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, 2013, 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 #ifdef __KERNEL__
  47 # include <libcfs/libcfs.h>
  48 #else
  49 # include <liblustre.h>
  50 #endif
  51 #include <obd.h>
  52 #include <obd_class.h>
  53 #include <lustre_dlm.h>
  54 #include <lustre_net.h>
  55 #include <lustre_sec.h>
  56 #include "ldlm_internal.h"
  57
  58 /* @priority: If non-zero, move the selected connection to the list head.
  59  * @create: If zero, only search in existing connections.
  60  */
  61 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  62                            int priority, int create)
  63 {
  64         struct ptlrpc_connection *ptlrpc_conn;
  65         struct obd_import_conn *imp_conn = NULL, *item;
  66         int rc = 0;
  67         ENTRY;
  68
  69         if (!create && !priority) {
  70                 CDEBUG(D_HA, "Nothing to do\n");
  71                 RETURN(-EINVAL);
  72         }
  73
  74         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  75         if (!ptlrpc_conn) {
  76                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  77                 RETURN (-ENOENT);
  78         }
  79
  80         if (create) {
  81                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  82                 if (!imp_conn) {
  83                         GOTO(out_put, rc = -ENOMEM);
  84                 }
  85         }
  86
  87         spin_lock(&imp->imp_lock);
  88         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  89                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  90                         if (priority) {
  91                                 cfs_list_del(&item->oic_item);
  92                                 cfs_list_add(&item->oic_item,
  93                                              &imp->imp_conn_list);
  94                                 item->oic_last_attempt = 0;
  95                         }
  96                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  97                                imp, imp->imp_obd->obd_name, uuid->uuid,
  98                                (priority ? ", moved to head" : ""));
  99                         spin_unlock(&imp->imp_lock);
 100                         GOTO(out_free, rc = 0);
 101                 }
 102         }
 103         /* No existing import connection found for \a uuid. */
 104         if (create) {
 105                 imp_conn->oic_conn = ptlrpc_conn;
 106                 imp_conn->oic_uuid = *uuid;
 107                 imp_conn->oic_last_attempt = 0;
 108                 if (priority)
 109                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 110                 else
 111                         cfs_list_add_tail(&imp_conn->oic_item,
 112                                           &imp->imp_conn_list);
 113                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 114                        imp, imp->imp_obd->obd_name, uuid->uuid,
 115                        (priority ? "head" : "tail"));
 116         } else {
 117                 spin_unlock(&imp->imp_lock);
 118                 GOTO(out_free, rc = -ENOENT);
 119         }
 120
 121         spin_unlock(&imp->imp_lock);
 122         RETURN(0);
 123 out_free:
 124         if (imp_conn)
 125                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 126 out_put:
 127         ptlrpc_connection_put(ptlrpc_conn);
 128         RETURN(rc);
 129 }
 130
 131 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 132 {
 133         return import_set_conn(imp, uuid, 1, 0);
 134 }
 135
 136 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 137                            int priority)
 138 {
 139         return import_set_conn(imp, uuid, priority, 1);
 140 }
 141 EXPORT_SYMBOL(client_import_add_conn);
 142
 143 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 144 {
 145         struct obd_import_conn *imp_conn;
 146         struct obd_export *dlmexp;
 147         int rc = -ENOENT;
 148         ENTRY;
 149
 150         spin_lock(&imp->imp_lock);
 151         if (cfs_list_empty(&imp->imp_conn_list)) {
 152                 LASSERT(!imp->imp_connection);
 153                 GOTO(out, rc);
 154         }
 155
 156         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 157                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 158                         continue;
 159                 LASSERT(imp_conn->oic_conn);
 160
 161                 if (imp_conn == imp->imp_conn_current) {
 162                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 163
 164                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 165                             imp->imp_state != LUSTRE_IMP_DISCON) {
 166                                 CERROR("can't remove current connection\n");
 167                                 GOTO(out, rc = -EBUSY);
 168                         }
 169
 170                         ptlrpc_connection_put(imp->imp_connection);
 171                         imp->imp_connection = NULL;
 172
 173                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 174                         if (dlmexp && dlmexp->exp_connection) {
 175                                 LASSERT(dlmexp->exp_connection ==
 176                                         imp_conn->oic_conn);
 177                                 ptlrpc_connection_put(dlmexp->exp_connection);
 178                                 dlmexp->exp_connection = NULL;
 179                         }
 180                 }
 181
 182                 cfs_list_del(&imp_conn->oic_item);
 183                 ptlrpc_connection_put(imp_conn->oic_conn);
 184                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 185                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 186                        imp, imp->imp_obd->obd_name, uuid->uuid);
 187                 rc = 0;
 188                 break;
 189         }
 190 out:
 191         spin_unlock(&imp->imp_lock);
 192         if (rc == -ENOENT)
 193                 CERROR("connection %s not found\n", uuid->uuid);
 194         RETURN(rc);
 195 }
 196 EXPORT_SYMBOL(client_import_del_conn);
 197
 198 /**
 199  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 200  * to find a conn uuid of \a imp which can reach \a peer.
 201  */
 202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 203                             struct obd_uuid *uuid)
 204 {
 205         struct obd_import_conn *conn;
 206         int rc = -ENOENT;
 207         ENTRY;
 208
 209         spin_lock(&imp->imp_lock);
 210         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 211                 /* Check if conn UUID does have this peer NID. */
 212                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 213                         *uuid = conn->oic_uuid;
 214                         rc = 0;
 215                         break;
 216                 }
 217         }
 218         spin_unlock(&imp->imp_lock);
 219         RETURN(rc);
 220 }
 221 EXPORT_SYMBOL(client_import_find_conn);
 222
 223 void client_destroy_import(struct obd_import *imp)
 224 {
 225         /* Drop security policy instance after all RPCs have finished/aborted
 226          * to let all busy contexts be released. */
 227         class_import_get(imp);
 228         class_destroy_import(imp);
 229         sptlrpc_import_sec_put(imp);
 230         class_import_put(imp);
 231 }
 232 EXPORT_SYMBOL(client_destroy_import);
 233
 234 /**
 235  * Check whether or not the OSC is on MDT.
 236  * In the config log,
 237  * osc on MDT
 238  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 239  * osc on client
 240  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 241  *
 242  **/
 243 static int osc_on_mdt(char *obdname)
 244 {
 245         char *ptr;
 246
 247         ptr = strrchr(obdname, '-');
 248         if (ptr == NULL)
 249                 return 0;
 250
 251         if (strncmp(ptr + 1, "MDT", 3) == 0)
 252                 return 1;
 253
 254         return 0;
 255 }
 256
 257 /* Configure an RPC client OBD device.
 258  *
 259  * lcfg parameters:
 260  * 1 - client UUID
 261  * 2 - server UUID
 262  * 3 - inactive-on-startup
 263  */
 264 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 265 {
 266         struct client_obd *cli = &obddev->u.cli;
 267         struct obd_import *imp;
 268         struct obd_uuid server_uuid;
 269         int rq_portal, rp_portal, connect_op;
 270         char *name = obddev->obd_type->typ_name;
 271         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 272         int rc;
 273         ENTRY;
 274
 275         /* In a more perfect world, we would hang a ptlrpc_client off of
 276          * obd_type and just use the values from there. */
 277         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 278                 rq_portal = OST_REQUEST_PORTAL;
 279                 rp_portal = OSC_REPLY_PORTAL;
 280                 connect_op = OST_CONNECT;
 281                 cli->cl_sp_me = LUSTRE_SP_CLI;
 282                 cli->cl_sp_to = LUSTRE_SP_OST;
 283                 ns_type = LDLM_NS_TYPE_OSC;
 284         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 285                    !strcmp(name, LUSTRE_LWP_NAME)) {
 286                 rq_portal = MDS_REQUEST_PORTAL;
 287                 rp_portal = MDC_REPLY_PORTAL;
 288                 connect_op = MDS_CONNECT;
 289                 cli->cl_sp_me = LUSTRE_SP_CLI;
 290                 cli->cl_sp_to = LUSTRE_SP_MDT;
 291                 ns_type = LDLM_NS_TYPE_MDC;
 292         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 293                 if (strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL) {
 294                         /* OSP_on_MDT for other MDTs */
 295                         connect_op = MDS_CONNECT;
 296                         cli->cl_sp_to = LUSTRE_SP_MDT;
 297                         ns_type = LDLM_NS_TYPE_MDC;
 298                         rq_portal = OUT_PORTAL;
 299                 } else {
 300                         /* OSP on MDT for OST */
 301                         connect_op = OST_CONNECT;
 302                         cli->cl_sp_to = LUSTRE_SP_OST;
 303                         ns_type = LDLM_NS_TYPE_OSC;
 304                         rq_portal = OST_REQUEST_PORTAL;
 305                 }
 306                 rp_portal = OSC_REPLY_PORTAL;
 307                 cli->cl_sp_me = LUSTRE_SP_CLI;
 308         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 309                 rq_portal = MGS_REQUEST_PORTAL;
 310                 rp_portal = MGC_REPLY_PORTAL;
 311                 connect_op = MGS_CONNECT;
 312                 cli->cl_sp_me = LUSTRE_SP_MGC;
 313                 cli->cl_sp_to = LUSTRE_SP_MGS;
 314                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 315                 ns_type = LDLM_NS_TYPE_MGC;
 316         } else {
 317                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 318                        name);
 319                 RETURN(-EINVAL);
 320         }
 321
 322         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 323                 CERROR("requires a TARGET UUID\n");
 324                 RETURN(-EINVAL);
 325         }
 326
 327         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 328                 CERROR("client UUID must be less than 38 characters\n");
 329                 RETURN(-EINVAL);
 330         }
 331
 332         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 333                 CERROR("setup requires a SERVER UUID\n");
 334                 RETURN(-EINVAL);
 335         }
 336
 337         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 338                 CERROR("target UUID must be less than 38 characters\n");
 339                 RETURN(-EINVAL);
 340         }
 341
 342         init_rwsem(&cli->cl_sem);
 343         sema_init(&cli->cl_mgc_sem, 1);
 344         cli->cl_conn_count = 0;
 345         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 346                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 347                      sizeof(server_uuid)));
 348
 349         cli->cl_dirty = 0;
 350         cli->cl_avail_grant = 0;
 351         /* FIXME: Should limit this for the sum of all cl_dirty_max. */
 352         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 353         if (cli->cl_dirty_max >> PAGE_CACHE_SHIFT > totalram_pages / 8)
 354                 cli->cl_dirty_max = totalram_pages << (PAGE_CACHE_SHIFT - 3);
 355         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 356         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 357         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 358         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 359         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 360         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 361         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
 362         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
 363         cli->cl_r_in_flight = 0;
 364         cli->cl_w_in_flight = 0;
 365
 366         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 367         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 368         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 369         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 370         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 371         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 372
 373         /* lru for osc. */
 374         CFS_INIT_LIST_HEAD(&cli->cl_lru_osc);
 375         cfs_atomic_set(&cli->cl_lru_shrinkers, 0);
 376         cfs_atomic_set(&cli->cl_lru_busy, 0);
 377         cfs_atomic_set(&cli->cl_lru_in_list, 0);
 378         CFS_INIT_LIST_HEAD(&cli->cl_lru_list);
 379         client_obd_list_lock_init(&cli->cl_lru_list_lock);
 380         cfs_atomic_set(&cli->cl_unstable_count, 0);
 381
 382         init_waitqueue_head(&cli->cl_destroy_waitq);
 383         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 384 #ifdef ENABLE_CHECKSUM
 385         /* Turn on checksumming by default. */
 386         cli->cl_checksum = 1;
 387         /*
 388          * The supported checksum types will be worked out at connect time
 389          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 390          * through procfs.
