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                 item->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 static void 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_write_list);
 281         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 282         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 283         cli->cl_r_in_flight = 0;
 284         cli->cl_w_in_flight = 0;
 285
 286         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 287         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 288         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 289         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 290         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 291         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 292         cfs_waitq_init(&cli->cl_destroy_waitq);
 293         atomic_set(&cli->cl_destroy_in_flight, 0);
 294 #ifdef ENABLE_CHECKSUM
 295         /* Turn on checksumming by default. */
 296         cli->cl_checksum = 1;
 297         /*
 298          * The supported checksum types will be worked out at connect time
 299          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 300          * through procfs.
 301          */
 302         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 303 #endif
 304         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 305
 306         /* This value may be changed at connect time in
 307            ptlrpc_connect_interpret. */
 308         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 309                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 310
 311         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 312                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 313         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 314                 cli->cl_max_rpcs_in_flight = 2;
 315         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 316                 cli->cl_max_rpcs_in_flight = 3;
 317         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 318                 cli->cl_max_rpcs_in_flight = 4;
 319         } else {
 320                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 321         }
 322
 323         rc = ldlm_get_ref();
 324         if (rc) {
 325                 CERROR("ldlm_get_ref failed: %d\n", rc);
 326                 GOTO(err, rc);
 327         }
 328
 329         ptlrpc_init_client(rq_portal, rp_portal, name,
 330                            &obddev->obd_ldlm_client);
 331
 332         imp = class_new_import(obddev);
 333         if (imp == NULL)
 334                 GOTO(err_ldlm, rc = -ENOENT);
 335         imp->imp_client = &obddev->obd_ldlm_client;
 336         imp->imp_connect_op = connect_op;
 337         imp->imp_initial_recov = 1;
 338         imp->imp_initial_recov_bk = 0;
 339         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 340         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 341                LUSTRE_CFG_BUFLEN(lcfg, 1));
 342         class_import_put(imp);
 343
 344         rc = client_import_add_conn(imp, &server_uuid, 1);
 345         if (rc) {
 346                 CERROR("can't add initial connection\n");
 347                 GOTO(err_import, rc);
 348         }
 349
 350         cli->cl_import = imp;
 351         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 352         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 353         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 354
 355         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 356                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 357                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 358                                name, obddev->obd_name,
 359                                cli->cl_target_uuid.uuid);
 360                         spin_lock(&imp->imp_lock);
 361                         imp->imp_invalid = 1;
 362                         spin_unlock(&imp->imp_lock);
 363                 }
 364         }
 365
 366         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 367
 368         RETURN(rc);
 369
 370 err_import:
 371         class_destroy_import(imp);
 372 err_ldlm:
 373         ldlm_put_ref();
 374 err:
 375         RETURN(rc);
 376
 377 }
 378
 379 int client_obd_cleanup(struct obd_device *obddev)
 380 {
 381         ENTRY;
 382         ldlm_put_ref();
 383         RETURN(0);
 384 }
 385
 386 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 387 int client_connect_import(const struct lu_env *env,
 388                           struct lustre_handle *dlm_handle,
 389                           struct obd_device *obd, struct obd_uuid *cluuid,
 390                           struct obd_connect_data *data, void *localdata)
 391 {
 392         struct client_obd *cli = &obd->u.cli;
 393         struct obd_import *imp = cli->cl_import;
 394         struct obd_export *exp;
 395         struct obd_connect_data *ocd;
 396         struct ldlm_namespace *to_be_freed = NULL;
 397         int rc;
 398         ENTRY;
 399
 400         down_write(&cli->cl_sem);
 401         rc = class_connect(dlm_handle, obd, cluuid);
 402         if (rc)
 403                 GOTO(out_sem, rc);
 404
 405         cli->cl_conn_count++;
 406         if (cli->cl_conn_count > 1)
 407                 GOTO(out_sem, rc);
 408         exp = class_conn2export(dlm_handle);
 409
 410         if (obd->obd_namespace != NULL)
 411                 CERROR("already have namespace!\n");
 412         obd->obd_namespace = ldlm_namespace_new(obd, obd->obd_name,
 413                                                 LDLM_NAMESPACE_CLIENT,
 414                                                 LDLM_NAMESPACE_GREEDY);
 415         if (obd->obd_namespace == NULL)
 416                 GOTO(out_disco, rc = -ENOMEM);
 417
 418         imp->imp_dlm_handle = *dlm_handle;
 419         rc = ptlrpc_init_import(imp);
 420         if (rc != 0)
 421                 GOTO(out_ldlm, rc);
 422
 423         ocd = &imp->imp_connect_data;
 424         if (data) {
 425                 *ocd = *data;
 426                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 427         }
 428
 429         rc = ptlrpc_connect_import(imp, NULL);
 430         if (rc != 0) {
 431                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 432                 GOTO(out_ldlm, rc);
 433         }
 434         LASSERT(exp->exp_connection);
 435
 436         if (data) {
 437                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 438                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 439                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 440                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 441         }
 442
 443         ptlrpc_pinger_add_import(imp);
 444
 445         EXIT;
 446
 447         if (rc) {
 448 out_ldlm:
 449                 ldlm_namespace_free_prior(obd->obd_namespace, imp, 0);
 450                 to_be_freed = obd->obd_namespace;
 451                 obd->obd_namespace = NULL;
 452 out_disco:
 453                 cli->cl_conn_count--;
 454                 class_disconnect(exp);
 455         } else {
 456                 class_export_put(exp);
 457         }
 458 out_sem:
 459         up_write(&cli->cl_sem);
 460         if (to_be_freed)
 461                 ldlm_namespace_free_post(to_be_freed);
 462
 463         return rc;
 464 }
 465
 466 int client_disconnect_export(struct obd_export *exp)
 467 {
 468         struct obd_device *obd = class_exp2obd(exp);
 469         struct client_obd *cli;
 470         struct obd_import *imp;
 471         int rc = 0, err;
 472         struct ldlm_namespace *to_be_freed = NULL;
 473         ENTRY;
 474
 475         if (!obd) {
 476                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 477                        exp, exp ? exp->exp_handle.h_cookie : -1);
 478                 RETURN(-EINVAL);
 479         }
 480
 481         cli = &obd->u.cli;
 482         imp = cli->cl_import;
 483
 484         down_write(&cli->cl_sem);
 485         if (!cli->cl_conn_count) {
 486                 CERROR("disconnecting disconnected device (%s)\n",
 487                        obd->obd_name);
 488                 GOTO(out_sem, rc = -EINVAL);
 489         }
 490
 491         cli->cl_conn_count--;
 492         if (cli->cl_conn_count)
 493                 GOTO(out_no_disconnect, rc = 0);
 494
 495         /* Mark import deactivated now, so we don't try to reconnect if any
 496          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 497          * fully deactivate the import, or that would drop all requests. */
 498         spin_lock(&imp->imp_lock);
 499         imp->imp_deactive = 1;
 500         spin_unlock(&imp->imp_lock);
 501
 502         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 503          * delete it regardless.  (It's safe to delete an import that was
 504          * never added.) */
 505         (void)ptlrpc_pinger_del_import(imp);
 506
 507         if (obd->obd_namespace != NULL) {
 508                 /* obd_force == local only */
 509                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 510                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 511                                        NULL);
 512                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 513                 to_be_freed = obd->obd_namespace;
 514         }
 515
 516         rc = ptlrpc_disconnect_import(imp, 0);
 517
 518         ptlrpc_invalidate_import(imp);
 519         /* set obd_namespace to NULL only after invalidate, because we can have
 520          * some connect requests in flight, and his need store a connect flags
 521          * in obd_namespace. bug 14260 */
 522         obd->obd_namespace = NULL;
 523
 524         if (imp->imp_rq_pool) {
 525                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 526                 imp->imp_rq_pool = NULL;
 527         }
 528         destroy_import(imp);
 529         cli->cl_import = NULL;
 530
 531         EXIT;
 532  out_no_disconnect:
 533         err = class_disconnect(exp);
 534         if (!rc && err)
 535                 rc = err;
 536  out_sem:
 537         up_write(&cli->cl_sem);
 538         if (to_be_freed)
 539                 ldlm_namespace_free_post(to_be_freed);
 540
 541         RETURN(rc);
 542 }
 543
 544 /* --------------------------------------------------------------------------
 545  * from old lib/target.c
 546  * -------------------------------------------------------------------------- */
 547
 548 int target_handle_reconnect(struct lustre_handle *conn, struct obd_export *exp,
 549                             struct obd_uuid *cluuid, int initial_conn)
 550 {
 551         ENTRY;
 552         if (exp->exp_connection && exp->exp_imp_reverse && !initial_conn) {
 553                 struct lustre_handle *hdl;
 554                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 555                 /* Might be a re-connect after a partition. */
 556                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 557                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 558                               cluuid->uuid);
 559                         conn->cookie = exp->exp_handle.h_cookie;
 560                         /* target_handle_connect() treats EALREADY and
 561                          * -EALREADY differently.  EALREADY means we are
 562                          * doing a valid reconnect from the same client. */
 563                         RETURN(EALREADY);
 564                 } else {
 565                         CERROR("%s reconnecting from %s, "
 566                                "handle mismatch (ours "LPX64", theirs "
 567                                LPX64")\n", cluuid->uuid,
 568                                exp->exp_connection->c_remote_uuid.uuid,
 569                                hdl->cookie, conn->cookie);
 570                         memset(conn, 0, sizeof *conn);
 571                         /* target_handle_connect() treats EALREADY and
 572                          * -EALREADY differently.  -EALREADY is an error
 573                          * (same UUID, different handle). */
 574                         RETURN(-EALREADY);
 575                 }
 576         }
 577
 578         conn->cookie = exp->exp_handle.h_cookie;
 579         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 580                cluuid->uuid, exp, conn->cookie);
 581         RETURN(0);
 582 }
 583
 584 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 585                           int error)
 586 {
 587         struct obd_export *exp = cb_data;
 588
 589         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 590                obd->obd_name, exp->exp_client_uuid.uuid);
 591
 592         spin_lock(&exp->exp_lock);
 593         exp->exp_need_sync = 0;
 594         spin_unlock(&exp->exp_lock);
 595 }
 596 EXPORT_SYMBOL(target_client_add_cb);
 597 static void
 598 target_start_and_reset_recovery_timer(struct obd_device *obd,
 599                                       struct ptlrpc_request *req,
 600                                       int new_client);
 601
 602 int target_handle_connect(struct ptlrpc_request *req)
 603 {
 604         struct obd_device *target, *targref = NULL;
 605         struct obd_export *export = NULL;
 606         struct obd_import *revimp;
 607         struct lustre_handle conn;
 608         struct lustre_handle *tmp;
 609         struct obd_uuid tgtuuid;
 610         struct obd_uuid cluuid;
 611         struct obd_uuid remote_uuid;
 612         char *str;
 613         int rc = 0;
 614         int initial_conn = 0;
 615         struct obd_connect_data *data, *tmpdata;
 616         lnet_nid_t *client_nid = NULL;
 617         ENTRY;
 618
 619         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 620
 621         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 622         if (str == NULL) {
 623                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 624                 GOTO(out, rc = -EINVAL);
 625         }
 626
 627         obd_str2uuid(&tgtuuid, str);
 628         target = class_uuid2obd(&tgtuuid);
 629         if (!target)
 630                 target = class_name2obd(str);
 631
 632         if (!target || target->obd_stopping || !target->obd_set_up) {
 633                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 634                                    " for connect (%s)\n", str,
 635                                    !target ? "no target" :
 636                                    (target->obd_stopping ? "stopping" :
 637                                    "not set up"));
 638                 GOTO(out, rc = -ENODEV);
 639         }
 640
 641         if (target->obd_no_conn) {
 642                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 643                               "from %s\n", target->obd_name,
 644                               libcfs_nid2str(req->rq_peer.nid));
 645                 GOTO(out, rc = -EAGAIN);
 646         }
 647
 648         /* Make sure the target isn't cleaned up while we're here. Yes,
 649            there's still a race between the above check and our incref here.
