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_disconnect, rc = -EINVAL);
 497         }
 498
 499         cli->cl_conn_count--;
 500         if (cli->cl_conn_count)
 501                 GOTO(out_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
 547  out_disconnect:
 548         /* use server style - class_disconnect should be always called for
 549          * o_disconnect */
 550         err = class_disconnect(exp);
 551         if (!rc && err)
 552                 rc = err;
 553
 554         up_write(&cli->cl_sem);
 555         if (to_be_freed)
 556                 ldlm_namespace_free_post(to_be_freed);
 557
 558         RETURN(rc);
 559 }
 560
 561 /* --------------------------------------------------------------------------
 562  * from old lib/target.c
 563  * -------------------------------------------------------------------------- */
 564
 565 static int target_handle_reconnect(struct lustre_handle *conn,
 566                                    struct obd_export *exp,
 567                                    struct obd_uuid *cluuid)
 568 {
 569         ENTRY;
 570         if (exp->exp_connection && exp->exp_imp_reverse) {
 571                 struct lustre_handle *hdl;
 572                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 573                 /* Might be a re-connect after a partition. */
 574                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 575                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 576                               cluuid->uuid);
 577                         conn->cookie = exp->exp_handle.h_cookie;
 578                         /* target_handle_connect() treats EALREADY and
 579                          * -EALREADY differently.  EALREADY means we are
 580                          * doing a valid reconnect from the same client. */
 581                         RETURN(EALREADY);
 582                 } else {
 583                         CERROR("%s reconnecting from %s, "
 584                                "handle mismatch (ours "LPX64", theirs "
 585                                LPX64")\n", cluuid->uuid,
 586                                exp->exp_connection->c_remote_uuid.uuid,
 587                                hdl->cookie, conn->cookie);
 588                         memset(conn, 0, sizeof *conn);
 589                         /* target_handle_connect() treats EALREADY and
 590                          * -EALREADY differently.  -EALREADY is an error
 591                          * (same UUID, different handle). */
 592                         RETURN(-EALREADY);
 593                 }
 594         }
 595
 596         conn->cookie = exp->exp_handle.h_cookie;
 597         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 598                cluuid->uuid, exp, conn->cookie);
 599         RETURN(0);
 600 }
 601
 602 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 603                           int error)
 604 {
 605         struct obd_export *exp = cb_data;
 606
 607         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 608                obd->obd_name, exp->exp_client_uuid.uuid);
 609
 610         spin_lock(&exp->exp_lock);
 611         exp->exp_need_sync = 0;
 612         spin_unlock(&exp->exp_lock);
 613 }
 614 EXPORT_SYMBOL(target_client_add_cb);
 615 static void
 616 target_start_and_reset_recovery_timer(struct obd_device *obd,
 617                                       struct ptlrpc_request *req,
 618                                       int new_client);
 619
 620 int target_handle_connect(struct ptlrpc_request *req)
 621 {
 622         struct obd_device *target, *targref = NULL;
 623         struct obd_export *export = NULL;
 624         struct obd_import *revimp;
 625         struct lustre_handle conn;
 626         struct lustre_handle *tmp;
 627         struct obd_uuid tgtuuid;
 628         struct obd_uuid cluuid;
 629         struct obd_uuid remote_uuid;
 630         char *str;
 631         int rc = 0;
 632         int mds_conn = 0;
 633         struct obd_connect_data *data, *tmpdata;
 634         lnet_nid_t *client_nid = NULL;
 635         ENTRY;
 636
 637         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 638
 639         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 640         if (str == NULL) {
 641                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 642                 GOTO(out, rc = -EINVAL);
 643         }
 644
 645         obd_str2uuid(&tgtuuid, str);
 646         target = class_uuid2obd(&tgtuuid);
 647         if (!target)
 648                 target = class_name2obd(str);
 649
 650         if (!target || target->obd_stopping || !target->obd_set_up) {
 651                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 652                                    " for connect (%s)\n", str,
 653                                    !target ? "no target" :
 654                                    (target->obd_stopping ? "stopping" :
 655                                    "not set up"));
 656                 GOTO(out, rc = -ENODEV);
 657         }
 658
 659         if (target->obd_no_conn) {
 660                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 661                               "from %s\n", target->obd_name,
 662                               libcfs_nid2str(req->rq_peer.nid));
 663                 GOTO(out, rc = -EAGAIN);
 664         }
 665
 666         /* Make sure the target isn't cleaned up while we're here. Yes,
 667            there's still a race between the above check and our incref here.
 668            Really, class_uuid2obd should take the ref. */
 669         targref = class_incref(target, __FUNCTION__, cfs_current());
 670
 671
 672         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 673         if (str == NULL) {
 674                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 675                 GOTO(out, rc = -EINVAL);
 676         }
 677
 678         obd_str2uuid(&cluuid, str);
 679
 680         /* XXX extract a nettype and format accordingly */
 681         switch (sizeof(lnet_nid_t)) {
 682                 /* NB the casts only avoid compiler warnings */
 683         case 8:
 684                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 685                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 686                 break;
 687         case 4:
 688                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 689                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 690                 break;
 691         default:
 692                 LBUG();
 693         }
 694
 695         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 696         if (tmp == NULL)
 697                 GOTO(out, rc = -EPROTO);
 698
 699         conn = *tmp;
 700
 701         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 702         if (!data)
 703                 GOTO(out, rc = -EPROTO);
 704
 705         rc = req_capsule_server_pack(&req->rq_pill);
 706         if (rc)
 707                 GOTO(out, rc);
 708
 709         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 710                 if (!data) {
 711                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 712                                   "libclient connection attempt");
 713                         GOTO(out, rc = -EPROTO);
 714                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 715                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 716                            data->ocd_version > LUSTRE_VERSION_CODE +
 717                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 718                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 719                                   "libclient connection attempt",
 720                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 721                                   "old" : "new",
 722                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 723                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 724                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 725                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 726                         data = req_capsule_server_sized_get(&req->rq_pill,
 727                                                         &RMF_CONNECT_DATA,
 728                                     offsetof(typeof(*data), ocd_version) +
 729                                              sizeof(data->ocd_version));
 730                         if (data) {
 731                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 732                                 data->ocd_version = LUSTRE_VERSION_CODE;
 733                         }
 734                         GOTO(out, rc = -EPROTO);
 735                 }
 736         }
 737
 738         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 739             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 740                 mds_conn = 1;
 741
 742         /* lctl gets a backstage, all-access pass. */
 743         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 744                 goto dont_check_exports;
 745
 746         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 747         if (!export)
 748                 goto no_export;
 749
 750         /* we've found an export in the hash */
 751         if (export->exp_connecting) { /* bug 9635, et. al. */
 752                 CWARN("%s: exp %p already connecting\n",
 753                       export->exp_obd->obd_name, export);
 754                 class_export_put(export);
 755                 export = NULL;
 756                 rc = -EALREADY;
 757         } else if (mds_conn && export->exp_connection) {
 758                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 759                         /* mds reconnected after failover */
 760                         CWARN("%s: received MDS connection from NID %s,"
 761                               " removing former export from NID %s\n",
 762                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 763                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 764                 else
 765                         /* new mds connection from the same nid */
 766                         CWARN("%s: received new MDS connection from NID %s,"
 767                               " removing former export from same NID\n",
 768                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 769                 class_fail_export(export);
 770                 class_export_put(export);
 771                 export = NULL;
 772                 rc = 0;
 773         } else if (export->exp_connection != NULL &&
 774                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 775                 /* in mds failover we have static uuid but nid can be
 776                  * changed*/
 777                 CWARN("%s: cookie %s seen on new NID %s when "
 778                       "existing NID %s is already connected\n",
 779                       target->obd_name, cluuid.uuid,
 780                       libcfs_nid2str(req->rq_peer.nid),
 781                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 782                 rc = -EALREADY;
 783                 class_export_put(export);
