lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #ifdef __KERNEL__
  43 # include <libcfs/libcfs.h>
  44 #else
  45 # include <liblustre.h>
  46 #endif
  47 #include <obd.h>
  48 #include <lustre_mds.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /* @priority: if non-zero, move the selected to the list head
  55  * @create: if zero, only search in existed connections
  56  */
  57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  58                            int priority, int create)
  59 {
  60         struct ptlrpc_connection *ptlrpc_conn;
  61         struct obd_import_conn *imp_conn = NULL, *item;
  62         int rc = 0;
  63         ENTRY;
  64
  65         if (!create && !priority) {
  66                 CDEBUG(D_HA, "Nothing to do\n");
  67                 RETURN(-EINVAL);
  68         }
  69
  70         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  71         if (!ptlrpc_conn) {
  72                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  73                 RETURN (-ENOENT);
  74         }
  75
  76         if (create) {
  77                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  78                 if (!imp_conn) {
  79                         GOTO(out_put, rc = -ENOMEM);
  80                 }
  81         }
  82
  83         spin_lock(&imp->imp_lock);
  84         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  85                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  86                         if (priority) {
  87                                 list_del(&item->oic_item);
  88                                 list_add(&item->oic_item, &imp->imp_conn_list);
  89                                 item->oic_last_attempt = 0;
  90                         }
  91                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  92                                imp, imp->imp_obd->obd_name, uuid->uuid,
  93                                (priority ? ", moved to head" : ""));
  94                         spin_unlock(&imp->imp_lock);
  95                         GOTO(out_free, rc = 0);
  96                 }
  97         }
  98         /* not found */
  99         if (create) {
 100                 imp_conn->oic_conn = ptlrpc_conn;
 101                 imp_conn->oic_uuid = *uuid;
 102                 imp_conn->oic_last_attempt = 0;
 103                 if (priority)
 104                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 105                 else
 106                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 108                        imp, imp->imp_obd->obd_name, uuid->uuid,
 109                        (priority ? "head" : "tail"));
 110         } else {
 111                 spin_unlock(&imp->imp_lock);
 112                 GOTO(out_free, rc = -ENOENT);
 113         }
 114
 115         spin_unlock(&imp->imp_lock);
 116         RETURN(0);
 117 out_free:
 118         if (imp_conn)
 119                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 120 out_put:
 121         ptlrpc_connection_put(ptlrpc_conn);
 122         RETURN(rc);
 123 }
 124
 125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 126 {
 127         return import_set_conn(imp, uuid, 1, 0);
 128 }
 129
 130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 131                            int priority)
 132 {
 133         return import_set_conn(imp, uuid, priority, 1);
 134 }
 135
 136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 137 {
 138         struct obd_import_conn *imp_conn;
 139         struct obd_export *dlmexp;
 140         int rc = -ENOENT;
 141         ENTRY;
 142
 143         spin_lock(&imp->imp_lock);
 144         if (list_empty(&imp->imp_conn_list)) {
 145                 LASSERT(!imp->imp_connection);
 146                 GOTO(out, rc);
 147         }
 148
 149         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 150                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 151                         continue;
 152                 LASSERT(imp_conn->oic_conn);
 153
 154                 /* is current conn? */
 155                 if (imp_conn == imp->imp_conn_current) {
 156                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 157
 158                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 159                             imp->imp_state != LUSTRE_IMP_DISCON) {
 160                                 CERROR("can't remove current connection\n");
 161                                 GOTO(out, rc = -EBUSY);
 162                         }
 163
 164                         ptlrpc_connection_put(imp->imp_connection);
 165                         imp->imp_connection = NULL;
 166
 167                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 168                         if (dlmexp && dlmexp->exp_connection) {
 169                                 LASSERT(dlmexp->exp_connection ==
 170                                         imp_conn->oic_conn);
 171                                 ptlrpc_connection_put(dlmexp->exp_connection);
 172                                 dlmexp->exp_connection = NULL;
 173                         }
 174                 }
 175
 176                 list_del(&imp_conn->oic_item);
 177                 ptlrpc_connection_put(imp_conn->oic_conn);
 178                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 179                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 180                        imp, imp->imp_obd->obd_name, uuid->uuid);
 181                 rc = 0;
 182                 break;
 183         }
 184 out:
 185         spin_unlock(&imp->imp_lock);
 186         if (rc == -ENOENT)
 187                 CERROR("connection %s not found\n", uuid->uuid);
 188         RETURN(rc);
 189 }
 190
 191 void client_destroy_import(struct obd_import *imp)
 192 {
 193         /* drop security policy instance after all rpc finished/aborted
 194          * to let all busy contexts be released. */
 195         class_import_get(imp);
 196         class_destroy_import(imp);
 197         sptlrpc_import_sec_put(imp);
 198         class_import_put(imp);
 199 }
 200
 201 /* configure an RPC client OBD device
 202  *
 203  * lcfg parameters:
 204  * 1 - client UUID
 205  * 2 - server UUID
 206  * 3 - inactive-on-startup
 207  */
 208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 209 {
 210         struct client_obd *cli = &obddev->u.cli;
 211         struct obd_import *imp;
 212         struct obd_uuid server_uuid;
 213         int rq_portal, rp_portal, connect_op;
 214         char *name = obddev->obd_type->typ_name;
 215         int rc;
 216         ENTRY;
 217
 218         /* In a more perfect world, we would hang a ptlrpc_client off of
 219          * obd_type and just use the values from there. */
 220         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 221                 rq_portal = OST_REQUEST_PORTAL;
 222                 rp_portal = OSC_REPLY_PORTAL;
 223                 connect_op = OST_CONNECT;
 224                 cli->cl_sp_me = LUSTRE_SP_CLI;
 225                 cli->cl_sp_to = LUSTRE_SP_OST;
 226         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 227                 rq_portal = MDS_REQUEST_PORTAL;
 228                 rp_portal = MDC_REPLY_PORTAL;
 229                 connect_op = MDS_CONNECT;
 230                 cli->cl_sp_me = LUSTRE_SP_CLI;
 231                 cli->cl_sp_to = LUSTRE_SP_MDT;
 232         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 233                 rq_portal = MGS_REQUEST_PORTAL;
 234                 rp_portal = MGC_REPLY_PORTAL;
 235                 connect_op = MGS_CONNECT;
 236                 cli->cl_sp_me = LUSTRE_SP_MGC;
 237                 cli->cl_sp_to = LUSTRE_SP_MGS;
 238                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 239         } else {
 240                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 241                        name);
 242                 RETURN(-EINVAL);
 243         }
 244
 245         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 246                 CERROR("requires a TARGET UUID\n");
 247                 RETURN(-EINVAL);
 248         }
 249
 250         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 251                 CERROR("client UUID must be less than 38 characters\n");
 252                 RETURN(-EINVAL);
 253         }
 254
 255         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 256                 CERROR("setup requires a SERVER UUID\n");
 257                 RETURN(-EINVAL);
 258         }
 259
 260         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 261                 CERROR("target UUID must be less than 38 characters\n");
 262                 RETURN(-EINVAL);
 263         }
 264
 265         init_rwsem(&cli->cl_sem);
 266         sema_init(&cli->cl_mgc_sem, 1);
 267         cli->cl_conn_count = 0;
 268         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 269                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 270                      sizeof(server_uuid)));
 271
 272         cli->cl_dirty = 0;
 273         cli->cl_avail_grant = 0;
 274         /* FIXME: should limit this for the sum of all cl_dirty_max */
 275         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 276         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 277                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 278         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 279         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 280         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 281         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 282         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 283         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 284         cli->cl_r_in_flight = 0;
 285         cli->cl_w_in_flight = 0;
 286
 287         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 288         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 289         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 290         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 291         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 292         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 293         cfs_waitq_init(&cli->cl_destroy_waitq);
 294         atomic_set(&cli->cl_destroy_in_flight, 0);
 295 #ifdef ENABLE_CHECKSUM
 296         /* Turn on checksumming by default. */
 297         cli->cl_checksum = 1;
 298         /*
 299          * The supported checksum types will be worked out at connect time
 300          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 301          * through procfs.
