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         if (cli->cl_conn_count > 0 )
 404                 GOTO(out_sem, rc = -EALREADY);
 405
 406         rc = class_connect(&conn, obd, cluuid);
 407         if (rc)
 408                 GOTO(out_sem, rc);
 409
 410         cli->cl_conn_count++;
 411         *exp = class_conn2export(&conn);
 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         /* this is a bit of a layering violation, but much less risk than
1007           * changing this very complex and race-prone code.  bug=16839 */
1008         if (data->ocd_connect_flags & OBD_CONNECT_MDS)
1009                 obd_set_info_async(export, sizeof(KEY_MDS_CONN), KEY_MDS_CONN,
1010                                    0, NULL, NULL);
1011
1012         /* unknown versions will be caught in
1013          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1014         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1015
1016         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1017             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1018                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1019         else
1020                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1021
1022         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1023         if (rc) {
1024                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1025                 export->exp_imp_reverse = NULL;
1026                 class_destroy_import(revimp);
1027         }
1028
1029         class_import_put(revimp);
1030 out:
1031         if (export) {
1032                 spin_lock(&export->exp_lock);
1033                 export->exp_connecting = 0;
1034                 spin_unlock(&export->exp_lock);
1035         }
1036         if (targref)
1037                 class_decref(targref, __FUNCTION__, cfs_current());
1038         if (rc)
1039                 req->rq_status = rc;
1040         RETURN(rc);
1041 }
1042
1043 int target_handle_disconnect(struct ptlrpc_request *req)
1044 {
1045         int rc;
1046         ENTRY;
1047
1048         rc = req_capsule_server_pack(&req->rq_pill);
1049         if (rc)
1050                 RETURN(rc);
1051
1052         /* keep the rq_export around so we can send the reply */
1053         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1054
1055         RETURN(0);
1056 }
1057
1058 void target_destroy_export(struct obd_export *exp)
1059 {
1060         /* exports created from last_rcvd data, and "fake"
1061            exports created by lctl don't have an import */
1062         if (exp->exp_imp_reverse != NULL)
1063                 client_destroy_import(exp->exp_imp_reverse);
1064
1065         /* We cancel locks at disconnect time, but this will catch any locks
1066          * granted in a race with recovery-induced disconnect. */
1067         if (exp->exp_obd->obd_namespace != NULL)
1068                 ldlm_cancel_locks_for_export(exp);
1069 }
1070
1071 /*
1072  * Recovery functions
1073  */
1074
1075 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1076 {
1077         struct ptlrpc_request *copy_req;
1078         struct lustre_msg *copy_reqmsg;
1079         struct ptlrpc_user_desc *udesc = NULL;
1080
1081         OBD_ALLOC_PTR(copy_req);
1082         if (!copy_req)
1083                 return NULL;
1084         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1085         if (!copy_reqmsg){
1086                 OBD_FREE_PTR(copy_req);
1087                 return NULL;
1088         }
1089
1090         if (orig_req->rq_user_desc) {
1091                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1092
1093                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1094                 if (!udesc) {
1095                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1096                         OBD_FREE_PTR(copy_req);
1097                         return NULL;
1098                 }
1099                 memcpy(udesc, orig_req->rq_user_desc,
1100                        sptlrpc_user_desc_size(ngroups));
1101         }
1102
1103         *copy_req = *orig_req;
1104         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1105         copy_req->rq_reqmsg = copy_reqmsg;
1106         copy_req->rq_user_desc = udesc;
1107
1108         class_export_rpc_get(copy_req->rq_export);
1109         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1110         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1111         sptlrpc_svc_ctx_addref(copy_req);
1112
1113         if (copy_req->rq_reply_state) {
1114                 /* the copied req takes over the reply state */
1115                 orig_req->rq_reply_state = NULL;
1116                 /* to catch further access */
1117                 orig_req->rq_repmsg = NULL;
1118                 orig_req->rq_replen = 0;
1119         }
1120
1121         return copy_req;
1122 }
1123
1124 void ptlrpc_free_clone(struct ptlrpc_request *req)
1125 {
1126         LASSERT(list_empty(&req->rq_replay_list));
1127
1128         ptlrpc_req_drop_rs(req);
1129         sptlrpc_svc_ctx_decref(req);
1130         class_export_rpc_put(req->rq_export);
1131         list_del_init(&req->rq_list);
1132
1133         if (req->rq_user_desc) {
1134                 int ngroups = req->rq_user_desc->pud_ngroups;
1135                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1136         }
1137         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1138         OBD_FREE_PTR(req);
1139 }
1140
1141 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1142 {
1143         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1144         struct obd_export     *exp = req->rq_export;
1145         struct ptlrpc_request *reqiter;
1146         int                    dup = 0;
1147
1148         LASSERT(exp);
1149
1150         spin_lock(&exp->exp_lock);
1151         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1152                             rq_replay_list) {
1153                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1154                         dup = 1;
1155                         break;
1156                 }
1157         }
1158
1159         if (dup) {
1160                 /* we expect it with RESENT and REPLAY flags */
1161                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1162                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1163                         CERROR("invalid flags %x of resent replay\n",
1164                                lustre_msg_get_flags(req->rq_reqmsg));
1165         } else {
1166                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1167         }
1168
1169         spin_unlock(&exp->exp_lock);
1170         return dup;
1171 }
1172
1173 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1174 {
1175         LASSERT(!list_empty(&req->rq_replay_list));
1176         LASSERT(req->rq_export);
1177
1178         spin_lock(&req->rq_export->exp_lock);
1179         list_del_init(&req->rq_replay_list);
1180         spin_unlock(&req->rq_export->exp_lock);
1181 }
1182
1183 #ifdef __KERNEL__
1184 static void target_finish_recovery(struct obd_device *obd)
1185 {
1186         ENTRY;
1187         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1188                       obd->obd_name);
1189
1190         ldlm_reprocess_all_ns(obd->obd_namespace);
1191         spin_lock_bh(&obd->obd_processing_task_lock);
1192         if (list_empty(&obd->obd_req_replay_queue) &&
1193             list_empty(&obd->obd_lock_replay_queue) &&
1194             list_empty(&obd->obd_final_req_queue)) {
1195                 obd->obd_processing_task = 0;
1196         } else {
1197                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1198                        obd->obd_name,
