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 (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  31  *
  32  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  33  */
  34 /*
  35  * This file is part of Lustre, http://www.lustre.org/
  36  * Lustre is a trademark of Sun Microsystems, Inc.
  37  */
  38
  39 #ifndef EXPORT_SYMTAB
  40 # define EXPORT_SYMTAB
  41 #endif
  42 #define DEBUG_SUBSYSTEM S_LDLM
  43
  44 #ifdef __KERNEL__
  45 # include <libcfs/libcfs.h>
  46 #else
  47 # include <liblustre.h>
  48 #endif
  49 #include <obd.h>
  50 #include <lustre_mds.h>
  51 #include <lustre_dlm.h>
  52 #include <lustre_net.h>
  53 #include <lustre_sec.h>
  54 #include "ldlm_internal.h"
  55
  56 /* @priority: if non-zero, move the selected to the list head
  57  * @create: if zero, only search in existed connections
  58  */
  59 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  60                            int priority, int create)
  61 {
  62         struct ptlrpc_connection *ptlrpc_conn;
  63         struct obd_import_conn *imp_conn = NULL, *item;
  64         int rc = 0;
  65         ENTRY;
  66
  67         if (!create && !priority) {
  68                 CDEBUG(D_HA, "Nothing to do\n");
  69                 RETURN(-EINVAL);
  70         }
  71
  72         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  73         if (!ptlrpc_conn) {
  74                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  75                 RETURN (-ENOENT);
  76         }
  77
  78         if (create) {
  79                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  80                 if (!imp_conn) {
  81                         GOTO(out_put, rc = -ENOMEM);
  82                 }
  83         }
  84
  85         cfs_spin_lock(&imp->imp_lock);
  86         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  87                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  88                         if (priority) {
  89                                 cfs_list_del(&item->oic_item);
  90                                 cfs_list_add(&item->oic_item,
  91                                              &imp->imp_conn_list);
  92                                 item->oic_last_attempt = 0;
  93                         }
  94                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  95                                imp, imp->imp_obd->obd_name, uuid->uuid,
  96                                (priority ? ", moved to head" : ""));
  97                         cfs_spin_unlock(&imp->imp_lock);
  98                         GOTO(out_free, rc = 0);
  99                 }
 100         }
 101         /* not found */
 102         if (create) {
 103                 imp_conn->oic_conn = ptlrpc_conn;
 104                 imp_conn->oic_uuid = *uuid;
 105                 imp_conn->oic_last_attempt = 0;
 106                 if (priority)
 107                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 108                 else
 109                         cfs_list_add_tail(&imp_conn->oic_item,
 110                                           &imp->imp_conn_list);
 111                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 112                        imp, imp->imp_obd->obd_name, uuid->uuid,
 113                        (priority ? "head" : "tail"));
 114         } else {
 115                 cfs_spin_unlock(&imp->imp_lock);
 116                 GOTO(out_free, rc = -ENOENT);
 117         }
 118
 119         cfs_spin_unlock(&imp->imp_lock);
 120         RETURN(0);
 121 out_free:
 122         if (imp_conn)
 123                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 124 out_put:
 125         ptlrpc_connection_put(ptlrpc_conn);
 126         RETURN(rc);
 127 }
 128
 129 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 130 {
 131         return import_set_conn(imp, uuid, 1, 0);
 132 }
 133
 134 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 135                            int priority)
 136 {
 137         return import_set_conn(imp, uuid, priority, 1);
 138 }
 139
 140 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 141 {
 142         struct obd_import_conn *imp_conn;
 143         struct obd_export *dlmexp;
 144         int rc = -ENOENT;
 145         ENTRY;
 146
 147         cfs_spin_lock(&imp->imp_lock);
 148         if (cfs_list_empty(&imp->imp_conn_list)) {
 149                 LASSERT(!imp->imp_connection);
 150                 GOTO(out, rc);
 151         }
 152
 153         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 154                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 155                         continue;
 156                 LASSERT(imp_conn->oic_conn);
 157
 158                 /* is current conn? */
 159                 if (imp_conn == imp->imp_conn_current) {
 160                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 161
 162                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 163                             imp->imp_state != LUSTRE_IMP_DISCON) {
 164                                 CERROR("can't remove current connection\n");
 165                                 GOTO(out, rc = -EBUSY);
 166                         }
 167
 168                         ptlrpc_connection_put(imp->imp_connection);
 169                         imp->imp_connection = NULL;
 170
 171                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 172                         if (dlmexp && dlmexp->exp_connection) {
 173                                 LASSERT(dlmexp->exp_connection ==
 174                                         imp_conn->oic_conn);
 175                                 ptlrpc_connection_put(dlmexp->exp_connection);
 176                                 dlmexp->exp_connection = NULL;
 177                         }
 178                 }
 179
 180                 cfs_list_del(&imp_conn->oic_item);
 181                 ptlrpc_connection_put(imp_conn->oic_conn);
 182                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 183                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 184                        imp, imp->imp_obd->obd_name, uuid->uuid);
 185                 rc = 0;
 186                 break;
 187         }
 188 out:
 189         cfs_spin_unlock(&imp->imp_lock);
 190         if (rc == -ENOENT)
 191                 CERROR("connection %s not found\n", uuid->uuid);
 192         RETURN(rc);
 193 }
 194
 195 /**
 196  * Find conn uuid by peer nid. @peer is a server nid. This function is used
 197  * to find a conn uuid of @imp which can reach @peer.
 198  */
 199 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 200                             struct obd_uuid *uuid)
 201 {
 202         struct obd_import_conn *conn;
 203         int rc = -ENOENT;
 204         ENTRY;
 205
 206         cfs_spin_lock(&imp->imp_lock);
 207         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 208                 /* check if conn uuid does have this peer nid */
 209                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 210                         *uuid = conn->oic_uuid;
 211                         rc = 0;
 212                         break;
 213                 }
 214         }
 215         cfs_spin_unlock(&imp->imp_lock);
 216         RETURN(rc);
 217 }
 218 EXPORT_SYMBOL(client_import_find_conn);
 219
 220 void client_destroy_import(struct obd_import *imp)
 221 {
 222         /* drop security policy instance after all rpc finished/aborted
 223          * to let all busy contexts be released. */
 224         class_import_get(imp);
 225         class_destroy_import(imp);
 226         sptlrpc_import_sec_put(imp);
 227         class_import_put(imp);
 228 }
 229 EXPORT_SYMBOL(client_destroy_import);
 230
 231 /* configure an RPC client OBD device
 232  *
 233  * lcfg parameters:
 234  * 1 - client UUID
 235  * 2 - server UUID
 236  * 3 - inactive-on-startup
 237  */
 238 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 239 {
 240         struct client_obd *cli = &obddev->u.cli;
 241         struct obd_import *imp;
 242         struct obd_uuid server_uuid;
 243         int rq_portal, rp_portal, connect_op;
 244         char *name = obddev->obd_type->typ_name;
 245         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 246         int rc;
 247         ENTRY;
 248
 249         /* In a more perfect world, we would hang a ptlrpc_client off of
 250          * obd_type and just use the values from there. */
 251         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 252                 rq_portal = OST_REQUEST_PORTAL;
 253                 rp_portal = OSC_REPLY_PORTAL;
 254                 connect_op = OST_CONNECT;
 255                 cli->cl_sp_me = LUSTRE_SP_CLI;
 256                 cli->cl_sp_to = LUSTRE_SP_OST;
 257                 ns_type = LDLM_NS_TYPE_OSC;
 258
 259         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 260                 rq_portal = MDS_REQUEST_PORTAL;
 261                 rp_portal = MDC_REPLY_PORTAL;
 262                 connect_op = MDS_CONNECT;
 263                 cli->cl_sp_me = LUSTRE_SP_CLI;
 264                 cli->cl_sp_to = LUSTRE_SP_MDT;
 265                 ns_type = LDLM_NS_TYPE_MDC;
 266
 267         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 268                 rq_portal = MGS_REQUEST_PORTAL;
 269                 rp_portal = MGC_REPLY_PORTAL;
 270                 connect_op = MGS_CONNECT;
 271                 cli->cl_sp_me = LUSTRE_SP_MGC;
 272                 cli->cl_sp_to = LUSTRE_SP_MGS;
 273                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 274                 ns_type = LDLM_NS_TYPE_MGC;
 275
 276         } else {
 277                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 278                        name);
 279                 RETURN(-EINVAL);
 280         }
 281
 282         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 283                 CERROR("requires a TARGET UUID\n");
 284                 RETURN(-EINVAL);
 285         }
 286
 287         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 288                 CERROR("client UUID must be less than 38 characters\n");
 289                 RETURN(-EINVAL);
 290         }
 291
 292         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 293                 CERROR("setup requires a SERVER UUID\n");
 294                 RETURN(-EINVAL);
 295         }
 296
 297         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 298                 CERROR("target UUID must be less than 38 characters\n");
 299                 RETURN(-EINVAL);
 300         }
 301
 302         cfs_init_rwsem(&cli->cl_sem);
 303         cfs_sema_init(&cli->cl_mgc_sem, 1);
 304         cli->cl_conn_count = 0;
 305         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 306                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 307                      sizeof(server_uuid)));
 308
 309         cli->cl_dirty = 0;
 310         cli->cl_avail_grant = 0;
 311         /* FIXME: should limit this for the sum of all cl_dirty_max */
 312         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 313         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
 314                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
 315         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 316         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 317         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 318         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 319         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 320         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 321         cli->cl_r_in_flight = 0;
 322         cli->cl_w_in_flight = 0;
 323
 324         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 325         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 326         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 327         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 328         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 329         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 330         cfs_waitq_init(&cli->cl_destroy_waitq);
 331         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 332 #ifdef ENABLE_CHECKSUM
 333         /* Turn on checksumming by default. */
 334         cli->cl_checksum = 1;
 335         /*
 336          * The supported checksum types will be worked out at connect time
 337          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 338          * through procfs.