 302          */
 303         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 304 #endif
 305         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 306
 307         /* This value may be changed at connect time in
 308            ptlrpc_connect_interpret. */
 309         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 310                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 311
 312         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 313                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 314         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 315                 cli->cl_max_rpcs_in_flight = 2;
 316         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 317                 cli->cl_max_rpcs_in_flight = 3;
 318         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 319                 cli->cl_max_rpcs_in_flight = 4;
 320         } else {
 321                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 322         }
 323
 324         rc = ldlm_get_ref();
 325         if (rc) {
 326                 CERROR("ldlm_get_ref failed: %d\n", rc);
 327                 GOTO(err, rc);
 328         }
 329
 330         ptlrpc_init_client(rq_portal, rp_portal, name,
 331                            &obddev->obd_ldlm_client);
 332
 333         imp = class_new_import(obddev);
 334         if (imp == NULL)
 335                 GOTO(err_ldlm, rc = -ENOENT);
 336         imp->imp_client = &obddev->obd_ldlm_client;
 337         imp->imp_connect_op = connect_op;
 338         imp->imp_initial_recov = 1;
 339         imp->imp_initial_recov_bk = 0;
 340         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 341         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 342                LUSTRE_CFG_BUFLEN(lcfg, 1));
 343         class_import_put(imp);
 344
 345         rc = client_import_add_conn(imp, &server_uuid, 1);
 346         if (rc) {
 347                 CERROR("can't add initial connection\n");
 348                 GOTO(err_import, rc);
 349         }
 350
 351         cli->cl_import = imp;
 352         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 353         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 354         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 355
 356         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 357                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 358                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 359                                name, obddev->obd_name,
 360                                cli->cl_target_uuid.uuid);
 361                         spin_lock(&imp->imp_lock);
 362                         imp->imp_invalid = 1;
 363                         spin_unlock(&imp->imp_lock);
 364                 }
 365         }
 366
 367         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 368
 369         RETURN(rc);
 370
 371 err_import:
 372         class_destroy_import(imp);
 373 err_ldlm:
 374         ldlm_put_ref();
 375 err:
 376         RETURN(rc);
 377
 378 }
 379
 380 int client_obd_cleanup(struct obd_device *obddev)
 381 {
 382         ENTRY;
 383         ldlm_put_ref();
 384         RETURN(0);
 385 }
 386
 387 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 388 int client_connect_import(const struct lu_env *env,
 389                           struct obd_export **exp,
 390                           struct obd_device *obd, struct obd_uuid *cluuid,
 391                           struct obd_connect_data *data, void *localdata)
 392 {
 393         struct client_obd *cli = &obd->u.cli;
 394         struct obd_import *imp = cli->cl_import;
 395         struct obd_connect_data *ocd;
 396         struct ldlm_namespace *to_be_freed = NULL;
 397         struct lustre_handle conn = { 0 };
 398         int rc;
 399         ENTRY;
 400
 401         *exp = NULL;
 402         down_write(&cli->cl_sem);
 403         if (cli->cl_conn_count > 0 )
 404                 GOTO(out_sem, rc = -EALREADY);
 405
 406         rc = class_connect(&conn, obd, cluuid);
 407         if (rc)
 408                 GOTO(out_sem, rc);
 409
 410         cli->cl_conn_count++;
 411         *exp = class_conn2export(&conn);
 412
 413         if (obd->obd_namespace != NULL)
 414                 CERROR("already have namespace!\n");
 415
 416         /*
 417          * Deadlock case - bug 18380
 418          */
 419         up_write(&cli->cl_sem);
 420         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 421                                                 LDLM_NAMESPACE_CLIENT,
 422                                                 LDLM_NAMESPACE_GREEDY);
 423         down_write(&cli->cl_sem);
 424         if (obd->obd_namespace == NULL)
 425                 GOTO(out_disco, rc = -ENOMEM);
 426
 427         imp->imp_dlm_handle = conn;
 428         rc = ptlrpc_init_import(imp);
 429         if (rc != 0)
 430                 GOTO(out_ldlm, rc);
 431
 432         ocd = &imp->imp_connect_data;
 433         if (data) {
 434                 *ocd = *data;
 435                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 436         }
 437
 438         rc = ptlrpc_connect_import(imp, NULL);
 439         if (rc != 0) {
 440                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 441                 GOTO(out_ldlm, rc);
 442         }
 443         LASSERT((*exp)->exp_connection);
 444
 445         if (data) {
 446                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 447                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 448                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 449                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 450         }
 451
 452         ptlrpc_pinger_add_import(imp);
 453
 454         EXIT;
 455
 456         if (rc) {
 457 out_ldlm:
 458                 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
 459                 to_be_freed = obd->obd_namespace;
 460                 obd->obd_namespace = NULL;
 461 out_disco:
 462                 cli->cl_conn_count--;
 463                 class_disconnect(*exp);
 464                 *exp = NULL;
 465         }
 466 out_sem:
 467         up_write(&cli->cl_sem);
 468         if (to_be_freed)
 469                 ldlm_namespace_free_post(to_be_freed);
 470
 471         return rc;
 472 }
 473
 474 int client_disconnect_export(struct obd_export *exp)
 475 {
 476         struct obd_device *obd = class_exp2obd(exp);
 477         struct client_obd *cli;
 478         struct obd_import *imp;
 479         int rc = 0, err;
 480         struct ldlm_namespace *to_be_freed = NULL;
 481         ENTRY;
 482
 483         if (!obd) {
 484                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 485                        exp, exp ? exp->exp_handle.h_cookie : -1);
 486                 RETURN(-EINVAL);
 487         }
 488
 489         cli = &obd->u.cli;
 490         imp = cli->cl_import;
 491
 492         down_write(&cli->cl_sem);
 493         if (!cli->cl_conn_count) {
 494                 CERROR("disconnecting disconnected device (%s)\n",
 495                        obd->obd_name);
 496                 GOTO(out_sem, rc = -EINVAL);
 497         }
 498
 499         cli->cl_conn_count--;
 500         if (cli->cl_conn_count)
 501                 GOTO(out_no_disconnect, rc = 0);
 502
 503         /* Mark import deactivated now, so we don't try to reconnect if any
 504          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 505          * fully deactivate the import, or that would drop all requests. */
 506         spin_lock(&imp->imp_lock);
 507         imp->imp_deactive = 1;
 508         spin_unlock(&imp->imp_lock);
 509
 510         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 511          * delete it regardless.  (It's safe to delete an import that was
 512          * never added.) */
 513         (void)ptlrpc_pinger_del_import(imp);
 514
 515         if (obd->obd_namespace != NULL) {
 516                 /* obd_force == local only */
 517                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 518                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 519                                        NULL);
 520                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 521                 to_be_freed = obd->obd_namespace;
 522         }
 523
 524         /*
 525          * there's no necessary to hold sem during diconnecting an import,
 526          * and actually it may cause deadlock in gss.
 527          */
 528         up_write(&cli->cl_sem);
 529         rc = ptlrpc_disconnect_import(imp, 0);
 530         down_write(&cli->cl_sem);
 531
 532         ptlrpc_invalidate_import(imp);
 533         /* set obd_namespace to NULL only after invalidate, because we can have
 534          * some connect requests in flight, and his need store a connect flags
 535          * in obd_namespace. bug 14260 */
 536         obd->obd_namespace = NULL;
 537
 538         if (imp->imp_rq_pool) {
 539                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 540                 imp->imp_rq_pool = NULL;
 541         }
 542         client_destroy_import(imp);
 543         cli->cl_import = NULL;
 544
 545         EXIT;
 546  out_no_disconnect:
 547         err = class_disconnect(exp);
 548         if (!rc && err)
 549                 rc = err;
 550  out_sem:
 551         up_write(&cli->cl_sem);
 552         if (to_be_freed)
 553                 ldlm_namespace_free_post(to_be_freed);
 554
 555         RETURN(rc);
 556 }
 557
 558 /* --------------------------------------------------------------------------
 559  * from old lib/target.c
 560  * -------------------------------------------------------------------------- */
 561
 562 static int target_handle_reconnect(struct lustre_handle *conn,
 563                                    struct obd_export *exp,
 564                                    struct obd_uuid *cluuid)
 565 {
 566         ENTRY;
 567         if (exp->exp_connection && exp->exp_imp_reverse) {
 568                 struct lustre_handle *hdl;
 569                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 570                 /* Might be a re-connect after a partition. */
 571                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 572                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 573                               cluuid->uuid);
 574                         conn->cookie = exp->exp_handle.h_cookie;
 575                         /* target_handle_connect() treats EALREADY and
 576                          * -EALREADY differently.  EALREADY means we are
 577                          * doing a valid reconnect from the same client. */
 578                         RETURN(EALREADY);
 579                 } else {
 580                         CERROR("%s reconnecting from %s, "
 581                                "handle mismatch (ours "LPX64", theirs "
 582                                LPX64")\n", cluuid->uuid,
 583                                exp->exp_connection->c_remote_uuid.uuid,
 584                                hdl->cookie, conn->cookie);
 585                         memset(conn, 0, sizeof *conn);
 586                         /* target_handle_connect() treats EALREADY and
 587                          * -EALREADY differently.  -EALREADY is an error
 588                          * (same UUID, different handle). */
 589                         RETURN(-EALREADY);
 590                 }
 591         }
 592
 593         conn->cookie = exp->exp_handle.h_cookie;
 594         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 595                cluuid->uuid, exp, conn->cookie);
 596         RETURN(0);
 597 }
 598
 599 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 600                           int error)
 601 {
 602         struct obd_export *exp = cb_data;
 603
 604         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 605                obd->obd_name, exp->exp_client_uuid.uuid);
 606
 607         spin_lock(&exp->exp_lock);
 608         exp->exp_need_sync = 0;
 609         spin_unlock(&exp->exp_lock);
 610 }
 611 EXPORT_SYMBOL(target_client_add_cb);
 612 static void
 613 target_start_and_reset_recovery_timer(struct obd_device *obd,
 614                                       struct ptlrpc_request *req,
 615                                       int new_client);
 616
 617 int target_handle_connect(struct ptlrpc_request *req)
 618 {
 619         struct obd_device *target, *targref = NULL;
 620         struct obd_export *export = NULL;
 621         struct obd_import *revimp;
 622         struct lustre_handle conn;
 623         struct lustre_handle *tmp;
 624         struct obd_uuid tgtuuid;
 625         struct obd_uuid cluuid;
 626         struct obd_uuid remote_uuid;
 627         char *str;
 628         int rc = 0;
 629         int mds_conn = 0;
 630         struct obd_connect_data *data, *tmpdata;
 631         lnet_nid_t *client_nid = NULL;
 632         ENTRY;
 633
 634         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 635
 636         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 637         if (str == NULL) {
 638                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 639                 GOTO(out, rc = -EINVAL);
 640         }
 641
 642         obd_str2uuid(&tgtuuid, str);
 643         target = class_uuid2obd(&tgtuuid);
 644         if (!target)
 645                 target = class_name2obd(str);
 646
 647         if (!target || target->obd_stopping || !target->obd_set_up) {
 648                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 649                                    " for connect (%s)\n", str,
 650                                    !target ? "no target" :
 651                                    (target->obd_stopping ? "stopping" :
 652                                    "not set up"));
 653                 GOTO(out, rc = -ENODEV);
 654         }
 655
 656         if (target->obd_no_conn) {
 657                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 658                               "from %s\n", target->obd_name,
 659                               libcfs_nid2str(req->rq_peer.nid));
 660                 GOTO(out, rc = -EAGAIN);
 661         }
 662
 663         /* Make sure the target isn't cleaned up while we're here. Yes,
 664            there's still a race between the above check and our incref here.