1199                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1200                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1201                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1202                 spin_unlock_bh(&obd->obd_processing_task_lock);
1203                 LBUG();
1204         }
1205         spin_unlock_bh(&obd->obd_processing_task_lock);
1206
1207         /* when recovery finished, cleanup orphans on mds and ost */
1208         if (OBT(obd) && OBP(obd, postrecov)) {
1209                 int rc = OBP(obd, postrecov)(obd);
1210                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1211                               rc < 0 ? "failed" : "complete", rc);
1212         }
1213
1214         obd->obd_recovery_end = cfs_time_current_sec();
1215         EXIT;
1216 }
1217
1218 static void abort_req_replay_queue(struct obd_device *obd)
1219 {
1220         struct ptlrpc_request *req, *n;
1221         struct list_head abort_list;
1222
1223         CFS_INIT_LIST_HEAD(&abort_list);
1224         spin_lock_bh(&obd->obd_processing_task_lock);
1225         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1226         spin_unlock_bh(&obd->obd_processing_task_lock);
1227         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1228                 DEBUG_REQ(D_WARNING, req, "aborted:");
1229                 req->rq_status = -ENOTCONN;
1230                 if (ptlrpc_error(req)) {
1231                         DEBUG_REQ(D_ERROR, req,
1232                                   "failed abort_req_reply; skipping");
1233                 }
1234                 target_exp_dequeue_req_replay(req);
1235                 ptlrpc_free_clone(req);
1236         }
1237 }
1238
1239 static void abort_lock_replay_queue(struct obd_device *obd)
1240 {
1241         struct ptlrpc_request *req, *n;
1242         struct list_head abort_list;
1243
1244         CFS_INIT_LIST_HEAD(&abort_list);
1245         spin_lock_bh(&obd->obd_processing_task_lock);
1246         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1247         spin_unlock_bh(&obd->obd_processing_task_lock);
1248         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1249                 DEBUG_REQ(D_ERROR, req, "aborted:");
1250                 req->rq_status = -ENOTCONN;
1251                 if (ptlrpc_error(req)) {
1252                         DEBUG_REQ(D_ERROR, req,
1253                                   "failed abort_lock_reply; skipping");
1254                 }
1255                 ptlrpc_free_clone(req);
1256         }
1257 }
1258 #endif
1259
1260 /* Called from a cleanup function if the device is being cleaned up
1261    forcefully.  The exports should all have been disconnected already,
1262    the only thing left to do is
1263      - clear the recovery flags
1264      - cancel the timer
1265      - free queued requests and replies, but don't send replies
1266    Because the obd_stopping flag is set, no new requests should be received.
1267
1268 */
1269 void target_cleanup_recovery(struct obd_device *obd)
1270 {
1271         struct ptlrpc_request *req, *n;
1272         struct list_head clean_list;
1273         ENTRY;
1274
1275         LASSERT(obd->obd_stopping);
1276
1277         CFS_INIT_LIST_HEAD(&clean_list);
1278         spin_lock_bh(&obd->obd_processing_task_lock);
1279         if (!obd->obd_recovering) {
1280                 spin_unlock_bh(&obd->obd_processing_task_lock);
1281                 EXIT;
1282                 return;
1283         }
1284         obd->obd_recovering = obd->obd_abort_recovery = 0;
1285         target_cancel_recovery_timer(obd);
1286
1287         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1288         spin_unlock_bh(&obd->obd_processing_task_lock);
1289
1290         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1291                 LASSERT(req->rq_reply_state == 0);
1292                 target_exp_dequeue_req_replay(req);
1293                 ptlrpc_free_clone(req);
1294         }
1295
1296         spin_lock_bh(&obd->obd_processing_task_lock);
1297         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1298         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1299         spin_unlock_bh(&obd->obd_processing_task_lock);
1300
1301         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1302                 LASSERT(req->rq_reply_state == 0);
1303                 ptlrpc_free_clone(req);
1304         }
1305
1306         EXIT;
1307 }
1308
1309 /* obd_processing_task_lock should be held */
1310 void target_cancel_recovery_timer(struct obd_device *obd)
1311 {
1312         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1313         cfs_timer_disarm(&obd->obd_recovery_timer);
1314 }
1315
1316 /* extend = 1 means require at least "duration" seconds left in the timer,
1317    extend = 0 means set the total duration (start_recovery_timer) */
1318 static void reset_recovery_timer(struct obd_device *obd, int duration,
1319                                  int extend)
1320 {
1321         cfs_time_t now = cfs_time_current_sec();
1322         cfs_duration_t left;
1323
1324         spin_lock_bh(&obd->obd_processing_task_lock);
1325         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1326                 spin_unlock_bh(&obd->obd_processing_task_lock);
1327                 return;
1328         }
1329
1330         left = cfs_time_sub(obd->obd_recovery_end, now);
1331
1332         if (extend && (duration > left))
1333                 obd->obd_recovery_timeout += duration - left;
1334         else if (!extend && (duration > obd->obd_recovery_timeout))
1335                 /* Track the client's largest expected replay time */
1336                 obd->obd_recovery_timeout = duration;
1337 #ifdef CRAY_XT3
1338         /*
1339          * If total recovery time already exceed the
1340          * obd_recovery_max_time, then CRAY XT3 will
1341          * abort the recovery
1342          */
1343         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1344                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1345 #endif
1346         obd->obd_recovery_end = obd->obd_recovery_start +
1347                                 obd->obd_recovery_timeout;
1348         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1349             cfs_time_before(now, obd->obd_recovery_end)) {
1350                 left = cfs_time_sub(obd->obd_recovery_end, now);
1351                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1352         }
1353         spin_unlock_bh(&obd->obd_processing_task_lock);
1354         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1355                obd->obd_name, (unsigned)left);
1356 }
1357
1358 static void check_and_start_recovery_timer(struct obd_device *obd)
1359 {
1360         spin_lock_bh(&obd->obd_processing_task_lock);
1361         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1362                 spin_unlock_bh(&obd->obd_processing_task_lock);
1363                 return;
1364         }
1365         CWARN("%s: starting recovery timer\n", obd->obd_name);
1366         obd->obd_recovery_start = cfs_time_current_sec();
1367         /* minimum */
1368         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1369         spin_unlock_bh(&obd->obd_processing_task_lock);
1370
1371         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1372 }
1373
1374 /* Reset the timer with each new client connection */
1375 /*
1376  * This timer is actually reconnect_timer, which is for making sure
1377  * the total recovery window is at least as big as my reconnect
1378  * attempt timing. So the initial recovery time_out will be set to
1379  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1380  * from client is bigger than this, then the recovery time_out will
1381  * be extend to make sure the client could be reconnected, in the
1382  * process, the timeout from the new client should be ignored.
