lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Whamcloud, Inc.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_LDLM
  38
  39 #ifdef __KERNEL__
  40 # include <libcfs/libcfs.h>
  41 #else
  42 # include <liblustre.h>
  43 #endif
  44 #include <obd.h>
  45 #include <obd_class.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_net.h>
  48 #include <lustre_sec.h>
  49 #include "ldlm_internal.h"
  50
  51 /* @priority: if non-zero, move the selected to the list head
  52  * @create: if zero, only search in existed connections
  53  */
  54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  55                            int priority, int create)
  56 {
  57         struct ptlrpc_connection *ptlrpc_conn;
  58         struct obd_import_conn *imp_conn = NULL, *item;
  59         int rc = 0;
  60         ENTRY;
  61
  62         if (!create && !priority) {
  63                 CDEBUG(D_HA, "Nothing to do\n");
  64                 RETURN(-EINVAL);
  65         }
  66
  67         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  68         if (!ptlrpc_conn) {
  69                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  70                 RETURN (-ENOENT);
  71         }
  72
  73         if (create) {
  74                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  75                 if (!imp_conn) {
  76                         GOTO(out_put, rc = -ENOMEM);
  77                 }
  78         }
  79
  80         spin_lock(&imp->imp_lock);
  81         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  82                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  83                         if (priority) {
  84                                 cfs_list_del(&item->oic_item);
  85                                 cfs_list_add(&item->oic_item,
  86                                              &imp->imp_conn_list);
  87                                 item->oic_last_attempt = 0;
  88                         }
  89                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  90                                imp, imp->imp_obd->obd_name, uuid->uuid,
  91                                (priority ? ", moved to head" : ""));
  92                         spin_unlock(&imp->imp_lock);
  93                         GOTO(out_free, rc = 0);
  94                 }
  95         }
  96         /* not found */
  97         if (create) {
  98                 imp_conn->oic_conn = ptlrpc_conn;
  99                 imp_conn->oic_uuid = *uuid;
 100                 imp_conn->oic_last_attempt = 0;
 101                 if (priority)
 102                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 103                 else
 104                         cfs_list_add_tail(&imp_conn->oic_item,
 105                                           &imp->imp_conn_list);
 106                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 107                        imp, imp->imp_obd->obd_name, uuid->uuid,
 108                        (priority ? "head" : "tail"));
 109         } else {
 110                 spin_unlock(&imp->imp_lock);
 111                 GOTO(out_free, rc = -ENOENT);
 112         }
 113
 114         spin_unlock(&imp->imp_lock);
 115         RETURN(0);
 116 out_free:
 117         if (imp_conn)
 118                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 119 out_put:
 120         ptlrpc_connection_put(ptlrpc_conn);
 121         RETURN(rc);
 122 }
 123
 124 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 125 {
 126         return import_set_conn(imp, uuid, 1, 0);
 127 }
 128
 129 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 130                            int priority)
 131 {
 132         return import_set_conn(imp, uuid, priority, 1);
 133 }
 134 EXPORT_SYMBOL(client_import_add_conn);
 135
 136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 137 {
 138         struct obd_import_conn *imp_conn;
 139         struct obd_export *dlmexp;
 140         int rc = -ENOENT;
 141         ENTRY;
 142
 143         spin_lock(&imp->imp_lock);
 144         if (cfs_list_empty(&imp->imp_conn_list)) {
 145                 LASSERT(!imp->imp_connection);
 146                 GOTO(out, rc);
 147         }
 148
 149         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 150                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 151                         continue;
 152                 LASSERT(imp_conn->oic_conn);
 153
 154                 /* is current conn? */
 155                 if (imp_conn == imp->imp_conn_current) {
 156                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 157
 158                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 159                             imp->imp_state != LUSTRE_IMP_DISCON) {
 160                                 CERROR("can't remove current connection\n");
 161                                 GOTO(out, rc = -EBUSY);
 162                         }
 163
 164                         ptlrpc_connection_put(imp->imp_connection);
 165                         imp->imp_connection = NULL;
 166
 167                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 168                         if (dlmexp && dlmexp->exp_connection) {
 169                                 LASSERT(dlmexp->exp_connection ==
 170                                         imp_conn->oic_conn);
 171                                 ptlrpc_connection_put(dlmexp->exp_connection);
 172                                 dlmexp->exp_connection = NULL;
 173                         }
 174                 }
 175
 176                 cfs_list_del(&imp_conn->oic_item);
 177                 ptlrpc_connection_put(imp_conn->oic_conn);
 178                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 179                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 180                        imp, imp->imp_obd->obd_name, uuid->uuid);
 181                 rc = 0;
 182                 break;
 183         }
 184 out:
 185         spin_unlock(&imp->imp_lock);
 186         if (rc == -ENOENT)
 187                 CERROR("connection %s not found\n", uuid->uuid);
 188         RETURN(rc);
 189 }
 190 EXPORT_SYMBOL(client_import_del_conn);
 191
 192 /**
 193  * Find conn uuid by peer nid. @peer is a server nid. This function is used
 194  * to find a conn uuid of @imp which can reach @peer.
