lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * Copyright (c) 2011 Whamcloud, Inc.
  34  */
  35 /*
  36  * This file is part of Lustre, http://www.lustre.org/
  37  * Lustre is a trademark of Sun Microsystems, Inc.
  38  */
  39
  40 #ifndef EXPORT_SYMTAB
  41 # define EXPORT_SYMTAB
  42 #endif
  43 #define DEBUG_SUBSYSTEM S_LDLM
  44
  45 #ifdef __KERNEL__
  46 # include <libcfs/libcfs.h>
  47 #else
  48 # include <liblustre.h>
  49 #endif
  50 #include <obd.h>
  51 #include <lustre_mds.h>
  52 #include <lustre_dlm.h>
  53 #include <lustre_net.h>
  54 #include <lustre_sec.h>
  55 #include "ldlm_internal.h"
  56
  57 /* @priority: if non-zero, move the selected to the list head
  58  * @create: if zero, only search in existed connections
  59  */
  60 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  61                            int priority, int create)
  62 {
  63         struct ptlrpc_connection *ptlrpc_conn;
  64         struct obd_import_conn *imp_conn = NULL, *item;
  65         int rc = 0;
  66         ENTRY;
  67
  68         if (!create && !priority) {
  69                 CDEBUG(D_HA, "Nothing to do\n");
  70                 RETURN(-EINVAL);
  71         }
  72
  73         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  74         if (!ptlrpc_conn) {
  75                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  76                 RETURN (-ENOENT);
  77         }
  78
  79         if (create) {
  80                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  81                 if (!imp_conn) {
  82                         GOTO(out_put, rc = -ENOMEM);
  83                 }
  84         }
  85
  86         cfs_spin_lock(&imp->imp_lock);
  87         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  88                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  89                         if (priority) {
  90                                 cfs_list_del(&item->oic_item);
  91                                 cfs_list_add(&item->oic_item,
  92                                              &imp->imp_conn_list);
  93                                 item->oic_last_attempt = 0;
  94                         }
  95                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  96                                imp, imp->imp_obd->obd_name, uuid->uuid,
  97                                (priority ? ", moved to head" : ""));
  98                         cfs_spin_unlock(&imp->imp_lock);
  99                         GOTO(out_free, rc = 0);
 100                 }
 101         }
 102         /* not found */
 103         if (create) {
 104                 imp_conn->oic_conn = ptlrpc_conn;
 105                 imp_conn->oic_uuid = *uuid;
 106                 imp_conn->oic_last_attempt = 0;
 107                 if (priority)
 108                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 109                 else
 110                         cfs_list_add_tail(&imp_conn->oic_item,
 111                                           &imp->imp_conn_list);
 112                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 113                        imp, imp->imp_obd->obd_name, uuid->uuid,
 114                        (priority ? "head" : "tail"));
 115         } else {
 116                 cfs_spin_unlock(&imp->imp_lock);
 117                 GOTO(out_free, rc = -ENOENT);
 118         }
 119
 120         cfs_spin_unlock(&imp->imp_lock);
 121         RETURN(0);
 122 out_free:
 123         if (imp_conn)
 124                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 125 out_put:
 126         ptlrpc_connection_put(ptlrpc_conn);
 127         RETURN(rc);
 128 }
 129
 130 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 131 {
 132         return import_set_conn(imp, uuid, 1, 0);
 133 }
 134
 135 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 136                            int priority)
 137 {
 138         return import_set_conn(imp, uuid, priority, 1);
 139 }
 140
 141 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 142 {
 143         struct obd_import_conn *imp_conn;
 144         struct obd_export *dlmexp;
 145         int rc = -ENOENT;
 146         ENTRY;
 147
 148         cfs_spin_lock(&imp->imp_lock);
 149         if (cfs_list_empty(&imp->imp_conn_list)) {
 150                 LASSERT(!imp->imp_connection);
 151                 GOTO(out, rc);
 152         }
 153
 154         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 155                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 156                         continue;
 157                 LASSERT(imp_conn->oic_conn);
 158
 159                 /* is current conn? */
 160                 if (imp_conn == imp->imp_conn_current) {
 161                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 162
 163                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 164                             imp->imp_state != LUSTRE_IMP_DISCON) {
 165                                 CERROR("can't remove current connection\n");
 166                                 GOTO(out, rc = -EBUSY);
 167                         }
 168
 169                         ptlrpc_connection_put(imp->imp_connection);
 170                         imp->imp_connection = NULL;
 171
 172                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 173                         if (dlmexp && dlmexp->exp_connection) {
 174                                 LASSERT(dlmexp->exp_connection ==
 175                                         imp_conn->oic_conn);
 176                                 ptlrpc_connection_put(dlmexp->exp_connection);
 177                                 dlmexp->exp_connection = NULL;
 178                         }
 179                 }
 180
 181                 cfs_list_del(&imp_conn->oic_item);
 182                 ptlrpc_connection_put(imp_conn->oic_conn);
 183                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 184                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 185                        imp, imp->imp_obd->obd_name, uuid->uuid);
 186                 rc = 0;
 187                 break;
 188         }
 189 out:
 190         cfs_spin_unlock(&imp->imp_lock);
 191         if (rc == -ENOENT)
 192                 CERROR("connection %s not found\n", uuid->uuid);
 193         RETURN(rc);
 194 }
 195
 196 /**
 197  * Find conn uuid by peer nid. @peer is a server nid. This function is used
 198  * to find a conn uuid of @imp which can reach @peer.
 199  */
 200 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 201                             struct obd_uuid *uuid)
 202 {
 203         struct obd_import_conn *conn;
 204         int rc = -ENOENT;
 205         ENTRY;
 206
 207         cfs_spin_lock(&imp->imp_lock);
 208         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 209                 /* check if conn uuid does have this peer nid */
 210                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 211                         *uuid = conn->oic_uuid;
 212                         rc = 0;
 213                         break;
 214                 }
 215         }
 216         cfs_spin_unlock(&imp->imp_lock);
 217         RETURN(rc);
 218 }
 219 EXPORT_SYMBOL(client_import_find_conn);
 220
 221 void client_destroy_import(struct obd_import *imp)
 222 {
 223         /* drop security policy instance after all rpc finished/aborted
 224          * to let all busy contexts be released. */
 225         class_import_get(imp);
 226         class_destroy_import(imp);
 227         sptlrpc_import_sec_put(imp);
 228         class_import_put(imp);
 229 }
 230
 231 /* configure an RPC client OBD device
 232  *
 233  * lcfg parameters:
 234  * 1 - client UUID
 235  * 2 - server UUID
 236  * 3 - inactive-on-startup
 237  */
 238 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 239 {
 240         struct client_obd *cli = &obddev->u.cli;
 241         struct obd_import *imp;
 242         struct obd_uuid server_uuid;
 243         int rq_portal, rp_portal, connect_op;
 244         char *name = obddev->obd_type->typ_name;
 245         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 246         int rc;
 247         ENTRY;
 248
 249         /* In a more perfect world, we would hang a ptlrpc_client off of
 250          * obd_type and just use the values from there. */
 251         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 252                 rq_portal = OST_REQUEST_PORTAL;
 253                 rp_portal = OSC_REPLY_PORTAL;
 254                 connect_op = OST_CONNECT;
 255                 cli->cl_sp_me = LUSTRE_SP_CLI;
 256                 cli->cl_sp_to = LUSTRE_SP_OST;
 257                 ns_type = LDLM_NS_TYPE_OSC;
 258
 259         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 260                 rq_portal = MDS_REQUEST_PORTAL;
 261                 rp_portal = MDC_REPLY_PORTAL;
 262                 connect_op = MDS_CONNECT;
 263                 cli->cl_sp_me = LUSTRE_SP_CLI;
 264                 cli->cl_sp_to = LUSTRE_SP_MDT;
 265                 ns_type = LDLM_NS_TYPE_MDC;
 266
 267         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 268                 rq_portal = MGS_REQUEST_PORTAL;
 269                 rp_portal = MGC_REPLY_PORTAL;
 270                 connect_op = MGS_CONNECT;
 271                 cli->cl_sp_me = LUSTRE_SP_MGC;
 272                 cli->cl_sp_to = LUSTRE_SP_MGS;
 273                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 274                 ns_type = LDLM_NS_TYPE_MGC;
 275
 276         } else {
 277                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 278                        name);
 279                 RETURN(-EINVAL);
 280         }
 281
 282         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 283                 CERROR("requires a TARGET UUID\n");
 284                 RETURN(-EINVAL);
 285         }
 286
 287         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 288                 CERROR("client UUID must be less than 38 characters\n");
 289                 RETURN(-EINVAL);
 290         }
 291
 292         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 293                 CERROR("setup requires a SERVER UUID\n");
 294                 RETURN(-EINVAL);
 295         }
 296
 297         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 298                 CERROR("target UUID must be less than 38 characters\n");
 299                 RETURN(-EINVAL);
 300         }
 301
 302         cfs_init_rwsem(&cli->cl_sem);
 303         cfs_sema_init(&cli->cl_mgc_sem, 1);
 304         cli->cl_conn_count = 0;
 305         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 306                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 307                      sizeof(server_uuid)));
 308
 309         cli->cl_dirty = 0;
 310         cli->cl_avail_grant = 0;
 311         /* FIXME: should limit this for the sum of all cl_dirty_max */
 312         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 313         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
 314                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
 315         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 316         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 317         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 318         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 319         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 320         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 321         cli->cl_r_in_flight = 0;
 322         cli->cl_w_in_flight = 0;
 323
 324         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 325         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 326         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 327         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 328         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 329         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 330         cfs_waitq_init(&cli->cl_destroy_waitq);
 331         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 332 #ifdef ENABLE_CHECKSUM
 333         /* Turn on checksumming by default. */
 334         cli->cl_checksum = 1;
 335         /*
 336          * The supported checksum types will be worked out at connect time
 337          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 338          * through procfs.
