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  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #ifdef __KERNEL__
  43 # include <libcfs/libcfs.h>
  44 #else
  45 # include <liblustre.h>
  46 #endif
  47 #include <obd.h>
  48 #include <lustre_mds.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /* @priority: if non-zero, move the selected to the list head
  55  * @create: if zero, only search in existed connections
  56  */
  57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  58                            int priority, int create)
  59 {
  60         struct ptlrpc_connection *ptlrpc_conn;
  61         struct obd_import_conn *imp_conn = NULL, *item;
  62         int rc = 0;
  63         ENTRY;
  64
  65         if (!create && !priority) {
  66                 CDEBUG(D_HA, "Nothing to do\n");
  67                 RETURN(-EINVAL);
  68         }
  69
  70         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  71         if (!ptlrpc_conn) {
  72                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  73                 RETURN (-ENOENT);
  74         }
  75
  76         if (create) {
  77                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  78                 if (!imp_conn) {
  79                         GOTO(out_put, rc = -ENOMEM);
  80                 }
  81         }
  82
  83         spin_lock(&imp->imp_lock);
  84         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  85                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  86                         if (priority) {
  87                                 list_del(&item->oic_item);
  88                                 list_add(&item->oic_item, &imp->imp_conn_list);
  89                                 item->oic_last_attempt = 0;
  90                         }
  91                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  92                                imp, imp->imp_obd->obd_name, uuid->uuid,
  93                                (priority ? ", moved to head" : ""));
  94                         spin_unlock(&imp->imp_lock);
  95                         GOTO(out_free, rc = 0);
  96                 }
  97         }
  98         /* not found */
  99         if (create) {
 100                 imp_conn->oic_conn = ptlrpc_conn;
 101                 imp_conn->oic_uuid = *uuid;
 102                 imp_conn->oic_last_attempt = 0;
 103                 if (priority)
 104                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 105                 else
 106                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 108                        imp, imp->imp_obd->obd_name, uuid->uuid,
 109                        (priority ? "head" : "tail"));
 110         } else {
 111                 spin_unlock(&imp->imp_lock);
 112                 GOTO(out_free, rc = -ENOENT);
 113         }
 114
 115         spin_unlock(&imp->imp_lock);
 116         RETURN(0);
 117 out_free:
 118         if (imp_conn)
 119                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 120 out_put:
 121         ptlrpc_connection_put(ptlrpc_conn);
 122         RETURN(rc);
 123 }
 124
 125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 126 {
 127         return import_set_conn(imp, uuid, 1, 0);
 128 }
 129
 130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 131                            int priority)
 132 {
 133         return import_set_conn(imp, uuid, priority, 1);
 134 }
 135
 136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 137 {
 138         struct obd_import_conn *imp_conn;
 139         struct obd_export *dlmexp;
 140         int rc = -ENOENT;
 141         ENTRY;
 142
 143         spin_lock(&imp->imp_lock);
 144         if (list_empty(&imp->imp_conn_list)) {
 145                 LASSERT(!imp->imp_connection);
 146                 GOTO(out, rc);
 147         }
 148
 149         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 150                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 151                         continue;
 152                 LASSERT(imp_conn->oic_conn);
 153
 154                 /* is current conn? */
 155                 if (imp_conn == imp->imp_conn_current) {
 156                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 157
 158                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 159                             imp->imp_state != LUSTRE_IMP_DISCON) {
 160                                 CERROR("can't remove current connection\n");
 161                                 GOTO(out, rc = -EBUSY);
 162                         }
 163
 164                         ptlrpc_connection_put(imp->imp_connection);
 165                         imp->imp_connection = NULL;
 166
 167                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 168                         if (dlmexp && dlmexp->exp_connection) {
 169                                 LASSERT(dlmexp->exp_connection ==
 170                                         imp_conn->oic_conn);
 171                                 ptlrpc_connection_put(dlmexp->exp_connection);
 172                                 dlmexp->exp_connection = NULL;
 173                         }
 174                 }
 175
 176                 list_del(&imp_conn->oic_item);
 177                 ptlrpc_connection_put(imp_conn->oic_conn);
 178                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 179                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 180                        imp, imp->imp_obd->obd_name, uuid->uuid);
 181                 rc = 0;
 182                 break;
 183         }
 184 out:
 185         spin_unlock(&imp->imp_lock);
 186         if (rc == -ENOENT)
 187                 CERROR("connection %s not found\n", uuid->uuid);
 188         RETURN(rc);
 189 }
 190
 191 void client_destroy_import(struct obd_import *imp)
 192 {
 193         /* drop security policy instance after all rpc finished/aborted
 194          * to let all busy contexts be released. */
 195         class_import_get(imp);
 196         class_destroy_import(imp);
 197         sptlrpc_import_sec_put(imp);
 198         class_import_put(imp);
 199 }
 200
 201 /* configure an RPC client OBD device
 202  *
 203  * lcfg parameters:
 204  * 1 - client UUID
 205  * 2 - server UUID
 206  * 3 - inactive-on-startup
 207  */
 208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 209 {
 210         struct client_obd *cli = &obddev->u.cli;
 211         struct obd_import *imp;
 212         struct obd_uuid server_uuid;
 213         int rq_portal, rp_portal, connect_op;
 214         char *name = obddev->obd_type->typ_name;
 215         int rc;
 216         ENTRY;
 217
 218         /* In a more perfect world, we would hang a ptlrpc_client off of
 219          * obd_type and just use the values from there. */
 220         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 221                 rq_portal = OST_REQUEST_PORTAL;
 222                 rp_portal = OSC_REPLY_PORTAL;
 223                 connect_op = OST_CONNECT;
 224                 cli->cl_sp_me = LUSTRE_SP_CLI;
 225                 cli->cl_sp_to = LUSTRE_SP_OST;
 226         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 227                 rq_portal = MDS_REQUEST_PORTAL;
 228                 rp_portal = MDC_REPLY_PORTAL;
 229                 connect_op = MDS_CONNECT;
 230                 cli->cl_sp_me = LUSTRE_SP_CLI;
 231                 cli->cl_sp_to = LUSTRE_SP_MDT;
 232         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 233                 rq_portal = MGS_REQUEST_PORTAL;
 234                 rp_portal = MGC_REPLY_PORTAL;
 235                 connect_op = MGS_CONNECT;
 236                 cli->cl_sp_me = LUSTRE_SP_MGC;
 237                 cli->cl_sp_to = LUSTRE_SP_MGS;
 238                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 239         } else {
 240                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 241                        name);
 242                 RETURN(-EINVAL);
 243         }
 244
 245         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 246                 CERROR("requires a TARGET UUID\n");
 247                 RETURN(-EINVAL);
 248         }
 249
 250         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 251                 CERROR("client UUID must be less than 38 characters\n");
 252                 RETURN(-EINVAL);
 253         }
 254
 255         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 256                 CERROR("setup requires a SERVER UUID\n");
 257                 RETURN(-EINVAL);
 258         }
 259
 260         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 261                 CERROR("target UUID must be less than 38 characters\n");
 262                 RETURN(-EINVAL);
 263         }
 264
 265         init_rwsem(&cli->cl_sem);
 266         sema_init(&cli->cl_mgc_sem, 1);
 267         cli->cl_conn_count = 0;
 268         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 269                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 270                      sizeof(server_uuid)));
 271
 272         cli->cl_dirty = 0;
 273         cli->cl_avail_grant = 0;
 274         /* FIXME: should limit this for the sum of all cl_dirty_max */
 275         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 276         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 277                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 278         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 279         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 280         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 281         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 282         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 283         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 284         cli->cl_r_in_flight = 0;
 285         cli->cl_w_in_flight = 0;
 286
 287         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 288         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 289         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 290         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 291         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 292         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 293         cfs_waitq_init(&cli->cl_destroy_waitq);
 294         atomic_set(&cli->cl_destroy_in_flight, 0);
 295 #ifdef ENABLE_CHECKSUM
 296         /* Turn on checksumming by default. */
 297         cli->cl_checksum = 1;
 298         /*
 299          * The supported checksum types will be worked out at connect time
 300          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 301          * through procfs.
