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