);
2343         }
2344
2345         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2346         if (rc) {
2347                 OBD_FREE_PTR(thread);
2348                 OBD_FREE_PTR(env);
2349                 RETURN(rc);
2350         }
2351
2352         thread->t_env = env;
2353         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2354         env->le_ctx.lc_thread = thread;
2355         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
2356         thread->t_watchdog = NULL;
2357
2358         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2359                current_pid());
2360         trd->trd_processing_task = current_pid();
2361
2362         spin_lock(&obd->obd_dev_lock);
2363         obd->obd_recovering = 1;
2364         spin_unlock(&obd->obd_dev_lock);
2365         complete(&trd->trd_starting);
2366
2367         /* first of all, we have to know the first transno to replay */
2368         if (target_recovery_overseer(lut, check_for_recovery_ready,
2369                                      exp_connect_healthy)) {
2370                 abort_req_replay_queue(obd);
2371                 abort_lock_replay_queue(obd);
2372                 if (lut->lut_tdtd != NULL)
2373                         dtrq_list_destroy(lut->lut_tdtd);
2374         }
2375
2376         /* next stage: replay requests or update */
2377         delta = jiffies;
2378         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2379                atomic_read(&obd->obd_req_replay_clients),
2380                obd->obd_next_recovery_transno);
2381         replay_request_or_update(env, lut, trd, thread);
2382
2383         /**
2384          * The second stage: replay locks
2385          */
2386         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2387                atomic_read(&obd->obd_lock_replay_clients));
2388         while ((req = target_next_replay_lock(lut))) {
2389                 LASSERT(trd->trd_processing_task == current_pid());
2390                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2391                           libcfs_nid2str(req->rq_peer.nid));
2392                 handle_recovery_req(thread, req,
2393                                     trd->trd_recovery_handler);
2394                 target_request_copy_put(req);
2395                 obd->obd_replayed_locks++;
2396         }
2397
2398         /**
2399          * The third stage: reply on final pings, at this moment all clients
2400          * must have request in final queue
2401          */
2402         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_RECONNECT, cfs_fail_val);
2403         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2404         /** Update server last boot epoch */
2405         tgt_boot_epoch_update(lut);
2406         /* We drop recoverying flag to forward all new requests
2407          * to regular mds_handle() since now */
2408         spin_lock(&obd->obd_dev_lock);
2409         obd->obd_recovering = obd->obd_abort_recovery = 0;
2410         spin_unlock(&obd->obd_dev_lock);
2411         spin_lock(&obd->obd_recovery_task_lock);
2412         target_cancel_recovery_timer(obd);
2413         spin_unlock(&obd->obd_recovery_task_lock);
2414         while ((req = target_next_final_ping(obd))) {
2415                 LASSERT(trd->trd_processing_task == current_pid());
2416                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2417                           libcfs_nid2str(req->rq_peer.nid));
2418                 handle_recovery_req(thread, req,
2419                                     trd->trd_recovery_handler);
2420                 /* Because the waiting client can not send ping to server,
2421                  * so we need refresh the last_request_time, to avoid the
2422                  * export is being evicted */
2423                 ptlrpc_update_export_timer(req->rq_export, 0);
2424                 target_request_copy_put(req);
2425         }
2426
2427         delta = jiffies_to_msecs(jiffies - delta) / MSEC_PER_SEC;
2428         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2429               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2430         if (delta > OBD_RECOVERY_TIME_SOFT) {
2431                 CWARN("too long recovery - read logs\n");
2432                 libcfs_debug_dumplog();
2433         }
2434
2435         target_finish_recovery(lut);
2436
2437         lu_context_fini(&env->le_ctx);
2438         trd->trd_processing_task = 0;
2439         complete(&trd->trd_finishing);
2440
2441         tgt_io_thread_done(thread);
2442         OBD_FREE_PTR(thread);
2443         OBD_FREE_PTR(env);