 650            Really, class_uuid2obd should take the ref. */
 651         targref = class_incref(target, __FUNCTION__, cfs_current());
 652
 653
 654         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 655         if (str == NULL) {
 656                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 657                 GOTO(out, rc = -EINVAL);
 658         }
 659
 660         obd_str2uuid(&cluuid, str);
 661
 662         /* XXX extract a nettype and format accordingly */
 663         switch (sizeof(lnet_nid_t)) {
 664                 /* NB the casts only avoid compiler warnings */
 665         case 8:
 666                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 667                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 668                 break;
 669         case 4:
 670                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 671                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 672                 break;
 673         default:
 674                 LBUG();
 675         }
 676
 677         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 678         if (tmp == NULL)
 679                 GOTO(out, rc = -EPROTO);
 680
 681         conn = *tmp;
 682
 683         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 684         if (!data)
 685                 GOTO(out, rc = -EPROTO);
 686
 687         rc = req_capsule_server_pack(&req->rq_pill);
 688         if (rc)
 689                 GOTO(out, rc);
 690
 691         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 692                 if (!data) {
 693                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 694                                   "libclient connection attempt");
 695                         GOTO(out, rc = -EPROTO);
 696                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 697                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 698                            data->ocd_version > LUSTRE_VERSION_CODE +
 699                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 700                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 701                                   "libclient connection attempt",
 702                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 703                                   "old" : "new",
 704                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 705                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 706                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 707                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 708                         data = req_capsule_server_sized_get(&req->rq_pill,
 709                                                         &RMF_CONNECT_DATA,
 710                                     offsetof(typeof(*data), ocd_version) +
 711                                              sizeof(data->ocd_version));
 712                         if (data) {
 713                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 714                                 data->ocd_version = LUSTRE_VERSION_CODE;
 715                         }
 716                         GOTO(out, rc = -EPROTO);
 717                 }
 718         }
 719
 720         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL)
 721                 initial_conn = 1;
 722
 723         /* lctl gets a backstage, all-access pass. */
 724         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 725                 goto dont_check_exports;
 726
 727         spin_lock(&target->obd_dev_lock);
 728         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 729
 730         if (export != NULL && export->exp_connecting) { /* bug 9635, et. al. */
 731                 CWARN("%s: exp %p already connecting\n",
 732                       export->exp_obd->obd_name, export);
 733                 class_export_put(export);
 734                 export = NULL;
 735                 rc = -EALREADY;
 736         } else if (export != NULL && export->exp_connection != NULL &&
 737                    req->rq_peer.nid != export->exp_connection->c_peer.nid) {
 738                 /* make darn sure this is coming from the same peer
 739                  * if the UUIDs matched */
 740                   CWARN("%s: cookie %s seen on new NID %s when "
 741                           "existing NID %s is already connected\n",
 742                         target->obd_name, cluuid.uuid,
 743                   libcfs_nid2str(req->rq_peer.nid),
 744                   libcfs_nid2str(export->exp_connection->c_peer.nid));
 745                   class_export_put(export);
 746                   export = NULL;
 747                   rc = -EALREADY;
 748         } else if (export != NULL) {
 749                 spin_lock(&export->exp_lock);
 750                 export->exp_connecting = 1;
 751                 spin_unlock(&export->exp_lock);
 752                 class_export_put(export);
 753                 spin_unlock(&target->obd_dev_lock);
 754                 LASSERT(export->exp_obd == target);
 755
 756                 rc = target_handle_reconnect(&conn, export, &cluuid, initial_conn);
 757         }
 758
 759         /* If we found an export, we already unlocked. */
 760         if (!export) {
 761                 spin_unlock(&target->obd_dev_lock);
 762                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 763         } else if (req->rq_export == NULL &&
 764                    atomic_read(&export->exp_rpc_count) > 0) {
 765                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 766                       target->obd_name, cluuid.uuid,
 767                       libcfs_nid2str(req->rq_peer.nid),
 768                       export, atomic_read(&export->exp_refcount));
 769                 GOTO(out, rc = -EBUSY);
 770         } else if (req->rq_export != NULL &&
 771                    (atomic_read(&export->exp_rpc_count) > 1)) {
 772                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 773                       target->obd_name, cluuid.uuid,
 774                       libcfs_nid2str(req->rq_peer.nid),
 775                       export, atomic_read(&export->exp_rpc_count));
 776                 GOTO(out, rc = -EBUSY);
 777         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1 &&
 778                    !initial_conn) {
 779                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 780                        "cookies not random?\n", target->obd_name,
 781                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 782                 GOTO(out, rc = -EALREADY);
 783         } else {
 784                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 785                 if (req->rq_export == NULL && initial_conn)
 786                        export->exp_last_request_time =
 787                                max(export->exp_last_request_time,
 788                                    (time_t)cfs_time_current_sec());
 789         }
 790
 791         if (rc < 0) {
 792                 GOTO(out, rc);
 793         }
 794
 795         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 796                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 797               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 798               export, (long)cfs_time_current_sec(),
 799               export ? (long)export->exp_last_request_time : 0);
 800
 801         /* Tell the client if we're in recovery. */
 802         if (target->obd_recovering) {
 803                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 804                 /* If this is the first time a client connects,
 805                    reset the recovery timer */
 806                 if (rc == 0)
 807                         target_start_and_reset_recovery_timer(target, req,
 808                                                               !export);
 809         }
 810
 811         /* We want to handle EALREADY but *not* -EALREADY from
 812          * target_handle_reconnect(), return reconnection state in a flag */
 813         if (rc == EALREADY) {
 814                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 815                 rc = 0;
 816         } else {
 817                 LASSERT(rc == 0);
 818         }
 819
 820         /* Tell the client if we support replayable requests */
 821         if (target->obd_replayable)
 822                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 823         client_nid = &req->rq_peer.nid;
 824
 825         if (export == NULL) {
 826                 if (target->obd_recovering) {
 827                         cfs_time_t t;
 828
 829                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 830                         t = cfs_time_sub(t, cfs_time_current());
 831                         CERROR("%s: denying connection for new client %s (%s): "
 832                                "%d clients in recovery for "CFS_TIME_T"s\n",
 833                                target->obd_name,
 834                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 835                                target->obd_recoverable_clients,
 836                                cfs_duration_sec(t));
 837                         rc = -EBUSY;
 838                 } else {
 839 dont_check_exports:
 840                         rc = obd_connect(req->rq_svc_thread->t_env,
 841                                          &conn, target, &cluuid, data,
 842                                          client_nid);
 843                 }
 844         } else {
 845                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 846                                    export, target, &cluuid, data, client_nid);
 847         }
 848         if (rc)
 849                 GOTO(out, rc);
 850         /* Return only the parts of obd_connect_data that we understand, so the
 851          * client knows that we don't understand the rest. */
 852         if (data) {
 853                  tmpdata = req_capsule_server_get(&req->rq_pill,
 854                                                   &RMF_CONNECT_DATA);
 855                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 856                  *tmpdata = *data;
 857         }
 858
 859         /* If all else goes well, this is our RPC return code. */
 860         req->rq_status = 0;
 861
 862         lustre_msg_set_handle(req->rq_repmsg, &conn);
 863
 864         /* ownership of this export ref transfers to the request AFTER we
 865          * drop any previous reference the request had, but we don't want
 866          * that to go to zero before we get our new export reference. */
 867         export = class_conn2export(&conn);
 868         if (!export) {
 869                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 870                 GOTO(out, rc = -ENODEV);
 871         }
 872
 873         /* If the client and the server are the same node, we will already
 874          * have an export that really points to the client's DLM export,
 875          * because we have a shared handles table.
