lustre/ldlm/ldlm_lib.c

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