 339          */
 340         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 341 #endif
 342         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 343
 344         /* This value may be changed at connect time in
 345            ptlrpc_connect_interpret. */
 346         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 347                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 348
 349         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 350                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 351         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 352                 cli->cl_max_rpcs_in_flight = 2;
 353         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 354                 cli->cl_max_rpcs_in_flight = 3;
 355         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 356                 cli->cl_max_rpcs_in_flight = 4;
 357         } else {
 358                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 359         }
 360
 361         rc = ldlm_get_ref();
 362         if (rc) {
 363                 CERROR("ldlm_get_ref failed: %d\n", rc);
 364                 GOTO(err, rc);
 365         }
 366
 367         ptlrpc_init_client(rq_portal, rp_portal, name,
 368                            &obddev->obd_ldlm_client);
 369
 370         imp = class_new_import(obddev);
 371         if (imp == NULL)
 372                 GOTO(err_ldlm, rc = -ENOENT);
 373         imp->imp_client = &obddev->obd_ldlm_client;
 374         imp->imp_connect_op = connect_op;
 375         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 376         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 377                LUSTRE_CFG_BUFLEN(lcfg, 1));
 378         class_import_put(imp);
 379
 380         rc = client_import_add_conn(imp, &server_uuid, 1);
 381         if (rc) {
 382                 CERROR("can't add initial connection\n");
 383                 GOTO(err_import, rc);
 384         }
 385
 386         cli->cl_import = imp;
 387         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 388         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 389         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 390
 391         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 392                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 393                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 394                                name, obddev->obd_name,
 395                                cli->cl_target_uuid.uuid);
 396                         cfs_spin_lock(&imp->imp_lock);
 397                         imp->imp_deactive = 1;
 398                         cfs_spin_unlock(&imp->imp_lock);
 399                 }
 400         }
 401
 402         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 403                                                    LDLM_NAMESPACE_CLIENT,
 404                                                    LDLM_NAMESPACE_GREEDY,
 405                                                    ns_type);
 406         if (obddev->obd_namespace == NULL) {
 407                 CERROR("Unable to create client namespace - %s\n",
 408                        obddev->obd_name);
 409                 GOTO(err_import, rc = -ENOMEM);
 410         }
 411
 412         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 413
 414         RETURN(rc);
 415
 416 err_import:
 417         class_destroy_import(imp);
 418 err_ldlm:
 419         ldlm_put_ref();
 420 err:
 421         RETURN(rc);
 422
 423 }
 424
 425 int client_obd_cleanup(struct obd_device *obddev)
 426 {
 427         ENTRY;
 428
 429         ldlm_namespace_free_post(obddev->obd_namespace);
 430         obddev->obd_namespace = NULL;
 431
 432         LASSERT(obddev->u.cli.cl_import == NULL);
 433
 434         ldlm_put_ref();
 435         RETURN(0);
 436 }
 437
 438 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 439 int client_connect_import(const struct lu_env *env,
 440                           struct obd_export **exp,
 441                           struct obd_device *obd, struct obd_uuid *cluuid,
 442                           struct obd_connect_data *data, void *localdata)
 443 {
 444         struct client_obd *cli = &obd->u.cli;
 445         struct obd_import *imp = cli->cl_import;
 446         struct obd_connect_data *ocd;
 447         struct lustre_handle conn = { 0 };
 448         int rc;
 449         ENTRY;
 450
 451         *exp = NULL;
 452         cfs_down_write(&cli->cl_sem);
 453         if (cli->cl_conn_count > 0 )
 454                 GOTO(out_sem, rc = -EALREADY);
 455
 456         rc = class_connect(&conn, obd, cluuid);
 457         if (rc)
 458                 GOTO(out_sem, rc);
 459
 460         cli->cl_conn_count++;
 461         *exp = class_conn2export(&conn);
 462
 463         LASSERT(obd->obd_namespace);
 464
 465         imp->imp_dlm_handle = conn;
 466         rc = ptlrpc_init_import(imp);
 467         if (rc != 0)
 468                 GOTO(out_ldlm, rc);
 469
 470         ocd = &imp->imp_connect_data;
 471         if (data) {
 472                 *ocd = *data;
 473                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 474         }
 475
 476         rc = ptlrpc_connect_import(imp);
 477         if (rc != 0) {
 478                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 479                 GOTO(out_ldlm, rc);
 480         }
 481         LASSERT((*exp)->exp_connection);
 482
 483         if (data) {
 484                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 485                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 486                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 487                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 488         }
 489
 490         ptlrpc_pinger_add_import(imp);
 491
 492         EXIT;
 493
 494         if (rc) {
 495 out_ldlm:
 496                 cli->cl_conn_count--;
 497                 class_disconnect(*exp);
 498                 *exp = NULL;
 499         }
 500 out_sem:
 501         cfs_up_write(&cli->cl_sem);
 502
 503         return rc;
 504 }
 505
 506 int client_disconnect_export(struct obd_export *exp)
 507 {
 508         struct obd_device *obd = class_exp2obd(exp);
 509         struct client_obd *cli;
 510         struct obd_import *imp;
 511         int rc = 0, err;
 512         ENTRY;
 513
 514         if (!obd) {
 515                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 516                        exp, exp ? exp->exp_handle.h_cookie : -1);
 517                 RETURN(-EINVAL);
 518         }
 519
 520         cli = &obd->u.cli;
 521         imp = cli->cl_import;
 522
 523         cfs_down_write(&cli->cl_sem);
 524         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 525                cli->cl_conn_count);
 526
 527         if (!cli->cl_conn_count) {
 528                 CERROR("disconnecting disconnected device (%s)\n",
 529                        obd->obd_name);
 530                 GOTO(out_disconnect, rc = -EINVAL);
 531         }
 532
 533         cli->cl_conn_count--;
 534         if (cli->cl_conn_count)
 535                 GOTO(out_disconnect, rc = 0);
 536
 537         /* Mark import deactivated now, so we don't try to reconnect if any
 538          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 539          * fully deactivate the import, or that would drop all requests. */
 540         cfs_spin_lock(&imp->imp_lock);
 541         imp->imp_deactive = 1;
 542         cfs_spin_unlock(&imp->imp_lock);
 543
 544         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 545          * delete it regardless.  (It's safe to delete an import that was
 546          * never added.) */
 547         (void)ptlrpc_pinger_del_import(imp);
 548
 549         if (obd->obd_namespace != NULL) {
 550                 /* obd_force == local only */
 551                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 552                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 553                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 554         }
 555
 556         /*
 557          * there's no need to hold sem during disconnecting an import,
 558          * and actually it may cause deadlock in gss.
 559          */
 560         cfs_up_write(&cli->cl_sem);
 561         rc = ptlrpc_disconnect_import(imp, 0);
 562         cfs_down_write(&cli->cl_sem);
 563
 564         ptlrpc_invalidate_import(imp);
 565
 566         EXIT;
 567
 568  out_disconnect:
 569         /* use server style - class_disconnect should be always called for
 570          * o_disconnect */
 571         err = class_disconnect(exp);
 572         if (!rc && err)
 573                 rc = err;
 574
 575         cfs_up_write(&cli->cl_sem);
 576
 577         RETURN(rc);
 578 }
 579
 580 int server_disconnect_export(struct obd_export *exp)
 581 {
 582         int rc;
 583         ENTRY;
 584
 585         /* Disconnect early so that clients can't keep using export */
 586         rc = class_disconnect(exp);
 587         /* close import for avoid sending any requests */
 588         if (exp->exp_imp_reverse)
 589                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 590
 591         if (exp->exp_obd->obd_namespace != NULL)
 592                 ldlm_cancel_locks_for_export(exp);
 593
 594         /* complete all outstanding replies */
 595         cfs_spin_lock(&exp->exp_lock);
 596         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 597                 struct ptlrpc_reply_state *rs =
 598                         cfs_list_entry(exp->exp_outstanding_replies.next,
 599                                        struct ptlrpc_reply_state, rs_exp_list);
 600                 struct ptlrpc_service *svc = rs->rs_service;
 601
 602                 cfs_spin_lock(&svc->srv_rs_lock);
 603                 cfs_list_del_init(&rs->rs_exp_list);
 604                 cfs_spin_lock(&rs->rs_lock);
 605                 ptlrpc_schedule_difficult_reply(rs);
 606                 cfs_spin_unlock(&rs->rs_lock);
 607                 cfs_spin_unlock(&svc->srv_rs_lock);
 608         }
 609         cfs_spin_unlock(&exp->exp_lock);
 610
 611         RETURN(rc);
 612 }
 613
 614 /* --------------------------------------------------------------------------
 615  * from old lib/target.c
 616  * -------------------------------------------------------------------------- */
 617
 618 static int target_handle_reconnect(struct lustre_handle *conn,
 619                                    struct obd_export *exp,
 620                                    struct obd_uuid *cluuid)
 621 {
 622         ENTRY;
 623
 624         if (exp->exp_connection && exp->exp_imp_reverse) {
 625                 struct lustre_handle *hdl;
 626                 struct obd_device *target;
 627
 628                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 629                 target = exp->exp_obd;
 630
 631                 /* Might be a re-connect after a partition. */
 632                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 633                         if (target->obd_recovering) {
 634                                 int timeout = cfs_duration_sec(cfs_time_sub(
 635                                         cfs_timer_deadline(
 636                                         &target->obd_recovery_timer),
 637                                         cfs_time_current()));
 638
 639                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 640                                         "ing, waiting for %d clients in recov"
 641                                         "ery for %d:%.02d\n", target->obd_name,
 642                                         obd_uuid2str(&exp->exp_client_uuid),
 643                                         obd_export_nid2str(exp),
 644                                         target->obd_max_recoverable_clients,
 645                                         timeout / 60, timeout % 60);
 646                         } else {
 647                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 648                                         "reconnecting\n", target->obd_name,
 649                                         obd_uuid2str(&exp->exp_client_uuid),
 650                                         obd_export_nid2str(exp));
 651                         }
 652
 653                         conn->cookie = exp->exp_handle.h_cookie;
 654                         /* target_handle_connect() treats EALREADY and
 655                          * -EALREADY differently.  EALREADY means we are
 656                          * doing a valid reconnect from the same client. */
 657                         RETURN(EALREADY);
 658                 } else {
 659                         LCONSOLE_WARN("%s: The server has already connected "
 660                                       "client %s (at %s) with handle " LPX64
 661                                       ", rejecting a client with the same "
 662                                       "uuid trying to reconnect with "
 663                                       "handle " LPX64, target->obd_name,
 664                                       obd_uuid2str(&exp->exp_client_uuid),
 665                                       obd_export_nid2str(exp),
 666                                       hdl->cookie, conn->cookie);
 667                         memset(conn, 0, sizeof *conn);
 668                         /* target_handle_connect() treats EALREADY and
 669                          * -EALREADY differently.  -EALREADY is an error
 670                          * (same UUID, different handle). */
 671                         RETURN(-EALREADY);
 672                 }
 673         }
 674
 675         conn->cookie = exp->exp_handle.h_cookie;
 676         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 677                cluuid->uuid, exp, conn->cookie);
 678         RETURN(0);
 679 }
 680
 681 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 682                           int error)
 683 {
 684         struct obd_export *exp = cb_data;
 685
 686         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 687                obd->obd_name, exp->exp_client_uuid.uuid);
 688
 689         cfs_spin_lock(&exp->exp_lock);
 690         exp->exp_need_sync = 0;
 691         cfs_spin_unlock(&exp->exp_lock);
 692         class_export_cb_put(exp);
 693 }
 694 EXPORT_SYMBOL(target_client_add_cb);
 695
 696 #ifdef __KERNEL__
 697 static void
 698 check_and_start_recovery_timer(struct obd_device *obd,
 699                                struct ptlrpc_request *req, int new_client);
 700 #else
 701 static inline void
 702 check_and_start_recovery_timer(struct obd_device *obd,
 703                                struct ptlrpc_request *req, int new_client)
 704 {
 705 }
 706 #endif
 707
 708 int target_handle_connect(struct ptlrpc_request *req)
 709 {
 710         struct obd_device *target, *targref = NULL;
 711         struct obd_export *export = NULL;
 712         struct obd_import *revimp;
 713         struct lustre_handle conn;
 714         struct lustre_handle *tmp;
 715         struct obd_uuid tgtuuid;
 716         struct obd_uuid cluuid;
 717         struct obd_uuid remote_uuid;
 718         char *str;
 719         int rc = 0;
 720         char *target_start;
 721         int target_len;
 722         int mds_conn = 0;
 723         struct obd_connect_data *data, *tmpdata;
 724         int size, tmpsize;
 725         lnet_nid_t *client_nid = NULL;
 726         ENTRY;
 727
 728         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 729
 730         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 731         if (str == NULL) {
 732                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 733                 GOTO(out, rc = -EINVAL);
 734         }
 735
 736         obd_str2uuid(&tgtuuid, str);
 737         target = class_uuid2obd(&tgtuuid);
 738         if (!target)
 739                 target = class_name2obd(str);
 740
 741         if (!target || target->obd_stopping || !target->obd_set_up) {
 742                 deuuidify(str, NULL, &target_start, &target_len);
 743                 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
 744                                    "from %s (%s)\n", target_len, target_start,
 745                                    libcfs_nid2str(req->rq_peer.nid), !target ?
