lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  */
  39
  40 #ifndef EXPORT_SYMTAB
  41 # define EXPORT_SYMTAB
  42 #endif
  43 #define DEBUG_SUBSYSTEM S_LDLM
  44
  45 #ifdef __KERNEL__
  46 # include <libcfs/libcfs.h>
  47 #else
  48 # include <liblustre.h>
  49 #endif
  50 #include <obd.h>
  51 #include <lustre_mds.h>
  52 #include <lustre_dlm.h>
  53 #include <lustre_net.h>
  54 #include <lustre_sec.h>
  55 #include "ldlm_internal.h"
  56
  57 /* @priority: if non-zero, move the selected to the list head
  58  * @create: if zero, only search in existed connections
  59  */
  60 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  61                            int priority, int create)
  62 {
  63         struct ptlrpc_connection *ptlrpc_conn;
  64         struct obd_import_conn *imp_conn = NULL, *item;
  65         int rc = 0;
  66         ENTRY;
  67
  68         if (!create && !priority) {
  69                 CDEBUG(D_HA, "Nothing to do\n");
  70                 RETURN(-EINVAL);
  71         }
  72
  73         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  74         if (!ptlrpc_conn) {
  75                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  76                 RETURN (-ENOENT);
  77         }
  78
  79         if (create) {
  80                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  81                 if (!imp_conn) {
  82                         GOTO(out_put, rc = -ENOMEM);
  83                 }
  84         }
  85
  86         cfs_spin_lock(&imp->imp_lock);
  87         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  88                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  89                         if (priority) {
  90                                 cfs_list_del(&item->oic_item);
  91                                 cfs_list_add(&item->oic_item,
  92                                              &imp->imp_conn_list);
  93                                 item->oic_last_attempt = 0;
  94                         }
  95                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  96                                imp, imp->imp_obd->obd_name, uuid->uuid,
  97                                (priority ? ", moved to head" : ""));
  98                         cfs_spin_unlock(&imp->imp_lock);
  99                         GOTO(out_free, rc = 0);
 100                 }
 101         }
 102         /* not found */
 103         if (create) {
 104                 imp_conn->oic_conn = ptlrpc_conn;
 105                 imp_conn->oic_uuid = *uuid;
 106                 imp_conn->oic_last_attempt = 0;
 107                 if (priority)
 108                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 109                 else
 110                         cfs_list_add_tail(&imp_conn->oic_item,
 111                                           &imp->imp_conn_list);
 112                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 113                        imp, imp->imp_obd->obd_name, uuid->uuid,
 114                        (priority ? "head" : "tail"));
 115         } else {
 116                 cfs_spin_unlock(&imp->imp_lock);
 117                 GOTO(out_free, rc = -ENOENT);
 118         }
 119
 120         cfs_spin_unlock(&imp->imp_lock);
 121         RETURN(0);
 122 out_free:
 123         if (imp_conn)
 124                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 125 out_put:
 126         ptlrpc_connection_put(ptlrpc_conn);
 127         RETURN(rc);
 128 }
 129
 130 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 131 {
 132         return import_set_conn(imp, uuid, 1, 0);
 133 }
 134
 135 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 136                            int priority)
 137 {
 138         return import_set_conn(imp, uuid, priority, 1);
 139 }
 140
 141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 142 {
 143         struct obd_import_conn *imp_conn;
 144         struct obd_export *dlmexp;
 145         int rc = -ENOENT;
 146         ENTRY;
 147
 148         cfs_spin_lock(&imp->imp_lock);
 149         if (cfs_list_empty(&imp->imp_conn_list)) {
 150                 LASSERT(!imp->imp_connection);
 151                 GOTO(out, rc);
 152         }
 153
 154         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 155                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 156                         continue;
 157                 LASSERT(imp_conn->oic_conn);
 158
 159                 /* is current conn? */
 160                 if (imp_conn == imp->imp_conn_current) {
 161                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 162
 163                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 164                             imp->imp_state != LUSTRE_IMP_DISCON) {
 165                                 CERROR("can't remove current connection\n");
 166                                 GOTO(out, rc = -EBUSY);
 167                         }
 168
 169                         ptlrpc_connection_put(imp->imp_connection);
 170                         imp->imp_connection = NULL;
 171
 172                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 173                         if (dlmexp && dlmexp->exp_connection) {
 174                                 LASSERT(dlmexp->exp_connection ==
 175                                         imp_conn->oic_conn);
 176                                 ptlrpc_connection_put(dlmexp->exp_connection);
 177                                 dlmexp->exp_connection = NULL;
 178                         }
 179                 }
 180
 181                 cfs_list_del(&imp_conn->oic_item);
 182                 ptlrpc_connection_put(imp_conn->oic_conn);
 183                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 184                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 185                        imp, imp->imp_obd->obd_name, uuid->uuid);
 186                 rc = 0;
 187                 break;
 188         }
 189 out:
 190         cfs_spin_unlock(&imp->imp_lock);
 191         if (rc == -ENOENT)
 192                 CERROR("connection %s not found\n", uuid->uuid);
 193         RETURN(rc);
 194 }
 195
 196 /**
 197  * Find conn uuid by peer nid. @peer is a server nid. This function is used
 198  * to find a conn uuid of @imp which can reach @peer.
 199  */
 200 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 201                             struct obd_uuid *uuid)
 202 {
 203         struct obd_import_conn *conn;
 204         int rc = -ENOENT;
 205         ENTRY;
 206
 207         cfs_spin_lock(&imp->imp_lock);
 208         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 209                 /* check if conn uuid does have this peer nid */
 210                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 211                         *uuid = conn->oic_uuid;
 212                         rc = 0;
 213                         break;
 214                 }
 215         }
 216         cfs_spin_unlock(&imp->imp_lock);
 217         RETURN(rc);
 218 }
 219 EXPORT_SYMBOL(client_import_find_conn);
 220
 221 void client_destroy_import(struct obd_import *imp)
 222 {
 223         /* drop security policy instance after all rpc finished/aborted
 224          * to let all busy contexts be released. */
 225         class_import_get(imp);
 226         class_destroy_import(imp);
 227         sptlrpc_import_sec_put(imp);
 228         class_import_put(imp);
 229 }
 230 EXPORT_SYMBOL(client_destroy_import);
 231
 232 /* configure an RPC client OBD device
 233  *
 234  * lcfg parameters:
 235  * 1 - client UUID
 236  * 2 - server UUID
 237  * 3 - inactive-on-startup
 238  */
 239 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 240 {
 241         struct client_obd *cli = &obddev->u.cli;
 242         struct obd_import *imp;
 243         struct obd_uuid server_uuid;
 244         int rq_portal, rp_portal, connect_op;
 245         char *name = obddev->obd_type->typ_name;
 246         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 247         int rc;
 248         ENTRY;
 249
 250         /* In a more perfect world, we would hang a ptlrpc_client off of
 251          * obd_type and just use the values from there. */
 252         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 253                 rq_portal = OST_REQUEST_PORTAL;
 254                 rp_portal = OSC_REPLY_PORTAL;
 255                 connect_op = OST_CONNECT;
 256                 cli->cl_sp_me = LUSTRE_SP_CLI;
 257                 cli->cl_sp_to = LUSTRE_SP_OST;
 258                 ns_type = LDLM_NS_TYPE_OSC;
 259
 260         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 261                 rq_portal = MDS_REQUEST_PORTAL;
 262                 rp_portal = MDC_REPLY_PORTAL;
 263                 connect_op = MDS_CONNECT;
 264                 cli->cl_sp_me = LUSTRE_SP_CLI;
 265                 cli->cl_sp_to = LUSTRE_SP_MDT;
 266                 ns_type = LDLM_NS_TYPE_MDC;
 267
 268         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 269                 rq_portal = MGS_REQUEST_PORTAL;
 270                 rp_portal = MGC_REPLY_PORTAL;
 271                 connect_op = MGS_CONNECT;
 272                 cli->cl_sp_me = LUSTRE_SP_MGC;
 273                 cli->cl_sp_to = LUSTRE_SP_MGS;
 274                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 275                 ns_type = LDLM_NS_TYPE_MGC;
 276
 277         } else {
 278                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 279                        name);
 280                 RETURN(-EINVAL);
 281         }
 282
 283         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 284                 CERROR("requires a TARGET UUID\n");
 285                 RETURN(-EINVAL);
 286         }
 287
 288         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 289                 CERROR("client UUID must be less than 38 characters\n");
 290                 RETURN(-EINVAL);
 291         }
 292
 293         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 294                 CERROR("setup requires a SERVER UUID\n");
 295                 RETURN(-EINVAL);
 296         }
 297
 298         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 299                 CERROR("target UUID must be less than 38 characters\n");
 300                 RETURN(-EINVAL);
 301         }
 302
 303         cfs_init_rwsem(&cli->cl_sem);
 304         cfs_sema_init(&cli->cl_mgc_sem, 1);
 305         cli->cl_conn_count = 0;
 306         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 307                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 308                      sizeof(server_uuid)));
 309
 310         cli->cl_dirty = 0;
 311         cli->cl_avail_grant = 0;
 312         /* FIXME: should limit this for the sum of all cl_dirty_max */
 313         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 314         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
 315                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
 316         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 317         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 318         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 319         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 320         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 321         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 322         cli->cl_r_in_flight = 0;
 323         cli->cl_w_in_flight = 0;
 324
 325         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 326         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 327         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 328         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 329         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 330         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 331         cfs_waitq_init(&cli->cl_destroy_waitq);
 332         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 333 #ifdef ENABLE_CHECKSUM
 334         /* Turn on checksumming by default. */
 335         cli->cl_checksum = 1;
 336         /*
 337          * The supported checksum types will be worked out at connect time
 338          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 339          * through procfs.