 339          */
 340         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 341 #endif
 342         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 343
 344         /* This value may be changed at connect time in
 345            ptlrpc_connect_interpret. */
 346         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 347                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 348
 349         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 350                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 351         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 352                 cli->cl_max_rpcs_in_flight = 2;
 353         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 354                 cli->cl_max_rpcs_in_flight = 3;
 355         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 356                 cli->cl_max_rpcs_in_flight = 4;
 357         } else {
 358                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 359         }
 360
 361         rc = ldlm_get_ref();
 362         if (rc) {
 363                 CERROR("ldlm_get_ref failed: %d\n", rc);
 364                 GOTO(err, rc);
 365         }
 366
 367         ptlrpc_init_client(rq_portal, rp_portal, name,
 368                            &obddev->obd_ldlm_client);
 369
 370         imp = class_new_import(obddev);
 371         if (imp == NULL)
 372                 GOTO(err_ldlm, rc = -ENOENT);
 373         imp->imp_client = &obddev->obd_ldlm_client;
 374         imp->imp_connect_op = connect_op;
 375         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 376         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 377                LUSTRE_CFG_BUFLEN(lcfg, 1));
 378         class_import_put(imp);
 379
 380         rc = client_import_add_conn(imp, &server_uuid, 1);
 381         if (rc) {
 382                 CERROR("can't add initial connection\n");
 383                 GOTO(err_import, rc);
 384         }
 385
 386         cli->cl_import = imp;
 387         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 388         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 389         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 390
 391         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 392                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 393                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 394                                name, obddev->obd_name,
 395                                cli->cl_target_uuid.uuid);
 396                         cfs_spin_lock(&imp->imp_lock);
 397                         imp->imp_deactive = 1;
 398                         cfs_spin_unlock(&imp->imp_lock);
 399                 }
 400         }
 401
 402         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 403                                                    LDLM_NAMESPACE_CLIENT,
 404                                                    LDLM_NAMESPACE_GREEDY,
 405                                                    ns_type);
 406         if (obddev->obd_namespace == NULL) {
 407                 CERROR("Unable to create client namespace - %s\n",
 408                        obddev->obd_name);
 409                 GOTO(err_import, rc = -ENOMEM);
 410         }
 411
 412         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 413
 414         RETURN(rc);
 415
 416 err_import:
 417         class_destroy_import(imp);
 418 err_ldlm:
 419         ldlm_put_ref();
 420 err:
 421         RETURN(rc);
 422
 423 }
 424
 425 int client_obd_cleanup(struct obd_device *obddev)
 426 {
 427         ENTRY;
 428
 429         ldlm_namespace_free_post(obddev->obd_namespace);
 430         obddev->obd_namespace = NULL;
 431
 432         ldlm_put_ref();
 433         RETURN(0);
 434 }
 435
 436 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 437 int client_connect_import(const struct lu_env *env,
 438                           struct obd_export **exp,
 439                           struct obd_device *obd, struct obd_uuid *cluuid,
 440                           struct obd_connect_data *data, void *localdata)
 441 {
 442         struct client_obd *cli = &obd->u.cli;
 443         struct obd_import *imp = cli->cl_import;
 444         struct obd_connect_data *ocd;
 445         struct lustre_handle conn = { 0 };
 446         int rc;
 447         ENTRY;
 448
 449         *exp = NULL;
 450         cfs_down_write(&cli->cl_sem);
 451         if (cli->cl_conn_count > 0 )
 452                 GOTO(out_sem, rc = -EALREADY);
 453
 454         rc = class_connect(&conn, obd, cluuid);
 455         if (rc)
 456                 GOTO(out_sem, rc);
 457
 458         cli->cl_conn_count++;
 459         *exp = class_conn2export(&conn);
 460
 461         LASSERT(obd->obd_namespace);
 462
 463         imp->imp_dlm_handle = conn;
 464         rc = ptlrpc_init_import(imp);
 465         if (rc != 0)
 466                 GOTO(out_ldlm, rc);
 467
 468         ocd = &imp->imp_connect_data;
 469         if (data) {
 470                 *ocd = *data;
 471                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 472         }
 473
 474         rc = ptlrpc_connect_import(imp);
 475         if (rc != 0) {
 476                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 477                 GOTO(out_ldlm, rc);
 478         }
 479         LASSERT((*exp)->exp_connection);
 480
 481         if (data) {
 482                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 483                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 484                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 485                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 486         }
 487
 488         ptlrpc_pinger_add_import(imp);
 489
 490         EXIT;
 491
 492         if (rc) {
 493 out_ldlm:
 494                 cli->cl_conn_count--;
 495                 class_disconnect(*exp);
 496                 *exp = NULL;
 497         }
 498 out_sem:
 499         cfs_up_write(&cli->cl_sem);
 500
 501         return rc;
 502 }
 503
 504 int client_disconnect_export(struct obd_export *exp)
 505 {
 506         struct obd_device *obd = class_exp2obd(exp);
 507         struct client_obd *cli;
 508         struct obd_import *imp;
 509         int rc = 0, err;
 510         ENTRY;
 511
 512         if (!obd) {
 513                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 514                        exp, exp ? exp->exp_handle.h_cookie : -1);
 515                 RETURN(-EINVAL);
 516         }
 517
 518         cli = &obd->u.cli;
 519         imp = cli->cl_import;
 520
 521         cfs_down_write(&cli->cl_sem);
 522         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 523                cli->cl_conn_count);
 524
 525         if (!cli->cl_conn_count) {
 526                 CERROR("disconnecting disconnected device (%s)\n",
 527                        obd->obd_name);
 528                 GOTO(out_disconnect, rc = -EINVAL);
 529         }
 530
 531         cli->cl_conn_count--;
 532         if (cli->cl_conn_count)
 533                 GOTO(out_disconnect, rc = 0);
 534
 535         /* Mark import deactivated now, so we don't try to reconnect if any
 536          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 537          * fully deactivate the import, or that would drop all requests. */
 538         cfs_spin_lock(&imp->imp_lock);
 539         imp->imp_deactive = 1;
 540         cfs_spin_unlock(&imp->imp_lock);
 541
 542         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 543          * delete it regardless.  (It's safe to delete an import that was
 544          * never added.) */
 545         (void)ptlrpc_pinger_del_import(imp);
 546
 547         if (obd->obd_namespace != NULL) {
 548                 /* obd_force == local only */
 549                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 550                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 551                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 552         }
 553
 554         /*
 555          * there's no need to hold sem during disconnecting an import,
 556          * and actually it may cause deadlock in gss.
 557          */
 558         cfs_up_write(&cli->cl_sem);
 559         rc = ptlrpc_disconnect_import(imp, 0);
 560         cfs_down_write(&cli->cl_sem);
 561
 562         ptlrpc_invalidate_import(imp);
 563
 564         if (imp->imp_rq_pool) {
 565                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 566                 imp->imp_rq_pool = NULL;
 567         }
 568         client_destroy_import(imp);
 569         cli->cl_import = NULL;
 570
 571         EXIT;
 572
 573  out_disconnect:
 574         /* use server style - class_disconnect should be always called for
 575          * o_disconnect */
 576         err = class_disconnect(exp);
 577         if (!rc && err)
 578                 rc = err;
 579
 580         cfs_up_write(&cli->cl_sem);
 581
 582         RETURN(rc);
 583 }
 584
 585 int server_disconnect_export(struct obd_export *exp)
 586 {
 587         int rc;
 588         ENTRY;
 589
 590         /* Disconnect early so that clients can't keep using export */
 591         rc = class_disconnect(exp);
 592         /* close import for avoid sending any requests */
 593         if (exp->exp_imp_reverse)
 594                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 595
 596         if (exp->exp_obd->obd_namespace != NULL)
 597                 ldlm_cancel_locks_for_export(exp);
 598
 599         /* complete all outstanding replies */
 600         cfs_spin_lock(&exp->exp_lock);
 601         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 602                 struct ptlrpc_reply_state *rs =
 603                         cfs_list_entry(exp->exp_outstanding_replies.next,
 604                                        struct ptlrpc_reply_state, rs_exp_list);
 605                 struct ptlrpc_service *svc = rs->rs_service;
 606
 607                 cfs_spin_lock(&svc->srv_rs_lock);
 608                 cfs_list_del_init(&rs->rs_exp_list);
 609                 cfs_spin_lock(&rs->rs_lock);
 610                 ptlrpc_schedule_difficult_reply(rs);
 611                 cfs_spin_unlock(&rs->rs_lock);
 612                 cfs_spin_unlock(&svc->srv_rs_lock);
 613         }
 614         cfs_spin_unlock(&exp->exp_lock);
 615
 616         RETURN(rc);
 617 }
 618
 619 /* --------------------------------------------------------------------------
 620  * from old lib/target.c
 621  * -------------------------------------------------------------------------- */
 622
 623 static int target_handle_reconnect(struct lustre_handle *conn,
 624                                    struct obd_export *exp,
 625                                    struct obd_uuid *cluuid)
 626 {
 627         ENTRY;
 628         if (exp->exp_connection && exp->exp_imp_reverse) {
 629                 struct lustre_handle *hdl;
 630                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 631                 /* Might be a re-connect after a partition. */
 632                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 633                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 634                               cluuid->uuid);
 635                         conn->cookie = exp->exp_handle.h_cookie;
 636                         /* target_handle_connect() treats EALREADY and
 637                          * -EALREADY differently.  EALREADY means we are
 638                          * doing a valid reconnect from the same client. */
 639                         RETURN(EALREADY);
 640                 } else {
 641                         CERROR("%s reconnecting from %s, "
 642                                "handle mismatch (ours "LPX64", theirs "
 643                                LPX64")\n", cluuid->uuid,
 644                                exp->exp_connection->c_remote_uuid.uuid,
 645                                hdl->cookie, conn->cookie);
 646                         memset(conn, 0, sizeof *conn);
 647                         /* target_handle_connect() treats EALREADY and
 648                          * -EALREADY differently.  -EALREADY is an error
 649                          * (same UUID, different handle). */
 650                         RETURN(-EALREADY);
 651                 }
 652         }
 653
 654         conn->cookie = exp->exp_handle.h_cookie;
 655         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 656                cluuid->uuid, exp, conn->cookie);
 657         RETURN(0);
 658 }
 659
 660 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 661                           int error)
 662 {
 663         struct obd_export *exp = cb_data;
 664
 665         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 666                obd->obd_name, exp->exp_client_uuid.uuid);
 667
 668         cfs_spin_lock(&exp->exp_lock);
 669         exp->exp_need_sync = 0;
 670         cfs_spin_unlock(&exp->exp_lock);
 671         class_export_cb_put(exp);
 672 }
 673 EXPORT_SYMBOL(target_client_add_cb);
 674
 675 #ifdef __KERNEL__
 676 static void
 677 check_and_extend_recovery_timer(struct obd_device *obd,
 678                                 struct ptlrpc_request *req);
 679 #else
 680 static inline void
 681 check_and_extend_recovery_timer(struct obd_device *obd,
 682                                 struct ptlrpc_request *req)
 683 {
 684 }
 685 #endif
 686
 687 int target_handle_connect(struct ptlrpc_request *req)
 688 {
 689         struct obd_device *target, *targref = NULL;
 690         struct obd_export *export = NULL;
 691         struct obd_import *revimp;
 692         struct lustre_handle conn;
 693         struct lustre_handle *tmp;
 694         struct obd_uuid tgtuuid;
 695         struct obd_uuid cluuid;
 696         struct obd_uuid remote_uuid;
 697         char *str;
 698         int rc = 0;
 699         int mds_conn = 0;
 700         struct obd_connect_data *data, *tmpdata;
 701         int size, tmpsize;
 702         lnet_nid_t *client_nid = NULL;
 703         ENTRY;
 704
 705         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 706
 707         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 708         if (str == NULL) {
 709                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 710                 GOTO(out, rc = -EINVAL);
 711         }
 712
 713         obd_str2uuid(&tgtuuid, str);
 714         target = class_uuid2obd(&tgtuuid);
 715         if (!target)
 716                 target = class_name2obd(str);
 717
 718         if (!target || target->obd_stopping || !target->obd_set_up) {
 719                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 720                                    "for connect (%s)\n", str,
 721                                    !target ? "no target" :
 722                                    (target->obd_stopping ? "stopping" :
 723                                    "not set up"));
 724                 GOTO(out, rc = -ENODEV);
 725         }
 726
 727         if (target->obd_no_conn) {
 728                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 729                               "from %s\n", target->obd_name,
 730                               libcfs_nid2str(req->rq_peer.nid));
 731                 GOTO(out, rc = -EAGAIN);
 732         }
 733
 734         /* Make sure the target isn't cleaned up while we're here. Yes,
 735            there's still a race between the above check and our incref here.