 784                 export = NULL;
 785         } else {
 786                 spin_lock(&export->exp_lock);
 787                 export->exp_connecting = 1;
 788                 spin_unlock(&export->exp_lock);
 789                 class_export_put(export);
 790                 LASSERT(export->exp_obd == target);
 791
 792                 rc = target_handle_reconnect(&conn, export, &cluuid);
 793         }
 794
 795         /* If we found an export, we already unlocked. */
 796         if (!export) {
 797 no_export:
 798                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 799         } else if (req->rq_export == NULL &&
 800                    atomic_read(&export->exp_rpc_count) > 0) {
 801                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 802                       target->obd_name, cluuid.uuid,
 803                       libcfs_nid2str(req->rq_peer.nid),
 804                       export, atomic_read(&export->exp_refcount));
 805                 GOTO(out, rc = -EBUSY);
 806         } else if (req->rq_export != NULL &&
 807                    (atomic_read(&export->exp_rpc_count) > 1)) {
 808                 /* the current connect rpc has increased exp_rpc_count */
 809                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 810                       target->obd_name, cluuid.uuid,
 811                       libcfs_nid2str(req->rq_peer.nid),
 812                       export, atomic_read(&export->exp_rpc_count) - 1);
 813                 spin_lock(&export->exp_lock);
 814                 if (req->rq_export->exp_conn_cnt <
 815                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 816                         /* try to abort active requests */
 817                         req->rq_export->exp_abort_active_req = 1;
 818                 spin_unlock(&export->exp_lock);
 819                 GOTO(out, rc = -EBUSY);
 820         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 821                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 822                        "cookies not random?\n", target->obd_name,
 823                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 824                 GOTO(out, rc = -EALREADY);
 825         } else {
 826                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 827         }
 828
 829         if (rc < 0) {
 830                 GOTO(out, rc);
 831         }
 832
 833         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 834                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 835               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 836               export, (long)cfs_time_current_sec(),
 837               export ? (long)export->exp_last_request_time : 0);
 838
 839         /* Tell the client if we're in recovery. */
 840         if (target->obd_recovering) {
 841                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 842                 /* If this is the first time a client connects,
 843                    reset the recovery timer */
 844                 if (rc == 0)
 845                         target_start_and_reset_recovery_timer(target, req,
 846                                                               !export);
 847         }
 848
 849         /* We want to handle EALREADY but *not* -EALREADY from
 850          * target_handle_reconnect(), return reconnection state in a flag */
 851         if (rc == EALREADY) {
 852                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 853                 rc = 0;
 854         } else {
 855                 LASSERT(rc == 0);
 856         }
 857
 858         /* Tell the client if we support replayable requests */
 859         if (target->obd_replayable)
 860                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 861         client_nid = &req->rq_peer.nid;
 862
 863         if (export == NULL) {
 864                 if (target->obd_recovering) {
 865                         cfs_time_t t;
 866
 867                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 868                         t = cfs_time_sub(t, cfs_time_current());
 869                         CERROR("%s: denying connection for new client %s (%s): "
 870                                "%d clients in recovery for "CFS_TIME_T"s\n",
 871                                target->obd_name,
 872                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 873                                target->obd_recoverable_clients,
 874                                cfs_duration_sec(t));
 875                         rc = -EBUSY;
 876                 } else {
 877 dont_check_exports:
 878                         rc = obd_connect(req->rq_svc_thread->t_env,
 879                                          &export, target, &cluuid, data,
 880                                          client_nid);
 881                         if (rc == 0)
 882                                 conn.cookie = export->exp_handle.h_cookie;
 883                 }
 884         } else {
 885                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 886                                    export, target, &cluuid, data, client_nid);
 887                 if (rc == 0)
 888                         /* prevous done via class_conn2export */
 889                         class_export_get(export);
 890         }
 891         if (rc)
 892                 GOTO(out, rc);
 893         /* Return only the parts of obd_connect_data that we understand, so the
 894          * client knows that we don't understand the rest. */
 895         if (data) {
 896                  tmpdata = req_capsule_server_get(&req->rq_pill,
 897                                                   &RMF_CONNECT_DATA);
 898                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 899                  *tmpdata = *data;
 900         }
 901
 902         /* If all else goes well, this is our RPC return code. */
 903         req->rq_status = 0;
 904
 905         lustre_msg_set_handle(req->rq_repmsg, &conn);
 906
 907         /* If the client and the server are the same node, we will already
 908          * have an export that really points to the client's DLM export,
 909          * because we have a shared handles table.
 910          *
 911          * XXX this will go away when shaver stops sending the "connect" handle
 912          * in the real "remote handle" field of the request --phik 24 Apr 2003
 913          */
 914         if (req->rq_export != NULL)
 915                 class_export_put(req->rq_export);
 916
 917         req->rq_export = export;
 918
 919         spin_lock(&export->exp_lock);
 920         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 921                 spin_unlock(&export->exp_lock);
 922                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 923                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 924                        export->exp_conn_cnt,
 925                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 926
 927                 GOTO(out, rc = -EALREADY);
 928         }
 929         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 930         export->exp_abort_active_req = 0;
 931
 932         /* request from liblustre?  Don't evict it for not pinging. */
 933         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 934                 export->exp_libclient = 1;
 935                 spin_unlock(&export->exp_lock);
 936
 937                 spin_lock(&target->obd_dev_lock);
 938                 list_del_init(&export->exp_obd_chain_timed);
 939                 spin_unlock(&target->obd_dev_lock);
 940         } else {
 941                 spin_unlock(&export->exp_lock);
 942         }
 943
 944         if (export->exp_connection != NULL)
 945                 ptlrpc_connection_put(export->exp_connection);
 946         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 947                                                        req->rq_self,
 948                                                        &remote_uuid);
 949         if (hlist_unhashed(&export->exp_nid_hash)) {
 950                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 951                                        &export->exp_connection->c_peer.nid,
 952                                        &export->exp_nid_hash);
 953         }
 954
 955         spin_lock_bh(&target->obd_processing_task_lock);
 956         if (target->obd_recovering && !export->exp_in_recovery) {
 957                 spin_lock(&export->exp_lock);
 958                 export->exp_in_recovery = 1;
 959                 export->exp_req_replay_needed = 1;
 960                 export->exp_lock_replay_needed = 1;
 961                 spin_unlock(&export->exp_lock);
 962                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 963                      && (data->ocd_transno == 0))
 964                         CWARN("Connect with zero transno!\n");
 965
 966                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 967                      && data->ocd_transno < target->obd_next_recovery_transno)
 968                         target->obd_next_recovery_transno = data->ocd_transno;
 969                 target->obd_connected_clients++;
 970                 /* each connected client is counted as recoverable */
 971                 target->obd_recoverable_clients++;
 972                 atomic_inc(&target->obd_req_replay_clients);
 973                 atomic_inc(&target->obd_lock_replay_clients);
 974                 if (target->obd_connected_clients ==
 975                     target->obd_max_recoverable_clients)
 976                         cfs_waitq_signal(&target->obd_next_transno_waitq);
 977         }
 978         spin_unlock_bh(&target->obd_processing_task_lock);
 979         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 980         conn = *tmp;
 981
 982         if (export->exp_imp_reverse != NULL) {
 983                 /* destroyed import can be still referenced in ctxt */
 984                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 985                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 986
 987                 /* in some recovery senarios, previous ctx init rpc handled
 988                  * in sptlrpc_target_export_check() might be used to install
 989                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 990                  * using the same gss ctx could reach here and following new
 991                  * reverse import. note all reverse ctx in new/old import are
 992                  * actually based on the same gss ctx. so we invalidate ctx
 993                  * here before destroy import, otherwise flush old import will
 994                  * lead to remote reverse ctx be destroied, thus the reverse
 995                  * ctx of new import will lost its peer.
 996                  * there might be a better way to deal with this???
 997                  */
 998                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 999
1000                 client_destroy_import(export->exp_imp_reverse);
1001         }
1002
1003         /* for the rest part, we return -ENOTCONN in case of errors
1004          * in order to let client initialize connection again.