 665            Really, class_uuid2obd should take the ref. */
 666         targref = class_incref(target, __FUNCTION__, cfs_current());
 667
 668
 669         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 670         if (str == NULL) {
 671                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 672                 GOTO(out, rc = -EINVAL);
 673         }
 674
 675         obd_str2uuid(&cluuid, str);
 676
 677         /* XXX extract a nettype and format accordingly */
 678         switch (sizeof(lnet_nid_t)) {
 679                 /* NB the casts only avoid compiler warnings */
 680         case 8:
 681                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 682                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 683                 break;
 684         case 4:
 685                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 686                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 687                 break;
 688         default:
 689                 LBUG();
 690         }
 691
 692         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 693         if (tmp == NULL)
 694                 GOTO(out, rc = -EPROTO);
 695
 696         conn = *tmp;
 697
 698         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 699         if (!data)
 700                 GOTO(out, rc = -EPROTO);
 701
 702         rc = req_capsule_server_pack(&req->rq_pill);
 703         if (rc)
 704                 GOTO(out, rc);
 705
 706         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 707                 if (!data) {
 708                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 709                                   "libclient connection attempt");
 710                         GOTO(out, rc = -EPROTO);
 711                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 712                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 713                            data->ocd_version > LUSTRE_VERSION_CODE +
 714                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 715                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 716                                   "libclient connection attempt",
 717                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 718                                   "old" : "new",
 719                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 720                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 721                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 722                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 723                         data = req_capsule_server_sized_get(&req->rq_pill,
 724                                                         &RMF_CONNECT_DATA,
 725                                     offsetof(typeof(*data), ocd_version) +
 726                                              sizeof(data->ocd_version));
 727                         if (data) {
 728                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 729                                 data->ocd_version = LUSTRE_VERSION_CODE;
 730                         }
 731                         GOTO(out, rc = -EPROTO);
 732                 }
 733         }
 734
 735         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 736             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 737                 mds_conn = 1;
 738
 739         /* lctl gets a backstage, all-access pass. */
 740         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 741                 goto dont_check_exports;
 742
 743         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 744         if (!export)
 745                 goto no_export;
 746
 747         /* we've found an export in the hash */
 748         if (export->exp_connecting) { /* bug 9635, et. al. */
 749                 CWARN("%s: exp %p already connecting\n",
 750                       export->exp_obd->obd_name, export);
 751                 class_export_put(export);
 752                 export = NULL;
 753                 rc = -EALREADY;
 754         } else if (mds_conn && export->exp_connection) {
 755                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 756                         /* mds reconnected after failover */
 757                         CWARN("%s: received MDS connection from NID %s,"
 758                               " removing former export from NID %s\n",
 759                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 760                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 761                 else
 762                         /* new mds connection from the same nid */
 763                         CWARN("%s: received new MDS connection from NID %s,"
 764                               " removing former export from same NID\n",
 765                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 766                 class_fail_export(export);
 767                 class_export_put(export);
 768                 export = NULL;
 769                 rc = 0;
 770         } else if (export->exp_connection != NULL &&
 771                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 772                 /* in mds failover we have static uuid but nid can be
 773                  * changed*/
 774                 CWARN("%s: cookie %s seen on new NID %s when "
 775                       "existing NID %s is already connected\n",
 776                       target->obd_name, cluuid.uuid,
 777                       libcfs_nid2str(req->rq_peer.nid),
 778                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 779                 rc = -EALREADY;
 780                 class_export_put(export);
 781                 export = NULL;
 782         } else {
 783                 spin_lock(&export->exp_lock);
 784                 export->exp_connecting = 1;
 785                 spin_unlock(&export->exp_lock);
 786                 class_export_put(export);
 787                 LASSERT(export->exp_obd == target);
 788
 789                 rc = target_handle_reconnect(&conn, export, &cluuid);
 790         }
 791
 792         /* If we found an export, we already unlocked. */
 793         if (!export) {
 794 no_export:
 795                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 796         } else if (req->rq_export == NULL &&
 797                    atomic_read(&export->exp_rpc_count) > 0) {
 798                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 799                       target->obd_name, cluuid.uuid,
 800                       libcfs_nid2str(req->rq_peer.nid),
 801                       export, atomic_read(&export->exp_refcount));
 802                 GOTO(out, rc = -EBUSY);
 803         } else if (req->rq_export != NULL &&
 804                    (atomic_read(&export->exp_rpc_count) > 1)) {
 805                 /* the current connect rpc has increased exp_rpc_count */
 806                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 807                       target->obd_name, cluuid.uuid,
 808                       libcfs_nid2str(req->rq_peer.nid),
 809                       export, atomic_read(&export->exp_rpc_count) - 1);
 810                 spin_lock(&export->exp_lock);
 811                 if (req->rq_export->exp_conn_cnt <
 812                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 813                         /* try to abort active requests */
 814                         req->rq_export->exp_abort_active_req = 1;
 815                 spin_unlock(&export->exp_lock);
 816                 GOTO(out, rc = -EBUSY);
 817         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 818                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 819                        "cookies not random?\n", target->obd_name,
 820                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 821                 GOTO(out, rc = -EALREADY);
 822         } else {
 823                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 824         }
 825
 826         if (rc < 0) {
 827                 GOTO(out, rc);
 828         }
 829
 830         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 831                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 832               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 833               export, (long)cfs_time_current_sec(),
 834               export ? (long)export->exp_last_request_time : 0);
 835
 836         /* Tell the client if we're in recovery. */
 837         if (target->obd_recovering) {
 838                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 839                 /* If this is the first time a client connects,
 840                    reset the recovery timer */
 841                 if (rc == 0)
 842                         target_start_and_reset_recovery_timer(target, req,
 843                                                               !export);
 844         }
 845
 846         /* We want to handle EALREADY but *not* -EALREADY from
 847          * target_handle_reconnect(), return reconnection state in a flag */
 848         if (rc == EALREADY) {
 849                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 850                 rc = 0;
 851         } else {
 852                 LASSERT(rc == 0);
 853         }
 854
 855         /* Tell the client if we support replayable requests */
 856         if (target->obd_replayable)
 857                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 858         client_nid = &req->rq_peer.nid;
 859
 860         if (export == NULL) {
 861                 if (target->obd_recovering) {
 862                         cfs_time_t t;
 863
 864                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 865                         t = cfs_time_sub(t, cfs_time_current());
 866                         CERROR("%s: denying connection for new client %s (%s): "
 867                                "%d clients in recovery for "CFS_TIME_T"s\n",
 868                                target->obd_name,
 869                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 870                                target->obd_recoverable_clients,
 871                                cfs_duration_sec(t));
 872                         rc = -EBUSY;
 873                 } else {
 874 dont_check_exports:
 875                         rc = obd_connect(req->rq_svc_thread->t_env,
 876                                          &export, target, &cluuid, data,
 877                                          client_nid);
 878                         if (rc == 0)
 879                                 conn.cookie = export->exp_handle.h_cookie;
 880                 }
 881         } else {
 882                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 883                                    export, target, &cluuid, data, client_nid);
 884                 if (rc == 0)
 885                         /* prevous done via class_conn2export */
 886                         class_export_get(export);
 887         }
 888         if (rc)
 889                 GOTO(out, rc);
 890         /* Return only the parts of obd_connect_data that we understand, so the
 891          * client knows that we don't understand the rest. */
 892         if (data) {
 893                  tmpdata = req_capsule_server_get(&req->rq_pill,
 894                                                   &RMF_CONNECT_DATA);
 895                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 896                  *tmpdata = *data;
 897         }
 898
 899         /* If all else goes well, this is our RPC return code. */
 900         req->rq_status = 0;
 901
 902         lustre_msg_set_handle(req->rq_repmsg, &conn);
 903
 904         /* If the client and the server are the same node, we will already
 905          * have an export that really points to the client's DLM export,
 906          * because we have a shared handles table.
 907          *
 908          * XXX this will go away when shaver stops sending the "connect" handle
 909          * in the real "remote handle" field of the request --phik 24 Apr 2003
 910          */
 911         if (req->rq_export != NULL)
 912                 class_export_put(req->rq_export);
 913
 914         req->rq_export = export;
 915
 916         spin_lock(&export->exp_lock);
 917         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 918                 spin_unlock(&export->exp_lock);
 919                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 920                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 921                        export->exp_conn_cnt,
 922                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 923
 924                 GOTO(out, rc = -EALREADY);
 925         }
 926         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 927         export->exp_abort_active_req = 0;
 928
 929         /* request from liblustre?  Don't evict it for not pinging. */
 930         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 931                 export->exp_libclient = 1;
 932                 spin_unlock(&export->exp_lock);
 933
 934                 spin_lock(&target->obd_dev_lock);
 935                 list_del_init(&export->exp_obd_chain_timed);
 936                 spin_unlock(&target->obd_dev_lock);
 937         } else {
 938                 spin_unlock(&export->exp_lock);
 939         }
 940
 941         if (export->exp_connection != NULL)
 942                 ptlrpc_connection_put(export->exp_connection);
 943         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 944                                                        req->rq_self,
 945                                                        &remote_uuid);
 946         if (hlist_unhashed(&export->exp_nid_hash)) {
 947                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 948                                        &export->exp_connection->c_peer.nid,
 949                                        &export->exp_nid_hash);
 950         }
 951
 952         spin_lock_bh(&target->obd_processing_task_lock);
 953         if (target->obd_recovering && !export->exp_in_recovery) {
 954                 spin_lock(&export->exp_lock);
 955                 export->exp_in_recovery = 1;
 956                 export->exp_req_replay_needed = 1;
 957                 export->exp_lock_replay_needed = 1;
 958                 spin_unlock(&export->exp_lock);
 959                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 960                      && (data->ocd_transno == 0))
 961                         CWARN("Connect with zero transno!\n");
 962
 963                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 964                      && data->ocd_transno < target->obd_next_recovery_transno)
 965                         target->obd_next_recovery_transno = data->ocd_transno;
 966                 target->obd_connected_clients++;
 967                 /* each connected client is counted as recoverable */
 968                 target->obd_recoverable_clients++;
 969                 atomic_inc(&target->obd_req_replay_clients);
 970                 atomic_inc(&target->obd_lock_replay_clients);
 971                 if (target->obd_connected_clients ==
 972                     target->obd_max_recoverable_clients)
 973                         cfs_waitq_signal(&target->obd_next_transno_waitq);
 974         }
 975         spin_unlock_bh(&target->obd_processing_task_lock);
 976         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 977         conn = *tmp;
 978
 979         if (export->exp_imp_reverse != NULL) {
 980                 /* destroyed import can be still referenced in ctxt */
 981                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 982                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 983
 984                 /* in some recovery senarios, previous ctx init rpc handled
 985                  * in sptlrpc_target_export_check() might be used to install
 986                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 987                  * using the same gss ctx could reach here and following new
 988                  * reverse import. note all reverse ctx in new/old import are
 989                  * actually based on the same gss ctx. so we invalidate ctx
 990                  * here before destroy import, otherwise flush old import will
 991                  * lead to remote reverse ctx be destroied, thus the reverse
 992                  * ctx of new import will lost its peer.
