lustre/ldlm/ldlm_lib.c

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