 746                                    "no target" : (target->obd_stopping ?
 747                                    "stopping" : "not set up"));
 748                 GOTO(out, rc = -ENODEV);
 749         }
 750
 751         if (target->obd_no_conn) {
 752                 LCONSOLE_WARN("%s: Temporarily refusing client connection "
 753                               "from %s\n", target->obd_name,
 754                               libcfs_nid2str(req->rq_peer.nid));
 755                 GOTO(out, rc = -EAGAIN);
 756         }
 757
 758         /* Make sure the target isn't cleaned up while we're here. Yes,
 759            there's still a race between the above check and our incref here.
 760            Really, class_uuid2obd should take the ref. */
 761         targref = class_incref(target, __FUNCTION__, cfs_current());
 762
 763
 764         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 765         if (str == NULL) {
 766                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 767                 GOTO(out, rc = -EINVAL);
 768         }
 769
 770         obd_str2uuid(&cluuid, str);
 771
 772         /* XXX extract a nettype and format accordingly */
 773         switch (sizeof(lnet_nid_t)) {
 774                 /* NB the casts only avoid compiler warnings */
 775         case 8:
 776                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 777                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 778                 break;
 779         case 4:
 780                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 781                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 782                 break;
 783         default:
 784                 LBUG();
 785         }
 786
 787         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 788         if (tmp == NULL)
 789                 GOTO(out, rc = -EPROTO);
 790
 791         conn = *tmp;
 792
 793         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 794                                     RCL_CLIENT);
 795         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 796         if (!data)
 797                 GOTO(out, rc = -EPROTO);
 798
 799         rc = req_capsule_server_pack(&req->rq_pill);
 800         if (rc)
 801                 GOTO(out, rc);
 802
 803         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 804                 if (!data) {
 805                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 806                                   "libclient connection attempt");
 807                         GOTO(out, rc = -EPROTO);
 808                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 809                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 810                            data->ocd_version > LUSTRE_VERSION_CODE +
 811                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 812                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 813                                   "libclient connection attempt",
 814                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 815                                   "old" : "new",
 816                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 817                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 818                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 819                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 820                         data = req_capsule_server_sized_get(&req->rq_pill,
 821                                                         &RMF_CONNECT_DATA,
 822                                     offsetof(typeof(*data), ocd_version) +
 823                                              sizeof(data->ocd_version));
 824                         if (data) {
 825                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 826                                 data->ocd_version = LUSTRE_VERSION_CODE;
 827                         }
 828                         GOTO(out, rc = -EPROTO);
 829                 }
 830         }
 831
 832         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 833             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 834                 mds_conn = 1;
 835
 836         /* lctl gets a backstage, all-access pass. */
 837         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 838                 goto dont_check_exports;
 839
 840         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 841         if (!export)
 842                 goto no_export;
 843
 844         /* we've found an export in the hash */
 845         if (export->exp_connecting) { /* bug 9635, et. al. */
 846                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 847                               export->exp_obd->obd_name, export,
 848                               libcfs_nid2str(req->rq_peer.nid));
 849                 class_export_put(export);
 850                 export = NULL;
 851                 rc = -EALREADY;
 852         } else if (mds_conn && export->exp_connection) {
 853                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 854                         /* mds reconnected after failover */
 855                         LCONSOLE_WARN("%s: Received MDS connection from "
 856                             "%s, removing former export from %s\n",
 857                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 858                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 859                 else
 860                         /* new mds connection from the same nid */
 861                         LCONSOLE_WARN("%s: Received new MDS connection from "
 862                             "%s, removing former export from same NID\n",
 863                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 864                 class_fail_export(export);
 865                 class_export_put(export);
 866                 export = NULL;
 867                 rc = 0;
 868         } else if (export->exp_connection != NULL &&
 869                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 870                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 871                     MSG_CONNECT_INITIAL)) {
 872                 /* in mds failover we have static uuid but nid can be
 873                  * changed*/
 874                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 875                               "existing nid %s is already connected\n",
 876                               target->obd_name, cluuid.uuid,
 877                               libcfs_nid2str(req->rq_peer.nid),
 878                               libcfs_nid2str(
 879                                       export->exp_connection->c_peer.nid));
 880                 rc = -EALREADY;
 881                 class_export_put(export);
 882                 export = NULL;
 883         } else {
 884                 cfs_spin_lock(&export->exp_lock);
 885                 export->exp_connecting = 1;
 886                 cfs_spin_unlock(&export->exp_lock);
 887                 LASSERT(export->exp_obd == target);
 888
 889                 rc = target_handle_reconnect(&conn, export, &cluuid);
 890         }
 891
 892         /* If we found an export, we already unlocked. */
 893         if (!export) {
 894 no_export:
 895                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 896         } else if (req->rq_export == NULL &&
 897                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 898                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 899                               "still busy with %d references\n",
 900                               target->obd_name, cluuid.uuid,
 901                               libcfs_nid2str(req->rq_peer.nid),
 902                               cfs_atomic_read(&export->exp_refcount));
 903                 GOTO(out, rc = -EBUSY);
 904         } else if (req->rq_export != NULL &&
 905                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 906                 /* the current connect rpc has increased exp_rpc_count */
 907                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
 908                               "still busy with %d active RPCs\n",
 909                               target->obd_name, cluuid.uuid,
 910                               libcfs_nid2str(req->rq_peer.nid),
 911                               cfs_atomic_read(&export->exp_rpc_count) - 1);
 912                 cfs_spin_lock(&export->exp_lock);
 913                 if (req->rq_export->exp_conn_cnt <
 914                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 915                         /* try to abort active requests */
 916                         req->rq_export->exp_abort_active_req = 1;
 917                 cfs_spin_unlock(&export->exp_lock);
 918                 GOTO(out, rc = -EBUSY);
 919         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 920                 if (!strstr(cluuid.uuid, "mdt"))
 921                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 922                                       "known client %s (at %s) because it "
 923                                       "is indicating it is a new client",
 924                                       target->obd_name, cluuid.uuid,
 925                                       libcfs_nid2str(req->rq_peer.nid));
 926                 GOTO(out, rc = -EALREADY);
 927         } else {
 928                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 929         }
 930
 931         if (rc < 0) {
 932                 GOTO(out, rc);
 933         }
 934
 935         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 936                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 937               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 938               export, (long)cfs_time_current_sec(),
 939               export ? (long)export->exp_last_request_time : 0);
 940
 941         /* If this is the first time a client connects, reset the recovery
 942          * timer */
 943         if (rc == 0 && target->obd_recovering)
 944                 check_and_start_recovery_timer(target, req, export == NULL);
 945
 946         /* We want to handle EALREADY but *not* -EALREADY from
 947          * target_handle_reconnect(), return reconnection state in a flag */
 948         if (rc == EALREADY) {
 949                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 950                 rc = 0;
 951         } else {
 952                 LASSERT(rc == 0);
 953         }
 954
 955         /* Tell the client if we support replayable requests */
 956         if (target->obd_replayable)
 957                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 958         client_nid = &req->rq_peer.nid;
 959
 960         if (export == NULL) {
 961                 if (target->obd_recovering) {
 962                         cfs_time_t t;
 963
 964                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 965                         t = cfs_time_sub(t, cfs_time_current());
 966                         t = cfs_duration_sec(t);
 967                         LCONSOLE_WARN("%s: Denying connection for new client "
 968                                       "%s (at %s), waiting for %d clients in "
 969                                       "recovery for %d:%.02d\n",
 970                                       target->obd_name,
 971                                       libcfs_nid2str(req->rq_peer.nid),
 972                                       cluuid.uuid,
 973                                       cfs_atomic_read(&target-> \
 974                                                       obd_lock_replay_clients),
 975                                       (int)t / 60, (int)t % 60);
 976                         rc = -EBUSY;
 977                 } else {
 978 dont_check_exports:
 979                         rc = obd_connect(req->rq_svc_thread->t_env,
 980                                          &export, target, &cluuid, data,
 981                                          client_nid);
 982                         if (rc == 0) {
 983                                 conn.cookie = export->exp_handle.h_cookie;
 984                                 /* LU-1092 reconnect put export refcount in the
 985                                  * end, connect needs take one here too. */
 986                                 class_export_get(export);
 987                         }
 988                 }
 989         } else {
 990                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 991                                    export, target, &cluuid, data, client_nid);
 992                 if (rc == 0)
 993                         /* prevous done via class_conn2export */
 994                         class_export_get(export);
 995         }
 996         if (rc)
 997                 GOTO(out, rc);
 998
 999         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1000         data->ocd_instance = target->u.obt.obt_instance;
1001
1002         /* Return only the parts of obd_connect_data that we understand, so the
1003          * client knows that we don't understand the rest. */
1004         if (data) {
1005                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1006                                                RCL_SERVER);
1007                 tmpdata = req_capsule_server_get(&req->rq_pill,
1008                                                  &RMF_CONNECT_DATA);
1009                 /* Don't use struct assignment here, because the client reply
1010                  * buffer may be smaller/larger than the local struct
1011                  * obd_connect_data. */
1012                 memcpy(tmpdata, data, min(tmpsize, size));
1013         }
1014
1015         /* If all else goes well, this is our RPC return code. */
1016         req->rq_status = 0;
1017
1018         lustre_msg_set_handle(req->rq_repmsg, &conn);
1019
1020         /* If the client and the server are the same node, we will already
1021          * have an export that really points to the client's DLM export,
1022          * because we have a shared handles table.