2444         RETURN(rc);
2445 }
2446
2447 static int target_start_recovery_thread(struct lu_target *lut,
2448                                         svc_handler_t handler)
2449 {
2450         struct obd_device *obd = lut->lut_obd;
2451         int rc = 0;
2452         struct target_recovery_data *trd = &obd->obd_recovery_data;
2453
2454         memset(trd, 0, sizeof(*trd));
2455         init_completion(&trd->trd_starting);
2456         init_completion(&trd->trd_finishing);
2457         trd->trd_recovery_handler = handler;
2458
2459         if (!IS_ERR(kthread_run(target_recovery_thread,
2460                                 lut, "tgt_recov"))) {
2461                 wait_for_completion(&trd->trd_starting);
2462                 LASSERT(obd->obd_recovering != 0);
2463         } else {
2464                 rc = -ECHILD;
2465         }
2466
2467         return rc;
2468 }
2469
2470 void target_stop_recovery_thread(struct obd_device *obd)
2471 {
2472         if (obd->obd_recovery_data.trd_processing_task > 0) {
2473                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2474                 /** recovery can be done but postrecovery is not yet */
2475                 spin_lock(&obd->obd_dev_lock);
2476                 if (obd->obd_recovering) {
2477                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2478                         obd->obd_abort_recovery = 1;
2479                         wake_up(&obd->obd_next_transno_waitq);
2480                 }
2481                 spin_unlock(&obd->obd_dev_lock);
2482                 wait_for_completion(&trd->trd_finishing);
2483         }
2484 }
2485 EXPORT_SYMBOL(target_stop_recovery_thread);
2486
2487 void target_recovery_fini(struct obd_device *obd)
2488 {
2489         class_disconnect_exports(obd);
2490         target_stop_recovery_thread(obd);
2491         target_cleanup_recovery(obd);
2492 }
2493 EXPORT_SYMBOL(target_recovery_fini);
2494
2495 static void target_recovery_expired(unsigned long castmeharder)
2496 {
2497         struct obd_device *obd = (struct obd_device *)castmeharder;
2498         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2499                " after %lds (%d clients connected)\n",
2500                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
2501                cfs_time_current_sec()- obd->obd_recovery_start,
2502                atomic_read(&obd->obd_connected_clients));
2503
2504         obd->obd_recovery_expired = 1;
2505         wake_up(&obd->obd_next_transno_waitq);
2506 }
2507
2508 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2509 {
2510         struct obd_device *obd = lut->lut_obd;
2511
2512         if (obd->obd_max_recoverable_clients == 0) {
2513                 /** Update server last boot epoch */
2514                 tgt_boot_epoch_update(lut);
2515                 return;
2516         }
2517
2518         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2519                "last_transno "LPU64"\n", obd->obd_name,
2520                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2521         LASSERT(obd->obd_stopping == 0);
2522         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2523         obd->obd_recovery_start = 0;
2524         obd->obd_recovery_end = 0;
2525
2526         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2527         target_start_recovery_thread(lut, handler);
2528 }
2529 EXPORT_SYMBOL(target_recovery_init);
2530
2531 static int target_process_req_flags(struct obd_device *obd,
2532                                     struct ptlrpc_request *req)
2533 {
2534         struct obd_export *exp = req->rq_export;
2535         LASSERT(exp != NULL);
2536         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2537                 /* client declares he's ready to replay locks */
2538                 spin_lock(&exp->exp_lock);
2539                 if (exp->exp_req_replay_needed) {
2540                         exp->exp_req_replay_needed = 0;
2541                         spin_unlock(&exp->exp_lock);
2542
2543                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2544                         atomic_dec(&obd->obd_req_replay_clients);
2545                 } else {
2546                         spin_unlock(&exp->exp_lock);
2547                 }
2548         }
2549         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2550                 /* client declares he's ready to complete recovery
2551                  * so, we put the request on th final queue */
2552                 spin_lock(&exp->exp_lock);
2553                 if (exp->exp_lock_replay_needed) {