 340          */
 341         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 342 #endif
 343         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 344
 345         /* This value may be changed at connect time in
 346            ptlrpc_connect_interpret. */
 347         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 348                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 349
 350         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 351                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 352         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 353                 cli->cl_max_rpcs_in_flight = 2;
 354         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 355                 cli->cl_max_rpcs_in_flight = 3;
 356         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 357                 cli->cl_max_rpcs_in_flight = 4;
 358         } else {
 359                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 360         }
 361
 362         rc = ldlm_get_ref();
 363         if (rc) {
 364                 CERROR("ldlm_get_ref failed: %d\n", rc);
 365                 GOTO(err, rc);
 366         }
 367
 368         ptlrpc_init_client(rq_portal, rp_portal, name,
 369                            &obddev->obd_ldlm_client);
 370
 371         imp = class_new_import(obddev);
 372         if (imp == NULL)
 373                 GOTO(err_ldlm, rc = -ENOENT);
 374         imp->imp_client = &obddev->obd_ldlm_client;
 375         imp->imp_connect_op = connect_op;
 376         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 377         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 378                LUSTRE_CFG_BUFLEN(lcfg, 1));
 379         class_import_put(imp);
 380
 381         rc = client_import_add_conn(imp, &server_uuid, 1);
 382         if (rc) {
 383                 CERROR("can't add initial connection\n");
 384                 GOTO(err_import, rc);
 385         }
 386
 387         cli->cl_import = imp;
 388         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 389         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 390         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 391
 392         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 393                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 394                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 395                                name, obddev->obd_name,
 396                                cli->cl_target_uuid.uuid);
 397                         cfs_spin_lock(&imp->imp_lock);
 398                         imp->imp_deactive = 1;
 399                         cfs_spin_unlock(&imp->imp_lock);
 400                 }
 401         }
 402
 403         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 404                                                    LDLM_NAMESPACE_CLIENT,
 405                                                    LDLM_NAMESPACE_GREEDY,
 406                                                    ns_type);
 407         if (obddev->obd_namespace == NULL) {
 408                 CERROR("Unable to create client namespace - %s\n",
 409                        obddev->obd_name);
 410                 GOTO(err_import, rc = -ENOMEM);
 411         }
 412
 413         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 414
 415         RETURN(rc);
 416
 417 err_import:
 418         class_destroy_import(imp);
 419 err_ldlm:
 420         ldlm_put_ref();
 421 err:
 422         RETURN(rc);
 423
 424 }
 425
 426 int client_obd_cleanup(struct obd_device *obddev)
 427 {
 428         ENTRY;
 429
 430         ldlm_namespace_free_post(obddev->obd_namespace);
 431         obddev->obd_namespace = NULL;
 432
 433         LASSERT(obddev->u.cli.cl_import == NULL);
 434
 435         ldlm_put_ref();
 436         RETURN(0);
 437 }
 438
 439 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 440 int client_connect_import(const struct lu_env *env,
 441                           struct obd_export **exp,
 442                           struct obd_device *obd, struct obd_uuid *cluuid,
 443                           struct obd_connect_data *data, void *localdata)
 444 {
 445         struct client_obd *cli = &obd->u.cli;
 446         struct obd_import *imp = cli->cl_import;
 447         struct obd_connect_data *ocd;
 448         struct lustre_handle conn = { 0 };
 449         int rc;
 450         ENTRY;
 451
 452         *exp = NULL;
 453         cfs_down_write(&cli->cl_sem);
 454         if (cli->cl_conn_count > 0 )
 455                 GOTO(out_sem, rc = -EALREADY);
 456
 457         rc = class_connect(&conn, obd, cluuid);
 458         if (rc)
 459                 GOTO(out_sem, rc);
 460
 461         cli->cl_conn_count++;
 462         *exp = class_conn2export(&conn);
 463
 464         LASSERT(obd->obd_namespace);
 465
 466         imp->imp_dlm_handle = conn;
 467         rc = ptlrpc_init_import(imp);
 468         if (rc != 0)
 469                 GOTO(out_ldlm, rc);
 470
 471         ocd = &imp->imp_connect_data;
 472         if (data) {
 473                 *ocd = *data;
 474                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 475         }
 476
 477         rc = ptlrpc_connect_import(imp, NULL);
 478         if (rc != 0) {
 479                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 480                 GOTO(out_ldlm, rc);
 481         }
 482         LASSERT((*exp)->exp_connection);
 483
 484         if (data) {
 485                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 486                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 487                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 488                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 489         }
 490
 491         ptlrpc_pinger_add_import(imp);
 492
 493         EXIT;
 494
 495         if (rc) {
 496 out_ldlm:
 497                 cli->cl_conn_count--;
 498                 class_disconnect(*exp);
 499                 *exp = NULL;
 500         }
 501 out_sem:
 502         cfs_up_write(&cli->cl_sem);
 503
 504         return rc;
 505 }
 506
 507 int client_disconnect_export(struct obd_export *exp)
 508 {
 509         struct obd_device *obd = class_exp2obd(exp);
 510         struct client_obd *cli;
 511         struct obd_import *imp;
 512         int rc = 0, err;
 513         ENTRY;
 514
 515         if (!obd) {
 516                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 517                        exp, exp ? exp->exp_handle.h_cookie : -1);
 518                 RETURN(-EINVAL);
 519         }
 520
 521         cli = &obd->u.cli;
 522         imp = cli->cl_import;
 523
 524         cfs_down_write(&cli->cl_sem);
 525         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 526                cli->cl_conn_count);
 527
 528         if (!cli->cl_conn_count) {
 529                 CERROR("disconnecting disconnected device (%s)\n",
 530                        obd->obd_name);
 531                 GOTO(out_disconnect, rc = -EINVAL);
 532         }
 533
 534         cli->cl_conn_count--;
 535         if (cli->cl_conn_count)
 536                 GOTO(out_disconnect, rc = 0);
 537
 538         /* Mark import deactivated now, so we don't try to reconnect if any
 539          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 540          * fully deactivate the import, or that would drop all requests. */
 541         cfs_spin_lock(&imp->imp_lock);
 542         imp->imp_deactive = 1;
 543         cfs_spin_unlock(&imp->imp_lock);
 544
 545         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 546          * delete it regardless.  (It's safe to delete an import that was
 547          * never added.) */
 548         (void)ptlrpc_pinger_del_import(imp);
 549
 550         if (obd->obd_namespace != NULL) {
 551                 /* obd_force == local only */
 552                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 553                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 554                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 555         }
 556
 557         /*
 558          * there's no need to hold sem during disconnecting an import,
 559          * and actually it may cause deadlock in gss.
 560          */
 561         cfs_up_write(&cli->cl_sem);
 562         rc = ptlrpc_disconnect_import(imp, 0);
 563         cfs_down_write(&cli->cl_sem);
 564
 565         ptlrpc_invalidate_import(imp);
 566
 567         EXIT;
 568
 569  out_disconnect:
 570         /* use server style - class_disconnect should be always called for
 571          * o_disconnect */
 572         err = class_disconnect(exp);
 573         if (!rc && err)
 574                 rc = err;
 575
 576         cfs_up_write(&cli->cl_sem);
 577
 578         RETURN(rc);
 579 }
 580
 581 int server_disconnect_export(struct obd_export *exp)
 582 {
 583         int rc;
 584         ENTRY;
 585
 586         /* Disconnect early so that clients can't keep using export */
 587         rc = class_disconnect(exp);
 588         /* close import for avoid sending any requests */
 589         if (exp->exp_imp_reverse)
 590                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 591
 592         if (exp->exp_obd->obd_namespace != NULL)
 593                 ldlm_cancel_locks_for_export(exp);
 594
 595         /* complete all outstanding replies */
 596         cfs_spin_lock(&exp->exp_lock);
 597         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 598                 struct ptlrpc_reply_state *rs =
 599                         cfs_list_entry(exp->exp_outstanding_replies.next,
 600                                        struct ptlrpc_reply_state, rs_exp_list);
 601                 struct ptlrpc_service *svc = rs->rs_service;
 602
 603                 cfs_spin_lock(&svc->srv_rs_lock);
 604                 cfs_list_del_init(&rs->rs_exp_list);
 605                 cfs_spin_lock(&rs->rs_lock);
 606                 ptlrpc_schedule_difficult_reply(rs);
 607                 cfs_spin_unlock(&rs->rs_lock);
 608                 cfs_spin_unlock(&svc->srv_rs_lock);
 609         }
 610         cfs_spin_unlock(&exp->exp_lock);
 611
 612         RETURN(rc);
 613 }
 614
 615 /* --------------------------------------------------------------------------
 616  * from old lib/target.c
 617  * -------------------------------------------------------------------------- */
 618
 619 static int target_handle_reconnect(struct lustre_handle *conn,
 620                                    struct obd_export *exp,
 621                                    struct obd_uuid *cluuid)
 622 {
 623         ENTRY;
 624
 625         if (exp->exp_connection && exp->exp_imp_reverse) {
 626                 struct lustre_handle *hdl;
 627                 struct obd_device *target;
 628
 629                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 630                 target = exp->exp_obd;
 631
 632                 /* Might be a re-connect after a partition. */
 633                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 634                         if (target->obd_recovering) {
 635                                 int timeout = cfs_duration_sec(cfs_time_sub(
 636                                         cfs_timer_deadline(
 637                                         &target->obd_recovery_timer),
 638                                         cfs_time_current()));
 639
 640                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 641                                         "ing, waiting for %d clients in recov"
 642                                         "ery for %d:%.02d\n", target->obd_name,
 643                                         obd_uuid2str(&exp->exp_client_uuid),
 644                                         obd_export_nid2str(exp),
 645                                         target->obd_max_recoverable_clients,
 646                                         timeout / 60, timeout % 60);
 647                         } else {
 648                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 649                                         "reconnecting\n", target->obd_name,
 650                                         obd_uuid2str(&exp->exp_client_uuid),
 651                                         obd_export_nid2str(exp));
 652                         }
 653
 654                         conn->cookie = exp->exp_handle.h_cookie;
 655                         /* target_handle_connect() treats EALREADY and
 656                          * -EALREADY differently.  EALREADY means we are
 657                          * doing a valid reconnect from the same client. */
 658                         RETURN(EALREADY);
 659                 } else {
 660                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 661                                       "with handle "LPX64". Rejecting client "
 662                                       "with the same UUID trying to reconnect "
 663                                       "with handle "LPX64"\n", target->obd_name,
 664                                       obd_uuid2str(&exp->exp_client_uuid),
 665                                       obd_export_nid2str(exp),
 666                                       hdl->cookie, conn->cookie);
 667                         memset(conn, 0, sizeof *conn);
 668                         /* target_handle_connect() treats EALREADY and
 669                          * -EALREADY differently.  -EALREADY is an error
 670                          * (same UUID, different handle). */
 671                         RETURN(-EALREADY);
 672                 }
 673         }
 674
 675         conn->cookie = exp->exp_handle.h_cookie;
 676         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 677                cluuid->uuid, exp, conn->cookie);
 678         RETURN(0);
 679 }
 680
 681 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 682                           int error)
 683 {
 684         struct obd_export *exp = cb_data;
 685
 686         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 687                obd->obd_name, exp->exp_client_uuid.uuid);
 688
 689         cfs_spin_lock(&exp->exp_lock);
 690         exp->exp_need_sync = 0;
 691         cfs_spin_unlock(&exp->exp_lock);
 692         class_export_cb_put(exp);
 693 }
 694 EXPORT_SYMBOL(target_client_add_cb);
 695
 696 static void
 697 target_start_and_reset_recovery_timer(struct obd_device *obd,
 698                                       struct ptlrpc_request *req,
 699                                       int new_client);
 700
 701 int target_handle_connect(struct ptlrpc_request *req)
 702 {
 703         struct obd_device *target = NULL, *targref = NULL;
 704         struct obd_export *export = NULL;
 705         struct obd_import *revimp;
 706         struct lustre_handle conn;
 707         struct lustre_handle *tmp;
 708         struct obd_uuid tgtuuid;
 709         struct obd_uuid cluuid;
 710         struct obd_uuid remote_uuid;
 711         char *str;
 712         int rc = 0;
 713         char *target_start;
 714         int target_len;
 715         int mds_conn = 0;
 716         struct obd_connect_data *data, *tmpdata;
 717         int size, tmpsize;
 718         lnet_nid_t *client_nid = NULL;
 719         ENTRY;
 720
 721         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 722
 723         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 724         if (str == NULL) {
 725                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 726                 GOTO(out, rc = -EINVAL);
 727         }
 728
 729         obd_str2uuid(&tgtuuid, str);
 730         target = class_uuid2obd(&tgtuuid);
 731         if (!target)
 732                 target = class_name2obd(str);
 733
 734         if (!target) {
 735                 deuuidify(str, NULL, &target_start, &target_len);
 736                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
 737                                    "connect (no target)\n", str);
 738                 GOTO(out, rc = -ENODEV);
 739         }
 740
 741         cfs_spin_lock(&target->obd_dev_lock);
 742         if (target->obd_stopping || !target->obd_set_up) {
 743                 cfs_spin_unlock(&target->obd_dev_lock);
 744
 745                 deuuidify(str, NULL, &target_start, &target_len);
 746                 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
 747                                    "from %s (%s)\n", target_len, target_start,
 748                                    libcfs_nid2str(req->rq_peer.nid),
 749                                    (target->obd_stopping ?