 993                  * there might be a better way to deal with this???
 994                  */
 995                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 996
 997                 client_destroy_import(export->exp_imp_reverse);
 998         }
 999
1000         /* for the rest part, we return -ENOTCONN in case of errors
1001          * in order to let client initialize connection again.
1002          */
1003         revimp = export->exp_imp_reverse = class_new_import(target);
1004         if (!revimp) {
1005                 CERROR("fail to alloc new reverse import.\n");
1006                 GOTO(out, rc = -ENOTCONN);
1007         }
1008
1009         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1010         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1011         revimp->imp_remote_handle = conn;
1012         revimp->imp_dlm_fake = 1;
1013         revimp->imp_state = LUSTRE_IMP_FULL;
1014
1015         /* unknown versions will be caught in
1016          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1017         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1018
1019         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1020             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1021                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1022         else
1023                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1024
1025         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1026         if (rc) {
1027                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1028                 export->exp_imp_reverse = NULL;
1029                 class_destroy_import(revimp);
1030         }
1031
1032         class_import_put(revimp);
1033 out:
1034         if (export) {
1035                 spin_lock(&export->exp_lock);
1036                 export->exp_connecting = 0;
1037                 spin_unlock(&export->exp_lock);
1038         }
1039         if (targref)
1040                 class_decref(targref, __FUNCTION__, cfs_current());
1041         if (rc)
1042                 req->rq_status = rc;
1043         RETURN(rc);
1044 }
1045
1046 int target_handle_disconnect(struct ptlrpc_request *req)
1047 {
1048         int rc;
1049         ENTRY;
1050
1051         rc = req_capsule_server_pack(&req->rq_pill);
1052         if (rc)
1053                 RETURN(rc);
1054
1055         /* keep the rq_export around so we can send the reply */
1056         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1057
1058         RETURN(0);
1059 }
1060
1061 void target_destroy_export(struct obd_export *exp)
1062 {
1063         /* exports created from last_rcvd data, and "fake"
1064            exports created by lctl don't have an import */
1065         if (exp->exp_imp_reverse != NULL)
1066                 client_destroy_import(exp->exp_imp_reverse);
1067
1068         /* We cancel locks at disconnect time, but this will catch any locks
1069          * granted in a race with recovery-induced disconnect. */
1070         if (exp->exp_obd->obd_namespace != NULL)
1071                 ldlm_cancel_locks_for_export(exp);
1072 }
1073
1074 /*
1075  * Recovery functions
1076  */
1077
1078 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1079 {
1080         struct ptlrpc_request *copy_req;
1081         struct lustre_msg *copy_reqmsg;
1082         struct ptlrpc_user_desc *udesc = NULL;
1083
1084         OBD_ALLOC_PTR(copy_req);
1085         if (!copy_req)
1086                 return NULL;
1087         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1088         if (!copy_reqmsg){
1089                 OBD_FREE_PTR(copy_req);
1090                 return NULL;
1091         }
1092
1093         if (orig_req->rq_user_desc) {
1094                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1095
1096                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1097                 if (!udesc) {
1098                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1099                         OBD_FREE_PTR(copy_req);
1100                         return NULL;
1101                 }
1102                 memcpy(udesc, orig_req->rq_user_desc,
1103                        sptlrpc_user_desc_size(ngroups));
1104         }
1105
1106         *copy_req = *orig_req;
1107         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1108         copy_req->rq_reqmsg = copy_reqmsg;
1109         copy_req->rq_user_desc = udesc;
1110
1111         class_export_rpc_get(copy_req->rq_export);
1112         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1113         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1114         sptlrpc_svc_ctx_addref(copy_req);
1115
1116         if (copy_req->rq_reply_state) {
1117                 /* the copied req takes over the reply state */
1118                 orig_req->rq_reply_state = NULL;
1119                 /* to catch further access */
1120                 orig_req->rq_repmsg = NULL;
1121                 orig_req->rq_replen = 0;
1122         }
1123
1124         return copy_req;
1125 }
1126
1127 void ptlrpc_free_clone(struct ptlrpc_request *req)
1128 {
1129         LASSERT(list_empty(&req->rq_replay_list));
1130
1131         ptlrpc_req_drop_rs(req);
1132         sptlrpc_svc_ctx_decref(req);
1133         class_export_rpc_put(req->rq_export);
1134         list_del_init(&req->rq_list);
1135
1136         if (req->rq_user_desc) {
1137                 int ngroups = req->rq_user_desc->pud_ngroups;
1138                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1139         }
1140         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1141         OBD_FREE_PTR(req);
1142 }
1143
1144 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1145 {
1146         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1147         struct obd_export     *exp = req->rq_export;
1148         struct ptlrpc_request *reqiter;
1149         int                    dup = 0;
1150
1151         LASSERT(exp);
1152
1153         spin_lock(&exp->exp_lock);
1154         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1155                             rq_replay_list) {
1156                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1157                         dup = 1;
1158                         break;
1159                 }
1160         }
1161
1162         if (dup) {
1163                 /* we expect it with RESENT and REPLAY flags */
1164                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1165                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1166                         CERROR("invalid flags %x of resent replay\n",
1167                                lustre_msg_get_flags(req->rq_reqmsg));
1168         } else {
1169                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1170         }
1171
1172         spin_unlock(&exp->exp_lock);
1173         return dup;
1174 }
1175
1176 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1177 {
1178         LASSERT(!list_empty(&req->rq_replay_list));
1179         LASSERT(req->rq_export);
1180
1181         spin_lock(&req->rq_export->exp_lock);
1182         list_del_init(&req->rq_replay_list);
1183         spin_unlock(&req->rq_export->exp_lock);
1184 }
1185
1186 #ifdef __KERNEL__
1187 static void target_finish_recovery(struct obd_device *obd)
1188 {
1189         ENTRY;
1190         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1191                       obd->obd_name);
1192
1193         ldlm_reprocess_all_ns(obd->obd_namespace);
1194         spin_lock_bh(&obd->obd_processing_task_lock);
1195         if (list_empty(&obd->obd_req_replay_queue) &&
1196             list_empty(&obd->obd_lock_replay_queue) &&
1197             list_empty(&obd->obd_final_req_queue)) {
1198                 obd->obd_processing_task = 0;
1199         } else {
1200                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1201                        obd->obd_name,
1202                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1203                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1204                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1205                 spin_unlock_bh(&obd->obd_processing_task_lock);
1206                 LBUG();
1207         }
1208         spin_unlock_bh(&obd->obd_processing_task_lock);
1209
1210         /* when recovery finished, cleanup orphans on mds and ost */
1211         if (OBT(obd) && OBP(obd, postrecov)) {
1212                 int rc = OBP(obd, postrecov)(obd);
1213                 if (rc < 0)
1214                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1215                                       obd->obd_name, rc);
1216         }
1217
1218         obd->obd_recovery_end = cfs_time_current_sec();
1219         EXIT;
1220 }
1221
1222 static void abort_req_replay_queue(struct obd_device *obd)
1223 {
1224         struct ptlrpc_request *req, *n;
1225         struct list_head abort_list;
1226
1227         CFS_INIT_LIST_HEAD(&abort_list);
1228         spin_lock_bh(&obd->obd_processing_task_lock);
1229         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1230         spin_unlock_bh(&obd->obd_processing_task_lock);
1231         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1232                 DEBUG_REQ(D_WARNING, req, "aborted:");
1233                 req->rq_status = -ENOTCONN;
1234                 if (ptlrpc_error(req)) {
1235                         DEBUG_REQ(D_ERROR, req,
1236                                   "failed abort_req_reply; skipping");
1237                 }
1238                 target_exp_dequeue_req_replay(req);
1239                 ptlrpc_free_clone(req);
1240         }
1241 }
1242
1243 static void abort_lock_replay_queue(struct obd_device *obd)
1244 {
1245         struct ptlrpc_request *req, *n;
1246         struct list_head abort_list;
1247
1248         CFS_INIT_LIST_HEAD(&abort_list);
1249         spin_lock_bh(&obd->obd_processing_task_lock);
1250         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1251         spin_unlock_bh(&obd->obd_processing_task_lock);
1252         list_for_each_entry_safe(req, n, &abort_list, rq_list){
1253                 DEBUG_REQ(D_ERROR, req, "aborted:");
1254                 req->rq_status = -ENOTCONN;
1255                 if (ptlrpc_error(req)) {
1256                         DEBUG_REQ(D_ERROR, req,
1257                                   "failed abort_lock_reply; skipping");
1258                 }
1259                 ptlrpc_free_clone(req);