2554                         exp->exp_lock_replay_needed = 0;
2555                         spin_unlock(&exp->exp_lock);
2556
2557                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2558                         atomic_dec(&obd->obd_lock_replay_clients);
2559                 } else {
2560                         spin_unlock(&exp->exp_lock);
2561                 }
2562         }
2563         return 0;
2564 }
2565
2566 int target_queue_recovery_request(struct ptlrpc_request *req,
2567                                   struct obd_device *obd)
2568 {
2569         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2570         struct ptlrpc_request *reqiter;
2571         int inserted = 0;
2572         ENTRY;
2573
2574         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2575                 /* Processing the queue right now, don't re-add. */
2576                 RETURN(1);
2577         }
2578
2579         target_process_req_flags(obd, req);
2580
2581         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2582                 /* client declares he's ready to complete recovery
2583                  * so, we put the request on th final queue */
2584                 target_request_copy_get(req);
2585                 DEBUG_REQ(D_HA, req, "queue final req");
2586                 wake_up(&obd->obd_next_transno_waitq);
2587                 spin_lock(&obd->obd_recovery_task_lock);
2588                 if (obd->obd_recovering) {
2589                         list_add_tail(&req->rq_list,
2590                                           &obd->obd_final_req_queue);
2591                 } else {
2592                         spin_unlock(&obd->obd_recovery_task_lock);
2593                         target_request_copy_put(req);
2594                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2595                 }
2596                 spin_unlock(&obd->obd_recovery_task_lock);
2597                 RETURN(0);
2598         }
2599         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2600                 /* client declares he's ready to replay locks */
2601                 target_request_copy_get(req);
2602                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2603                 wake_up(&obd->obd_next_transno_waitq);
2604                 spin_lock(&obd->obd_recovery_task_lock);
2605                 LASSERT(obd->obd_recovering);
2606                 /* usually due to recovery abort */
2607                 if (!req->rq_export->exp_in_recovery) {
2608                         spin_unlock(&obd->obd_recovery_task_lock);
2609                         target_request_copy_put(req);
2610                         RETURN(-ENOTCONN);
2611                 }
2612                 LASSERT(req->rq_export->exp_lock_replay_needed);
2613                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2614                 spin_unlock(&obd->obd_recovery_task_lock);
2615                 RETURN(0);
2616         }
2617
2618         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2619          * (i.e. buflens etc are in my own byte order), but type-dependent
2620          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2621
2622         if (!transno) {
2623                 INIT_LIST_HEAD(&req->rq_list);
2624                 DEBUG_REQ(D_HA, req, "not queueing");
2625                 RETURN(1);
2626         }
2627
2628         /* If we're processing the queue, we want don't want to queue this
2629          * message.
2630          *
2631          * Also, if this request has a transno less than the one we're waiting
2632          * for, we should process it now.  It could (and currently always will)
2633          * be an open request for a descriptor that was opened some time ago.
2634          *
2635          * Also, a resent, replayed request that has already been
2636          * handled will pass through here and be processed immediately.
2637          */
2638         CDEBUG(D_HA, "Next recovery transno: "LPU64
2639                ", current: "LPU64", replaying\n",
2640                obd->obd_next_recovery_transno, transno);
2641
2642         /* If the request has been replayed by update replay, then sends this
2643          * request to the recovery thread (replay_request_or_update()), where
2644          * it will be handled */
2645         spin_lock(&obd->obd_recovery_task_lock);
2646         if (transno < obd->obd_next_recovery_transno &&
2647             !is_req_replayed_by_update(req)) {
2648                 /* Processing the queue right now, don't re-add. */
2649                 LASSERT(list_empty(&req->rq_list));
2650                 spin_unlock(&obd->obd_recovery_task_lock);