 195  */
 196 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 197                             struct obd_uuid *uuid)
 198 {
 199         struct obd_import_conn *conn;
 200         int rc = -ENOENT;
 201         ENTRY;
 202
 203         spin_lock(&imp->imp_lock);
 204         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 205                 /* check if conn uuid does have this peer nid */
 206                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 207                         *uuid = conn->oic_uuid;
 208                         rc = 0;
 209                         break;
 210                 }
 211         }
 212         spin_unlock(&imp->imp_lock);
 213         RETURN(rc);
 214 }
 215 EXPORT_SYMBOL(client_import_find_conn);
 216
 217 void client_destroy_import(struct obd_import *imp)
 218 {
 219         /* drop security policy instance after all rpc finished/aborted
 220          * to let all busy contexts be released. */
 221         class_import_get(imp);
 222         class_destroy_import(imp);
 223         sptlrpc_import_sec_put(imp);
 224         class_import_put(imp);
 225 }
 226 EXPORT_SYMBOL(client_destroy_import);
 227
 228 /**
 229  * check whether the osc is on MDT or not
 230  * In the config log,
 231  * osc on MDT
 232  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 233  * osc on client
 234  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 235  *
 236  **/
 237 static int osc_on_mdt(char *obdname)
 238 {
 239         char *ptr;
 240
 241         ptr = strrchr(obdname, '-');
 242         if (ptr == NULL)
 243                 return 0;
 244
 245         if (strncmp(ptr + 1, "MDT", 3) == 0)
 246                 return 1;
 247
 248         return 0;
 249 }
 250
 251 /* configure an RPC client OBD device
 252  *
 253  * lcfg parameters:
 254  * 1 - client UUID
 255  * 2 - server UUID
 256  * 3 - inactive-on-startup
 257  */
 258 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 259 {
 260         struct client_obd *cli = &obddev->u.cli;
 261         struct obd_import *imp;
 262         struct obd_uuid server_uuid;
 263         int rq_portal, rp_portal, connect_op;
 264         char *name = obddev->obd_type->typ_name;
 265         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 266         int rc;
 267         ENTRY;
 268
 269         /* In a more perfect world, we would hang a ptlrpc_client off of
 270          * obd_type and just use the values from there. */
 271         if (!strcmp(name, LUSTRE_OSC_NAME) ||
 272             (!strcmp(name, LUSTRE_OSP_NAME) &&
 273              !is_osp_on_ost(lustre_cfg_buf(lcfg, 0)))) {
 274                 rq_portal = OST_REQUEST_PORTAL;
 275                 rp_portal = OSC_REPLY_PORTAL;
 276                 connect_op = OST_CONNECT;
 277                 cli->cl_sp_me = LUSTRE_SP_CLI;
 278                 cli->cl_sp_to = LUSTRE_SP_OST;
 279                 ns_type = LDLM_NS_TYPE_OSC;
 280
 281         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 282                    (!strcmp(name, LUSTRE_OSP_NAME) &&
 283                     is_osp_on_ost(lustre_cfg_buf(lcfg, 0)))) {
 284                 rq_portal = MDS_REQUEST_PORTAL;
 285                 rp_portal = MDC_REPLY_PORTAL;
 286                 connect_op = MDS_CONNECT;
 287                 cli->cl_sp_me = LUSTRE_SP_CLI;
 288                 cli->cl_sp_to = LUSTRE_SP_MDT;
 289                 ns_type = LDLM_NS_TYPE_MDC;
 290
 291         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 292                 rq_portal = MGS_REQUEST_PORTAL;
 293                 rp_portal = MGC_REPLY_PORTAL;
 294                 connect_op = MGS_CONNECT;
 295                 cli->cl_sp_me = LUSTRE_SP_MGC;
 296                 cli->cl_sp_to = LUSTRE_SP_MGS;
 297                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 298                 ns_type = LDLM_NS_TYPE_MGC;
 299
 300         } else if (!strcmp(name, LUSTRE_OSP_NAME)) {
 301                 rq_portal = OST_REQUEST_PORTAL;
 302                 rp_portal = OSC_REPLY_PORTAL;
 303                 connect_op = OST_CONNECT;
 304                 cli->cl_sp_me = LUSTRE_SP_CLI;
 305                 cli->cl_sp_to = LUSTRE_SP_OST;
 306                 ns_type = LDLM_NS_TYPE_OSC;
 307
 308         } else {
 309                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 310                        name);
 311                 RETURN(-EINVAL);
 312         }
 313
 314         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 315                 CERROR("requires a TARGET UUID\n");
 316                 RETURN(-EINVAL);
 317         }
 318
 319         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 320                 CERROR("client UUID must be less than 38 characters\n");
 321                 RETURN(-EINVAL);
 322         }
 323
 324         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 325                 CERROR("setup requires a SERVER UUID\n");
 326                 RETURN(-EINVAL);
 327         }
 328
 329         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 330                 CERROR("target UUID must be less than 38 characters\n");
 331                 RETURN(-EINVAL);
 332         }
 333
 334         init_rwsem(&cli->cl_sem);
 335         sema_init(&cli->cl_mgc_sem, 1);
 336         cli->cl_conn_count = 0;
 337         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 338                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 339                      sizeof(server_uuid)));
 340
 341         cli->cl_dirty = 0;
 342         cli->cl_avail_grant = 0;
 343         /* FIXME: should limit this for the sum of all cl_dirty_max */
 344         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 345         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