 302          */
 303         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 304 #endif
 305         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 306
 307         /* This value may be changed at connect time in
 308            ptlrpc_connect_interpret. */
 309         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 310                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 311
 312         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 313                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 314         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 315                 cli->cl_max_rpcs_in_flight = 2;
 316         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 317                 cli->cl_max_rpcs_in_flight = 3;
 318         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 319                 cli->cl_max_rpcs_in_flight = 4;
 320         } else {
 321                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 322         }
 323
 324         rc = ldlm_get_ref();
 325         if (rc) {
 326                 CERROR("ldlm_get_ref failed: %d\n", rc);
 327                 GOTO(err, rc);
 328         }
 329
 330         ptlrpc_init_client(rq_portal, rp_portal, name,
 331                            &obddev->obd_ldlm_client);
 332
 333         imp = class_new_import(obddev);
 334         if (imp == NULL)
 335                 GOTO(err_ldlm, rc = -ENOENT);
 336         imp->imp_client = &obddev->obd_ldlm_client;
 337         imp->imp_connect_op = connect_op;
 338         imp->imp_initial_recov = 1;
 339         imp->imp_initial_recov_bk = 0;
 340         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 341         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 342                LUSTRE_CFG_BUFLEN(lcfg, 1));
 343         class_import_put(imp);
 344
 345         rc = client_import_add_conn(imp, &server_uuid, 1);
 346         if (rc) {
 347                 CERROR("can't add initial connection\n");
 348                 GOTO(err_import, rc);
 349         }
 350
 351         cli->cl_import = imp;
 352         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 353         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 354         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 355
 356         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 357                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 358                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 359                                name, obddev->obd_name,
 360                                cli->cl_target_uuid.uuid);
 361                         spin_lock(&imp->imp_lock);
 362                         imp->imp_invalid = 1;
 363                         spin_unlock(&imp->imp_lock);
 364                 }
 365         }
 366
 367         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 368
 369         RETURN(rc);
 370
 371 err_import:
 372         class_destroy_import(imp);
 373 err_ldlm:
 374         ldlm_put_ref();
 375 err:
 376         RETURN(rc);
 377
 378 }
 379
 380 int client_obd_cleanup(struct obd_device *obddev)
 381 {
 382         ENTRY;
 383         ldlm_put_ref();
 384         RETURN(0);
 385 }
 386
 387 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 388 int client_connect_import(const struct lu_env *env,
 389                           struct obd_export **exp,
 390                           struct obd_device *obd, struct obd_uuid *cluuid,
 391                           struct obd_connect_data *data, void *localdata)
 392 {
 393         struct client_obd *cli = &obd->u.cli;
 394         struct obd_import *imp = cli->cl_import;
 395         struct obd_connect_data *ocd;
 396         struct ldlm_namespace *to_be_freed = NULL;
 397         struct lustre_handle conn = { 0 };
 398         int rc;
 399         ENTRY;
 400
 401         *exp = NULL;
 402         down_write(&cli->cl_sem);
 403         rc = class_connect(&conn, obd, cluuid);
 404         if (rc)
 405                 GOTO(out_sem, rc);
 406
 407         *exp = class_conn2export(&conn);
 408
 409         cli->cl_conn_count++;
 410         if (cli->cl_conn_count > 1)
 411                 GOTO(out_sem, rc);
 412
 413         if (obd->obd_namespace != NULL)
 414                 CERROR("already have namespace!\n");
 415
 416         /*
 417          * Deadlock case - bug 18380
 418          */
 419         up_write(&cli->cl_sem);
 420         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 421                                                 LDLM_NAMESPACE_CLIENT,
 422                                                 LDLM_NAMESPACE_GREEDY);
 423         down_write(&cli->cl_sem);
 424         if (obd->obd_namespace == NULL)
 425                 GOTO(out_disco, rc = -ENOMEM);
 426
 427         imp->imp_dlm_handle = conn;
 428         rc = ptlrpc_init_import(imp);
 429         if (rc != 0)
 430                 GOTO(out_ldlm, rc);
 431
 432         ocd = &imp->imp_connect_data;
 433         if (data) {
 434                 *ocd = *data;
 435                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 436         }
 437
 438         rc = ptlrpc_connect_import(imp, NULL);
 439         if (rc != 0) {
 440                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 441                 GOTO(out_ldlm, rc);
 442         }
 443         LASSERT((*exp)->exp_connection);
 444
 445         if (data) {
 446                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 447                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 448                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 449                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 450         }
 451
 452         ptlrpc_pinger_add_import(imp);
 453
 454         EXIT;
 455
 456         if (rc) {
 457 out_ldlm:
 458                 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
 459                 to_be_freed = obd->obd_namespace;
 460                 obd->obd_namespace = NULL;
 461 out_disco:
 462                 cli->cl_conn_count--;
 463                 class_disconnect(*exp);
 464                 *exp = NULL;
 465         }
 466 out_sem:
 467         up_write(&cli->cl_sem);
 468         if (to_be_freed)
 469                 ldlm_namespace_free_post(to_be_freed);
 470
 471         return rc;
 472 }
 473
 474 int client_disconnect_export(struct obd_export *exp)
 475 {
 476         struct obd_device *obd = class_exp2obd(exp);
 477         struct client_obd *cli;
 478         struct obd_import *imp;
 479         int rc = 0, err;
 480         struct ldlm_namespace *to_be_freed = NULL;
 481         ENTRY;
 482
 483         if (!obd) {
 484                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 485                        exp, exp ? exp->exp_handle.h_cookie : -1);
 486                 RETURN(-EINVAL);
 487         }
 488
 489         cli = &obd->u.cli;
 490         imp = cli->cl_import;
 491
 492         down_write(&cli->cl_sem);
 493         if (!cli->cl_conn_count) {
 494                 CERROR("disconnecting disconnected device (%s)\n",
 495                        obd->obd_name);
 496                 GOTO(out_sem, rc = -EINVAL);
 497         }
 498
 499         cli->cl_conn_count--;
 500         if (cli->cl_conn_count)
 501                 GOTO(out_no_disconnect, rc = 0);
 502
 503         /* Mark import deactivated now, so we don't try to reconnect if any
 504          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 505          * fully deactivate the import, or that would drop all requests. */
 506         spin_lock(&imp->imp_lock);
 507         imp->imp_deactive = 1;
 508         spin_unlock(&imp->imp_lock);
 509
 510         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 511          * delete it regardless.  (It's safe to delete an import that was
 512          * never added.) */
 513         (void)ptlrpc_pinger_del_import(imp);
 514
 515         if (obd->obd_namespace != NULL) {
 516                 /* obd_force == local only */
 517                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 518                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 519                                        NULL);
 520                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 521                 to_be_freed = obd->obd_namespace;
 522         }
 523
 524         /*
 525          * there's no necessary to hold sem during diconnecting an import,
 526          * and actually it may cause deadlock in gss.
 527          */
 528         up_write(&cli->cl_sem);
 529         rc = ptlrpc_disconnect_import(imp, 0);
 530         down_write(&cli->cl_sem);
 531
 532         ptlrpc_invalidate_import(imp);
 533         /* set obd_namespace to NULL only after invalidate, because we can have
 534          * some connect requests in flight, and his need store a connect flags
 535          * in obd_namespace. bug 14260 */
 536         obd->obd_namespace = NULL;
 537
 538         if (imp->imp_rq_pool) {
 539                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 540                 imp->imp_rq_pool = NULL;
 541         }
 542         client_destroy_import(imp);
 543         cli->cl_import = NULL;
 544
 545         EXIT;
 546  out_no_disconnect:
 547         err = class_disconnect(exp);
 548         if (!rc && err)
 549                 rc = err;
 550  out_sem:
 551         up_write(&cli->cl_sem);
 552         if (to_be_freed)
 553                 ldlm_namespace_free_post(to_be_freed);
 554
 555         RETURN(rc);
 556 }
 557
 558 /* --------------------------------------------------------------------------
 559  * from old lib/target.c
 560  * -------------------------------------------------------------------------- */
 561
 562 static int target_handle_reconnect(struct lustre_handle *conn,
 563                                    struct obd_export *exp,
 564                                    struct obd_uuid *cluuid)
 565 {
 566         ENTRY;
 567         if (exp->exp_connection && exp->exp_imp_reverse) {
 568                 struct lustre_handle *hdl;
 569                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 570                 /* Might be a re-connect after a partition. */
 571                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 572                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 573                               cluuid->uuid);
 574                         conn->cookie = exp->exp_handle.h_cookie;
 575                         /* target_handle_connect() treats EALREADY and
 576                          * -EALREADY differently.  EALREADY means we are
 577                          * doing a valid reconnect from the same client. */
 578                         RETURN(EALREADY);
 579                 } else {
 580                         CERROR("%s reconnecting from %s, "
 581                                "handle mismatch (ours "LPX64", theirs "
 582                                LPX64")\n", cluuid->uuid,
 583                                exp->exp_connection->c_remote_uuid.uuid,
 584                                hdl->cookie, conn->cookie);
 585                         memset(conn, 0, sizeof *conn);
 586                         /* target_handle_connect() treats EALREADY and
 587                          * -EALREADY differently.  -EALREADY is an error
 588                          * (same UUID, different handle). */
 589                         RETURN(-EALREADY);
 590                 }
 591         }
 592
 593         conn->cookie = exp->exp_handle.h_cookie;
 594         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 595                cluuid->uuid, exp, conn->cookie);
 596         RETURN(0);
 597 }
 598
 599 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 600                           int error)
 601 {
 602         struct obd_export *exp = cb_data;
 603
 604         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 605                obd->obd_name, exp->exp_client_uuid.uuid);
 606
 607         spin_lock(&exp->exp_lock);
 608         exp->exp_need_sync = 0;
 609         spin_unlock(&exp->exp_lock);
 610 }
 611 EXPORT_SYMBOL(target_client_add_cb);
 612 static void
 613 target_start_and_reset_recovery_timer(struct obd_device *obd,
 614                                       struct ptlrpc_request *req,
 615                                       int new_client);
 616
 617 int target_handle_connect(struct ptlrpc_request *req)
 618 {
 619         struct obd_device *target, *targref = NULL;
 620         struct obd_export *export = NULL;
 621         struct obd_import *revimp;
 622         struct lustre_handle conn;
 623         struct lustre_handle *tmp;
 624         struct obd_uuid tgtuuid;
 625         struct obd_uuid cluuid;
 626         struct obd_uuid remote_uuid;
 627         char *str;
 628         int rc = 0;
 629         int mds_conn = 0;
 630         struct obd_connect_data *data, *tmpdata;
 631         lnet_nid_t *client_nid = NULL;
 632         ENTRY;
 633
 634         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 635
 636         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 637         if (str == NULL) {
 638                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 639                 GOTO(out, rc = -EINVAL);
 640         }
 641
 642         obd_str2uuid(&tgtuuid, str);
 643         target = class_uuid2obd(&tgtuuid);
 644         if (!target)
 645                 target = class_name2obd(str);
 646
 647         if (!target || target->obd_stopping || !target->obd_set_up) {
 648                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 649                                    " for connect (%s)\n", str,
 650                                    !target ? "no target" :
 651                                    (target->obd_stopping ? "stopping" :
 652                                    "not set up"));
 653                 GOTO(out, rc = -ENODEV);
 654         }
 655
 656         if (target->obd_no_conn) {
 657                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 658                               "from %s\n", target->obd_name,
 659                               libcfs_nid2str(req->rq_peer.nid));
 660                 GOTO(out, rc = -EAGAIN);
 661         }
 662
 663         /* Make sure the target isn't cleaned up while we're here. Yes,
 664            there's still a race between the above check and our incref here.