2651                 RETURN(1);
2652         }
2653         spin_unlock(&obd->obd_recovery_task_lock);
2654
2655         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2656                 RETURN(0);
2657
2658         target_request_copy_get(req);
2659         if (!req->rq_export->exp_in_recovery) {
2660                 target_request_copy_put(req);
2661                 RETURN(-ENOTCONN);
2662         }
2663         LASSERT(req->rq_export->exp_req_replay_needed);
2664
2665         if (target_exp_enqueue_req_replay(req)) {
2666                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2667                 target_request_copy_put(req);
2668                 RETURN(0);
2669         }
2670
2671         /* XXX O(n^2) */
2672         spin_lock(&obd->obd_recovery_task_lock);
2673         LASSERT(obd->obd_recovering);
2674         list_for_each_entry(reqiter, &obd->obd_req_replay_queue, rq_list) {
2675                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2676                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2677                         inserted = 1;
2678                         goto added;
2679                 }
2680
2681                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2682                              transno)) {
2683                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2684                                   "has been claimed by another client");
2685                         spin_unlock(&obd->obd_recovery_task_lock);
2686                         target_exp_dequeue_req_replay(req);
2687                         target_request_copy_put(req);
2688                         RETURN(0);
2689                 }
2690         }
2691 added:
2692         if (!inserted)
2693                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2694
2695         obd->obd_requests_queued_for_recovery++;
2696         spin_unlock(&obd->obd_recovery_task_lock);
2697         wake_up(&obd->obd_next_transno_waitq);
2698         RETURN(0);
2699 }
2700
2701 int target_handle_ping(struct ptlrpc_request *req)
2702 {
2703         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2704         return req_capsule_server_pack(&req->rq_pill);
2705 }
2706
2707 void target_committed_to_req(struct ptlrpc_request *req)
2708 {
2709         struct obd_export *exp = req->rq_export;
2710
2711         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2712                 lustre_msg_set_last_committed(req->rq_repmsg,
2713                                               exp->exp_last_committed);
2714         else
2715                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2716                           "%d)", exp->exp_obd->obd_no_transno,
2717                           req->rq_repmsg == NULL);
2718
2719         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2720                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2721 }
2722
2723 #endif /* HAVE_SERVER_SUPPORT */
2724
2725 /**
2726  * Packs current SLV and Limit into \a req.
2727  */
2728 int target_pack_pool_reply(struct ptlrpc_request *req)
2729 {
2730         struct obd_device *obd;
2731         ENTRY;
2732
2733         /* Check that we still have all structures alive as this may
2734          * be some late RPC at shutdown time. */
2735         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2736                      !exp_connect_lru_resize(req->rq_export))) {
2737                 lustre_msg_set_slv(req->rq_repmsg, 0);
2738                 lustre_msg_set_limit(req->rq_repmsg, 0);
2739                 RETURN(0);
2740         }
2741
2742         /* OBD is alive here as export is alive, which we checked above. */
2743         obd = req->rq_export->exp_obd;
2744
2745         read_lock(&obd->obd_pool_lock);
2746         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2747         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2748         read_unlock(&obd->obd_pool_lock);
2749
2750         RETURN(0);
2751 }
2752
2753 static int target_send_reply_msg(struct ptlrpc_request *req,
2754                                  int rc, int fail_id)
2755 {
2756         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2757                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2758                 return -ECOMM;
2759         }
2760         /* We can have a null rq_reqmsg in the event of bad signature or
2761          * no context when unwrapping */
2762         if (req->rq_reqmsg &&
2763             unlikely(lustre_msg_get_opc(req->rq_reqmsg) == MDS_REINT &&