1383  */
1384
1385 static void
1386 target_start_and_reset_recovery_timer(struct obd_device *obd,
1387                                       struct ptlrpc_request *req,
1388                                       int new_client)
1389 {
1390         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1391
1392         if (!new_client && service_time)
1393                 /* Teach server about old server's estimates, as first guess
1394                  * at how long new requests will take. */
1395                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1396                        service_time);
1397
1398         check_and_start_recovery_timer(obd);
1399
1400         /* convert the service time to rpc timeout,
1401          * reuse service_time to limit stack usage */
1402         service_time = at_est2timeout(service_time);
1403
1404         /* We expect other clients to timeout within service_time, then try
1405          * to reconnect, then try the failover server.  The max delay between
1406          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1407         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1408                              INITIAL_CONNECT_TIMEOUT);
1409         if (service_time > obd->obd_recovery_timeout && !new_client)
1410                 reset_recovery_timer(obd, service_time, 0);
1411 }
1412
1413 #ifdef __KERNEL__
1414 static int check_for_next_transno(struct obd_device *obd)
1415 {
1416         struct ptlrpc_request *req = NULL;
1417         int wake_up = 0, connected, completed, queue_len, max;
1418         __u64 next_transno, req_transno;
1419         ENTRY;
1420         spin_lock_bh(&obd->obd_processing_task_lock);
1421
1422         if (!list_empty(&obd->obd_req_replay_queue)) {
1423                 req = list_entry(obd->obd_req_replay_queue.next,
1424                                  struct ptlrpc_request, rq_list);
1425                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1426         } else {
1427                 req_transno = 0;
1428         }
1429
1430         max = obd->obd_max_recoverable_clients;
1431         connected = obd->obd_connected_clients;
1432         completed = connected - obd->obd_recoverable_clients;
1433         queue_len = obd->obd_requests_queued_for_recovery;
1434         next_transno = obd->obd_next_recovery_transno;
1435
1436         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1437                "req_transno: "LPU64", next_transno: "LPU64"\n",
1438                max, connected, completed, queue_len, req_transno, next_transno);
1439
1440         if (obd->obd_abort_recovery) {
1441                 CDEBUG(D_HA, "waking for aborted recovery\n");
1442                 wake_up = 1;
1443         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1444                 CDEBUG(D_HA, "waking for completed recovery\n");
1445                 wake_up = 1;
1446         } else if (req_transno == next_transno) {
1447                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1448                 wake_up = 1;
1449         } else if (queue_len + completed == max) {
1450                 /* handle gaps occured due to lost reply. It is allowed gaps
1451                  * because all clients are connected and there will be resend
1452                  * for missed transaction */
1453                 LASSERTF(req_transno >= next_transno,
1454                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1455                          req_transno, next_transno);
1456
1457                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1458                        "waking for skipped transno (skip: "LPD64
1459                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1460                        next_transno, queue_len, completed, connected, req_transno);
1461                 obd->obd_next_recovery_transno = req_transno;
1462                 wake_up = 1;
1463         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1464                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1465                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1466                 obd->obd_next_recovery_transno = req_transno;
1467                 wake_up = 1;
1468         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1469                 /* some clients haven't connected in time, but we can try
1470                  * to replay requests that demand on already committed ones
1471                  * also, we can replay first non-committed transation */
1472                 LASSERT(req_transno != 0);
1473                 if (obd->obd_version_recov ||
1474                     req_transno == obd->obd_last_committed + 1) {
1475                         obd->obd_next_recovery_transno = req_transno;
1476                 } else if (req_transno > obd->obd_last_committed) {
1477                         /* can't continue recovery: have no needed transno */
1478                         obd->obd_abort_recovery = 1;
1479                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1480                                "connected: %d, completed: %d, queue_len: %d, "
1481                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1482                                max, connected, completed, queue_len,
1483                                req_transno, next_transno);
1484                 }
1485                 wake_up = 1;
1486         }
1487
1488         spin_unlock_bh(&obd->obd_processing_task_lock);
1489         return wake_up;
1490 }
1491
1492 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1493 {
1494         struct l_wait_info lwi = { 0 };
1495         struct ptlrpc_request *req;
1496
1497         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1498                obd->obd_next_recovery_transno);
1499         l_wait_event(obd->obd_next_transno_waitq,
1500                      check_for_next_transno(obd), &lwi);
1501
1502         spin_lock_bh(&obd->obd_processing_task_lock);
1503         if (obd->obd_abort_recovery) {
1504                 req = NULL;
1505         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1506                 req = list_entry(obd->obd_req_replay_queue.next,
1507                                  struct ptlrpc_request, rq_list);
1508                 target_exp_dequeue_req_replay(req);
1509                 list_del_init(&req->rq_list);
1510                 obd->obd_requests_queued_for_recovery--;
1511         } else {
1512                 req = NULL;
1513         }
1514         spin_unlock_bh(&obd->obd_processing_task_lock);
1515         RETURN(req);
1516 }
1517
1518 static int check_for_next_lock(struct obd_device *obd)
1519 {
1520         struct ptlrpc_request *req = NULL;
1521         int wake_up = 0;
1522
1523         spin_lock_bh(&obd->obd_processing_task_lock);
1524         if (!list_empty(&obd->obd_lock_replay_queue)) {
1525                 req = list_entry(obd->obd_lock_replay_queue.next,