1023          *
1024          * XXX this will go away when shaver stops sending the "connect" handle
1025          * in the real "remote handle" field of the request --phik 24 Apr 2003
1026          */
1027         if (req->rq_export != NULL)
1028                 class_export_put(req->rq_export);
1029
1030         req->rq_export = export;
1031
1032         cfs_spin_lock(&export->exp_lock);
1033         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1034                 cfs_spin_unlock(&export->exp_lock);
1035                 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1036                        "conn_cnt: %d > %d\n",
1037                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1038                        export->exp_conn_cnt,
1039                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1040
1041                 GOTO(out, rc = -EALREADY);
1042         }
1043         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1044         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1045         export->exp_abort_active_req = 0;
1046
1047         /* request from liblustre?  Don't evict it for not pinging. */
1048         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1049                 export->exp_libclient = 1;
1050                 cfs_spin_unlock(&export->exp_lock);
1051
1052                 cfs_spin_lock(&target->obd_dev_lock);
1053                 cfs_list_del_init(&export->exp_obd_chain_timed);
1054                 cfs_spin_unlock(&target->obd_dev_lock);
1055         } else {
1056                 cfs_spin_unlock(&export->exp_lock);
1057         }
1058
1059         if (export->exp_connection != NULL) {
1060                 /* Check to see if connection came from another NID */
1061                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1062                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1063                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1064                                      &export->exp_connection->c_peer.nid,
1065                                      &export->exp_nid_hash);
1066
1067                 ptlrpc_connection_put(export->exp_connection);
1068         }
1069
1070         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1071                                                        req->rq_self,
1072                                                        &remote_uuid);
1073         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1074                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1075                              &export->exp_connection->c_peer.nid,
1076                              &export->exp_nid_hash);
1077         }
1078         /**
1079           class_disconnect->class_export_recovery_cleanup() race
1080          */
1081         if (target->obd_recovering && !export->exp_in_recovery) {
1082                 int has_transno;
1083                 __u64 transno = data->ocd_transno;
1084
1085                 cfs_spin_lock(&export->exp_lock);
1086                 export->exp_in_recovery = 1;
1087                 export->exp_req_replay_needed = 1;
1088                 export->exp_lock_replay_needed = 1;
1089                 cfs_spin_unlock(&export->exp_lock);
1090
1091                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1092                                  MSG_CONNECT_TRANSNO);
1093                 if (has_transno && transno == 0)
1094                         CWARN("Connect with zero transno!\n");
1095
1096                 if (has_transno && transno > 0 &&
1097                     transno < target->obd_next_recovery_transno &&
1098                     transno > target->obd_last_committed) {
1099                         /* another way is to use cmpxchg() so it will be
1100                          * lock free */
1101                         cfs_spin_lock(&target->obd_recovery_task_lock);
1102                         if (transno < target->obd_next_recovery_transno)
1103                                 target->obd_next_recovery_transno = transno;
1104                         cfs_spin_unlock(&target->obd_recovery_task_lock);
1105                 }
1106
1107                 cfs_atomic_inc(&target->obd_req_replay_clients);
1108                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1109                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1110                     target->obd_max_recoverable_clients)
1111                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1112         }
1113
1114         /* Tell the client we're in recovery, when client is involved in it. */
1115         if (target->obd_recovering)
1116                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1117
1118         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1119         conn = *tmp;
1120
1121         if (export->exp_imp_reverse != NULL) {
1122                 /* destroyed import can be still referenced in ctxt */
1123                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1124                                    KEY_REVIMP_UPD, 0, NULL, NULL);
1125
1126                 client_destroy_import(export->exp_imp_reverse);
1127         }
1128
1129         /* for the rest part, we return -ENOTCONN in case of errors
1130          * in order to let client initialize connection again.
1131          */
1132         revimp = export->exp_imp_reverse = class_new_import(target);
1133         if (!revimp) {
1134                 CERROR("fail to alloc new reverse import.\n");
1135                 GOTO(out, rc = -ENOTCONN);
1136         }
1137
1138         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1139         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1140         revimp->imp_remote_handle = conn;
1141         revimp->imp_dlm_fake = 1;
1142         revimp->imp_state = LUSTRE_IMP_FULL;
1143
1144         /* unknown versions will be caught in
1145          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1146         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1147
1148         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1149             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1150                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1151         else
1152                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1153
1154         if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1155             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1156                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1157         else
1158                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1159
1160         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1161         if (rc) {
1162                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1163                 export->exp_imp_reverse = NULL;
1164                 class_destroy_import(revimp);
1165         }
1166
1167         class_import_put(revimp);
1168 out:
1169         if (export) {
1170                 cfs_spin_lock(&export->exp_lock);
1171                 export->exp_connecting = 0;
1172                 cfs_spin_unlock(&export->exp_lock);
1173
1174                 class_export_put(export);
1175         }
1176         if (targref)
1177                 class_decref(targref, __FUNCTION__, cfs_current());
1178         if (rc)
1179                 req->rq_status = rc;
1180         RETURN(rc);
1181 }
1182
1183 int target_handle_disconnect(struct ptlrpc_request *req)
1184 {
1185         int rc;
1186         ENTRY;
1187
1188         rc = req_capsule_server_pack(&req->rq_pill);
1189         if (rc)
1190                 RETURN(rc);
1191
1192         /* keep the rq_export around so we can send the reply */
1193         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1194
1195         RETURN(0);
1196 }
1197
1198 void target_destroy_export(struct obd_export *exp)
1199 {
1200         /* exports created from last_rcvd data, and "fake"
1201            exports created by lctl don't have an import */
1202         if (exp->exp_imp_reverse != NULL)
1203                 client_destroy_import(exp->exp_imp_reverse);
1204
1205         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1206         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1207         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1208         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1209 }
1210
1211 /*
1212  * Recovery functions
1213  */
1214 static void target_request_copy_get(struct ptlrpc_request *req)
1215 {
1216         class_export_rpc_get(req->rq_export);
1217         LASSERT(cfs_list_empty(&req->rq_list));
1218         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1219
1220         /* increase refcount to keep request in queue */
1221         cfs_atomic_inc(&req->rq_refcount);
1222         /** let export know it has replays to be handled */
1223         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1224 }
1225
1226 static void target_request_copy_put(struct ptlrpc_request *req)
1227 {
1228         LASSERT(cfs_list_empty(&req->rq_replay_list));
1229         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1230
1231         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1232         class_export_rpc_put(req->rq_export);
1233         ptlrpc_server_drop_request(req);
1234 }
1235
1236 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1237 {
1238         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1239         struct obd_export     *exp = req->rq_export;
1240         struct ptlrpc_request *reqiter;
1241         int                    dup = 0;
1242
1243         LASSERT(exp);
1244
1245         cfs_spin_lock(&exp->exp_lock);
1246         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1247                                 rq_replay_list) {
1248                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1249                         dup = 1;
1250                         break;
1251                 }
1252         }
1253
1254         if (dup) {
1255                 /* we expect it with RESENT and REPLAY flags */
1256                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1257                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1258                         CERROR("invalid flags %x of resent replay\n",
1259                                lustre_msg_get_flags(req->rq_reqmsg));
1260         } else {
1261                 cfs_list_add_tail(&req->rq_replay_list,
1262                                   &exp->exp_req_replay_queue);
1263         }
1264
1265         cfs_spin_unlock(&exp->exp_lock);
1266         return dup;
1267 }
1268
1269 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1270 {
1271         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1272         LASSERT(req->rq_export);
1273
1274         cfs_spin_lock(&req->rq_export->exp_lock);
1275         cfs_list_del_init(&req->rq_replay_list);
1276         cfs_spin_unlock(&req->rq_export->exp_lock);
1277 }
1278
1279 #ifdef __KERNEL__
1280 static void target_finish_recovery(struct obd_device *obd)
1281 {
1282         time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1283                                     obd->obd_recovery_start);
1284         ENTRY;
1285
1286         LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1287                       "%d recovered and %d %s evicted.\n", obd->obd_name,
1288                       (int)elapsed_time / 60, (int)elapsed_time % 60,
1289                       obd->obd_max_recoverable_clients,
1290                       cfs_atomic_read(&obd->obd_connected_clients),
1291                       obd->obd_stale_clients,
1292                       obd->obd_stale_clients == 1 ? "was" : "were");
1293
1294         ldlm_reprocess_all_ns(obd->obd_namespace);
1295         cfs_spin_lock(&obd->obd_recovery_task_lock);
1296         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1297             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1298             !cfs_list_empty(&obd->obd_final_req_queue)) {
1299                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1300                        obd->obd_name,
1301                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1302                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1303                                "" : "lock ",
1304                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1305                                "" : "final ");
1306                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1307                 LBUG();
1308         }
1309         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1310
1311         obd->obd_recovery_end = cfs_time_current_sec();
1312
1313         /* when recovery finished, cleanup orphans on mds and ost */
1314         if (OBT(obd) && OBP(obd, postrecov)) {
1315                 int rc = OBP(obd, postrecov)(obd);
1316                 if (rc < 0)
1317                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1318                                       obd->obd_name, rc);
1319         }
1320         EXIT;
1321 }
1322
1323 static void abort_req_replay_queue(struct obd_device *obd)
1324 {
1325         struct ptlrpc_request *req, *n;
1326         cfs_list_t abort_list;
1327
1328         CFS_INIT_LIST_HEAD(&abort_list);
1329         cfs_spin_lock(&obd->obd_recovery_task_lock);
1330         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1331         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1332         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1333                 DEBUG_REQ(D_WARNING, req, "aborted:");
1334                 req->rq_status = -ENOTCONN;
1335                 if (ptlrpc_error(req)) {
1336                         DEBUG_REQ(D_ERROR, req,
1337                                   "failed abort_req_reply; skipping");
1338                 }
1339                 target_exp_dequeue_req_replay(req);
1340                 target_request_copy_put(req);
1341         }
1342 }
1343
1344 static void abort_lock_replay_queue(struct obd_device *obd)
1345 {
1346         struct ptlrpc_request *req, *n;
1347         cfs_list_t abort_list;
1348
1349         CFS_INIT_LIST_HEAD(&abort_list);
1350         cfs_spin_lock(&obd->obd_recovery_task_lock);
1351         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1352         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1353         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1354                 DEBUG_REQ(D_ERROR, req, "aborted:");
1355                 req->rq_status = -ENOTCONN;
1356                 if (ptlrpc_error(req)) {
1357                         DEBUG_REQ(D_ERROR, req,
1358                                   "failed abort_lock_reply; skipping");
1359                 }
1360                 target_request_copy_put(req);