 391          */
 392         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 393 #endif
 394         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 395
 396         /* This value may be reduced at connect time in
 397          * ptlrpc_connect_interpret() . We initialize it to only
 398          * 1MB until we know what the performance looks like.
 399          * In the future this should likely be increased. LU-1431 */
 400         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
 401                                           LNET_MTU >> PAGE_CACHE_SHIFT);
 402
 403         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 404          * it will be updated at OSC connection time. */
 405         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 406
 407         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 408                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 409         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 410                 cli->cl_max_rpcs_in_flight = 2;
 411         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 412                 cli->cl_max_rpcs_in_flight = 3;
 413         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 414                 cli->cl_max_rpcs_in_flight = 4;
 415         } else {
 416                 if (osc_on_mdt(obddev->obd_name))
 417                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
 418                 else
 419                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 420         }
 421         rc = ldlm_get_ref();
 422         if (rc) {
 423                 CERROR("ldlm_get_ref failed: %d\n", rc);
 424                 GOTO(err, rc);
 425         }
 426
 427         ptlrpc_init_client(rq_portal, rp_portal, name,
 428                            &obddev->obd_ldlm_client);
 429
 430         imp = class_new_import(obddev);
 431         if (imp == NULL)
 432                 GOTO(err_ldlm, rc = -ENOENT);
 433         imp->imp_client = &obddev->obd_ldlm_client;
 434         imp->imp_connect_op = connect_op;
 435         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 436                LUSTRE_CFG_BUFLEN(lcfg, 1));
 437         class_import_put(imp);
 438
 439         rc = client_import_add_conn(imp, &server_uuid, 1);
 440         if (rc) {
 441                 CERROR("can't add initial connection\n");
 442                 GOTO(err_import, rc);
 443         }
 444
 445         cli->cl_import = imp;
 446         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 447         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 448         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 449
 450         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 451                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 452                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 453                                name, obddev->obd_name,
 454                                cli->cl_target_uuid.uuid);
 455                         spin_lock(&imp->imp_lock);
 456                         imp->imp_deactive = 1;
 457                         spin_unlock(&imp->imp_lock);
 458                 }
 459         }
 460
 461         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 462                                                    LDLM_NAMESPACE_CLIENT,
 463                                                    LDLM_NAMESPACE_GREEDY,
 464                                                    ns_type);
 465         if (obddev->obd_namespace == NULL) {
 466                 CERROR("Unable to create client namespace - %s\n",
 467                        obddev->obd_name);
 468                 GOTO(err_import, rc = -ENOMEM);
 469         }
 470
 471         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 472
 473         RETURN(rc);
 474
 475 err_import:
 476         class_destroy_import(imp);
 477 err_ldlm:
 478         ldlm_put_ref();
 479 err:
 480         RETURN(rc);
 481
 482 }
 483 EXPORT_SYMBOL(client_obd_setup);
 484
 485 int client_obd_cleanup(struct obd_device *obddev)
 486 {
 487         ENTRY;
 488
 489         ldlm_namespace_free_post(obddev->obd_namespace);
 490         obddev->obd_namespace = NULL;
 491
 492         obd_cleanup_client_import(obddev);
 493         LASSERT(obddev->u.cli.cl_import == NULL);
 494
 495         ldlm_put_ref();
 496         RETURN(0);
 497 }
 498 EXPORT_SYMBOL(client_obd_cleanup);
 499
 500 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 501 int client_connect_import(const struct lu_env *env,
 502                           struct obd_export **exp,
 503                           struct obd_device *obd, struct obd_uuid *cluuid,
 504                           struct obd_connect_data *data, void *localdata)
 505 {
 506         struct client_obd       *cli    = &obd->u.cli;
 507         struct obd_import       *imp    = cli->cl_import;
 508         struct obd_connect_data *ocd;
 509         struct lustre_handle    conn    = { 0 };
 510         int                     rc;
 511         ENTRY;
 512
 513         *exp = NULL;
 514         down_write(&cli->cl_sem);
 515         if (cli->cl_conn_count > 0 )
 516                 GOTO(out_sem, rc = -EALREADY);
 517
 518         rc = class_connect(&conn, obd, cluuid);
 519         if (rc)
 520                 GOTO(out_sem, rc);
 521
 522         cli->cl_conn_count++;
 523         *exp = class_conn2export(&conn);
 524
 525         LASSERT(obd->obd_namespace);
 526
 527         imp->imp_dlm_handle = conn;
 528         rc = ptlrpc_init_import(imp);
 529         if (rc != 0)
 530                 GOTO(out_ldlm, rc);
 531
 532         ocd = &imp->imp_connect_data;
 533         if (data) {
 534                 *ocd = *data;
 535                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 536         }
 537
 538         rc = ptlrpc_connect_import(imp);
 539         if (rc != 0) {
 540                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 541                 GOTO(out_ldlm, rc);
 542         }
 543         LASSERT((*exp)->exp_connection);
 544
 545         if (data) {
 546                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 547                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 548                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 549                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 550         }
 551
 552         ptlrpc_pinger_add_import(imp);
 553
 554         EXIT;
 555
 556         if (rc) {
 557 out_ldlm:
 558                 cli->cl_conn_count--;
 559                 class_disconnect(*exp);
 560                 *exp = NULL;
 561         }
 562 out_sem:
 563         up_write(&cli->cl_sem);
 564
 565         return rc;
 566 }
 567 EXPORT_SYMBOL(client_connect_import);
 568
 569 int client_disconnect_export(struct obd_export *exp)
 570 {
 571         struct obd_device *obd = class_exp2obd(exp);
 572         struct client_obd *cli;
 573         struct obd_import *imp;
 574         int rc = 0, err;
 575         ENTRY;
 576
 577         if (!obd) {
 578                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 579                        exp, exp ? exp->exp_handle.h_cookie : -1);
 580                 RETURN(-EINVAL);
 581         }
 582
 583         cli = &obd->u.cli;
 584         imp = cli->cl_import;
 585
 586         down_write(&cli->cl_sem);
 587         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 588                cli->cl_conn_count);
 589
 590         if (!cli->cl_conn_count) {
 591                 CERROR("disconnecting disconnected device (%s)\n",
 592                        obd->obd_name);
 593                 GOTO(out_disconnect, rc = -EINVAL);
 594         }
 595
 596         cli->cl_conn_count--;
 597         if (cli->cl_conn_count)
 598                 GOTO(out_disconnect, rc = 0);
 599
 600         /* Mark import deactivated now, so we don't try to reconnect if any
 601          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 602          * fully deactivate the import, or that would drop all requests. */
 603         spin_lock(&imp->imp_lock);
 604         imp->imp_deactive = 1;
 605         spin_unlock(&imp->imp_lock);
 606
 607         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 608          * delete it regardless.  (It's safe to delete an import that was
 609          * never added.) */
 610         (void)ptlrpc_pinger_del_import(imp);
 611
 612         if (obd->obd_namespace != NULL) {
 613                 /* obd_force == local only */
 614                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 615                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 616                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 617         }
 618
 619         /* There's no need to hold sem while disconnecting an import,
 620          * and it may actually cause deadlock in GSS. */
 621         up_write(&cli->cl_sem);
 622         rc = ptlrpc_disconnect_import(imp, 0);
 623         down_write(&cli->cl_sem);
 624
 625         ptlrpc_invalidate_import(imp);
 626
 627         EXIT;
 628
 629 out_disconnect:
 630         /* Use server style - class_disconnect should be always called for
 631          * o_disconnect. */
 632         err = class_disconnect(exp);
 633         if (!rc && err)
 634                 rc = err;
 635
 636         up_write(&cli->cl_sem);
 637
 638         RETURN(rc);
 639 }
 640 EXPORT_SYMBOL(client_disconnect_export);
 641
 642 #ifdef HAVE_SERVER_SUPPORT
 643 int server_disconnect_export(struct obd_export *exp)
 644 {
 645         int rc;
 646         ENTRY;
 647
 648         /* Disconnect early so that clients can't keep using export. */
 649         rc = class_disconnect(exp);
 650         /* Close import to avoid sending any requests. */
 651         if (exp->exp_imp_reverse)
 652                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 653
 654         if (exp->exp_obd->obd_namespace != NULL)
 655                 ldlm_cancel_locks_for_export(exp);
 656
 657         /* complete all outstanding replies */
 658         spin_lock(&exp->exp_lock);
 659         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 660                 struct ptlrpc_reply_state *rs =
 661                         cfs_list_entry(exp->exp_outstanding_replies.next,
 662                                        struct ptlrpc_reply_state, rs_exp_list);
 663                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 664
 665                 spin_lock(&svcpt->scp_rep_lock);
 666
 667                 cfs_list_del_init(&rs->rs_exp_list);
 668                 spin_lock(&rs->rs_lock);
 669                 ptlrpc_schedule_difficult_reply(rs);
 670                 spin_unlock(&rs->rs_lock);
 671
 672                 spin_unlock(&svcpt->scp_rep_lock);
 673         }
 674         spin_unlock(&exp->exp_lock);
 675
 676         RETURN(rc);
 677 }
 678 EXPORT_SYMBOL(server_disconnect_export);
 679
 680 /* --------------------------------------------------------------------------
 681  * from old lib/target.c
 682  * -------------------------------------------------------------------------- */
 683
 684 static int target_handle_reconnect(struct lustre_handle *conn,
 685                                    struct obd_export *exp,
 686                                    struct obd_uuid *cluuid)
 687 {
 688         ENTRY;
 689
 690         if (exp->exp_connection && exp->exp_imp_reverse) {
 691                 struct lustre_handle *hdl;
 692                 struct obd_device *target;
 693
 694                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 695                 target = exp->exp_obd;
 696
 697                 /* Might be a re-connect after a partition. */
 698                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 699                         if (target->obd_recovering) {
 700                                 int timeout = cfs_duration_sec(cfs_time_sub(
 701                                         cfs_timer_deadline(
 702                                         &target->obd_recovery_timer),
 703                                         cfs_time_current()));
 704
 705                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 706                                         "ing, waiting for %d clients in recov"
 707                                         "ery for %d:%.02d\n", target->obd_name,
 708                                         obd_uuid2str(&exp->exp_client_uuid),
 709                                         obd_export_nid2str(exp),
 710                                         target->obd_max_recoverable_clients,
 711                                         timeout / 60, timeout % 60);
 712                         } else {
 713                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 714                                         "reconnecting\n", target->obd_name,
 715                                         obd_uuid2str(&exp->exp_client_uuid),
 716                                         obd_export_nid2str(exp));
 717                         }
 718
 719                         conn->cookie = exp->exp_handle.h_cookie;
 720                         /* target_handle_connect() treats EALREADY and
 721                          * -EALREADY differently.  EALREADY means we are
 722                          * doing a valid reconnect from the same client. */
 723                         RETURN(EALREADY);
 724                 } else {
 725                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 726                                       "with handle "LPX64". Rejecting client "
 727                                       "with the same UUID trying to reconnect "
 728                                       "with handle "LPX64"\n", target->obd_name,
 729                                       obd_uuid2str(&exp->exp_client_uuid),
 730                                       obd_export_nid2str(exp),
 731                                       hdl->cookie, conn->cookie);
 732                         memset(conn, 0, sizeof *conn);
 733                         /* target_handle_connect() treats EALREADY and
 734                          * -EALREADY differently.  -EALREADY is an error
 735                          * (same UUID, different handle). */
 736                         RETURN(-EALREADY);
 737                 }
 738         }
 739
 740         conn->cookie = exp->exp_handle.h_cookie;
 741         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 742                cluuid->uuid, exp, conn->cookie);
 743         RETURN(0);
 744 }
 745
 746 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 747                           int error)
 748 {
 749         struct obd_export *exp = cb_data;
 750
 751         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 752                obd->obd_name, exp->exp_client_uuid.uuid);
 753
 754         spin_lock(&exp->exp_lock);
 755         exp->exp_need_sync = 0;
 756         spin_unlock(&exp->exp_lock);
 757         class_export_cb_put(exp);
 758 }
 759 EXPORT_SYMBOL(target_client_add_cb);
 760
 761 #ifdef __KERNEL__
 762 static void
 763 check_and_start_recovery_timer(struct obd_device *obd,
 764                                struct ptlrpc_request *req, int new_client);
 765 #else
 766 static inline void
 767 check_and_start_recovery_timer(struct obd_device *obd,
 768                                struct ptlrpc_request *req, int new_client)
 769 {
 770 }
 771 #endif
 772
 773 int target_handle_connect(struct ptlrpc_request *req)
 774 {
 775         struct obd_device *target = NULL, *targref = NULL;
 776         struct obd_export *export = NULL;
 777         struct obd_import *revimp;
 778         struct obd_import *tmp_imp = NULL;
 779         struct lustre_handle conn;
 780         struct lustre_handle *tmp;
 781         struct obd_uuid tgtuuid;
 782         struct obd_uuid cluuid;
 783         struct obd_uuid remote_uuid;
 784         char *str;
 785         int rc = 0;
 786         char *target_start;
 787         int target_len;
 788         bool     mds_conn = false, lw_client = false;
 789         struct obd_connect_data *data, *tmpdata;
 790         int size, tmpsize;
 791         lnet_nid_t *client_nid = NULL;
 792         ENTRY;
 793
 794         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 795
 796         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 797         if (str == NULL) {
 798                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 799                 GOTO(out, rc = -EINVAL);
 800         }
 801
 802         obd_str2uuid(&tgtuuid, str);
 803         target = class_uuid2obd(&tgtuuid);
 804         if (!target)
 805                 target = class_name2obd(str);
 806
 807         if (!target) {
 808                 deuuidify(str, NULL, &target_start, &target_len);
 809                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 810                                    "from %s (no target). If you are running "
 811                                    "an HA pair check that the target is "
 812                                    "mounted on the other server.\n", str,
 813                                    libcfs_nid2str(req->rq_peer.nid));
 814                 GOTO(out, rc = -ENODEV);
 815         }
 816
 817         spin_lock(&target->obd_dev_lock);
 818         if (target->obd_stopping || !target->obd_set_up) {
 819                 spin_unlock(&target->obd_dev_lock);
 820
 821                 deuuidify(str, NULL, &target_start, &target_len);
 822                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 823                               target_len, target_start,
 824                               libcfs_nid2str(req->rq_peer.nid),
 825                               (target->obd_stopping ?