1526                                  struct ptlrpc_request, rq_list);
1527                 CDEBUG(D_HA, "waking for next lock\n");
1528                 wake_up = 1;
1529         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1530                 CDEBUG(D_HA, "waking for completed lock replay\n");
1531                 wake_up = 1;
1532         } else if (obd->obd_abort_recovery) {
1533                 CDEBUG(D_HA, "waking for aborted recovery\n");
1534                 wake_up = 1;
1535         }
1536         spin_unlock_bh(&obd->obd_processing_task_lock);
1537
1538         return wake_up;
1539 }
1540
1541 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1542 {
1543         struct l_wait_info lwi = { 0 };
1544         struct ptlrpc_request *req;
1545
1546         CDEBUG(D_HA, "Waiting for lock\n");
1547         l_wait_event(obd->obd_next_transno_waitq,
1548                      check_for_next_lock(obd), &lwi);
1549
1550         spin_lock_bh(&obd->obd_processing_task_lock);
1551         if (obd->obd_abort_recovery) {
1552                 req = NULL;
1553         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1554                 req = list_entry(obd->obd_lock_replay_queue.next,
1555                                  struct ptlrpc_request, rq_list);
1556                 list_del_init(&req->rq_list);
1557         } else {
1558                 req = NULL;
1559         }
1560         spin_unlock_bh(&obd->obd_processing_task_lock);
1561         return req;
1562 }
1563
1564 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1565 {
1566         struct ptlrpc_request *req;
1567
1568         spin_lock_bh(&obd->obd_processing_task_lock);
1569         if (!list_empty(&obd->obd_final_req_queue)) {
1570                 req = list_entry(obd->obd_final_req_queue.next,
1571                                  struct ptlrpc_request, rq_list);
1572                 list_del_init(&req->rq_list);
1573                 if (req->rq_export->exp_in_recovery) {
1574                         spin_lock(&req->rq_export->exp_lock);
1575                         req->rq_export->exp_in_recovery = 0;
1576                         spin_unlock(&req->rq_export->exp_lock);
1577                 }
1578         } else {
1579                 req = NULL;
1580         }
1581         spin_unlock_bh(&obd->obd_processing_task_lock);
1582         return req;
1583 }
1584
1585 static inline int req_vbr_done(struct obd_export *exp)
1586 {
1587         return (exp->exp_vbr_failed == 0);
1588 }
1589
1590 static inline int req_replay_done(struct obd_export *exp)
1591 {
1592         return (exp->exp_req_replay_needed == 0);
1593 }
1594
1595 static inline int lock_replay_done(struct obd_export *exp)
1596 {
1597         return (exp->exp_lock_replay_needed == 0);
1598 }
1599
1600 static inline int connect_done(struct obd_export *exp)
1601 {
1602         return (exp->exp_in_recovery != 0);
1603 }
1604
1605 static int check_for_clients(struct obd_device *obd)
1606 {
1607         if (obd->obd_abort_recovery || obd->obd_version_recov)
1608                 return 1;
1609         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1610         if (obd->obd_no_conn == 0 &&
1611             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1612                 return 1;
1613         return 0;
1614 }
1615
1616 static int handle_recovery_req(struct ptlrpc_thread *thread,
1617                                struct ptlrpc_request *req,
1618                                svc_handler_t handler)
1619 {
1620         int rc;
1621         ENTRY;
1622
1623         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1624         if (rc) {
1625                 CERROR("Failure to initialize session: %d\n", rc);
1626                 return rc;
1627         }
1628         req->rq_session.lc_thread = thread;
1629         lu_context_enter(&req->rq_session);
1630         req->rq_svc_thread = thread;
1631         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1632
1633         /* thread context */
1634         lu_context_enter(&thread->t_env->le_ctx);
1635         (void)handler(req);
1636         lu_context_exit(&thread->t_env->le_ctx);
1637
1638         lu_context_exit(&req->rq_session);
1639         lu_context_fini(&req->rq_session);
1640         /* don't reset timer for final stage */
1641         if (!req_replay_done(req->rq_export) ||
1642             !lock_replay_done(req->rq_export))
1643                 reset_recovery_timer(class_exp2obd(req->rq_export),
1644                        AT_OFF ? obd_timeout :
1645                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1646         ptlrpc_free_clone(req);
1647         RETURN(0);
1648 }
1649
1650 static int target_recovery_thread(void *arg)
1651 {
1652         struct lu_target *lut = arg;
1653         struct obd_device *obd = lut->lut_obd;
1654         struct ptlrpc_request *req;
1655         struct target_recovery_data *trd = &obd->obd_recovery_data;
1656         struct l_wait_info lwi = { 0 };
1657         unsigned long delta;
1658         unsigned long flags;
1659         struct lu_env env;
1660         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1661         int rc = 0;
1662         ENTRY;
1663
1664         cfs_daemonize("tgt_recov");
1665
1666         SIGNAL_MASK_LOCK(current, flags);
1667         sigfillset(&current->blocked);
1668         RECALC_SIGPENDING;
1669         SIGNAL_MASK_UNLOCK(current, flags);
1670
1671         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1672         if (rc)
1673                 RETURN(rc);
1674
1675         thread->t_env = &env;
1676         env.le_ctx.lc_thread = thread;
1677
1678         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1679                cfs_curproc_pid());
1680         trd->trd_processing_task = cfs_curproc_pid();
1681
1682         obd->obd_recovering = 1;
1683         complete(&trd->trd_starting);
1684
1685         /* first of all, we have to know the first transno to replay */
1686         obd->obd_abort_recovery = 0;
1687         l_wait_event(obd->obd_next_transno_waitq,
1688                      check_for_clients(obd), &lwi);
1689
1690         /* If some clients haven't connected in time, evict them */
1691         if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) {
1692                 CWARN("Some clients haven't connect in time (%d/%d),"
1693                        "evict them\n", obd->obd_connected_clients,
1694                        obd->obd_max_recoverable_clients);
1695                 class_disconnect_stale_exports(obd, connect_done,
1696                                                exp_flags_from_obd(obd) |
1697                                                OBD_OPT_ABORT_RECOV);
1698                 /**
1699                  * if recovery proceeds with versions then some clients may be
1700                  * timed out waiting for others and trying to reconnect.
1701                  * Extend timer for such reconnect cases.