 736            Really, class_uuid2obd should take the ref. */
 737         targref = class_incref(target, __FUNCTION__, cfs_current());
 738
 739
 740         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 741         if (str == NULL) {
 742                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 743                 GOTO(out, rc = -EINVAL);
 744         }
 745
 746         obd_str2uuid(&cluuid, str);
 747
 748         /* XXX extract a nettype and format accordingly */
 749         switch (sizeof(lnet_nid_t)) {
 750                 /* NB the casts only avoid compiler warnings */
 751         case 8:
 752                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 753                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 754                 break;
 755         case 4:
 756                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 757                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 758                 break;
 759         default:
 760                 LBUG();
 761         }
 762
 763         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 764         if (tmp == NULL)
 765                 GOTO(out, rc = -EPROTO);
 766
 767         conn = *tmp;
 768
 769         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 770                                     RCL_CLIENT);
 771         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 772         if (!data)
 773                 GOTO(out, rc = -EPROTO);
 774
 775         rc = req_capsule_server_pack(&req->rq_pill);
 776         if (rc)
 777                 GOTO(out, rc);
 778
 779         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 780                 if (!data) {
 781                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 782                                   "libclient connection attempt");
 783                         GOTO(out, rc = -EPROTO);
 784                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 785                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 786                            data->ocd_version > LUSTRE_VERSION_CODE +
 787                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 788                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 789                                   "libclient connection attempt",
 790                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 791                                   "old" : "new",
 792                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 793                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 794                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 795                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 796                         data = req_capsule_server_sized_get(&req->rq_pill,
 797                                                         &RMF_CONNECT_DATA,
 798                                     offsetof(typeof(*data), ocd_version) +
 799                                              sizeof(data->ocd_version));
 800                         if (data) {
 801                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 802                                 data->ocd_version = LUSTRE_VERSION_CODE;
 803                         }
 804                         GOTO(out, rc = -EPROTO);
 805                 }
 806         }
 807
 808         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 809             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 810                 mds_conn = 1;
 811
 812         /* lctl gets a backstage, all-access pass. */
 813         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 814                 goto dont_check_exports;
 815
 816         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 817         if (!export)
 818                 goto no_export;
 819
 820         /* we've found an export in the hash */
 821         if (export->exp_connecting) { /* bug 9635, et. al. */
 822                 CWARN("%s: exp %p already connecting\n",
 823                       export->exp_obd->obd_name, export);
 824                 class_export_put(export);
 825                 export = NULL;
 826                 rc = -EALREADY;
 827         } else if (mds_conn && export->exp_connection) {
 828                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 829                         /* mds reconnected after failover */
 830                         CWARN("%s: received MDS connection from NID %s,"
 831                               " removing former export from NID %s\n",
 832                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 833                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 834                 else
 835                         /* new mds connection from the same nid */
 836                         CWARN("%s: received new MDS connection from NID %s,"
 837                               " removing former export from same NID\n",
 838                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 839                 class_fail_export(export);
 840                 class_export_put(export);
 841                 export = NULL;
 842                 rc = 0;
 843         } else if (export->exp_connection != NULL &&
 844                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 845                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 846                     MSG_CONNECT_INITIAL)) {
 847                 /* in mds failover we have static uuid but nid can be
 848                  * changed*/
 849                 CWARN("%s: cookie %s seen on new NID %s when "
 850                       "existing NID %s is already connected\n",
 851                       target->obd_name, cluuid.uuid,
 852                       libcfs_nid2str(req->rq_peer.nid),
 853                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 854                 rc = -EALREADY;
 855                 class_export_put(export);
 856                 export = NULL;
 857         } else {
 858                 cfs_spin_lock(&export->exp_lock);
 859                 export->exp_connecting = 1;
 860                 cfs_spin_unlock(&export->exp_lock);
 861                 class_export_put(export);
 862                 LASSERT(export->exp_obd == target);
 863
 864                 rc = target_handle_reconnect(&conn, export, &cluuid);
 865         }
 866
 867         /* If we found an export, we already unlocked. */
 868         if (!export) {
 869 no_export:
 870                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 871         } else if (req->rq_export == NULL &&
 872                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 873                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 874                       target->obd_name, cluuid.uuid,
 875                       libcfs_nid2str(req->rq_peer.nid),
 876                       export, cfs_atomic_read(&export->exp_refcount));
 877                 GOTO(out, rc = -EBUSY);
 878         } else if (req->rq_export != NULL &&
 879                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 880                 /* the current connect rpc has increased exp_rpc_count */
 881                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 882                       target->obd_name, cluuid.uuid,
 883                       libcfs_nid2str(req->rq_peer.nid),
 884                       export, cfs_atomic_read(&export->exp_rpc_count) - 1);
 885                 cfs_spin_lock(&export->exp_lock);
 886                 if (req->rq_export->exp_conn_cnt <
 887                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 888                         /* try to abort active requests */
 889                         req->rq_export->exp_abort_active_req = 1;
 890                 cfs_spin_unlock(&export->exp_lock);
 891                 GOTO(out, rc = -EBUSY);
 892         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 893                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 894                        "cookies not random?\n", target->obd_name,
 895                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 896                 GOTO(out, rc = -EALREADY);
 897         } else {
 898                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 899         }
 900
 901         if (rc < 0) {
 902                 GOTO(out, rc);
 903         }
 904
 905         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 906                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 907               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 908               export, (long)cfs_time_current_sec(),
 909               export ? (long)export->exp_last_request_time : 0);
 910
 911         /* If this is the first time a client connects, reset the recovery
 912          * timer */
 913         if (rc == 0 && target->obd_recovering && export)
 914                 check_and_extend_recovery_timer(target, req);
 915
 916         /* We want to handle EALREADY but *not* -EALREADY from
 917          * target_handle_reconnect(), return reconnection state in a flag */
 918         if (rc == EALREADY) {
 919                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 920                 rc = 0;
 921         } else {
 922                 LASSERT(rc == 0);
 923         }
 924
 925         /* Tell the client if we support replayable requests */
 926         if (target->obd_replayable)
 927                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 928         client_nid = &req->rq_peer.nid;
 929
 930         if (export == NULL) {
 931                 if (target->obd_recovering) {
 932                         cfs_time_t t;
 933
 934                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 935                         t = cfs_time_sub(t, cfs_time_current());
 936                         CERROR("%s: denying connection for new client %s (%s): "
 937                                "%d clients in recovery for "CFS_TIME_T"s\n",
 938                                target->obd_name,
 939                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 940                                cfs_atomic_read(&target-> \
 941                                                obd_lock_replay_clients),
 942                                cfs_duration_sec(t));
 943                         rc = -EBUSY;
 944                 } else {
 945 dont_check_exports:
 946                         rc = obd_connect(req->rq_svc_thread->t_env,
 947                                          &export, target, &cluuid, data,
 948                                          client_nid);
 949                         if (rc == 0)
 950                                 conn.cookie = export->exp_handle.h_cookie;
 951                 }
 952         } else {
 953                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 954                                    export, target, &cluuid, data, client_nid);
 955                 if (rc == 0)
 956                         /* prevous done via class_conn2export */
 957                         class_export_get(export);
 958         }
 959         if (rc)
 960                 GOTO(out, rc);
 961
 962         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
 963         data->ocd_instance = target->u.obt.obt_instance;
 964
 965         /* Return only the parts of obd_connect_data that we understand, so the
 966          * client knows that we don't understand the rest. */
 967         if (data) {
 968                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 969                                                RCL_SERVER);
 970                 tmpdata = req_capsule_server_get(&req->rq_pill,
 971                                                  &RMF_CONNECT_DATA);
 972                 /* Don't use struct assignment here, because the client reply
 973                  * buffer may be smaller/larger than the local struct
 974                  * obd_connect_data. */
 975                 memcpy(tmpdata, data, min(tmpsize, size));
 976         }
 977
 978         /* If all else goes well, this is our RPC return code. */
 979         req->rq_status = 0;
 980
 981         lustre_msg_set_handle(req->rq_repmsg, &conn);
 982
 983         /* If the client and the server are the same node, we will already
 984          * have an export that really points to the client's DLM export,
 985          * because we have a shared handles table.