 826                                "stopping" : "not set up"));
 827                 GOTO(out, rc = -ENODEV);
 828         }
 829
 830         if (target->obd_no_conn) {
 831                 spin_unlock(&target->obd_dev_lock);
 832
 833                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 834                                "from %s\n", target->obd_name,
 835                                libcfs_nid2str(req->rq_peer.nid));
 836                 GOTO(out, rc = -EAGAIN);
 837         }
 838
 839         /* Make sure the target isn't cleaned up while we're here. Yes,
 840          * there's still a race between the above check and our incref here.
 841          * Really, class_uuid2obd should take the ref. */
 842         targref = class_incref(target, __FUNCTION__, current);
 843
 844         target->obd_conn_inprogress++;
 845         spin_unlock(&target->obd_dev_lock);
 846
 847         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 848         if (str == NULL) {
 849                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 850                 GOTO(out, rc = -EINVAL);
 851         }
 852
 853         obd_str2uuid(&cluuid, str);
 854
 855         /* XXX Extract a nettype and format accordingly. */
 856         switch (sizeof(lnet_nid_t)) {
 857         /* NB the casts only avoid compiler warnings. */
 858         case 8:
 859                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 860                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 861                 break;
 862         case 4:
 863                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 864                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 865                 break;
 866         default:
 867                 LBUG();
 868         }
 869
 870         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 871         if (tmp == NULL)
 872                 GOTO(out, rc = -EPROTO);
 873
 874         conn = *tmp;
 875
 876         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 877                                     RCL_CLIENT);
 878         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 879         if (!data)
 880                 GOTO(out, rc = -EPROTO);
 881
 882         rc = req_capsule_server_pack(&req->rq_pill);
 883         if (rc)
 884                 GOTO(out, rc);
 885
 886         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 887                 if (data->ocd_version < LUSTRE_VERSION_CODE -
 888                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 889                     data->ocd_version > LUSTRE_VERSION_CODE +
 890                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 891                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 892                                   "libclient connection attempt",
 893                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 894                                   "old" : "new",
 895                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 896                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 897                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 898                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 899                         data = req_capsule_server_sized_get(&req->rq_pill,
 900                                                             &RMF_CONNECT_DATA,
 901                                     offsetof(typeof(*data), ocd_version) +
 902                                              sizeof(data->ocd_version));
 903                         if (data) {
 904                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 905                                 data->ocd_version = LUSTRE_VERSION_CODE;
 906                         }
 907                         GOTO(out, rc = -EPROTO);
 908                 }
 909         }
 910
 911         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 912             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 913                 mds_conn = true;
 914
 915         if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
 916                 lw_client = true;
 917
 918         /* lctl gets a backstage, all-access pass. */
 919         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 920                 goto dont_check_exports;
 921
 922         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 923         if (!export)
 924                 goto no_export;
 925
 926         /* We've found an export in the hash. */
 927
 928         spin_lock(&export->exp_lock);
 929
 930         if (export->exp_connecting) { /* bug 9635, et. al. */
 931                 spin_unlock(&export->exp_lock);
 932                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 933                               export->exp_obd->obd_name, export,
 934                               libcfs_nid2str(req->rq_peer.nid));
 935                 class_export_put(export);
 936                 export = NULL;
 937                 rc = -EALREADY;
 938         } else if (mds_conn && export->exp_connection) {
 939                 spin_unlock(&export->exp_lock);
 940                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 941                         /* MDS reconnected after failover. */
 942                         LCONSOLE_WARN("%s: Received MDS connection from "
 943                             "%s, removing former export from %s\n",
 944                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 945                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 946                 else
 947                         /* New MDS connection from the same NID. */
 948                         LCONSOLE_WARN("%s: Received new MDS connection from "
 949                             "%s, removing former export from same NID\n",
 950                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 951                 class_fail_export(export);
 952                 class_export_put(export);
 953                 export = NULL;
 954                 rc = 0;
 955         } else if (export->exp_connection != NULL &&
 956                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 957                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 958                     MSG_CONNECT_INITIAL)) {
 959                 spin_unlock(&export->exp_lock);
 960                 /* In MDS failover we have static UUID but NID can change. */
 961                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 962                               "existing nid %s is already connected\n",
 963                               target->obd_name, cluuid.uuid,
 964                               libcfs_nid2str(req->rq_peer.nid),
 965                               libcfs_nid2str(
 966                                       export->exp_connection->c_peer.nid));
 967                 rc = -EALREADY;
 968                 class_export_put(export);
 969                 export = NULL;
 970         } else {
 971                 export->exp_connecting = 1;
 972                 spin_unlock(&export->exp_lock);
 973                 LASSERT(export->exp_obd == target);
 974
 975                 rc = target_handle_reconnect(&conn, export, &cluuid);
 976         }
 977
 978         /* If we found an export, we already unlocked. */
 979         if (!export) {
 980 no_export:
 981                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 982         } else if (req->rq_export == NULL &&
 983                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 984                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 985                               "still busy with %d references\n",
 986                               target->obd_name, cluuid.uuid,
 987                               libcfs_nid2str(req->rq_peer.nid),
 988                               cfs_atomic_read(&export->exp_refcount));
 989                 GOTO(out, rc = -EBUSY);
 990         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 991                 if (!strstr(cluuid.uuid, "mdt"))
 992                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 993                                       "known client %s (at %s) because it "
 994                                       "is indicating it is a new client",
 995                                       target->obd_name, cluuid.uuid,
 996                                       libcfs_nid2str(req->rq_peer.nid));
 997                 GOTO(out, rc = -EALREADY);
 998         } else {
 999                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1000         }
1001
1002         if (rc < 0) {
1003                 GOTO(out, rc);
1004         }
1005
1006         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1007                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1008               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1009               export, (long)cfs_time_current_sec(),
1010               export ? (long)export->exp_last_request_time : 0);
1011
1012         /* If this is the first time a client connects, reset the recovery
1013          * timer. Discard lightweight connections which might be local. */
1014         if (!lw_client && rc == 0 && target->obd_recovering)
1015                 check_and_start_recovery_timer(target, req, export == NULL);
1016
1017         /* We want to handle EALREADY but *not* -EALREADY from
1018          * target_handle_reconnect(), return reconnection state in a flag. */
1019         if (rc == EALREADY) {
1020                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1021                 rc = 0;
1022         } else {
1023                 LASSERT(rc == 0);
1024         }
1025
1026         /* Tell the client if we support replayable requests. */
1027         if (target->obd_replayable)
1028                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1029         client_nid = &req->rq_peer.nid;
1030
1031         if (export == NULL) {
1032                 /* allow lightweight connections during recovery */
1033                 if (target->obd_recovering && !lw_client) {
1034                         cfs_time_t t;
1035                         int     c; /* connected */
1036                         int     i; /* in progress */
1037                         int     k; /* known */
1038                         int     s; /* stale/evicted */
1039
1040                         c = cfs_atomic_read(&target->obd_connected_clients);
1041                         i = cfs_atomic_read(&target->obd_lock_replay_clients);
1042                         k = target->obd_max_recoverable_clients;
1043                         s = target->obd_stale_clients;
1044                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1045                         t = cfs_time_sub(t, cfs_time_current());
1046                         t = cfs_duration_sec(t);
1047                         LCONSOLE_WARN("%s: Denying connection for new client "
1048                                       "%s (at %s), waiting for all %d known "
1049                                       "clients (%d recovered, %d in progress, "
1050                                       "and %d evicted) to recover in %d:%.02d\n",
1051                                       target->obd_name, cluuid.uuid,
1052                                       libcfs_nid2str(req->rq_peer.nid), k,
1053                                       c - i, i, s, (int)t / 60,
1054                                       (int)t % 60);
1055                         rc = -EBUSY;
1056                 } else {
1057 dont_check_exports:
1058                         rc = obd_connect(req->rq_svc_thread->t_env,
1059                                          &export, target, &cluuid, data,
1060                                          client_nid);
1061                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1062                                 lustre_msg_add_op_flags(req->rq_repmsg,
1063                                                 MSG_CONNECT_RECOVERING);
1064                         if (rc == 0)
1065                                 conn.cookie = export->exp_handle.h_cookie;
1066                 }
1067         } else {
1068                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1069                                    export, target, &cluuid, data, client_nid);
1070         }
1071         if (rc)
1072                 GOTO(out, rc);
1073
1074 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1075         /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1076          * will do the swabbing for if the client is using the same endianness.