1005          */
1006         revimp = export->exp_imp_reverse = class_new_import(target);
1007         if (!revimp) {
1008                 CERROR("fail to alloc new reverse import.\n");
1009                 GOTO(out, rc = -ENOTCONN);
1010         }
1011
1012         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1013         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1014         revimp->imp_remote_handle = conn;
1015         revimp->imp_dlm_fake = 1;
1016         revimp->imp_state = LUSTRE_IMP_FULL;
1017
1018         /* unknown versions will be caught in
1019          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1020         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1021
1022         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1023             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1024                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1025         else
1026                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1027
1028         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1029         if (rc) {
1030                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1031                 export->exp_imp_reverse = NULL;
1032                 class_destroy_import(revimp);
1033         }
1034
1035         class_import_put(revimp);
1036 out:
1037         if (export) {
1038                 spin_lock(&export->exp_lock);
1039                 export->exp_connecting = 0;
1040                 spin_unlock(&export->exp_lock);
1041         }
1042         if (targref)
1043                 class_decref(targref, __FUNCTION__, cfs_current());
1044         if (rc)
1045                 req->rq_status = rc;
1046         RETURN(rc);
1047 }
1048
1049 int target_handle_disconnect(struct ptlrpc_request *req)
1050 {
1051         int rc;
1052         ENTRY;
1053
1054         rc = req_capsule_server_pack(&req->rq_pill);
1055         if (rc)
1056                 RETURN(rc);
1057
1058         /* keep the rq_export around so we can send the reply */
1059         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1060
1061         RETURN(0);
1062 }
1063
1064 void target_destroy_export(struct obd_export *exp)
1065 {
1066         /* exports created from last_rcvd data, and "fake"
1067            exports created by lctl don't have an import */
1068         if (exp->exp_imp_reverse != NULL)
1069                 client_destroy_import(exp->exp_imp_reverse);
1070
1071         /* We cancel locks at disconnect time, but this will catch any locks
1072          * granted in a race with recovery-induced disconnect. */
1073         if (exp->exp_obd->obd_namespace != NULL)
1074                 ldlm_cancel_locks_for_export(exp);
1075 }
1076
1077 /*
1078  * Recovery functions
1079  */
1080
1081 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1082 {
1083         struct ptlrpc_request *copy_req;
1084         struct lustre_msg *copy_reqmsg;
1085         struct ptlrpc_user_desc *udesc = NULL;
1086
1087         OBD_ALLOC_PTR(copy_req);
1088         if (!copy_req)
1089                 return NULL;
1090         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1091         if (!copy_reqmsg){
1092                 OBD_FREE_PTR(copy_req);
1093                 return NULL;
1094         }
1095
1096         if (orig_req->rq_user_desc) {
1097                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1098
1099                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1100                 if (!udesc) {
1101                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1102                         OBD_FREE_PTR(copy_req);
1103                         return NULL;
1104                 }
1105                 memcpy(udesc, orig_req->rq_user_desc,
1106                        sptlrpc_user_desc_size(ngroups));
1107         }
1108
1109         *copy_req = *orig_req;
1110         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1111         copy_req->rq_reqmsg = copy_reqmsg;
1112         copy_req->rq_user_desc = udesc;
1113
1114         class_export_rpc_get(copy_req->rq_export);
1115         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1116         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1117         sptlrpc_svc_ctx_addref(copy_req);
1118
1119         if (copy_req->rq_reply_state) {
1120                 /* the copied req takes over the reply state */
1121                 orig_req->rq_reply_state = NULL;
1122                 /* to catch further access */
1123                 orig_req->rq_repmsg = NULL;
1124                 orig_req->rq_replen = 0;
1125         }
1126
1127         return copy_req;
1128 }
1129
1130 void ptlrpc_free_clone(struct ptlrpc_request *req)
1131 {
1132         LASSERT(list_empty(&req->rq_replay_list));
1133
1134         ptlrpc_req_drop_rs(req);
1135         sptlrpc_svc_ctx_decref(req);
1136         class_export_rpc_put(req->rq_export);
1137         list_del_init(&req->rq_list);
1138
1139         if (req->rq_user_desc) {
1140                 int ngroups = req->rq_user_desc->pud_ngroups;
1141                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1142         }
1143         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1144         OBD_FREE_PTR(req);
1145 }
1146
1147 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1148 {
1149         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1150         struct obd_export     *exp = req->rq_export;
1151         struct ptlrpc_request *reqiter;
1152         int                    dup = 0;
1153
1154         LASSERT(exp);
1155
1156         spin_lock(&exp->exp_lock);
1157         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1158                             rq_replay_list) {
1159                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1160                         dup = 1;
1161                         break;
1162                 }
1163         }
1164
1165         if (dup) {
1166                 /* we expect it with RESENT and REPLAY flags */
1167                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1168                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1169                         CERROR("invalid flags %x of resent replay\n",
1170                                lustre_msg_get_flags(req->rq_reqmsg));
1171         } else {
1172                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1173         }
1174
1175         spin_unlock(&exp->exp_lock);
1176         return dup;
1177 }
1178
1179 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1180 {
1181         LASSERT(!list_empty(&req->rq_replay_list));
1182         LASSERT(req->rq_export);
1183
1184         spin_lock(&req->rq_export->exp_lock);
1185         list_del_init(&req->rq_replay_list);
1186         spin_unlock(&req->rq_export->exp_lock);
1187 }
1188
1189 #ifdef __KERNEL__
1190 static void target_finish_recovery(struct obd_device *obd)
1191 {
1192         ENTRY;
1193         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1194                       obd->obd_name);
1195
1196         ldlm_reprocess_all_ns(obd->obd_namespace);
1197         spin_lock_bh(&obd->obd_processing_task_lock);
1198         if (list_empty(&obd->obd_req_replay_queue) &&
1199             list_empty(&obd->obd_lock_replay_queue) &&
1200             list_empty(&obd->obd_final_req_queue)) {
1201                 obd->obd_processing_task = 0;
1202         } else {
1203                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1204                        obd->obd_name,
1205                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1206                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1207                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1208                 spin_unlock_bh(&obd->obd_processing_task_lock);
1209                 LBUG();
1210         }
1211         spin_unlock_bh(&obd->obd_processing_task_lock);
1212
1213         /* when recovery finished, cleanup orphans on mds and ost */
1214         if (OBT(obd) && OBP(obd, postrecov)) {
1215                 int rc = OBP(obd, postrecov)(obd);
1216                 if (rc < 0)
1217                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1218                                       obd->obd_name, rc);
1219         }
1220
1221         obd->obd_recovery_end = cfs_time_current_sec();
1222         EXIT;
1223 }
1224
1225 static void abort_req_replay_queue(struct obd_device *obd)
1226 {
1227         struct ptlrpc_request *req, *n;
1228         struct list_head abort_list;
1229
1230         CFS_INIT_LIST_HEAD(&abort_list);
1231         spin_lock_bh(&obd->obd_processing_task_lock);
1232         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1233         spin_unlock_bh(&obd->obd_processing_task_lock);
1234         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1235                 DEBUG_REQ(D_WARNING, req, "aborted:");
1236                 req->rq_status = -ENOTCONN;
1237                 if (ptlrpc_error(req)) {
1238                         DEBUG_REQ(D_ERROR, req,
1239                                   "failed abort_req_reply; skipping");
1240                 }
1241                 target_exp_dequeue_req_replay(req);
1242                 ptlrpc_free_clone(req);
1243         }
1244 }
1245
1246 static void abort_lock_replay_queue(struct obd_device *obd)
1247 {
1248         struct ptlrpc_request *req, *n;
1249         struct list_head abort_list;
1250
1251         CFS_INIT_LIST_HEAD(&abort_list);
1252         spin_lock_bh(&obd->obd_processing_task_lock);
1253         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1254         spin_unlock_bh(&obd->obd_processing_task_lock);
1255         list_for_each_entry_safe(req, n, &abort_list, rq_list){
1256                 DEBUG_REQ(D_ERROR, req, "aborted:");
1257                 req->rq_status = -ENOTCONN;
1258                 if (ptlrpc_error(req)) {
1259                         DEBUG_REQ(D_ERROR, req,
1260                                   "failed abort_lock_reply; skipping");
1261                 }
1262                 ptlrpc_free_clone(req);