1361         }
1362 }
1363
1364 /* Called from a cleanup function if the device is being cleaned up
1365    forcefully.  The exports should all have been disconnected already,
1366    the only thing left to do is
1367      - clear the recovery flags
1368      - cancel the timer
1369      - free queued requests and replies, but don't send replies
1370    Because the obd_stopping flag is set, no new requests should be received.
1371
1372 */
1373 void target_cleanup_recovery(struct obd_device *obd)
1374 {
1375         struct ptlrpc_request *req, *n;
1376         cfs_list_t clean_list;
1377         ENTRY;
1378
1379         CFS_INIT_LIST_HEAD(&clean_list);
1380         cfs_spin_lock(&obd->obd_dev_lock);
1381         if (!obd->obd_recovering) {
1382                 cfs_spin_unlock(&obd->obd_dev_lock);
1383                 EXIT;
1384                 return;
1385         }
1386         obd->obd_recovering = obd->obd_abort_recovery = 0;
1387         cfs_spin_unlock(&obd->obd_dev_lock);
1388
1389         cfs_spin_lock(&obd->obd_recovery_task_lock);
1390         target_cancel_recovery_timer(obd);
1391         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1392         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1393
1394         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1395                 LASSERT(req->rq_reply_state == 0);
1396                 target_exp_dequeue_req_replay(req);
1397                 target_request_copy_put(req);
1398         }
1399
1400         cfs_spin_lock(&obd->obd_recovery_task_lock);
1401         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1402         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1403         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1404
1405         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1406                 LASSERT(req->rq_reply_state == 0);
1407                 target_request_copy_put(req);
1408         }
1409
1410         EXIT;
1411 }
1412
1413 /* obd_recovery_task_lock should be held */
1414 void target_cancel_recovery_timer(struct obd_device *obd)
1415 {
1416         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1417         cfs_timer_disarm(&obd->obd_recovery_timer);
1418 }
1419
1420 static void target_start_recovery_timer(struct obd_device *obd)
1421 {
1422         if (obd->obd_recovery_start != 0)
1423                 return;
1424
1425         cfs_spin_lock(&obd->obd_dev_lock);
1426         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1427                 cfs_spin_unlock(&obd->obd_dev_lock);
1428                 return;
1429         }
1430
1431         LASSERT(obd->obd_recovery_timeout != 0);
1432
1433         if (obd->obd_recovery_start != 0) {
1434                 cfs_spin_unlock(&obd->obd_dev_lock);
1435                 return;
1436         }
1437
1438         cfs_timer_arm(&obd->obd_recovery_timer,
1439                       cfs_time_shift(obd->obd_recovery_timeout));
1440         obd->obd_recovery_start = cfs_time_current_sec();
1441         cfs_spin_unlock(&obd->obd_dev_lock);
1442
1443         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1444                       "or until %d client%s reconnect%s\n",
1445                       obd->obd_name,
1446                       obd->obd_recovery_timeout / 60,
1447                       obd->obd_recovery_timeout % 60,
1448                       obd->obd_max_recoverable_clients,
1449                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1450                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1451 }
1452
1453 /**
1454  * extend recovery window.
1455  *
1456  * if @extend is true, extend recovery window to have @drt remaining at least;
1457  * otherwise, make sure the recovery timeout value is not less than @drt.
1458  */
1459 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1460 {
1461         cfs_time_t now;
1462         cfs_time_t end;
1463         cfs_duration_t left;
1464         int to;
1465
1466         cfs_spin_lock(&obd->obd_dev_lock);
1467         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1468                 cfs_spin_unlock(&obd->obd_dev_lock);
1469                 return;
1470         }
1471         LASSERT(obd->obd_recovery_start != 0);
1472
1473         now  = cfs_time_current_sec();
1474         to   = obd->obd_recovery_timeout;
1475         end  = obd->obd_recovery_start + to;
1476         left = cfs_time_sub(end, now);
1477
1478         if (extend && (drt > left)) {
1479                 to += drt - left;
1480         } else if (!extend && (drt > to)) {
1481                 to = drt;
1482                 /* reduce drt by already passed time */
1483                 drt -= obd->obd_recovery_timeout - left;
1484         }
1485
1486         if (to > obd->obd_recovery_time_hard)
1487                 to = obd->obd_recovery_time_hard;
1488         if (obd->obd_recovery_timeout < to) {
1489                 obd->obd_recovery_timeout = to;
1490                 cfs_timer_arm(&obd->obd_recovery_timer,
1491                               cfs_time_shift(drt));
1492         }
1493         cfs_spin_unlock(&obd->obd_dev_lock);
1494
1495         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1496                obd->obd_name, (unsigned)drt);
1497 }
1498
1499 /* Reset the timer with each new client connection */
1500 /*
1501  * This timer is actually reconnect_timer, which is for making sure
1502  * the total recovery window is at least as big as my reconnect
1503  * attempt timing. So the initial recovery time_out will be set to
1504  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1505  * from client is bigger than this, then the recovery time_out will
1506  * be extended to make sure the client could be reconnected, in the
1507  * process, the timeout from the new client should be ignored.
1508  */
1509
1510 static void
1511 check_and_start_recovery_timer(struct obd_device *obd,
1512                                struct ptlrpc_request *req,
1513                                int new_client)
1514 {
1515         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1516         struct obd_device_target *obt = &obd->u.obt;
1517         struct lustre_sb_info *lsi;
1518
1519         if (!new_client && service_time)
1520                 /* Teach server about old server's estimates, as first guess
1521                  * at how long new requests will take. */
1522                 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1523                             service_time);
1524
1525         target_start_recovery_timer(obd);
1526
1527         /* convert the service time to rpc timeout,
1528          * reuse service_time to limit stack usage */
1529         service_time = at_est2timeout(service_time);
1530
1531         /* We expect other clients to timeout within service_time, then try
1532          * to reconnect, then try the failover server.  The max delay between
1533          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1534         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1535
1536         LASSERT(obt->obt_magic == OBT_MAGIC);
1537         lsi = s2lsi(obt->obt_sb);
1538         if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1539                 service_time += 2 * (CONNECTION_SWITCH_MAX +
1540                                      CONNECTION_SWITCH_INC);
1541         if (service_time > obd->obd_recovery_timeout && !new_client)
1542                 extend_recovery_timer(obd, service_time, false);
1543 }
1544
1545 /** Health checking routines */
1546 static inline int exp_connect_healthy(struct obd_export *exp)
1547 {
1548         return (exp->exp_in_recovery);
1549 }
1550
1551 /** if export done req_replay or has replay in queue */
1552 static inline int exp_req_replay_healthy(struct obd_export *exp)
1553 {
1554         return (!exp->exp_req_replay_needed ||
1555                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1556 }
1557 /** if export done lock_replay or has replay in queue */
1558 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1559 {
1560         return (!exp->exp_lock_replay_needed ||
1561                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1562 }
1563
1564 static inline int exp_vbr_healthy(struct obd_export *exp)
1565 {
1566         return (!exp->exp_vbr_failed);
1567 }
1568
1569 static inline int exp_finished(struct obd_export *exp)
1570 {
1571         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1572 }
1573
1574 /** Checking routines for recovery */
1575 static int check_for_clients(struct obd_device *obd)
1576 {
1577         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1578
1579         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1580                 return 1;
1581         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1582         if (obd->obd_no_conn == 0 &&
1583             clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1584                 return 1;
1585         return 0;
1586 }
1587
1588 static int check_for_next_transno(struct obd_device *obd)
1589 {
1590         struct ptlrpc_request *req = NULL;
1591         int wake_up = 0, connected, completed, queue_len;
1592         __u64 next_transno, req_transno;
1593         ENTRY;
1594
1595         cfs_spin_lock(&obd->obd_recovery_task_lock);
1596         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1597                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1598                                      struct ptlrpc_request, rq_list);
1599                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1600         } else {
1601                 req_transno = 0;
1602         }
1603
1604         connected = cfs_atomic_read(&obd->obd_connected_clients);
1605         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1606         queue_len = obd->obd_requests_queued_for_recovery;
1607         next_transno = obd->obd_next_recovery_transno;
1608
1609         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1610                "req_transno: "LPU64", next_transno: "LPU64"\n",
1611                obd->obd_max_recoverable_clients, connected, completed,
1612                queue_len, req_transno, next_transno);
1613
1614         if (obd->obd_abort_recovery) {
1615                 CDEBUG(D_HA, "waking for aborted recovery\n");
1616                 wake_up = 1;
1617         } else if (obd->obd_recovery_expired) {
1618                 CDEBUG(D_HA, "waking for expired recovery\n");
1619                 wake_up = 1;
1620         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1621                 CDEBUG(D_HA, "waking for completed recovery\n");
1622                 wake_up = 1;
1623         } else if (req_transno == next_transno) {
1624                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1625                 wake_up = 1;
1626         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1627                 int d_lvl = D_HA;
1628                 /** handle gaps occured due to lost reply or VBR */
1629                 LASSERTF(req_transno >= next_transno,
1630                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1631                          req_transno, next_transno);
1632                 if (req_transno > obd->obd_last_committed &&
1633                     !obd->obd_version_recov)
1634                         d_lvl = D_ERROR;
1635                 CDEBUG(d_lvl,
1636                        "%s: waking for gap in transno, VBR is %s (skip: "
1637                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1638                        ", last_committed: "LPD64")\n",
1639                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1640                        next_transno, queue_len, completed, connected,
1641                        req_transno, obd->obd_last_committed);
1642                 obd->obd_next_recovery_transno = req_transno;
1643                 wake_up = 1;
1644         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1645                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1646                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1647                 obd->obd_next_recovery_transno = req_transno;
1648                 wake_up = 1;
1649         }
1650         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1651         return wake_up;
1652 }
1653
1654 static int check_for_next_lock(struct obd_device *obd)
1655 {
1656         int wake_up = 0;
1657
1658         cfs_spin_lock(&obd->obd_recovery_task_lock);
1659         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1660                 CDEBUG(D_HA, "waking for next lock\n");
1661                 wake_up = 1;
1662         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1663                 CDEBUG(D_HA, "waking for completed lock replay\n");
1664                 wake_up = 1;
1665         } else if (obd->obd_abort_recovery) {
1666                 CDEBUG(D_HA, "waking for aborted recovery\n");
1667                 wake_up = 1;
1668         } else if (obd->obd_recovery_expired) {
1669                 CDEBUG(D_HA, "waking for expired recovery\n");
1670                 wake_up = 1;
1671         }
1672         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1673
1674         return wake_up;
1675 }
1676
1677 /**
1678  * wait for recovery events,
1679  * check its status with help of check_routine
1680  * evict dead clients via health_check
1681  */
1682 static int target_recovery_overseer(struct obd_device *obd,
1683                                     int (*check_routine)(struct obd_device *),
1684                                     int (*health_check)(struct obd_export *))
1685 {
1686 repeat:
1687         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1688         if (obd->obd_abort_recovery) {
1689                 CWARN("recovery is aborted, evict exports in recovery\n");
1690                 /** evict exports which didn't finish recovery yet */
1691                 class_disconnect_stale_exports(obd, exp_finished);
1692                 return 1;
1693         } else if (obd->obd_recovery_expired) {
1694                 obd->obd_recovery_expired = 0;
1695                 /** If some clients died being recovered, evict them */
1696                 CDEBUG(D_WARNING,
1697                        "recovery is timed out, evict stale exports\n");
1698                 /** evict cexports with no replay in queue, they are stalled */