1077          *
1078          * This fixup is version-limited, because we don't want to carry the
1079          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1080          * interop with unpatched 2.2 clients.  For newer clients, servers
1081          * will never do MNE swabbing, let the client handle that.  LU-1644 */
1082         export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1083                         !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1084 #else
1085 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1086 #endif
1087
1088         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1089         data->ocd_instance = target->u.obt.obt_instance;
1090
1091         /* Return only the parts of obd_connect_data that we understand, so the
1092          * client knows that we don't understand the rest. */
1093         if (data) {
1094                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1095                                                RCL_SERVER);
1096                 tmpdata = req_capsule_server_get(&req->rq_pill,
1097                                                  &RMF_CONNECT_DATA);
1098                 /* Don't use struct assignment here, because the client reply
1099                  * buffer may be smaller/larger than the local struct
1100                  * obd_connect_data. */
1101                 memcpy(tmpdata, data, min(tmpsize, size));
1102         }
1103
1104         /* If all else goes well, this is our RPC return code. */
1105         req->rq_status = 0;
1106
1107         lustre_msg_set_handle(req->rq_repmsg, &conn);
1108
1109         /* If the client and the server are the same node, we will already
1110          * have an export that really points to the client's DLM export,
1111          * because we have a shared handles table.
1112          *
1113          * XXX this will go away when shaver stops sending the "connect" handle
1114          * in the real "remote handle" field of the request --phik 24 Apr 2003
1115          */
1116         ptlrpc_request_change_export(req, export);
1117
1118         spin_lock(&export->exp_lock);
1119         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1120                 spin_unlock(&export->exp_lock);
1121                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1122                        "or equal conn_cnt: %d >= %d\n",
1123                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1124                        export->exp_conn_cnt,
1125                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1126
1127                 GOTO(out, rc = -EALREADY);
1128         }
1129         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1130         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1131
1132         /* Don't evict liblustre clients for not pinging. */
1133         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1134                 export->exp_libclient = 1;
1135                 spin_unlock(&export->exp_lock);
1136
1137                 spin_lock(&target->obd_dev_lock);
1138                 cfs_list_del_init(&export->exp_obd_chain_timed);
1139                 spin_unlock(&target->obd_dev_lock);
1140         } else {
1141                 spin_unlock(&export->exp_lock);
1142         }
1143
1144         if (export->exp_connection != NULL) {
1145                 /* Check to see if connection came from another NID. */
1146                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1147                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1148                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1149                                      &export->exp_connection->c_peer.nid,
1150                                      &export->exp_nid_hash);
1151
1152                 ptlrpc_connection_put(export->exp_connection);
1153         }
1154
1155         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1156                                                        req->rq_self,
1157                                                        &remote_uuid);
1158         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1159                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1160                              &export->exp_connection->c_peer.nid,
1161                              &export->exp_nid_hash);
1162         }
1163
1164         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1165                 int has_transno;
1166                 __u64 transno = data->ocd_transno;
1167
1168                 spin_lock(&export->exp_lock);
1169                 /* possible race with class_disconnect_stale_exports,
1170                  * export may be already in the eviction process */
1171                 if (export->exp_failed) {
1172                         spin_unlock(&export->exp_lock);
1173                         GOTO(out, rc = -ENODEV);
1174                 }
1175                 export->exp_in_recovery = 1;
1176                 export->exp_req_replay_needed = 1;
1177                 export->exp_lock_replay_needed = 1;
1178                 spin_unlock(&export->exp_lock);
1179
1180                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1181                                  MSG_CONNECT_TRANSNO);
1182                 if (has_transno && transno == 0)
1183                         CWARN("Connect with zero transno!\n");
1184
1185                 if (has_transno && transno > 0 &&
1186                     transno < target->obd_next_recovery_transno &&
1187                     transno > target->obd_last_committed) {
1188                         /* Another way is to use cmpxchg() to be lock-free. */
1189                         spin_lock(&target->obd_recovery_task_lock);
1190                         if (transno < target->obd_next_recovery_transno)
1191                                 target->obd_next_recovery_transno = transno;
1192                         spin_unlock(&target->obd_recovery_task_lock);
1193                 }
1194
1195                 cfs_atomic_inc(&target->obd_req_replay_clients);
1196                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1197                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1198                     target->obd_max_recoverable_clients)
1199                         wake_up(&target->obd_next_transno_waitq);
1200         }
1201
1202         /* Tell the client we're in recovery, when client is involved in it. */
1203         if (target->obd_recovering && !lw_client)
1204                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1205
1206         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1207         conn = *tmp;
1208
1209         /* Return -ENOTCONN in case of errors to let client reconnect. */
1210         revimp = class_new_import(target);
1211         if (revimp == NULL) {
1212                 CERROR("fail to alloc new reverse import.\n");
1213                 GOTO(out, rc = -ENOTCONN);
1214         }
1215
1216         spin_lock(&export->exp_lock);
1217         if (export->exp_imp_reverse != NULL)
1218                 /* destroyed import can be still referenced in ctxt */
1219                 tmp_imp = export->exp_imp_reverse;
1220         export->exp_imp_reverse = revimp;
1221         spin_unlock(&export->exp_lock);
1222
1223         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1224         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1225         revimp->imp_remote_handle = conn;
1226         revimp->imp_dlm_fake = 1;
1227         revimp->imp_state = LUSTRE_IMP_FULL;
1228
1229         /* Unknown versions will be caught in
1230          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1231         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1232
1233         if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1234             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1235                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1236         else
1237                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1238
1239         if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1240             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1241                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1242         else
1243                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1244
1245         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1246         if (rc) {
1247                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1248                 spin_lock(&export->exp_lock);
1249                 export->exp_imp_reverse = NULL;
1250                 spin_unlock(&export->exp_lock);
1251                 class_destroy_import(revimp);
1252         }
1253
1254         class_import_put(revimp);
1255
1256 out:
1257         if (tmp_imp != NULL)
1258                 client_destroy_import(tmp_imp);
1259         if (export) {
1260                 spin_lock(&export->exp_lock);
1261                 export->exp_connecting = 0;
1262                 spin_unlock(&export->exp_lock);
1263
1264                 class_export_put(export);
1265         }
1266         if (targref) {
1267                 spin_lock(&target->obd_dev_lock);
1268                 target->obd_conn_inprogress--;
1269                 spin_unlock(&target->obd_dev_lock);
1270
1271                 class_decref(targref, __func__, current);
1272         }
1273         if (rc)
1274                 req->rq_status = rc;
1275         RETURN(rc);
1276 }
1277 EXPORT_SYMBOL(target_handle_connect);
1278
1279 int target_handle_disconnect(struct ptlrpc_request *req)
1280 {
1281         int rc;
1282         ENTRY;
1283
1284         rc = req_capsule_server_pack(&req->rq_pill);
1285         if (rc)
1286                 RETURN(rc);
1287
1288         /* Keep the rq_export around so we can send the reply. */
1289         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1290
1291         RETURN(0);
1292 }
1293 EXPORT_SYMBOL(target_handle_disconnect);
1294
1295 void target_destroy_export(struct obd_export *exp)
1296 {
1297         struct obd_import       *imp = NULL;
1298         /* exports created from last_rcvd data, and "fake"
1299            exports created by lctl don't have an import */
1300         spin_lock(&exp->exp_lock);
1301         if (exp->exp_imp_reverse != NULL) {
1302                 imp = exp->exp_imp_reverse;
1303                 exp->exp_imp_reverse = NULL;
1304         }
1305         spin_unlock(&exp->exp_lock);
1306         if (imp != NULL)
1307                 client_destroy_import(imp);
1308
1309         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1310         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1311         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1312         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1313 }
1314 EXPORT_SYMBOL(target_destroy_export);
1315
1316 /*
1317  * Recovery functions
1318  */
1319 static void target_request_copy_get(struct ptlrpc_request *req)
1320 {
1321         class_export_rpc_inc(req->rq_export);
1322         LASSERT(cfs_list_empty(&req->rq_list));
1323         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1324
1325         /* Increase refcount to keep request in queue. */
1326         cfs_atomic_inc(&req->rq_refcount);
1327         /* Let export know it has replays to be handled. */
1328         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1329 }
1330
1331 static void target_request_copy_put(struct ptlrpc_request *req)
1332 {
1333         LASSERT(cfs_list_empty(&req->rq_replay_list));
1334         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1335
1336         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1337         class_export_rpc_dec(req->rq_export);
1338         ptlrpc_server_drop_request(req);
1339 }
1340
1341 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1342 {
1343         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1344         struct obd_export     *exp = req->rq_export;
1345         struct ptlrpc_request *reqiter;
1346         int                    dup = 0;
1347
1348         LASSERT(exp);
1349
1350         spin_lock(&exp->exp_lock);
1351         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1352                                 rq_replay_list) {
1353                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1354                         dup = 1;
1355                         break;
1356                 }
1357         }
1358
1359         if (dup) {
1360                 /* We expect it with RESENT and REPLAY flags. */
1361                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1362                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1363                         CERROR("invalid flags %x of resent replay\n",
1364                                lustre_msg_get_flags(req->rq_reqmsg));
1365         } else {
1366                 cfs_list_add_tail(&req->rq_replay_list,
1367                                   &exp->exp_req_replay_queue);
1368         }
1369
1370         spin_unlock(&exp->exp_lock);
1371         return dup;
1372 }
1373
1374 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1375 {
1376         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1377         LASSERT(req->rq_export);
1378
1379         spin_lock(&req->rq_export->exp_lock);
1380         cfs_list_del_init(&req->rq_replay_list);
1381         spin_unlock(&req->rq_export->exp_lock);
1382 }
1383
1384 #ifdef __KERNEL__
1385 static void target_finish_recovery(struct obd_device *obd)
1386 {
1387         ENTRY;
1388
1389         /* Only log a recovery message when recovery has occurred. */
1390         if (obd->obd_recovery_start) {
1391                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1392                                         obd->obd_recovery_start);
1393                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1394                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1395                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1396                         obd->obd_max_recoverable_clients,
1397                         cfs_atomic_read(&obd->obd_connected_clients),
1398                         obd->obd_stale_clients,
1399                         obd->obd_stale_clients == 1 ? "was" : "were");
1400         }
1401
1402         ldlm_reprocess_all_ns(obd->obd_namespace);
1403         spin_lock(&obd->obd_recovery_task_lock);
1404         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1405             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1406             !cfs_list_empty(&obd->obd_final_req_queue)) {
1407                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1408                        obd->obd_name,
1409                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1410                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1411                                "" : "lock ",
1412                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1413                                "" : "final ");
1414                 spin_unlock(&obd->obd_recovery_task_lock);
1415                 LBUG();
1416         }
1417         spin_unlock(&obd->obd_recovery_task_lock);
1418
1419         obd->obd_recovery_end = cfs_time_current_sec();
1420
1421         /* When recovery finished, cleanup orphans on MDS and OST. */
1422         if (OBT(obd) && OBP(obd, postrecov)) {
1423                 int rc = OBP(obd, postrecov)(obd);
1424                 if (rc < 0)
1425                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1426                                       obd->obd_name, rc);
1427         }
1428         EXIT;
1429 }
1430
1431 static void abort_req_replay_queue(struct obd_device *obd)
1432 {
1433         struct ptlrpc_request *req, *n;
1434         cfs_list_t abort_list;
1435
1436         CFS_INIT_LIST_HEAD(&abort_list);
1437         spin_lock(&obd->obd_recovery_task_lock);
1438         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1439         spin_unlock(&obd->obd_recovery_task_lock);
1440         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1441                 DEBUG_REQ(D_WARNING, req, "aborted:");
1442                 req->rq_status = -ENOTCONN;
1443                 if (ptlrpc_error(req)) {
1444                         DEBUG_REQ(D_ERROR, req,
1445                                   "failed abort_req_reply; skipping");
1446                 }
1447                 target_exp_dequeue_req_replay(req);
1448                 target_request_copy_put(req);
1449         }
1450 }
1451
1452 static void abort_lock_replay_queue(struct obd_device *obd)
1453 {
1454         struct ptlrpc_request *req, *n;
1455         cfs_list_t abort_list;
1456
1457         CFS_INIT_LIST_HEAD(&abort_list);
1458         spin_lock(&obd->obd_recovery_task_lock);
1459         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1460         spin_unlock(&obd->obd_recovery_task_lock);
1461         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1462                 DEBUG_REQ(D_ERROR, req, "aborted:");
1463                 req->rq_status = -ENOTCONN;
1464                 if (ptlrpc_error(req)) {
1465                         DEBUG_REQ(D_ERROR, req,
1466                                   "failed abort_lock_reply; skipping");
1467                 }
1468                 target_request_copy_put(req);