2764             OBD_FAIL_CHECK(OBD_FAIL_MDS_REINT_MULTI_NET_REP)))
2765                 return -ECOMM;
2766
2767         if (unlikely(rc)) {
2768                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2769                 req->rq_status = rc;
2770                 return ptlrpc_send_error(req, 1);
2771         } else {
2772                 DEBUG_REQ(D_NET, req, "sending reply");
2773         }
2774
2775         return ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT);
2776 }
2777
2778 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2779 {
2780         struct ptlrpc_service_part *svcpt;
2781         int                        netrc;
2782         struct ptlrpc_reply_state *rs;
2783         struct obd_export         *exp;
2784         ENTRY;
2785
2786         if (req->rq_no_reply) {
2787                 EXIT;
2788                 return;
2789         }
2790
2791         svcpt = req->rq_rqbd->rqbd_svcpt;
2792         rs = req->rq_reply_state;
2793         if (rs == NULL || !rs->rs_difficult) {
2794                 /* no notifiers */
2795                 target_send_reply_msg (req, rc, fail_id);
2796                 EXIT;
2797                 return;
2798         }
2799
2800         /* must be an export if locks saved */
2801         LASSERT(req->rq_export != NULL);
2802         /* req/reply consistent */
2803         LASSERT(rs->rs_svcpt == svcpt);
2804
2805         /* "fresh" reply */
2806         LASSERT(!rs->rs_scheduled);
2807         LASSERT(!rs->rs_scheduled_ever);
2808         LASSERT(!rs->rs_handled);
2809         LASSERT(!rs->rs_on_net);
2810         LASSERT(rs->rs_export == NULL);
2811         LASSERT(list_empty(&rs->rs_obd_list));
2812         LASSERT(list_empty(&rs->rs_exp_list));
2813
2814         exp = class_export_get(req->rq_export);
2815
2816         /* disable reply scheduling while I'm setting up */
2817         rs->rs_scheduled = 1;
2818         rs->rs_on_net    = 1;
2819         rs->rs_xid       = req->rq_xid;
2820         rs->rs_transno   = req->rq_transno;
2821         rs->rs_export    = exp;
2822         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2823
2824         spin_lock(&exp->exp_uncommitted_replies_lock);
2825         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2826                rs->rs_transno, exp->exp_last_committed);
2827         if (rs->rs_transno > exp->exp_last_committed) {
2828                 /* not committed already */
2829                 list_add_tail(&rs->rs_obd_list,
2830                                   &exp->exp_uncommitted_replies);
2831         }
2832         spin_unlock(&exp->exp_uncommitted_replies_lock);
2833
2834         spin_lock(&exp->exp_lock);
2835         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2836         spin_unlock(&exp->exp_lock);
2837
2838         netrc = target_send_reply_msg(req, rc, fail_id);
2839
2840         spin_lock(&svcpt->scp_rep_lock);
2841
2842         atomic_inc(&svcpt->scp_nreps_difficult);
2843
2844         if (netrc != 0) {
2845                 /* error sending: reply is off the net.  Also we need +1
2846                  * reply ref until ptlrpc_handle_rs() is done
2847                  * with the reply state (if the send was successful, there
2848                  * would have been +1 ref for the net, which
2849                  * reply_out_callback leaves alone) */
2850                 rs->rs_on_net = 0;
2851                 ptlrpc_rs_addref(rs);
2852         }
2853
2854         spin_lock(&rs->rs_lock);
2855         if (rs->rs_transno <= exp->exp_last_committed ||
2856             (!rs->rs_on_net && !rs->rs_no_ack) ||
2857             list_empty(&rs->rs_exp_list) ||     /* completed already */
2858             list_empty(&rs->rs_obd_list)) {
2859                 CDEBUG(D_HA, "Schedule reply immediately\n");
2860                 ptlrpc_dispatch_difficult_reply(rs);
2861         } else {
2862                 list_add(&rs->rs_list, &svcpt->scp_rep_active);
2863                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2864         }
2865         spin_unlock(&rs->rs_lock);
2866         spin_unlock(&svcpt->scp_rep_lock);
2867         EXIT;
2868 }
2869
2870 enum ldlm_mode lck_compat_array[] = {
2871         [LCK_EX]    = LCK_COMPAT_EX,
2872         [LCK_PW]    = LCK_COMPAT_PW,
2873         [LCK_PR]    = LCK_COMPAT_PR,
2874         [LCK_CW]    = LCK_COMPAT_CW,
2875         [LCK_CR]    = LCK_COMPAT_CR,
2876         [LCK_NL]    = LCK_COMPAT_NL,
2877         [LCK_GROUP] = LCK_COMPAT_GROUP,
2878         [LCK_COS]   = LCK_COMPAT_COS,
2879 };
2880
2881 /**
2882  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2883  * not escape to the user level.