 876          *
 877          * XXX this will go away when shaver stops sending the "connect" handle
 878          * in the real "remote handle" field of the request --phik 24 Apr 2003
 879          */
 880         if (req->rq_export != NULL)
 881                 class_export_put(req->rq_export);
 882
 883         req->rq_export = export;
 884
 885         spin_lock(&export->exp_lock);
 886         if (initial_conn) {
 887                 lustre_msg_set_conn_cnt(req->rq_repmsg, export->exp_conn_cnt + 1);
 888         } else if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 889                 spin_unlock(&export->exp_lock);
 890                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 891                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 892                        export->exp_conn_cnt,
 893                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 894
 895                 GOTO(out, rc = -EALREADY);
 896         }
 897         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 898
 899         /* request from liblustre?  Don't evict it for not pinging. */
 900         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 901                 export->exp_libclient = 1;
 902                 spin_unlock(&export->exp_lock);
 903
 904                 spin_lock(&target->obd_dev_lock);
 905                 list_del_init(&export->exp_obd_chain_timed);
 906                 spin_unlock(&target->obd_dev_lock);
 907         } else {
 908                 spin_unlock(&export->exp_lock);
 909         }
 910
 911         if (export->exp_connection != NULL)
 912                 ptlrpc_connection_put(export->exp_connection);
 913         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 914                                                        req->rq_self,
 915                                                        &remote_uuid);
 916         if (hlist_unhashed(&export->exp_nid_hash)) {
 917                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 918                                        &export->exp_connection->c_peer.nid,
 919                                        &export->exp_nid_hash);
 920         }
 921
 922         spin_lock_bh(&target->obd_processing_task_lock);
 923         if (target->obd_recovering && !export->exp_in_recovery) {
 924                 spin_lock(&export->exp_lock);
 925                 export->exp_in_recovery = 1;
 926                 export->exp_req_replay_needed = 1;
 927                 export->exp_lock_replay_needed = 1;
 928                 spin_unlock(&export->exp_lock);
 929                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 930                      && (data->ocd_transno == 0))
 931                         CWARN("Connect with zero transno!\n");
 932
 933                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
 934                      && data->ocd_transno < target->obd_next_recovery_transno)
 935                         target->obd_next_recovery_transno = data->ocd_transno;
 936                 target->obd_connected_clients++;
 937                 /* each connected client is counted as recoverable */
 938                 target->obd_recoverable_clients++;
 939                 atomic_inc(&target->obd_req_replay_clients);
 940                 atomic_inc(&target->obd_lock_replay_clients);
 941                 if (target->obd_connected_clients ==
 942                     target->obd_max_recoverable_clients)
 943                         wake_up(&target->obd_next_transno_waitq);
 944         }
 945         spin_unlock_bh(&target->obd_processing_task_lock);
 946         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 947         conn = *tmp;
 948
 949         if (export->exp_imp_reverse != NULL) {
 950                 /* destroyed import can be still referenced in ctxt */
 951                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
 952                                    KEY_REVIMP_UPD, 0, NULL, NULL);
 953
 954                 /* in some recovery senarios, previous ctx init rpc handled
 955                  * in sptlrpc_target_export_check() might be used to install
 956                  * a reverse ctx in this reverse import, and later OBD_CONNECT
 957                  * using the same gss ctx could reach here and following new
 958                  * reverse import. note all reverse ctx in new/old import are
 959                  * actually based on the same gss ctx. so we invalidate ctx
 960                  * here before destroy import, otherwise flush old import will
 961                  * lead to remote reverse ctx be destroied, thus the reverse
 962                  * ctx of new import will lost its peer.
 963                  * there might be a better way to deal with this???
 964                  */
 965                 sptlrpc_import_inval_all_ctx(export->exp_imp_reverse);
 966
 967                 destroy_import(export->exp_imp_reverse);
 968         }
 969
 970         /* for the rest part, we return -ENOTCONN in case of errors
 971          * in order to let client initialize connection again.
 972          */
 973         revimp = export->exp_imp_reverse = class_new_import(target);
 974         if (!revimp) {
 975                 CERROR("fail to alloc new reverse import.\n");
 976                 GOTO(out, rc = -ENOTCONN);
 977         }
 978
 979         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
 980         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
 981         revimp->imp_remote_handle = conn;
 982         revimp->imp_dlm_fake = 1;
 983         revimp->imp_state = LUSTRE_IMP_FULL;
 984
 985         /* unknown versions will be caught in
 986          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
 987         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
 988
 989         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
 990             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
 991                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
 992         else
 993                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
 994
 995         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx,
 996                                       req->rq_flvr.sf_rpc);
 997         if (rc) {
 998                 CERROR("Failed to get sec for reverse import: %d\n", rc);
 999                 export->exp_imp_reverse = NULL;
1000                 class_destroy_import(revimp);
1001         }
1002
1003         class_import_put(revimp);
1004 out:
1005         if (export) {
1006                 spin_lock(&export->exp_lock);
1007                 export->exp_connecting = 0;
1008                 spin_unlock(&export->exp_lock);
1009         }
1010         if (targref)
1011                 class_decref(targref, __FUNCTION__, cfs_current());
1012         if (rc)
1013                 req->rq_status = rc;
1014         RETURN(rc);
1015 }
1016
1017 int target_handle_disconnect(struct ptlrpc_request *req)
1018 {
1019         int rc;
1020         ENTRY;
1021
1022         rc = req_capsule_server_pack(&req->rq_pill);
1023         if (rc)
1024                 RETURN(rc);
1025
1026         /* keep the rq_export around so we can send the reply */
1027         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1028
1029         RETURN(0);
1030 }
1031
1032 void target_destroy_export(struct obd_export *exp)
1033 {
1034         /* exports created from last_rcvd data, and "fake"
1035            exports created by lctl don't have an import */
1036         if (exp->exp_imp_reverse != NULL)
1037                 destroy_import(exp->exp_imp_reverse);
1038
1039         /* We cancel locks at disconnect time, but this will catch any locks
1040          * granted in a race with recovery-induced disconnect. */
1041         if (exp->exp_obd->obd_namespace != NULL)
1042                 ldlm_cancel_locks_for_export(exp);
1043 }
1044
1045 /*
1046  * Recovery functions
1047  */
1048
1049 struct ptlrpc_request *ptlrpc_clone_req( struct ptlrpc_request *orig_req)
1050 {
1051         struct ptlrpc_request *copy_req;
1052         struct lustre_msg *copy_reqmsg;
1053         struct ptlrpc_user_desc *udesc = NULL;
1054
1055         OBD_ALLOC_PTR(copy_req);
1056         if (!copy_req)
1057                 return NULL;
1058         OBD_ALLOC(copy_reqmsg, orig_req->rq_reqlen);
1059         if (!copy_reqmsg){
1060                 OBD_FREE_PTR(copy_req);
1061                 return NULL;
1062         }
1063
1064         if (orig_req->rq_user_desc) {
1065                 int ngroups = orig_req->rq_user_desc->pud_ngroups;
1066
1067                 OBD_ALLOC(udesc, sptlrpc_user_desc_size(ngroups));
1068                 if (!udesc) {
1069                         OBD_FREE(copy_reqmsg, orig_req->rq_reqlen);
1070                         OBD_FREE_PTR(copy_req);
1071                         return NULL;
1072                 }
1073                 memcpy(udesc, orig_req->rq_user_desc,
1074                        sptlrpc_user_desc_size(ngroups));
1075         }
1076
1077         *copy_req = *orig_req;
1078         memcpy(copy_reqmsg, orig_req->rq_reqmsg, orig_req->rq_reqlen);
1079         copy_req->rq_reqmsg = copy_reqmsg;
1080         copy_req->rq_user_desc = udesc;
1081
1082         class_export_rpc_get(copy_req->rq_export);
1083         CFS_INIT_LIST_HEAD(&copy_req->rq_list);
1084         CFS_INIT_LIST_HEAD(&copy_req->rq_replay_list);
1085         sptlrpc_svc_ctx_addref(copy_req);
1086
1087         if (copy_req->rq_reply_state) {
1088                 /* the copied req takes over the reply state */
1089                 orig_req->rq_reply_state = NULL;
1090                 /* to catch further access */
1091                 orig_req->rq_repmsg = NULL;
1092                 orig_req->rq_replen = 0;
1093         }
1094
1095         return copy_req;
1096 }
1097
1098 void ptlrpc_free_clone(struct ptlrpc_request *req)
1099 {
1100         LASSERT(list_empty(&req->rq_replay_list));
1101
1102         ptlrpc_req_drop_rs(req);
1103         sptlrpc_svc_ctx_decref(req);
1104         class_export_rpc_put(req->rq_export);
1105         list_del_init(&req->rq_list);
1106
1107         if (req->rq_user_desc) {
1108                 int ngroups = req->rq_user_desc->pud_ngroups;
1109                 OBD_FREE(req->rq_user_desc, sptlrpc_user_desc_size(ngroups));