1263         }
1264 }
1265 #endif
1266
1267 /* Called from a cleanup function if the device is being cleaned up
1268    forcefully.  The exports should all have been disconnected already,
1269    the only thing left to do is
1270      - clear the recovery flags
1271      - cancel the timer
1272      - free queued requests and replies, but don't send replies
1273    Because the obd_stopping flag is set, no new requests should be received.
1274
1275 */
1276 void target_cleanup_recovery(struct obd_device *obd)
1277 {
1278         struct ptlrpc_request *req, *n;
1279         struct list_head clean_list;
1280         ENTRY;
1281
1282         LASSERT(obd->obd_stopping);
1283
1284         CFS_INIT_LIST_HEAD(&clean_list);
1285         spin_lock_bh(&obd->obd_processing_task_lock);
1286         if (!obd->obd_recovering) {
1287                 spin_unlock_bh(&obd->obd_processing_task_lock);
1288                 EXIT;
1289                 return;
1290         }
1291         obd->obd_recovering = obd->obd_abort_recovery = 0;
1292         target_cancel_recovery_timer(obd);
1293
1294         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1295         spin_unlock_bh(&obd->obd_processing_task_lock);
1296
1297         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1298                 LASSERT(req->rq_reply_state == 0);
1299                 target_exp_dequeue_req_replay(req);
1300                 ptlrpc_free_clone(req);
1301         }
1302
1303         spin_lock_bh(&obd->obd_processing_task_lock);
1304         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1305         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1306         spin_unlock_bh(&obd->obd_processing_task_lock);
1307
1308         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1309                 LASSERT(req->rq_reply_state == 0);
1310                 ptlrpc_free_clone(req);
1311         }
1312
1313         EXIT;
1314 }
1315
1316 /* obd_processing_task_lock should be held */
1317 void target_cancel_recovery_timer(struct obd_device *obd)
1318 {
1319         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1320         cfs_timer_disarm(&obd->obd_recovery_timer);
1321 }
1322
1323 /* extend = 1 means require at least "duration" seconds left in the timer,
1324    extend = 0 means set the total duration (start_recovery_timer) */
1325 static void reset_recovery_timer(struct obd_device *obd, int duration,
1326                                  int extend)
1327 {
1328         cfs_time_t now = cfs_time_current_sec();
1329         cfs_duration_t left;
1330
1331         spin_lock_bh(&obd->obd_processing_task_lock);
1332         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1333                 spin_unlock_bh(&obd->obd_processing_task_lock);
1334                 return;
1335         }
1336
1337         left = cfs_time_sub(obd->obd_recovery_end, now);
1338
1339         if (extend && (duration > left))
1340                 obd->obd_recovery_timeout += duration - left;
1341         else if (!extend && (duration > obd->obd_recovery_timeout))
1342                 /* Track the client's largest expected replay time */
1343                 obd->obd_recovery_timeout = duration;
1344 #ifdef CRAY_XT3
1345         /*
1346          * If total recovery time already exceed the
1347          * obd_recovery_max_time, then CRAY XT3 will
1348          * abort the recovery
1349          */
1350         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1351                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1352 #endif
1353         obd->obd_recovery_end = obd->obd_recovery_start +
1354                                 obd->obd_recovery_timeout;
1355         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1356             cfs_time_before(now, obd->obd_recovery_end)) {
1357                 left = cfs_time_sub(obd->obd_recovery_end, now);
1358                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1359         }
1360         spin_unlock_bh(&obd->obd_processing_task_lock);
1361         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1362                obd->obd_name, (unsigned)left);
1363 }
1364
1365 static void check_and_start_recovery_timer(struct obd_device *obd)
1366 {
1367         spin_lock_bh(&obd->obd_processing_task_lock);
1368         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1369                 spin_unlock_bh(&obd->obd_processing_task_lock);
1370                 return;
1371         }
1372         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1373         obd->obd_recovery_start = cfs_time_current_sec();
1374         /* minimum */
1375         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1376         spin_unlock_bh(&obd->obd_processing_task_lock);
1377
1378         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1379 }
1380
1381 /* Reset the timer with each new client connection */
1382 /*
1383  * This timer is actually reconnect_timer, which is for making sure
1384  * the total recovery window is at least as big as my reconnect
1385  * attempt timing. So the initial recovery time_out will be set to
1386  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1387  * from client is bigger than this, then the recovery time_out will
1388  * be extend to make sure the client could be reconnected, in the
1389  * process, the timeout from the new client should be ignored.
1390  */
1391
1392 static void
1393 target_start_and_reset_recovery_timer(struct obd_device *obd,
1394                                       struct ptlrpc_request *req,
1395                                       int new_client)
1396 {
1397         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1398
1399         if (!new_client && service_time)
1400                 /* Teach server about old server's estimates, as first guess
1401                  * at how long new requests will take. */
1402                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1403                        service_time);
1404
1405         check_and_start_recovery_timer(obd);
1406
1407         /* convert the service time to rpc timeout,
1408          * reuse service_time to limit stack usage */
1409         service_time = at_est2timeout(service_time);
1410
1411         /* We expect other clients to timeout within service_time, then try
1412          * to reconnect, then try the failover server.  The max delay between
1413          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1414         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1415                              INITIAL_CONNECT_TIMEOUT);
1416         if (service_time > obd->obd_recovery_timeout && !new_client)
1417                 reset_recovery_timer(obd, service_time, 0);
1418 }
1419
1420 #ifdef __KERNEL__
1421 static int check_for_next_transno(struct obd_device *obd)
1422 {
1423         struct ptlrpc_request *req = NULL;
1424         int wake_up = 0, connected, completed, queue_len, max;
1425         __u64 next_transno, req_transno;
1426         ENTRY;
1427         spin_lock_bh(&obd->obd_processing_task_lock);
1428
1429         if (!list_empty(&obd->obd_req_replay_queue)) {
1430                 req = list_entry(obd->obd_req_replay_queue.next,
1431                                  struct ptlrpc_request, rq_list);
1432                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1433         } else {
1434                 req_transno = 0;
1435         }
1436
1437         max = obd->obd_max_recoverable_clients;
1438         connected = obd->obd_connected_clients;
1439         completed = connected - obd->obd_recoverable_clients;
1440         queue_len = obd->obd_requests_queued_for_recovery;
1441         next_transno = obd->obd_next_recovery_transno;
1442
1443         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1444                "req_transno: "LPU64", next_transno: "LPU64"\n",
1445                max, connected, completed, queue_len, req_transno, next_transno);
1446
1447         if (obd->obd_abort_recovery) {
1448                 CDEBUG(D_HA, "waking for aborted recovery\n");
1449                 wake_up = 1;
1450         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1451                 CDEBUG(D_HA, "waking for completed recovery\n");
1452                 wake_up = 1;
1453         } else if (req_transno == next_transno) {
1454                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1455                 wake_up = 1;
1456         } else if (queue_len + completed == max) {
1457                 /* handle gaps occured due to lost reply. It is allowed gaps
1458                  * because all clients are connected and there will be resend
1459                  * for missed transaction */
1460                 LASSERTF(req_transno >= next_transno,
1461                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1462                          req_transno, next_transno);
1463
1464                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1465                        "waking for skipped transno (skip: "LPD64
1466                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1467                        next_transno, queue_len, completed, connected, req_transno);
1468                 obd->obd_next_recovery_transno = req_transno;
1469                 wake_up = 1;
1470         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1471                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1472                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1473                 obd->obd_next_recovery_transno = req_transno;
1474                 wake_up = 1;
1475         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1476                 /* some clients haven't connected in time, but we can try
1477                  * to replay requests that demand on already committed ones
1478                  * also, we can replay first non-committed transation */