2884  */
2885 int ldlm_error2errno(enum ldlm_error error)
2886 {
2887         int result;
2888
2889         switch (error) {
2890         case ELDLM_OK:
2891         case ELDLM_LOCK_MATCHED:
2892                 result = 0;
2893                 break;
2894         case ELDLM_LOCK_CHANGED:
2895                 result = -ESTALE;
2896                 break;
2897         case ELDLM_LOCK_ABORTED:
2898                 result = -ENAVAIL;
2899                 break;
2900         case ELDLM_LOCK_REPLACED:
2901                 result = -ESRCH;
2902                 break;
2903         case ELDLM_NO_LOCK_DATA:
2904                 result = -ENOENT;
2905                 break;
2906         case ELDLM_NAMESPACE_EXISTS:
2907                 result = -EEXIST;
2908                 break;
2909         case ELDLM_BAD_NAMESPACE:
2910                 result = -EBADF;
2911                 break;
2912         default:
2913                 if (((int)error) < 0) { /* cast to signed type */
2914                         result = error; /* as ldlm_error can be unsigned */
2915                 } else {
2916                         CERROR("Invalid DLM result code: %d\n", error);
2917                         result = -EPROTO;
2918                 }
2919         }
2920         return result;
2921 }
2922 EXPORT_SYMBOL(ldlm_error2errno);
2923
2924 /**
2925  * Dual to ldlm_error2errno(): maps errno values back to enum ldlm_error.
2926  */
2927 enum ldlm_error ldlm_errno2error(int err_no)
2928 {
2929         int error;
2930
2931         switch (err_no) {
2932         case 0:
2933                 error = ELDLM_OK;
2934                 break;
2935         case -ESTALE:
2936                 error = ELDLM_LOCK_CHANGED;
2937                 break;
2938         case -ENAVAIL:
2939                 error = ELDLM_LOCK_ABORTED;
2940                 break;
2941         case -ESRCH:
2942                 error = ELDLM_LOCK_REPLACED;
2943                 break;
2944         case -ENOENT:
2945                 error = ELDLM_NO_LOCK_DATA;
2946                 break;
2947         case -EEXIST:
2948                 error = ELDLM_NAMESPACE_EXISTS;
2949                 break;
2950         case -EBADF:
2951                 error = ELDLM_BAD_NAMESPACE;
2952                 break;
2953         default:
2954                 error = err_no;
2955         }
2956         return error;
2957 }
2958
2959 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2960 void ldlm_dump_export_locks(struct obd_export *exp)
2961 {
2962         spin_lock(&exp->exp_locks_list_guard);
2963         if (!list_empty(&exp->exp_locks_list)) {
2964                 struct ldlm_lock *lock;
2965
2966                 CERROR("dumping locks for export %p,"
2967                        "ignore if the unmount doesn't hang\n", exp);
2968                 list_for_each_entry(lock, &exp->exp_locks_list,
2969                                         l_exp_refs_link)
2970                         LDLM_ERROR(lock, "lock:");
2971         }
2972         spin_unlock(&exp->exp_locks_list_guard);
2973 }
2974 #endif
2975
2976 #ifdef HAVE_SERVER_SUPPORT
2977 static int target_bulk_timeout(void *data)
2978 {
2979         ENTRY;
2980         /* We don't fail the connection here, because having the export
2981          * killed makes the (vital) call to commitrw very sad.