1110         }
1111         OBD_FREE(req->rq_reqmsg, req->rq_reqlen);
1112         OBD_FREE_PTR(req);
1113 }
1114
1115 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1116 {
1117         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1118         struct obd_export     *exp = req->rq_export;
1119         struct ptlrpc_request *reqiter;
1120         int                    dup = 0;
1121
1122         LASSERT(exp);
1123
1124         spin_lock(&exp->exp_lock);
1125         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1126                             rq_replay_list) {
1127                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1128                         dup = 1;
1129                         break;
1130                 }
1131         }
1132
1133         if (dup) {
1134                 /* we expect it with RESENT and REPLAY flags */
1135                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1136                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1137                         CERROR("invalid flags %x of resent replay\n",
1138                                lustre_msg_get_flags(req->rq_reqmsg));
1139         } else {
1140                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1141         }
1142
1143         spin_unlock(&exp->exp_lock);
1144         return dup;
1145 }
1146
1147 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1148 {
1149         LASSERT(!list_empty(&req->rq_replay_list));
1150         LASSERT(req->rq_export);
1151
1152         spin_lock(&req->rq_export->exp_lock);
1153         list_del_init(&req->rq_replay_list);
1154         spin_unlock(&req->rq_export->exp_lock);
1155 }
1156
1157 #ifdef __KERNEL__
1158 static void target_finish_recovery(struct obd_device *obd)
1159 {
1160         ENTRY;
1161         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1162                       obd->obd_name);
1163
1164         ldlm_reprocess_all_ns(obd->obd_namespace);
1165         spin_lock_bh(&obd->obd_processing_task_lock);
1166         if (list_empty(&obd->obd_req_replay_queue) &&
1167             list_empty(&obd->obd_lock_replay_queue) &&
1168             list_empty(&obd->obd_final_req_queue)) {
1169                 obd->obd_processing_task = 0;
1170         } else {
1171                 CERROR("%s: Recovery queues ( %s%s%s) are empty\n",
1172                        obd->obd_name,
1173                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1174                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1175                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1176                 spin_unlock_bh(&obd->obd_processing_task_lock);
1177                 LBUG();
1178         }
1179         spin_unlock_bh(&obd->obd_processing_task_lock);
1180
1181         /* when recovery finished, cleanup orphans on mds and ost */
1182         if (OBT(obd) && OBP(obd, postrecov)) {
1183                 int rc = OBP(obd, postrecov)(obd);
1184                 LCONSOLE_WARN("%s: recovery %s: rc %d\n", obd->obd_name,
1185                               rc < 0 ? "failed" : "complete", rc);
1186         }
1187
1188         obd->obd_recovery_end = cfs_time_current_sec();
1189         EXIT;
1190 }
1191
1192 static void abort_req_replay_queue(struct obd_device *obd)
1193 {
1194         struct ptlrpc_request *req, *n;
1195         struct list_head abort_list;
1196
1197         CFS_INIT_LIST_HEAD(&abort_list);
1198         spin_lock_bh(&obd->obd_processing_task_lock);
1199         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1200         spin_unlock_bh(&obd->obd_processing_task_lock);
1201         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1202                 DEBUG_REQ(D_WARNING, req, "aborted:");
1203                 req->rq_status = -ENOTCONN;
1204                 if (ptlrpc_error(req)) {
1205                         DEBUG_REQ(D_ERROR, req,
1206                                   "failed abort_req_reply; skipping");
1207                 }
1208                 target_exp_dequeue_req_replay(req);
1209                 ptlrpc_free_clone(req);
1210         }
1211 }
1212
1213 static void abort_lock_replay_queue(struct obd_device *obd)
1214 {
1215         struct ptlrpc_request *req, *n;
1216         struct list_head abort_list;
1217
1218         CFS_INIT_LIST_HEAD(&abort_list);
1219         spin_lock_bh(&obd->obd_processing_task_lock);
1220         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1221         spin_unlock_bh(&obd->obd_processing_task_lock);
1222         list_for_each_entry_safe(req, n, &obd->obd_lock_replay_queue, rq_list){
1223                 DEBUG_REQ(D_ERROR, req, "aborted:");
1224                 req->rq_status = -ENOTCONN;
1225                 if (ptlrpc_error(req)) {
1226                         DEBUG_REQ(D_ERROR, req,
1227                                   "failed abort_lock_reply; skipping");
1228                 }
1229                 ptlrpc_free_clone(req);
1230         }
1231 }
1232 #endif
1233
1234 /* Called from a cleanup function if the device is being cleaned up
1235    forcefully.  The exports should all have been disconnected already,
1236    the only thing left to do is
1237      - clear the recovery flags
1238      - cancel the timer
1239      - free queued requests and replies, but don't send replies
1240    Because the obd_stopping flag is set, no new requests should be received.
1241
1242 */
1243 void target_cleanup_recovery(struct obd_device *obd)
1244 {
1245         struct ptlrpc_request *req, *n;
1246         struct list_head clean_list;
1247         ENTRY;
1248
1249         LASSERT(obd->obd_stopping);
1250
1251         CFS_INIT_LIST_HEAD(&clean_list);
1252         spin_lock_bh(&obd->obd_processing_task_lock);
1253         if (!obd->obd_recovering) {
1254                 spin_unlock_bh(&obd->obd_processing_task_lock);
1255                 EXIT;
1256                 return;
1257         }
1258         obd->obd_recovering = obd->obd_abort_recovery = 0;
1259         target_cancel_recovery_timer(obd);
1260
1261         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1262         spin_unlock_bh(&obd->obd_processing_task_lock);
1263
1264         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1265                 LASSERT(req->rq_reply_state == 0);
1266                 target_exp_dequeue_req_replay(req);
1267                 ptlrpc_free_clone(req);
1268         }
1269
1270         spin_lock_bh(&obd->obd_processing_task_lock);
1271         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1272         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1273         spin_unlock_bh(&obd->obd_processing_task_lock);
1274
1275         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1276                 LASSERT(req->rq_reply_state == 0);
1277                 ptlrpc_free_clone(req);
1278         }
1279
1280         EXIT;
1281 }
1282
1283 /* obd_processing_task_lock should be held */
1284 void target_cancel_recovery_timer(struct obd_device *obd)
1285 {
1286         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1287         cfs_timer_disarm(&obd->obd_recovery_timer);
1288 }
1289
1290 /* extend = 1 means require at least "duration" seconds left in the timer,
1291    extend = 0 means set the total duration (start_recovery_timer) */
1292 static void reset_recovery_timer(struct obd_device *obd, int duration,
1293                                  int extend)
1294 {
1295         cfs_time_t now = cfs_time_current_sec();
1296         cfs_duration_t left;
1297
1298         spin_lock_bh(&obd->obd_processing_task_lock);
1299         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1300                 spin_unlock_bh(&obd->obd_processing_task_lock);
1301                 return;
1302         }
1303
1304         left = cfs_time_sub(obd->obd_recovery_end, now);
1305
1306         if (extend && (duration > left))
1307                 obd->obd_recovery_timeout += duration - left;
1308         else if (!extend && (duration > obd->obd_recovery_timeout))
1309                 /* Track the client's largest expected replay time */
1310                 obd->obd_recovery_timeout = duration;
1311 #ifdef CRAY_XT3
1312         /*
1313          * If total recovery time already exceed the
1314          * obd_recovery_max_time, then CRAY XT3 will
1315          * abort the recovery
1316          */
1317         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1318                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1319 #endif
1320         obd->obd_recovery_end = obd->obd_recovery_start +
1321                                 obd->obd_recovery_timeout;
1322         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1323             cfs_time_before(now, obd->obd_recovery_end)) {
1324                 left = cfs_time_sub(obd->obd_recovery_end, now);
1325                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1326         }
1327         spin_unlock_bh(&obd->obd_processing_task_lock);
1328         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1329                obd->obd_name, (unsigned)left);
1330 }
1331
1332 static void check_and_start_recovery_timer(struct obd_device *obd)
1333 {
1334         spin_lock_bh(&obd->obd_processing_task_lock);
1335         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1336                 spin_unlock_bh(&obd->obd_processing_task_lock);
1337                 return;
1338         }
1339         CWARN("%s: starting recovery timer\n", obd->obd_name);
1340         obd->obd_recovery_start = cfs_time_current_sec();
1341         /* minimum */
1342         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1343         spin_unlock_bh(&obd->obd_processing_task_lock);
1344
1345         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1346 }
1347
1348 /* Reset the timer with each new client connection */
1349 /*
1350  * This timer is actually reconnect_timer, which is for making sure
1351  * the total recovery window is at least as big as my reconnect
1352  * attempt timing. So the initial recovery time_out will be set to
1353  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1354  * from client is bigger than this, then the recovery time_out will
1355  * be extend to make sure the client could be reconnected, in the
1356  * process, the timeout from the new client should be ignored.
