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