 665            Really, class_uuid2obd should take the ref. */
 666         targref = class_incref(target, __FUNCTION__, cfs_current());
 667
 668
 669         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 670         if (str == NULL) {
 671                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 672                 GOTO(out, rc = -EINVAL);
 673         }
 674
 675         obd_str2uuid(&cluuid, str);
 676
 677         /* XXX extract a nettype and format accordingly */
 678         switch (sizeof(lnet_nid_t)) {
 679                 /* NB the casts only avoid compiler warnings */
 680         case 8:
 681                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 682                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 683                 break;
 684         case 4:
 685                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 686                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 687                 break;
 688         default:
 689                 LBUG();
 690         }
 691
 692         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 693         if (tmp == NULL)
 694                 GOTO(out, rc = -EPROTO);
 695
 696         conn = *tmp;
 697
 698         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 699         if (!data)
 700                 GOTO(out, rc = -EPROTO);
 701
 702         rc = req_capsule_server_pack(&req->rq_pill);
 703         if (rc)
 704                 GOTO(out, rc);
 705
 706         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 707                 if (!data) {
 708                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 709                                   "libclient connection attempt");
 710                         GOTO(out, rc = -EPROTO);
 711                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 712                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 713                            data->ocd_version > LUSTRE_VERSION_CODE +
 714                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 715                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 716                                   "libclient connection attempt",
 717                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 718                                   "old" : "new",
 719                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 720                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 721                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 722                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 723                         data = req_capsule_server_sized_get(&req->rq_pill,
 724                                                         &RMF_CONNECT_DATA,
 725                                     offsetof(typeof(*data), ocd_version) +
 726                                              sizeof(data->ocd_version));
 727                         if (data) {
 728                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 729                                 data->ocd_version = LUSTRE_VERSION_CODE;
 730                         }
 731                         GOTO(out, rc = -EPROTO);
 732                 }
 733         }
 734
 735         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 736             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 737                 mds_conn = 1;
 738
 739         /* lctl gets a backstage, all-access pass. */
 740         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 741                 goto dont_check_exports;
 742
 743         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 744         if (!export)
 745                 goto no_export;
 746
 747         /* we've found an export in the hash */
 748         if (export->exp_connecting) { /* bug 9635, et. al. */
 749                 CWARN("%s: exp %p already connecting\n",
 750                       export->exp_obd->obd_name, export);
 751                 class_export_put(export);
 752                 export = NULL;
 753                 rc = -EALREADY;
 754         } else if (mds_conn && export->exp_connection) {
 755                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 756                         /* mds reconnected after failover */
 757                         CWARN("%s: received MDS connection from NID %s,"
 758                               " removing former export from NID %s\n",
 759                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 760                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 761                 else
 762                         /* new mds connection from the same nid */
 763                         CWARN("%s: received new MDS connection from NID %s,"
 764                               " removing former export from same NID\n",
 765                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 766                 class_fail_export(export);
 767                 class_export_put(export);
 768                 export = NULL;
 769                 rc = 0;
 770         } else if (export->exp_connection != NULL &&
 771                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 772                 /* in mds failover we have static uuid but nid can be
 773                  * changed*/
 774                 CWARN("%s: cookie %s seen on new NID %s when "
 775                       "existing NID %s is already connected\n",
 776                       target->obd_name, cluuid.uuid,
 777                       libcfs_nid2str(req->rq_peer.nid),
 778                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 779                 rc = -EALREADY;
 780                 class_export_put(export);
 781                 export = NULL;
 782         } else {
 783                 spin_lock(&export->exp_lock);
 784                 export->exp_connecting = 1;
 785                 spin_unlock(&export->exp_lock);
 786                 class_export_put(export);
 787                 LASSERT(export->exp_obd == target);
 788
 789                 rc = target_handle_reconnect(&conn, export, &cluuid);
 790         }
 791
 792         /* If we found an export, we already unlocked. */
 793         if (!export) {
 794 no_export:
 795                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 796         } else if (req->rq_export == NULL &&
 797                    atomic_read(&export->exp_rpc_count) > 0) {
 798                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 799                       target->obd_name, cluuid.uuid,
 800                       libcfs_nid2str(req->rq_peer.nid),
 801                       export, atomic_read(&export->exp_refcount));
 802                 GOTO(out, rc = -EBUSY);
 803         } else if (req->rq_export != NULL &&
 804                    (atomic_read(&export->exp_rpc_count) > 1)) {
 805                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 806                       target->obd_name, cluuid.uuid,
 807                       libcfs_nid2str(req->rq_peer.nid),
 808                       export, atomic_read(&export->exp_rpc_count));
 809                 GOTO(out, rc = -EBUSY);
 810         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 811                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 812                        "cookies not random?\n", target->obd_name,
 813                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 814                 GOTO(out, rc = -EALREADY);
 815         } else {
 816                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 817         }
 818
 819         if (rc < 0) {
 820                 GOTO(out, rc);
 821         }
 822
 823         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 824                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 825               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 826               export, (long)cfs_time_current_sec(),
 827               export ? (long)export->exp_last_request_time : 0);
 828
 829         /* Tell the client if we're in recovery. */
 830         if (target->obd_recovering) {
 831                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 832                 /* If this is the first time a client connects,
 833                    reset the recovery timer */
 834                 if (rc == 0)
 835                         target_start_and_reset_recovery_timer(target, req,
 836                                                               !export);
 837         }
 838
 839         /* We want to handle EALREADY but *not* -EALREADY from
 840          * target_handle_reconnect(), return reconnection state in a flag */
 841         if (rc == EALREADY) {
 842                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 843                 rc = 0;
 844         } else {
 845                 LASSERT(rc == 0);
 846         }
 847
 848         /* Tell the client if we support replayable requests */
 849         if (target->obd_replayable)
 850                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 851         client_nid = &req->rq_peer.nid;
 852
 853         if (export == NULL) {
 854                 if (target->obd_recovering) {
 855                         cfs_time_t t;
 856
 857                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 858                         t = cfs_time_sub(t, cfs_time_current());
 859                         CERROR("%s: denying connection for new client %s (%s): "
 860                                "%d clients in recovery for "CFS_TIME_T"s\n",
 861                                target->obd_name,
 862                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 863                                target->obd_recoverable_clients,
 864                                cfs_duration_sec(t));
 865                         rc = -EBUSY;
 866                 } else {
 867 dont_check_exports:
 868                         rc = obd_connect(req->rq_svc_thread->t_env,
 869                                          &export, target, &cluuid, data,
 870                                          client_nid);
 871                         if (rc == 0)
 872                                 conn.cookie = export->exp_handle.h_cookie;
 873                 }
 874         } else {
 875                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 876                                    export, target, &cluuid, data, client_nid);
 877                 if (rc == 0)
 878                         /* prevous done via class_conn2export */
 879                         class_export_get(export);
 880         }
 881         if (rc)
 882                 GOTO(out, rc);
 883         /* Return only the parts of obd_connect_data that we understand, so the
 884          * client knows that we don't understand the rest. */
 885         if (data) {
 886                  tmpdata = req_capsule_server_get(&req->rq_pill,
 887                                                   &RMF_CONNECT_DATA);
 888                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 889                  *tmpdata = *data;
 890         }
 891
 892         /* If all else goes well, this is our RPC return code. */
 893         req->rq_status = 0;
 894
 895         lustre_msg_set_handle(req->rq_repmsg, &conn);
 896
 897         /* If the client and the server are the same node, we will already
 898          * have an export that really points to the client's DLM export,
 899          * because we have a shared handles table.
 900          *
 901          * XXX this will go away when shaver stops sending the "connect" handle
 902          * in the real "remote handle" field of the request --phik 24 Apr 2003
 903          */
 904         if (req->rq_export != NULL)
 905                 class_export_put(req->rq_export);
 906
 907         req->rq_export = export;
 908
 909         spin_lock(&export->exp_lock);
 910         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 911                 spin_unlock(&export->exp_lock);
 912                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 913                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 914                        export->exp_conn_cnt,
 915                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 916
 917                 GOTO(out, rc = -EALREADY);
 918         }
 919         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 920
 921         /* request from liblustre?  Don't evict it for not pinging. */
 922         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 923                 export->exp_libclient = 1;
 924                 spin_unlock(&export->exp_lock);
 925
 926                 spin_lock(&target->obd_dev_lock);
 927                 list_del_init(&export->exp_obd_chain_timed);
 928                 spin_unlock(&target->obd_dev_lock);
 929         } else {
 930                 spin_unlock(&export->exp_lock);
 931         }
 932
 933         if (export->exp_connection != NULL)
 934                 ptlrpc_connection_put(export->exp_connection);
 935         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 936                                                        req->rq_self,
 937                                                        &remote_uuid);
 938         if (hlist_unhashed(&export->exp_nid_hash)) {
 939                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 940                                        &export->exp_connection->c_peer.nid,
 941                                        &export->exp_nid_hash);
 942         }
 943
 944         spin_lock_bh(&target->obd_processing_task_lock);
 945         if (target->obd_recovering && !export->exp_in_recovery) {
 946                 spin_lock(&export->exp_lock);
 947                 export->exp_in_recovery = 1;
 948                 export->exp_req_replay_needed = 1;
 949                 export->exp_lock_replay_needed = 1;
 950                 spin_unlock(&export->exp_lock);
 951                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 952                      && (data->ocd_transno == 0))
 953                         CWARN("Connect with zero transno!\n");
 954
 955                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 956                      && data->ocd_transno < target->obd_next_recovery_transno)
 957                         target->obd_next_recovery_transno = data->ocd_transno;
 958                 target->obd_connected_clients++;
 959                 /* each connected client is counted as recoverable */
 960                 target->obd_recoverable_clients++;
 961                 atomic_inc(&target->obd_req_replay_clients);
 962                 atomic_inc(&target->obd_lock_replay_clients);
 963                 if (target->obd_connected_clients ==
 964                     target->obd_max_recoverable_clients)
 965                         cfs_waitq_signal(&target->obd_next_transno_waitq);
 966         }
 967         spin_unlock_bh(&target->obd_processing_task_lock);
 968         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 969         conn = *tmp;
 970
 971         if (export->exp_imp_reverse != NULL) {
 972                 /* destroyed import can be still referenced in ctxt */
 973                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 974                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 975
 976                 /* in some recovery senarios, previous ctx init rpc handled
 977                  * in sptlrpc_target_export_check() might be used to install
 978                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 979                  * using the same gss ctx could reach here and following new
 980                  * reverse import. note all reverse ctx in new/old import are
 981                  * actually based on the same gss ctx. so we invalidate ctx
 982                  * here before destroy import, otherwise flush old import will
 983                  * lead to remote reverse ctx be destroied, thus the reverse
 984                  * ctx of new import will lost its peer.
 985                  * there might be a better way to deal with this???