1357  */
1358
1359 static void
1360 target_start_and_reset_recovery_timer(struct obd_device *obd,
1361                                       struct ptlrpc_request *req,
1362                                       int new_client)
1363 {
1364         int req_timeout = lustre_msg_get_timeout(req->rq_reqmsg);
1365
1366         /* teach server about old server's estimates */
1367         if (!new_client)
1368                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1369                        at_timeout2est(req_timeout));
1370
1371         check_and_start_recovery_timer(obd);
1372
1373         req_timeout *= OBD_RECOVERY_FACTOR;
1374         if (req_timeout > obd->obd_recovery_timeout && !new_client)
1375                 reset_recovery_timer(obd, req_timeout, 0);
1376 }
1377
1378 #ifdef __KERNEL__
1379 static int check_for_next_transno(struct obd_device *obd)
1380 {
1381         struct ptlrpc_request *req = NULL;
1382         int wake_up = 0, connected, completed, queue_len, max;
1383         __u64 next_transno, req_transno;
1384         ENTRY;
1385         spin_lock_bh(&obd->obd_processing_task_lock);
1386
1387         if (!list_empty(&obd->obd_req_replay_queue)) {
1388                 req = list_entry(obd->obd_req_replay_queue.next,
1389                                  struct ptlrpc_request, rq_list);
1390                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1391         } else {
1392                 req_transno = 0;
1393         }
1394
1395         max = obd->obd_max_recoverable_clients;
1396         connected = obd->obd_connected_clients;
1397         completed = connected - obd->obd_recoverable_clients;
1398         queue_len = obd->obd_requests_queued_for_recovery;
1399         next_transno = obd->obd_next_recovery_transno;
1400
1401         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1402                "req_transno: "LPU64", next_transno: "LPU64"\n",
1403                max, connected, completed, queue_len, req_transno, next_transno);
1404
1405         if (obd->obd_abort_recovery) {
1406                 CDEBUG(D_HA, "waking for aborted recovery\n");
1407                 wake_up = 1;
1408         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1409                 CDEBUG(D_HA, "waking for completed recovery\n");
1410                 wake_up = 1;
1411         } else if (req_transno == next_transno) {
1412                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1413                 wake_up = 1;
1414         } else if (queue_len + completed == max) {
1415                 /* handle gaps occured due to lost reply. It is allowed gaps
1416                  * because all clients are connected and there will be resend
1417                  * for missed transaction */
1418                 LASSERTF(req_transno >= next_transno,
1419                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1420                          req_transno, next_transno);
1421
1422                 CDEBUG(req_transno > obd->obd_last_committed ? D_ERROR : D_HA,
1423                        "waking for skipped transno (skip: "LPD64
1424                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1425                        next_transno, queue_len, completed, connected, req_transno);
1426                 obd->obd_next_recovery_transno = req_transno;
1427                 wake_up = 1;
1428         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1429                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1430                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1431                 obd->obd_next_recovery_transno = req_transno;
1432                 wake_up = 1;
1433         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1434                 /* some clients haven't connected in time, but we can try
1435                  * to replay requests that demand on already committed ones
1436                  * also, we can replay first non-committed transation */
1437                 LASSERT(req_transno != 0);
1438                 if (req_transno == obd->obd_last_committed + 1) {
1439                         obd->obd_next_recovery_transno = req_transno;
1440                 } else if (req_transno > obd->obd_last_committed) {
1441                         /* can't continue recovery: have no needed transno */
1442                         obd->obd_abort_recovery = 1;
1443                         CDEBUG(D_ERROR, "abort due to missed clients. max: %d, "
1444                                "connected: %d, completed: %d, queue_len: %d, "
1445                                "req_transno: "LPU64", next_transno: "LPU64"\n",
1446                                max, connected, completed, queue_len,
1447                                req_transno, next_transno);
1448                 }
1449                 wake_up = 1;
1450         }
1451
1452         spin_unlock_bh(&obd->obd_processing_task_lock);
1453         return wake_up;
1454 }
1455
1456 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1457 {
1458         struct l_wait_info lwi = { 0 };
1459         struct ptlrpc_request *req;
1460
1461         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1462                obd->obd_next_recovery_transno);
1463         l_wait_event(obd->obd_next_transno_waitq,
1464                      check_for_next_transno(obd), &lwi);
1465
1466         spin_lock_bh(&obd->obd_processing_task_lock);
1467         if (obd->obd_abort_recovery) {
1468                 req = NULL;
1469         } else if (!list_empty(&obd->obd_req_replay_queue)) {
1470                 req = list_entry(obd->obd_req_replay_queue.next,
1471                                  struct ptlrpc_request, rq_list);
1472                 target_exp_dequeue_req_replay(req);
1473                 list_del_init(&req->rq_list);
1474                 obd->obd_requests_queued_for_recovery--;
1475         } else {
1476                 req = NULL;
1477         }
1478         spin_unlock_bh(&obd->obd_processing_task_lock);
1479         RETURN(req);
1480 }
1481
1482 static int check_for_next_lock(struct obd_device *obd)
1483 {
1484         struct ptlrpc_request *req = NULL;
1485         int wake_up = 0;
1486
1487         spin_lock_bh(&obd->obd_processing_task_lock);
1488         if (!list_empty(&obd->obd_lock_replay_queue)) {
1489                 req = list_entry(obd->obd_lock_replay_queue.next,
1490                                  struct ptlrpc_request, rq_list);
1491                 CDEBUG(D_HA, "waking for next lock\n");
1492                 wake_up = 1;
1493         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1494                 CDEBUG(D_HA, "waking for completed lock replay\n");
1495                 wake_up = 1;
1496         } else if (obd->obd_abort_recovery) {
1497                 CDEBUG(D_HA, "waking for aborted recovery\n");
1498                 wake_up = 1;
1499         }
1500         spin_unlock_bh(&obd->obd_processing_task_lock);
1501
1502         return wake_up;
1503 }
1504
1505 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1506 {
1507         struct l_wait_info lwi = { 0 };
1508         struct ptlrpc_request *req;
1509
1510         CDEBUG(D_HA, "Waiting for lock\n");
1511         l_wait_event(obd->obd_next_transno_waitq,
1512                      check_for_next_lock(obd), &lwi);
1513
1514         spin_lock_bh(&obd->obd_processing_task_lock);
1515         if (obd->obd_abort_recovery) {
1516                 req = NULL;
1517         } else if (!list_empty(&obd->obd_lock_replay_queue)) {
1518                 req = list_entry(obd->obd_lock_replay_queue.next,
1519                                  struct ptlrpc_request, rq_list);
1520                 list_del_init(&req->rq_list);
1521         } else {
1522                 req = NULL;
1523         }
1524         spin_unlock_bh(&obd->obd_processing_task_lock);
1525         return req;
1526 }
1527
1528 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1529 {
1530         struct ptlrpc_request *req;
1531
1532         spin_lock_bh(&obd->obd_processing_task_lock);
1533         if (!list_empty(&obd->obd_final_req_queue)) {
1534                 req = list_entry(obd->obd_final_req_queue.next,
1535                                  struct ptlrpc_request, rq_list);
1536                 list_del_init(&req->rq_list);
1537         } else {
1538                 req = NULL;
1539         }
1540         spin_unlock_bh(&obd->obd_processing_task_lock);
1541         return req;
1542 }
1543
1544 static inline int req_replay_done(struct obd_export *exp)
1545 {
1546         return (exp->exp_req_replay_needed == 0);
1547 }
1548
1549 static inline int lock_replay_done(struct obd_export *exp)
1550 {
1551         return (exp->exp_lock_replay_needed == 0);
1552 }
1553
1554 static inline int connect_done(struct obd_export *exp)
1555 {
1556         return (exp->exp_in_recovery != 0);
1557 }
1558
1559 static int check_for_clients(struct obd_device *obd)
1560 {
1561         if (obd->obd_abort_recovery)
1562                 return 1;
1563         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1564         if (obd->obd_no_conn == 0 &&
1565             obd->obd_connected_clients == obd->obd_max_recoverable_clients)
1566                 return 1;
1567         return 0;
1568 }
1569
1570 static int handle_recovery_req(struct ptlrpc_thread *thread,
1571                                struct ptlrpc_request *req,
1572                                svc_handler_t handler)
1573 {
1574         int rc;
1575         ENTRY;
1576
1577         rc = lu_context_init(&req->rq_session, LCT_SESSION);
1578         if (rc) {
1579                 CERROR("Failure to initialize session: %d\n", rc);
1580                 return rc;
1581         }
1582         req->rq_session.lc_thread = thread;
1583         lu_context_enter(&req->rq_session);
1584         req->rq_svc_thread = thread;
1585         req->rq_svc_thread->t_env->le_ses = &req->rq_session;
1586
1587         /* thread context */
1588         lu_context_enter(&thread->t_env->le_ctx);
1589         (void)handler(req);
1590         lu_context_exit(&thread->t_env->le_ctx);
1591
1592         lu_context_exit(&req->rq_session);
1593         lu_context_fini(&req->rq_session);