1469         }
1470 }
1471
1472 /* Called from a cleanup function if the device is being cleaned up
1473    forcefully.  The exports should all have been disconnected already,
1474    the only thing left to do is
1475      - clear the recovery flags
1476      - cancel the timer
1477      - free queued requests and replies, but don't send replies
1478    Because the obd_stopping flag is set, no new requests should be received.
1479
1480 */
1481 void target_cleanup_recovery(struct obd_device *obd)
1482 {
1483         struct ptlrpc_request *req, *n;
1484         cfs_list_t clean_list;
1485         ENTRY;
1486
1487         CFS_INIT_LIST_HEAD(&clean_list);
1488         spin_lock(&obd->obd_dev_lock);
1489         if (!obd->obd_recovering) {
1490                 spin_unlock(&obd->obd_dev_lock);
1491                 EXIT;
1492                 return;
1493         }
1494         obd->obd_recovering = obd->obd_abort_recovery = 0;
1495         spin_unlock(&obd->obd_dev_lock);
1496
1497         spin_lock(&obd->obd_recovery_task_lock);
1498         target_cancel_recovery_timer(obd);
1499         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1500         spin_unlock(&obd->obd_recovery_task_lock);
1501
1502         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1503                 LASSERT(req->rq_reply_state == 0);
1504                 target_exp_dequeue_req_replay(req);
1505                 target_request_copy_put(req);
1506         }
1507
1508         spin_lock(&obd->obd_recovery_task_lock);
1509         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1510         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1511         spin_unlock(&obd->obd_recovery_task_lock);
1512
1513         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1514                 LASSERT(req->rq_reply_state == 0);
1515                 target_request_copy_put(req);
1516         }
1517
1518         EXIT;
1519 }
1520 EXPORT_SYMBOL(target_cleanup_recovery);
1521
1522 /* obd_recovery_task_lock should be held */
1523 void target_cancel_recovery_timer(struct obd_device *obd)
1524 {
1525         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1526         cfs_timer_disarm(&obd->obd_recovery_timer);
1527 }
1528 EXPORT_SYMBOL(target_cancel_recovery_timer);
1529
1530 static void target_start_recovery_timer(struct obd_device *obd)
1531 {
1532         if (obd->obd_recovery_start != 0)
1533                 return;
1534
1535         spin_lock(&obd->obd_dev_lock);
1536         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1537                 spin_unlock(&obd->obd_dev_lock);
1538                 return;
1539         }
1540
1541         LASSERT(obd->obd_recovery_timeout != 0);
1542
1543         if (obd->obd_recovery_start != 0) {
1544                 spin_unlock(&obd->obd_dev_lock);
1545                 return;
1546         }
1547
1548         cfs_timer_arm(&obd->obd_recovery_timer,
1549                       cfs_time_shift(obd->obd_recovery_timeout));
1550         obd->obd_recovery_start = cfs_time_current_sec();
1551         spin_unlock(&obd->obd_dev_lock);
1552
1553         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1554                       "or until %d client%s reconnect%s\n",
1555                       obd->obd_name,
1556                       obd->obd_recovery_timeout / 60,
1557                       obd->obd_recovery_timeout % 60,
1558                       obd->obd_max_recoverable_clients,
1559                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1560                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1561 }
1562
1563 /**
1564  * extend recovery window.
1565  *
1566  * if @extend is true, extend recovery window to have @drt remaining at least;
1567  * otherwise, make sure the recovery timeout value is not less than @drt.
1568  */
1569 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1570 {
1571         cfs_time_t now;
1572         cfs_time_t end;
1573         cfs_duration_t left;
1574         int to;
1575
1576         spin_lock(&obd->obd_dev_lock);
1577         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1578                 spin_unlock(&obd->obd_dev_lock);
1579                 return;
1580         }
1581         LASSERT(obd->obd_recovery_start != 0);
1582
1583         now  = cfs_time_current_sec();
1584         to   = obd->obd_recovery_timeout;
1585         end  = obd->obd_recovery_start + to;
1586         left = cfs_time_sub(end, now);
1587
1588         if (extend && (drt > left)) {
1589                 to += drt - left;
1590         } else if (!extend && (drt > to)) {
1591                 to = drt;
1592                 /* reduce drt by already passed time */
1593                 drt -= obd->obd_recovery_timeout - left;
1594         }
1595
1596         if (to > obd->obd_recovery_time_hard)
1597                 to = obd->obd_recovery_time_hard;
1598         if (obd->obd_recovery_timeout < to ||
1599             obd->obd_recovery_timeout == obd->obd_recovery_time_hard) {
1600                 obd->obd_recovery_timeout = to;
1601                 cfs_timer_arm(&obd->obd_recovery_timer,
1602                               cfs_time_shift(drt));
1603         }
1604         spin_unlock(&obd->obd_dev_lock);
1605
1606         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1607                obd->obd_name, (unsigned)drt);
1608 }
1609
1610 /* Reset the timer with each new client connection */
1611 /*
1612  * This timer is actually reconnect_timer, which is for making sure
1613  * the total recovery window is at least as big as my reconnect
1614  * attempt timing. So the initial recovery time_out will be set to
1615  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1616  * from client is bigger than this, then the recovery time_out will
1617  * be extended to make sure the client could be reconnected, in the
1618  * process, the timeout from the new client should be ignored.
1619  */
1620
1621 static void
1622 check_and_start_recovery_timer(struct obd_device *obd,
1623                                struct ptlrpc_request *req,
1624                                int new_client)
1625 {
1626         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1627         struct obd_device_target *obt = &obd->u.obt;
1628
1629         if (!new_client && service_time)
1630                 /* Teach server about old server's estimates, as first guess
1631                  * at how long new requests will take. */
1632                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1633                             service_time);
1634
1635         target_start_recovery_timer(obd);
1636
1637         /* Convert the service time to RPC timeout,
1638          * and reuse service_time to limit stack usage. */
1639         service_time = at_est2timeout(service_time);
1640
1641         /* We expect other clients to timeout within service_time, then try
1642          * to reconnect, then try the failover server.  The max delay between
1643          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1644         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1645
1646         LASSERT(obt->obt_magic == OBT_MAGIC);
1647         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1648         if (service_time > obd->obd_recovery_timeout && !new_client)
1649                 extend_recovery_timer(obd, service_time, false);
1650 }
1651
1652 /** Health checking routines */
1653 static inline int exp_connect_healthy(struct obd_export *exp)
1654 {
1655         return (exp->exp_in_recovery);
1656 }
1657
1658 /** if export done req_replay or has replay in queue */
1659 static inline int exp_req_replay_healthy(struct obd_export *exp)
1660 {
1661         return (!exp->exp_req_replay_needed ||
1662                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1663 }
1664 /** if export done lock_replay or has replay in queue */
1665 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1666 {
1667         return (!exp->exp_lock_replay_needed ||
1668                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1669 }
1670
1671 static inline int exp_vbr_healthy(struct obd_export *exp)
1672 {
1673         return (!exp->exp_vbr_failed);
1674 }
1675
1676 static inline int exp_finished(struct obd_export *exp)
1677 {
1678         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1679 }
1680
1681 /** Checking routines for recovery */
1682 static int check_for_clients(struct obd_device *obd)
1683 {
1684         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1685
1686         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1687                 return 1;
1688         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1689         return (clnts + obd->obd_stale_clients ==
1690                 obd->obd_max_recoverable_clients);
1691 }
1692
1693 static int check_for_next_transno(struct obd_device *obd)
1694 {
1695         struct ptlrpc_request *req = NULL;
1696         int wake_up = 0, connected, completed, queue_len;
1697         __u64 next_transno, req_transno;
1698         ENTRY;
1699
1700         spin_lock(&obd->obd_recovery_task_lock);
1701         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1702                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1703                                      struct ptlrpc_request, rq_list);
1704                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1705         } else {
1706                 req_transno = 0;
1707         }
1708
1709         connected = cfs_atomic_read(&obd->obd_connected_clients);
1710         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1711         queue_len = obd->obd_requests_queued_for_recovery;
1712         next_transno = obd->obd_next_recovery_transno;
1713
1714         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1715                "req_transno: "LPU64", next_transno: "LPU64"\n",
1716                obd->obd_max_recoverable_clients, connected, completed,
1717                queue_len, req_transno, next_transno);
1718
1719         if (obd->obd_abort_recovery) {
1720                 CDEBUG(D_HA, "waking for aborted recovery\n");
1721                 wake_up = 1;
1722         } else if (obd->obd_recovery_expired) {
1723                 CDEBUG(D_HA, "waking for expired recovery\n");
1724                 wake_up = 1;
1725         } else if (req_transno == next_transno) {
1726                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1727                 wake_up = 1;
1728         } else if (queue_len > 0 &&
1729                    queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1730                 int d_lvl = D_HA;
1731                 /** handle gaps occured due to lost reply or VBR */
1732                 LASSERTF(req_transno >= next_transno,
1733                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1734                          req_transno, next_transno);
1735                 if (req_transno > obd->obd_last_committed &&
1736                     !obd->obd_version_recov)
1737                         d_lvl = D_ERROR;
1738                 CDEBUG(d_lvl,
1739                        "%s: waking for gap in transno, VBR is %s (skip: "
1740                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1741                        ", last_committed: "LPD64")\n",
1742                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1743                        next_transno, queue_len, completed, connected,
1744                        req_transno, obd->obd_last_committed);
1745                 obd->obd_next_recovery_transno = req_transno;
1746                 wake_up = 1;
1747         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1748                 CDEBUG(D_HA, "waking for completed recovery\n");
1749                 wake_up = 1;
1750         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1751                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1752                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1753                 obd->obd_next_recovery_transno = req_transno;
1754                 wake_up = 1;
1755         }
1756         spin_unlock(&obd->obd_recovery_task_lock);
1757         return wake_up;
1758 }
1759
1760 static int check_for_next_lock(struct obd_device *obd)
1761 {
1762         int wake_up = 0;
1763
1764         spin_lock(&obd->obd_recovery_task_lock);
1765         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1766                 CDEBUG(D_HA, "waking for next lock\n");
1767                 wake_up = 1;
1768         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1769                 CDEBUG(D_HA, "waking for completed lock replay\n");
1770                 wake_up = 1;
1771         } else if (obd->obd_abort_recovery) {
1772                 CDEBUG(D_HA, "waking for aborted recovery\n");
1773                 wake_up = 1;
1774         } else if (obd->obd_recovery_expired) {
1775                 CDEBUG(D_HA, "waking for expired recovery\n");
1776                 wake_up = 1;
1777         }
1778         spin_unlock(&obd->obd_recovery_task_lock);
1779
1780         return wake_up;
1781 }
1782
1783 /**
1784  * wait for recovery events,