 986                  */
 987                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 988
 989                 client_destroy_import(export->exp_imp_reverse);
 990         }
 991
 992         /* for the rest part, we return -ENOTCONN in case of errors
 993          * in order to let client initialize connection again.
 994          */
 995         revimp = export->exp_imp_reverse = class_new_import(target);
 996         if (!revimp) {
 997                 CERROR("fail to alloc new reverse import.\n");
 998                 GOTO(out, rc = -ENOTCONN);
 999         }
1000
1001         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1002         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1003         revimp->imp_remote_handle = conn;
1004         revimp->imp_dlm_fake = 1;
1005         revimp->imp_state = LUSTRE_IMP_FULL;
1006
1007         /* unknown versions will be caught in
1008          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1009         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1010
1011         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1012             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1013                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1014         else
1015                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1016
1017         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1018         if (rc) {
1019                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1020                 export->exp_imp_reverse = NULL;
1021                 class_destroy_import(revimp);
1022         }
1023
1024         class_import_put(revimp);
1025 out:
1026         if (export) {
1027                 spin_lock(&export->exp_lock);
1028                 export->exp_connecting = 0;
1029                 spin_unlock(&export->exp_lock);
1030         }
1031         if (targref)
1032                 class_decref(targref, __FUNCTION__, cfs_current());
1033         if (rc)
1034                 req->rq_status = rc;
1035         RETURN(rc);
1036 }
1037
1038 int target_handle_disconnect(struct ptlrpc_request *req)
1039 {
1040         int rc;
1041         ENTRY;
1042
1043         rc = req_capsule_server_pack(&req->rq_pill);
1044         if (rc)
1045                 RETURN(rc);
1046
1047         /* keep the rq_export around so we can send the reply */
1048         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1049
1050         RETURN(0);
1051 }
1052
1053 void target_destroy_export(struct obd_export *exp)
1054 {
1055         /* exports created from last_rcvd data, and "fake"
1056            exports created by lctl don't have an import */
1057         if (exp->exp_imp_reverse != NULL)
1058                 client_destroy_import(exp->exp_imp_reverse);
1059
1060         /* We cancel locks at disconnect time, but this will catch any locks
1061          * granted in a race with recovery-induced disconnect. */
1062         if (exp->exp_obd->obd_namespace != NULL)
1063                 ldlm_cancel_locks_for_export(exp);
1064 }
1065
1066 /*
1067  * Recovery functions
1068  */
1069
1070 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1071 {
1072         struct ptlrpc_request *copy_req;
1073         struct lustre_msg *copy_reqmsg;
1074         struct ptlrpc_user_desc *udesc = NULL;
1075
1076         OBD_ALLOC_PTR(copy_req);
1077         if (!copy_req)
1078                 return NULL;
1079         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1080         if (!copy_reqmsg){
1081                 OBD_FREE_PTR(copy_req);
1082                 return NULL;
1083         }
1084
1085         if (orig_req->rq_user_desc) {
1086                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1087
1088                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1089                 if (!udesc) {
1090                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1091                         OBD_FREE_PTR(copy_req);
1092                         return NULL;
1093                 }
1094                 memcpy(udesc, orig_req->rq_user_desc,
1095                        sptlrpc_user_desc_size(ngroups));
1096         }
1097
1098         *copy_req = *orig_req;
1099         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1100         copy_req->rq_reqmsg = copy_reqmsg;
1101         copy_req->rq_user_desc = udesc;
1102
1103         class_export_rpc_get(copy_req->rq_export);
1104         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1105         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1106         sptlrpc_svc_ctx_addref(copy_req);
1107
1108         if (copy_req->rq_reply_state) {
1109                 /* the copied req takes over the reply state */
1110                 orig_req->rq_reply_state = NULL;
1111                 /* to catch further access */
1112                 orig_req->rq_repmsg = NULL;
1113                 orig_req->rq_replen = 0;
1114         }
1115
1116         return copy_req;
1117 }
1118
1119 void ptlrpc_free_clone(struct ptlrpc_request *req)
1120 {
1121         LASSERT(list_empty(&req->rq_replay_list));
1122
1123         ptlrpc_req_drop_rs(req);
1124         sptlrpc_svc_ctx_decref(req);
1125         class_export_rpc_put(req->rq_export);
1126         list_del_init(&req->rq_list);
1127
1128         if (req->rq_user_desc) {
1129                 int ngroups = req->rq_user_desc->pud_ngroups;
1130                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1131         }
1132         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1133         OBD_FREE_PTR(req);
1134 }
1135
1136 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1137 {
1138         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1139         struct obd_export     *exp = req->rq_export;
1140         struct ptlrpc_request *reqiter;
1141         int                    dup = 0;
1142
1143         LASSERT(exp);
1144
1145         spin_lock(&exp->exp_lock);
1146         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1147                             rq_replay_list) {
1148                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1149                         dup = 1;
1150                         break;
1151                 }
1152         }
1153
1154         if (dup) {
1155                 /* we expect it with RESENT and REPLAY flags */
1156                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1157                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1158                         CERROR("invalid flags %x of resent replay\n",
1159                                lustre_msg_get_flags(req->rq_reqmsg));
1160         } else {
1161                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1162         }
1163
1164         spin_unlock(&exp->exp_lock);
1165         return dup;
1166 }
1167
1168 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1169 {
1170         LASSERT(!list_empty(&req->rq_replay_list));
1171         LASSERT(req->rq_export);
1172
1173         spin_lock(&req->rq_export->exp_lock);
1174         list_del_init(&req->rq_replay_list);
1175         spin_unlock(&req->rq_export->exp_lock);
1176 }
1177
1178 #ifdef __KERNEL__
1179 static void target_finish_recovery(struct obd_device *obd)
1180 {
1181         ENTRY;
1182         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1183                       obd->obd_name);
1184
1185         ldlm_reprocess_all_ns(obd->obd_namespace);
1186         spin_lock_bh(&obd->obd_processing_task_lock);
1187         if (list_empty(&obd->obd_req_replay_queue) &&
1188             list_empty(&obd->obd_lock_replay_queue) &&
1189             list_empty(&obd->obd_final_req_queue)) {
1190                 obd->obd_processing_task = 0;
1191         } else {
1192                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1193                        obd->obd_name,
1194                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1195                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1196                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1197                 spin_unlock_bh(&obd->obd_processing_task_lock);
1198                 LBUG();
1199         }
1200         spin_unlock_bh(&obd->obd_processing_task_lock);
1201
1202         /* when recovery finished, cleanup orphans on mds and ost */
1203         if (OBT(obd) && OBP(obd, postrecov)) {
1204                 int rc = OBP(obd, postrecov)(obd);
1205                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1206                               rc < 0 ? "failed" : "complete", rc);
1207         }
1208
1209         obd->obd_recovery_end = cfs_time_current_sec();
1210         EXIT;
1211 }
1212
1213 static void abort_req_replay_queue(struct obd_device *obd)
1214 {
1215         struct ptlrpc_request *req, *n;
1216         struct list_head abort_list;
1217
1218         CFS_INIT_LIST_HEAD(&abort_list);
1219         spin_lock_bh(&obd->obd_processing_task_lock);
1220         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1221         spin_unlock_bh(&obd->obd_processing_task_lock);
1222         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1223                 DEBUG_REQ(D_WARNING, req, "aborted:");
1224                 req->rq_status = -ENOTCONN;
1225                 if (ptlrpc_error(req)) {
1226                         DEBUG_REQ(D_ERROR, req,
1227                                   "failed abort_req_reply; skipping");
1228                 }
1229                 target_exp_dequeue_req_replay(req);
1230                 ptlrpc_free_clone(req);
1231         }
1232 }
1233
1234 static void abort_lock_replay_queue(struct obd_device *obd)
1235 {
1236         struct ptlrpc_request *req, *n;
1237         struct list_head abort_list;
1238
1239         CFS_INIT_LIST_HEAD(&abort_list);
1240         spin_lock_bh(&obd->obd_processing_task_lock);
1241         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1242         spin_unlock_bh(&obd->obd_processing_task_lock);
1243         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1244                 DEBUG_REQ(D_ERROR, req, "aborted:");
1245                 req->rq_status = -ENOTCONN;
1246                 if (ptlrpc_error(req)) {
1247                         DEBUG_REQ(D_ERROR, req,
1248                                   "failed abort_lock_reply; skipping");
1249                 }
1250                 ptlrpc_free_clone(req);