1479                 LASSERT(req_transno != 0);
1480                 if (obd->obd_version_recov ||
1481                     req_transno == obd->obd_last_committed + 1) {
1482                         obd->obd_next_recovery_transno = req_transno;
1483                 } else if (req_transno > obd->obd_last_committed) {
1484                         /* can't continue recovery: have no needed transno */
1485                         obd->obd_abort_recovery = 1;
1486                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1487                                "connected: %d, completed: %d, queue_len: %d, "
1488                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1489                                max, connected, completed, queue_len,
1490                                req_transno, next_transno);
1491                 }
1492                 wake_up = 1;
1493         }
1494
1495         spin_unlock_bh(&obd->obd_processing_task_lock);
1496         return wake_up;
1497 }
1498
1499 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1500 {
1501         struct l_wait_info lwi = { 0 };
1502         struct ptlrpc_request *req;
1503
1504         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1505                obd->obd_next_recovery_transno);
1506         l_wait_event(obd->obd_next_transno_waitq,
1507                      check_for_next_transno(obd), &lwi);
1508
1509         spin_lock_bh(&obd->obd_processing_task_lock);
1510         if (obd->obd_abort_recovery) {
1511                 req = NULL;
1512         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1513                 req = list_entry(obd->obd_req_replay_queue.next,
1514                                  struct ptlrpc_request, rq_list);
1515                 list_del_init(&req->rq_list);
1516                 obd->obd_requests_queued_for_recovery--;
1517         } else {
1518                 req = NULL;
1519         }
1520         spin_unlock_bh(&obd->obd_processing_task_lock);
1521         RETURN(req);
1522 }
1523
1524 static int check_for_next_lock(struct obd_device *obd)
1525 {
1526         struct ptlrpc_request *req = NULL;
1527         int wake_up = 0;
1528
1529         spin_lock_bh(&obd->obd_processing_task_lock);
1530         if (!list_empty(&obd->obd_lock_replay_queue)) {
1531                 req = list_entry(obd->obd_lock_replay_queue.next,
1532                                  struct ptlrpc_request, rq_list);
1533                 CDEBUG(D_HA, "waking for next lock\n");
1534                 wake_up = 1;
1535         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1536                 CDEBUG(D_HA, "waking for completed lock replay\n");
1537                 wake_up = 1;
1538         } else if (obd->obd_abort_recovery) {
1539                 CDEBUG(D_HA, "waking for aborted recovery\n");
1540                 wake_up = 1;
1541         }
1542         spin_unlock_bh(&obd->obd_processing_task_lock);
1543
1544         return wake_up;
1545 }
1546
1547 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1548 {
1549         struct l_wait_info lwi = { 0 };
1550         struct ptlrpc_request *req;
1551
1552         CDEBUG(D_HA, "Waiting for lock\n");
1553         l_wait_event(obd->obd_next_transno_waitq,
1554                      check_for_next_lock(obd), &lwi);
1555
1556         spin_lock_bh(&obd->obd_processing_task_lock);
1557         if (obd->obd_abort_recovery) {
1558                 req = NULL;
1559         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1560                 req = list_entry(obd->obd_lock_replay_queue.next,
1561                                  struct ptlrpc_request, rq_list);
1562                 list_del_init(&req->rq_list);
1563         } else {
1564                 req = NULL;
1565         }
1566         spin_unlock_bh(&obd->obd_processing_task_lock);
1567         return req;
1568 }
1569
1570 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1571 {
1572         struct ptlrpc_request *req;
1573
1574         spin_lock_bh(&obd->obd_processing_task_lock);
1575         if (!list_empty(&obd->obd_final_req_queue)) {
1576                 req = list_entry(obd->obd_final_req_queue.next,
1577                                  struct ptlrpc_request, rq_list);
1578                 list_del_init(&req->rq_list);
1579                 if (req->rq_export->exp_in_recovery) {
1580                         spin_lock(&req->rq_export->exp_lock);
1581                         req->rq_export->exp_in_recovery = 0;
1582                         spin_unlock(&req->rq_export->exp_lock);
1583                 }
1584         } else {
1585                 req = NULL;
1586         }
1587         spin_unlock_bh(&obd->obd_processing_task_lock);
1588         return req;
1589 }
1590
1591 static inline int req_vbr_done(struct obd_export *exp)
1592 {
1593         return (exp->exp_vbr_failed == 0);
1594 }
1595
1596 static inline int req_replay_done(struct obd_export *exp)
1597 {
1598         return (exp->exp_req_replay_needed == 0);
1599 }
1600
1601 static inline int lock_replay_done(struct obd_export *exp)
1602 {
1603         return (exp->exp_lock_replay_needed == 0);
1604 }
1605
1606 static inline int connect_done(struct obd_export *exp)
1607 {
1608         return (exp->exp_in_recovery != 0);
1609 }
1610
1611 static int check_for_clients(struct obd_device *obd)
1612 {
1613         if (obd->obd_abort_recovery || obd->obd_version_recov)
1614                 return 1;
1615         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1616         if (obd->obd_no_conn == 0 &&
1617             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1618                 return 1;
1619         return 0;
1620 }
1621
1622 static int handle_recovery_req(struct ptlrpc_thread *thread,
1623                                struct ptlrpc_request *req,
1624                                svc_handler_t handler)
1625 {
1626         int rc;
1627         ENTRY;
1628
1629         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1630         if (rc) {
1631                 CERROR("Failure to initialize session: %d\n", rc);
1632                 return rc;
1633         }
1634         req->rq_session.lc_thread = thread;
1635         lu_context_enter(&req->rq_session);
1636         req->rq_svc_thread = thread;
1637         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1638
1639         /* thread context */
1640         lu_context_enter(&thread->t_env->le_ctx);
1641         (void)handler(req);
1642         lu_context_exit(&thread->t_env->le_ctx);
1643
1644         lu_context_exit(&req->rq_session);
1645         lu_context_fini(&req->rq_session);
1646         /* don't reset timer for final stage */
1647         if (!req_replay_done(req->rq_export) ||
1648             !lock_replay_done(req->rq_export))
1649                 reset_recovery_timer(class_exp2obd(req->rq_export),
1650                                      AT_OFF ? obd_timeout :
1651                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1652
1653         /**
1654          * bz18031: increase next_recovery_transno before ptlrpc_free_clone()
1655          * will drop exp_rpc reference
1656          */
1657         if (!req_replay_done(req->rq_export)) {
1658                 spin_lock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1659                 req->rq_export->exp_obd->obd_next_recovery_transno++;
1660                 spin_unlock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1661                 target_exp_dequeue_req_replay(req);
1662         }
1663         ptlrpc_free_clone(req);
1664         RETURN(0);
1665 }
1666
1667 static int target_recovery_thread(void *arg)
1668 {
1669         struct lu_target *lut = arg;
1670         struct obd_device *obd = lut->lut_obd;
1671         struct ptlrpc_request *req;
1672         struct target_recovery_data *trd = &obd->obd_recovery_data;
1673         struct l_wait_info lwi = { 0 };
1674         unsigned long delta;
1675         unsigned long flags;
1676         struct lu_env env;
1677         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1678         int rc = 0;
1679         ENTRY;
1680
1681         cfs_daemonize("tgt_recov");
1682
1683         SIGNAL_MASK_LOCK(current, flags);
1684         sigfillset(&current->blocked);
1685         RECALC_SIGPENDING;
1686         SIGNAL_MASK_UNLOCK(current, flags);
1687
1688         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1689         if (rc)
1690                 RETURN(rc);
1691
1692         thread->t_env = &env;
1693         env.le_ctx.lc_thread = thread;
1694
1695         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1696                cfs_curproc_pid());
1697         trd->trd_processing_task = cfs_curproc_pid();
1698
1699         obd->obd_recovering = 1;
1700         complete(&trd->trd_starting);
1701
1702         /* first of all, we have to know the first transno to replay */
1703         obd->obd_abort_recovery = 0;
1704         l_wait_event(obd->obd_next_transno_waitq,
1705                      check_for_clients(obd), &lwi);
1706
1707         /* If some clients haven't connected in time, evict them */
1708         if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) {
1709                 CWARN("Some clients haven't connect in time (%d/%d),"
1710                        "evict them\n", obd->obd_connected_clients,
1711                        obd->obd_max_recoverable_clients);
1712                 class_disconnect_stale_exports(obd, connect_done,
1713                                                exp_flags_from_obd(obd) |
1714                                                OBD_OPT_ABORT_RECOV);
1715                 /**
1716                  * if recovery proceeds with versions then some clients may be
1717                  * timed out waiting for others and trying to reconnect.
1718                  * Extend timer for such reconnect cases.