1702                  */
1703                 if (obd->obd_version_recov)
1704                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1);
1705         }
1706
1707         /* next stage: replay requests */
1708         delta = jiffies;
1709         obd->obd_req_replaying = 1;
1710         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1711                atomic_read(&obd->obd_req_replay_clients),
1712                obd->obd_next_recovery_transno);
1713         while ((req = target_next_replay_req(obd))) {
1714                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1715                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1716                           lustre_msg_get_transno(req->rq_reqmsg),
1717                           libcfs_nid2str(req->rq_peer.nid));
1718                 handle_recovery_req(thread, req,
1719                                     trd->trd_recovery_handler);
1720                 obd->obd_replayed_requests++;
1721                 spin_lock_bh(&obd->obd_processing_task_lock);
1722                 obd->obd_next_recovery_transno++;
1723                 spin_unlock_bh(&obd->obd_processing_task_lock);
1724         }
1725
1726         /* If some clients haven't replayed requests in time, evict them */
1727         if (obd->obd_abort_recovery) {
1728                 CDEBUG(D_WARNING, "req replay is aborted\n");
1729                 class_disconnect_stale_exports(obd, req_replay_done,
1730                                                exp_flags_from_obd(obd) |
1731                                                OBD_OPT_ABORT_RECOV);
1732                 abort_req_replay_queue(obd);
1733         }
1734         LASSERT(list_empty(&obd->obd_req_replay_queue));
1735
1736         /* The second stage: replay locks */
1737         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1738                atomic_read(&obd->obd_lock_replay_clients));
1739         while ((req = target_next_replay_lock(obd))) {
1740                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1741                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1742                           libcfs_nid2str(req->rq_peer.nid));
1743                 handle_recovery_req(thread, req,
1744                                     trd->trd_recovery_handler);
1745                 obd->obd_replayed_locks++;
1746         }
1747
1748         /* If some clients haven't replayed requests in time, evict them */
1749         if (obd->obd_abort_recovery) {
1750                 int stale;
1751                 CERROR("lock replay is aborted\n");
1752                 stale = class_disconnect_stale_exports(obd, lock_replay_done,
1753                                                        exp_flags_from_obd(obd) |
1754                                                        OBD_OPT_ABORT_RECOV);
1755                 abort_lock_replay_queue(obd);
1756         }
1757         LASSERT(list_empty(&obd->obd_lock_replay_queue));
1758
1759         /* The third stage: reply on final pings */
1760         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1761         /** evict exports failed VBR */
1762         class_disconnect_stale_exports(obd, req_vbr_done,
1763                                        exp_flags_from_obd(obd) |
1764                                        OBD_OPT_ABORT_RECOV);
1765         /** Update server last boot epoch */
1766         lut_boot_epoch_update(lut);
1767         /* We drop recoverying flag to forward all new requests
1768          * to regular mds_handle() since now */
1769         spin_lock_bh(&obd->obd_processing_task_lock);
1770         obd->obd_recovering = obd->obd_abort_recovery = 0;
1771         target_cancel_recovery_timer(obd);
1772         spin_unlock_bh(&obd->obd_processing_task_lock);
1773         while ((req = target_next_final_ping(obd))) {
1774                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1775                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1776                           libcfs_nid2str(req->rq_peer.nid));
1777                 handle_recovery_req(thread, req,
1778                                     trd->trd_recovery_handler);
1779         }
1780
1781         delta = (jiffies - delta) / HZ;
1782         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1783               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1784         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1785         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1786         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1787                 CWARN("too long recovery - read logs\n");
1788                 libcfs_debug_dumplog();
1789         }
1790
1791         target_finish_recovery(obd);
1792
1793         lu_context_fini(&env.le_ctx);
1794         trd->trd_processing_task = 0;
1795         complete(&trd->trd_finishing);
1796         RETURN(rc);
1797 }
1798
1799 static int target_start_recovery_thread(struct lu_target *lut,
1800                                         svc_handler_t handler)
1801 {
1802         struct obd_device *obd = lut->lut_obd;
1803         int rc = 0;
1804         struct target_recovery_data *trd = &obd->obd_recovery_data;
1805
1806         memset(trd, 0, sizeof(*trd));
1807         init_completion(&trd->trd_starting);
1808         init_completion(&trd->trd_finishing);
1809         trd->trd_recovery_handler = handler;
1810
1811         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1812                 wait_for_completion(&trd->trd_starting);
1813                 LASSERT(obd->obd_recovering != 0);
1814         } else
1815                 rc = -ECHILD;
1816
1817         return rc;
1818 }
1819
1820 void target_stop_recovery_thread(struct obd_device *obd)
1821 {
1822         spin_lock_bh(&obd->obd_processing_task_lock);
1823         if (obd->obd_recovery_data.trd_processing_task > 0) {
1824                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1825                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1826                 obd->obd_abort_recovery = 1;
1827                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1828                 spin_unlock_bh(&obd->obd_processing_task_lock);
1829                 wait_for_completion(&trd->trd_finishing);
1830         } else {
1831                 spin_unlock_bh(&obd->obd_processing_task_lock);
1832         }
1833 }
1834
1835 void target_recovery_fini(struct obd_device *obd)
1836 {
1837         class_disconnect_exports(obd);
1838         target_stop_recovery_thread(obd);
1839         target_cleanup_recovery(obd);
1840 }
1841 EXPORT_SYMBOL(target_recovery_fini);
1842
1843 static void target_recovery_expired(unsigned long castmeharder)
1844 {
1845         struct obd_device *obd = (struct obd_device *)castmeharder;
1846         LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1847                       "after %lds (%d clients did)\n",
1848                       obd->obd_name, obd->obd_recoverable_clients,
1849                       cfs_time_current_sec()- obd->obd_recovery_start,
1850                       obd->obd_connected_clients);
1851
1852         spin_lock_bh(&obd->obd_processing_task_lock);
1853         obd->obd_version_recov = 1;
1854         CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n",
1855                atomic_read(&obd->obd_req_replay_clients),
1856                obd->obd_next_recovery_transno);
1857         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1858         spin_unlock_bh(&obd->obd_processing_task_lock);
1859 }
1860
1861 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1862 {
1863         struct obd_device *obd = lut->lut_obd;
1864         if (obd->obd_max_recoverable_clients == 0)
1865                 return;
1866
1867         CWARN("RECOVERY: service %s, %d recoverable clients, "
1868               "last_transno "LPU64"\n", obd->obd_name,
1869               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1870         LASSERT(obd->obd_stopping == 0);
1871         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1872         obd->obd_recovery_start = 0;
1873         obd->obd_recovery_end = 0;
1874         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1875         /* bz13079: this should be set to desired value for ost but not for mds */
1876         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1877         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1878         target_start_recovery_thread(lut, handler);
1879 }
1880 EXPORT_SYMBOL(target_recovery_init);
1881
1882 #endif
1883
1884 static int target_process_req_flags(struct obd_device *obd,
1885                                     struct ptlrpc_request *req)
1886 {
1887         struct obd_export *exp = req->rq_export;
1888         LASSERT(exp != NULL);