1699                 class_disconnect_stale_exports(obd, health_check);
1700                 /** continue with VBR */
1701                 cfs_spin_lock(&obd->obd_dev_lock);
1702                 obd->obd_version_recov = 1;
1703                 cfs_spin_unlock(&obd->obd_dev_lock);
1704                 /**
1705                  * reset timer, recovery will proceed with versions now,
1706                  * timeout is set just to handle reconnection delays
1707                  */
1708                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1709                 /** Wait for recovery events again, after evicting bad clients */
1710                 goto repeat;
1711         }
1712         return 0;
1713 }
1714
1715 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1716 {
1717         struct ptlrpc_request *req = NULL;
1718         ENTRY;
1719
1720         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1721                obd->obd_next_recovery_transno);
1722
1723         if (target_recovery_overseer(obd, check_for_next_transno,
1724                                      exp_req_replay_healthy)) {
1725                 abort_req_replay_queue(obd);
1726                 abort_lock_replay_queue(obd);
1727         }
1728
1729         cfs_spin_lock(&obd->obd_recovery_task_lock);
1730         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1731                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1732                                      struct ptlrpc_request, rq_list);
1733                 cfs_list_del_init(&req->rq_list);
1734                 obd->obd_requests_queued_for_recovery--;
1735                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1736         } else {
1737                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1738                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1739                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1740                 /** evict exports failed VBR */
1741                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1742         }
1743         RETURN(req);
1744 }
1745
1746 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1747 {
1748         struct ptlrpc_request *req = NULL;
1749
1750         CDEBUG(D_HA, "Waiting for lock\n");
1751         if (target_recovery_overseer(obd, check_for_next_lock,
1752                                      exp_lock_replay_healthy))
1753                 abort_lock_replay_queue(obd);
1754
1755         cfs_spin_lock(&obd->obd_recovery_task_lock);
1756         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1757                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1758                                      struct ptlrpc_request, rq_list);
1759                 cfs_list_del_init(&req->rq_list);
1760                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1761         } else {
1762                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1763                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1764                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1765                 /** evict exports failed VBR */
1766                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1767         }
1768         return req;
1769 }
1770
1771 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1772 {
1773         struct ptlrpc_request *req = NULL;
1774
1775         cfs_spin_lock(&obd->obd_recovery_task_lock);
1776         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1777                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1778                                      struct ptlrpc_request, rq_list);
1779                 cfs_list_del_init(&req->rq_list);
1780                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1781                 if (req->rq_export->exp_in_recovery) {
1782                         cfs_spin_lock(&req->rq_export->exp_lock);
1783                         req->rq_export->exp_in_recovery = 0;
1784                         cfs_spin_unlock(&req->rq_export->exp_lock);
1785                 }
1786         } else {
1787                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1788         }
1789         return req;
1790 }
1791
1792 static int handle_recovery_req(struct ptlrpc_thread *thread,
1793                                struct ptlrpc_request *req,
1794                                svc_handler_t handler)
1795 {
1796         int rc;
1797         ENTRY;
1798
1799         /**
1800          * export can be evicted during recovery, no need to handle replays for
1801          * it after that, discard such request silently
1802          */
1803         if (req->rq_export->exp_disconnected)
1804                 GOTO(reqcopy_put, rc = 0);
1805
1806         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1807         if (rc) {
1808                 CERROR("Failure to initialize session: %d\n", rc);
1809                 GOTO(reqcopy_put, rc);
1810         }
1811
1812         req->rq_recov_session.lc_thread = thread;
1813         lu_context_enter(&req->rq_recov_session);
1814         req->rq_svc_thread = thread;
1815         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1816
1817         /* thread context */
1818         lu_context_enter(&thread->t_env->le_ctx);
1819         (void)handler(req);
1820         lu_context_exit(&thread->t_env->le_ctx);
1821
1822         lu_context_exit(&req->rq_recov_session);
1823         lu_context_fini(&req->rq_recov_session);
1824         /* don't reset timer for final stage */
1825         if (!exp_finished(req->rq_export)) {
1826                 int to = obd_timeout;
1827
1828                 /**
1829                  * Add request timeout to the recovery time so next request from
1830                  * this client may come in recovery time
1831                  */
1832                 if (!AT_OFF) {
1833                         struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1834                         /* If the server sent early reply for this request,
1835                          * the client will recalculate the timeout according to
1836                          * current server estimate service time, so we will
1837                          * use the maxium timeout here for waiting the client
1838                          * sending the next req */
1839                         to = max((int)at_est2timeout(
1840                                  at_get(&svc->srv_at_estimate)),
1841                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1842                         /* Add net_latency (see ptlrpc_replay_req) */
1843                         to += lustre_msg_get_service_time(req->rq_reqmsg);
1844                 }
1845                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1846         }
1847 reqcopy_put:
1848         RETURN(rc);
1849 }
1850
1851 static int target_recovery_thread(void *arg)
1852 {
1853         struct lu_target *lut = arg;
1854         struct obd_device *obd = lut->lut_obd;
1855         struct ptlrpc_request *req;
1856         struct target_recovery_data *trd = &obd->obd_recovery_data;
1857         unsigned long delta;
1858         unsigned long flags;
1859         struct lu_env *env;
1860         struct ptlrpc_thread *thread = NULL;
1861         int rc = 0;
1862         ENTRY;
1863
1864         cfs_daemonize_ctxt("tgt_recov");
1865
1866         SIGNAL_MASK_LOCK(current, flags);
1867         sigfillset(&current->blocked);
1868         RECALC_SIGPENDING;
1869         SIGNAL_MASK_UNLOCK(current, flags);
1870
1871         OBD_ALLOC_PTR(thread);
1872         if (thread == NULL)
1873                 RETURN(-ENOMEM);
1874
1875         OBD_ALLOC_PTR(env);
1876         if (env == NULL) {
1877                 OBD_FREE_PTR(thread);
1878                 RETURN(-ENOMEM);
1879         }
1880
1881         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD);
1882         if (rc) {
1883                 OBD_FREE_PTR(thread);
1884                 OBD_FREE_PTR(env);
1885                 RETURN(rc);
1886         }
1887
1888         thread->t_env = env;
1889         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1890         env->le_ctx.lc_thread = thread;
1891         thread->t_data = NULL;
1892         thread->t_watchdog = NULL;
1893
1894         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1895                cfs_curproc_pid());
1896         trd->trd_processing_task = cfs_curproc_pid();
1897
1898         cfs_spin_lock(&obd->obd_dev_lock);
1899         obd->obd_recovering = 1;
1900         cfs_spin_unlock(&obd->obd_dev_lock);
1901         cfs_complete(&trd->trd_starting);
1902
1903         /* first of all, we have to know the first transno to replay */
1904         if (target_recovery_overseer(obd, check_for_clients,
1905                                      exp_connect_healthy)) {
1906                 abort_req_replay_queue(obd);
1907                 abort_lock_replay_queue(obd);
1908         }
1909
1910         /* next stage: replay requests */
1911         delta = jiffies;
1912         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1913                cfs_atomic_read(&obd->obd_req_replay_clients),
1914                obd->obd_next_recovery_transno);
1915         while ((req = target_next_replay_req(obd))) {
1916                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1917                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1918                           lustre_msg_get_transno(req->rq_reqmsg),
1919                           libcfs_nid2str(req->rq_peer.nid));
1920                 handle_recovery_req(thread, req,
1921                                     trd->trd_recovery_handler);
1922                 /**
1923                  * bz18031: increase next_recovery_transno before
1924                  * target_request_copy_put() will drop exp_rpc reference
1925                  */
1926                 cfs_spin_lock(&obd->obd_recovery_task_lock);
1927                 obd->obd_next_recovery_transno++;
1928                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1929                 target_exp_dequeue_req_replay(req);
1930                 target_request_copy_put(req);
1931                 obd->obd_replayed_requests++;
1932         }
1933
1934         /**
1935          * The second stage: replay locks
1936          */
1937         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1938                cfs_atomic_read(&obd->obd_lock_replay_clients));
1939         while ((req = target_next_replay_lock(obd))) {
1940                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1941                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1942                           libcfs_nid2str(req->rq_peer.nid));
1943                 handle_recovery_req(thread, req,
1944                                     trd->trd_recovery_handler);
1945                 target_request_copy_put(req);
1946                 obd->obd_replayed_locks++;
1947         }
1948
1949         /**
1950          * The third stage: reply on final pings, at this moment all clients
1951          * must have request in final queue
1952          */
1953         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1954         /** Update server last boot epoch */
1955         lut_boot_epoch_update(lut);
1956         /* We drop recoverying flag to forward all new requests
1957          * to regular mds_handle() since now */
1958         cfs_spin_lock(&obd->obd_dev_lock);
1959         obd->obd_recovering = obd->obd_abort_recovery = 0;
1960         cfs_spin_unlock(&obd->obd_dev_lock);
1961         cfs_spin_lock(&obd->obd_recovery_task_lock);
1962         target_cancel_recovery_timer(obd);
1963         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1964         while ((req = target_next_final_ping(obd))) {
1965                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1966                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1967                           libcfs_nid2str(req->rq_peer.nid));
1968                 handle_recovery_req(thread, req,
1969                                     trd->trd_recovery_handler);
1970                 target_request_copy_put(req);
1971         }
1972
1973         delta = (jiffies - delta) / CFS_HZ;
1974         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1975               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1976         if (delta > OBD_RECOVERY_TIME_SOFT) {
1977                 CWARN("too long recovery - read logs\n");
1978                 libcfs_debug_dumplog();
1979         }
1980
1981         target_finish_recovery(obd);
1982
1983         lu_context_fini(&env->le_ctx);
1984         trd->trd_processing_task = 0;
1985         cfs_complete(&trd->trd_finishing);
1986
1987         OBD_FREE_PTR(thread);
1988         OBD_FREE_PTR(env);
1989         RETURN(rc);
1990 }
1991
1992 static int target_start_recovery_thread(struct lu_target *lut,
1993                                         svc_handler_t handler)
1994 {
1995         struct obd_device *obd = lut->lut_obd;
1996         int rc = 0;
1997         struct target_recovery_data *trd = &obd->obd_recovery_data;
1998
1999         memset(trd, 0, sizeof(*trd));
2000         cfs_init_completion(&trd->trd_starting);
2001         cfs_init_completion(&trd->trd_finishing);
2002         trd->trd_recovery_handler = handler;
2003
2004         if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2005                 cfs_wait_for_completion(&trd->trd_starting);
2006                 LASSERT(obd->obd_recovering != 0);
2007         } else
2008                 rc = -ECHILD;
2009
2010         return rc;
2011 }
2012
2013 void target_stop_recovery_thread(struct obd_device *obd)
2014 {
2015         if (obd->obd_recovery_data.trd_processing_task > 0) {
2016                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2017                 /** recovery can be done but postrecovery is not yet */
2018                 cfs_spin_lock(&obd->obd_dev_lock);
2019                 if (obd->obd_recovering) {
2020                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2021                         obd->obd_abort_recovery = 1;
2022                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2023                 }
2024                 cfs_spin_unlock(&obd->obd_dev_lock);