1594         /* don't reset timer for final stage */
1595         if (!req_replay_done(req->rq_export) ||
1596             !lock_replay_done(req->rq_export))
1597                 reset_recovery_timer(class_exp2obd(req->rq_export),
1598                        OBD_RECOVERY_FACTOR * AT_OFF ? obd_timeout :
1599                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1600         ptlrpc_free_clone(req);
1601         RETURN(0);
1602 }
1603
1604 static void resume_recovery_timer(struct obd_device *obd)
1605 {
1606         /* to be safe, make it at least OBD_RECOVERY_FACTOR * obd_timeout */
1607         reset_recovery_timer(obd, OBD_RECOVERY_FACTOR * obd_timeout, 1);
1608 }
1609
1610 static int target_recovery_thread(void *arg)
1611 {
1612         struct obd_device *obd = arg;
1613         struct ptlrpc_request *req;
1614         struct target_recovery_data *trd = &obd->obd_recovery_data;
1615         struct l_wait_info lwi = { 0 };
1616         unsigned long delta;
1617         unsigned long flags;
1618         struct lu_env env;
1619         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1620         int rc = 0;
1621         ENTRY;
1622
1623         cfs_daemonize("tgt_recov");
1624
1625         SIGNAL_MASK_LOCK(current, flags);
1626         sigfillset(&current->blocked);
1627         RECALC_SIGPENDING;
1628         SIGNAL_MASK_UNLOCK(current, flags);
1629
1630         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1631         if (rc)
1632                 RETURN(rc);
1633
1634         thread->t_env = &env;
1635         env.le_ctx.lc_thread = thread;
1636
1637         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1638                current->pid);
1639         trd->trd_processing_task = current->pid;
1640
1641         obd->obd_recovering = 1;
1642         complete(&trd->trd_starting);
1643
1644         /* first of all, we have to know the first transno to replay */
1645         obd->obd_abort_recovery = 0;
1646         l_wait_event(obd->obd_next_transno_waitq,
1647                      check_for_clients(obd), &lwi);
1648
1649         spin_lock_bh(&obd->obd_processing_task_lock);
1650         target_cancel_recovery_timer(obd);
1651         spin_unlock_bh(&obd->obd_processing_task_lock);
1652
1653         /* If some clients haven't connected in time, evict them */
1654         if (obd->obd_abort_recovery) {
1655                 CWARN("Some clients haven't connect in time (%d/%d),"
1656                        "evict them\n", obd->obd_connected_clients,
1657                        obd->obd_max_recoverable_clients);
1658                 obd->obd_abort_recovery = obd->obd_stopping;
1659                 class_disconnect_stale_exports(obd, connect_done);
1660         }
1661         /* next stage: replay requests */
1662         delta = jiffies;
1663         obd->obd_req_replaying = 1;
1664         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1665               atomic_read(&obd->obd_req_replay_clients),
1666               obd->obd_next_recovery_transno);
1667         resume_recovery_timer(obd);
1668         while ((req = target_next_replay_req(obd))) {
1669                 LASSERT(trd->trd_processing_task == current->pid);
1670                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1671                           lustre_msg_get_transno(req->rq_reqmsg),
1672                           libcfs_nid2str(req->rq_peer.nid));
1673                 handle_recovery_req(thread, req,
1674                                     trd->trd_recovery_handler);
1675                 obd->obd_replayed_requests++;
1676                 spin_lock_bh(&obd->obd_processing_task_lock);
1677                 obd->obd_next_recovery_transno++;
1678                 spin_unlock_bh(&obd->obd_processing_task_lock);
1679         }
1680
1681         spin_lock_bh(&obd->obd_processing_task_lock);
1682         target_cancel_recovery_timer(obd);
1683         spin_unlock_bh(&obd->obd_processing_task_lock);
1684
1685         /* If some clients haven't replayed requests in time, evict them */
1686         if (obd->obd_abort_recovery) {
1687                 CDEBUG(D_ERROR, "req replay timed out, aborting ...\n");
1688                 obd->obd_abort_recovery = obd->obd_stopping;
1689                 class_disconnect_stale_exports(obd, req_replay_done);
1690                 abort_req_replay_queue(obd);
1691         }
1692
1693         /* The second stage: replay locks */
1694         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1695                atomic_read(&obd->obd_lock_replay_clients));
1696         resume_recovery_timer(obd);
1697         while ((req = target_next_replay_lock(obd))) {
1698                 LASSERT(trd->trd_processing_task == current->pid);
1699                 DEBUG_REQ(D_HA|D_WARNING, req, "processing lock from %s: ",
1700                           libcfs_nid2str(req->rq_peer.nid));
1701                 handle_recovery_req(thread, req,
1702                                     trd->trd_recovery_handler);
1703                 obd->obd_replayed_locks++;
1704         }
1705
1706         spin_lock_bh(&obd->obd_processing_task_lock);
1707         target_cancel_recovery_timer(obd);
1708         spin_unlock_bh(&obd->obd_processing_task_lock);
1709         /* If some clients haven't replayed requests in time, evict them */
1710         if (obd->obd_abort_recovery) {
1711                 int stale;
1712                 CERROR("lock replay timed out, aborting ...\n");
1713                 obd->obd_abort_recovery = obd->obd_stopping;
1714                 stale = class_disconnect_stale_exports(obd, lock_replay_done);
1715                 abort_lock_replay_queue(obd);
1716         }
1717
1718         /* We drop recoverying flag to forward all new requests
1719          * to regular mds_handle() since now */
1720         spin_lock_bh(&obd->obd_processing_task_lock);
1721         obd->obd_recovering = obd->obd_abort_recovery = 0;
1722         spin_unlock_bh(&obd->obd_processing_task_lock);
1723         /* The third stage: reply on final pings */
1724         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1725         while ((req = target_next_final_ping(obd))) {
1726                 LASSERT(trd->trd_processing_task == current->pid);
1727                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1728                           libcfs_nid2str(req->rq_peer.nid));
1729                 handle_recovery_req(thread, req,
1730                                     trd->trd_recovery_handler);
1731         }
1732
1733         delta = (jiffies - delta) / HZ;
1734         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1735               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1736         LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1737         LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1738         if (delta > obd_timeout * 2) {
1739                 CWARN("too long recovery - read logs\n");
1740                 libcfs_debug_dumplog();
1741         }
1742
1743         target_finish_recovery(obd);
1744
1745         lu_context_fini(&env.le_ctx);
1746         trd->trd_processing_task = 0;
1747         complete(&trd->trd_finishing);
1748         RETURN(rc);
1749 }
1750
1751 int target_start_recovery_thread(struct obd_device *obd, svc_handler_t handler)
1752 {
1753         int rc = 0;
1754         struct target_recovery_data *trd = &obd->obd_recovery_data;
1755
1756         memset(trd, 0, sizeof(*trd));
1757         init_completion(&trd->trd_starting);
1758         init_completion(&trd->trd_finishing);
1759         trd->trd_recovery_handler = handler;
1760
1761         if (kernel_thread(target_recovery_thread, obd, 0) > 0) {
1762                 wait_for_completion(&trd->trd_starting);
1763                 LASSERT(obd->obd_recovering != 0);
1764         } else
1765                 rc = -ECHILD;
1766
1767         return rc;
1768 }
1769
1770 void target_stop_recovery_thread(struct obd_device *obd)
1771 {
1772         spin_lock_bh(&obd->obd_processing_task_lock);
1773         if (obd->obd_recovery_data.trd_processing_task > 0) {
1774                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1775                 CERROR("%s: Aborting recovery\n", obd->obd_name);
1776                 obd->obd_abort_recovery = 1;
1777                 wake_up(&obd->obd_next_transno_waitq);
1778                 spin_unlock_bh(&obd->obd_processing_task_lock);
1779                 wait_for_completion(&trd->trd_finishing);
1780         } else {
1781                 spin_unlock_bh(&obd->obd_processing_task_lock);
1782         }
1783 }
1784
1785 void target_recovery_fini(struct obd_device *obd)
1786 {
1787         class_disconnect_exports(obd);
1788         target_stop_recovery_thread(obd);
1789         target_cleanup_recovery(obd);
1790 }
1791 EXPORT_SYMBOL(target_recovery_fini);
1792
1793 static void target_recovery_expired(unsigned long castmeharder)
1794 {
1795         struct obd_device *obd = (struct obd_device *)castmeharder;
1796         LCONSOLE_WARN("%s: recovery timed out; %d clients never reconnected "
1797                       "after %lds (%d clients did)\n",
1798                       obd->obd_name, obd->obd_recoverable_clients,
1799                       cfs_time_current_sec()- obd->obd_recovery_start,
1800                       obd->obd_connected_clients);
1801         spin_lock_bh(&obd->obd_processing_task_lock);
1802         if (obd->obd_recovering)
1803                 obd->obd_abort_recovery = 1;
1804         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1805         spin_unlock_bh(&obd->obd_processing_task_lock);
1806 }
1807
1808 void target_recovery_init(struct obd_device *obd, svc_handler_t handler)
1809 {
1810         if (obd->obd_max_recoverable_clients == 0)
1811                 return;
1812
1813         CWARN("RECOVERY: service %s, %d recoverable clients, "
1814               "last_transno "LPU64"\n", obd->obd_name,
1815               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1816         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1817         obd->obd_recovery_start = 0;
1818         obd->obd_recovery_end = 0;