1889         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1890                 /* client declares he's ready to replay locks */
1891                 spin_lock_bh(&obd->obd_processing_task_lock);
1892                 if (exp->exp_req_replay_needed) {
1893                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1894                         spin_lock(&exp->exp_lock);
1895                         exp->exp_req_replay_needed = 0;
1896                         spin_unlock(&exp->exp_lock);
1897                         atomic_dec(&obd->obd_req_replay_clients);
1898                         LASSERT(obd->obd_recoverable_clients > 0);
1899                         obd->obd_recoverable_clients--;
1900                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1901                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1902                 }
1903                 spin_unlock_bh(&obd->obd_processing_task_lock);
1904         }
1905         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1906                 /* client declares he's ready to complete recovery
1907                  * so, we put the request on th final queue */
1908                 spin_lock_bh(&obd->obd_processing_task_lock);
1909                 if (exp->exp_lock_replay_needed) {
1910                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1911                         spin_lock(&exp->exp_lock);
1912                         exp->exp_lock_replay_needed = 0;
1913                         spin_unlock(&exp->exp_lock);
1914                         atomic_dec(&obd->obd_lock_replay_clients);
1915                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1916                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1917                 }
1918                 spin_unlock_bh(&obd->obd_processing_task_lock);
1919         }
1920
1921         return 0;
1922 }
1923
1924 int target_queue_recovery_request(struct ptlrpc_request *req,
1925                                   struct obd_device *obd)
1926 {
1927         struct list_head *tmp;
1928         int inserted = 0;
1929         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1930         ENTRY;
1931
1932         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1933                 /* Processing the queue right now, don't re-add. */
1934                 RETURN(1);
1935         }
1936
1937         target_process_req_flags(obd, req);
1938
1939         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1940                 /* client declares he's ready to complete recovery
1941                  * so, we put the request on th final queue */
1942                 req = ptlrpc_clone_req(req);
1943                 if (req == NULL)
1944                         RETURN(-ENOMEM);
1945                 DEBUG_REQ(D_HA, req, "queue final req");
1946                 spin_lock_bh(&obd->obd_processing_task_lock);
1947                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1948                 if (obd->obd_recovering)
1949                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1950                 else {
1951                         spin_unlock_bh(&obd->obd_processing_task_lock);
1952                         ptlrpc_free_clone(req);
1953                         if (obd->obd_stopping) {
1954                                 RETURN(-ENOTCONN);
1955                         } else {
1956                                 RETURN(1);
1957                         }
1958                 }
1959                 spin_unlock_bh(&obd->obd_processing_task_lock);
1960                 RETURN(0);
1961         }
1962         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1963                 /* client declares he's ready to replay locks */
1964                 req = ptlrpc_clone_req(req);
1965                 if (req == NULL)
1966                         RETURN(-ENOMEM);
1967                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1968                 spin_lock_bh(&obd->obd_processing_task_lock);
1969                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1970                 LASSERT(obd->obd_recovering);
1971                 /* usually due to recovery abort */
1972                 if (!req->rq_export->exp_in_recovery) {
1973                         spin_unlock_bh(&obd->obd_processing_task_lock);
1974                         ptlrpc_free_clone(req);
1975                         RETURN(-ENOTCONN);
1976                 }
1977                 LASSERT(req->rq_export->exp_lock_replay_needed);
1978                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1979                 spin_unlock_bh(&obd->obd_processing_task_lock);
1980                 RETURN(0);
1981         }
1982
1983         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1984          * (i.e. buflens etc are in my own byte order), but type-dependent
1985          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1986
1987         if (!transno) {
1988                 CFS_INIT_LIST_HEAD(&req->rq_list);
1989                 DEBUG_REQ(D_HA, req, "not queueing");
1990                 RETURN(1);
1991         }
1992
1993         spin_lock_bh(&obd->obd_processing_task_lock);
1994
1995         /* If we're processing the queue, we want don't want to queue this
1996          * message.
1997          *
1998          * Also, if this request has a transno less than the one we're waiting
1999          * for, we should process it now.  It could (and currently always will)
2000          * be an open request for a descriptor that was opened some time ago.
2001          *
2002          * Also, a resent, replayed request that has already been
2003          * handled will pass through here and be processed immediately.
2004          */
2005         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2006               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2007         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2008                 /* Processing the queue right now, don't re-add. */
2009                 LASSERT(list_empty(&req->rq_list));
2010                 spin_unlock_bh(&obd->obd_processing_task_lock);
2011                 RETURN(1);
2012         }
2013         spin_unlock_bh(&obd->obd_processing_task_lock);
2014
2015         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2016                 RETURN(0);
2017
2018         req = ptlrpc_clone_req(req);
2019         if (req == NULL)
2020                 RETURN(-ENOMEM);
2021
2022         spin_lock_bh(&obd->obd_processing_task_lock);
2023         LASSERT(obd->obd_recovering);
2024         if (!req->rq_export->exp_in_recovery) {
2025                 spin_unlock_bh(&obd->obd_processing_task_lock);
2026                 ptlrpc_free_clone(req);
2027                 RETURN(-ENOTCONN);
2028         }
2029         LASSERT(req->rq_export->exp_req_replay_needed);
2030
2031         if (target_exp_enqueue_req_replay(req)) {
2032                 spin_unlock_bh(&obd->obd_processing_task_lock);
2033                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2034                 ptlrpc_free_clone(req);
2035                 RETURN(0);
2036         }
2037
2038         /* XXX O(n^2) */
2039         list_for_each(tmp, &obd->obd_req_replay_queue) {
2040                 struct ptlrpc_request *reqiter =
2041                         list_entry(tmp, struct ptlrpc_request, rq_list);
2042
2043                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2044                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2045                         inserted = 1;
2046                         break;
2047                 }
2048
2049                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2050                              transno)) {
2051                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2052                                   "has been claimed by another client");
2053                         spin_unlock_bh(&obd->obd_processing_task_lock);
2054                         target_exp_dequeue_req_replay(req);
2055                         ptlrpc_free_clone(req);
2056                         RETURN(0);
2057                 }
2058         }
2059
2060         if (!inserted)
2061                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2062
2063         obd->obd_requests_queued_for_recovery++;
2064         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2065         spin_unlock_bh(&obd->obd_processing_task_lock);
2066         RETURN(0);
2067 }
2068
2069 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2070 {
2071         return req->rq_export->exp_obd;
2072 }
2073
2074 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2075 {
2076         LASSERT(exp != NULL);
2077         return &exp->exp_obd->obd_namespace->ns_pool;
2078 }
2079
2080 /**
2081  * Packs current SLV and Limit into \a req.