 986          *
 987          * XXX this will go away when shaver stops sending the "connect" handle
 988          * in the real "remote handle" field of the request --phik 24 Apr 2003
 989          */
 990         if (req->rq_export != NULL)
 991                 class_export_put(req->rq_export);
 992
 993         req->rq_export = export;
 994
 995         cfs_spin_lock(&export->exp_lock);
 996         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 997                 cfs_spin_unlock(&export->exp_lock);
 998                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 999                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1000                        export->exp_conn_cnt,
1001                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1002
1003                 GOTO(out, rc = -EALREADY);
1004         }
1005         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1006         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1007         export->exp_abort_active_req = 0;
1008
1009         /* request from liblustre?  Don't evict it for not pinging. */
1010         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1011                 export->exp_libclient = 1;
1012                 cfs_spin_unlock(&export->exp_lock);
1013
1014                 cfs_spin_lock(&target->obd_dev_lock);
1015                 cfs_list_del_init(&export->exp_obd_chain_timed);
1016                 cfs_spin_unlock(&target->obd_dev_lock);
1017         } else {
1018                 cfs_spin_unlock(&export->exp_lock);
1019         }
1020
1021         if (export->exp_connection != NULL) {
1022                 /* Check to see if connection came from another NID */
1023                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1024                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1025                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1026                                      &export->exp_connection->c_peer.nid,
1027                                      &export->exp_nid_hash);
1028
1029                 ptlrpc_connection_put(export->exp_connection);
1030         }
1031
1032         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1033                                                        req->rq_self,
1034                                                        &remote_uuid);
1035         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1036                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1037                              &export->exp_connection->c_peer.nid,
1038                              &export->exp_nid_hash);
1039         }
1040
1041         if (target->obd_recovering && !export->exp_in_recovery) {
1042                 int has_transno;
1043                 __u64 transno = data->ocd_transno;
1044
1045                 cfs_spin_lock(&export->exp_lock);
1046                 export->exp_in_recovery = 1;
1047                 export->exp_req_replay_needed = 1;
1048                 export->exp_lock_replay_needed = 1;
1049                 cfs_spin_unlock(&export->exp_lock);
1050
1051                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1052                                  MSG_CONNECT_TRANSNO);
1053                 if (has_transno && transno == 0)
1054                         CWARN("Connect with zero transno!\n");
1055
1056                 if (has_transno && transno > 0 &&
1057                     transno < target->obd_next_recovery_transno &&
1058                     transno > target->obd_last_committed) {
1059                         /* another way is to use cmpxchg() so it will be
1060                          * lock free */
1061                         cfs_spin_lock(&target->obd_recovery_task_lock);
1062                         if (transno < target->obd_next_recovery_transno)
1063                                 target->obd_next_recovery_transno = transno;
1064                         cfs_spin_unlock(&target->obd_recovery_task_lock);
1065                 }
1066
1067                 cfs_atomic_inc(&target->obd_req_replay_clients);
1068                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1069                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1070                     target->obd_max_recoverable_clients)
1071                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1072         }
1073
1074         /* Tell the client we're in recovery, when client is involved in it. */
1075         if (target->obd_recovering)
1076                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1077
1078         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1079         conn = *tmp;
1080
1081         if (export->exp_imp_reverse != NULL) {
1082                 /* destroyed import can be still referenced in ctxt */
1083                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1084                                    KEY_REVIMP_UPD, 0, NULL, NULL);
1085
1086                 client_destroy_import(export->exp_imp_reverse);
1087         }
1088
1089         /* for the rest part, we return -ENOTCONN in case of errors
1090          * in order to let client initialize connection again.
1091          */
1092         revimp = export->exp_imp_reverse = class_new_import(target);
1093         if (!revimp) {
1094                 CERROR("fail to alloc new reverse import.\n");
1095                 GOTO(out, rc = -ENOTCONN);
1096         }
1097
1098         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1099         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1100         revimp->imp_remote_handle = conn;
1101         revimp->imp_dlm_fake = 1;
1102         revimp->imp_state = LUSTRE_IMP_FULL;
1103
1104         /* unknown versions will be caught in
1105          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1106         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1107
1108         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1109             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1110                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1111         else
1112                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1113
1114         if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1115             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1116                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1117         else
1118                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1119
1120         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1121         if (rc) {
1122                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1123                 export->exp_imp_reverse = NULL;
1124                 class_destroy_import(revimp);
1125         }
1126
1127         class_import_put(revimp);
1128 out:
1129         if (export) {
1130                 cfs_spin_lock(&export->exp_lock);
1131                 export->exp_connecting = 0;
1132                 cfs_spin_unlock(&export->exp_lock);
1133         }
1134         if (targref)
1135                 class_decref(targref, __FUNCTION__, cfs_current());
1136         if (rc)
1137                 req->rq_status = rc;
1138         RETURN(rc);
1139 }
1140
1141 int target_handle_disconnect(struct ptlrpc_request *req)
1142 {
1143         int rc;
1144         ENTRY;
1145
1146         rc = req_capsule_server_pack(&req->rq_pill);
1147         if (rc)
1148                 RETURN(rc);
1149
1150         /* keep the rq_export around so we can send the reply */
1151         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1152
1153         RETURN(0);
1154 }
1155
1156 void target_destroy_export(struct obd_export *exp)
1157 {
1158         /* exports created from last_rcvd data, and "fake"
1159            exports created by lctl don't have an import */
1160         if (exp->exp_imp_reverse != NULL)
1161                 client_destroy_import(exp->exp_imp_reverse);
1162
1163         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1164         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1165         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1166         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1167 }
1168
1169 /*
1170  * Recovery functions
1171  */
1172 static void target_request_copy_get(struct ptlrpc_request *req)
1173 {
1174         class_export_rpc_get(req->rq_export);
1175         LASSERT(cfs_list_empty(&req->rq_list));
1176         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1177
1178         /* increase refcount to keep request in queue */
1179         cfs_atomic_inc(&req->rq_refcount);
1180         /** let export know it has replays to be handled */
1181         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1182 }
1183
1184 static void target_request_copy_put(struct ptlrpc_request *req)
1185 {
1186         LASSERT(cfs_list_empty(&req->rq_replay_list));
1187         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1188
1189         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1190         class_export_rpc_put(req->rq_export);
1191         ptlrpc_server_drop_request(req);
1192 }
1193
1194 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1195 {
1196         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1197         struct obd_export     *exp = req->rq_export;
1198         struct ptlrpc_request *reqiter;
1199         int                    dup = 0;
1200
1201         LASSERT(exp);
1202
1203         cfs_spin_lock(&exp->exp_lock);
1204         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1205                                 rq_replay_list) {
1206                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1207                         dup = 1;
1208                         break;
1209                 }
1210         }
1211
1212         if (dup) {
1213                 /* we expect it with RESENT and REPLAY flags */
1214                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1215                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1216                         CERROR("invalid flags %x of resent replay\n",
1217                                lustre_msg_get_flags(req->rq_reqmsg));
1218         } else {
1219                 cfs_list_add_tail(&req->rq_replay_list,
1220                                   &exp->exp_req_replay_queue);
1221         }
1222
1223         cfs_spin_unlock(&exp->exp_lock);
1224         return dup;
1225 }
1226
1227 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1228 {
1229         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1230         LASSERT(req->rq_export);
1231
1232         cfs_spin_lock(&req->rq_export->exp_lock);
1233         cfs_list_del_init(&req->rq_replay_list);
1234         cfs_spin_unlock(&req->rq_export->exp_lock);
1235 }
1236
1237 #ifdef __KERNEL__
1238 static void target_finish_recovery(struct obd_device *obd)
1239 {
1240         time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1241                                     obd->obd_recovery_start);
1242         ENTRY;
1243
1244         LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1245                       "%d recovered and %d %s evicted.\n", obd->obd_name,
1246                       (int)elapsed_time / 60, (int)elapsed_time % 60,
1247                       obd->obd_max_recoverable_clients,
1248                       cfs_atomic_read(&obd->obd_connected_clients),
1249                       obd->obd_stale_clients,
1250                       obd->obd_stale_clients == 1 ? "was" : "were");
1251
1252         ldlm_reprocess_all_ns(obd->obd_namespace);
1253         cfs_spin_lock(&obd->obd_recovery_task_lock);
1254         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1255             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1256             !cfs_list_empty(&obd->obd_final_req_queue)) {
1257                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1258                        obd->obd_name,
1259                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1260                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1261                                "" : "lock ",
1262                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1263                                "" : "final ");
1264                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1265                 LBUG();
1266         }
1267         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1268
1269         obd->obd_recovery_end = cfs_time_current_sec();
1270
1271         /* when recovery finished, cleanup orphans on mds and ost */
1272         if (OBT(obd) && OBP(obd, postrecov)) {
1273                 int rc = OBP(obd, postrecov)(obd);
1274                 if (rc < 0)
1275                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1276                                       obd->obd_name, rc);
1277         }
1278         EXIT;
1279 }
1280
1281 static void abort_req_replay_queue(struct obd_device *obd)
1282 {
1283         struct ptlrpc_request *req, *n;
1284         cfs_list_t abort_list;
1285
1286         CFS_INIT_LIST_HEAD(&abort_list);
1287         cfs_spin_lock(&obd->obd_recovery_task_lock);
1288         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1289         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1290         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1291                 DEBUG_REQ(D_WARNING, req, "aborted:");
1292                 req->rq_status = -ENOTCONN;
1293                 if (ptlrpc_error(req)) {
1294                         DEBUG_REQ(D_ERROR, req,
1295                                   "failed abort_req_reply; skipping");
1296                 }
1297                 target_exp_dequeue_req_replay(req);
1298                 target_request_copy_put(req);
1299         }
1300 }
1301
1302 static void abort_lock_replay_queue(struct obd_device *obd)
1303 {
1304         struct ptlrpc_request *req, *n;
1305         cfs_list_t abort_list;
1306
1307         CFS_INIT_LIST_HEAD(&abort_list);
1308         cfs_spin_lock(&obd->obd_recovery_task_lock);
1309         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1310         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1311         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1312                 DEBUG_REQ(D_ERROR, req, "aborted:");
1313                 req->rq_status = -ENOTCONN;
1314                 if (ptlrpc_error(req)) {
1315                         DEBUG_REQ(D_ERROR, req,
1316                                   "failed abort_lock_reply; skipping");
1317                 }
1318                 target_request_copy_put(req);