1819         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1820         /* bz13079: this should be set to desired value for ost but not for mds */
1821         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1822         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1823         target_start_recovery_thread(obd, handler);
1824 }
1825 EXPORT_SYMBOL(target_recovery_init);
1826
1827 #endif
1828
1829 int target_process_req_flags(struct obd_device *obd, struct ptlrpc_request *req)
1830 {
1831         struct obd_export *exp = req->rq_export;
1832         LASSERT(exp != NULL);
1833         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1834                 /* client declares he's ready to replay locks */
1835                 spin_lock_bh(&obd->obd_processing_task_lock);
1836                 if (exp->exp_req_replay_needed) {
1837                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1838                         spin_lock(&exp->exp_lock);
1839                         exp->exp_req_replay_needed = 0;
1840                         spin_unlock(&exp->exp_lock);
1841                         atomic_dec(&obd->obd_req_replay_clients);
1842                         LASSERT(obd->obd_recoverable_clients > 0);
1843                         obd->obd_recoverable_clients--;
1844                         if (atomic_read(&obd->obd_req_replay_clients) == 0)
1845                                 CDEBUG(D_HA, "all clients have replayed reqs\n");
1846                         wake_up(&obd->obd_next_transno_waitq);
1847                 }
1848                 spin_unlock_bh(&obd->obd_processing_task_lock);
1849         }
1850         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1851                 /* client declares he's ready to complete recovery
1852                  * so, we put the request on th final queue */
1853                 spin_lock_bh(&obd->obd_processing_task_lock);
1854                 if (exp->exp_lock_replay_needed) {
1855                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1856                         spin_lock(&exp->exp_lock);
1857                         exp->exp_lock_replay_needed = 0;
1858                         spin_unlock(&exp->exp_lock);
1859                         atomic_dec(&obd->obd_lock_replay_clients);
1860                         if (atomic_read(&obd->obd_lock_replay_clients) == 0)
1861                                 CDEBUG(D_HA, "all clients have replayed locks\n");
1862                         wake_up(&obd->obd_next_transno_waitq);
1863                 }
1864                 spin_unlock_bh(&obd->obd_processing_task_lock);
1865         }
1866
1867         return 0;
1868 }
1869
1870 int target_queue_recovery_request(struct ptlrpc_request *req,
1871                                   struct obd_device *obd)
1872 {
1873         struct list_head *tmp;
1874         int inserted = 0;
1875         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1876
1877         ENTRY;
1878
1879         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1880                 /* Processing the queue right now, don't re-add. */
1881                 RETURN(1);
1882         }
1883
1884         target_process_req_flags(obd, req);
1885
1886         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1887                 /* client declares he's ready to complete recovery
1888                  * so, we put the request on th final queue */
1889                 req = ptlrpc_clone_req(req);
1890                 if (req == NULL)
1891                         RETURN(-ENOMEM);
1892                 DEBUG_REQ(D_HA, req, "queue final req");
1893                 spin_lock_bh(&obd->obd_processing_task_lock);
1894                 if (obd->obd_recovering)
1895                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1896                 else {
1897                         spin_unlock_bh(&obd->obd_processing_task_lock);
1898                         ptlrpc_free_clone(req);
1899                         if (obd->obd_stopping) {
1900                                 RETURN(-ENOTCONN);
1901                         } else {
1902                                 RETURN(1);
1903                         }
1904                 }
1905                 spin_unlock_bh(&obd->obd_processing_task_lock);
1906                 RETURN(0);
1907         }
1908         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1909                 /* client declares he's ready to replay locks */
1910                 req = ptlrpc_clone_req(req);
1911                 if (req == NULL)
1912                         RETURN(-ENOMEM);
1913                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1914                 spin_lock_bh(&obd->obd_processing_task_lock);
1915                 LASSERT(obd->obd_recovering);
1916                 /* usually due to recovery abort */
1917                 if (!req->rq_export->exp_in_recovery) {
1918                         spin_unlock_bh(&obd->obd_processing_task_lock);
1919                         ptlrpc_free_clone(req);
1920                         RETURN(-ENOTCONN);
1921                 }
1922                 LASSERT(req->rq_export->exp_lock_replay_needed);
1923                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1924                 spin_unlock_bh(&obd->obd_processing_task_lock);
1925                 wake_up(&obd->obd_next_transno_waitq);
1926                 RETURN(0);
1927         }
1928
1929         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1930          * (i.e. buflens etc are in my own byte order), but type-dependent
1931          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1932
1933         if (!transno) {
1934                 CFS_INIT_LIST_HEAD(&req->rq_list);
1935                 DEBUG_REQ(D_HA, req, "not queueing");
1936                 RETURN(1);
1937         }
1938
1939         spin_lock_bh(&obd->obd_processing_task_lock);
1940
1941         /* If we're processing the queue, we want don't want to queue this
1942          * message.
1943          *
1944          * Also, if this request has a transno less than the one we're waiting
1945          * for, we should process it now.  It could (and currently always will)
1946          * be an open request for a descriptor that was opened some time ago.
1947          *
1948          * Also, a resent, replayed request that has already been
1949          * handled will pass through here and be processed immediately.
1950          */
1951         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
1952               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
1953         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
1954                 /* Processing the queue right now, don't re-add. */
1955                 LASSERT(list_empty(&req->rq_list));
1956                 spin_unlock_bh(&obd->obd_processing_task_lock);
1957                 RETURN(1);
1958         }
1959         spin_unlock_bh(&obd->obd_processing_task_lock);
1960
1961         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
1962                 RETURN(0);
1963
1964         req = ptlrpc_clone_req(req);
1965         if (req == NULL)
1966                 RETURN(-ENOMEM);
1967
1968         spin_lock_bh(&obd->obd_processing_task_lock);
1969         LASSERT(obd->obd_recovering);
1970         if (!req->rq_export->exp_in_recovery) {
1971                 spin_unlock_bh(&obd->obd_processing_task_lock);
1972                 ptlrpc_free_clone(req);
1973                 RETURN(-ENOTCONN);
1974         }
1975         LASSERT(req->rq_export->exp_req_replay_needed);
1976
1977         if (target_exp_enqueue_req_replay(req)) {
1978                 spin_unlock_bh(&obd->obd_processing_task_lock);
1979                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
1980                 ptlrpc_free_clone(req);
1981                 RETURN(0);
1982         }
1983
1984         /* XXX O(n^2) */
1985         list_for_each(tmp, &obd->obd_req_replay_queue) {
1986                 struct ptlrpc_request *reqiter =
1987                         list_entry(tmp, struct ptlrpc_request, rq_list);
1988
1989                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1990                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1991                         inserted = 1;
1992                         break;
1993                 }
1994
1995                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
1996                              transno)) {
1997                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
1998                                   "has been claimed by another client");
1999                         spin_unlock_bh(&obd->obd_processing_task_lock);
2000                         target_exp_dequeue_req_replay(req);
2001                         ptlrpc_free_clone(req);
2002                         RETURN(0);
2003                 }
2004         }
2005
2006         if (!inserted)
2007                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2008
2009         obd->obd_requests_queued_for_recovery++;
2010         wake_up(&obd->obd_next_transno_waitq);
2011         spin_unlock_bh(&obd->obd_processing_task_lock);
2012         RETURN(0);
2013 }
2014
2015 struct obd_device * target_req2obd(struct ptlrpc_request *req)
2016 {
2017         return req->rq_export->exp_obd;
2018 }
2019
2020 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
2021 {
2022         LASSERT(exp != NULL);
2023         return &exp->exp_obd->obd_namespace->ns_pool;
2024 }
2025
2026 /**
2027  * Packs current SLV and Limit into \a req.
2028  */
2029 int target_pack_pool_reply(struct ptlrpc_request *req)
2030 {
2031         struct obd_device *obd;
2032         ENTRY;
2033
2034         /*
2035          * Check that we still have all structures alive as this may
2036          * be some late rpc in shutdown time.
2037          */
2038         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2039                      !exp_connect_lru_resize(req->rq_export))) {
2040                 lustre_msg_set_slv(req->rq_repmsg, 0);
2041                 lustre_msg_set_limit(req->rq_repmsg, 0);
2042                 RETURN(0);
2043         }
2044
2045         /*
2046          * OBD is alive here as export is alive, which we checked above.