2082  */
2083 int target_pack_pool_reply(struct ptlrpc_request *req)
2084 {
2085         struct obd_device *obd;
2086         ENTRY;
2087
2088         /*
2089          * Check that we still have all structures alive as this may
2090          * be some late rpc in shutdown time.
2091          */
2092         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2093                      !exp_connect_lru_resize(req->rq_export))) {
2094                 lustre_msg_set_slv(req->rq_repmsg, 0);
2095                 lustre_msg_set_limit(req->rq_repmsg, 0);
2096                 RETURN(0);
2097         }
2098
2099         /*
2100          * OBD is alive here as export is alive, which we checked above.
2101          */
2102         obd = req->rq_export->exp_obd;
2103
2104         read_lock(&obd->obd_pool_lock);
2105         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2106         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2107         read_unlock(&obd->obd_pool_lock);
2108
2109         RETURN(0);
2110 }
2111
2112 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2113 {
2114         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2115                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2116                 return (-ECOMM);
2117         }
2118
2119         if (unlikely(rc)) {
2120                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2121                 req->rq_status = rc;
2122                 return (ptlrpc_send_error(req, 1));
2123         } else {
2124                 DEBUG_REQ(D_NET, req, "sending reply");
2125         }
2126
2127         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2128 }
2129
2130 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2131 {
2132         int                        netrc;
2133         struct ptlrpc_reply_state *rs;
2134         struct obd_device         *obd;
2135         struct obd_export         *exp;
2136         struct ptlrpc_service     *svc;
2137         ENTRY;
2138
2139         if (req->rq_no_reply) {
2140                 EXIT;
2141                 return;
2142         }
2143
2144         svc = req->rq_rqbd->rqbd_service;
2145         rs = req->rq_reply_state;
2146         if (rs == NULL || !rs->rs_difficult) {
2147                 /* no notifiers */
2148                 target_send_reply_msg (req, rc, fail_id);
2149                 EXIT;
2150                 return;
2151         }
2152
2153         /* must be an export if locks saved */
2154         LASSERT (req->rq_export != NULL);
2155         /* req/reply consistent */
2156         LASSERT (rs->rs_service == svc);
2157
2158         /* "fresh" reply */
2159         LASSERT (!rs->rs_scheduled);
2160         LASSERT (!rs->rs_scheduled_ever);
2161         LASSERT (!rs->rs_handled);
2162         LASSERT (!rs->rs_on_net);
2163         LASSERT (rs->rs_export == NULL);
2164         LASSERT (list_empty(&rs->rs_obd_list));
2165         LASSERT (list_empty(&rs->rs_exp_list));
2166
2167         exp = class_export_get (req->rq_export);
2168         obd = exp->exp_obd;
2169
2170         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2171         rs->rs_scheduled = 1;
2172         rs->rs_on_net    = 1;
2173         rs->rs_xid       = req->rq_xid;
2174         rs->rs_transno   = req->rq_transno;
2175         rs->rs_export    = exp;
2176
2177         spin_lock(&exp->exp_uncommitted_replies_lock);
2178         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2179                rs->rs_transno, exp->exp_last_committed);
2180         if (rs->rs_transno > exp->exp_last_committed) {
2181                 /* not committed already */
2182                 list_add_tail(&rs->rs_obd_list,
2183                               &exp->exp_uncommitted_replies);
2184         }
2185         spin_unlock (&exp->exp_uncommitted_replies_lock);
2186
2187         spin_lock(&exp->exp_lock);
2188         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2189         spin_unlock(&exp->exp_lock);
2190
2191         netrc = target_send_reply_msg (req, rc, fail_id);
2192
2193         spin_lock(&svc->srv_lock);
2194
2195         atomic_inc(&svc->srv_n_difficult_replies);
2196
2197         if (netrc != 0) {
2198                 /* error sending: reply is off the net.  Also we need +1
2199                  * reply ref until ptlrpc_server_handle_reply() is done
2200                  * with the reply state (if the send was successful, there
2201                  * would have been +1 ref for the net, which
2202                  * reply_out_callback leaves alone) */
2203                 rs->rs_on_net = 0;
2204                 ptlrpc_rs_addref(rs);
2205                 atomic_inc (&svc->srv_outstanding_replies);
2206         }
2207
2208         spin_lock(&rs->rs_lock);
2209         if (rs->rs_transno <= exp->exp_last_committed ||
2210             (!rs->rs_on_net && !rs->rs_no_ack) ||
2211              list_empty(&rs->rs_exp_list) ||     /* completed already */
2212              list_empty(&rs->rs_obd_list)) {
2213                 CDEBUG(D_HA, "Schedule reply immediately\n");
2214                 ptlrpc_dispatch_difficult_reply(rs);
2215         } else {
2216                 list_add (&rs->rs_list, &svc->srv_active_replies);
2217                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2218         }
2219         spin_unlock(&rs->rs_lock);
2220         spin_unlock(&svc->srv_lock);
2221         EXIT;
2222 }
2223
2224 int target_handle_ping(struct ptlrpc_request *req)
2225 {
2226         obd_ping(req->rq_export);
2227         return req_capsule_server_pack(&req->rq_pill);
2228 }
2229
2230 void target_committed_to_req(struct ptlrpc_request *req)
2231 {
2232         struct obd_export *exp = req->rq_export;
2233
2234         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2235                 lustre_msg_set_last_committed(req->rq_repmsg,
2236                                               exp->exp_last_committed);
2237         else
2238                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2239                           "%d)", exp->exp_obd->obd_no_transno,
2240                           req->rq_repmsg == NULL);
2241
2242         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2243                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2244 }
2245 EXPORT_SYMBOL(target_committed_to_req);
2246
2247 int target_handle_qc_callback(struct ptlrpc_request *req)
2248 {
2249         struct obd_quotactl *oqctl;
2250         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2251
2252         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2253         if (oqctl == NULL) {