1785  * check its status with help of check_routine
1786  * evict dead clients via health_check
1787  */
1788 static int target_recovery_overseer(struct obd_device *obd,
1789                                     int (*check_routine)(struct obd_device *),
1790                                     int (*health_check)(struct obd_export *))
1791 {
1792 repeat:
1793         wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1794         if (obd->obd_abort_recovery) {
1795                 CWARN("recovery is aborted, evict exports in recovery\n");
1796                 /** evict exports which didn't finish recovery yet */
1797                 class_disconnect_stale_exports(obd, exp_finished);
1798                 return 1;
1799         } else if (obd->obd_recovery_expired) {
1800                 obd->obd_recovery_expired = 0;
1801                 /** If some clients died being recovered, evict them */
1802                 LCONSOLE_WARN("%s: recovery is timed out, "
1803                               "evict stale exports\n", obd->obd_name);
1804                 /** evict cexports with no replay in queue, they are stalled */
1805                 class_disconnect_stale_exports(obd, health_check);
1806                 /** continue with VBR */
1807                 spin_lock(&obd->obd_dev_lock);
1808                 obd->obd_version_recov = 1;
1809                 spin_unlock(&obd->obd_dev_lock);
1810                 /**
1811                  * reset timer, recovery will proceed with versions now,
1812                  * timeout is set just to handle reconnection delays
1813                  */
1814                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1815                 /** Wait for recovery events again, after evicting bad clients */
1816                 goto repeat;
1817         }
1818         return 0;
1819 }
1820
1821 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1822 {
1823         struct ptlrpc_request *req = NULL;
1824         ENTRY;
1825
1826         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1827                obd->obd_next_recovery_transno);
1828
1829         if (target_recovery_overseer(obd, check_for_next_transno,
1830                                      exp_req_replay_healthy)) {
1831                 abort_req_replay_queue(obd);
1832                 abort_lock_replay_queue(obd);
1833         }
1834
1835         spin_lock(&obd->obd_recovery_task_lock);
1836         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1837                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1838                                      struct ptlrpc_request, rq_list);
1839                 cfs_list_del_init(&req->rq_list);
1840                 obd->obd_requests_queued_for_recovery--;
1841                 spin_unlock(&obd->obd_recovery_task_lock);
1842         } else {
1843                 spin_unlock(&obd->obd_recovery_task_lock);
1844                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1845                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1846                 /** evict exports failed VBR */
1847                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1848         }
1849         RETURN(req);
1850 }
1851
1852 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1853 {
1854         struct ptlrpc_request *req = NULL;
1855
1856         CDEBUG(D_HA, "Waiting for lock\n");
1857         if (target_recovery_overseer(obd, check_for_next_lock,
1858                                      exp_lock_replay_healthy))
1859                 abort_lock_replay_queue(obd);
1860
1861         spin_lock(&obd->obd_recovery_task_lock);
1862         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1863                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1864                                      struct ptlrpc_request, rq_list);
1865                 cfs_list_del_init(&req->rq_list);
1866                 spin_unlock(&obd->obd_recovery_task_lock);
1867         } else {
1868                 spin_unlock(&obd->obd_recovery_task_lock);
1869                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1870                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1871                 /** evict exports failed VBR */
1872                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1873         }
1874         return req;
1875 }
1876
1877 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1878 {
1879         struct ptlrpc_request *req = NULL;
1880
1881         spin_lock(&obd->obd_recovery_task_lock);
1882         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1883                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1884                                      struct ptlrpc_request, rq_list);
1885                 cfs_list_del_init(&req->rq_list);
1886                 spin_unlock(&obd->obd_recovery_task_lock);
1887                 if (req->rq_export->exp_in_recovery) {
1888                         spin_lock(&req->rq_export->exp_lock);
1889                         req->rq_export->exp_in_recovery = 0;
1890                         spin_unlock(&req->rq_export->exp_lock);
1891                 }
1892         } else {
1893                 spin_unlock(&obd->obd_recovery_task_lock);
1894         }
1895         return req;
1896 }
1897
1898 static int handle_recovery_req(struct ptlrpc_thread *thread,
1899                                struct ptlrpc_request *req,
1900                                svc_handler_t handler)
1901 {
1902         int rc;
1903         ENTRY;
1904
1905         /**
1906          * export can be evicted during recovery, no need to handle replays for
1907          * it after that, discard such request silently
1908          */
1909         if (req->rq_export->exp_disconnected)
1910                 GOTO(reqcopy_put, rc = 0);
1911
1912         rc = lu_context_init(&req->rq_recov_session, LCT_SERVER_SESSION);
1913         if (rc) {
1914                 CERROR("Failure to initialize session: %d\n", rc);
1915                 GOTO(reqcopy_put, rc);
1916         }
1917
1918         req->rq_recov_session.lc_thread = thread;
1919         lu_context_enter(&req->rq_recov_session);
1920         req->rq_svc_thread = thread;
1921         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1922
1923         /* thread context */
1924         lu_context_enter(&thread->t_env->le_ctx);
1925         (void)handler(req);
1926         lu_context_exit(&thread->t_env->le_ctx);
1927
1928         lu_context_exit(&req->rq_recov_session);
1929         lu_context_fini(&req->rq_recov_session);
1930         /* don't reset timer for final stage */
1931         if (!exp_finished(req->rq_export)) {
1932                 int to = obd_timeout;
1933
1934                 /**
1935                  * Add request timeout to the recovery time so next request from
1936                  * this client may come in recovery time
1937                  */
1938                 if (!AT_OFF) {
1939                         struct ptlrpc_service_part *svcpt;
1940
1941                         svcpt = req->rq_rqbd->rqbd_svcpt;
1942                         /* If the server sent early reply for this request,
1943                          * the client will recalculate the timeout according to
1944                          * current server estimate service time, so we will
1945                          * use the maxium timeout here for waiting the client
1946                          * sending the next req */
1947                         to = max((int)at_est2timeout(
1948                                  at_get(&svcpt->scp_at_estimate)),
1949                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1950                         /* Add 2 net_latency, one for balance rq_deadline
1951                          * (see ptl_send_rpc), one for resend the req to server,
1952                          * Note: client will pack net_latency in replay req
1953                          * (see ptlrpc_replay_req) */
1954                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1955                 }
1956                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1957         }
1958 reqcopy_put:
1959         RETURN(rc);
1960 }
1961
1962 static int target_recovery_thread(void *arg)
1963 {
1964         struct lu_target *lut = arg;
1965         struct obd_device *obd = lut->lut_obd;
1966         struct ptlrpc_request *req;
1967         struct target_recovery_data *trd = &obd->obd_recovery_data;
1968         unsigned long delta;
1969         struct lu_env *env;
1970         struct ptlrpc_thread *thread = NULL;
1971         int rc = 0;
1972         ENTRY;
1973
1974         unshare_fs_struct();
1975         OBD_ALLOC_PTR(thread);
1976         if (thread == NULL)
1977                 RETURN(-ENOMEM);
1978
1979         OBD_ALLOC_PTR(env);
1980         if (env == NULL) {
1981                 OBD_FREE_PTR(thread);
1982                 RETURN(-ENOMEM);
1983         }
1984
1985         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1986         if (rc) {
1987                 OBD_FREE_PTR(thread);
1988                 OBD_FREE_PTR(env);
1989                 RETURN(rc);
1990         }
1991
1992         thread->t_env = env;
1993         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1994         env->le_ctx.lc_thread = thread;
1995         tgt_io_thread_init(thread); /* init thread_big_cache for IO requests */
1996         thread->t_watchdog = NULL;
1997
1998         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1999                current_pid());
2000         trd->trd_processing_task = current_pid();
2001
2002         spin_lock(&obd->obd_dev_lock);
2003         obd->obd_recovering = 1;
2004         spin_unlock(&obd->obd_dev_lock);
2005         complete(&trd->trd_starting);
2006
2007         /* first of all, we have to know the first transno to replay */
2008         if (target_recovery_overseer(obd, check_for_clients,
2009                                      exp_connect_healthy)) {
2010                 abort_req_replay_queue(obd);
2011                 abort_lock_replay_queue(obd);
2012         }
2013
2014         /* next stage: replay requests */
2015         delta = jiffies;
2016         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2017                cfs_atomic_read(&obd->obd_req_replay_clients),
2018                obd->obd_next_recovery_transno);
2019         while ((req = target_next_replay_req(obd))) {
2020                 LASSERT(trd->trd_processing_task == current_pid());
2021                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2022                           lustre_msg_get_transno(req->rq_reqmsg),
2023                           libcfs_nid2str(req->rq_peer.nid));
2024                 handle_recovery_req(thread, req,
2025                                     trd->trd_recovery_handler);
2026                 /**
2027                  * bz18031: increase next_recovery_transno before
2028                  * target_request_copy_put() will drop exp_rpc reference
2029                  */
2030                 spin_lock(&obd->obd_recovery_task_lock);
2031                 obd->obd_next_recovery_transno++;
2032                 spin_unlock(&obd->obd_recovery_task_lock);
2033                 target_exp_dequeue_req_replay(req);
2034                 target_request_copy_put(req);
2035                 obd->obd_replayed_requests++;
2036         }
2037
2038         /**
2039          * The second stage: replay locks
2040          */
2041         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2042                cfs_atomic_read(&obd->obd_lock_replay_clients));
2043         while ((req = target_next_replay_lock(obd))) {
2044                 LASSERT(trd->trd_processing_task == current_pid());
2045                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2046                           libcfs_nid2str(req->rq_peer.nid));
2047                 handle_recovery_req(thread, req,
2048                                     trd->trd_recovery_handler);
2049                 target_request_copy_put(req);
2050                 obd->obd_replayed_locks++;
2051         }
2052
2053         /**
2054          * The third stage: reply on final pings, at this moment all clients
2055          * must have request in final queue
2056          */
2057         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2058         /** Update server last boot epoch */
2059         tgt_boot_epoch_update(lut);
2060         /* We drop recoverying flag to forward all new requests
2061          * to regular mds_handle() since now */
2062         spin_lock(&obd->obd_dev_lock);
2063         obd->obd_recovering = obd->obd_abort_recovery = 0;
2064         spin_unlock(&obd->obd_dev_lock);
2065         spin_lock(&obd->obd_recovery_task_lock);
2066         target_cancel_recovery_timer(obd);
2067         spin_unlock(&obd->obd_recovery_task_lock);
2068         while ((req = target_next_final_ping(obd))) {
2069                 LASSERT(trd->trd_processing_task == current_pid());
2070                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2071                           libcfs_nid2str(req->rq_peer.nid));