 750                                    "stopping" : "not set up"));
 751                 GOTO(out, rc = -ENODEV);
 752         }
 753
 754         if (target->obd_no_conn) {
 755                 cfs_spin_unlock(&target->obd_dev_lock);
 756
 757                 LCONSOLE_WARN("%s: Temporarily refusing client connection "
 758                               "from %s\n", target->obd_name,
 759                               libcfs_nid2str(req->rq_peer.nid));
 760                 GOTO(out, rc = -EAGAIN);
 761         }
 762
 763         /* Make sure the target isn't cleaned up while we're here. Yes,
 764            there's still a race between the above check and our incref here.
 765            Really, class_uuid2obd should take the ref. */
 766         targref = class_incref(target, __FUNCTION__, cfs_current());
 767
 768         target->obd_conn_inprogress++;
 769         cfs_spin_unlock(&target->obd_dev_lock);
 770
 771         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 772         if (str == NULL) {
 773                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 774                 GOTO(out, rc = -EINVAL);
 775         }
 776
 777         obd_str2uuid(&cluuid, str);
 778
 779         /* XXX extract a nettype and format accordingly */
 780         switch (sizeof(lnet_nid_t)) {
 781                 /* NB the casts only avoid compiler warnings */
 782         case 8:
 783                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 784                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 785                 break;
 786         case 4:
 787                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 788                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 789                 break;
 790         default:
 791                 LBUG();
 792         }
 793
 794         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 795         if (tmp == NULL)
 796                 GOTO(out, rc = -EPROTO);
 797
 798         conn = *tmp;
 799
 800         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 801                                     RCL_CLIENT);
 802         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 803         if (!data)
 804                 GOTO(out, rc = -EPROTO);
 805
 806         rc = req_capsule_server_pack(&req->rq_pill);
 807         if (rc)
 808                 GOTO(out, rc);
 809
 810         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 811                 if (!data) {
 812                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 813                                   "libclient connection attempt");
 814                         GOTO(out, rc = -EPROTO);
 815                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 816                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 817                            data->ocd_version > LUSTRE_VERSION_CODE +
 818                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 819                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 820                                   "libclient connection attempt",
 821                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 822                                   "old" : "new",
 823                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 824                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 825                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 826                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 827                         data = req_capsule_server_sized_get(&req->rq_pill,
 828                                                         &RMF_CONNECT_DATA,
 829                                     offsetof(typeof(*data), ocd_version) +
 830                                              sizeof(data->ocd_version));
 831                         if (data) {
 832                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 833                                 data->ocd_version = LUSTRE_VERSION_CODE;
 834                         }
 835                         GOTO(out, rc = -EPROTO);
 836                 }
 837         }
 838
 839         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 840             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 841                 mds_conn = 1;
 842
 843         /* lctl gets a backstage, all-access pass. */
 844         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 845                 goto dont_check_exports;
 846
 847         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 848         if (!export)
 849                 goto no_export;
 850
 851         /* we've found an export in the hash */
 852         if (export->exp_connecting) { /* bug 9635, et. al. */
 853                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 854                               export->exp_obd->obd_name, export,
 855                               libcfs_nid2str(req->rq_peer.nid));
 856                 class_export_put(export);
 857                 export = NULL;
 858                 rc = -EALREADY;
 859         } else if (mds_conn && export->exp_connection) {
 860                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 861                         /* mds reconnected after failover */
 862                         LCONSOLE_WARN("%s: Received MDS connection from "
 863                             "%s, removing former export from %s\n",
 864                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 865                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 866                 else
 867                         /* new mds connection from the same nid */
 868                         LCONSOLE_WARN("%s: Received new MDS connection from "
 869                             "%s, removing former export from same NID\n",
 870                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 871                 class_fail_export(export);
 872                 class_export_put(export);
 873                 export = NULL;
 874                 rc = 0;
 875         } else if (export->exp_connection != NULL &&
 876                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 877                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 878                     MSG_CONNECT_INITIAL)) {
 879                 /* in mds failover we have static uuid but nid can be
 880                  * changed*/
 881                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 882                               "existing nid %s is already connected\n",
 883                               target->obd_name, cluuid.uuid,
 884                               libcfs_nid2str(req->rq_peer.nid),
 885                               libcfs_nid2str(
 886                                       export->exp_connection->c_peer.nid));
 887                 rc = -EALREADY;
 888                 class_export_put(export);
 889                 export = NULL;
 890         } else {
 891                 cfs_spin_lock(&export->exp_lock);
 892                 export->exp_connecting = 1;
 893                 cfs_spin_unlock(&export->exp_lock);
 894                 LASSERT(export->exp_obd == target);
 895
 896                 rc = target_handle_reconnect(&conn, export, &cluuid);
 897         }
 898
 899         /* If we found an export, we already unlocked. */
 900         if (!export) {
 901 no_export:
 902                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 903         } else if (req->rq_export == NULL &&
 904                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 905                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 906                               "still busy with %d references\n",
 907                               target->obd_name, cluuid.uuid,
 908                               libcfs_nid2str(req->rq_peer.nid),
 909                               cfs_atomic_read(&export->exp_refcount));
 910                 GOTO(out, rc = -EBUSY);
 911         } else if (req->rq_export != NULL &&
 912                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 913                 /* the current connect rpc has increased exp_rpc_count */
 914                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
 915                               "still busy with %d active RPCs\n",
 916                               target->obd_name, cluuid.uuid,
 917                               libcfs_nid2str(req->rq_peer.nid),
 918                               cfs_atomic_read(&export->exp_rpc_count) - 1);
 919                 cfs_spin_lock(&export->exp_lock);
 920                 if (req->rq_export->exp_conn_cnt <
 921                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 922                         /* try to abort active requests */
 923                         req->rq_export->exp_abort_active_req = 1;
 924                 cfs_spin_unlock(&export->exp_lock);
 925                 GOTO(out, rc = -EBUSY);
 926         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 927                 if (!strstr(cluuid.uuid, "mdt"))
 928                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 929                                       "known client %s (at %s) because it "
 930                                       "is indicating it is a new client",
 931                                       target->obd_name, cluuid.uuid,
 932                                       libcfs_nid2str(req->rq_peer.nid));
 933                 GOTO(out, rc = -EALREADY);
 934         } else {
 935                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 936         }
 937
 938         if (rc < 0) {
 939                 GOTO(out, rc);
 940         }
 941
 942         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 943                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 944               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 945               export, (long)cfs_time_current_sec(),
 946               export ? (long)export->exp_last_request_time : 0);
 947
 948         /* If this is the first time a client connects,
 949          * reset the recovery timer */
 950         if (rc == 0 && target->obd_recovering)
 951                 target_start_and_reset_recovery_timer(target, req, !export);
 952
 953         /* We want to handle EALREADY but *not* -EALREADY from
 954          * target_handle_reconnect(), return reconnection state in a flag */
 955         if (rc == EALREADY) {
 956                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 957                 rc = 0;
 958         } else {
 959                 LASSERT(rc == 0);
 960         }
 961
 962         /* Tell the client if we support replayable requests */
 963         if (target->obd_replayable)
 964                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 965         client_nid = &req->rq_peer.nid;
 966
 967         if (export == NULL) {
 968                 if (target->obd_recovering) {
 969                         cfs_time_t t;
 970
 971                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 972                         t = cfs_time_sub(t, cfs_time_current());
 973                         t = cfs_duration_sec(t);
 974                         LCONSOLE_WARN("%s: Denying connection for new client "
 975                                       "%s (at %s), waiting for %d clients in "
 976                                       "recovery for %d:%.02d\n",
 977                                       target->obd_name,
 978                                       libcfs_nid2str(req->rq_peer.nid),
 979                                       cluuid.uuid,
 980                                       cfs_atomic_read(&target-> \
 981                                                       obd_lock_replay_clients),
 982                                       (int)t / 60, (int)t % 60);
 983                         rc = -EBUSY;
 984                 } else {
 985 dont_check_exports:
 986                         rc = obd_connect(req->rq_svc_thread->t_env,
 987                                          &export, target, &cluuid, data,
 988                                          client_nid);
 989                         if (rc == 0)
 990                                 conn.cookie = export->exp_handle.h_cookie;
 991                 }
 992         } else {
 993                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 994                                    export, target, &cluuid, data, client_nid);
 995         }
 996         if (rc)
 997                 GOTO(out, rc);
 998         /* Return only the parts of obd_connect_data that we understand, so the
 999          * client knows that we don't understand the rest. */
1000         if (data) {
1001                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1002                                                RCL_SERVER);
1003                 tmpdata = req_capsule_server_get(&req->rq_pill,
1004                                                  &RMF_CONNECT_DATA);
1005                 /* Don't use struct assignment here, because the client reply
1006                  * buffer may be smaller/larger than the local struct
1007                  * obd_connect_data. */
1008                 memcpy(tmpdata, data, min(tmpsize, size));
1009         }
1010
1011         /* If all else goes well, this is our RPC return code. */
1012         req->rq_status = 0;
1013
1014         lustre_msg_set_handle(req->rq_repmsg, &conn);
1015
1016         /* If the client and the server are the same node, we will already
1017          * have an export that really points to the client's DLM export,
1018          * because we have a shared handles table.