1251         }
1252 }
1253 #endif
1254
1255 /* Called from a cleanup function if the device is being cleaned up
1256    forcefully.  The exports should all have been disconnected already,
1257    the only thing left to do is
1258      - clear the recovery flags
1259      - cancel the timer
1260      - free queued requests and replies, but don't send replies
1261    Because the obd_stopping flag is set, no new requests should be received.
1262
1263 */
1264 void target_cleanup_recovery(struct obd_device *obd)
1265 {
1266         struct ptlrpc_request *req, *n;
1267         struct list_head clean_list;
1268         ENTRY;
1269
1270         LASSERT(obd->obd_stopping);
1271
1272         CFS_INIT_LIST_HEAD(&clean_list);
1273         spin_lock_bh(&obd->obd_processing_task_lock);
1274         if (!obd->obd_recovering) {
1275                 spin_unlock_bh(&obd->obd_processing_task_lock);
1276                 EXIT;
1277                 return;
1278         }
1279         obd->obd_recovering = obd->obd_abort_recovery = 0;
1280         target_cancel_recovery_timer(obd);
1281
1282         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1283         spin_unlock_bh(&obd->obd_processing_task_lock);
1284
1285         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1286                 LASSERT(req->rq_reply_state == 0);
1287                 target_exp_dequeue_req_replay(req);
1288                 ptlrpc_free_clone(req);
1289         }
1290
1291         spin_lock_bh(&obd->obd_processing_task_lock);
1292         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1293         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1294         spin_unlock_bh(&obd->obd_processing_task_lock);
1295
1296         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1297                 LASSERT(req->rq_reply_state == 0);
1298                 ptlrpc_free_clone(req);
1299         }
1300
1301         EXIT;
1302 }
1303
1304 /* obd_processing_task_lock should be held */
1305 void target_cancel_recovery_timer(struct obd_device *obd)
1306 {
1307         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1308         cfs_timer_disarm(&obd->obd_recovery_timer);
1309 }
1310
1311 /* extend = 1 means require at least "duration" seconds left in the timer,
1312    extend = 0 means set the total duration (start_recovery_timer) */
1313 static void reset_recovery_timer(struct obd_device *obd, int duration,
1314                                  int extend)
1315 {
1316         cfs_time_t now = cfs_time_current_sec();
1317         cfs_duration_t left;
1318
1319         spin_lock_bh(&obd->obd_processing_task_lock);
1320         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1321                 spin_unlock_bh(&obd->obd_processing_task_lock);
1322                 return;
1323         }
1324
1325         left = cfs_time_sub(obd->obd_recovery_end, now);
1326
1327         if (extend && (duration > left))
1328                 obd->obd_recovery_timeout += duration - left;
1329         else if (!extend && (duration > obd->obd_recovery_timeout))
1330                 /* Track the client's largest expected replay time */
1331                 obd->obd_recovery_timeout = duration;
1332 #ifdef CRAY_XT3
1333         /*
1334          * If total recovery time already exceed the
1335          * obd_recovery_max_time, then CRAY XT3 will
1336          * abort the recovery
1337          */
1338         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1339                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1340 #endif
1341         obd->obd_recovery_end = obd->obd_recovery_start +
1342                                 obd->obd_recovery_timeout;
1343         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1344             cfs_time_before(now, obd->obd_recovery_end)) {
1345                 left = cfs_time_sub(obd->obd_recovery_end, now);
1346                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1347         }
1348         spin_unlock_bh(&obd->obd_processing_task_lock);
1349         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1350                obd->obd_name, (unsigned)left);
1351 }
1352
1353 static void check_and_start_recovery_timer(struct obd_device *obd)
1354 {
1355         spin_lock_bh(&obd->obd_processing_task_lock);
1356         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1357                 spin_unlock_bh(&obd->obd_processing_task_lock);
1358                 return;
1359         }
1360         CWARN("%s: starting recovery timer\n", obd->obd_name);
1361         obd->obd_recovery_start = cfs_time_current_sec();
1362         /* minimum */
1363         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1364         spin_unlock_bh(&obd->obd_processing_task_lock);
1365
1366         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1367 }
1368
1369 /* Reset the timer with each new client connection */
1370 /*
1371  * This timer is actually reconnect_timer, which is for making sure
1372  * the total recovery window is at least as big as my reconnect
1373  * attempt timing. So the initial recovery time_out will be set to
1374  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1375  * from client is bigger than this, then the recovery time_out will
1376  * be extend to make sure the client could be reconnected, in the
1377  * process, the timeout from the new client should be ignored.
1378  */
1379
1380 static void
1381 target_start_and_reset_recovery_timer(struct obd_device *obd,
1382                                       struct ptlrpc_request *req,
1383                                       int new_client)
1384 {
1385         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1386
1387         if (!new_client && service_time)
1388                 /* Teach server about old server's estimates, as first guess
1389                  * at how long new requests will take. */
1390                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1391                        service_time);
1392
1393         check_and_start_recovery_timer(obd);
1394
1395         /* convert the service time to rpc timeout,
1396          * reuse service_time to limit stack usage */
1397         service_time = at_est2timeout(service_time);
1398
1399         /* We expect other clients to timeout within service_time, then try
1400          * to reconnect, then try the failover server.  The max delay between
1401          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1402         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1403                              INITIAL_CONNECT_TIMEOUT);
1404         if (service_time > obd->obd_recovery_timeout && !new_client)
1405                 reset_recovery_timer(obd, service_time, 0);
1406 }
1407
1408 #ifdef __KERNEL__
1409 static int check_for_next_transno(struct obd_device *obd)
1410 {
1411         struct ptlrpc_request *req = NULL;
1412         int wake_up = 0, connected, completed, queue_len, max;
1413         __u64 next_transno, req_transno;
1414         ENTRY;
1415         spin_lock_bh(&obd->obd_processing_task_lock);
1416
1417         if (!list_empty(&obd->obd_req_replay_queue)) {
1418                 req = list_entry(obd->obd_req_replay_queue.next,
1419                                  struct ptlrpc_request, rq_list);
1420                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1421         } else {
1422                 req_transno = 0;
1423         }
1424
1425         max = obd->obd_max_recoverable_clients;
1426         connected = obd->obd_connected_clients;
1427         completed = connected - obd->obd_recoverable_clients;
1428         queue_len = obd->obd_requests_queued_for_recovery;
1429         next_transno = obd->obd_next_recovery_transno;
1430
1431         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1432                "req_transno: "LPU64", next_transno: "LPU64"\n",
1433                max, connected, completed, queue_len, req_transno, next_transno);
1434
1435         if (obd->obd_abort_recovery) {
1436                 CDEBUG(D_HA, "waking for aborted recovery\n");
1437                 wake_up = 1;
1438         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1439                 CDEBUG(D_HA, "waking for completed recovery\n");
1440                 wake_up = 1;
1441         } else if (req_transno == next_transno) {
1442                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1443                 wake_up = 1;
1444         } else if (queue_len + completed == max) {
1445                 /* handle gaps occured due to lost reply. It is allowed gaps
1446                  * because all clients are connected and there will be resend
1447                  * for missed transaction */
1448                 LASSERTF(req_transno >= next_transno,
1449                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1450                          req_transno, next_transno);
1451
1452                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1453                        "waking for skipped transno (skip: "LPD64
1454                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1455                        next_transno, queue_len, completed, connected, req_transno);
1456                 obd->obd_next_recovery_transno = req_transno;
1457                 wake_up = 1;
1458         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1459                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1460                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1461                 obd->obd_next_recovery_transno = req_transno;
1462                 wake_up = 1;
1463         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1464                 /* some clients haven't connected in time, but we can try
1465                  * to replay requests that demand on already committed ones
1466                  * also, we can replay first non-committed transation */
1467                 LASSERT(req_transno != 0);
1468                 if (obd->obd_version_recov ||
1469                     req_transno == obd->obd_last_committed + 1) {
1470                         obd->obd_next_recovery_transno = req_transno;
1471                 } else if (req_transno > obd->obd_last_committed) {
1472                         /* can't continue recovery: have no needed transno */
1473                         obd->obd_abort_recovery = 1;
1474                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1475                                "connected: %d, completed: %d, queue_len: %d, "
1476                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1477                                max, connected, completed, queue_len,
1478                                req_transno, next_transno);
1479                 }
1480                 wake_up = 1;
1481         }
1482
1483         spin_unlock_bh(&obd->obd_processing_task_lock);
1484         return wake_up;
1485 }
1486
1487 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1488 {
1489         struct l_wait_info lwi = { 0 };
1490         struct ptlrpc_request *req;
1491
1492         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1493                obd->obd_next_recovery_transno);
1494         l_wait_event(obd->obd_next_transno_waitq,
1495                      check_for_next_transno(obd), &lwi);
1496
1497         spin_lock_bh(&obd->obd_processing_task_lock);
1498         if (obd->obd_abort_recovery) {
1499                 req = NULL;
1500         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1501                 req = list_entry(obd->obd_req_replay_queue.next,
1502                                  struct ptlrpc_request, rq_list);
1503                 target_exp_dequeue_req_replay(req);
1504                 list_del_init(&req->rq_list);
1505                 obd->obd_requests_queued_for_recovery--;
1506         } else {
1507                 req = NULL;
1508         }
1509         spin_unlock_bh(&obd->obd_processing_task_lock);
1510         RETURN(req);
1511 }
1512
1513 static int check_for_next_lock(struct obd_device *obd)
1514 {
1515         struct ptlrpc_request *req = NULL;
1516         int wake_up = 0;
1517
1518         spin_lock_bh(&obd->obd_processing_task_lock);
1519         if (!list_empty(&obd->obd_lock_replay_queue)) {
1520                 req = list_entry(obd->obd_lock_replay_queue.next,
1521                                  struct ptlrpc_request, rq_list);
1522                 CDEBUG(D_HA, "waking for next lock\n");
1523                 wake_up = 1;
1524         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1525                 CDEBUG(D_HA, "waking for completed lock replay\n");
1526                 wake_up = 1;
1527         } else if (obd->obd_abort_recovery) {
1528                 CDEBUG(D_HA, "waking for aborted recovery\n");