2254                 CERROR("Can't unpack obd_quotactl\n");
2255                 RETURN(-EPROTO);
2256         }
2257
2258         cli->cl_qchk_stat = oqctl->qc_stat;
2259
2260         return 0;
2261 }
2262
2263 #ifdef HAVE_QUOTA_SUPPORT
2264 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2265 {
2266 #ifdef __KERNEL__
2267         struct obd_device *obd = req->rq_export->exp_obd;
2268         struct obd_device *master_obd;
2269         struct obd_device_target *obt;
2270         struct lustre_quota_ctxt *qctxt;
2271         struct qunit_data *qdata = NULL;
2272         int rc = 0;
2273         ENTRY;
2274
2275         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2276                 RETURN(rc);
2277
2278         rc = req_capsule_server_pack(&req->rq_pill);
2279         if (rc) {
2280                 CERROR("packing reply failed!: rc = %d\n", rc);
2281                 RETURN(rc);
2282         }
2283
2284         LASSERT(req->rq_export);
2285
2286         OBD_ALLOC(qdata, sizeof(struct qunit_data));
2287         if (!qdata)
2288                 RETURN(-ENOMEM);
2289         rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2290         if (rc < 0) {
2291                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2292                 GOTO(out, rc);
2293         }
2294
2295         /* we use the observer */
2296         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2297                 CERROR("Can't find the observer, it is recovering\n");
2298                 req->rq_status = -EAGAIN;
2299                 GOTO(send_reply, rc = -EAGAIN);
2300         }
2301
2302         master_obd = obd->obd_observer->obd_observer;
2303         obt = &master_obd->u.obt;
2304         qctxt = &obt->obt_qctxt;
2305
2306         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2307                 /* quota_type has not been processed yet, return EAGAIN
2308                  * until we know whether or not quotas are supposed to
2309                  * be enabled */
2310                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2311                        "-EAGAIN\n");
2312                 req->rq_status = -EAGAIN;
2313                 rc = ptlrpc_reply(req);
2314                 GOTO(out, rc);
2315         }
2316
2317         down_read(&obt->obt_rwsem);
2318         if (qctxt->lqc_lqs_hash == NULL) {
2319                 up_read(&obt->obt_rwsem);
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_ctxt is not ready yet, return "
2324                        "-EAGAIN\n");
2325                 req->rq_status = -EAGAIN;
2326                 rc = ptlrpc_reply(req);
2327                 GOTO(out, rc);
2328         }
2329
2330         LASSERT(qctxt->lqc_handler);
2331         rc = qctxt->lqc_handler(master_obd, qdata,
2332                                 lustre_msg_get_opc(req->rq_reqmsg));
2333         up_read(&obt->obt_rwsem);
2334         if (rc && rc != -EDQUOT)
2335                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2336                        "dqacq failed! (rc:%d)\n", rc);
2337         req->rq_status = rc;
2338
2339         /* there are three forms of qunit(historic causes), so we need to
2340          * adjust the same form to different forms slaves needed */
2341         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2342         if (rc < 0) {
2343                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2344                 GOTO(out, rc);
2345         }
2346
2347         /* Block the quota req. b=14840 */
2348         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2349 send_reply:
2350         rc = ptlrpc_reply(req);
2351 out:
2352         OBD_FREE(qdata, sizeof(struct qunit_data));
2353         RETURN(rc);
2354 #else
2355         return 0;
2356 #endif /* !__KERNEL__ */
2357 }
2358 #endif /* HAVE_QUOTA_SUPPORT */
2359
2360 ldlm_mode_t lck_compat_array[] = {
2361         [LCK_EX] LCK_COMPAT_EX,
2362         [LCK_PW] LCK_COMPAT_PW,
2363         [LCK_PR] LCK_COMPAT_PR,
2364         [LCK_CW] LCK_COMPAT_CW,
2365         [LCK_CR] LCK_COMPAT_CR,
2366         [LCK_NL] LCK_COMPAT_NL,
2367         [LCK_GROUP] LCK_COMPAT_GROUP,
2368         [LCK_COS] LCK_COMPAT_COS,
2369 };
2370
2371 /**
2372  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2373  * not escape to the user level.
2374  */
2375 int ldlm_error2errno(ldlm_error_t error)
2376 {
2377         int result;
2378
2379         switch (error) {
2380         case ELDLM_OK:
2381                 result = 0;
2382                 break;
2383         case ELDLM_LOCK_CHANGED:
2384                 result = -ESTALE;
2385                 break;
2386         case ELDLM_LOCK_ABORTED:
2387                 result = -ENAVAIL;
2388                 break;
2389         case ELDLM_LOCK_REPLACED:
2390                 result = -ESRCH;
2391                 break;
2392         case ELDLM_NO_LOCK_DATA:
2393                 result = -ENOENT;
2394                 break;
2395         case ELDLM_NAMESPACE_EXISTS:
2396                 result = -EEXIST;
2397                 break;
2398         case ELDLM_BAD_NAMESPACE:
2399                 result = -EBADF;
2400                 break;
2401         default:
2402                 if (((int)error) < 0)  /* cast to signed type */
2403                         result = error; /* as ldlm_error_t can be unsigned */
2404                 else {
2405                         CERROR("Invalid DLM result code: %i\n", error);
2406                         result = -EPROTO;
2407                 }
2408         }
2409         return result;
2410 }
2411 EXPORT_SYMBOL(ldlm_error2errno);
2412
2413 /**
2414  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2415  */
2416 ldlm_error_t ldlm_errno2error(int err_no)
2417 {
2418         int error;
2419
2420         switch (err_no) {
2421         case 0:
2422                 error = ELDLM_OK;
2423                 break;
2424         case -ESTALE:
2425                 error = ELDLM_LOCK_CHANGED;
2426                 break;
2427         case -ENAVAIL:
2428                 error = ELDLM_LOCK_ABORTED;
2429                 break;
2430         case -ESRCH:
2431                 error = ELDLM_LOCK_REPLACED;
2432                 break;
2433         case -ENOENT:
2434                 error = ELDLM_NO_LOCK_DATA;
2435                 break;
2436         case -EEXIST:
2437                 error = ELDLM_NAMESPACE_EXISTS;
2438                 break;
2439         case -EBADF:
2440                 error = ELDLM_BAD_NAMESPACE;
2441                 break;
2442         default:
2443                 error = err_no;
2444         }
2445         return error;
2446 }
2447 EXPORT_SYMBOL(ldlm_errno2error);
2448