1019          *
1020          * XXX this will go away when shaver stops sending the "connect" handle
1021          * in the real "remote handle" field of the request --phik 24 Apr 2003
1022          */
1023         if (req->rq_export != NULL)
1024                 class_export_put(req->rq_export);
1025
1026         /* request takes one export refcount */
1027         req->rq_export = class_export_get(export);
1028
1029         cfs_spin_lock(&export->exp_lock);
1030         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1031                 cfs_spin_unlock(&export->exp_lock);
1032                 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1033                        "conn_cnt: %d > %d\n",
1034                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1035                        export->exp_conn_cnt,
1036                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1037
1038                 GOTO(out, rc = -EALREADY);
1039         }
1040         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1041         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1042         export->exp_abort_active_req = 0;
1043
1044         /* request from liblustre?  Don't evict it for not pinging. */
1045         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1046                 export->exp_libclient = 1;
1047                 cfs_spin_unlock(&export->exp_lock);
1048
1049                 cfs_spin_lock(&target->obd_dev_lock);
1050                 cfs_list_del_init(&export->exp_obd_chain_timed);
1051                 cfs_spin_unlock(&target->obd_dev_lock);
1052         } else {
1053                 cfs_spin_unlock(&export->exp_lock);
1054         }
1055
1056         if (export->exp_connection != NULL) {
1057                 /* Check to see if connection came from another NID */
1058                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1059                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1060                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1061                                      &export->exp_connection->c_peer.nid,
1062                                      &export->exp_nid_hash);
1063
1064                 ptlrpc_connection_put(export->exp_connection);
1065         }
1066
1067         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1068                                                        req->rq_self,
1069                                                        &remote_uuid);
1070         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1071                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1072                              &export->exp_connection->c_peer.nid,
1073                              &export->exp_nid_hash);
1074         }
1075         /**
1076           class_disconnect->class_export_recovery_cleanup() race
1077          */
1078         cfs_spin_lock(&target->obd_recovery_task_lock);
1079         if (target->obd_recovering && !export->exp_in_recovery &&
1080             !export->exp_disconnected) {
1081                 cfs_spin_lock(&export->exp_lock);
1082                 /* possible race with class_disconnect_stale_exports,
1083                  * export may be already in the eviction process */
1084                 if (export->exp_failed) {
1085                         cfs_spin_unlock(&export->exp_lock);
1086                         GOTO(out, rc = -ENODEV);
1087                 }
1088                 export->exp_in_recovery = 1;
1089                 export->exp_req_replay_needed = 1;
1090                 export->exp_lock_replay_needed = 1;
1091                 cfs_spin_unlock(&export->exp_lock);
1092                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1093                      && (data->ocd_transno == 0))
1094                         CWARN("Connect with zero transno!\n");
1095
1096                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1097                      && data->ocd_transno < target->obd_next_recovery_transno &&
1098                      data->ocd_transno > target->obd_last_committed)
1099                         target->obd_next_recovery_transno = data->ocd_transno;
1100                 target->obd_connected_clients++;
1101                 cfs_atomic_inc(&target->obd_req_replay_clients);
1102                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1103                 if (target->obd_connected_clients ==
1104                     target->obd_max_recoverable_clients)
1105                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1106         }
1107         cfs_spin_unlock(&target->obd_recovery_task_lock);
1108
1109         /* Tell the client we're in recovery, when client is involved in it. */
1110         if (target->obd_recovering)
1111                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1112
1113         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1114         conn = *tmp;
1115
1116         if (export->exp_imp_reverse != NULL) {
1117                 /* destroyed import can be still referenced in ctxt */
1118                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1119                                    KEY_REVIMP_UPD, 0, NULL, NULL);
1120
1121                 client_destroy_import(export->exp_imp_reverse);
1122         }
1123
1124         /* for the rest part, we return -ENOTCONN in case of errors
1125          * in order to let client initialize connection again.
1126          */
1127         revimp = export->exp_imp_reverse = class_new_import(target);
1128         if (!revimp) {
1129                 CERROR("fail to alloc new reverse import.\n");
1130                 GOTO(out, rc = -ENOTCONN);
1131         }
1132
1133         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1134         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1135         revimp->imp_remote_handle = conn;
1136         revimp->imp_dlm_fake = 1;
1137         revimp->imp_state = LUSTRE_IMP_FULL;
1138
1139         /* unknown versions will be caught in
1140          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1141         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1142
1143         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1144             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1145                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1146         else
1147                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1148
1149         if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1150             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1151                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1152         else
1153                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1154
1155         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1156         if (rc) {
1157                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1158                 export->exp_imp_reverse = NULL;
1159                 class_destroy_import(revimp);
1160         }
1161
1162         class_import_put(revimp);
1163 out:
1164         if (export) {
1165                 cfs_spin_lock(&export->exp_lock);
1166                 export->exp_connecting = 0;
1167                 cfs_spin_unlock(&export->exp_lock);
1168
1169                 class_export_put(export);
1170         }
1171         if (targref) {
1172                 cfs_spin_lock(&target->obd_dev_lock);
1173                 target->obd_conn_inprogress--;
1174                 cfs_spin_unlock(&target->obd_dev_lock);
1175
1176                 class_decref(targref, __FUNCTION__, cfs_current());
1177         }
1178         if (rc)
1179                 req->rq_status = rc;
1180         RETURN(rc);
1181 }
1182
1183 int target_handle_disconnect(struct ptlrpc_request *req)
1184 {
1185         int rc;
1186         ENTRY;
1187
1188         rc = req_capsule_server_pack(&req->rq_pill);
1189         if (rc)
1190                 RETURN(rc);
1191
1192         /* keep the rq_export around so we can send the reply */
1193         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1194
1195         RETURN(0);
1196 }
1197
1198 void target_destroy_export(struct obd_export *exp)
1199 {
1200         /* exports created from last_rcvd data, and "fake"
1201            exports created by lctl don't have an import */
1202         if (exp->exp_imp_reverse != NULL)
1203                 client_destroy_import(exp->exp_imp_reverse);
1204
1205         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1206         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1207         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1208         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1209 }
1210
1211 /*
1212  * Recovery functions
1213  */
1214 static void target_request_copy_get(struct ptlrpc_request *req)
1215 {
1216         class_export_rpc_get(req->rq_export);
1217         LASSERT(cfs_list_empty(&req->rq_list));
1218         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1219
1220         /* increase refcount to keep request in queue */
1221         cfs_atomic_inc(&req->rq_refcount);
1222         /** let export know it has replays to be handled */
1223         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1224 }
1225
1226 static void target_request_copy_put(struct ptlrpc_request *req)
1227 {
1228         LASSERT(cfs_list_empty(&req->rq_replay_list));
1229         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1230
1231         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1232         class_export_rpc_put(req->rq_export);
1233         ptlrpc_server_drop_request(req);
1234 }
1235
1236 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1237 {
1238         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1239         struct obd_export     *exp = req->rq_export;
1240         struct ptlrpc_request *reqiter;
1241         int                    dup = 0;
1242
1243         LASSERT(exp);
1244
1245         cfs_spin_lock(&exp->exp_lock);
1246         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1247                                 rq_replay_list) {
1248                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1249                         dup = 1;
1250                         break;
1251                 }
1252         }
1253
1254         if (dup) {
1255                 /* we expect it with RESENT and REPLAY flags */
1256                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1257                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1258                         CERROR("invalid flags %x of resent replay\n",
1259                                lustre_msg_get_flags(req->rq_reqmsg));
1260         } else {
1261                 cfs_list_add_tail(&req->rq_replay_list,
1262                                   &exp->exp_req_replay_queue);
1263         }
1264
1265         cfs_spin_unlock(&exp->exp_lock);
1266         return dup;
1267 }
1268
1269 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1270 {
1271         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1272         LASSERT(req->rq_export);
1273
1274         cfs_spin_lock(&req->rq_export->exp_lock);
1275         cfs_list_del_init(&req->rq_replay_list);
1276         cfs_spin_unlock(&req->rq_export->exp_lock);
1277 }
1278
1279 #ifdef __KERNEL__
1280 static void target_finish_recovery(struct obd_device *obd)
1281 {
1282         ENTRY;
1283         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1284                       obd->obd_name);
1285
1286         ldlm_reprocess_all_ns(obd->obd_namespace);
1287         cfs_spin_lock(&obd->obd_recovery_task_lock);
1288         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1289             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1290             !cfs_list_empty(&obd->obd_final_req_queue)) {
1291                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1292                        obd->obd_name,
1293                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1294                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1295                                "" : "lock ",
1296                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1297                                "" : "final ");
1298                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1299                 LBUG();
1300         }
1301         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1302
1303         obd->obd_recovery_end = cfs_time_current_sec();
1304
1305         /* when recovery finished, cleanup orphans on mds and ost */
1306         if (OBT(obd) && OBP(obd, postrecov)) {
1307                 int rc = OBP(obd, postrecov)(obd);
1308                 if (rc < 0)
1309                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1310                                       obd->obd_name, rc);
1311         }
1312         EXIT;
1313 }
1314
1315 static void abort_req_replay_queue(struct obd_device *obd)
1316 {
1317         struct ptlrpc_request *req, *n;
1318         cfs_list_t abort_list;
1319
1320         CFS_INIT_LIST_HEAD(&abort_list);
1321         cfs_spin_lock(&obd->obd_recovery_task_lock);
1322         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1323         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1324         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1325                 DEBUG_REQ(D_WARNING, req, "aborted:");
1326                 req->rq_status = -ENOTCONN;
1327                 if (ptlrpc_error(req)) {
1328                         DEBUG_REQ(D_ERROR, req,
1329                                   "failed abort_req_reply; skipping");
1330                 }
1331                 target_exp_dequeue_req_replay(req);
1332                 target_request_copy_put(req);
1333         }
1334 }
1335
1336 static void abort_lock_replay_queue(struct obd_device *obd)
1337 {
1338         struct ptlrpc_request *req, *n;
1339         cfs_list_t abort_list;
1340
1341         CFS_INIT_LIST_HEAD(&abort_list);
1342         cfs_spin_lock(&obd->obd_recovery_task_lock);
1343         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1344         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1345         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1346                 DEBUG_REQ(D_ERROR, req, "aborted:");
1347                 req->rq_status = -ENOTCONN;
1348                 if (ptlrpc_error(req)) {
1349                         DEBUG_REQ(D_ERROR, req,
1350                                   "failed abort_lock_reply; skipping");
1351                 }
1352                 target_request_copy_put(req);