1719                  */
1720                 if (obd->obd_version_recov)
1721                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1);
1722         }
1723
1724         /* next stage: replay requests */
1725         delta = jiffies;
1726         obd->obd_req_replaying = 1;
1727         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1728                atomic_read(&obd->obd_req_replay_clients),
1729                obd->obd_next_recovery_transno);
1730         while ((req = target_next_replay_req(obd))) {
1731                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1732                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1733                           lustre_msg_get_transno(req->rq_reqmsg),
1734                           libcfs_nid2str(req->rq_peer.nid));
1735                 handle_recovery_req(thread, req,
1736                                     trd->trd_recovery_handler);
1737                 obd->obd_replayed_requests++;
1738         }
1739
1740         /* If some clients haven't replayed requests in time, evict them */
1741         if (obd->obd_abort_recovery) {
1742                 CDEBUG(D_WARNING, "req replay is aborted\n");
1743                 class_disconnect_stale_exports(obd, req_replay_done,
1744                                                exp_flags_from_obd(obd) |
1745                                                OBD_OPT_ABORT_RECOV);
1746                 abort_req_replay_queue(obd);
1747         }
1748         LASSERT(list_empty(&obd->obd_req_replay_queue));
1749
1750         /* The second stage: replay locks */
1751         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1752                atomic_read(&obd->obd_lock_replay_clients));
1753         while ((req = target_next_replay_lock(obd))) {
1754                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1755                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1756                           libcfs_nid2str(req->rq_peer.nid));
1757                 handle_recovery_req(thread, req,
1758                                     trd->trd_recovery_handler);
1759                 obd->obd_replayed_locks++;
1760         }
1761
1762         /* If some clients haven't replayed requests in time, evict them */
1763         if (obd->obd_abort_recovery) {
1764                 CERROR("lock replay is aborted\n");
1765                 class_disconnect_stale_exports(obd, lock_replay_done,
1766                                                exp_flags_from_obd(obd) |
1767                                                OBD_OPT_ABORT_RECOV);
1768                 abort_lock_replay_queue(obd);
1769         }
1770         LASSERT(list_empty(&obd->obd_lock_replay_queue));
1771
1772         /* The third stage: reply on final pings */
1773         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1774         /** evict exports failed VBR */
1775         class_disconnect_stale_exports(obd, req_vbr_done,
1776                                        exp_flags_from_obd(obd) |
1777                                        OBD_OPT_ABORT_RECOV);
1778         /** Update server last boot epoch */
1779         lut_boot_epoch_update(lut);
1780         /* We drop recoverying flag to forward all new requests
1781          * to regular mds_handle() since now */
1782         spin_lock_bh(&obd->obd_processing_task_lock);
1783         obd->obd_recovering = obd->obd_abort_recovery = 0;
1784         target_cancel_recovery_timer(obd);
1785         spin_unlock_bh(&obd->obd_processing_task_lock);
1786         while ((req = target_next_final_ping(obd))) {
1787                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1788                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1789                           libcfs_nid2str(req->rq_peer.nid));
1790                 handle_recovery_req(thread, req,
1791                                     trd->trd_recovery_handler);
1792         }
1793
1794         delta = (jiffies - delta) / HZ;
1795         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1796               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1797         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1798         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1799         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1800                 CWARN("too long recovery - read logs\n");
1801                 libcfs_debug_dumplog();
1802         }
1803
1804         target_finish_recovery(obd);
1805
1806         lu_context_fini(&env.le_ctx);
1807         trd->trd_processing_task = 0;
1808         complete(&trd->trd_finishing);
1809         RETURN(rc);
1810 }
1811
1812 static int target_start_recovery_thread(struct lu_target *lut,
1813                                         svc_handler_t handler)
1814 {
1815         struct obd_device *obd = lut->lut_obd;
1816         int rc = 0;
1817         struct target_recovery_data *trd = &obd->obd_recovery_data;
1818
1819         memset(trd, 0, sizeof(*trd));
1820         init_completion(&trd->trd_starting);
1821         init_completion(&trd->trd_finishing);
1822         trd->trd_recovery_handler = handler;
1823
1824         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1825                 wait_for_completion(&trd->trd_starting);
1826                 LASSERT(obd->obd_recovering != 0);
1827         } else
1828                 rc = -ECHILD;
1829
1830         return rc;
1831 }
1832
1833 void target_stop_recovery_thread(struct obd_device *obd)
1834 {
1835         spin_lock_bh(&obd->obd_processing_task_lock);
1836         if (obd->obd_recovery_data.trd_processing_task > 0) {
1837                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1838                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1839                 obd->obd_abort_recovery = 1;
1840                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1841                 spin_unlock_bh(&obd->obd_processing_task_lock);
1842                 wait_for_completion(&trd->trd_finishing);
1843         } else {
1844                 spin_unlock_bh(&obd->obd_processing_task_lock);
1845         }
1846 }
1847
1848 void target_recovery_fini(struct obd_device *obd)
1849 {
1850         class_disconnect_exports(obd);
1851         target_stop_recovery_thread(obd);
1852         target_cleanup_recovery(obd);
1853 }
1854 EXPORT_SYMBOL(target_recovery_fini);
1855
1856 static void target_recovery_expired(unsigned long castmeharder)
1857 {
1858         struct obd_device *obd = (struct obd_device *)castmeharder;
1859         CDEBUG(D_HA, "%s: recovery timed out; %d clients never reconnected "
1860                "after %lds (%d clients did)\n",
1861                obd->obd_name, obd->obd_recoverable_clients,
1862                cfs_time_current_sec()- obd->obd_recovery_start,
1863                obd->obd_connected_clients);
1864
1865         spin_lock_bh(&obd->obd_processing_task_lock);
1866         obd->obd_version_recov = 1;
1867         CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n",
1868                atomic_read(&obd->obd_req_replay_clients),
1869                obd->obd_next_recovery_transno);
1870         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1871         spin_unlock_bh(&obd->obd_processing_task_lock);
1872 }
1873
1874 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1875 {
1876         struct obd_device *obd = lut->lut_obd;
1877         if (obd->obd_max_recoverable_clients == 0)
1878                 return;
1879
1880         CWARN("RECOVERY: service %s, %d recoverable clients, "
1881               "last_transno "LPU64"\n", obd->obd_name,
1882               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1883         LASSERT(obd->obd_stopping == 0);
1884         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1885         obd->obd_recovery_start = 0;
1886         obd->obd_recovery_end = 0;
1887         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1888         /* bz13079: this should be set to desired value for ost but not for mds */
1889         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1890         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1891         target_start_recovery_thread(lut, handler);
1892 }
1893 EXPORT_SYMBOL(target_recovery_init);
1894
1895 #endif
1896
1897 static int target_process_req_flags(struct obd_device *obd,
1898                                     struct ptlrpc_request *req)
1899 {
1900         struct obd_export *exp = req->rq_export;
1901         LASSERT(exp != NULL);
1902         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1903                 /* client declares he's ready to replay locks */
1904                 spin_lock_bh(&obd->obd_processing_task_lock);
1905                 if (exp->exp_req_replay_needed) {
1906                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1907                         spin_lock(&exp->exp_lock);
1908                         exp->exp_req_replay_needed = 0;
1909                         spin_unlock(&exp->exp_lock);
1910                         atomic_dec(&obd->obd_req_replay_clients);
1911                         LASSERT(obd->obd_recoverable_clients > 0);
1912                         obd->obd_recoverable_clients--;
1913                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1914                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1915                 }
1916                 spin_unlock_bh(&obd->obd_processing_task_lock);
1917         }
1918         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1919                 /* client declares he's ready to complete recovery
1920                  * so, we put the request on th final queue */
1921                 spin_lock_bh(&obd->obd_processing_task_lock);
1922                 if (exp->exp_lock_replay_needed) {
1923                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1924                         spin_lock(&exp->exp_lock);
1925                         exp->exp_lock_replay_needed = 0;
1926                         spin_unlock(&exp->exp_lock);
1927                         atomic_dec(&obd->obd_lock_replay_clients);
1928                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1929                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1930                 }
1931                 spin_unlock_bh(&obd->obd_processing_task_lock);
1932         }
1933
1934         return 0;
1935 }
1936
1937 int target_queue_recovery_request(struct ptlrpc_request *req,
1938                                   struct obd_device *obd)
1939 {
1940         struct list_head *tmp;
1941         int inserted = 0;
1942         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1943         ENTRY;
1944
1945         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1946                 /* Processing the queue right now, don't re-add. */
1947                 RETURN(1);
1948         }
1949
1950         target_process_req_flags(obd, req);
1951
1952         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1953                 /* client declares he's ready to complete recovery
1954                  * so, we put the request on th final queue */
1955                 req = ptlrpc_clone_req(req);
1956                 if (req == NULL)
1957                         RETURN(-ENOMEM);
1958                 DEBUG_REQ(D_HA, req, "queue final req");
1959                 spin_lock_bh(&obd->obd_processing_task_lock);
1960                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1961                 if (obd->obd_recovering)
1962                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1963                 else {
1964                         spin_unlock_bh(&obd->obd_processing_task_lock);
1965                         ptlrpc_free_clone(req);
1966                         if (obd->obd_stopping) {
1967                                 RETURN(-ENOTCONN);
1968                         } else {
1969                                 RETURN(1);
1970                         }
1971                 }
1972                 spin_unlock_bh(&obd->obd_processing_task_lock);
1973                 RETURN(0);
1974         }
1975         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1976                 /* client declares he's ready to replay locks */
1977                 req = ptlrpc_clone_req(req);
1978                 if (req == NULL)
1979                         RETURN(-ENOMEM);
1980                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1981                 spin_lock_bh(&obd->obd_processing_task_lock);
1982                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1983                 LASSERT(obd->obd_recovering);
1984                 /* usually due to recovery abort */
1985                 if (!req->rq_export->exp_in_recovery) {
1986                         spin_unlock_bh(&obd->obd_processing_task_lock);
1987                         ptlrpc_free_clone(req);
1988                         RETURN(-ENOTCONN);
1989                 }
1990                 LASSERT(req->rq_export->exp_lock_replay_needed);
1991                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1992                 spin_unlock_bh(&obd->obd_processing_task_lock);
1993                 RETURN(0);
1994         }
1995
1996         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1997          * (i.e. buflens etc are in my own byte order), but type-dependent
1998          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1999
2000         if (!transno) {
2001                 CFS_INIT_LIST_HEAD(&req->rq_list);
2002                 DEBUG_REQ(D_HA, req, "not queueing");
2003                 RETURN(1);
2004         }
2005
2006         spin_lock_bh(&obd->obd_processing_task_lock);
2007
2008         /* If we're processing the queue, we want don't want to queue this
2009          * message.