2025                 cfs_wait_for_completion(&trd->trd_finishing);
2026         }
2027 }
2028
2029 void target_recovery_fini(struct obd_device *obd)
2030 {
2031         class_disconnect_exports(obd);
2032         target_stop_recovery_thread(obd);
2033         target_cleanup_recovery(obd);
2034 }
2035 EXPORT_SYMBOL(target_recovery_fini);
2036
2037 static void target_recovery_expired(unsigned long castmeharder)
2038 {
2039         struct obd_device *obd = (struct obd_device *)castmeharder;
2040         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2041                " after %lds (%d clients connected)\n",
2042                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2043                cfs_time_current_sec()- obd->obd_recovery_start,
2044                cfs_atomic_read(&obd->obd_connected_clients));
2045
2046         obd->obd_recovery_expired = 1;
2047         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2048 }
2049
2050 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2051 {
2052         struct obd_device *obd = lut->lut_obd;
2053         if (obd->obd_max_recoverable_clients == 0) {
2054                 /** Update server last boot epoch */
2055                 lut_boot_epoch_update(lut);
2056                 return;
2057         }
2058
2059         CWARN("RECOVERY: service %s, %d recoverable clients, "
2060               "last_transno "LPU64"\n", obd->obd_name,
2061               obd->obd_max_recoverable_clients, obd->obd_last_committed);
2062         LASSERT(obd->obd_stopping == 0);
2063         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2064         obd->obd_recovery_start = 0;
2065         obd->obd_recovery_end = 0;
2066
2067         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2068         target_start_recovery_thread(lut, handler);
2069 }
2070 EXPORT_SYMBOL(target_recovery_init);
2071
2072 #endif
2073
2074 static int target_process_req_flags(struct obd_device *obd,
2075                                     struct ptlrpc_request *req)
2076 {
2077         struct obd_export *exp = req->rq_export;
2078         LASSERT(exp != NULL);
2079         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2080                 /* client declares he's ready to replay locks */
2081                 cfs_spin_lock(&exp->exp_lock);
2082                 if (exp->exp_req_replay_needed) {
2083                         exp->exp_req_replay_needed = 0;
2084                         cfs_spin_unlock(&exp->exp_lock);
2085
2086                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2087                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2088                 } else {
2089                         cfs_spin_unlock(&exp->exp_lock);
2090                 }
2091         }
2092         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2093                 /* client declares he's ready to complete recovery
2094                  * so, we put the request on th final queue */
2095                 cfs_spin_lock(&exp->exp_lock);
2096                 if (exp->exp_lock_replay_needed) {
2097                         exp->exp_lock_replay_needed = 0;
2098                         cfs_spin_unlock(&exp->exp_lock);
2099
2100                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2101                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2102                 } else {
2103                         cfs_spin_unlock(&exp->exp_lock);
2104                 }
2105         }
2106         return 0;
2107 }
2108
2109 int target_queue_recovery_request(struct ptlrpc_request *req,
2110                                   struct obd_device *obd)
2111 {
2112         cfs_list_t *tmp;
2113         int inserted = 0;
2114         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2115         ENTRY;
2116
2117         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2118                 /* Processing the queue right now, don't re-add. */
2119                 RETURN(1);
2120         }
2121
2122         target_process_req_flags(obd, req);
2123
2124         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2125                 /* client declares he's ready to complete recovery
2126                  * so, we put the request on th final queue */
2127                 target_request_copy_get(req);
2128                 DEBUG_REQ(D_HA, req, "queue final req");
2129                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2130                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2131                 if (obd->obd_recovering) {
2132                         cfs_list_add_tail(&req->rq_list,
2133                                           &obd->obd_final_req_queue);
2134                 } else {
2135                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2136                         target_request_copy_put(req);
2137                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2138                 }
2139                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2140                 RETURN(0);
2141         }
2142         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2143                 /* client declares he's ready to replay locks */
2144                 target_request_copy_get(req);
2145                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2146                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2147                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2148                 LASSERT(obd->obd_recovering);
2149                 /* usually due to recovery abort */
2150                 if (!req->rq_export->exp_in_recovery) {
2151                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2152                         target_request_copy_put(req);
2153                         RETURN(-ENOTCONN);
2154                 }
2155                 LASSERT(req->rq_export->exp_lock_replay_needed);
2156                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2157                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2158                 RETURN(0);
2159         }
2160
2161         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2162          * (i.e. buflens etc are in my own byte order), but type-dependent
2163          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2164
2165         if (!transno) {
2166                 CFS_INIT_LIST_HEAD(&req->rq_list);
2167                 DEBUG_REQ(D_HA, req, "not queueing");
2168                 RETURN(1);
2169         }
2170
2171         /* If we're processing the queue, we want don't want to queue this
2172          * message.
2173          *
2174          * Also, if this request has a transno less than the one we're waiting
2175          * for, we should process it now.  It could (and currently always will)
2176          * be an open request for a descriptor that was opened some time ago.
2177          *
2178          * Also, a resent, replayed request that has already been
2179          * handled will pass through here and be processed immediately.
2180          */
2181         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying\n",
2182               obd->obd_next_recovery_transno, transno);
2183         cfs_spin_lock(&obd->obd_recovery_task_lock);
2184         if (transno < obd->obd_next_recovery_transno) {
2185                 /* Processing the queue right now, don't re-add. */
2186                 LASSERT(cfs_list_empty(&req->rq_list));
2187                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2188                 RETURN(1);
2189         }
2190         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2191
2192         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2193                 RETURN(0);
2194
2195         target_request_copy_get(req);
2196         if (!req->rq_export->exp_in_recovery) {
2197                 target_request_copy_put(req);
2198                 RETURN(-ENOTCONN);
2199         }
2200         LASSERT(req->rq_export->exp_req_replay_needed);
2201
2202         if (target_exp_enqueue_req_replay(req)) {
2203                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2204                 target_request_copy_put(req);
2205                 RETURN(0);
2206         }
2207
2208         /* XXX O(n^2) */
2209         cfs_spin_lock(&obd->obd_recovery_task_lock);
2210         LASSERT(obd->obd_recovering);
2211         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2212                 struct ptlrpc_request *reqiter =
2213                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2214
2215                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2216                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2217                         inserted = 1;
2218                         break;
2219                 }
2220
2221                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2222                              transno)) {
2223                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2224                                   "has been claimed by another client");
2225                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2226                         target_exp_dequeue_req_replay(req);
2227                         target_request_copy_put(req);
2228                         RETURN(0);
2229                 }
2230         }
2231
2232         if (!inserted)
2233                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2234
2235         obd->obd_requests_queued_for_recovery++;
2236         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2237         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2238         RETURN(0);
2239 }
2240
2241 /**
2242  * Packs current SLV and Limit into \a req.
2243  */
2244 int target_pack_pool_reply(struct ptlrpc_request *req)
2245 {
2246         struct obd_device *obd;
2247         ENTRY;
2248
2249         /*
2250          * Check that we still have all structures alive as this may
2251          * be some late rpc in shutdown time.
2252          */
2253         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2254                      !exp_connect_lru_resize(req->rq_export))) {
2255                 lustre_msg_set_slv(req->rq_repmsg, 0);
2256                 lustre_msg_set_limit(req->rq_repmsg, 0);
2257                 RETURN(0);
2258         }
2259
2260         /*
2261          * OBD is alive here as export is alive, which we checked above.
2262          */
2263         obd = req->rq_export->exp_obd;
2264
2265         cfs_read_lock(&obd->obd_pool_lock);
2266         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2267         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2268         cfs_read_unlock(&obd->obd_pool_lock);
2269
2270         RETURN(0);
2271 }
2272
2273 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2274 {
2275         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2276                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2277                 return (-ECOMM);
2278         }
2279
2280         if (unlikely(rc)) {
2281                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2282                 req->rq_status = rc;
2283                 return (ptlrpc_send_error(req, 1));
2284         } else {
2285                 DEBUG_REQ(D_NET, req, "sending reply");
2286         }
2287
2288         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2289 }
2290
2291 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2292 {
2293         int                        netrc;
2294         struct ptlrpc_reply_state *rs;
2295         struct obd_export         *exp;
2296         struct ptlrpc_service     *svc;
2297         ENTRY;
2298
2299         if (req->rq_no_reply) {
2300                 EXIT;
2301                 return;
2302         }
2303
2304         svc = req->rq_rqbd->rqbd_service;
2305         rs = req->rq_reply_state;
2306         if (rs == NULL || !rs->rs_difficult) {
2307                 /* no notifiers */
2308                 target_send_reply_msg (req, rc, fail_id);
2309                 EXIT;
2310                 return;
2311         }
2312
2313         /* must be an export if locks saved */
2314         LASSERT (req->rq_export != NULL);
2315         /* req/reply consistent */
2316         LASSERT (rs->rs_service == svc);
2317
2318         /* "fresh" reply */
2319         LASSERT (!rs->rs_scheduled);
2320         LASSERT (!rs->rs_scheduled_ever);
2321         LASSERT (!rs->rs_handled);
2322         LASSERT (!rs->rs_on_net);
2323         LASSERT (rs->rs_export == NULL);
2324         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2325         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2326
2327         exp = class_export_get (req->rq_export);
2328
2329         /* disable reply scheduling while I'm setting up */
2330         rs->rs_scheduled = 1;
2331         rs->rs_on_net    = 1;
2332         rs->rs_xid       = req->rq_xid;
2333         rs->rs_transno   = req->rq_transno;
2334         rs->rs_export    = exp;
2335         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2336
2337         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2338         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2339                rs->rs_transno, exp->exp_last_committed);
2340         if (rs->rs_transno > exp->exp_last_committed) {
2341                 /* not committed already */
2342                 cfs_list_add_tail(&rs->rs_obd_list,
2343                                   &exp->exp_uncommitted_replies);
2344         }
2345         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2346
2347         cfs_spin_lock(&exp->exp_lock);
2348         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2349         cfs_spin_unlock(&exp->exp_lock);
2350
2351         netrc = target_send_reply_msg (req, rc, fail_id);
2352
2353         cfs_spin_lock(&svc->srv_rs_lock);
2354
2355         cfs_atomic_inc(&svc->srv_n_difficult_replies);
2356
2357         if (netrc != 0) {
2358                 /* error sending: reply is off the net.  Also we need +1
2359                  * reply ref until ptlrpc_handle_rs() is done
2360                  * with the reply state (if the send was successful, there
2361                  * would have been +1 ref for the net, which
2362                  * reply_out_callback leaves alone) */