1353         }
1354 }
1355 #endif
1356
1357 /* Called from a cleanup function if the device is being cleaned up
1358    forcefully.  The exports should all have been disconnected already,
1359    the only thing left to do is
1360      - clear the recovery flags
1361      - cancel the timer
1362      - free queued requests and replies, but don't send replies
1363    Because the obd_stopping flag is set, no new requests should be received.
1364
1365 */
1366 void target_cleanup_recovery(struct obd_device *obd)
1367 {
1368         struct ptlrpc_request *req, *n;
1369         cfs_list_t clean_list;
1370         ENTRY;
1371
1372         CFS_INIT_LIST_HEAD(&clean_list);
1373         cfs_spin_lock(&obd->obd_dev_lock);
1374         if (!obd->obd_recovering) {
1375                 cfs_spin_unlock(&obd->obd_dev_lock);
1376                 EXIT;
1377                 return;
1378         }
1379         obd->obd_recovering = obd->obd_abort_recovery = 0;
1380         cfs_spin_unlock(&obd->obd_dev_lock);
1381
1382         cfs_spin_lock(&obd->obd_recovery_task_lock);
1383         target_cancel_recovery_timer(obd);
1384         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1385         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1386
1387         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1388                 LASSERT(req->rq_reply_state == 0);
1389                 target_exp_dequeue_req_replay(req);
1390                 target_request_copy_put(req);
1391         }
1392
1393         cfs_spin_lock(&obd->obd_recovery_task_lock);
1394         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1395         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1396         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1397
1398         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1399                 LASSERT(req->rq_reply_state == 0);
1400                 target_request_copy_put(req);
1401         }
1402
1403         EXIT;
1404 }
1405
1406 /* obd_recovery_task_lock should be held */
1407 void target_cancel_recovery_timer(struct obd_device *obd)
1408 {
1409         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1410         cfs_timer_disarm(&obd->obd_recovery_timer);
1411 }
1412
1413 /* extend = 1 means require at least "duration" seconds left in the timer,
1414    extend = 0 means set the total duration (start_recovery_timer) */
1415 static void reset_recovery_timer(struct obd_device *obd, int duration,
1416                                  int extend)
1417 {
1418         cfs_time_t now = cfs_time_current_sec();
1419         cfs_duration_t left;
1420
1421         cfs_spin_lock(&obd->obd_recovery_task_lock);
1422         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1423                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1424                 return;
1425         }
1426
1427         left = cfs_time_sub(obd->obd_recovery_end, now);
1428
1429         if (extend && (duration > left))
1430                 obd->obd_recovery_timeout += duration - left;
1431         else if (!extend && (duration > obd->obd_recovery_timeout))
1432                 /* Track the client's largest expected replay time */
1433                 obd->obd_recovery_timeout = duration;
1434
1435         /* Hard limit of obd_recovery_time_hard which should not happen */
1436         if (obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1437                 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1438
1439         obd->obd_recovery_end = obd->obd_recovery_start +
1440                                 obd->obd_recovery_timeout;
1441         if (cfs_time_before(now, obd->obd_recovery_end)) {
1442                 left = cfs_time_sub(obd->obd_recovery_end, now);
1443                 LASSERT(left != 0);
1444                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1445         }
1446         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1447         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1448                obd->obd_name, (unsigned)left);
1449 }
1450
1451 static void check_and_start_recovery_timer(struct obd_device *obd)
1452 {
1453         cfs_spin_lock(&obd->obd_recovery_task_lock);
1454         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1455                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1456                 return;
1457         }
1458         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1459         obd->obd_recovery_start = cfs_time_current_sec();
1460         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1461
1462         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1463 }
1464
1465 /* Reset the timer with each new client connection */
1466 /*
1467  * This timer is actually reconnect_timer, which is for making sure
1468  * the total recovery window is at least as big as my reconnect
1469  * attempt timing. So the initial recovery time_out will be set to
1470  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1471  * from client is bigger than this, then the recovery time_out will
1472  * be extended to make sure the client could be reconnected, in the
1473  * process, the timeout from the new client should be ignored.
1474  */
1475
1476 static void
1477 target_start_and_reset_recovery_timer(struct obd_device *obd,
1478                                       struct ptlrpc_request *req,
1479                                       int new_client)
1480 {
1481         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1482
1483         if (!new_client && service_time)
1484                 /* Teach server about old server's estimates, as first guess
1485                  * at how long new requests will take. */
1486                 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1487                             service_time);
1488
1489         check_and_start_recovery_timer(obd);
1490
1491         /* convert the service time to rpc timeout,
1492          * reuse service_time to limit stack usage */
1493         service_time = at_est2timeout(service_time);
1494
1495         /* We expect other clients to timeout within service_time, then try
1496          * to reconnect, then try the failover server.  The max delay between
1497          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1498         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1499                              INITIAL_CONNECT_TIMEOUT);
1500         if (service_time > obd->obd_recovery_timeout && !new_client)
1501                 reset_recovery_timer(obd, service_time, 0);
1502 }
1503
1504 #ifdef __KERNEL__
1505
1506 /** Health checking routines */
1507 static inline int exp_connect_healthy(struct obd_export *exp)
1508 {
1509         return (exp->exp_in_recovery);
1510 }
1511
1512 /** if export done req_replay or has replay in queue */
1513 static inline int exp_req_replay_healthy(struct obd_export *exp)
1514 {
1515         return (!exp->exp_req_replay_needed ||
1516                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1517 }
1518 /** if export done lock_replay or has replay in queue */
1519 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1520 {
1521         return (!exp->exp_lock_replay_needed ||
1522                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1523 }
1524
1525 static inline int exp_vbr_healthy(struct obd_export *exp)
1526 {
1527         return (!exp->exp_vbr_failed);
1528 }
1529
1530 static inline int exp_finished(struct obd_export *exp)
1531 {
1532         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1533 }
1534
1535 /** Checking routines for recovery */
1536 static int check_for_clients(struct obd_device *obd)
1537 {
1538         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1539                 return 1;
1540         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1541         if (obd->obd_no_conn == 0 &&
1542             obd->obd_connected_clients + obd->obd_stale_clients ==
1543             obd->obd_max_recoverable_clients)
1544                 return 1;
1545         return 0;
1546 }
1547
1548 static int check_for_next_transno(struct obd_device *obd)
1549 {
1550         struct ptlrpc_request *req = NULL;
1551         int wake_up = 0, connected, completed, queue_len;
1552         __u64 next_transno, req_transno;
1553         ENTRY;
1554
1555         cfs_spin_lock(&obd->obd_recovery_task_lock);
1556         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1557                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1558                                      struct ptlrpc_request, rq_list);
1559                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1560         } else {
1561                 req_transno = 0;
1562         }
1563
1564         connected = obd->obd_connected_clients;
1565         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1566         queue_len = obd->obd_requests_queued_for_recovery;
1567         next_transno = obd->obd_next_recovery_transno;
1568
1569         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1570                "req_transno: "LPU64", next_transno: "LPU64"\n",
1571                obd->obd_max_recoverable_clients, connected, completed,
1572                queue_len, req_transno, next_transno);
1573
1574         if (obd->obd_abort_recovery) {
1575                 CDEBUG(D_HA, "waking for aborted recovery\n");
1576                 wake_up = 1;
1577         } else if (obd->obd_recovery_expired) {
1578                 CDEBUG(D_HA, "waking for expired recovery\n");
1579                 wake_up = 1;
1580         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1581                 CDEBUG(D_HA, "waking for completed recovery\n");
1582                 wake_up = 1;
1583         } else if (req_transno == next_transno) {
1584                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1585                 wake_up = 1;
1586         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1587                 int d_lvl = D_HA;
1588                 /** handle gaps occured due to lost reply or VBR */
1589                 LASSERTF(req_transno >= next_transno,
1590                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1591                          req_transno, next_transno);
1592                 if (req_transno > obd->obd_last_committed &&
1593                     !obd->obd_version_recov)
1594                         d_lvl = D_ERROR;
1595                 CDEBUG(d_lvl,
1596                        "%s: waking for gap in transno, VBR is %s (skip: "
1597                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1598                        ", last_committed: "LPD64")\n",
1599                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1600                        next_transno, queue_len, completed, connected,
1601                        req_transno, obd->obd_last_committed);
1602                 obd->obd_next_recovery_transno = req_transno;
1603                 wake_up = 1;
1604         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1605                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1606                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1607                 obd->obd_next_recovery_transno = req_transno;
1608                 wake_up = 1;
1609         }
1610         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1611         return wake_up;
1612 }
1613
1614 static int check_for_next_lock(struct obd_device *obd)
1615 {
1616         int wake_up = 0;
1617
1618         cfs_spin_lock(&obd->obd_recovery_task_lock);
1619         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1620                 CDEBUG(D_HA, "waking for next lock\n");
1621                 wake_up = 1;
1622         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1623                 CDEBUG(D_HA, "waking for completed lock replay\n");
1624                 wake_up = 1;
1625         } else if (obd->obd_abort_recovery) {
1626                 CDEBUG(D_HA, "waking for aborted recovery\n");
1627                 wake_up = 1;
1628         } else if (obd->obd_recovery_expired) {
1629                 CDEBUG(D_HA, "waking for expired recovery\n");
1630                 wake_up = 1;
1631         }
1632         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1633
1634         return wake_up;
1635 }
1636
1637 /**
1638  * wait for recovery events,
1639  * check its status with help of check_routine
1640  * evict dead clients via health_check
1641  */
1642 static int target_recovery_overseer(struct obd_device *obd,
1643                                     int (*check_routine)(struct obd_device *),
1644                                     int (*health_check)(struct obd_export *))
1645 {
1646 repeat:
1647         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1648         if (obd->obd_abort_recovery) {
1649                 CWARN("recovery is aborted, evict exports in recovery\n");
1650                 /** evict exports which didn't finish recovery yet */
1651                 class_disconnect_stale_exports(obd, exp_finished);
1652                 return 1;
1653         } else if (obd->obd_recovery_expired) {
1654                 obd->obd_recovery_expired = 0;
1655                 /** If some clients died being recovered, evict them */
1656                 LCONSOLE_WARN("%s: recovery is timed out, "
1657                               "evict stale exports\n", obd->obd_name);
1658                 /** evict cexports with no replay in queue, they are stalled */
1659                 class_disconnect_stale_exports(obd, health_check);
1660                 /** continue with VBR */
1661                 cfs_spin_lock(&obd->obd_dev_lock);
1662                 obd->obd_version_recov = 1;
1663                 cfs_spin_unlock(&obd->obd_dev_lock);
1664                 /**
1665                  * reset timer, recovery will proceed with versions now,
1666                  * timeout is set just to handle reconnection delays