1260         }
1261 }
1262 #endif
1263
1264 /* Called from a cleanup function if the device is being cleaned up
1265    forcefully.  The exports should all have been disconnected already,
1266    the only thing left to do is
1267      - clear the recovery flags
1268      - cancel the timer
1269      - free queued requests and replies, but don't send replies
1270    Because the obd_stopping flag is set, no new requests should be received.
1271
1272 */
1273 void target_cleanup_recovery(struct obd_device *obd)
1274 {
1275         struct ptlrpc_request *req, *n;
1276         struct list_head clean_list;
1277         ENTRY;
1278
1279         LASSERT(obd->obd_stopping);
1280
1281         CFS_INIT_LIST_HEAD(&clean_list);
1282         spin_lock_bh(&obd->obd_processing_task_lock);
1283         if (!obd->obd_recovering) {
1284                 spin_unlock_bh(&obd->obd_processing_task_lock);
1285                 EXIT;
1286                 return;
1287         }
1288         obd->obd_recovering = obd->obd_abort_recovery = 0;
1289         target_cancel_recovery_timer(obd);
1290
1291         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1292         spin_unlock_bh(&obd->obd_processing_task_lock);
1293
1294         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1295                 LASSERT(req->rq_reply_state == 0);
1296                 target_exp_dequeue_req_replay(req);
1297                 ptlrpc_free_clone(req);
1298         }
1299
1300         spin_lock_bh(&obd->obd_processing_task_lock);
1301         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1302         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1303         spin_unlock_bh(&obd->obd_processing_task_lock);
1304
1305         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1306                 LASSERT(req->rq_reply_state == 0);
1307                 ptlrpc_free_clone(req);
1308         }
1309
1310         EXIT;
1311 }
1312
1313 /* obd_processing_task_lock should be held */
1314 void target_cancel_recovery_timer(struct obd_device *obd)
1315 {
1316         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1317         cfs_timer_disarm(&obd->obd_recovery_timer);
1318 }
1319
1320 /* extend = 1 means require at least "duration" seconds left in the timer,
1321    extend = 0 means set the total duration (start_recovery_timer) */
1322 static void reset_recovery_timer(struct obd_device *obd, int duration,
1323                                  int extend)
1324 {
1325         cfs_time_t now = cfs_time_current_sec();
1326         cfs_duration_t left;
1327
1328         spin_lock_bh(&obd->obd_processing_task_lock);
1329         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1330                 spin_unlock_bh(&obd->obd_processing_task_lock);
1331                 return;
1332         }
1333
1334         left = cfs_time_sub(obd->obd_recovery_end, now);
1335
1336         if (extend && (duration > left))
1337                 obd->obd_recovery_timeout += duration - left;
1338         else if (!extend && (duration > obd->obd_recovery_timeout))
1339                 /* Track the client's largest expected replay time */
1340                 obd->obd_recovery_timeout = duration;
1341 #ifdef CRAY_XT3
1342         /*
1343          * If total recovery time already exceed the
1344          * obd_recovery_max_time, then CRAY XT3 will
1345          * abort the recovery
1346          */
1347         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1348                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1349 #endif
1350         obd->obd_recovery_end = obd->obd_recovery_start +
1351                                 obd->obd_recovery_timeout;
1352         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1353             cfs_time_before(now, obd->obd_recovery_end)) {
1354                 left = cfs_time_sub(obd->obd_recovery_end, now);
1355                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1356         }
1357         spin_unlock_bh(&obd->obd_processing_task_lock);
1358         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1359                obd->obd_name, (unsigned)left);
1360 }
1361
1362 static void check_and_start_recovery_timer(struct obd_device *obd)
1363 {
1364         spin_lock_bh(&obd->obd_processing_task_lock);
1365         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1366                 spin_unlock_bh(&obd->obd_processing_task_lock);
1367                 return;
1368         }
1369         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1370         obd->obd_recovery_start = cfs_time_current_sec();
1371         /* minimum */
1372         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1373         spin_unlock_bh(&obd->obd_processing_task_lock);
1374
1375         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1376 }
1377
1378 /* Reset the timer with each new client connection */
1379 /*
1380  * This timer is actually reconnect_timer, which is for making sure
1381  * the total recovery window is at least as big as my reconnect
1382  * attempt timing. So the initial recovery time_out will be set to
1383  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1384  * from client is bigger than this, then the recovery time_out will
1385  * be extend to make sure the client could be reconnected, in the
1386  * process, the timeout from the new client should be ignored.
1387  */
1388
1389 static void
1390 target_start_and_reset_recovery_timer(struct obd_device *obd,
1391                                       struct ptlrpc_request *req,
1392                                       int new_client)
1393 {
1394         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1395
1396         if (!new_client && service_time)
1397                 /* Teach server about old server's estimates, as first guess
1398                  * at how long new requests will take. */
1399                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1400                        service_time);
1401
1402         check_and_start_recovery_timer(obd);
1403
1404         /* convert the service time to rpc timeout,
1405          * reuse service_time to limit stack usage */
1406         service_time = at_est2timeout(service_time);
1407
1408         /* We expect other clients to timeout within service_time, then try
1409          * to reconnect, then try the failover server.  The max delay between
1410          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1411         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1412                              INITIAL_CONNECT_TIMEOUT);
1413         if (service_time > obd->obd_recovery_timeout && !new_client)
1414                 reset_recovery_timer(obd, service_time, 0);
1415 }
1416
1417 #ifdef __KERNEL__
1418 static int check_for_next_transno(struct obd_device *obd)
1419 {
1420         struct ptlrpc_request *req = NULL;
1421         int wake_up = 0, connected, completed, queue_len, max;
1422         __u64 next_transno, req_transno;
1423         ENTRY;
1424         spin_lock_bh(&obd->obd_processing_task_lock);
1425
1426         if (!list_empty(&obd->obd_req_replay_queue)) {
1427                 req = list_entry(obd->obd_req_replay_queue.next,
1428                                  struct ptlrpc_request, rq_list);
1429                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1430         } else {
1431                 req_transno = 0;
1432         }
1433
1434         max = obd->obd_max_recoverable_clients;
1435         connected = obd->obd_connected_clients;
1436         completed = connected - obd->obd_recoverable_clients;
1437         queue_len = obd->obd_requests_queued_for_recovery;
1438         next_transno = obd->obd_next_recovery_transno;
1439
1440         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1441                "req_transno: "LPU64", next_transno: "LPU64"\n",
1442                max, connected, completed, queue_len, req_transno, next_transno);
1443
1444         if (obd->obd_abort_recovery) {
1445                 CDEBUG(D_HA, "waking for aborted recovery\n");
1446                 wake_up = 1;
1447         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1448                 CDEBUG(D_HA, "waking for completed recovery\n");
1449                 wake_up = 1;
1450         } else if (req_transno == next_transno) {
1451                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1452                 wake_up = 1;
1453         } else if (queue_len + completed == max) {
1454                 /* handle gaps occured due to lost reply. It is allowed gaps
1455                  * because all clients are connected and there will be resend
1456                  * for missed transaction */
1457                 LASSERTF(req_transno >= next_transno,
1458                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1459                          req_transno, next_transno);
1460
1461                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1462                        "waking for skipped transno (skip: "LPD64
1463                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1464                        next_transno, queue_len, completed, connected, req_transno);
1465                 obd->obd_next_recovery_transno = req_transno;
1466                 wake_up = 1;
1467         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1468                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1469                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1470                 obd->obd_next_recovery_transno = req_transno;
1471                 wake_up = 1;
1472         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1473                 /* some clients haven't connected in time, but we can try
1474                  * to replay requests that demand on already committed ones
1475                  * also, we can replay first non-committed transation */
1476                 LASSERT(req_transno != 0);
1477                 if (obd->obd_version_recov ||
1478                     req_transno == obd->obd_last_committed + 1) {
1479                         obd->obd_next_recovery_transno = req_transno;
1480                 } else if (req_transno > obd->obd_last_committed) {
1481                         /* can't continue recovery: have no needed transno */
1482                         obd->obd_abort_recovery = 1;
1483                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1484                                "connected: %d, completed: %d, queue_len: %d, "
1485                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1486                                max, connected, completed, queue_len,
1487                                req_transno, next_transno);
1488                 }
1489                 wake_up = 1;
1490         }
1491
1492         spin_unlock_bh(&obd->obd_processing_task_lock);
1493         return wake_up;
1494 }
1495
1496 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1497 {
1498         struct l_wait_info lwi = { 0 };
1499         struct ptlrpc_request *req;
1500
1501         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1502                obd->obd_next_recovery_transno);
1503         l_wait_event(obd->obd_next_transno_waitq,
1504                      check_for_next_transno(obd), &lwi);
1505
1506         spin_lock_bh(&obd->obd_processing_task_lock);
1507         if (obd->obd_abort_recovery) {
1508                 req = NULL;
1509         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1510                 req = list_entry(obd->obd_req_replay_queue.next,
1511                                  struct ptlrpc_request, rq_list);
1512                 target_exp_dequeue_req_replay(req);
1513                 list_del_init(&req->rq_list);
1514                 obd->obd_requests_queued_for_recovery--;
1515         } else {
1516                 req = NULL;
1517         }
1518         spin_unlock_bh(&obd->obd_processing_task_lock);
1519         RETURN(req);
1520 }
1521
1522 static int check_for_next_lock(struct obd_device *obd)
1523 {
1524         struct ptlrpc_request *req = NULL;
1525         int wake_up = 0;
1526
1527         spin_lock_bh(&obd->obd_processing_task_lock);
1528         if (!list_empty(&obd->obd_lock_replay_queue)) {
1529                 req = list_entry(obd->obd_lock_replay_queue.next,
1530                                  struct ptlrpc_request, rq_list);
1531                 CDEBUG(D_HA, "waking for next lock\n");
1532                 wake_up = 1;