1319         }
1320 }
1321
1322 /* Called from a cleanup function if the device is being cleaned up
1323    forcefully.  The exports should all have been disconnected already,
1324    the only thing left to do is
1325      - clear the recovery flags
1326      - cancel the timer
1327      - free queued requests and replies, but don't send replies
1328    Because the obd_stopping flag is set, no new requests should be received.
1329
1330 */
1331 void target_cleanup_recovery(struct obd_device *obd)
1332 {
1333         struct ptlrpc_request *req, *n;
1334         cfs_list_t clean_list;
1335         ENTRY;
1336
1337         CFS_INIT_LIST_HEAD(&clean_list);
1338         cfs_spin_lock(&obd->obd_dev_lock);
1339         if (!obd->obd_recovering) {
1340                 cfs_spin_unlock(&obd->obd_dev_lock);
1341                 EXIT;
1342                 return;
1343         }
1344         obd->obd_recovering = obd->obd_abort_recovery = 0;
1345         cfs_spin_unlock(&obd->obd_dev_lock);
1346
1347         cfs_spin_lock(&obd->obd_recovery_task_lock);
1348         target_cancel_recovery_timer(obd);
1349         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1350         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1351
1352         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1353                 LASSERT(req->rq_reply_state == 0);
1354                 target_exp_dequeue_req_replay(req);
1355                 target_request_copy_put(req);
1356         }
1357
1358         cfs_spin_lock(&obd->obd_recovery_task_lock);
1359         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1360         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1361         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1362
1363         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1364                 LASSERT(req->rq_reply_state == 0);
1365                 target_request_copy_put(req);
1366         }
1367
1368         EXIT;
1369 }
1370
1371 /* obd_recovery_task_lock should be held */
1372 void target_cancel_recovery_timer(struct obd_device *obd)
1373 {
1374         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1375         cfs_timer_disarm(&obd->obd_recovery_timer);
1376 }
1377
1378 void target_start_recovery_timer(struct obd_device *obd)
1379 {
1380         cfs_spin_lock(&obd->obd_dev_lock);
1381         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1382                 cfs_spin_unlock(&obd->obd_dev_lock);
1383                 return;
1384         }
1385
1386         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1387                 cfs_spin_unlock(&obd->obd_dev_lock);
1388                 return;
1389         }
1390
1391         cfs_timer_arm(&obd->obd_recovery_timer,
1392                       cfs_time_shift(obd->obd_recovery_timeout));
1393         obd->obd_recovery_start = cfs_time_current_sec();
1394         cfs_spin_unlock(&obd->obd_dev_lock);
1395         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1396 }
1397 EXPORT_SYMBOL(target_start_recovery_timer);
1398
1399 /* extend recovery window to have extra @duration seconds at least. */
1400 static void extend_recovery_timer(struct obd_device *obd, int drt)
1401 {
1402         cfs_time_t now = cfs_time_current_sec();
1403         cfs_duration_t left;
1404
1405         if (!cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1406                 cfs_spin_lock(&obd->obd_dev_lock);
1407                 if (obd->obd_recovery_timeout < drt)
1408                         obd->obd_recovery_timeout = drt;
1409                 cfs_spin_unlock(&obd->obd_dev_lock);
1410                 return;
1411         }
1412
1413         left = obd->obd_recovery_timeout;
1414         left -= cfs_time_sub(now, obd->obd_recovery_start);
1415         if (drt > left) {
1416                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(drt));
1417                 CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1418                        obd->obd_name, (unsigned)drt);
1419         }
1420 }
1421
1422 /* Reset the timer with each new client connection */
1423 /*
1424  * This timer is actually reconnect_timer, which is for making sure
1425  * the total recovery window is at least as big as my reconnect
1426  * attempt timing. So the initial recovery time_out will be set to
1427  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1428  * from client is bigger than this, then the recovery time_out will
1429  * be extended to make sure the client could be reconnected, in the
1430  * process, the timeout from the new client should be ignored.
1431  */
1432
1433 static void
1434 check_and_extend_recovery_timer(struct obd_device *obd,
1435                                 struct ptlrpc_request *req)
1436 {
1437         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1438         struct obd_device_target *obt = &obd->u.obt;
1439         struct lustre_sb_info *lsi;
1440
1441         if (service_time)
1442                 /* Teach server about old server's estimates, as first guess
1443                  * at how long new requests will take. */
1444                 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1445                             service_time);
1446
1447         /* convert the service time to rpc timeout,
1448          * reuse service_time to limit stack usage */
1449         service_time = at_est2timeout(service_time);
1450
1451         /* We expect other clients to timeout within service_time, then try
1452          * to reconnect, then try the failover server.  The max delay between
1453          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1454         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1455
1456         LASSERT(obt->obt_magic == OBT_MAGIC);
1457         lsi = s2lsi(obt->obt_sb);
1458         if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1459                 service_time += 2 * (CONNECTION_SWITCH_MAX +
1460                                      CONNECTION_SWITCH_INC);
1461         service_time -= obd->obd_recovery_timeout;
1462         if (service_time > 0)
1463                 extend_recovery_timer(obd, service_time);
1464 }
1465
1466 /** Health checking routines */
1467 static inline int exp_connect_healthy(struct obd_export *exp)
1468 {
1469         return (exp->exp_in_recovery);
1470 }
1471
1472 /** if export done req_replay or has replay in queue */
1473 static inline int exp_req_replay_healthy(struct obd_export *exp)
1474 {
1475         return (!exp->exp_req_replay_needed ||
1476                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1477 }
1478 /** if export done lock_replay or has replay in queue */
1479 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1480 {
1481         return (!exp->exp_lock_replay_needed ||
1482                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1483 }
1484
1485 static inline int exp_vbr_healthy(struct obd_export *exp)
1486 {
1487         return (!exp->exp_vbr_failed);
1488 }
1489
1490 static inline int exp_finished(struct obd_export *exp)
1491 {
1492         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1493 }
1494
1495 /** Checking routines for recovery */
1496 static int check_for_clients(struct obd_device *obd)
1497 {
1498         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1499
1500         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1501                 return 1;
1502         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1503         if (obd->obd_no_conn == 0 &&
1504             clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1505                 return 1;
1506         return 0;
1507 }
1508
1509 static int check_for_next_transno(struct obd_device *obd)
1510 {
1511         struct ptlrpc_request *req = NULL;
1512         int wake_up = 0, connected, completed, queue_len;
1513         __u64 next_transno, req_transno;
1514         ENTRY;
1515
1516         cfs_spin_lock(&obd->obd_recovery_task_lock);
1517         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1518                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1519                                      struct ptlrpc_request, rq_list);
1520                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1521         } else {
1522                 req_transno = 0;
1523         }
1524
1525         connected = cfs_atomic_read(&obd->obd_connected_clients);
1526         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1527         queue_len = obd->obd_requests_queued_for_recovery;
1528         next_transno = obd->obd_next_recovery_transno;
1529
1530         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1531                "req_transno: "LPU64", next_transno: "LPU64"\n",
1532                obd->obd_max_recoverable_clients, connected, completed,
1533                queue_len, req_transno, next_transno);
1534
1535         if (obd->obd_abort_recovery) {
1536                 CDEBUG(D_HA, "waking for aborted recovery\n");
1537                 wake_up = 1;
1538         } else if (obd->obd_recovery_expired) {
1539                 CDEBUG(D_HA, "waking for expired recovery\n");
1540                 wake_up = 1;
1541         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1542                 CDEBUG(D_HA, "waking for completed recovery\n");
1543                 wake_up = 1;
1544         } else if (req_transno == next_transno) {
1545                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1546                 wake_up = 1;
1547         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1548                 int d_lvl = D_HA;
1549                 /** handle gaps occured due to lost reply or VBR */
1550                 LASSERTF(req_transno >= next_transno,
1551                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1552                          req_transno, next_transno);
1553                 if (req_transno > obd->obd_last_committed &&
1554                     !obd->obd_version_recov)
1555                         d_lvl = D_ERROR;
1556                 CDEBUG(d_lvl,
1557                        "%s: waking for gap in transno, VBR is %s (skip: "
1558                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1559                        ", last_committed: "LPD64")\n",
1560                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1561                        next_transno, queue_len, completed, connected,
1562                        req_transno, obd->obd_last_committed);
1563                 obd->obd_next_recovery_transno = req_transno;
1564                 wake_up = 1;
1565         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1566                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1567                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1568                 obd->obd_next_recovery_transno = req_transno;
1569                 wake_up = 1;
1570         }
1571         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1572         return wake_up;
1573 }
1574
1575 static int check_for_next_lock(struct obd_device *obd)
1576 {
1577         int wake_up = 0;
1578
1579         cfs_spin_lock(&obd->obd_recovery_task_lock);
1580         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1581                 CDEBUG(D_HA, "waking for next lock\n");
1582                 wake_up = 1;
1583         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1584                 CDEBUG(D_HA, "waking for completed lock replay\n");
1585                 wake_up = 1;
1586         } else if (obd->obd_abort_recovery) {
1587                 CDEBUG(D_HA, "waking for aborted recovery\n");
1588                 wake_up = 1;
1589         } else if (obd->obd_recovery_expired) {
1590                 CDEBUG(D_HA, "waking for expired recovery\n");
1591                 wake_up = 1;
1592         }
1593         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1594
1595         return wake_up;
1596 }
1597
1598 /**
1599  * wait for recovery events,
1600  * check its status with help of check_routine
1601  * evict dead clients via health_check
1602  */
1603 static int target_recovery_overseer(struct obd_device *obd,
1604                                     int (*check_routine)(struct obd_device *),
1605                                     int (*health_check)(struct obd_export *))
1606 {
1607 repeat:
1608         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1609         if (obd->obd_abort_recovery) {
1610                 CWARN("recovery is aborted, evict exports in recovery\n");
1611                 /** evict exports which didn't finish recovery yet */
1612                 class_disconnect_stale_exports(obd, exp_finished);
1613                 return 1;
1614         } else if (obd->obd_recovery_expired) {
1615                 obd->obd_recovery_expired = 0;
1616                 /** If some clients died being recovered, evict them */
1617                 CDEBUG(D_WARNING,
1618                        "recovery is timed out, evict stale exports\n");
1619                 /** evict cexports with no replay in queue, they are stalled */
1620                 class_disconnect_stale_exports(obd, health_check);
1621                 /** continue with VBR */
1622                 cfs_spin_lock(&obd->obd_dev_lock);
1623                 obd->obd_version_recov = 1;
1624                 cfs_spin_unlock(&obd->obd_dev_lock);
1625                 /**
1626                  * reset timer, recovery will proceed with versions now,
1627                  * timeout is set just to handle reconnection delays
1628                  */
1629                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX);
1630                 /** Wait for recovery events again, after evicting bad clients */
1631                 goto repeat;
1632         }
1633         return 0;
1634 }
1635
1636 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1637 {
1638         struct ptlrpc_request *req = NULL;
1639         ENTRY;
1640
1641         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1642                obd->obd_next_recovery_transno);
1643
1644         if (target_recovery_overseer(obd, check_for_next_transno,
1645                                      exp_req_replay_healthy)) {
1646                 abort_req_replay_queue(obd);
1647                 abort_lock_replay_queue(obd);
1648         }
1649
1650         cfs_spin_lock(&obd->obd_recovery_task_lock);
1651         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1652                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1653                                      struct ptlrpc_request, rq_list);
1654                 cfs_list_del_init(&req->rq_list);
1655                 obd->obd_requests_queued_for_recovery--;
1656                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1657         } else {
1658                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1659                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1660                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1661                 /** evict exports failed VBR */
1662                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1663         }
1664         RETURN(req);
1665 }
1666
1667 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1668 {
1669         struct ptlrpc_request *req = NULL;
1670
1671         CDEBUG(D_HA, "Waiting for lock\n");
1672         if (target_recovery_overseer(obd, check_for_next_lock,
1673                                      exp_lock_replay_healthy))
1674                 abort_lock_replay_queue(obd);
1675
1676         cfs_spin_lock(&obd->obd_recovery_task_lock);
1677         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1678                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1679                                      struct ptlrpc_request, rq_list);
1680                 cfs_list_del_init(&req->rq_list);
1681                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1682         } else {
1683                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1684                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1685                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1686                 /** evict exports failed VBR */
1687                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1688         }
1689         return req;
1690 }
1691
1692 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1693 {