2010          *
2011          * Also, if this request has a transno less than the one we're waiting
2012          * for, we should process it now.  It could (and currently always will)
2013          * be an open request for a descriptor that was opened some time ago.
2014          *
2015          * Also, a resent, replayed request that has already been
2016          * handled will pass through here and be processed immediately.
2017          */
2018         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2019               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2020         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2021                 /* Processing the queue right now, don't re-add. */
2022                 LASSERT(list_empty(&req->rq_list));
2023                 spin_unlock_bh(&obd->obd_processing_task_lock);
2024                 RETURN(1);
2025         }
2026         spin_unlock_bh(&obd->obd_processing_task_lock);
2027
2028         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2029                 RETURN(0);
2030
2031         req = ptlrpc_clone_req(req);
2032         if (req == NULL)
2033                 RETURN(-ENOMEM);
2034
2035         spin_lock_bh(&obd->obd_processing_task_lock);
2036         LASSERT(obd->obd_recovering);
2037         if (!req->rq_export->exp_in_recovery) {
2038                 spin_unlock_bh(&obd->obd_processing_task_lock);
2039                 ptlrpc_free_clone(req);
2040                 RETURN(-ENOTCONN);
2041         }
2042         LASSERT(req->rq_export->exp_req_replay_needed);
2043
2044         if (target_exp_enqueue_req_replay(req)) {
2045                 spin_unlock_bh(&obd->obd_processing_task_lock);
2046                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2047                 ptlrpc_free_clone(req);
2048                 RETURN(0);
2049         }
2050
2051         /* XXX O(n^2) */
2052         list_for_each(tmp, &obd->obd_req_replay_queue) {
2053                 struct ptlrpc_request *reqiter =
2054                         list_entry(tmp, struct ptlrpc_request, rq_list);
2055
2056                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2057                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2058                         inserted = 1;
2059                         break;
2060                 }
2061
2062                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2063                              transno)) {
2064                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2065                                   "has been claimed by another client");
2066                         spin_unlock_bh(&obd->obd_processing_task_lock);
2067                         target_exp_dequeue_req_replay(req);
2068                         ptlrpc_free_clone(req);
2069                         RETURN(0);
2070                 }
2071         }
2072
2073         if (!inserted)
2074                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2075
2076         obd->obd_requests_queued_for_recovery++;
2077         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2078         spin_unlock_bh(&obd->obd_processing_task_lock);
2079         RETURN(0);
2080 }
2081
2082 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2083 {
2084         return req->rq_export->exp_obd;
2085 }
2086
2087 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2088 {
2089         LASSERT(exp != NULL);
2090         return &exp->exp_obd->obd_namespace->ns_pool;
2091 }
2092
2093 /**
2094  * Packs current SLV and Limit into \a req.
2095  */
2096 int target_pack_pool_reply(struct ptlrpc_request *req)
2097 {
2098         struct obd_device *obd;
2099         ENTRY;
2100
2101         /*
2102          * Check that we still have all structures alive as this may
2103          * be some late rpc in shutdown time.
2104          */
2105         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2106                      !exp_connect_lru_resize(req->rq_export))) {
2107                 lustre_msg_set_slv(req->rq_repmsg, 0);
2108                 lustre_msg_set_limit(req->rq_repmsg, 0);
2109                 RETURN(0);
2110         }
2111
2112         /*
2113          * OBD is alive here as export is alive, which we checked above.
2114          */
2115         obd = req->rq_export->exp_obd;
2116
2117         read_lock(&obd->obd_pool_lock);
2118         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2119         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2120         read_unlock(&obd->obd_pool_lock);
2121
2122         RETURN(0);
2123 }
2124
2125 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2126 {
2127         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2128                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2129                 return (-ECOMM);
2130         }
2131
2132         if (unlikely(rc)) {
2133                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2134                 req->rq_status = rc;
2135                 return (ptlrpc_send_error(req, 1));
2136         } else {
2137                 DEBUG_REQ(D_NET, req, "sending reply");
2138         }
2139
2140         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2141 }
2142
2143 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2144 {
2145         int                        netrc;
2146         struct ptlrpc_reply_state *rs;
2147         struct obd_device         *obd;
2148         struct obd_export         *exp;
2149         struct ptlrpc_service     *svc;
2150         ENTRY;
2151
2152         if (req->rq_no_reply) {
2153                 EXIT;
2154                 return;
2155         }
2156
2157         svc = req->rq_rqbd->rqbd_service;
2158         rs = req->rq_reply_state;
2159         if (rs == NULL || !rs->rs_difficult) {
2160                 /* no notifiers */
2161                 target_send_reply_msg (req, rc, fail_id);
2162                 EXIT;
2163                 return;
2164         }
2165
2166         /* must be an export if locks saved */
2167         LASSERT (req->rq_export != NULL);
2168         /* req/reply consistent */
2169         LASSERT (rs->rs_service == svc);
2170
2171         /* "fresh" reply */
2172         LASSERT (!rs->rs_scheduled);
2173         LASSERT (!rs->rs_scheduled_ever);
2174         LASSERT (!rs->rs_handled);
2175         LASSERT (!rs->rs_on_net);
2176         LASSERT (rs->rs_export == NULL);
2177         LASSERT (list_empty(&rs->rs_obd_list));
2178         LASSERT (list_empty(&rs->rs_exp_list));
2179
2180         exp = class_export_get (req->rq_export);
2181         obd = exp->exp_obd;
2182
2183         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2184         rs->rs_scheduled = 1;
2185         rs->rs_on_net    = 1;
2186         rs->rs_xid       = req->rq_xid;
2187         rs->rs_transno   = req->rq_transno;
2188         rs->rs_export    = exp;
2189         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2190
2191         spin_lock(&exp->exp_uncommitted_replies_lock);
2192         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2193                rs->rs_transno, exp->exp_last_committed);
2194         if (rs->rs_transno > exp->exp_last_committed) {
2195                 /* not committed already */
2196                 list_add_tail(&rs->rs_obd_list,
2197                               &exp->exp_uncommitted_replies);
2198         }
2199         spin_unlock (&exp->exp_uncommitted_replies_lock);
2200
2201         spin_lock(&exp->exp_lock);
2202         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2203         spin_unlock(&exp->exp_lock);
2204
2205         netrc = target_send_reply_msg (req, rc, fail_id);
2206
2207         spin_lock(&svc->srv_lock);
2208
2209         atomic_inc(&svc->srv_n_difficult_replies);
2210
2211         if (netrc != 0) {
2212                 /* error sending: reply is off the net.  Also we need +1
2213                  * reply ref until ptlrpc_server_handle_reply() is done
2214                  * with the reply state (if the send was successful, there
2215                  * would have been +1 ref for the net, which
2216                  * reply_out_callback leaves alone) */
2217                 rs->rs_on_net = 0;
2218                 ptlrpc_rs_addref(rs);
2219                 atomic_inc (&svc->srv_outstanding_replies);
2220         }
2221
2222         spin_lock(&rs->rs_lock);