1667                  */
1668                 reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1669                 /** Wait for recovery events again, after evicting bad clients */
1670                 goto repeat;
1671         }
1672         return 0;
1673 }
1674
1675 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1676 {
1677         struct ptlrpc_request *req = NULL;
1678         ENTRY;
1679
1680         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1681                obd->obd_next_recovery_transno);
1682
1683         if (target_recovery_overseer(obd, check_for_next_transno,
1684                                      exp_req_replay_healthy)) {
1685                 abort_req_replay_queue(obd);
1686                 abort_lock_replay_queue(obd);
1687         }
1688
1689         cfs_spin_lock(&obd->obd_recovery_task_lock);
1690         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1691                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1692                                      struct ptlrpc_request, rq_list);
1693                 cfs_list_del_init(&req->rq_list);
1694                 obd->obd_requests_queued_for_recovery--;
1695                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1696         } else {
1697                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1698                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1699                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1700                 /** evict exports failed VBR */
1701                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1702         }
1703         RETURN(req);
1704 }
1705
1706 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1707 {
1708         struct ptlrpc_request *req = NULL;
1709
1710         CDEBUG(D_HA, "Waiting for lock\n");
1711         if (target_recovery_overseer(obd, check_for_next_lock,
1712                                      exp_lock_replay_healthy))
1713                 abort_lock_replay_queue(obd);
1714
1715         cfs_spin_lock(&obd->obd_recovery_task_lock);
1716         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1717                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1718                                      struct ptlrpc_request, rq_list);
1719                 cfs_list_del_init(&req->rq_list);
1720                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1721         } else {
1722                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1723                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1724                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1725                 /** evict exports failed VBR */
1726                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1727         }
1728         return req;
1729 }
1730
1731 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1732 {
1733         struct ptlrpc_request *req = NULL;
1734
1735         cfs_spin_lock(&obd->obd_recovery_task_lock);
1736         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1737                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1738                                      struct ptlrpc_request, rq_list);
1739                 cfs_list_del_init(&req->rq_list);
1740                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1741                 if (req->rq_export->exp_in_recovery) {
1742                         cfs_spin_lock(&req->rq_export->exp_lock);
1743                         req->rq_export->exp_in_recovery = 0;
1744                         cfs_spin_unlock(&req->rq_export->exp_lock);
1745                 }
1746         } else {
1747                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1748         }
1749         return req;
1750 }
1751
1752 static int handle_recovery_req(struct ptlrpc_thread *thread,
1753                                struct ptlrpc_request *req,
1754                                svc_handler_t handler)
1755 {
1756         int rc;
1757         ENTRY;
1758
1759         /**
1760          * export can be evicted during recovery, no need to handle replays for
1761          * it after that, discard such request silently
1762          */
1763         if (req->rq_export->exp_disconnected)
1764                 GOTO(reqcopy_put, rc = 0);
1765
1766         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1767         if (rc) {
1768                 CERROR("Failure to initialize session: %d\n", rc);
1769                 GOTO(reqcopy_put, rc);
1770         }
1771
1772         req->rq_recov_session.lc_thread = thread;
1773         lu_context_enter(&req->rq_recov_session);
1774         req->rq_svc_thread = thread;
1775         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1776
1777         /* thread context */
1778         lu_context_enter(&thread->t_env->le_ctx);
1779         (void)handler(req);
1780         lu_context_exit(&thread->t_env->le_ctx);
1781
1782         lu_context_exit(&req->rq_recov_session);
1783         lu_context_fini(&req->rq_recov_session);
1784         /* don't reset timer for final stage */
1785         if (!exp_finished(req->rq_export)) {
1786                 /**
1787                  * Add request timeout to the recovery time so next request from
1788                  * this client may come in recovery time
1789                  */
1790                  reset_recovery_timer(class_exp2obd(req->rq_export),
1791                                       AT_OFF ? obd_timeout :
1792                                       lustre_msg_get_timeout(req->rq_reqmsg), 1);
1793         }
1794 reqcopy_put:
1795         RETURN(rc);
1796 }
1797
1798 static int target_recovery_thread(void *arg)
1799 {
1800         struct lu_target *lut = arg;
1801         struct obd_device *obd = lut->lut_obd;
1802         struct ptlrpc_request *req;
1803         struct target_recovery_data *trd = &obd->obd_recovery_data;
1804         unsigned long delta;
1805         unsigned long flags;
1806         struct lu_env env;
1807         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1808         int rc = 0;
1809         ENTRY;
1810
1811         cfs_daemonize_ctxt("tgt_recov");
1812
1813         SIGNAL_MASK_LOCK(current, flags);
1814         sigfillset(&current->blocked);
1815         RECALC_SIGPENDING;
1816         SIGNAL_MASK_UNLOCK(current, flags);
1817
1818         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1819         if (rc)
1820                 RETURN(rc);
1821
1822         thread->t_env = &env;
1823         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1824         env.le_ctx.lc_thread = thread;
1825         thread->t_data = NULL;
1826         thread->t_watchdog = NULL;
1827
1828         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1829                cfs_curproc_pid());
1830         trd->trd_processing_task = cfs_curproc_pid();
1831
1832         cfs_spin_lock(&obd->obd_dev_lock);
1833         obd->obd_recovering = 1;
1834         cfs_spin_unlock(&obd->obd_dev_lock);
1835         cfs_complete(&trd->trd_starting);
1836
1837         /* first of all, we have to know the first transno to replay */
1838         if (target_recovery_overseer(obd, check_for_clients,
1839                                      exp_connect_healthy)) {
1840                 abort_req_replay_queue(obd);
1841                 abort_lock_replay_queue(obd);
1842         }
1843
1844         /* next stage: replay requests */
1845         delta = jiffies;
1846         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1847                cfs_atomic_read(&obd->obd_req_replay_clients),
1848                obd->obd_next_recovery_transno);
1849         while ((req = target_next_replay_req(obd))) {
1850                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1851                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1852                           lustre_msg_get_transno(req->rq_reqmsg),
1853                           libcfs_nid2str(req->rq_peer.nid));
1854                 handle_recovery_req(thread, req,
1855                                     trd->trd_recovery_handler);
1856                 /**
1857                  * bz18031: increase next_recovery_transno before
1858                  * target_request_copy_put() will drop exp_rpc reference
1859                  */
1860                 cfs_spin_lock(&obd->obd_recovery_task_lock);
1861                 obd->obd_next_recovery_transno++;
1862                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1863                 target_exp_dequeue_req_replay(req);
1864                 target_request_copy_put(req);
1865                 obd->obd_replayed_requests++;
1866         }
1867
1868         /**
1869          * The second stage: replay locks
1870          */
1871         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1872                cfs_atomic_read(&obd->obd_lock_replay_clients));
1873         while ((req = target_next_replay_lock(obd))) {
1874                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1875                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1876                           libcfs_nid2str(req->rq_peer.nid));
1877                 handle_recovery_req(thread, req,
1878                                     trd->trd_recovery_handler);
1879                 target_request_copy_put(req);
1880                 obd->obd_replayed_locks++;
1881         }
1882
1883         /**
1884          * The third stage: reply on final pings, at this moment all clients
1885          * must have request in final queue
1886          */
1887         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1888         /** Update server last boot epoch */
1889         lut_boot_epoch_update(lut);
1890         /* We drop recoverying flag to forward all new requests
1891          * to regular mds_handle() since now */
1892         cfs_spin_lock(&obd->obd_dev_lock);
1893         obd->obd_recovering = obd->obd_abort_recovery = 0;
1894         cfs_spin_unlock(&obd->obd_dev_lock);
1895         cfs_spin_lock(&obd->obd_recovery_task_lock);
1896         target_cancel_recovery_timer(obd);
1897         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1898         while ((req = target_next_final_ping(obd))) {
1899                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1900                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1901                           libcfs_nid2str(req->rq_peer.nid));
1902                 handle_recovery_req(thread, req,
1903                                     trd->trd_recovery_handler);
1904                 target_request_copy_put(req);
1905         }
1906
1907         delta = (jiffies - delta) / CFS_HZ;
1908         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1909               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1910         if (delta > OBD_RECOVERY_TIME_SOFT) {
1911                 CWARN("too long recovery - read logs\n");
1912                 libcfs_debug_dumplog();
1913         }
1914
1915         target_finish_recovery(obd);
1916
1917         lu_context_fini(&env.le_ctx);
1918         trd->trd_processing_task = 0;
1919         cfs_complete(&trd->trd_finishing);
1920         RETURN(rc);
1921 }
1922
1923 static int target_start_recovery_thread(struct lu_target *lut,
1924                                         svc_handler_t handler)
1925 {
1926         struct obd_device *obd = lut->lut_obd;
1927         int rc = 0;
1928         struct target_recovery_data *trd = &obd->obd_recovery_data;
1929
1930         memset(trd, 0, sizeof(*trd));
1931         cfs_init_completion(&trd->trd_starting);
1932         cfs_init_completion(&trd->trd_finishing);
1933         trd->trd_recovery_handler = handler;
1934
1935         if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
1936                 cfs_wait_for_completion(&trd->trd_starting);
1937                 LASSERT(obd->obd_recovering != 0);
1938         } else
1939                 rc = -ECHILD;
1940
1941         return rc;
1942 }
1943
1944 void target_stop_recovery_thread(struct obd_device *obd)
1945 {
1946         if (obd->obd_recovery_data.trd_processing_task > 0) {
1947                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1948                 /** recovery can be done but postrecovery is not yet */
1949                 cfs_spin_lock(&obd->obd_dev_lock);
1950                 if (obd->obd_recovering) {
1951                         CERROR("%s: Aborting recovery\n", obd->obd_name);
1952                         obd->obd_abort_recovery = 1;
1953                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1954                 }
1955                 cfs_spin_unlock(&obd->obd_dev_lock);
1956                 cfs_wait_for_completion(&trd->trd_finishing);
1957         }
1958 }
1959
1960 void target_recovery_fini(struct obd_device *obd)
1961 {
1962         class_disconnect_exports(obd);
1963         target_stop_recovery_thread(obd);
1964         target_cleanup_recovery(obd);
1965 }
1966 EXPORT_SYMBOL(target_recovery_fini);
1967
1968 static void target_recovery_expired(unsigned long castmeharder)
1969 {
1970         struct obd_device *obd = (struct obd_device *)castmeharder;
1971         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1972                " after %lds (%d clients connected)\n",
1973                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1974                cfs_time_current_sec()- obd->obd_recovery_start,
1975                obd->obd_connected_clients);
1976
1977         obd->obd_recovery_expired = 1;
1978         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1979 }
1980
1981 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1982 {
1983         struct obd_device *obd = lut->lut_obd;
1984         if (obd->obd_max_recoverable_clients == 0) {
1985                 /** Update server last boot epoch */
1986                 lut_boot_epoch_update(lut);