2047          */
2048         obd = req->rq_export->exp_obd;
2049
2050         read_lock(&obd->obd_pool_lock);
2051         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2052         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2053         read_unlock(&obd->obd_pool_lock);
2054
2055         RETURN(0);
2056 }
2057
2058 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2059 {
2060         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2061                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2062                 return (-ECOMM);
2063         }
2064
2065         if (unlikely(rc)) {
2066                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2067                 req->rq_status = rc;
2068                 return (ptlrpc_send_error(req, 1));
2069         } else {
2070                 DEBUG_REQ(D_NET, req, "sending reply");
2071         }
2072
2073         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2074 }
2075
2076 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2077 {
2078         int                        netrc;
2079         struct ptlrpc_reply_state *rs;
2080         struct obd_device         *obd;
2081         struct obd_export         *exp;
2082         struct ptlrpc_service     *svc;
2083         ENTRY;
2084
2085         if (req->rq_no_reply) {
2086                 EXIT;
2087                 return;
2088         }
2089
2090         svc = req->rq_rqbd->rqbd_service;
2091         rs = req->rq_reply_state;
2092         if (rs == NULL || !rs->rs_difficult) {
2093                 /* no notifiers */
2094                 target_send_reply_msg (req, rc, fail_id);
2095                 EXIT;
2096                 return;
2097         }
2098
2099         /* must be an export if locks saved */
2100         LASSERT (req->rq_export != NULL);
2101         /* req/reply consistent */
2102         LASSERT (rs->rs_service == svc);
2103
2104         /* "fresh" reply */
2105         LASSERT (!rs->rs_scheduled);
2106         LASSERT (!rs->rs_scheduled_ever);
2107         LASSERT (!rs->rs_handled);
2108         LASSERT (!rs->rs_on_net);
2109         LASSERT (rs->rs_export == NULL);
2110         LASSERT (list_empty(&rs->rs_obd_list));
2111         LASSERT (list_empty(&rs->rs_exp_list));
2112
2113         exp = class_export_get (req->rq_export);
2114         obd = exp->exp_obd;
2115
2116         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
2117         rs->rs_scheduled = 1;
2118         rs->rs_on_net    = 1;
2119         rs->rs_xid       = req->rq_xid;
2120         rs->rs_transno   = req->rq_transno;
2121         rs->rs_export    = exp;
2122
2123         spin_lock(&obd->obd_uncommitted_replies_lock);
2124
2125         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2126                rs->rs_transno, obd->obd_last_committed);
2127         if (rs->rs_transno > obd->obd_last_committed) {
2128                 /* not committed already */
2129                 list_add_tail (&rs->rs_obd_list,
2130                                &obd->obd_uncommitted_replies);
2131         }
2132
2133         spin_unlock (&obd->obd_uncommitted_replies_lock);
2134         spin_lock (&exp->exp_lock);
2135
2136         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
2137
2138         spin_unlock(&exp->exp_lock);
2139
2140         netrc = target_send_reply_msg (req, rc, fail_id);
2141
2142         spin_lock(&svc->srv_lock);
2143
2144         svc->srv_n_difficult_replies++;
2145
2146         if (netrc != 0) {
2147                 /* error sending: reply is off the net.  Also we need +1
2148                  * reply ref until ptlrpc_server_handle_reply() is done
2149                  * with the reply state (if the send was successful, there
2150                  * would have been +1 ref for the net, which
2151                  * reply_out_callback leaves alone) */
2152                 rs->rs_on_net = 0;
2153                 ptlrpc_rs_addref(rs);
2154                 atomic_inc (&svc->srv_outstanding_replies);
2155         }
2156
2157         if (rs->rs_transno <= obd->obd_last_committed ||
2158             (!rs->rs_on_net && !rs->rs_no_ack) ||
2159              list_empty(&rs->rs_exp_list) ||     /* completed already */
2160              list_empty(&rs->rs_obd_list)) {
2161                 CDEBUG(D_HA, "Schedule reply immediately\n");
2162                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
2163                 cfs_waitq_signal (&svc->srv_waitq);
2164         } else {
2165                 list_add (&rs->rs_list, &svc->srv_active_replies);
2166                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2167         }
2168
2169         spin_unlock(&svc->srv_lock);
2170         EXIT;
2171 }
2172
2173 int target_handle_ping(struct ptlrpc_request *req)
2174 {
2175         obd_ping(req->rq_export);
2176         return req_capsule_server_pack(&req->rq_pill);
2177 }
2178
2179 void target_committed_to_req(struct ptlrpc_request *req)
2180 {
2181         struct obd_device *obd;
2182
2183         if (req == NULL || req->rq_export == NULL)
2184                 return;
2185
2186         obd = req->rq_export->exp_obd;
2187         if (obd == NULL)
2188                 return;
2189
2190         if (!obd->obd_no_transno && req->rq_repmsg != NULL)
2191                 lustre_msg_set_last_committed(req->rq_repmsg,
2192                                               obd->obd_last_committed);
2193         else
2194                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2195                           "%d)", obd->obd_no_transno, req->rq_repmsg == NULL);
2196
2197         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2198                obd->obd_last_committed, req->rq_transno, req->rq_xid);
2199 }
2200
2201 EXPORT_SYMBOL(target_committed_to_req);
2202
2203 int target_handle_qc_callback(struct ptlrpc_request *req)
2204 {
2205         struct obd_quotactl *oqctl;
2206         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2207
2208         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2209         if (oqctl == NULL) {
2210                 CERROR("Can't unpack obd_quotactl\n");
2211                 RETURN(-EPROTO);
2212         }
2213
2214         cli->cl_qchk_stat = oqctl->qc_stat;
2215
2216         return 0;
2217 }
2218
2219 #ifdef HAVE_QUOTA_SUPPORT
2220 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2221 {
2222 #ifdef __KERNEL__
2223         struct obd_device *obd = req->rq_export->exp_obd;
2224         struct obd_device *master_obd;
2225         struct obd_device_target *obt;
2226         struct lustre_quota_ctxt *qctxt;
2227         struct qunit_data *qdata = NULL;
2228         int rc = 0;
2229         ENTRY;
2230
2231         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2232                 RETURN(rc);
2233
2234         rc = req_capsule_server_pack(&req->rq_pill);
2235         if (rc) {
2236                 CERROR("packing reply failed!: rc = %d\n", rc);
2237                 RETURN(rc);
2238         }
2239
2240         LASSERT(req->rq_export);
2241
2242         OBD_ALLOC(qdata, sizeof(struct qunit_data));
2243         if (!qdata)
2244                 RETURN(-ENOMEM);
2245         rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2246         if (rc < 0) {
2247                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2248                 GOTO(out, rc);
2249         }
2250
2251         /* we use the observer */
2252         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2253                 CERROR("Can't find the observer, it is recovering\n");
2254                 req->rq_status = -EAGAIN;
2255                 GOTO(send_reply, rc = -EAGAIN);
2256         }
2257
2258         master_obd = obd->obd_observer->obd_observer;
2259         obt = &master_obd->u.obt;
2260         qctxt = &obt->obt_qctxt;
2261
2262         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2263                 /* quota_type has not been processed yet, return EAGAIN
2264                  * until we know whether or not quotas are supposed to
2265                  * be enabled */
2266                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2267                        "-EAGAIN\n");
2268                 req->rq_status = -EAGAIN;
2269                 rc = ptlrpc_reply(req);
2270                 GOTO(out, rc);
2271         }
2272
2273         down_read(&obt->obt_rwsem);
2274         if (qctxt->lqc_lqs_hash == NULL) {
2275                 up_read(&obt->obt_rwsem);
2276                 /* quota_type has not been processed yet, return EAGAIN
2277                  * until we know whether or not quotas are supposed to
2278                  * be enabled */
2279                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2280                        "-EAGAIN\n");
2281                 req->rq_status = -EAGAIN;
2282                 rc = ptlrpc_reply(req);
2283                 GOTO(out, rc);
2284         }
2285
2286         LASSERT(qctxt->lqc_handler);
2287         rc = qctxt->lqc_handler(master_obd, qdata,
2288                                 lustre_msg_get_opc(req->rq_reqmsg));
2289         up_read(&obt->obt_rwsem);
2290         if (rc && rc != -EDQUOT)
2291                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2292                        "dqacq failed! (rc:%d)\n", rc);
2293         req->rq_status = rc;
2294
2295         /* there are three forms of qunit(historic causes), so we need to
2296          * adjust the same form to different forms slaves needed */
2297         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2298         if (rc < 0) {
2299                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2300                 GOTO(out, rc);
2301         }
2302
2303         /* Block the quota req. b=14840 */
2304         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2305 send_reply:
2306         rc = ptlrpc_reply(req);
2307 out:
2308         OBD_FREE(qdata, sizeof(struct qunit_data));
2309         RETURN(rc);
2310 #else
2311         return 0;
2312 #endif /* !__KERNEL__ */
2313 }
2314 #endif /* HAVE_QUOTA_SUPPORT */
2315
2316 ldlm_mode_t lck_compat_array[] = {
2317         [LCK_EX] LCK_COMPAT_EX,
2318         [LCK_PW] LCK_COMPAT_PW,
2319         [LCK_PR] LCK_COMPAT_PR,
2320         [LCK_CW] LCK_COMPAT_CW,
2321         [LCK_CR] LCK_COMPAT_CR,
2322         [LCK_NL] LCK_COMPAT_NL,
2323         [LCK_GROUP] LCK_COMPAT_GROUP,
2324         [LCK_COS] LCK_COMPAT_COS,
2325 };
2326
2327 /**
2328  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2329  * not escape to the user level.
2330  */
2331 int ldlm_error2errno(ldlm_error_t error)
2332 {
2333         int result;
2334
2335         switch (error) {
2336         case ELDLM_OK:
2337                 result = 0;
2338                 break;
2339         case ELDLM_LOCK_CHANGED:
2340                 result = -ESTALE;
2341                 break;
2342         case ELDLM_LOCK_ABORTED:
2343                 result = -ENAVAIL;
2344                 break;
2345         case ELDLM_LOCK_REPLACED:
2346                 result = -ESRCH;
2347                 break;
2348         case ELDLM_NO_LOCK_DATA:
2349                 result = -ENOENT;
2350                 break;
2351         case ELDLM_NAMESPACE_EXISTS:
2352                 result = -EEXIST;
2353                 break;
2354         case ELDLM_BAD_NAMESPACE:
2355                 result = -EBADF;
2356                 break;
2357         default:
2358                 if (((int)error) < 0)  /* cast to signed type */
2359                         result = error; /* as ldlm_error_t can be unsigned */
2360                 else {
2361                         CERROR("Invalid DLM result code: %i\n", error);
2362                         result = -EPROTO;
2363                 }
2364         }
2365         return result;
2366 }
2367 EXPORT_SYMBOL(ldlm_error2errno);
2368
2369 /**
2370  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2371  */
2372 ldlm_error_t ldlm_errno2error(int err_no)
2373 {
2374         int error;
2375
2376         switch (err_no) {
2377         case 0:
2378                 error = ELDLM_OK;
2379                 break;
2380         case -ESTALE:
2381                 error = ELDLM_LOCK_CHANGED;
2382                 break;
2383         case -ENAVAIL:
2384                 error = ELDLM_LOCK_ABORTED;
2385                 break;
2386         case -ESRCH:
2387                 error = ELDLM_LOCK_REPLACED;
2388                 break;
2389         case -ENOENT:
2390                 error = ELDLM_NO_LOCK_DATA;
2391                 break;
2392         case -EEXIST:
2393                 error = ELDLM_NAMESPACE_EXISTS;
2394                 break;
2395         case -EBADF:
2396                 error = ELDLM_BAD_NAMESPACE;
2397                 break;
2398         default:
2399                 error = err_no;
2400         }
2401         return error;
2402 }
2403 EXPORT_SYMBOL(ldlm_errno2error);
2404