2072                 handle_recovery_req(thread, req,
2073                                     trd->trd_recovery_handler);
2074                 target_request_copy_put(req);
2075         }
2076
2077         delta = (jiffies - delta) / HZ;
2078         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2079               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2080         if (delta > OBD_RECOVERY_TIME_SOFT) {
2081                 CWARN("too long recovery - read logs\n");
2082                 libcfs_debug_dumplog();
2083         }
2084
2085         target_finish_recovery(obd);
2086
2087         lu_context_fini(&env->le_ctx);
2088         trd->trd_processing_task = 0;
2089         complete(&trd->trd_finishing);
2090
2091         tgt_io_thread_done(thread);
2092         OBD_FREE_PTR(thread);
2093         OBD_FREE_PTR(env);
2094         RETURN(rc);
2095 }
2096
2097 static int target_start_recovery_thread(struct lu_target *lut,
2098                                         svc_handler_t handler)
2099 {
2100         struct obd_device *obd = lut->lut_obd;
2101         int rc = 0;
2102         struct target_recovery_data *trd = &obd->obd_recovery_data;
2103
2104         memset(trd, 0, sizeof(*trd));
2105         init_completion(&trd->trd_starting);
2106         init_completion(&trd->trd_finishing);
2107         trd->trd_recovery_handler = handler;
2108
2109         if (!IS_ERR(kthread_run(target_recovery_thread,
2110                                 lut, "tgt_recov"))) {
2111                 wait_for_completion(&trd->trd_starting);
2112                 LASSERT(obd->obd_recovering != 0);
2113         } else {
2114                 rc = -ECHILD;
2115         }
2116
2117         return rc;
2118 }
2119
2120 void target_stop_recovery_thread(struct obd_device *obd)
2121 {
2122         if (obd->obd_recovery_data.trd_processing_task > 0) {
2123                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2124                 /** recovery can be done but postrecovery is not yet */
2125                 spin_lock(&obd->obd_dev_lock);
2126                 if (obd->obd_recovering) {
2127                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2128                         obd->obd_abort_recovery = 1;
2129                         wake_up(&obd->obd_next_transno_waitq);
2130                 }
2131                 spin_unlock(&obd->obd_dev_lock);
2132                 wait_for_completion(&trd->trd_finishing);
2133         }
2134 }
2135 EXPORT_SYMBOL(target_stop_recovery_thread);
2136
2137 void target_recovery_fini(struct obd_device *obd)
2138 {
2139         class_disconnect_exports(obd);
2140         target_stop_recovery_thread(obd);
2141         target_cleanup_recovery(obd);
2142 }
2143 EXPORT_SYMBOL(target_recovery_fini);
2144
2145 static void target_recovery_expired(unsigned long castmeharder)
2146 {
2147         struct obd_device *obd = (struct obd_device *)castmeharder;
2148         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2149                " after %lds (%d clients connected)\n",
2150                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2151                cfs_time_current_sec()- obd->obd_recovery_start,
2152                cfs_atomic_read(&obd->obd_connected_clients));
2153
2154         obd->obd_recovery_expired = 1;
2155         wake_up(&obd->obd_next_transno_waitq);
2156 }
2157
2158 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2159 {
2160         struct obd_device *obd = lut->lut_obd;
2161         if (obd->obd_max_recoverable_clients == 0) {
2162                 /** Update server last boot epoch */
2163                 tgt_boot_epoch_update(lut);
2164                 return;
2165         }
2166
2167         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2168                "last_transno "LPU64"\n", obd->obd_name,
2169                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2170         LASSERT(obd->obd_stopping == 0);
2171         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2172         obd->obd_recovery_start = 0;
2173         obd->obd_recovery_end = 0;
2174
2175         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2176         target_start_recovery_thread(lut, handler);
2177 }
2178 EXPORT_SYMBOL(target_recovery_init);
2179
2180 #endif /* __KERNEL__ */
2181
2182 static int target_process_req_flags(struct obd_device *obd,
2183                                     struct ptlrpc_request *req)
2184 {
2185         struct obd_export *exp = req->rq_export;
2186         LASSERT(exp != NULL);
2187         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2188                 /* client declares he's ready to replay locks */
2189                 spin_lock(&exp->exp_lock);
2190                 if (exp->exp_req_replay_needed) {
2191                         exp->exp_req_replay_needed = 0;
2192                         spin_unlock(&exp->exp_lock);
2193
2194                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2195                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2196                 } else {
2197                         spin_unlock(&exp->exp_lock);
2198                 }
2199         }
2200         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2201                 /* client declares he's ready to complete recovery
2202                  * so, we put the request on th final queue */
2203                 spin_lock(&exp->exp_lock);
2204                 if (exp->exp_lock_replay_needed) {
2205                         exp->exp_lock_replay_needed = 0;
2206                         spin_unlock(&exp->exp_lock);
2207
2208                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2209                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2210                 } else {
2211                         spin_unlock(&exp->exp_lock);
2212                 }
2213         }
2214         return 0;
2215 }
2216
2217 int target_queue_recovery_request(struct ptlrpc_request *req,
2218                                   struct obd_device *obd)
2219 {
2220         cfs_list_t *tmp;
2221         int inserted = 0;
2222         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2223         ENTRY;
2224
2225         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2226                 /* Processing the queue right now, don't re-add. */
2227                 RETURN(1);
2228         }
2229
2230         target_process_req_flags(obd, req);
2231
2232         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2233                 /* client declares he's ready to complete recovery
2234                  * so, we put the request on th final queue */
2235                 target_request_copy_get(req);
2236                 DEBUG_REQ(D_HA, req, "queue final req");
2237                 wake_up(&obd->obd_next_transno_waitq);
2238                 spin_lock(&obd->obd_recovery_task_lock);
2239                 if (obd->obd_recovering) {
2240                         cfs_list_add_tail(&req->rq_list,
2241                                           &obd->obd_final_req_queue);
2242                 } else {
2243                         spin_unlock(&obd->obd_recovery_task_lock);
2244                         target_request_copy_put(req);
2245                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2246                 }
2247                 spin_unlock(&obd->obd_recovery_task_lock);
2248                 RETURN(0);
2249         }
2250         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2251                 /* client declares he's ready to replay locks */
2252                 target_request_copy_get(req);
2253                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2254                 wake_up(&obd->obd_next_transno_waitq);
2255                 spin_lock(&obd->obd_recovery_task_lock);
2256                 LASSERT(obd->obd_recovering);
2257                 /* usually due to recovery abort */
2258                 if (!req->rq_export->exp_in_recovery) {
2259                         spin_unlock(&obd->obd_recovery_task_lock);
2260                         target_request_copy_put(req);
2261                         RETURN(-ENOTCONN);
2262                 }
2263                 LASSERT(req->rq_export->exp_lock_replay_needed);
2264                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2265                 spin_unlock(&obd->obd_recovery_task_lock);
2266                 RETURN(0);
2267         }
2268
2269         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2270          * (i.e. buflens etc are in my own byte order), but type-dependent
2271          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2272
2273         if (!transno) {
2274                 CFS_INIT_LIST_HEAD(&req->rq_list);
2275                 DEBUG_REQ(D_HA, req, "not queueing");
2276                 RETURN(1);
2277         }
2278
2279         /* If we're processing the queue, we want don't want to queue this
2280          * message.
2281          *
2282          * Also, if this request has a transno less than the one we're waiting
2283          * for, we should process it now.  It could (and currently always will)
2284          * be an open request for a descriptor that was opened some time ago.
2285          *
2286          * Also, a resent, replayed request that has already been
2287          * handled will pass through here and be processed immediately.
2288          */
2289         CDEBUG(D_HA, "Next recovery transno: "LPU64
2290                ", current: "LPU64", replaying\n",
2291                obd->obd_next_recovery_transno, transno);
2292         spin_lock(&obd->obd_recovery_task_lock);
2293         if (transno < obd->obd_next_recovery_transno) {
2294                 /* Processing the queue right now, don't re-add. */
2295                 LASSERT(cfs_list_empty(&req->rq_list));
2296                 spin_unlock(&obd->obd_recovery_task_lock);
2297                 RETURN(1);
2298         }
2299         spin_unlock(&obd->obd_recovery_task_lock);
2300
2301         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2302                 RETURN(0);
2303
2304         target_request_copy_get(req);
2305         if (!req->rq_export->exp_in_recovery) {
2306                 target_request_copy_put(req);
2307                 RETURN(-ENOTCONN);
2308         }
2309         LASSERT(req->rq_export->exp_req_replay_needed);
2310
2311         if (target_exp_enqueue_req_replay(req)) {
2312                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2313                 target_request_copy_put(req);
2314                 RETURN(0);
2315         }
2316
2317         /* XXX O(n^2) */
2318         spin_lock(&obd->obd_recovery_task_lock);
2319         LASSERT(obd->obd_recovering);
2320         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2321                 struct ptlrpc_request *reqiter =
2322                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2323
2324                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2325                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2326                         inserted = 1;
2327                         break;
2328                 }
2329
2330                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2331                              transno)) {
2332                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2333                                   "has been claimed by another client");
2334                         spin_unlock(&obd->obd_recovery_task_lock);
2335                         target_exp_dequeue_req_replay(req);
2336                         target_request_copy_put(req);
2337                         RETURN(0);
2338                 }
2339         }
2340
2341         if (!inserted)
2342                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2343
2344         obd->obd_requests_queued_for_recovery++;
2345         spin_unlock(&obd->obd_recovery_task_lock);
2346         wake_up(&obd->obd_next_transno_waitq);
2347         RETURN(0);
2348 }
2349 EXPORT_SYMBOL(target_queue_recovery_request);
2350
2351 int target_handle_ping(struct ptlrpc_request *req)
2352 {
2353         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2354         return req_capsule_server_pack(&req->rq_pill);
2355 }
2356 EXPORT_SYMBOL(target_handle_ping);
2357
2358 void target_committed_to_req(struct ptlrpc_request *req)
2359 {
2360         struct obd_export *exp = req->rq_export;
2361
2362         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2363                 lustre_msg_set_last_committed(req->rq_repmsg,
2364                                               exp->exp_last_committed);
2365         else
2366                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2367                           "%d)", exp->exp_obd->obd_no_transno,
2368                           req->rq_repmsg == NULL);
2369
2370         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2371                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2372 }
2373 EXPORT_SYMBOL(target_committed_to_req);
2374
2375 #endif /* HAVE_SERVER_SUPPORT */
2376
2377 /**
2378  * Packs current SLV and Limit into \a req.