1529                 wake_up = 1;
1530         }
1531         spin_unlock_bh(&obd->obd_processing_task_lock);
1532
1533         return wake_up;
1534 }
1535
1536 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1537 {
1538         struct l_wait_info lwi = { 0 };
1539         struct ptlrpc_request *req;
1540
1541         CDEBUG(D_HA, "Waiting for lock\n");
1542         l_wait_event(obd->obd_next_transno_waitq,
1543                      check_for_next_lock(obd), &lwi);
1544
1545         spin_lock_bh(&obd->obd_processing_task_lock);
1546         if (obd->obd_abort_recovery) {
1547                 req = NULL;
1548         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1549                 req = list_entry(obd->obd_lock_replay_queue.next,
1550                                  struct ptlrpc_request, rq_list);
1551                 list_del_init(&req->rq_list);
1552         } else {
1553                 req = NULL;
1554         }
1555         spin_unlock_bh(&obd->obd_processing_task_lock);
1556         return req;
1557 }
1558
1559 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1560 {
1561         struct ptlrpc_request *req;
1562
1563         spin_lock_bh(&obd->obd_processing_task_lock);
1564         if (!list_empty(&obd->obd_final_req_queue)) {
1565                 req = list_entry(obd->obd_final_req_queue.next,
1566                                  struct ptlrpc_request, rq_list);
1567                 list_del_init(&req->rq_list);
1568                 if (req->rq_export->exp_in_recovery) {
1569                         spin_lock(&req->rq_export->exp_lock);
1570                         req->rq_export->exp_in_recovery = 0;
1571                         spin_unlock(&req->rq_export->exp_lock);
1572                 }
1573         } else {
1574                 req = NULL;
1575         }
1576         spin_unlock_bh(&obd->obd_processing_task_lock);
1577         return req;
1578 }
1579
1580 static inline int req_vbr_done(struct obd_export *exp)
1581 {
1582         return (exp->exp_vbr_failed == 0);
1583 }
1584
1585 static inline int req_replay_done(struct obd_export *exp)
1586 {
1587         return (exp->exp_req_replay_needed == 0);
1588 }
1589
1590 static inline int lock_replay_done(struct obd_export *exp)
1591 {
1592         return (exp->exp_lock_replay_needed == 0);
1593 }
1594
1595 static inline int connect_done(struct obd_export *exp)
1596 {
1597         return (exp->exp_in_recovery != 0);
1598 }
1599
1600 static int check_for_clients(struct obd_device *obd)
1601 {
1602         if (obd->obd_abort_recovery || obd->obd_version_recov)
1603                 return 1;
1604         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1605         if (obd->obd_no_conn == 0 &&
1606             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1607                 return 1;
1608         return 0;
1609 }
1610
1611 static int handle_recovery_req(struct ptlrpc_thread *thread,
1612                                struct ptlrpc_request *req,
1613                                svc_handler_t handler)
1614 {
1615         int rc;
1616         ENTRY;
1617
1618         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1619         if (rc) {
1620                 CERROR("Failure to initialize session: %d\n", rc);
1621                 return rc;
1622         }
1623         req->rq_session.lc_thread = thread;
1624         lu_context_enter(&req->rq_session);
1625         req->rq_svc_thread = thread;
1626         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1627
1628         /* thread context */
1629         lu_context_enter(&thread->t_env->le_ctx);
1630         (void)handler(req);
1631         lu_context_exit(&thread->t_env->le_ctx);
1632
1633         lu_context_exit(&req->rq_session);
1634         lu_context_fini(&req->rq_session);
1635         /* don't reset timer for final stage */
1636         if (!req_replay_done(req->rq_export) ||
1637             !lock_replay_done(req->rq_export))
1638                 reset_recovery_timer(class_exp2obd(req->rq_export),
1639                        AT_OFF ? obd_timeout :
1640                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1641         ptlrpc_free_clone(req);
1642         RETURN(0);
1643 }
1644
1645 static void resume_recovery_timer(struct obd_device *obd)
1646 {
1647         /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1648         reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1649 }
1650
1651 static int target_recovery_thread(void *arg)
1652 {
1653         struct lu_target *lut = arg;
1654         struct obd_device *obd = lut->lut_obd;
1655         struct ptlrpc_request *req;
1656         struct target_recovery_data *trd = &obd->obd_recovery_data;
1657         struct l_wait_info lwi = { 0 };
1658         unsigned long delta;
1659         unsigned long flags;
1660         struct lu_env env;
1661         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1662         int rc = 0;
1663         ENTRY;
1664
1665         cfs_daemonize("tgt_recov");
1666
1667         SIGNAL_MASK_LOCK(current, flags);
1668         sigfillset(&current->blocked);
1669         RECALC_SIGPENDING;
1670         SIGNAL_MASK_UNLOCK(current, flags);
1671
1672         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1673         if (rc)
1674                 RETURN(rc);
1675
1676         thread->t_env = &env;
1677         env.le_ctx.lc_thread = thread;
1678
1679         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1680                cfs_curproc_pid());
1681         trd->trd_processing_task = cfs_curproc_pid();
1682
1683         obd->obd_recovering = 1;
1684         complete(&trd->trd_starting);
1685
1686         /* first of all, we have to know the first transno to replay */
1687         obd->obd_abort_recovery = 0;
1688         l_wait_event(obd->obd_next_transno_waitq,
1689                      check_for_clients(obd), &lwi);
1690
1691         spin_lock_bh(&obd->obd_processing_task_lock);
1692         target_cancel_recovery_timer(obd);
1693         spin_unlock_bh(&obd->obd_processing_task_lock);
1694
1695         /* If some clients haven't connected in time, evict them */
1696         if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) {
1697                 CWARN("Some clients haven't connect in time (%d/%d),"
1698                        "evict them\n", obd->obd_connected_clients,
1699                        obd->obd_max_recoverable_clients);
1700                 class_disconnect_stale_exports(obd, connect_done,
1701                                                exp_flags_from_obd(obd) |
1702                                                OBD_OPT_ABORT_RECOV);
1703                 /**
1704                  * if recovery proceeds with versions then some clients may be
1705                  * timed out waiting for others and trying to reconnect.
1706                  * Extend timer for such reconnect cases.
1707                  */
1708                 if (obd->obd_version_recov)
1709                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1);
1710         }
1711
1712         /* next stage: replay requests */
1713         delta = jiffies;
1714         obd->obd_req_replaying = 1;
1715         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1716                atomic_read(&obd->obd_req_replay_clients),
1717                obd->obd_next_recovery_transno);
1718         resume_recovery_timer(obd);
1719         while ((req = target_next_replay_req(obd))) {
1720                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1721                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1722                           lustre_msg_get_transno(req->rq_reqmsg),
1723                           libcfs_nid2str(req->rq_peer.nid));
1724                 handle_recovery_req(thread, req,
1725                                     trd->trd_recovery_handler);
1726                 obd->obd_replayed_requests++;
1727                 spin_lock_bh(&obd->obd_processing_task_lock);
1728                 obd->obd_next_recovery_transno++;
1729                 spin_unlock_bh(&obd->obd_processing_task_lock);
1730         }
1731
1732         spin_lock_bh(&obd->obd_processing_task_lock);
1733         target_cancel_recovery_timer(obd);
1734         spin_unlock_bh(&obd->obd_processing_task_lock);
1735
1736         /* If some clients haven't replayed requests in time, evict them */
1737         if (obd->obd_abort_recovery) {
1738                 CDEBUG(D_WARNING, "req replay is aborted\n");
1739                 class_disconnect_stale_exports(obd, req_replay_done,
1740                                                exp_flags_from_obd(obd) |
1741                                                OBD_OPT_ABORT_RECOV);
1742                 abort_req_replay_queue(obd);
1743         }
1744         LASSERT(list_empty(&obd->obd_req_replay_queue));
1745
1746         /* The second stage: replay locks */
1747         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1748                atomic_read(&obd->obd_lock_replay_clients));
1749         resume_recovery_timer(obd);
1750         while ((req = target_next_replay_lock(obd))) {
1751                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1752                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1753                           libcfs_nid2str(req->rq_peer.nid));
1754                 handle_recovery_req(thread, req,
1755                                     trd->trd_recovery_handler);
1756                 obd->obd_replayed_locks++;
1757         }
1758
1759         spin_lock_bh(&obd->obd_processing_task_lock);
1760         target_cancel_recovery_timer(obd);
1761         spin_unlock_bh(&obd->obd_processing_task_lock);
1762         /* If some clients haven't replayed requests in time, evict them */
1763         if (obd->obd_abort_recovery) {
1764                 int stale;
1765                 CERROR("lock replay is aborted\n");
1766                 stale = class_disconnect_stale_exports(obd, lock_replay_done,
1767                                                        exp_flags_from_obd(obd) |
1768                                                        OBD_OPT_ABORT_RECOV);
1769                 abort_lock_replay_queue(obd);
1770         }
1771         LASSERT(list_empty(&obd->obd_lock_replay_queue));
1772
1773         /* The third stage: reply on final pings */
1774         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1775         /** evict exports failed VBR */
1776         class_disconnect_stale_exports(obd, req_vbr_done,
1777                                        exp_flags_from_obd(obd) |
1778                                        OBD_OPT_ABORT_RECOV);
1779         /** Update server last boot epoch */
1780         lut_boot_epoch_update(lut);
1781         /* We drop recoverying flag to forward all new requests
1782          * to regular mds_handle() since now */
1783         spin_lock_bh(&obd->obd_processing_task_lock);
1784         obd->obd_recovering = obd->obd_abort_recovery = 0;
1785         spin_unlock_bh(&obd->obd_processing_task_lock);
1786         while ((req = target_next_final_ping(obd))) {
1787                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1788                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1789                           libcfs_nid2str(req->rq_peer.nid));
1790                 handle_recovery_req(thread, req,
1791                                     trd->trd_recovery_handler);
1792         }
1793
1794         delta = (jiffies - delta) / HZ;
1795         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1796               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1797         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1798         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1799         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1800                 CWARN("too long recovery - read logs\n");
1801                 libcfs_debug_dumplog();
1802         }
1803
1804         target_finish_recovery(obd);
1805
1806         lu_context_fini(&env.le_ctx);
1807         trd->trd_processing_task = 0;
1808         complete(&trd->trd_finishing);
1809         RETURN(rc);
1810 }
1811
1812 static int target_start_recovery_thread(struct lu_target *lut,
1813                                         svc_handler_t handler)
1814 {
1815         struct obd_device *obd = lut->lut_obd;
1816         int rc = 0;
1817         struct target_recovery_data *trd = &obd->obd_recovery_data;
1818
1819         memset(trd, 0, sizeof(*trd));
1820         init_completion(&trd->trd_starting);
1821         init_completion(&trd->trd_finishing);
1822         trd->trd_recovery_handler = handler;
1823
1824         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1825                 wait_for_completion(&trd->trd_starting);
1826                 LASSERT(obd->obd_recovering != 0);