1533         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1534                 CDEBUG(D_HA, "waking for completed lock replay\n");
1535                 wake_up = 1;
1536         } else if (obd->obd_abort_recovery) {
1537                 CDEBUG(D_HA, "waking for aborted recovery\n");
1538                 wake_up = 1;
1539         }
1540         spin_unlock_bh(&obd->obd_processing_task_lock);
1541
1542         return wake_up;
1543 }
1544
1545 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1546 {
1547         struct l_wait_info lwi = { 0 };
1548         struct ptlrpc_request *req;
1549
1550         CDEBUG(D_HA, "Waiting for lock\n");
1551         l_wait_event(obd->obd_next_transno_waitq,
1552                      check_for_next_lock(obd), &lwi);
1553
1554         spin_lock_bh(&obd->obd_processing_task_lock);
1555         if (obd->obd_abort_recovery) {
1556                 req = NULL;
1557         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1558                 req = list_entry(obd->obd_lock_replay_queue.next,
1559                                  struct ptlrpc_request, rq_list);
1560                 list_del_init(&req->rq_list);
1561         } else {
1562                 req = NULL;
1563         }
1564         spin_unlock_bh(&obd->obd_processing_task_lock);
1565         return req;
1566 }
1567
1568 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1569 {
1570         struct ptlrpc_request *req;
1571
1572         spin_lock_bh(&obd->obd_processing_task_lock);
1573         if (!list_empty(&obd->obd_final_req_queue)) {
1574                 req = list_entry(obd->obd_final_req_queue.next,
1575                                  struct ptlrpc_request, rq_list);
1576                 list_del_init(&req->rq_list);
1577                 if (req->rq_export->exp_in_recovery) {
1578                         spin_lock(&req->rq_export->exp_lock);
1579                         req->rq_export->exp_in_recovery = 0;
1580                         spin_unlock(&req->rq_export->exp_lock);
1581                 }
1582         } else {
1583                 req = NULL;
1584         }
1585         spin_unlock_bh(&obd->obd_processing_task_lock);
1586         return req;
1587 }
1588
1589 static inline int req_vbr_done(struct obd_export *exp)
1590 {
1591         return (exp->exp_vbr_failed == 0);
1592 }
1593
1594 static inline int req_replay_done(struct obd_export *exp)
1595 {
1596         return (exp->exp_req_replay_needed == 0);
1597 }
1598
1599 static inline int lock_replay_done(struct obd_export *exp)
1600 {
1601         return (exp->exp_lock_replay_needed == 0);
1602 }
1603
1604 static inline int connect_done(struct obd_export *exp)
1605 {
1606         return (exp->exp_in_recovery != 0);
1607 }
1608
1609 static int check_for_clients(struct obd_device *obd)
1610 {
1611         if (obd->obd_abort_recovery || obd->obd_version_recov)
1612                 return 1;
1613         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1614         if (obd->obd_no_conn == 0 &&
1615             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1616                 return 1;
1617         return 0;
1618 }
1619
1620 static int handle_recovery_req(struct ptlrpc_thread *thread,
1621                                struct ptlrpc_request *req,
1622                                svc_handler_t handler)
1623 {
1624         int rc;
1625         ENTRY;
1626
1627         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1628         if (rc) {
1629                 CERROR("Failure to initialize session: %d\n", rc);
1630                 return rc;
1631         }
1632         req->rq_session.lc_thread = thread;
1633         lu_context_enter(&req->rq_session);
1634         req->rq_svc_thread = thread;
1635         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1636
1637         /* thread context */
1638         lu_context_enter(&thread->t_env->le_ctx);
1639         (void)handler(req);
1640         lu_context_exit(&thread->t_env->le_ctx);
1641
1642         lu_context_exit(&req->rq_session);
1643         lu_context_fini(&req->rq_session);
1644         /* don't reset timer for final stage */
1645         if (!req_replay_done(req->rq_export) ||
1646             !lock_replay_done(req->rq_export))
1647                 reset_recovery_timer(class_exp2obd(req->rq_export),
1648                        AT_OFF ? obd_timeout :
1649                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1650         ptlrpc_free_clone(req);
1651         RETURN(0);
1652 }
1653
1654 static int target_recovery_thread(void *arg)
1655 {
1656         struct lu_target *lut = arg;
1657         struct obd_device *obd = lut->lut_obd;
1658         struct ptlrpc_request *req;
1659         struct target_recovery_data *trd = &obd->obd_recovery_data;
1660         struct l_wait_info lwi = { 0 };
1661         unsigned long delta;
1662         unsigned long flags;
1663         struct lu_env env;
1664         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1665         int rc = 0;
1666         ENTRY;
1667
1668         cfs_daemonize("tgt_recov");
1669
1670         SIGNAL_MASK_LOCK(current, flags);
1671         sigfillset(&current->blocked);
1672         RECALC_SIGPENDING;
1673         SIGNAL_MASK_UNLOCK(current, flags);
1674
1675         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1676         if (rc)
1677                 RETURN(rc);
1678
1679         thread->t_env = &env;
1680         env.le_ctx.lc_thread = thread;
1681
1682         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1683                cfs_curproc_pid());
1684         trd->trd_processing_task = cfs_curproc_pid();
1685
1686         obd->obd_recovering = 1;
1687         complete(&trd->trd_starting);
1688
1689         /* first of all, we have to know the first transno to replay */
1690         obd->obd_abort_recovery = 0;
1691         l_wait_event(obd->obd_next_transno_waitq,
1692                      check_for_clients(obd), &lwi);
1693
1694         /* If some clients haven't connected in time, evict them */
1695         if (obd->obd_connected_clients < obd->obd_max_recoverable_clients) {
1696                 CWARN("Some clients haven't connect in time (%d/%d),"
1697                        "evict them\n", obd->obd_connected_clients,
1698                        obd->obd_max_recoverable_clients);
1699                 class_disconnect_stale_exports(obd, connect_done,
1700                                                exp_flags_from_obd(obd) |
1701                                                OBD_OPT_ABORT_RECOV);
1702                 /**
1703                  * if recovery proceeds with versions then some clients may be
1704                  * timed out waiting for others and trying to reconnect.
1705                  * Extend timer for such reconnect cases.
1706                  */
1707                 if (obd->obd_version_recov)
1708                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX * 2, 1);
1709         }
1710
1711         /* next stage: replay requests */
1712         delta = jiffies;
1713         obd->obd_req_replaying = 1;
1714         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1715                atomic_read(&obd->obd_req_replay_clients),
1716                obd->obd_next_recovery_transno);
1717         while ((req = target_next_replay_req(obd))) {
1718                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1719                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1720                           lustre_msg_get_transno(req->rq_reqmsg),
1721                           libcfs_nid2str(req->rq_peer.nid));
1722                 handle_recovery_req(thread, req,
1723                                     trd->trd_recovery_handler);
1724                 obd->obd_replayed_requests++;
1725                 spin_lock_bh(&obd->obd_processing_task_lock);
1726                 obd->obd_next_recovery_transno++;
1727                 spin_unlock_bh(&obd->obd_processing_task_lock);
1728         }
1729
1730         /* If some clients haven't replayed requests in time, evict them */
1731         if (obd->obd_abort_recovery) {
1732                 CDEBUG(D_WARNING, "req replay is aborted\n");
1733                 class_disconnect_stale_exports(obd, req_replay_done,
1734                                                exp_flags_from_obd(obd) |
1735                                                OBD_OPT_ABORT_RECOV);
1736                 abort_req_replay_queue(obd);
1737         }
1738         LASSERT(list_empty(&obd->obd_req_replay_queue));
1739
1740         /* The second stage: replay locks */
1741         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1742                atomic_read(&obd->obd_lock_replay_clients));
1743         while ((req = target_next_replay_lock(obd))) {
1744                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1745                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1746                           libcfs_nid2str(req->rq_peer.nid));
1747                 handle_recovery_req(thread, req,
1748                                     trd->trd_recovery_handler);
1749                 obd->obd_replayed_locks++;
1750         }
1751
1752         /* If some clients haven't replayed requests in time, evict them */
1753         if (obd->obd_abort_recovery) {
1754                 CERROR("lock replay is aborted\n");
1755                 class_disconnect_stale_exports(obd, lock_replay_done,
1756                                                exp_flags_from_obd(obd) |
1757                                                OBD_OPT_ABORT_RECOV);
1758                 abort_lock_replay_queue(obd);
1759         }
1760         LASSERT(list_empty(&obd->obd_lock_replay_queue));
1761
1762         /* The third stage: reply on final pings */
1763         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1764         /** evict exports failed VBR */
1765         class_disconnect_stale_exports(obd, req_vbr_done,
1766                                        exp_flags_from_obd(obd) |
1767                                        OBD_OPT_ABORT_RECOV);
1768         /** Update server last boot epoch */
1769         lut_boot_epoch_update(lut);
1770         /* We drop recoverying flag to forward all new requests
1771          * to regular mds_handle() since now */
1772         spin_lock_bh(&obd->obd_processing_task_lock);
1773         obd->obd_recovering = obd->obd_abort_recovery = 0;
1774         target_cancel_recovery_timer(obd);
1775         spin_unlock_bh(&obd->obd_processing_task_lock);
1776         while ((req = target_next_final_ping(obd))) {
1777                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1778                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1779                           libcfs_nid2str(req->rq_peer.nid));
1780                 handle_recovery_req(thread, req,
1781                                     trd->trd_recovery_handler);
1782         }
1783
1784         delta = (jiffies - delta) / HZ;
1785         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1786               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1787         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1788         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1789         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1790                 CWARN("too long recovery - read logs\n");
1791                 libcfs_debug_dumplog();
1792         }
1793
1794         target_finish_recovery(obd);
1795
1796         lu_context_fini(&env.le_ctx);
1797         trd->trd_processing_task = 0;
1798         complete(&trd->trd_finishing);
1799         RETURN(rc);
1800 }
1801
1802 static int target_start_recovery_thread(struct lu_target *lut,
1803                                         svc_handler_t handler)
1804 {
1805         struct obd_device *obd = lut->lut_obd;
1806         int rc = 0;
1807         struct target_recovery_data *trd = &obd->obd_recovery_data;
1808
1809         memset(trd, 0, sizeof(*trd));
1810         init_completion(&trd->trd_starting);
1811         init_completion(&trd->trd_finishing);
1812         trd->trd_recovery_handler = handler;
1813
1814         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1815                 wait_for_completion(&trd->trd_starting);
1816                 LASSERT(obd->obd_recovering != 0);
1817         } else
1818                 rc = -ECHILD;
1819
1820         return rc;
1821 }
1822
1823 void target_stop_recovery_thread(struct obd_device *obd)
1824 {
1825         spin_lock_bh(&obd->obd_processing_task_lock);