2363                 rs->rs_on_net = 0;
2364                 ptlrpc_rs_addref(rs);
2365         }
2366
2367         cfs_spin_lock(&rs->rs_lock);
2368         if (rs->rs_transno <= exp->exp_last_committed ||
2369             (!rs->rs_on_net && !rs->rs_no_ack) ||
2370              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2371              cfs_list_empty(&rs->rs_obd_list)) {
2372                 CDEBUG(D_HA, "Schedule reply immediately\n");
2373                 ptlrpc_dispatch_difficult_reply(rs);
2374         } else {
2375                 cfs_list_add (&rs->rs_list, &svc->srv_active_replies);
2376                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2377         }
2378         cfs_spin_unlock(&rs->rs_lock);
2379         cfs_spin_unlock(&svc->srv_rs_lock);
2380         EXIT;
2381 }
2382
2383 int target_handle_ping(struct ptlrpc_request *req)
2384 {
2385         obd_ping(req->rq_export);
2386         return req_capsule_server_pack(&req->rq_pill);
2387 }
2388
2389 void target_committed_to_req(struct ptlrpc_request *req)
2390 {
2391         struct obd_export *exp = req->rq_export;
2392
2393         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2394                 lustre_msg_set_last_committed(req->rq_repmsg,
2395                                               exp->exp_last_committed);
2396         else
2397                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2398                           "%d)", exp->exp_obd->obd_no_transno,
2399                           req->rq_repmsg == NULL);
2400
2401         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2402                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2403 }
2404 EXPORT_SYMBOL(target_committed_to_req);
2405
2406 int target_handle_qc_callback(struct ptlrpc_request *req)
2407 {
2408         struct obd_quotactl *oqctl;
2409         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2410
2411         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2412         if (oqctl == NULL) {
2413                 CERROR("Can't unpack obd_quotactl\n");
2414                 RETURN(-EPROTO);
2415         }
2416
2417         cli->cl_qchk_stat = oqctl->qc_stat;
2418
2419         return 0;
2420 }
2421
2422 #ifdef HAVE_QUOTA_SUPPORT
2423 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2424 {
2425 #ifdef __KERNEL__
2426         struct obd_device *obd = req->rq_export->exp_obd;
2427         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2428         struct obd_device_target *obt;
2429         struct lustre_quota_ctxt *qctxt;
2430         struct qunit_data *qdata = NULL;
2431         int rc = 0;
2432         ENTRY;
2433
2434         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2435                 RETURN(rc);
2436
2437         rc = req_capsule_server_pack(&req->rq_pill);
2438         if (rc) {
2439                 CERROR("packing reply failed!: rc = %d\n", rc);
2440                 RETURN(rc);
2441         }
2442
2443         LASSERT(req->rq_export);
2444
2445         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2446         if (IS_ERR(qdata)) {
2447                 rc = PTR_ERR(qdata);
2448                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2449                 req->rq_status = rc;
2450                 GOTO(out, rc);
2451         }
2452
2453         /* we use the observer */
2454         if (obd_pin_observer(obd, &lov_obd) ||
2455             obd_pin_observer(lov_obd, &master_obd)) {
2456                 CERROR("Can't find the observer, it is recovering\n");
2457                 req->rq_status = -EAGAIN;
2458                 GOTO(out, rc);
2459         }
2460
2461         obt = &master_obd->u.obt;
2462         qctxt = &obt->obt_qctxt;
2463
2464         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2465                 /* quota_type has not been processed yet, return EAGAIN
2466                  * until we know whether or not quotas are supposed to
2467                  * be enabled */
2468                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2469                        "-EAGAIN\n");
2470                 req->rq_status = -EAGAIN;
2471                 GOTO(out, rc);
2472         }
2473
2474         cfs_down_read(&obt->obt_rwsem);
2475         if (qctxt->lqc_lqs_hash == NULL) {
2476                 cfs_up_read(&obt->obt_rwsem);
2477                 /* quota_type has not been processed yet, return EAGAIN
2478                  * until we know whether or not quotas are supposed to
2479                  * be enabled */
2480                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2481                        "-EAGAIN\n");
2482                 req->rq_status = -EAGAIN;
2483                 GOTO(out, rc);
2484         }
2485
2486         LASSERT(qctxt->lqc_handler);
2487         rc = qctxt->lqc_handler(master_obd, qdata,
2488                                 lustre_msg_get_opc(req->rq_reqmsg));
2489         cfs_up_read(&obt->obt_rwsem);
2490         if (rc && rc != -EDQUOT)
2491                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2492                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2493         req->rq_status = rc;
2494
2495         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2496         if (rc < 0) {
2497                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2498                 GOTO(out, rc);
2499         }
2500
2501         /* Block the quota req. b=14840 */
2502         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2503         EXIT;
2504
2505 out:
2506         if (master_obd)
2507                 obd_unpin_observer(lov_obd);
2508         if (lov_obd)
2509                 obd_unpin_observer(obd);
2510
2511         rc = ptlrpc_reply(req);
2512         return rc;
2513 #else
2514         return 0;
2515 #endif /* !__KERNEL__ */
2516 }
2517 #endif /* HAVE_QUOTA_SUPPORT */
2518
2519 ldlm_mode_t lck_compat_array[] = {
2520         [LCK_EX] LCK_COMPAT_EX,
2521         [LCK_PW] LCK_COMPAT_PW,
2522         [LCK_PR] LCK_COMPAT_PR,
2523         [LCK_CW] LCK_COMPAT_CW,
2524         [LCK_CR] LCK_COMPAT_CR,
2525         [LCK_NL] LCK_COMPAT_NL,
2526         [LCK_GROUP] LCK_COMPAT_GROUP,
2527         [LCK_COS] LCK_COMPAT_COS,
2528 };
2529
2530 /**
2531  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2532  * not escape to the user level.
2533  */
2534 int ldlm_error2errno(ldlm_error_t error)
2535 {
2536         int result;
2537
2538         switch (error) {
2539         case ELDLM_OK:
2540                 result = 0;
2541                 break;
2542         case ELDLM_LOCK_CHANGED:
2543                 result = -ESTALE;
2544                 break;
2545         case ELDLM_LOCK_ABORTED:
2546                 result = -ENAVAIL;
2547                 break;
2548         case ELDLM_LOCK_REPLACED:
2549                 result = -ESRCH;
2550                 break;
2551         case ELDLM_NO_LOCK_DATA:
2552                 result = -ENOENT;
2553                 break;
2554         case ELDLM_NAMESPACE_EXISTS:
2555                 result = -EEXIST;
2556                 break;
2557         case ELDLM_BAD_NAMESPACE:
2558                 result = -EBADF;
2559                 break;
2560         default:
2561                 if (((int)error) < 0)  /* cast to signed type */
2562                         result = error; /* as ldlm_error_t can be unsigned */
2563                 else {
2564                         CERROR("Invalid DLM result code: %d\n", error);
2565                         result = -EPROTO;
2566                 }
2567         }
2568         return result;
2569 }
2570 EXPORT_SYMBOL(ldlm_error2errno);
2571
2572 /**
2573  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2574  */
2575 ldlm_error_t ldlm_errno2error(int err_no)
2576 {
2577         int error;
2578
2579         switch (err_no) {
2580         case 0:
2581                 error = ELDLM_OK;
2582                 break;
2583         case -ESTALE:
2584                 error = ELDLM_LOCK_CHANGED;
2585                 break;
2586         case -ENAVAIL:
2587                 error = ELDLM_LOCK_ABORTED;
2588                 break;
2589         case -ESRCH:
2590                 error = ELDLM_LOCK_REPLACED;
2591                 break;
2592         case -ENOENT:
2593                 error = ELDLM_NO_LOCK_DATA;
2594                 break;
2595         case -EEXIST:
2596                 error = ELDLM_NAMESPACE_EXISTS;
2597                 break;
2598         case -EBADF:
2599                 error = ELDLM_BAD_NAMESPACE;
2600                 break;
2601         default:
2602                 error = err_no;
2603         }
2604         return error;
2605 }
2606 EXPORT_SYMBOL(ldlm_errno2error);
2607
2608 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2609 void ldlm_dump_export_locks(struct obd_export *exp)
2610 {
2611         cfs_spin_lock(&exp->exp_locks_list_guard);
2612         if (!cfs_list_empty(&exp->exp_locks_list)) {
2613             struct ldlm_lock *lock;
2614
2615             CERROR("dumping locks for export %p,"
2616                    "ignore if the unmount doesn't hang\n", exp);
2617             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2618                 ldlm_lock_dump(D_ERROR, lock, 0);
2619         }
2620         cfs_spin_unlock(&exp->exp_locks_list_guard);
2621 }
2622 #endif
2623
2624 static int target_bulk_timeout(void *data)
2625 {
2626         ENTRY;
2627         /* We don't fail the connection here, because having the export
2628          * killed makes the (vital) call to commitrw very sad.
2629          */
2630         RETURN(1);
2631 }
2632
2633 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2634 {
2635         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2636 }
2637
2638 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2639                    struct l_wait_info *lwi)
2640 {
2641         struct ptlrpc_request *req = desc->bd_req;
2642         int rc = 0;
2643         ENTRY;
2644
2645         /* Check if there is eviction in progress, and if so, wait for
2646          * it to finish */
2647         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2648                 *lwi = LWI_INTR(NULL, NULL);
2649                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2650                                   !cfs_atomic_read(&exp->exp_obd->
2651                                                    obd_evict_inprogress),
2652                                   lwi);
2653         }
2654
2655         /* Check if client was evicted or tried to reconnect already */
2656         if (exp->exp_failed || exp->exp_abort_active_req) {
2657                 rc = -ENOTCONN;
2658         } else {
2659                 if (desc->bd_type == BULK_PUT_SINK)
2660                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2661                 if (rc == 0)
2662                         rc = ptlrpc_start_bulk_transfer(desc);
2663         }
2664
2665         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2666                 ptlrpc_abort_bulk(desc);
2667         } else if (rc == 0) {
2668                 time_t start = cfs_time_current_sec();
2669                 do {
2670                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2671                         cfs_duration_t timeout = timeoutl <= 0 ?
2672                                 CFS_TICK : cfs_time_seconds(timeoutl);
2673                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2674                                                     cfs_time_seconds(1),
2675                                                    target_bulk_timeout,
2676                                                    desc);
2677                         rc = l_wait_event(desc->bd_waitq,
2678                                           !ptlrpc_server_bulk_active(desc) ||
2679                                           exp->exp_failed ||
2680                                           exp->exp_abort_active_req,
2681                                           lwi);
2682                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2683                         /* Wait again if we changed deadline */
2684                 } while ((rc == -ETIMEDOUT) &&
2685                          (req->rq_deadline > cfs_time_current_sec()));
2686
2687                 if (rc == -ETIMEDOUT) {
2688                         DEBUG_REQ(D_ERROR, req,
2689                                   "timeout on bulk %s after %ld%+lds",
2690                                   bulk2type(desc),
2691                                   req->rq_deadline - start,
2692                                   cfs_time_current_sec() -
2693                                   req->rq_deadline);
2694                         ptlrpc_abort_bulk(desc);
2695                 } else if (exp->exp_failed) {
2696                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2697                                   bulk2type(desc));
2698                         rc = -ENOTCONN;
2699                         ptlrpc_abort_bulk(desc);
2700                 } else if (exp->exp_abort_active_req) {
2701                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2702                                   bulk2type(desc));
2703                         /* we don't reply anyway */
2704                         rc = -ETIMEDOUT;
2705                         ptlrpc_abort_bulk(desc);
2706                 } else if (!desc->bd_success ||
2707                            desc->bd_nob_transferred != desc->bd_nob) {
2708                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2709                                   desc->bd_success ?
2710                                   "truncated" : "network error on",
2711                                   bulk2type(desc),
2712                                   desc->bd_nob_transferred,
2713                                   desc->bd_nob);
2714                         /* XXX should this be a different errno? */
2715                         rc = -ETIMEDOUT;
2716                 } else if (desc->bd_type == BULK_GET_SINK) {
2717                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2718                 }
2719         } else {
2720                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2721                           bulk2type(desc), rc);
2722         }
2723
2724         RETURN(rc);
2725 }
2726 EXPORT_SYMBOL(target_bulk_io);