1987                 return;
1988         }
1989
1990         CWARN("RECOVERY: service %s, %d recoverable clients, "
1991               "last_transno "LPU64"\n", obd->obd_name,
1992               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1993         LASSERT(obd->obd_stopping == 0);
1994         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1995         obd->obd_recovery_start = 0;
1996         obd->obd_recovery_end = 0;
1997
1998         /* both values can be get from mount data already */
1999         if (obd->obd_recovery_timeout == 0)
2000                 obd->obd_recovery_timeout = OBD_RECOVERY_TIME_SOFT;
2001         if (obd->obd_recovery_time_hard == 0)
2002                 obd->obd_recovery_time_hard = OBD_RECOVERY_TIME_HARD;
2003         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2004         target_start_recovery_thread(lut, handler);
2005 }
2006 EXPORT_SYMBOL(target_recovery_init);
2007
2008 #endif
2009
2010 static int target_process_req_flags(struct obd_device *obd,
2011                                     struct ptlrpc_request *req)
2012 {
2013         struct obd_export *exp = req->rq_export;
2014         LASSERT(exp != NULL);
2015         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2016                 /* client declares he's ready to replay locks */
2017                 cfs_spin_lock(&exp->exp_lock);
2018                 if (exp->exp_req_replay_needed) {
2019                         exp->exp_req_replay_needed = 0;
2020                         cfs_spin_unlock(&exp->exp_lock);
2021
2022                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2023                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2024                 } else {
2025                         cfs_spin_unlock(&exp->exp_lock);
2026                 }
2027         }
2028         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2029                 /* client declares he's ready to complete recovery
2030                  * so, we put the request on th final queue */
2031                 cfs_spin_lock(&exp->exp_lock);
2032                 if (exp->exp_lock_replay_needed) {
2033                         exp->exp_lock_replay_needed = 0;
2034                         cfs_spin_unlock(&exp->exp_lock);
2035
2036                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2037                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2038                 } else {
2039                         cfs_spin_unlock(&exp->exp_lock);
2040                 }
2041         }
2042         return 0;
2043 }
2044
2045 int target_queue_recovery_request(struct ptlrpc_request *req,
2046                                   struct obd_device *obd)
2047 {
2048         cfs_list_t *tmp;
2049         int inserted = 0;
2050         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2051         ENTRY;
2052
2053         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2054                 /* Processing the queue right now, don't re-add. */
2055                 RETURN(1);
2056         }
2057
2058         target_process_req_flags(obd, req);
2059
2060         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2061                 /* client declares he's ready to complete recovery
2062                  * so, we put the request on th final queue */
2063                 target_request_copy_get(req);
2064                 DEBUG_REQ(D_HA, req, "queue final req");
2065                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2066                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2067                 if (obd->obd_recovering) {
2068                         cfs_list_add_tail(&req->rq_list,
2069                                           &obd->obd_final_req_queue);
2070                 } else {
2071                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2072                         target_request_copy_put(req);
2073                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2074                 }
2075                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2076                 RETURN(0);
2077         }
2078         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2079                 /* client declares he's ready to replay locks */
2080                 target_request_copy_get(req);
2081                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2082                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2083                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2084                 LASSERT(obd->obd_recovering);
2085                 /* usually due to recovery abort */
2086                 if (!req->rq_export->exp_in_recovery) {
2087                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2088                         target_request_copy_put(req);
2089                         RETURN(-ENOTCONN);
2090                 }
2091                 LASSERT(req->rq_export->exp_lock_replay_needed);
2092                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2093                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2094                 RETURN(0);
2095         }
2096
2097         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2098          * (i.e. buflens etc are in my own byte order), but type-dependent
2099          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2100
2101         if (!transno) {
2102                 CFS_INIT_LIST_HEAD(&req->rq_list);
2103                 DEBUG_REQ(D_HA, req, "not queueing");
2104                 RETURN(1);
2105         }
2106
2107         /* If we're processing the queue, we want don't want to queue this
2108          * message.
2109          *
2110          * Also, if this request has a transno less than the one we're waiting
2111          * for, we should process it now.  It could (and currently always will)
2112          * be an open request for a descriptor that was opened some time ago.
2113          *
2114          * Also, a resent, replayed request that has already been
2115          * handled will pass through here and be processed immediately.
2116          */
2117         CDEBUG(D_HA, "Next recovery transno: "LPU64
2118                ", current: "LPU64", replaying\n",
2119                obd->obd_next_recovery_transno, transno);
2120         cfs_spin_lock(&obd->obd_recovery_task_lock);
2121         if (transno < obd->obd_next_recovery_transno) {
2122                 /* Processing the queue right now, don't re-add. */
2123                 LASSERT(cfs_list_empty(&req->rq_list));
2124                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2125                 RETURN(1);
2126         }
2127         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2128
2129         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2130                 RETURN(0);
2131
2132         target_request_copy_get(req);
2133         if (!req->rq_export->exp_in_recovery) {
2134                 target_request_copy_put(req);
2135                 RETURN(-ENOTCONN);
2136         }
2137         LASSERT(req->rq_export->exp_req_replay_needed);
2138
2139         if (target_exp_enqueue_req_replay(req)) {
2140                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2141                 target_request_copy_put(req);
2142                 RETURN(0);
2143         }
2144
2145         /* XXX O(n^2) */
2146         cfs_spin_lock(&obd->obd_recovery_task_lock);
2147         LASSERT(obd->obd_recovering);
2148         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2149                 struct ptlrpc_request *reqiter =
2150                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2151
2152                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2153                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2154                         inserted = 1;
2155                         break;
2156                 }
2157
2158                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2159                              transno)) {
2160                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2161                                   "has been claimed by another client");
2162                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2163                         target_exp_dequeue_req_replay(req);
2164                         target_request_copy_put(req);
2165                         RETURN(0);
2166                 }
2167         }
2168
2169         if (!inserted)
2170                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2171
2172         obd->obd_requests_queued_for_recovery++;
2173         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2174         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2175         RETURN(0);
2176 }
2177
2178 /**
2179  * Packs current SLV and Limit into \a req.
2180  */
2181 int target_pack_pool_reply(struct ptlrpc_request *req)
2182 {
2183         struct obd_device *obd;
2184         ENTRY;
2185
2186         /*
2187          * Check that we still have all structures alive as this may
2188          * be some late rpc in shutdown time.
2189          */
2190         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2191                      !exp_connect_lru_resize(req->rq_export))) {
2192                 lustre_msg_set_slv(req->rq_repmsg, 0);
2193                 lustre_msg_set_limit(req->rq_repmsg, 0);
2194                 RETURN(0);
2195         }
2196
2197         /*
2198          * OBD is alive here as export is alive, which we checked above.
2199          */
2200         obd = req->rq_export->exp_obd;
2201
2202         cfs_read_lock(&obd->obd_pool_lock);
2203         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2204         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2205         cfs_read_unlock(&obd->obd_pool_lock);
2206
2207         RETURN(0);
2208 }
2209
2210 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2211 {
2212         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2213                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2214                 return (-ECOMM);
2215         }
2216
2217         if (unlikely(rc)) {
2218                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2219                 req->rq_status = rc;
2220                 return (ptlrpc_send_error(req, 1));
2221         } else {
2222                 DEBUG_REQ(D_NET, req, "sending reply");
2223         }
2224
2225         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2226 }
2227
2228 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2229 {
2230         int                        netrc;
2231         struct ptlrpc_reply_state *rs;
2232         struct obd_device         *obd;
2233         struct obd_export         *exp;
2234         struct ptlrpc_service     *svc;
2235         ENTRY;
2236
2237         if (req->rq_no_reply) {
2238                 EXIT;
2239                 return;
2240         }
2241
2242         svc = req->rq_rqbd->rqbd_service;
2243         rs = req->rq_reply_state;
2244         if (rs == NULL || !rs->rs_difficult) {
2245                 /* no notifiers */
2246                 target_send_reply_msg (req, rc, fail_id);
2247                 EXIT;
2248                 return;
2249         }
2250
2251         /* must be an export if locks saved */
2252         LASSERT (req->rq_export != NULL);
2253         /* req/reply consistent */
2254         LASSERT (rs->rs_service == svc);
2255
2256         /* "fresh" reply */
2257         LASSERT (!rs->rs_scheduled);
2258         LASSERT (!rs->rs_scheduled_ever);
2259         LASSERT (!rs->rs_handled);
2260         LASSERT (!rs->rs_on_net);
2261         LASSERT (rs->rs_export == NULL);
2262         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2263         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2264
2265         exp = class_export_get (req->rq_export);
2266         obd = exp->exp_obd;
2267
2268         /* disable reply scheduling while I'm setting up */
2269         rs->rs_scheduled = 1;
2270         rs->rs_on_net    = 1;
2271         rs->rs_xid       = req->rq_xid;
2272         rs->rs_transno   = req->rq_transno;
2273         rs->rs_export    = exp;
2274         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2275
2276         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2277         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2278                rs->rs_transno, exp->exp_last_committed);
2279         if (rs->rs_transno > exp->exp_last_committed) {
2280                 /* not committed already */
2281                 cfs_list_add_tail(&rs->rs_obd_list,
2282                                   &exp->exp_uncommitted_replies);
2283         }
2284         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2285
2286         cfs_spin_lock(&exp->exp_lock);
2287         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2288         cfs_spin_unlock(&exp->exp_lock);
2289
2290         netrc = target_send_reply_msg (req, rc, fail_id);
2291
2292         cfs_spin_lock(&svc->srv_rs_lock);
2293
2294         cfs_atomic_inc(&svc->srv_n_difficult_replies);
2295
2296         if (netrc != 0) {
2297                 /* error sending: reply is off the net.  Also we need +1
2298                  * reply ref until ptlrpc_handle_rs() is done
2299                  * with the reply state (if the send was successful, there
2300                  * would have been +1 ref for the net, which
2301                  * reply_out_callback leaves alone) */
2302                 rs->rs_on_net = 0;
2303                 ptlrpc_rs_addref(rs);
2304         }
2305
2306         cfs_spin_lock(&rs->rs_lock);
2307         if (rs->rs_transno <= exp->exp_last_committed ||
2308             (!rs->rs_on_net && !rs->rs_no_ack) ||
2309              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2310              cfs_list_empty(&rs->rs_obd_list)) {
2311                 CDEBUG(D_HA, "Schedule reply immediately\n");