1694         struct ptlrpc_request *req = NULL;
1695
1696         cfs_spin_lock(&obd->obd_recovery_task_lock);
1697         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1698                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1699                                      struct ptlrpc_request, rq_list);
1700                 cfs_list_del_init(&req->rq_list);
1701                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1702                 if (req->rq_export->exp_in_recovery) {
1703                         cfs_spin_lock(&req->rq_export->exp_lock);
1704                         req->rq_export->exp_in_recovery = 0;
1705                         cfs_spin_unlock(&req->rq_export->exp_lock);
1706                 }
1707         } else {
1708                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1709         }
1710         return req;
1711 }
1712
1713 static int handle_recovery_req(struct ptlrpc_thread *thread,
1714                                struct ptlrpc_request *req,
1715                                svc_handler_t handler)
1716 {
1717         int rc;
1718         ENTRY;
1719
1720         /**
1721          * export can be evicted during recovery, no need to handle replays for
1722          * it after that, discard such request silently
1723          */
1724         if (req->rq_export->exp_disconnected)
1725                 GOTO(reqcopy_put, rc = 0);
1726
1727         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1728         if (rc) {
1729                 CERROR("Failure to initialize session: %d\n", rc);
1730                 GOTO(reqcopy_put, rc);
1731         }
1732
1733         req->rq_recov_session.lc_thread = thread;
1734         lu_context_enter(&req->rq_recov_session);
1735         req->rq_svc_thread = thread;
1736         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1737
1738         /* thread context */
1739         lu_context_enter(&thread->t_env->le_ctx);
1740         (void)handler(req);
1741         lu_context_exit(&thread->t_env->le_ctx);
1742
1743         lu_context_exit(&req->rq_recov_session);
1744         lu_context_fini(&req->rq_recov_session);
1745         /* don't reset timer for final stage */
1746         if (!exp_finished(req->rq_export)) {
1747                 int to = obd_timeout;
1748
1749                 /**
1750                  * Add request timeout to the recovery time so next request from
1751                  * this client may come in recovery time
1752                  */
1753                 if (!AT_OFF)
1754                         to = lustre_msg_get_timeout(req->rq_reqmsg);
1755                  extend_recovery_timer(class_exp2obd(req->rq_export), to);
1756         }
1757 reqcopy_put:
1758         RETURN(rc);
1759 }
1760
1761 static int target_recovery_thread(void *arg)
1762 {
1763         struct lu_target *lut = arg;
1764         struct obd_device *obd = lut->lut_obd;
1765         struct ptlrpc_request *req;
1766         struct target_recovery_data *trd = &obd->obd_recovery_data;
1767         unsigned long delta;
1768         unsigned long flags;
1769         struct lu_env env;
1770         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1771         int rc = 0;
1772         ENTRY;
1773
1774         cfs_daemonize_ctxt("tgt_recov");
1775
1776         SIGNAL_MASK_LOCK(current, flags);
1777         sigfillset(&current->blocked);
1778         RECALC_SIGPENDING;
1779         SIGNAL_MASK_UNLOCK(current, flags);
1780
1781         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1782         if (rc)
1783                 RETURN(rc);
1784
1785         thread->t_env = &env;
1786         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1787         env.le_ctx.lc_thread = thread;
1788         thread->t_data = NULL;
1789
1790         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1791                cfs_curproc_pid());
1792         trd->trd_processing_task = cfs_curproc_pid();
1793
1794         cfs_spin_lock(&obd->obd_dev_lock);
1795         obd->obd_recovering = 1;
1796         cfs_spin_unlock(&obd->obd_dev_lock);
1797         cfs_complete(&trd->trd_starting);
1798
1799         /* first of all, we have to know the first transno to replay */
1800         if (target_recovery_overseer(obd, check_for_clients,
1801                                      exp_connect_healthy)) {
1802                 abort_req_replay_queue(obd);
1803                 abort_lock_replay_queue(obd);
1804         }
1805
1806         /* next stage: replay requests */
1807         delta = jiffies;
1808         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1809                cfs_atomic_read(&obd->obd_req_replay_clients),
1810                obd->obd_next_recovery_transno);
1811         while ((req = target_next_replay_req(obd))) {
1812                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1813                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1814                           lustre_msg_get_transno(req->rq_reqmsg),
1815                           libcfs_nid2str(req->rq_peer.nid));
1816                 handle_recovery_req(thread, req,
1817                                     trd->trd_recovery_handler);
1818                 /**
1819                  * bz18031: increase next_recovery_transno before
1820                  * target_request_copy_put() will drop exp_rpc reference
1821                  */
1822                 cfs_spin_lock(&obd->obd_recovery_task_lock);
1823                 obd->obd_next_recovery_transno++;
1824                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1825                 target_exp_dequeue_req_replay(req);
1826                 target_request_copy_put(req);
1827                 obd->obd_replayed_requests++;
1828         }
1829
1830         /**
1831          * The second stage: replay locks
1832          */
1833         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1834                cfs_atomic_read(&obd->obd_lock_replay_clients));
1835         while ((req = target_next_replay_lock(obd))) {
1836                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1837                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1838                           libcfs_nid2str(req->rq_peer.nid));
1839                 handle_recovery_req(thread, req,
1840                                     trd->trd_recovery_handler);
1841                 target_request_copy_put(req);
1842                 obd->obd_replayed_locks++;
1843         }
1844
1845         /**
1846          * The third stage: reply on final pings, at this moment all clients
1847          * must have request in final queue
1848          */
1849         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1850         /** Update server last boot epoch */
1851         lut_boot_epoch_update(lut);
1852         /* We drop recoverying flag to forward all new requests
1853          * to regular mds_handle() since now */
1854         cfs_spin_lock(&obd->obd_dev_lock);
1855         obd->obd_recovering = obd->obd_abort_recovery = 0;
1856         cfs_spin_unlock(&obd->obd_dev_lock);
1857         cfs_spin_lock(&obd->obd_recovery_task_lock);
1858         target_cancel_recovery_timer(obd);
1859         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1860         while ((req = target_next_final_ping(obd))) {
1861                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1862                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1863                           libcfs_nid2str(req->rq_peer.nid));
1864                 handle_recovery_req(thread, req,
1865                                     trd->trd_recovery_handler);
1866                 target_request_copy_put(req);
1867         }
1868
1869         delta = (jiffies - delta) / CFS_HZ;
1870         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1871               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1872         if (delta > OBD_RECOVERY_TIME_SOFT) {
1873                 CWARN("too long recovery - read logs\n");
1874                 libcfs_debug_dumplog();
1875         }
1876
1877         target_finish_recovery(obd);
1878
1879         lu_context_fini(&env.le_ctx);
1880         trd->trd_processing_task = 0;
1881         cfs_complete(&trd->trd_finishing);
1882         RETURN(rc);
1883 }
1884
1885 static int target_start_recovery_thread(struct lu_target *lut,
1886                                         svc_handler_t handler)
1887 {
1888         struct obd_device *obd = lut->lut_obd;
1889         int rc = 0;
1890         struct target_recovery_data *trd = &obd->obd_recovery_data;
1891
1892         memset(trd, 0, sizeof(*trd));
1893         cfs_init_completion(&trd->trd_starting);
1894         cfs_init_completion(&trd->trd_finishing);
1895         trd->trd_recovery_handler = handler;
1896
1897         if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
1898                 cfs_wait_for_completion(&trd->trd_starting);
1899                 LASSERT(obd->obd_recovering != 0);
1900         } else
1901                 rc = -ECHILD;
1902
1903         return rc;
1904 }
1905
1906 void target_stop_recovery_thread(struct obd_device *obd)
1907 {
1908         if (obd->obd_recovery_data.trd_processing_task > 0) {
1909                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1910                 /** recovery can be done but postrecovery is not yet */
1911                 cfs_spin_lock(&obd->obd_dev_lock);
1912                 if (obd->obd_recovering) {
1913                         CERROR("%s: Aborting recovery\n", obd->obd_name);
1914                         obd->obd_abort_recovery = 1;
1915                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1916                 }
1917                 cfs_spin_unlock(&obd->obd_dev_lock);
1918                 cfs_wait_for_completion(&trd->trd_finishing);
1919         }
1920 }
1921
1922 void target_recovery_fini(struct obd_device *obd)
1923 {
1924         class_disconnect_exports(obd);
1925         target_stop_recovery_thread(obd);
1926         target_cleanup_recovery(obd);
1927 }
1928 EXPORT_SYMBOL(target_recovery_fini);
1929
1930 static void target_recovery_expired(unsigned long castmeharder)
1931 {
1932         struct obd_device *obd = (struct obd_device *)castmeharder;
1933         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1934                " after %lds (%d clients connected)\n",
1935                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
1936                cfs_time_current_sec()- obd->obd_recovery_start,
1937                cfs_atomic_read(&obd->obd_connected_clients));
1938
1939         obd->obd_recovery_expired = 1;
1940         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1941 }
1942
1943 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1944 {
1945         struct obd_device *obd = lut->lut_obd;
1946         if (obd->obd_max_recoverable_clients == 0) {
1947                 /** Update server last boot epoch */
1948                 lut_boot_epoch_update(lut);
1949                 return;
1950         }
1951
1952         CWARN("RECOVERY: service %s, %d recoverable clients, "
1953               "last_transno "LPU64"\n", obd->obd_name,
1954               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1955         LASSERT(obd->obd_stopping == 0);
1956         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1957         obd->obd_recovery_start = 0;
1958         obd->obd_recovery_end = 0;
1959
1960         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1961         target_start_recovery_thread(lut, handler);
1962 }
1963 EXPORT_SYMBOL(target_recovery_init);
1964
1965 #endif
1966
1967 static int target_process_req_flags(struct obd_device *obd,
1968                                     struct ptlrpc_request *req)
1969 {
1970         struct obd_export *exp = req->rq_export;
1971         LASSERT(exp != NULL);
1972         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1973                 /* client declares he's ready to replay locks */
1974                 cfs_spin_lock(&exp->exp_lock);
1975                 if (exp->exp_req_replay_needed) {
1976                         exp->exp_req_replay_needed = 0;
1977                         cfs_spin_unlock(&exp->exp_lock);
1978
1979                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
1980                         cfs_atomic_dec(&obd->obd_req_replay_clients);
1981                 } else {
1982                         cfs_spin_unlock(&exp->exp_lock);
1983                 }
1984         }
1985         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1986                 /* client declares he's ready to complete recovery
1987                  * so, we put the request on th final queue */
1988                 cfs_spin_lock(&exp->exp_lock);
1989                 if (exp->exp_lock_replay_needed) {
1990                         exp->exp_lock_replay_needed = 0;
1991                         cfs_spin_unlock(&exp->exp_lock);
1992
1993                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
1994                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
1995                 } else {
1996                         cfs_spin_unlock(&exp->exp_lock);
1997                 }
1998         }
1999         return 0;
2000 }
2001
2002 int target_queue_recovery_request(struct ptlrpc_request *req,
2003                                   struct obd_device *obd)
2004 {
2005         cfs_list_t *tmp;
2006         int inserted = 0;
2007         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2008         ENTRY;
2009
2010         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2011                 /* Processing the queue right now, don't re-add. */
2012                 RETURN(1);
2013         }
2014
2015         target_process_req_flags(obd, req);
2016
2017         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2018                 /* client declares he's ready to complete recovery
2019                  * so, we put the request on th final queue */
2020                 target_request_copy_get(req);
2021                 DEBUG_REQ(D_HA, req, "queue final req");
2022                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2023                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2024                 if (obd->obd_recovering) {
2025                         cfs_list_add_tail(&req->rq_list,
2026                                           &obd->obd_final_req_queue);
2027                 } else {
2028                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2029                         target_request_copy_put(req);
2030                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2031                 }
2032                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2033                 RETURN(0);
2034         }
2035         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2036                 /* client declares he's ready to replay locks */
2037                 target_request_copy_get(req);
2038                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2039                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2040                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2041                 LASSERT(obd->obd_recovering);
2042                 /* usually due to recovery abort */
2043                 if (!req->rq_export->exp_in_recovery) {
2044                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2045                         target_request_copy_put(req);
2046                         RETURN(-ENOTCONN);
2047                 }
2048                 LASSERT(req->rq_export->exp_lock_replay_needed);
2049                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2050                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2051                 RETURN(0);
2052         }
2053
2054         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2055          * (i.e. buflens etc are in my own byte order), but type-dependent
2056          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
2057
2058         if (!transno) {
2059                 CFS_INIT_LIST_HEAD(&req->rq_list);
2060                 DEBUG_REQ(D_HA, req, "not queueing");
2061                 RETURN(1);
2062         }
2063
2064         /* If we're processing the queue, we want don't want to queue this
2065          * message.