1827         } else
1828                 rc = -ECHILD;
1829
1830         return rc;
1831 }
1832
1833 void target_stop_recovery_thread(struct obd_device *obd)
1834 {
1835         spin_lock_bh(&obd->obd_processing_task_lock);
1836         if (obd->obd_recovery_data.trd_processing_task > 0) {
1837                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1838                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1839                 obd->obd_abort_recovery = 1;
1840                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1841                 spin_unlock_bh(&obd->obd_processing_task_lock);
1842                 wait_for_completion(&trd->trd_finishing);
1843         } else {
1844                 spin_unlock_bh(&obd->obd_processing_task_lock);
1845         }
1846 }
1847
1848 void target_recovery_fini(struct obd_device *obd)
1849 {
1850         class_disconnect_exports(obd);
1851         target_stop_recovery_thread(obd);
1852         target_cleanup_recovery(obd);
1853 }
1854 EXPORT_SYMBOL(target_recovery_fini);
1855
1856 static void target_recovery_expired(unsigned long castmeharder)
1857 {
1858         struct obd_device *obd = (struct obd_device *)castmeharder;
1859         LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1860                       "after %lds (%d clients did)\n",
1861                       obd->obd_name, obd->obd_recoverable_clients,
1862                       cfs_time_current_sec()- obd->obd_recovery_start,
1863                       obd->obd_connected_clients);
1864
1865         spin_lock_bh(&obd->obd_processing_task_lock);
1866         obd->obd_version_recov = 1;
1867         CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n",
1868                atomic_read(&obd->obd_req_replay_clients),
1869                obd->obd_next_recovery_transno);
1870         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1871         spin_unlock_bh(&obd->obd_processing_task_lock);
1872 }
1873
1874 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1875 {
1876         struct obd_device *obd = lut->lut_obd;
1877         if (obd->obd_max_recoverable_clients == 0)
1878                 return;
1879
1880         CWARN("RECOVERY: service %s, %d recoverable clients, "
1881               "last_transno "LPU64"\n", obd->obd_name,
1882               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1883         LASSERT(obd->obd_stopping == 0);
1884         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1885         obd->obd_recovery_start = 0;
1886         obd->obd_recovery_end = 0;
1887         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1888         /* bz13079: this should be set to desired value for ost but not for mds */
1889         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1890         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1891         target_start_recovery_thread(lut, handler);
1892 }
1893 EXPORT_SYMBOL(target_recovery_init);
1894
1895 #endif
1896
1897 static int target_process_req_flags(struct obd_device *obd,
1898                                     struct ptlrpc_request *req)
1899 {
1900         struct obd_export *exp = req->rq_export;
1901         LASSERT(exp != NULL);
1902         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1903                 /* client declares he's ready to replay locks */
1904                 spin_lock_bh(&obd->obd_processing_task_lock);
1905                 if (exp->exp_req_replay_needed) {
1906                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1907                         spin_lock(&exp->exp_lock);
1908                         exp->exp_req_replay_needed = 0;
1909                         spin_unlock(&exp->exp_lock);
1910                         atomic_dec(&obd->obd_req_replay_clients);
1911                         LASSERT(obd->obd_recoverable_clients > 0);
1912                         obd->obd_recoverable_clients--;
1913                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1914                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1915                 }
1916                 spin_unlock_bh(&obd->obd_processing_task_lock);
1917         }
1918         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1919                 /* client declares he's ready to complete recovery
1920                  * so, we put the request on th final queue */
1921                 spin_lock_bh(&obd->obd_processing_task_lock);
1922                 if (exp->exp_lock_replay_needed) {
1923                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1924                         spin_lock(&exp->exp_lock);
1925                         exp->exp_lock_replay_needed = 0;
1926                         spin_unlock(&exp->exp_lock);
1927                         atomic_dec(&obd->obd_lock_replay_clients);
1928                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1929                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1930                 }
1931                 spin_unlock_bh(&obd->obd_processing_task_lock);
1932         }
1933
1934         return 0;
1935 }
1936
1937 int target_queue_recovery_request(struct ptlrpc_request *req,
1938                                   struct obd_device *obd)
1939 {
1940         struct list_head *tmp;
1941         int inserted = 0;
1942         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1943         ENTRY;
1944
1945         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1946                 /* Processing the queue right now, don't re-add. */
1947                 RETURN(1);
1948         }
1949
1950         target_process_req_flags(obd, req);
1951
1952         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1953                 /* client declares he's ready to complete recovery
1954                  * so, we put the request on th final queue */
1955                 req = ptlrpc_clone_req(req);
1956                 if (req == NULL)
1957                         RETURN(-ENOMEM);
1958                 DEBUG_REQ(D_HA, req, "queue final req");
1959                 spin_lock_bh(&obd->obd_processing_task_lock);
1960                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1961                 if (obd->obd_recovering)
1962                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1963                 else {
1964                         spin_unlock_bh(&obd->obd_processing_task_lock);
1965                         ptlrpc_free_clone(req);
1966                         if (obd->obd_stopping) {
1967                                 RETURN(-ENOTCONN);
1968                         } else {
1969                                 RETURN(1);
1970                         }
1971                 }
1972                 spin_unlock_bh(&obd->obd_processing_task_lock);
1973                 RETURN(0);
1974         }
1975         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1976                 /* client declares he's ready to replay locks */
1977                 req = ptlrpc_clone_req(req);
1978                 if (req == NULL)
1979                         RETURN(-ENOMEM);
1980                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1981                 spin_lock_bh(&obd->obd_processing_task_lock);
1982                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1983                 LASSERT(obd->obd_recovering);
1984                 /* usually due to recovery abort */
1985                 if (!req->rq_export->exp_in_recovery) {
1986                         spin_unlock_bh(&obd->obd_processing_task_lock);
1987                         ptlrpc_free_clone(req);
1988                         RETURN(-ENOTCONN);
1989                 }
1990                 LASSERT(req->rq_export->exp_lock_replay_needed);
1991                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1992                 spin_unlock_bh(&obd->obd_processing_task_lock);
1993                 RETURN(0);
1994         }
1995
1996         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1997          * (i.e. buflens etc are in my own byte order), but type-dependent
1998          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1999
2000         if (!transno) {
2001                 CFS_INIT_LIST_HEAD(&req->rq_list);
2002                 DEBUG_REQ(D_HA, req, "not queueing");
2003                 RETURN(1);
2004         }
2005
2006         spin_lock_bh(&obd->obd_processing_task_lock);
2007
2008         /* If we're processing the queue, we want don't want to queue this
2009          * message.
2010          *
2011          * Also, if this request has a transno less than the one we're waiting
2012          * for, we should process it now.  It could (and currently always will)
2013          * be an open request for a descriptor that was opened some time ago.
2014          *
2015          * Also, a resent, replayed request that has already been
2016          * handled will pass through here and be processed immediately.
2017          */
2018         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2019               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2020         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2021                 /* Processing the queue right now, don't re-add. */
2022                 LASSERT(list_empty(&req->rq_list));
2023                 spin_unlock_bh(&obd->obd_processing_task_lock);
2024                 RETURN(1);
2025         }
2026         spin_unlock_bh(&obd->obd_processing_task_lock);
2027
2028         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2029                 RETURN(0);
2030
2031         req = ptlrpc_clone_req(req);
2032         if (req == NULL)
2033                 RETURN(-ENOMEM);
2034
2035         spin_lock_bh(&obd->obd_processing_task_lock);
2036         LASSERT(obd->obd_recovering);
2037         if (!req->rq_export->exp_in_recovery) {
2038                 spin_unlock_bh(&obd->obd_processing_task_lock);
2039                 ptlrpc_free_clone(req);
2040                 RETURN(-ENOTCONN);
2041         }
2042         LASSERT(req->rq_export->exp_req_replay_needed);
2043
2044         if (target_exp_enqueue_req_replay(req)) {
2045                 spin_unlock_bh(&obd->obd_processing_task_lock);
2046                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2047                 ptlrpc_free_clone(req);
2048                 RETURN(0);
2049         }
2050
2051         /* XXX O(n^2) */
2052         list_for_each(tmp, &obd->obd_req_replay_queue) {
2053                 struct ptlrpc_request *reqiter =
2054                         list_entry(tmp, struct ptlrpc_request, rq_list);
2055
2056                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2057                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2058                         inserted = 1;
2059                         break;
2060                 }
2061
2062                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2063                              transno)) {
2064                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2065                                   "has been claimed by another client");
2066                         spin_unlock_bh(&obd->obd_processing_task_lock);
2067                         target_exp_dequeue_req_replay(req);
2068                         ptlrpc_free_clone(req);
2069                         RETURN(0);
2070                 }
2071         }
2072
2073         if (!inserted)
2074                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2075
2076         obd->obd_requests_queued_for_recovery++;
2077         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2078         spin_unlock_bh(&obd->obd_processing_task_lock);
2079         RETURN(0);
2080 }
2081
2082 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2083 {
2084         return req->rq_export->exp_obd;
2085 }
2086
2087 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2088 {
2089         LASSERT(exp != NULL);
2090         return &exp->exp_obd->obd_namespace->ns_pool;
2091 }
2092
2093 /**
2094  * Packs current SLV and Limit into \a req.
2095  */
2096 int target_pack_pool_reply(struct ptlrpc_request *req)
2097 {
2098         struct obd_device *obd;
2099         ENTRY;
2100
2101         /*
2102          * Check that we still have all structures alive as this may
2103          * be some late rpc in shutdown time.
2104          */
2105         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2106                      !exp_connect_lru_resize(req->rq_export))) {
2107                 lustre_msg_set_slv(req->rq_repmsg, 0);
2108                 lustre_msg_set_limit(req->rq_repmsg, 0);
2109                 RETURN(0);
2110         }
2111
2112         /*
2113          * OBD is alive here as export is alive, which we checked above.