2982          */
2983         RETURN(1);
2984 }
2985
2986 static inline const char *bulk2type(struct ptlrpc_request *req)
2987 {
2988         if (req->rq_bulk_read)
2989                 return "READ";
2990         if (req->rq_bulk_write)
2991                 return "WRITE";
2992         return "UNKNOWN";
2993 }
2994
2995 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2996                    struct l_wait_info *lwi)
2997 {
2998         struct ptlrpc_request   *req = desc->bd_req;
2999         time_t                   start = cfs_time_current_sec();
3000         time_t                   deadline;
3001         int                      rc = 0;
3002
3003         ENTRY;
3004
3005         /* If there is eviction in progress, wait for it to finish. */
3006         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
3007                 *lwi = LWI_INTR(NULL, NULL);
3008                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
3009                                   !atomic_read(&exp->exp_obd->
3010                                                    obd_evict_inprogress),
3011                                   lwi);
3012         }
3013
3014         /* Check if client was evicted or reconnected already. */
3015         if (exp->exp_failed ||
3016             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3017                 rc = -ENOTCONN;
3018         } else {
3019                 if (req->rq_bulk_read)
3020                         rc = sptlrpc_svc_wrap_bulk(req, desc);
3021
3022                 if ((exp->exp_connect_data.ocd_connect_flags &
3023                      OBD_CONNECT_BULK_MBITS) != 0)
3024                         req->rq_mbits = lustre_msg_get_mbits(req->rq_reqmsg);
3025                 else /* old version, bulk matchbits is rq_xid */
3026                         req->rq_mbits = req->rq_xid;
3027
3028                 if (rc == 0)
3029                         rc = ptlrpc_start_bulk_transfer(desc);
3030         }
3031
3032         if (rc < 0) {
3033                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
3034                           bulk2type(req), rc);
3035                 RETURN(rc);
3036         }
3037
3038         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
3039                 ptlrpc_abort_bulk(desc);
3040                 RETURN(0);
3041         }
3042
3043         /* limit actual bulk transfer to bulk_timeout seconds */
3044         deadline = start + bulk_timeout;
3045         if (deadline > req->rq_deadline)
3046                 deadline = req->rq_deadline;
3047
3048         do {
3049                 long timeoutl = deadline - cfs_time_current_sec();
3050                 cfs_duration_t timeout = timeoutl <= 0 ?
3051                                          CFS_TICK : cfs_time_seconds(timeoutl);
3052                 time_t  rq_deadline;
3053
3054                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
3055                                             target_bulk_timeout, desc);
3056                 rc = l_wait_event(desc->bd_waitq,
3057                                   !ptlrpc_server_bulk_active(desc) ||
3058                                   exp->exp_failed ||
3059                                   exp->exp_conn_cnt >
3060                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
3061                                   lwi);
3062                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
3063                 /* Wait again if we changed rq_deadline. */
3064                 rq_deadline = ACCESS_ONCE(req->rq_deadline);
3065                 deadline = start + bulk_timeout;
3066                 if (deadline > rq_deadline)
3067                         deadline = rq_deadline;
3068         } while ((rc == -ETIMEDOUT) &&
3069                  (deadline > cfs_time_current_sec()));
3070
3071         if (rc == -ETIMEDOUT) {
3072                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
3073                           bulk2type(req), deadline - start,
3074                           cfs_time_current_sec() - deadline);
3075                 ptlrpc_abort_bulk(desc);
3076         } else if (exp->exp_failed) {
3077                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
3078                           bulk2type(req));
3079                 rc = -ENOTCONN;
3080                 ptlrpc_abort_bulk(desc);
3081         } else if (exp->exp_conn_cnt >
3082                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
3083                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
3084                           bulk2type(req));
3085                 /* We don't reply anyway. */
3086                 rc = -ETIMEDOUT;
3087                 ptlrpc_abort_bulk(desc);
3088         } else if (desc->bd_failure) {
3089                 DEBUG_REQ(D_ERROR, req, "network error on bulk %s",
3090                           bulk2type(req));
3091                 /* XXX should this be a different errno? */
3092                 rc = -ETIMEDOUT;
3093         } else {
3094                 if (req->rq_bulk_write)
3095                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
3096                 if (rc == 0 && desc->bd_nob_transferred != desc->bd_nob) {
3097                         DEBUG_REQ(D_ERROR, req, "truncated bulk %s %d(%d)",
3098                                   bulk2type(req), desc->bd_nob_transferred,
3099                                   desc->bd_nob);
3100                         /* XXX should this be a different errno? */
3101                         rc = -ETIMEDOUT;
3102                 }
3103         }
3104
3105         RETURN(rc);
3106 }
3107 EXPORT_SYMBOL(target_bulk_io);
3108
3109 #endif /* HAVE_SERVER_SUPPORT */