2066          *
2067          * Also, if this request has a transno less than the one we're waiting
2068          * for, we should process it now.  It could (and currently always will)
2069          * be an open request for a descriptor that was opened some time ago.
2070          *
2071          * Also, a resent, replayed request that has already been
2072          * handled will pass through here and be processed immediately.
2073          */
2074         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2075               obd->obd_next_recovery_transno, transno);
2076         cfs_spin_lock(&obd->obd_recovery_task_lock);
2077         if (transno < obd->obd_next_recovery_transno) {
2078                 /* Processing the queue right now, don't re-add. */
2079                 LASSERT(cfs_list_empty(&req->rq_list));
2080                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2081                 RETURN(1);
2082         }
2083         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2084
2085         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2086                 RETURN(0);
2087
2088         target_request_copy_get(req);
2089         if (!req->rq_export->exp_in_recovery) {
2090                 target_request_copy_put(req);
2091                 RETURN(-ENOTCONN);
2092         }
2093         LASSERT(req->rq_export->exp_req_replay_needed);
2094
2095         if (target_exp_enqueue_req_replay(req)) {
2096                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2097                 target_request_copy_put(req);
2098                 RETURN(0);
2099         }
2100
2101         /* XXX O(n^2) */
2102         cfs_spin_lock(&obd->obd_recovery_task_lock);
2103         LASSERT(obd->obd_recovering);
2104         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2105                 struct ptlrpc_request *reqiter =
2106                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2107
2108                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2109                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2110                         inserted = 1;
2111                         break;
2112                 }
2113
2114                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2115                              transno)) {
2116                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2117                                   "has been claimed by another client");
2118                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2119                         target_exp_dequeue_req_replay(req);
2120                         target_request_copy_put(req);
2121                         RETURN(0);
2122                 }
2123         }
2124
2125         if (!inserted)
2126                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2127
2128         obd->obd_requests_queued_for_recovery++;
2129         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2130         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2131         RETURN(0);
2132 }
2133
2134 /**
2135  * Packs current SLV and Limit into \a req.
2136  */
2137 int target_pack_pool_reply(struct ptlrpc_request *req)
2138 {
2139         struct obd_device *obd;
2140         ENTRY;
2141
2142         /*
2143          * Check that we still have all structures alive as this may
2144          * be some late rpc in shutdown time.
2145          */
2146         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2147                      !exp_connect_lru_resize(req->rq_export))) {
2148                 lustre_msg_set_slv(req->rq_repmsg, 0);
2149                 lustre_msg_set_limit(req->rq_repmsg, 0);
2150                 RETURN(0);
2151         }
2152
2153         /*
2154          * OBD is alive here as export is alive, which we checked above.
2155          */
2156         obd = req->rq_export->exp_obd;
2157
2158         cfs_read_lock(&obd->obd_pool_lock);
2159         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2160         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2161         cfs_read_unlock(&obd->obd_pool_lock);
2162
2163         RETURN(0);
2164 }
2165
2166 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2167 {
2168         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2169                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2170                 return (-ECOMM);
2171         }
2172
2173         if (unlikely(rc)) {
2174                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2175                 req->rq_status = rc;
2176                 return (ptlrpc_send_error(req, 1));
2177         } else {
2178                 DEBUG_REQ(D_NET, req, "sending reply");
2179         }
2180
2181         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2182 }
2183
2184 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2185 {
2186         int                        netrc;
2187         struct ptlrpc_reply_state *rs;
2188         struct obd_device         *obd;
2189         struct obd_export         *exp;
2190         struct ptlrpc_service     *svc;
2191         ENTRY;
2192
2193         if (req->rq_no_reply) {
2194                 EXIT;
2195                 return;
2196         }
2197
2198         svc = req->rq_rqbd->rqbd_service;
2199         rs = req->rq_reply_state;
2200         if (rs == NULL || !rs->rs_difficult) {
2201                 /* no notifiers */
2202                 target_send_reply_msg (req, rc, fail_id);
2203                 EXIT;
2204                 return;
2205         }
2206
2207         /* must be an export if locks saved */
2208         LASSERT (req->rq_export != NULL);
2209         /* req/reply consistent */
2210         LASSERT (rs->rs_service == svc);
2211
2212         /* "fresh" reply */
2213         LASSERT (!rs->rs_scheduled);
2214         LASSERT (!rs->rs_scheduled_ever);
2215         LASSERT (!rs->rs_handled);
2216         LASSERT (!rs->rs_on_net);
2217         LASSERT (rs->rs_export == NULL);
2218         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2219         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2220
2221         exp = class_export_get (req->rq_export);
2222         obd = exp->exp_obd;
2223
2224         /* disable reply scheduling while I'm setting up */
2225         rs->rs_scheduled = 1;
2226         rs->rs_on_net    = 1;
2227         rs->rs_xid       = req->rq_xid;
2228         rs->rs_transno   = req->rq_transno;
2229         rs->rs_export    = exp;
2230         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2231
2232         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2233         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2234                rs->rs_transno, exp->exp_last_committed);
2235         if (rs->rs_transno > exp->exp_last_committed) {
2236                 /* not committed already */
2237                 cfs_list_add_tail(&rs->rs_obd_list,
2238                                   &exp->exp_uncommitted_replies);
2239         }
2240         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2241
2242         cfs_spin_lock(&exp->exp_lock);
2243         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2244         cfs_spin_unlock(&exp->exp_lock);
2245
2246         netrc = target_send_reply_msg (req, rc, fail_id);
2247
2248         cfs_spin_lock(&svc->srv_rs_lock);
2249
2250         cfs_atomic_inc(&svc->srv_n_difficult_replies);
2251
2252         if (netrc != 0) {
2253                 /* error sending: reply is off the net.  Also we need +1
2254                  * reply ref until ptlrpc_handle_rs() is done
2255                  * with the reply state (if the send was successful, there
2256                  * would have been +1 ref for the net, which
2257                  * reply_out_callback leaves alone) */
2258                 rs->rs_on_net = 0;
2259                 ptlrpc_rs_addref(rs);
2260         }
2261
2262         cfs_spin_lock(&rs->rs_lock);
2263         if (rs->rs_transno <= exp->exp_last_committed ||
2264             (!rs->rs_on_net && !rs->rs_no_ack) ||
2265              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2266              cfs_list_empty(&rs->rs_obd_list)) {
2267                 CDEBUG(D_HA, "Schedule reply immediately\n");
2268                 ptlrpc_dispatch_difficult_reply(rs);
2269         } else {
2270                 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2271                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2272         }
2273         cfs_spin_unlock(&rs->rs_lock);
2274         cfs_spin_unlock(&svc->srv_rs_lock);
2275         EXIT;
2276 }
2277
2278 int target_handle_ping(struct ptlrpc_request *req)
2279 {
2280         obd_ping(req->rq_export);
2281         return req_capsule_server_pack(&req->rq_pill);
2282 }
2283
2284 void target_committed_to_req(struct ptlrpc_request *req)
2285 {
2286         struct obd_export *exp = req->rq_export;
2287
2288         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2289                 lustre_msg_set_last_committed(req->rq_repmsg,
2290                                               exp->exp_last_committed);
2291         else
2292                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2293                           "%d)", exp->exp_obd->obd_no_transno,
2294                           req->rq_repmsg == NULL);
2295
2296         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2297                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2298 }
2299 EXPORT_SYMBOL(target_committed_to_req);
2300
2301 int target_handle_qc_callback(struct ptlrpc_request *req)
2302 {
2303         struct obd_quotactl *oqctl;
2304         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2305
2306         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2307         if (oqctl == NULL) {
2308                 CERROR("Can't unpack obd_quotactl\n");
2309                 RETURN(-EPROTO);
2310         }
2311
2312         cli->cl_qchk_stat = oqctl->qc_stat;
2313
2314         return 0;
2315 }
2316
2317 #ifdef HAVE_QUOTA_SUPPORT
2318 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2319 {
2320 #ifdef __KERNEL__
2321         struct obd_device *obd = req->rq_export->exp_obd;
2322         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2323         struct obd_device_target *obt;
2324         struct lustre_quota_ctxt *qctxt;
2325         struct qunit_data *qdata = NULL;
2326         int rc = 0;
2327         ENTRY;
2328
2329         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2330                 RETURN(rc);
2331
2332         rc = req_capsule_server_pack(&req->rq_pill);
2333         if (rc) {
2334                 CERROR("packing reply failed!: rc = %d\n", rc);
2335                 RETURN(rc);
2336         }
2337
2338         LASSERT(req->rq_export);
2339
2340         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2341         if (IS_ERR(qdata)) {
2342                 rc = PTR_ERR(qdata);
2343                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2344                 req->rq_status = rc;
2345                 GOTO(out, rc);
