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