2114          */
2115         obd = req->rq_export->exp_obd;
2116
2117         read_lock(&obd->obd_pool_lock);
2118         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2119         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2120         read_unlock(&obd->obd_pool_lock);
2121
2122         RETURN(0);
2123 }
2124
2125 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2126 {
2127         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2128                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2129                 return (-ECOMM);
2130         }
2131
2132         if (unlikely(rc)) {
2133                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2134                 req->rq_status = rc;
2135                 return (ptlrpc_send_error(req, 1));
2136         } else {
2137                 DEBUG_REQ(D_NET, req, "sending reply");
2138         }
2139
2140         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2141 }
2142
2143 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2144 {
2145         int                        netrc;
2146         struct ptlrpc_reply_state *rs;
2147         struct obd_device         *obd;
2148         struct obd_export         *exp;
2149         struct ptlrpc_service     *svc;
2150         ENTRY;
2151
2152         if (req->rq_no_reply) {
2153                 EXIT;
2154                 return;
2155         }
2156
2157         svc = req->rq_rqbd->rqbd_service;
2158         rs = req->rq_reply_state;
2159         if (rs == NULL || !rs->rs_difficult) {
2160                 /* no notifiers */
2161                 target_send_reply_msg (req, rc, fail_id);
2162                 EXIT;
2163                 return;
2164         }
2165
2166         /* must be an export if locks saved */
2167         LASSERT (req->rq_export != NULL);
2168         /* req/reply consistent */
2169         LASSERT (rs->rs_service == svc);
2170
2171         /* "fresh" reply */
2172         LASSERT (!rs->rs_scheduled);
2173         LASSERT (!rs->rs_scheduled_ever);
2174         LASSERT (!rs->rs_handled);
2175         LASSERT (!rs->rs_on_net);
2176         LASSERT (rs->rs_export == NULL);
2177         LASSERT (list_empty(&rs->rs_obd_list));
2178         LASSERT (list_empty(&rs->rs_exp_list));
2179
2180         exp = class_export_get (req->rq_export);
2181         obd = exp->exp_obd;
2182
2183         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2184         rs->rs_scheduled = 1;
2185         rs->rs_on_net    = 1;
2186         rs->rs_xid       = req->rq_xid;
2187         rs->rs_transno   = req->rq_transno;
2188         rs->rs_export    = exp;
2189
2190         spin_lock(&exp->exp_uncommitted_replies_lock);
2191         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2192                rs->rs_transno, exp->exp_last_committed);
2193         if (rs->rs_transno > exp->exp_last_committed) {
2194                 /* not committed already */
2195                 list_add_tail(&rs->rs_obd_list,
2196                               &exp->exp_uncommitted_replies);
2197         }
2198         spin_unlock (&exp->exp_uncommitted_replies_lock);
2199
2200         spin_lock(&exp->exp_lock);
2201         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2202         spin_unlock(&exp->exp_lock);
2203
2204         netrc = target_send_reply_msg (req, rc, fail_id);
2205
2206         spin_lock(&svc->srv_lock);
2207
2208         atomic_inc(&svc->srv_n_difficult_replies);
2209
2210         if (netrc != 0) {
2211                 /* error sending: reply is off the net.  Also we need +1
2212                  * reply ref until ptlrpc_server_handle_reply() is done
2213                  * with the reply state (if the send was successful, there
2214                  * would have been +1 ref for the net, which
2215                  * reply_out_callback leaves alone) */
2216                 rs->rs_on_net = 0;
2217                 ptlrpc_rs_addref(rs);
2218                 atomic_inc (&svc->srv_outstanding_replies);
2219         }
2220
2221         spin_lock(&rs->rs_lock);
2222         if (rs->rs_transno <= exp->exp_last_committed ||
2223             (!rs->rs_on_net && !rs->rs_no_ack) ||
2224              list_empty(&rs->rs_exp_list) ||     /* completed already */
2225              list_empty(&rs->rs_obd_list)) {
2226                 CDEBUG(D_HA, "Schedule reply immediately\n");
2227                 ptlrpc_dispatch_difficult_reply(rs);
2228         } else {
2229                 list_add (&rs->rs_list, &svc->srv_active_replies);
2230                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2231         }
2232         spin_unlock(&rs->rs_lock);
2233         spin_unlock(&svc->srv_lock);
2234         EXIT;
2235 }
2236
2237 int target_handle_ping(struct ptlrpc_request *req)
2238 {
2239         obd_ping(req->rq_export);
2240         return req_capsule_server_pack(&req->rq_pill);
2241 }
2242
2243 void target_committed_to_req(struct ptlrpc_request *req)
2244 {
2245         struct obd_export *exp = req->rq_export;
2246
2247         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2248                 lustre_msg_set_last_committed(req->rq_repmsg,
2249                                               exp->exp_last_committed);
2250         else
2251                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2252                           "%d)", exp->exp_obd->obd_no_transno,
2253                           req->rq_repmsg == NULL);
2254
2255         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2256                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2257 }
2258 EXPORT_SYMBOL(target_committed_to_req);
2259
2260 int target_handle_qc_callback(struct ptlrpc_request *req)
2261 {
2262         struct obd_quotactl *oqctl;
2263         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2264
2265         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2266         if (oqctl == NULL) {
2267                 CERROR("Can't unpack obd_quotactl\n");
2268                 RETURN(-EPROTO);
2269         }
2270
2271         cli->cl_qchk_stat = oqctl->qc_stat;
2272
2273         return 0;
2274 }
2275
2276 #ifdef HAVE_QUOTA_SUPPORT
2277 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2278 {
2279 #ifdef __KERNEL__
2280         struct obd_device *obd = req->rq_export->exp_obd;
2281         struct obd_device *master_obd;
2282         struct obd_device_target *obt;
2283         struct lustre_quota_ctxt *qctxt;
2284         struct qunit_data *qdata = NULL;
2285         int rc = 0;
2286         ENTRY;
2287
2288         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2289                 RETURN(rc);
2290
2291         rc = req_capsule_server_pack(&req->rq_pill);
2292         if (rc) {
2293                 CERROR("packing reply failed!: rc = %d\n", rc);
2294                 RETURN(rc);
2295         }
2296
2297         LASSERT(req->rq_export);
2298
2299         OBD_ALLOC(qdata, sizeof(struct qunit_data));
2300         if (!qdata)
2301                 RETURN(-ENOMEM);
2302         rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2303         if (rc < 0) {
2304                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2305                 GOTO(out, rc);
2306         }
2307
2308         /* we use the observer */
2309         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2310                 CERROR("Can't find the observer, it is recovering\n");
2311                 req->rq_status = -EAGAIN;
2312                 GOTO(send_reply, rc = -EAGAIN);
2313         }
2314
2315         master_obd = obd->obd_observer->obd_observer;
2316         obt = &master_obd->u.obt;
2317         qctxt = &obt->obt_qctxt;
2318
2319         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2320                 /* quota_type has not been processed yet, return EAGAIN
2321                  * until we know whether or not quotas are supposed to
2322                  * be enabled */
2323                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2324                        "-EAGAIN\n");
2325                 req->rq_status = -EAGAIN;
2326                 rc = ptlrpc_reply(req);
2327                 GOTO(out, rc);
2328         }
2329
2330         down_read(&obt->obt_rwsem);
2331         if (qctxt->lqc_lqs_hash == NULL) {
2332                 up_read(&obt->obt_rwsem);
2333                 /* quota_type has not been processed yet, return EAGAIN
2334                  * until we know whether or not quotas are supposed to
2335                  * be enabled */
2336                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2337                        "-EAGAIN\n");
2338                 req->rq_status = -EAGAIN;
2339                 rc = ptlrpc_reply(req);
2340                 GOTO(out, rc);
2341         }
2342
2343         LASSERT(qctxt->lqc_handler);
2344         rc = qctxt->lqc_handler(master_obd, qdata,
2345                                 lustre_msg_get_opc(req->rq_reqmsg));
2346         up_read(&obt->obt_rwsem);
2347         if (rc && rc != -EDQUOT)
2348                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2349                        "dqacq failed! (rc:%d)\n", rc);
2350         req->rq_status = rc;
2351
2352         /* there are three forms of qunit(historic causes), so we need to
2353          * adjust the same form to different forms slaves needed */
2354         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2355         if (rc < 0) {
2356                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2357                 GOTO(out, rc);
2358         }
2359
2360         /* Block the quota req. b=14840 */
2361         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2362 send_reply:
2363         rc = ptlrpc_reply(req);
2364 out:
2365         OBD_FREE(qdata, sizeof(struct qunit_data));
2366         RETURN(rc);
2367 #else
2368         return 0;
2369 #endif /* !__KERNEL__ */
2370 }
2371 #endif /* HAVE_QUOTA_SUPPORT */
2372
2373 ldlm_mode_t lck_compat_array[] = {
2374         [LCK_EX] LCK_COMPAT_EX,
2375         [LCK_PW] LCK_COMPAT_PW,
2376         [LCK_PR] LCK_COMPAT_PR,
2377         [LCK_CW] LCK_COMPAT_CW,
2378         [LCK_CR] LCK_COMPAT_CR,
2379         [LCK_NL] LCK_COMPAT_NL,
2380         [LCK_GROUP] LCK_COMPAT_GROUP,
2381         [LCK_COS] LCK_COMPAT_COS,
2382 };
2383
2384 /**
2385  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2386  * not escape to the user level.
2387  */
2388 int ldlm_error2errno(ldlm_error_t error)
2389 {
2390         int result;
2391
2392         switch (error) {
2393         case ELDLM_OK:
2394                 result = 0;
2395                 break;
2396         case ELDLM_LOCK_CHANGED:
2397                 result = -ESTALE;
2398                 break;
2399         case ELDLM_LOCK_ABORTED:
2400                 result = -ENAVAIL;
2401                 break;
2402         case ELDLM_LOCK_REPLACED:
2403                 result = -ESRCH;
2404                 break;
2405         case ELDLM_NO_LOCK_DATA:
2406                 result = -ENOENT;
2407                 break;
2408         case ELDLM_NAMESPACE_EXISTS:
2409                 result = -EEXIST;
2410                 break;
2411         case ELDLM_BAD_NAMESPACE:
2412                 result = -EBADF;
2413                 break;
2414         default:
2415                 if (((int)error) < 0)  /* cast to signed type */
2416                         result = error; /* as ldlm_error_t can be unsigned */
2417                 else {
2418                         CERROR("Invalid DLM result code: %i\n", error);
2419                         result = -EPROTO;
2420                 }
2421         }
2422         return result;
2423 }
2424 EXPORT_SYMBOL(ldlm_error2errno);
2425
2426 /**
2427  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2428  */
2429 ldlm_error_t ldlm_errno2error(int err_no)
2430 {
2431         int error;
2432
2433         switch (err_no) {
2434         case 0:
2435                 error = ELDLM_OK;
2436                 break;
2437         case -ESTALE:
2438                 error = ELDLM_LOCK_CHANGED;
2439                 break;
2440         case -ENAVAIL:
2441                 error = ELDLM_LOCK_ABORTED;
2442                 break;
2443         case -ESRCH:
2444                 error = ELDLM_LOCK_REPLACED;
2445                 break;
2446         case -ENOENT:
2447                 error = ELDLM_NO_LOCK_DATA;
2448                 break;
2449         case -EEXIST:
2450                 error = ELDLM_NAMESPACE_EXISTS;
2451                 break;
2452         case -EBADF:
2453                 error = ELDLM_BAD_NAMESPACE;
2454                 break;
2455         default:
2456                 error = err_no;
2457         }
2458         return error;
2459 }
2460 EXPORT_SYMBOL(ldlm_errno2error);
2461