2346         }
2347
2348         /* we use the observer */
2349         if (obd_pin_observer(obd, &lov_obd) ||
2350             obd_pin_observer(lov_obd, &master_obd)) {
2351                 CERROR("Can't find the observer, it is recovering\n");
2352                 req->rq_status = -EAGAIN;
2353                 GOTO(out, rc);
2354         }
2355
2356         obt = &master_obd->u.obt;
2357         qctxt = &obt->obt_qctxt;
2358
2359         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2360                 /* quota_type has not been processed yet, return EAGAIN
2361                  * until we know whether or not quotas are supposed to
2362                  * be enabled */
2363                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2364                        "-EAGAIN\n");
2365                 req->rq_status = -EAGAIN;
2366                 GOTO(out, rc);
2367         }
2368
2369         cfs_down_read(&obt->obt_rwsem);
2370         if (qctxt->lqc_lqs_hash == NULL) {
2371                 cfs_up_read(&obt->obt_rwsem);
2372                 /* quota_type has not been processed yet, return EAGAIN
2373                  * until we know whether or not quotas are supposed to
2374                  * be enabled */
2375                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2376                        "-EAGAIN\n");
2377                 req->rq_status = -EAGAIN;
2378                 GOTO(out, rc);
2379         }
2380
2381         LASSERT(qctxt->lqc_handler);
2382         rc = qctxt->lqc_handler(master_obd, qdata,
2383                                 lustre_msg_get_opc(req->rq_reqmsg));
2384         cfs_up_read(&obt->obt_rwsem);
2385         if (rc && rc != -EDQUOT)
2386                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2387                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2388         req->rq_status = rc;
2389
2390         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2391         if (rc < 0) {
2392                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2393                 GOTO(out, rc);
2394         }
2395
2396         /* Block the quota req. b=14840 */
2397         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2398         EXIT;
2399
2400 out:
2401         if (master_obd)
2402                 obd_unpin_observer(lov_obd);
2403         if (lov_obd)
2404                 obd_unpin_observer(obd);
2405
2406         rc = ptlrpc_reply(req);
2407         return rc;
2408 #else
2409         return 0;
2410 #endif /* !__KERNEL__ */
2411 }
2412 #endif /* HAVE_QUOTA_SUPPORT */
2413
2414 ldlm_mode_t lck_compat_array[] = {
2415         [LCK_EX] LCK_COMPAT_EX,
2416         [LCK_PW] LCK_COMPAT_PW,
2417         [LCK_PR] LCK_COMPAT_PR,
2418         [LCK_CW] LCK_COMPAT_CW,
2419         [LCK_CR] LCK_COMPAT_CR,
2420         [LCK_NL] LCK_COMPAT_NL,
2421         [LCK_GROUP] LCK_COMPAT_GROUP,
2422         [LCK_COS] LCK_COMPAT_COS,
2423 };
2424
2425 /**
2426  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2427  * not escape to the user level.
2428  */
2429 int ldlm_error2errno(ldlm_error_t error)
2430 {
2431         int result;
2432
2433         switch (error) {
2434         case ELDLM_OK:
2435                 result = 0;
2436                 break;
2437         case ELDLM_LOCK_CHANGED:
2438                 result = -ESTALE;
2439                 break;
2440         case ELDLM_LOCK_ABORTED:
2441                 result = -ENAVAIL;
2442                 break;
2443         case ELDLM_LOCK_REPLACED:
2444                 result = -ESRCH;
2445                 break;
2446         case ELDLM_NO_LOCK_DATA:
2447                 result = -ENOENT;
2448                 break;
2449         case ELDLM_NAMESPACE_EXISTS:
2450                 result = -EEXIST;
2451                 break;
2452         case ELDLM_BAD_NAMESPACE:
2453                 result = -EBADF;
2454                 break;
2455         default:
2456                 if (((int)error) < 0)  /* cast to signed type */
2457                         result = error; /* as ldlm_error_t can be unsigned */
2458                 else {
2459                         CERROR("Invalid DLM result code: %d\n", error);
2460                         result = -EPROTO;
2461                 }
2462         }
2463         return result;
2464 }
2465 EXPORT_SYMBOL(ldlm_error2errno);
2466
2467 /**
2468  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2469  */
2470 ldlm_error_t ldlm_errno2error(int err_no)
2471 {
2472         int error;
2473
2474         switch (err_no) {
2475         case 0:
2476                 error = ELDLM_OK;
2477                 break;
2478         case -ESTALE:
2479                 error = ELDLM_LOCK_CHANGED;
2480                 break;
2481         case -ENAVAIL:
2482                 error = ELDLM_LOCK_ABORTED;
2483                 break;
2484         case -ESRCH:
2485                 error = ELDLM_LOCK_REPLACED;
2486                 break;
2487         case -ENOENT:
2488                 error = ELDLM_NO_LOCK_DATA;
2489                 break;
2490         case -EEXIST:
2491                 error = ELDLM_NAMESPACE_EXISTS;
2492                 break;
2493         case -EBADF:
2494                 error = ELDLM_BAD_NAMESPACE;
2495                 break;
2496         default:
2497                 error = err_no;
2498         }
2499         return error;
2500 }
2501 EXPORT_SYMBOL(ldlm_errno2error);
2502
2503 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2504 void ldlm_dump_export_locks(struct obd_export *exp)
2505 {
2506         cfs_spin_lock(&exp->exp_locks_list_guard);
2507         if (!cfs_list_empty(&exp->exp_locks_list)) {
2508             struct ldlm_lock *lock;
2509
2510             CERROR("dumping locks for export %p,"
2511                    "ignore if the unmount doesn't hang\n", exp);
2512             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2513                 ldlm_lock_dump(D_ERROR, lock, 0);
2514         }
2515         cfs_spin_unlock(&exp->exp_locks_list_guard);
2516 }
2517 #endif
2518
2519 static int target_bulk_timeout(void *data)
2520 {
2521         ENTRY;
2522         /* We don't fail the connection here, because having the export
2523          * killed makes the (vital) call to commitrw very sad.
2524          */
2525         RETURN(1);
2526 }
2527
2528 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2529 {
2530         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2531 }
2532
2533 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2534                    struct l_wait_info *lwi)
2535 {
2536         struct ptlrpc_request *req = desc->bd_req;
2537         int rc = 0;
2538         ENTRY;
2539
2540         /* Check if there is eviction in progress, and if so, wait for
2541          * it to finish */
2542         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2543                 *lwi = LWI_INTR(NULL, NULL);
2544                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2545                                   !cfs_atomic_read(&exp->exp_obd->
2546                                                    obd_evict_inprogress),
2547                                   lwi);
2548         }
2549
2550         /* Check if client was evicted or tried to reconnect already */
2551         if (exp->exp_failed || exp->exp_abort_active_req) {
2552                 rc = -ENOTCONN;
2553         } else {
2554                 if (desc->bd_type == BULK_PUT_SINK)
2555                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2556                 if (rc == 0)
2557                         rc = ptlrpc_start_bulk_transfer(desc);
2558         }
2559
2560         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2561                 ptlrpc_abort_bulk(desc);
2562         } else if (rc == 0) {
2563                 time_t start = cfs_time_current_sec();
2564                 do {
2565                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2566                         cfs_duration_t timeout = timeoutl <= 0 ?
2567                                 CFS_TICK : cfs_time_seconds(timeoutl);
2568                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2569                                                     cfs_time_seconds(1),
2570                                                    target_bulk_timeout,
2571                                                    desc);
2572                         rc = l_wait_event(desc->bd_waitq,
2573                                           !ptlrpc_server_bulk_active(desc) ||
2574                                           exp->exp_failed ||
2575                                           exp->exp_abort_active_req,
2576                                           lwi);
2577                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2578                         /* Wait again if we changed deadline */
2579                 } while ((rc == -ETIMEDOUT) &&
2580                          (req->rq_deadline > cfs_time_current_sec()));
2581
2582                 if (rc == -ETIMEDOUT) {
2583                         DEBUG_REQ(D_ERROR, req,
2584                                   "timeout on bulk %s after %ld%+lds",
2585                                   bulk2type(desc),
2586                                   req->rq_deadline - start,
2587                                   cfs_time_current_sec() -
2588                                   req->rq_deadline);
2589                         ptlrpc_abort_bulk(desc);
2590                 } else if (exp->exp_failed) {
2591                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2592                                   bulk2type(desc));
2593                         rc = -ENOTCONN;
2594                         ptlrpc_abort_bulk(desc);
2595                 } else if (exp->exp_abort_active_req) {
2596                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2597                                   bulk2type(desc));
2598                         /* we don't reply anyway */
2599                         rc = -ETIMEDOUT;
2600                         ptlrpc_abort_bulk(desc);
2601                 } else if (!desc->bd_success ||
2602                            desc->bd_nob_transferred != desc->bd_nob) {
2603                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2604                                   desc->bd_success ?
2605                                   "truncated" : "network error on",
2606                                   bulk2type(desc),
2607                                   desc->bd_nob_transferred,
2608                                   desc->bd_nob);
2609                         /* XXX should this be a different errno? */
2610                         rc = -ETIMEDOUT;
2611                 } else if (desc->bd_type == BULK_GET_SINK) {
2612                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2613                 }
2614         } else {
2615                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2616                           bulk2type(desc), rc);
2617         }
2618
2619         RETURN(rc);
2620 }
2621 EXPORT_SYMBOL(target_bulk_io);