2223         if (rs->rs_transno <= exp->exp_last_committed ||
2224             (!rs->rs_on_net && !rs->rs_no_ack) ||
2225              list_empty(&rs->rs_exp_list) ||     /* completed already */
2226              list_empty(&rs->rs_obd_list)) {
2227                 CDEBUG(D_HA, "Schedule reply immediately\n");
2228                 ptlrpc_dispatch_difficult_reply(rs);
2229         } else {
2230                 list_add (&rs->rs_list, &svc->srv_active_replies);
2231                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2232         }
2233         spin_unlock(&rs->rs_lock);
2234         spin_unlock(&svc->srv_lock);
2235         EXIT;
2236 }
2237
2238 int target_handle_ping(struct ptlrpc_request *req)
2239 {
2240         obd_ping(req->rq_export);
2241         return req_capsule_server_pack(&req->rq_pill);
2242 }
2243
2244 void target_committed_to_req(struct ptlrpc_request *req)
2245 {
2246         struct obd_export *exp = req->rq_export;
2247
2248         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2249                 lustre_msg_set_last_committed(req->rq_repmsg,
2250                                               exp->exp_last_committed);
2251         else
2252                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2253                           "%d)", exp->exp_obd->obd_no_transno,
2254                           req->rq_repmsg == NULL);
2255
2256         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2257                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2258 }
2259 EXPORT_SYMBOL(target_committed_to_req);
2260
2261 int target_handle_qc_callback(struct ptlrpc_request *req)
2262 {
2263         struct obd_quotactl *oqctl;
2264         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2265
2266         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2267         if (oqctl == NULL) {
2268                 CERROR("Can't unpack obd_quotactl\n");
2269                 RETURN(-EPROTO);
2270         }
2271
2272         cli->cl_qchk_stat = oqctl->qc_stat;
2273
2274         return 0;
2275 }
2276
2277 #ifdef HAVE_QUOTA_SUPPORT
2278 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2279 {
2280 #ifdef __KERNEL__
2281         struct obd_device *obd = req->rq_export->exp_obd;
2282         struct obd_device *master_obd;
2283         struct obd_device_target *obt;
2284         struct lustre_quota_ctxt *qctxt;
2285         struct qunit_data *qdata = NULL;
2286         int rc = 0;
2287         ENTRY;
2288
2289         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2290                 RETURN(rc);
2291
2292         rc = req_capsule_server_pack(&req->rq_pill);
2293         if (rc) {
2294                 CERROR("packing reply failed!: rc = %d\n", rc);
2295                 RETURN(rc);
2296         }
2297
2298         LASSERT(req->rq_export);
2299
2300         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2301         if (IS_ERR(qdata)) {
2302                 rc = PTR_ERR(qdata);
2303                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2304                 req->rq_status = rc;
2305                 GOTO(out, rc);
2306         }
2307
2308         /* we use the observer */
2309         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2310                 CERROR("Can't find the observer, it is recovering\n");
2311                 req->rq_status = -EAGAIN;
2312                 GOTO(out, rc);
2313         }
2314
2315         master_obd = obd->obd_observer->obd_observer;
2316         obt = &master_obd->u.obt;
2317         qctxt = &obt->obt_qctxt;
2318
2319         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2320                 /* quota_type has not been processed yet, return EAGAIN
2321                  * until we know whether or not quotas are supposed to
2322                  * be enabled */
2323                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2324                        "-EAGAIN\n");
2325                 req->rq_status = -EAGAIN;
2326                 GOTO(out, rc);
2327         }
2328
2329         down_read(&obt->obt_rwsem);
2330         if (qctxt->lqc_lqs_hash == NULL) {
2331                 up_read(&obt->obt_rwsem);
2332                 /* quota_type has not been processed yet, return EAGAIN
2333                  * until we know whether or not quotas are supposed to
2334                  * be enabled */
2335                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2336                        "-EAGAIN\n");
2337                 req->rq_status = -EAGAIN;
2338                 GOTO(out, rc);
2339         }
2340
2341         LASSERT(qctxt->lqc_handler);
2342         rc = qctxt->lqc_handler(master_obd, qdata,
2343                                 lustre_msg_get_opc(req->rq_reqmsg));
2344         up_read(&obt->obt_rwsem);
2345         if (rc && rc != -EDQUOT)
2346                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2347                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2348         req->rq_status = rc;
2349
2350         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2351         if (rc < 0) {
2352                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2353                 GOTO(out, rc);
2354         }
2355
2356         /* Block the quota req. b=14840 */
2357         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2358         EXIT;
2359
2360 out:
2361         rc = ptlrpc_reply(req);
2362         return rc;
2363 #else
2364         return 0;
2365 #endif /* !__KERNEL__ */
2366 }
2367 #endif /* HAVE_QUOTA_SUPPORT */
2368
2369 ldlm_mode_t lck_compat_array[] = {
2370         [LCK_EX] LCK_COMPAT_EX,
2371         [LCK_PW] LCK_COMPAT_PW,
2372         [LCK_PR] LCK_COMPAT_PR,
2373         [LCK_CW] LCK_COMPAT_CW,
2374         [LCK_CR] LCK_COMPAT_CR,
2375         [LCK_NL] LCK_COMPAT_NL,
2376         [LCK_GROUP] LCK_COMPAT_GROUP,
2377         [LCK_COS] LCK_COMPAT_COS,
2378 };
2379
2380 /**
2381  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2382  * not escape to the user level.
2383  */
2384 int ldlm_error2errno(ldlm_error_t error)
2385 {
2386         int result;
2387
2388         switch (error) {
2389         case ELDLM_OK:
2390                 result = 0;
2391                 break;
2392         case ELDLM_LOCK_CHANGED:
2393                 result = -ESTALE;
2394                 break;
2395         case ELDLM_LOCK_ABORTED:
2396                 result = -ENAVAIL;
2397                 break;
2398         case ELDLM_LOCK_REPLACED:
2399                 result = -ESRCH;
2400                 break;
2401         case ELDLM_NO_LOCK_DATA:
2402                 result = -ENOENT;
2403                 break;
2404         case ELDLM_NAMESPACE_EXISTS:
2405                 result = -EEXIST;
2406                 break;
2407         case ELDLM_BAD_NAMESPACE:
2408                 result = -EBADF;
2409                 break;
2410         default:
2411                 if (((int)error) < 0)  /* cast to signed type */
2412                         result = error; /* as ldlm_error_t can be unsigned */
2413                 else {
2414                         CERROR("Invalid DLM result code: %i\n", error);
2415                         result = -EPROTO;
2416                 }
2417         }
2418         return result;
2419 }
2420 EXPORT_SYMBOL(ldlm_error2errno);
2421
2422 /**
2423  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2424  */
2425 ldlm_error_t ldlm_errno2error(int err_no)
2426 {
2427         int error;
2428
2429         switch (err_no) {
2430         case 0:
2431                 error = ELDLM_OK;
2432                 break;
2433         case -ESTALE:
2434                 error = ELDLM_LOCK_CHANGED;
2435                 break;
2436         case -ENAVAIL:
2437                 error = ELDLM_LOCK_ABORTED;
2438                 break;
2439         case -ESRCH:
2440                 error = ELDLM_LOCK_REPLACED;
2441                 break;
2442         case -ENOENT:
2443                 error = ELDLM_NO_LOCK_DATA;
2444                 break;
2445         case -EEXIST:
2446                 error = ELDLM_NAMESPACE_EXISTS;
2447                 break;
2448         case -EBADF:
2449                 error = ELDLM_BAD_NAMESPACE;
2450                 break;
2451         default:
2452                 error = err_no;
2453         }
2454         return error;
2455 }
2456 EXPORT_SYMBOL(ldlm_errno2error);
2457