2312                 ptlrpc_dispatch_difficult_reply(rs);
2313         } else {
2314                 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2315                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2316         }
2317         cfs_spin_unlock(&rs->rs_lock);
2318         cfs_spin_unlock(&svc->srv_rs_lock);
2319         EXIT;
2320 }
2321
2322 int target_handle_ping(struct ptlrpc_request *req)
2323 {
2324         obd_ping(req->rq_export);
2325         return req_capsule_server_pack(&req->rq_pill);
2326 }
2327
2328 void target_committed_to_req(struct ptlrpc_request *req)
2329 {
2330         struct obd_export *exp = req->rq_export;
2331
2332         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2333                 lustre_msg_set_last_committed(req->rq_repmsg,
2334                                               exp->exp_last_committed);
2335         else
2336                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2337                           "%d)", exp->exp_obd->obd_no_transno,
2338                           req->rq_repmsg == NULL);
2339
2340         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2341                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2342 }
2343 EXPORT_SYMBOL(target_committed_to_req);
2344
2345 int target_handle_qc_callback(struct ptlrpc_request *req)
2346 {
2347         struct obd_quotactl *oqctl;
2348         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2349
2350         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2351         if (oqctl == NULL) {
2352                 CERROR("Can't unpack obd_quotactl\n");
2353                 RETURN(-EPROTO);
2354         }
2355
2356         cli->cl_qchk_stat = oqctl->qc_stat;
2357
2358         return 0;
2359 }
2360
2361 #ifdef HAVE_QUOTA_SUPPORT
2362 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2363 {
2364 #ifdef __KERNEL__
2365         struct obd_device *obd = req->rq_export->exp_obd;
2366         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2367         struct obd_device_target *obt;
2368         struct lustre_quota_ctxt *qctxt;
2369         struct qunit_data *qdata = NULL;
2370         int rc = 0;
2371         ENTRY;
2372
2373         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2374                 RETURN(rc);
2375
2376         rc = req_capsule_server_pack(&req->rq_pill);
2377         if (rc) {
2378                 CERROR("packing reply failed!: rc = %d\n", rc);
2379                 RETURN(rc);
2380         }
2381
2382         LASSERT(req->rq_export);
2383
2384         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2385         if (IS_ERR(qdata)) {
2386                 rc = PTR_ERR(qdata);
2387                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2388                 req->rq_status = rc;
2389                 GOTO(out, rc);
2390         }
2391
2392         /* we use the observer */
2393         if (obd_pin_observer(obd, &lov_obd) ||
2394             obd_pin_observer(lov_obd, &master_obd)) {
2395                 CERROR("Can't find the observer, it is recovering\n");
2396                 req->rq_status = -EAGAIN;
2397                 GOTO(out, rc);
2398         }
2399
2400         obt = &master_obd->u.obt;
2401         qctxt = &obt->obt_qctxt;
2402
2403         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2404                 /* quota_type has not been processed yet, return EAGAIN
2405                  * until we know whether or not quotas are supposed to
2406                  * be enabled */
2407                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2408                        "-EAGAIN\n");
2409                 req->rq_status = -EAGAIN;
2410                 GOTO(out, rc);
2411         }
2412
2413         cfs_down_read(&obt->obt_rwsem);
2414         if (qctxt->lqc_lqs_hash == NULL) {
2415                 cfs_up_read(&obt->obt_rwsem);
2416                 /* quota_type has not been processed yet, return EAGAIN
2417                  * until we know whether or not quotas are supposed to
2418                  * be enabled */
2419                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2420                        "-EAGAIN\n");
2421                 req->rq_status = -EAGAIN;
2422                 GOTO(out, rc);
2423         }
2424
2425         LASSERT(qctxt->lqc_handler);
2426         rc = qctxt->lqc_handler(master_obd, qdata,
2427                                 lustre_msg_get_opc(req->rq_reqmsg));
2428         cfs_up_read(&obt->obt_rwsem);
2429         if (rc && rc != -EDQUOT)
2430                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2431                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2432         req->rq_status = rc;
2433
2434         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2435         if (rc < 0) {
2436                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2437                 GOTO(out, rc);
2438         }
2439
2440         /* Block the quota req. b=14840 */
2441         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2442         EXIT;
2443
2444 out:
2445         if (master_obd)
2446                 obd_unpin_observer(lov_obd);
2447         if (lov_obd)
2448                 obd_unpin_observer(obd);
2449
2450         rc = ptlrpc_reply(req);
2451         return rc;
2452 #else
2453         return 0;
2454 #endif /* !__KERNEL__ */
2455 }
2456 #endif /* HAVE_QUOTA_SUPPORT */
2457
2458 ldlm_mode_t lck_compat_array[] = {
2459         [LCK_EX] LCK_COMPAT_EX,
2460         [LCK_PW] LCK_COMPAT_PW,
2461         [LCK_PR] LCK_COMPAT_PR,
2462         [LCK_CW] LCK_COMPAT_CW,
2463         [LCK_CR] LCK_COMPAT_CR,
2464         [LCK_NL] LCK_COMPAT_NL,
2465         [LCK_GROUP] LCK_COMPAT_GROUP,
2466         [LCK_COS] LCK_COMPAT_COS,
2467 };
2468
2469 /**
2470  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2471  * not escape to the user level.
2472  */
2473 int ldlm_error2errno(ldlm_error_t error)
2474 {
2475         int result;
2476
2477         switch (error) {
2478         case ELDLM_OK:
2479                 result = 0;
2480                 break;
2481         case ELDLM_LOCK_CHANGED:
2482                 result = -ESTALE;
2483                 break;
2484         case ELDLM_LOCK_ABORTED:
2485                 result = -ENAVAIL;
2486                 break;
2487         case ELDLM_LOCK_REPLACED:
2488                 result = -ESRCH;
2489                 break;
2490         case ELDLM_NO_LOCK_DATA:
2491                 result = -ENOENT;
2492                 break;
2493         case ELDLM_NAMESPACE_EXISTS:
2494                 result = -EEXIST;
2495                 break;
2496         case ELDLM_BAD_NAMESPACE:
2497                 result = -EBADF;
2498                 break;
2499         default:
2500                 if (((int)error) < 0)  /* cast to signed type */
2501                         result = error; /* as ldlm_error_t can be unsigned */
2502                 else {
2503                         CERROR("Invalid DLM result code: %d\n", error);
2504                         result = -EPROTO;
2505                 }
2506         }
2507         return result;
2508 }
2509 EXPORT_SYMBOL(ldlm_error2errno);
2510
2511 /**
2512  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2513  */
2514 ldlm_error_t ldlm_errno2error(int err_no)
2515 {
2516         int error;
2517
2518         switch (err_no) {
2519         case 0:
2520                 error = ELDLM_OK;
2521                 break;
2522         case -ESTALE:
2523                 error = ELDLM_LOCK_CHANGED;
2524                 break;
2525         case -ENAVAIL:
2526                 error = ELDLM_LOCK_ABORTED;
2527                 break;
2528         case -ESRCH:
2529                 error = ELDLM_LOCK_REPLACED;
2530                 break;
2531         case -ENOENT:
2532                 error = ELDLM_NO_LOCK_DATA;
2533                 break;
2534         case -EEXIST:
2535                 error = ELDLM_NAMESPACE_EXISTS;
2536                 break;
2537         case -EBADF:
2538                 error = ELDLM_BAD_NAMESPACE;
2539                 break;
2540         default:
2541                 error = err_no;
2542         }
2543         return error;
2544 }
2545 EXPORT_SYMBOL(ldlm_errno2error);
2546
2547 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2548 void ldlm_dump_export_locks(struct obd_export *exp)
2549 {
2550         cfs_spin_lock(&exp->exp_locks_list_guard);
2551         if (!cfs_list_empty(&exp->exp_locks_list)) {
2552             struct ldlm_lock *lock;
2553
2554             CERROR("dumping locks for export %p,"
2555                    "ignore if the unmount doesn't hang\n", exp);
2556             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2557                 ldlm_lock_dump(D_ERROR, lock, 0);
2558         }
2559         cfs_spin_unlock(&exp->exp_locks_list_guard);
2560 }
2561 #endif
2562
2563 static int target_bulk_timeout(void *data)
2564 {
2565         ENTRY;
2566         /* We don't fail the connection here, because having the export
2567          * killed makes the (vital) call to commitrw very sad.
2568          */
2569         RETURN(1);
2570 }
2571
2572 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2573 {
2574         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2575 }
2576
2577 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2578                    struct l_wait_info *lwi)
2579 {
2580         struct ptlrpc_request *req = desc->bd_req;
2581         int rc = 0;
2582         ENTRY;
2583
2584         /* Check if there is eviction in progress, and if so, wait for
2585          * it to finish */
2586         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2587                 *lwi = LWI_INTR(NULL, NULL);
2588                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2589                                   !cfs_atomic_read(&exp->exp_obd->
2590                                                    obd_evict_inprogress),
2591                                   lwi);
2592         }
2593
2594         /* Check if client was evicted or tried to reconnect already */
2595         if (exp->exp_failed || exp->exp_abort_active_req) {
2596                 rc = -ENOTCONN;
2597         } else {
2598                 if (desc->bd_type == BULK_PUT_SINK)
2599                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2600                 if (rc == 0)
2601                         rc = ptlrpc_start_bulk_transfer(desc);
2602         }
2603
2604         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2605                 ptlrpc_abort_bulk(desc);
2606         } else if (rc == 0) {
2607                 time_t start = cfs_time_current_sec();
2608                 do {
2609                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2610                         cfs_duration_t timeout = timeoutl <= 0 ?
2611                                 CFS_TICK : cfs_time_seconds(timeoutl);
2612                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2613                                                     cfs_time_seconds(1),
2614                                                    target_bulk_timeout,
2615                                                    desc);
2616                         rc = l_wait_event(desc->bd_waitq,
2617                                           !ptlrpc_server_bulk_active(desc) ||
2618                                           exp->exp_failed ||
2619                                           exp->exp_abort_active_req,
2620                                           lwi);
2621                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2622                         /* Wait again if we changed deadline */
2623                 } while ((rc == -ETIMEDOUT) &&
2624                          (req->rq_deadline > cfs_time_current_sec()));
2625
2626                 if (rc == -ETIMEDOUT) {
2627                         DEBUG_REQ(D_ERROR, req,
2628                                   "timeout on bulk %s after %ld%+lds",
2629                                   bulk2type(desc),
2630                                   req->rq_deadline - start,
2631                                   cfs_time_current_sec() -
2632                                   req->rq_deadline);
2633                         ptlrpc_abort_bulk(desc);
2634                 } else if (exp->exp_failed) {
2635                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2636                                   bulk2type(desc));
2637                         rc = -ENOTCONN;
2638                         ptlrpc_abort_bulk(desc);
2639                 } else if (exp->exp_abort_active_req) {
2640                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2641                                   bulk2type(desc));
2642                         /* we don't reply anyway */
2643                         rc = -ETIMEDOUT;
2644                         ptlrpc_abort_bulk(desc);
2645                 } else if (!desc->bd_success ||
2646                            desc->bd_nob_transferred != desc->bd_nob) {
2647                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2648                                   desc->bd_success ?
2649                                   "truncated" : "network error on",
2650                                   bulk2type(desc),
2651                                   desc->bd_nob_transferred,
2652                                   desc->bd_nob);
2653                         /* XXX should this be a different errno? */
2654                         rc = -ETIMEDOUT;
2655                 } else if (desc->bd_type == BULK_GET_SINK) {
2656                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2657                 }
2658         } else {
2659                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2660                           bulk2type(desc), rc);
2661         }
2662
2663         RETURN(rc);
2664 }
2665 EXPORT_SYMBOL(target_bulk_io);