1826         if (obd->obd_recovery_data.trd_processing_task > 0) {
1827                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1828                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1829                 obd->obd_abort_recovery = 1;
1830                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1831                 spin_unlock_bh(&obd->obd_processing_task_lock);
1832                 wait_for_completion(&trd->trd_finishing);
1833         } else {
1834                 spin_unlock_bh(&obd->obd_processing_task_lock);
1835         }
1836 }
1837
1838 void target_recovery_fini(struct obd_device *obd)
1839 {
1840         class_disconnect_exports(obd);
1841         target_stop_recovery_thread(obd);
1842         target_cleanup_recovery(obd);
1843 }
1844 EXPORT_SYMBOL(target_recovery_fini);
1845
1846 static void target_recovery_expired(unsigned long castmeharder)
1847 {
1848         struct obd_device *obd = (struct obd_device *)castmeharder;
1849         CDEBUG(D_HA, "%s: recovery timed out; %d clients never reconnected "
1850                "after %lds (%d clients did)\n",
1851                obd->obd_name, obd->obd_recoverable_clients,
1852                cfs_time_current_sec()- obd->obd_recovery_start,
1853                obd->obd_connected_clients);
1854
1855         spin_lock_bh(&obd->obd_processing_task_lock);
1856         obd->obd_version_recov = 1;
1857         CDEBUG(D_INFO, "VBR is used for %d clients from t"LPU64"\n",
1858                atomic_read(&obd->obd_req_replay_clients),
1859                obd->obd_next_recovery_transno);
1860         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1861         spin_unlock_bh(&obd->obd_processing_task_lock);
1862 }
1863
1864 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1865 {
1866         struct obd_device *obd = lut->lut_obd;
1867         if (obd->obd_max_recoverable_clients == 0)
1868                 return;
1869
1870         CWARN("RECOVERY: service %s, %d recoverable clients, "
1871               "last_transno "LPU64"\n", obd->obd_name,
1872               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1873         LASSERT(obd->obd_stopping == 0);
1874         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1875         obd->obd_recovery_start = 0;
1876         obd->obd_recovery_end = 0;
1877         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1878         /* bz13079: this should be set to desired value for ost but not for mds */
1879         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1880         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1881         target_start_recovery_thread(lut, handler);
1882 }
1883 EXPORT_SYMBOL(target_recovery_init);
1884
1885 #endif
1886
1887 static int target_process_req_flags(struct obd_device *obd,
1888                                     struct ptlrpc_request *req)
1889 {
1890         struct obd_export *exp = req->rq_export;
1891         LASSERT(exp != NULL);
1892         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1893                 /* client declares he's ready to replay locks */
1894                 spin_lock_bh(&obd->obd_processing_task_lock);
1895                 if (exp->exp_req_replay_needed) {
1896                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1897                         spin_lock(&exp->exp_lock);
1898                         exp->exp_req_replay_needed = 0;
1899                         spin_unlock(&exp->exp_lock);
1900                         atomic_dec(&obd->obd_req_replay_clients);
1901                         LASSERT(obd->obd_recoverable_clients > 0);
1902                         obd->obd_recoverable_clients--;
1903                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1904                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1905                 }
1906                 spin_unlock_bh(&obd->obd_processing_task_lock);
1907         }
1908         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1909                 /* client declares he's ready to complete recovery
1910                  * so, we put the request on th final queue */
1911                 spin_lock_bh(&obd->obd_processing_task_lock);
1912                 if (exp->exp_lock_replay_needed) {
1913                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1914                         spin_lock(&exp->exp_lock);
1915                         exp->exp_lock_replay_needed = 0;
1916                         spin_unlock(&exp->exp_lock);
1917                         atomic_dec(&obd->obd_lock_replay_clients);
1918                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1919                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1920                 }
1921                 spin_unlock_bh(&obd->obd_processing_task_lock);
1922         }
1923
1924         return 0;
1925 }
1926
1927 int target_queue_recovery_request(struct ptlrpc_request *req,
1928                                   struct obd_device *obd)
1929 {
1930         struct list_head *tmp;
1931         int inserted = 0;
1932         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1933         ENTRY;
1934
1935         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1936                 /* Processing the queue right now, don't re-add. */
1937                 RETURN(1);
1938         }
1939
1940         target_process_req_flags(obd, req);
1941
1942         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1943                 /* client declares he's ready to complete recovery
1944                  * so, we put the request on th final queue */
1945                 req = ptlrpc_clone_req(req);
1946                 if (req == NULL)
1947                         RETURN(-ENOMEM);
1948                 DEBUG_REQ(D_HA, req, "queue final req");
1949                 spin_lock_bh(&obd->obd_processing_task_lock);
1950                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1951                 if (obd->obd_recovering)
1952                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1953                 else {
1954                         spin_unlock_bh(&obd->obd_processing_task_lock);
1955                         ptlrpc_free_clone(req);
1956                         if (obd->obd_stopping) {
1957                                 RETURN(-ENOTCONN);
1958                         } else {
1959                                 RETURN(1);
1960                         }
1961                 }
1962                 spin_unlock_bh(&obd->obd_processing_task_lock);
1963                 RETURN(0);
1964         }
1965         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1966                 /* client declares he's ready to replay locks */
1967                 req = ptlrpc_clone_req(req);
1968                 if (req == NULL)
1969                         RETURN(-ENOMEM);
1970                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1971                 spin_lock_bh(&obd->obd_processing_task_lock);
1972                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1973                 LASSERT(obd->obd_recovering);
1974                 /* usually due to recovery abort */
1975                 if (!req->rq_export->exp_in_recovery) {
1976                         spin_unlock_bh(&obd->obd_processing_task_lock);
1977                         ptlrpc_free_clone(req);
1978                         RETURN(-ENOTCONN);
1979                 }
1980                 LASSERT(req->rq_export->exp_lock_replay_needed);
1981                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1982                 spin_unlock_bh(&obd->obd_processing_task_lock);
1983                 RETURN(0);
1984         }
1985
1986         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1987          * (i.e. buflens etc are in my own byte order), but type-dependent
1988          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1989
1990         if (!transno) {
1991                 CFS_INIT_LIST_HEAD(&req->rq_list);
1992                 DEBUG_REQ(D_HA, req, "not queueing");
1993                 RETURN(1);
1994         }
1995
1996         spin_lock_bh(&obd->obd_processing_task_lock);
1997
1998         /* If we're processing the queue, we want don't want to queue this
1999          * message.
2000          *
2001          * Also, if this request has a transno less than the one we're waiting
2002          * for, we should process it now.  It could (and currently always will)
2003          * be an open request for a descriptor that was opened some time ago.
2004          *
2005          * Also, a resent, replayed request that has already been
2006          * handled will pass through here and be processed immediately.
2007          */
2008         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2009               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2010         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2011                 /* Processing the queue right now, don't re-add. */
2012                 LASSERT(list_empty(&req->rq_list));
2013                 spin_unlock_bh(&obd->obd_processing_task_lock);
2014                 RETURN(1);
2015         }
2016         spin_unlock_bh(&obd->obd_processing_task_lock);
2017
2018         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2019                 RETURN(0);
2020
2021         req = ptlrpc_clone_req(req);
2022         if (req == NULL)
2023                 RETURN(-ENOMEM);
2024
2025         spin_lock_bh(&obd->obd_processing_task_lock);
2026         LASSERT(obd->obd_recovering);
2027         if (!req->rq_export->exp_in_recovery) {
2028                 spin_unlock_bh(&obd->obd_processing_task_lock);
2029                 ptlrpc_free_clone(req);
2030                 RETURN(-ENOTCONN);
2031         }
2032         LASSERT(req->rq_export->exp_req_replay_needed);
2033
2034         if (target_exp_enqueue_req_replay(req)) {
2035                 spin_unlock_bh(&obd->obd_processing_task_lock);
2036                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2037                 ptlrpc_free_clone(req);
2038                 RETURN(0);
2039         }
2040
2041         /* XXX O(n^2) */
2042         list_for_each(tmp, &obd->obd_req_replay_queue) {
2043                 struct ptlrpc_request *reqiter =
2044                         list_entry(tmp, struct ptlrpc_request, rq_list);
2045
2046                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2047                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2048                         inserted = 1;
2049                         break;
2050                 }
2051
2052                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2053                              transno)) {
2054                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2055                                   "has been claimed by another client");
2056                         spin_unlock_bh(&obd->obd_processing_task_lock);
2057                         target_exp_dequeue_req_replay(req);
2058                         ptlrpc_free_clone(req);
2059                         RETURN(0);
2060                 }
2061         }
2062
2063         if (!inserted)
2064                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2065
2066         obd->obd_requests_queued_for_recovery++;
2067         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2068         spin_unlock_bh(&obd->obd_processing_task_lock);
2069         RETURN(0);
2070 }
2071
2072 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2073 {
2074         return req->rq_export->exp_obd;
2075 }
2076
2077 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2078 {
2079         LASSERT(exp != NULL);
2080         return &exp->exp_obd->obd_namespace->ns_pool;
2081 }
2082
2083 /**
2084  * Packs current SLV and Limit into \a req.
2085  */
2086 int target_pack_pool_reply(struct ptlrpc_request *req)
2087 {
2088         struct obd_device *obd;
2089         ENTRY;
2090
2091         /*
2092          * Check that we still have all structures alive as this may
2093          * be some late rpc in shutdown time.
2094          */
2095         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2096                      !exp_connect_lru_resize(req->rq_export))) {
2097                 lustre_msg_set_slv(req->rq_repmsg, 0);
2098                 lustre_msg_set_limit(req->rq_repmsg, 0);
2099                 RETURN(0);
2100         }
2101
2102         /*
2103          * OBD is alive here as export is alive, which we checked above.