2379  */
2380 int target_pack_pool_reply(struct ptlrpc_request *req)
2381 {
2382         struct obd_device *obd;
2383         ENTRY;
2384
2385         /* Check that we still have all structures alive as this may
2386          * be some late RPC at shutdown time. */
2387         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2388                      !exp_connect_lru_resize(req->rq_export))) {
2389                 lustre_msg_set_slv(req->rq_repmsg, 0);
2390                 lustre_msg_set_limit(req->rq_repmsg, 0);
2391                 RETURN(0);
2392         }
2393
2394         /* OBD is alive here as export is alive, which we checked above. */
2395         obd = req->rq_export->exp_obd;
2396
2397         read_lock(&obd->obd_pool_lock);
2398         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2399         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2400         read_unlock(&obd->obd_pool_lock);
2401
2402         RETURN(0);
2403 }
2404 EXPORT_SYMBOL(target_pack_pool_reply);
2405
2406 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2407 {
2408         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2409                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2410                 return (-ECOMM);
2411         }
2412
2413         if (unlikely(rc)) {
2414                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2415                 req->rq_status = rc;
2416                 return (ptlrpc_send_error(req, 1));
2417         } else {
2418                 DEBUG_REQ(D_NET, req, "sending reply");
2419         }
2420
2421         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2422 }
2423
2424 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2425 {
2426         struct ptlrpc_service_part *svcpt;
2427         int                        netrc;
2428         struct ptlrpc_reply_state *rs;
2429         struct obd_export         *exp;
2430         ENTRY;
2431
2432         if (req->rq_no_reply) {
2433                 EXIT;
2434                 return;
2435         }
2436
2437         svcpt = req->rq_rqbd->rqbd_svcpt;
2438         rs = req->rq_reply_state;
2439         if (rs == NULL || !rs->rs_difficult) {
2440                 /* no notifiers */
2441                 target_send_reply_msg (req, rc, fail_id);
2442                 EXIT;
2443                 return;
2444         }
2445
2446         /* must be an export if locks saved */
2447         LASSERT (req->rq_export != NULL);
2448         /* req/reply consistent */
2449         LASSERT(rs->rs_svcpt == svcpt);
2450
2451         /* "fresh" reply */
2452         LASSERT (!rs->rs_scheduled);
2453         LASSERT (!rs->rs_scheduled_ever);
2454         LASSERT (!rs->rs_handled);
2455         LASSERT (!rs->rs_on_net);
2456         LASSERT (rs->rs_export == NULL);
2457         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2458         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2459
2460         exp = class_export_get (req->rq_export);
2461
2462         /* disable reply scheduling while I'm setting up */
2463         rs->rs_scheduled = 1;
2464         rs->rs_on_net    = 1;
2465         rs->rs_xid       = req->rq_xid;
2466         rs->rs_transno   = req->rq_transno;
2467         rs->rs_export    = exp;
2468         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2469
2470         spin_lock(&exp->exp_uncommitted_replies_lock);
2471         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2472                rs->rs_transno, exp->exp_last_committed);
2473         if (rs->rs_transno > exp->exp_last_committed) {
2474                 /* not committed already */
2475                 cfs_list_add_tail(&rs->rs_obd_list,
2476                                   &exp->exp_uncommitted_replies);
2477         }
2478         spin_unlock(&exp->exp_uncommitted_replies_lock);
2479
2480         spin_lock(&exp->exp_lock);
2481         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2482         spin_unlock(&exp->exp_lock);
2483
2484         netrc = target_send_reply_msg(req, rc, fail_id);
2485
2486         spin_lock(&svcpt->scp_rep_lock);
2487
2488         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2489
2490         if (netrc != 0) {
2491                 /* error sending: reply is off the net.  Also we need +1
2492                  * reply ref until ptlrpc_handle_rs() is done
2493                  * with the reply state (if the send was successful, there
2494                  * would have been +1 ref for the net, which
2495                  * reply_out_callback leaves alone) */
2496                 rs->rs_on_net = 0;
2497                 ptlrpc_rs_addref(rs);
2498         }
2499
2500         spin_lock(&rs->rs_lock);
2501         if (rs->rs_transno <= exp->exp_last_committed ||
2502             (!rs->rs_on_net && !rs->rs_no_ack) ||
2503             cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2504             cfs_list_empty(&rs->rs_obd_list)) {
2505                 CDEBUG(D_HA, "Schedule reply immediately\n");
2506                 ptlrpc_dispatch_difficult_reply(rs);
2507         } else {
2508                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2509                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2510         }
2511         spin_unlock(&rs->rs_lock);
2512         spin_unlock(&svcpt->scp_rep_lock);
2513         EXIT;
2514 }
2515 EXPORT_SYMBOL(target_send_reply);
2516
2517 ldlm_mode_t lck_compat_array[] = {
2518         [LCK_EX] LCK_COMPAT_EX,
2519         [LCK_PW] LCK_COMPAT_PW,
2520         [LCK_PR] LCK_COMPAT_PR,
2521         [LCK_CW] LCK_COMPAT_CW,
2522         [LCK_CR] LCK_COMPAT_CR,
2523         [LCK_NL] LCK_COMPAT_NL,
2524         [LCK_GROUP] LCK_COMPAT_GROUP,
2525         [LCK_COS] LCK_COMPAT_COS,
2526 };
2527
2528 /**
2529  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2530  * not escape to the user level.
2531  */
2532 int ldlm_error2errno(ldlm_error_t error)
2533 {
2534         int result;
2535
2536         switch (error) {
2537         case ELDLM_OK:
2538                 result = 0;
2539                 break;
2540         case ELDLM_LOCK_CHANGED:
2541                 result = -ESTALE;
2542                 break;
2543         case ELDLM_LOCK_ABORTED:
2544                 result = -ENAVAIL;
2545                 break;
2546         case ELDLM_LOCK_REPLACED:
2547                 result = -ESRCH;
2548                 break;
2549         case ELDLM_NO_LOCK_DATA:
2550                 result = -ENOENT;
2551                 break;
2552         case ELDLM_NAMESPACE_EXISTS:
2553                 result = -EEXIST;
2554                 break;
2555         case ELDLM_BAD_NAMESPACE:
2556                 result = -EBADF;
2557                 break;
2558         default:
2559                 if (((int)error) < 0)  /* cast to signed type */
2560                         result = error; /* as ldlm_error_t can be unsigned */
2561                 else {
2562                         CERROR("Invalid DLM result code: %d\n", error);
2563                         result = -EPROTO;
2564                 }
2565         }
2566         return result;
2567 }
2568 EXPORT_SYMBOL(ldlm_error2errno);
2569
2570 /**
2571  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2572  */
2573 ldlm_error_t ldlm_errno2error(int err_no)
2574 {
2575         int error;
2576
2577         switch (err_no) {
2578         case 0:
2579                 error = ELDLM_OK;
2580                 break;
2581         case -ESTALE:
2582                 error = ELDLM_LOCK_CHANGED;
2583                 break;
2584         case -ENAVAIL:
2585                 error = ELDLM_LOCK_ABORTED;
2586                 break;
2587         case -ESRCH:
2588                 error = ELDLM_LOCK_REPLACED;
2589                 break;
2590         case -ENOENT:
2591                 error = ELDLM_NO_LOCK_DATA;
2592                 break;
2593         case -EEXIST:
2594                 error = ELDLM_NAMESPACE_EXISTS;
2595                 break;
2596         case -EBADF:
2597                 error = ELDLM_BAD_NAMESPACE;
2598                 break;
2599         default:
2600                 error = err_no;
2601         }
2602         return error;
2603 }
2604 EXPORT_SYMBOL(ldlm_errno2error);
2605
2606 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2607 void ldlm_dump_export_locks(struct obd_export *exp)
2608 {
2609         spin_lock(&exp->exp_locks_list_guard);
2610         if (!cfs_list_empty(&exp->exp_locks_list)) {
2611                 struct ldlm_lock *lock;
2612
2613                 CERROR("dumping locks for export %p,"
2614                        "ignore if the unmount doesn't hang\n", exp);
2615                 cfs_list_for_each_entry(lock, &exp->exp_locks_list,
2616                                         l_exp_refs_link)
2617                         LDLM_ERROR(lock, "lock:");
2618         }
2619         spin_unlock(&exp->exp_locks_list_guard);
2620 }
2621 #endif
2622
2623 #ifdef HAVE_SERVER_SUPPORT
2624 static int target_bulk_timeout(void *data)
2625 {
2626         ENTRY;
2627         /* We don't fail the connection here, because having the export
2628          * killed makes the (vital) call to commitrw very sad.
2629          */
2630         RETURN(1);
2631 }
2632
2633 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2634 {
2635         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2636 }
2637
2638 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2639                    struct l_wait_info *lwi)
2640 {
2641         struct ptlrpc_request   *req = desc->bd_req;
2642         time_t                   start = cfs_time_current_sec();
2643         int                      rc = 0;
2644
2645         ENTRY;
2646
2647         /* If there is eviction in progress, wait for it to finish. */
2648         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2649                 *lwi = LWI_INTR(NULL, NULL);
2650                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2651                                   !cfs_atomic_read(&exp->exp_obd->
2652                                                    obd_evict_inprogress),
2653                                   lwi);
2654         }
2655
2656         /* Check if client was evicted or reconnected already. */
2657         if (exp->exp_failed ||
2658             exp->exp_conn_cnt > lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2659                 rc = -ENOTCONN;
2660         } else {
2661                 if (desc->bd_type == BULK_PUT_SINK)
2662                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2663                 if (rc == 0)
2664                         rc = ptlrpc_start_bulk_transfer(desc);
2665         }
2666
2667         if (rc < 0) {
2668                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2669                           bulk2type(desc), rc);
2670                 RETURN(rc);
2671         }
2672
2673         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2674                 ptlrpc_abort_bulk(desc);
2675                 RETURN(0);
2676         }
2677
2678         do {
2679                 long timeoutl = req->rq_deadline - cfs_time_current_sec();
2680                 cfs_duration_t timeout = timeoutl <= 0 ?
2681                                          CFS_TICK : cfs_time_seconds(timeoutl);
2682
2683                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2684                                             target_bulk_timeout, desc);
2685                 rc = l_wait_event(desc->bd_waitq,
2686                                   !ptlrpc_server_bulk_active(desc) ||
2687                                   exp->exp_failed ||
2688                                   exp->exp_conn_cnt >
2689                                   lustre_msg_get_conn_cnt(req->rq_reqmsg),
2690                                   lwi);
2691                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2692                 /* Wait again if we changed deadline. */
2693         } while ((rc == -ETIMEDOUT) &&
2694                  (req->rq_deadline > cfs_time_current_sec()));
2695
2696         if (rc == -ETIMEDOUT) {
2697                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
2698                           bulk2type(desc), req->rq_deadline - start,
2699                           cfs_time_current_sec() - req->rq_deadline);
2700                 ptlrpc_abort_bulk(desc);
2701         } else if (exp->exp_failed) {
2702                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2703                           bulk2type(desc));
2704                 rc = -ENOTCONN;
2705                 ptlrpc_abort_bulk(desc);
2706         } else if (exp->exp_conn_cnt >
2707                    lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
2708                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2709                           bulk2type(desc));
2710                 /* We don't reply anyway. */
2711                 rc = -ETIMEDOUT;
2712                 ptlrpc_abort_bulk(desc);
2713         } else if (desc->bd_failure ||
2714                    desc->bd_nob_transferred != desc->bd_nob) {
2715                 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2716                           desc->bd_failure ? "network error on" : "truncated",
2717                           bulk2type(desc), desc->bd_nob_transferred,
2718                           desc->bd_nob);
2719                 /* XXX Should this be a different errno? */
2720                 rc = -ETIMEDOUT;
2721         } else if (desc->bd_type == BULK_GET_SINK) {
2722                 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2723         }
2724
2725         RETURN(rc);
2726 }
2727 EXPORT_SYMBOL(target_bulk_io);
2728
2729 #endif /* HAVE_SERVER_SUPPORT */