2104          */
2105         obd = req->rq_export->exp_obd;
2106
2107         read_lock(&obd->obd_pool_lock);
2108         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2109         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2110         read_unlock(&obd->obd_pool_lock);
2111
2112         RETURN(0);
2113 }
2114
2115 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2116 {
2117         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2118                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2119                 return (-ECOMM);
2120         }
2121
2122         if (unlikely(rc)) {
2123                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2124                 req->rq_status = rc;
2125                 return (ptlrpc_send_error(req, 1));
2126         } else {
2127                 DEBUG_REQ(D_NET, req, "sending reply");
2128         }
2129
2130         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2131 }
2132
2133 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2134 {
2135         int                        netrc;
2136         struct ptlrpc_reply_state *rs;
2137         struct obd_device         *obd;
2138         struct obd_export         *exp;
2139         struct ptlrpc_service     *svc;
2140         ENTRY;
2141
2142         if (req->rq_no_reply) {
2143                 EXIT;
2144                 return;
2145         }
2146
2147         svc = req->rq_rqbd->rqbd_service;
2148         rs = req->rq_reply_state;
2149         if (rs == NULL || !rs->rs_difficult) {
2150                 /* no notifiers */
2151                 target_send_reply_msg (req, rc, fail_id);
2152                 EXIT;
2153                 return;
2154         }
2155
2156         /* must be an export if locks saved */
2157         LASSERT (req->rq_export != NULL);
2158         /* req/reply consistent */
2159         LASSERT (rs->rs_service == svc);
2160
2161         /* "fresh" reply */
2162         LASSERT (!rs->rs_scheduled);
2163         LASSERT (!rs->rs_scheduled_ever);
2164         LASSERT (!rs->rs_handled);
2165         LASSERT (!rs->rs_on_net);
2166         LASSERT (rs->rs_export == NULL);
2167         LASSERT (list_empty(&rs->rs_obd_list));
2168         LASSERT (list_empty(&rs->rs_exp_list));
2169
2170         exp = class_export_get (req->rq_export);
2171         obd = exp->exp_obd;
2172
2173         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2174         rs->rs_scheduled = 1;
2175         rs->rs_on_net    = 1;
2176         rs->rs_xid       = req->rq_xid;
2177         rs->rs_transno   = req->rq_transno;
2178         rs->rs_export    = exp;
2179
2180         spin_lock(&exp->exp_uncommitted_replies_lock);
2181         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2182                rs->rs_transno, exp->exp_last_committed);
2183         if (rs->rs_transno > exp->exp_last_committed) {
2184                 /* not committed already */
2185                 list_add_tail(&rs->rs_obd_list,
2186                               &exp->exp_uncommitted_replies);
2187         }
2188         spin_unlock (&exp->exp_uncommitted_replies_lock);
2189
2190         spin_lock(&exp->exp_lock);
2191         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2192         spin_unlock(&exp->exp_lock);
2193
2194         netrc = target_send_reply_msg (req, rc, fail_id);
2195
2196         spin_lock(&svc->srv_lock);
2197
2198         atomic_inc(&svc->srv_n_difficult_replies);
2199
2200         if (netrc != 0) {
2201                 /* error sending: reply is off the net.  Also we need +1
2202                  * reply ref until ptlrpc_server_handle_reply() is done
2203                  * with the reply state (if the send was successful, there
2204                  * would have been +1 ref for the net, which
2205                  * reply_out_callback leaves alone) */
2206                 rs->rs_on_net = 0;
2207                 ptlrpc_rs_addref(rs);
2208                 atomic_inc (&svc->srv_outstanding_replies);
2209         }
2210
2211         spin_lock(&rs->rs_lock);
2212         if (rs->rs_transno <= exp->exp_last_committed ||
2213             (!rs->rs_on_net && !rs->rs_no_ack) ||
2214              list_empty(&rs->rs_exp_list) ||     /* completed already */
2215              list_empty(&rs->rs_obd_list)) {
2216                 CDEBUG(D_HA, "Schedule reply immediately\n");
2217                 ptlrpc_dispatch_difficult_reply(rs);
2218         } else {
2219                 list_add (&rs->rs_list, &svc->srv_active_replies);
2220                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2221         }
2222         spin_unlock(&rs->rs_lock);
2223         spin_unlock(&svc->srv_lock);
2224         EXIT;
2225 }
2226
2227 int target_handle_ping(struct ptlrpc_request *req)
2228 {
2229         obd_ping(req->rq_export);
2230         return req_capsule_server_pack(&req->rq_pill);
2231 }
2232
2233 void target_committed_to_req(struct ptlrpc_request *req)
2234 {
2235         struct obd_export *exp = req->rq_export;
2236
2237         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2238                 lustre_msg_set_last_committed(req->rq_repmsg,
2239                                               exp->exp_last_committed);
2240         else
2241                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2242                           "%d)", exp->exp_obd->obd_no_transno,
2243                           req->rq_repmsg == NULL);
2244
2245         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2246                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2247 }
2248 EXPORT_SYMBOL(target_committed_to_req);
2249
2250 int target_handle_qc_callback(struct ptlrpc_request *req)
2251 {
2252         struct obd_quotactl *oqctl;
2253         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2254
2255         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2256         if (oqctl == NULL) {
2257                 CERROR("Can't unpack obd_quotactl\n");
2258                 RETURN(-EPROTO);
2259         }
2260
2261         cli->cl_qchk_stat = oqctl->qc_stat;
2262
2263         return 0;
2264 }
2265
2266 #ifdef HAVE_QUOTA_SUPPORT
2267 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2268 {
2269 #ifdef __KERNEL__
2270         struct obd_device *obd = req->rq_export->exp_obd;
2271         struct obd_device *master_obd;
2272         struct obd_device_target *obt;
2273         struct lustre_quota_ctxt *qctxt;
2274         struct qunit_data *qdata = NULL;
2275         int rc = 0;
2276         ENTRY;
2277
2278         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2279                 RETURN(rc);
2280
2281         rc = req_capsule_server_pack(&req->rq_pill);
2282         if (rc) {
2283                 CERROR("packing reply failed!: rc = %d\n", rc);
2284                 RETURN(rc);
2285         }
2286
2287         LASSERT(req->rq_export);
2288
2289         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2290         if (IS_ERR(qdata)) {
2291                 rc = PTR_ERR(qdata);
2292                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2293                 req->rq_status = rc;
2294                 GOTO(out, rc);
2295         }
2296
2297         /* we use the observer */
2298         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2299                 CERROR("Can't find the observer, it is recovering\n");
2300                 req->rq_status = -EAGAIN;
2301                 GOTO(out, rc);
2302         }
2303
2304         master_obd = obd->obd_observer->obd_observer;
2305         obt = &master_obd->u.obt;
2306         qctxt = &obt->obt_qctxt;
2307
2308         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2309                 /* quota_type has not been processed yet, return EAGAIN
2310                  * until we know whether or not quotas are supposed to
2311                  * be enabled */
2312                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2313                        "-EAGAIN\n");
2314                 req->rq_status = -EAGAIN;
2315                 GOTO(out, rc);
2316         }
2317
2318         down_read(&obt->obt_rwsem);
2319         if (qctxt->lqc_lqs_hash == NULL) {
2320                 up_read(&obt->obt_rwsem);
2321                 /* quota_type has not been processed yet, return EAGAIN
2322                  * until we know whether or not quotas are supposed to
2323                  * be enabled */
2324                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2325                        "-EAGAIN\n");
2326                 req->rq_status = -EAGAIN;
2327                 GOTO(out, rc);
2328         }
2329
2330         LASSERT(qctxt->lqc_handler);
2331         rc = qctxt->lqc_handler(master_obd, qdata,
2332                                 lustre_msg_get_opc(req->rq_reqmsg));
2333         up_read(&obt->obt_rwsem);
2334         if (rc && rc != -EDQUOT)
2335                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2336                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2337         req->rq_status = rc;
2338
2339         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2340         if (rc < 0) {
2341                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2342                 GOTO(out, rc);
2343         }
2344
2345         /* Block the quota req. b=14840 */
2346         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2347         EXIT;
2348
2349 out:
2350         rc = ptlrpc_reply(req);
2351         return rc;
2352 #else
2353         return 0;
2354 #endif /* !__KERNEL__ */
2355 }
2356 #endif /* HAVE_QUOTA_SUPPORT */
2357
2358 ldlm_mode_t lck_compat_array[] = {
2359         [LCK_EX] LCK_COMPAT_EX,
2360         [LCK_PW] LCK_COMPAT_PW,
2361         [LCK_PR] LCK_COMPAT_PR,
2362         [LCK_CW] LCK_COMPAT_CW,
2363         [LCK_CR] LCK_COMPAT_CR,
2364         [LCK_NL] LCK_COMPAT_NL,
2365         [LCK_GROUP] LCK_COMPAT_GROUP,
2366         [LCK_COS] LCK_COMPAT_COS,
2367 };
2368
2369 /**
2370  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2371  * not escape to the user level.
2372  */
2373 int ldlm_error2errno(ldlm_error_t error)
2374 {
2375         int result;
2376
2377         switch (error) {
2378         case ELDLM_OK:
2379                 result = 0;
2380                 break;
2381         case ELDLM_LOCK_CHANGED:
2382                 result = -ESTALE;
2383                 break;
2384         case ELDLM_LOCK_ABORTED:
2385                 result = -ENAVAIL;
2386                 break;
2387         case ELDLM_LOCK_REPLACED:
2388                 result = -ESRCH;
2389                 break;
2390         case ELDLM_NO_LOCK_DATA:
2391                 result = -ENOENT;
2392                 break;
2393         case ELDLM_NAMESPACE_EXISTS:
2394                 result = -EEXIST;
2395                 break;
2396         case ELDLM_BAD_NAMESPACE:
2397                 result = -EBADF;
2398                 break;
2399         default:
2400                 if (((int)error) < 0)  /* cast to signed type */
2401                         result = error; /* as ldlm_error_t can be unsigned */
2402                 else {
2403                         CERROR("Invalid DLM result code: %i\n", error);
2404                         result = -EPROTO;
2405                 }
2406         }
2407         return result;
2408 }
2409 EXPORT_SYMBOL(ldlm_error2errno);
2410
2411 /**
2412  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2413  */
2414 ldlm_error_t ldlm_errno2error(int err_no)
2415 {
2416         int error;
2417
2418         switch (err_no) {
2419         case 0:
2420                 error = ELDLM_OK;
2421                 break;
2422         case -ESTALE:
2423                 error = ELDLM_LOCK_CHANGED;
2424                 break;
2425         case -ENAVAIL:
2426                 error = ELDLM_LOCK_ABORTED;
2427                 break;
2428         case -ESRCH:
2429                 error = ELDLM_LOCK_REPLACED;
2430                 break;
2431         case -ENOENT:
2432                 error = ELDLM_NO_LOCK_DATA;
2433                 break;
2434         case -EEXIST:
2435                 error = ELDLM_NAMESPACE_EXISTS;
2436                 break;
2437         case -EBADF:
2438                 error = ELDLM_BAD_NAMESPACE;
2439                 break;
2440         default:
2441                 error = err_no;
2442         }
2443         return error;
2444 }
2445 EXPORT_SYMBOL(ldlm_errno2error);
2446