 346                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
 347         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 348         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 349         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 350         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 351         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 352         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 353         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
 354         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
 355         cli->cl_r_in_flight = 0;
 356         cli->cl_w_in_flight = 0;
 357
 358         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 359         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 360         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 361         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 362         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 363         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 364
 365         /* lru for osc. */
 366         CFS_INIT_LIST_HEAD(&cli->cl_lru_osc);
 367         cfs_atomic_set(&cli->cl_lru_shrinkers, 0);
 368         cfs_atomic_set(&cli->cl_lru_busy, 0);
 369         cfs_atomic_set(&cli->cl_lru_in_list, 0);
 370         CFS_INIT_LIST_HEAD(&cli->cl_lru_list);
 371         client_obd_list_lock_init(&cli->cl_lru_list_lock);
 372
 373         cfs_waitq_init(&cli->cl_destroy_waitq);
 374         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 375 #ifdef ENABLE_CHECKSUM
 376         /* Turn on checksumming by default. */
 377         cli->cl_checksum = 1;
 378         /*
 379          * The supported checksum types will be worked out at connect time
 380          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 381          * through procfs.
 382          */
 383         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 384 #endif
 385         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 386
 387         /* This value may be changed at connect time in
 388            ptlrpc_connect_interpret. */
 389         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 390                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 391
 392         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 393                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 394         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 395                 cli->cl_max_rpcs_in_flight = 2;
 396         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 397                 cli->cl_max_rpcs_in_flight = 3;
 398         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 399                 cli->cl_max_rpcs_in_flight = 4;
 400         } else {
 401                 if (osc_on_mdt(obddev->obd_name))
 402                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
 403                 else
 404                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 405         }
 406         rc = ldlm_get_ref();
 407         if (rc) {
 408                 CERROR("ldlm_get_ref failed: %d\n", rc);
 409                 GOTO(err, rc);
 410         }
 411
 412         ptlrpc_init_client(rq_portal, rp_portal, name,
 413                            &obddev->obd_ldlm_client);
 414
 415         imp = class_new_import(obddev);
 416         if (imp == NULL)
 417                 GOTO(err_ldlm, rc = -ENOENT);
 418         imp->imp_client = &obddev->obd_ldlm_client;
 419         imp->imp_connect_op = connect_op;
 420         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 421         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 422                LUSTRE_CFG_BUFLEN(lcfg, 1));
 423         class_import_put(imp);
 424
 425         rc = client_import_add_conn(imp, &server_uuid, 1);
 426         if (rc) {
 427                 CERROR("can't add initial connection\n");
 428                 GOTO(err_import, rc);
 429         }
 430
 431         cli->cl_import = imp;
 432         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 433         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 434         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 435
 436         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 437                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 438                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 439                                name, obddev->obd_name,
 440                                cli->cl_target_uuid.uuid);
 441                         spin_lock(&imp->imp_lock);
 442                         imp->imp_deactive = 1;
 443                         spin_unlock(&imp->imp_lock);
 444                 }
 445         }
 446
 447         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 448                                                    LDLM_NAMESPACE_CLIENT,
 449                                                    LDLM_NAMESPACE_GREEDY,
 450                                                    ns_type);
 451         if (obddev->obd_namespace == NULL) {
 452                 CERROR("Unable to create client namespace - %s\n",
 453                        obddev->obd_name);
 454                 GOTO(err_import, rc = -ENOMEM);
 455         }
 456
 457         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 458
 459         RETURN(rc);
 460
 461 err_import:
 462         class_destroy_import(imp);
 463 err_ldlm:
 464         ldlm_put_ref();
 465 err:
 466         RETURN(rc);
 467
 468 }
 469 EXPORT_SYMBOL(client_obd_setup);
 470
 471 int client_obd_cleanup(struct obd_device *obddev)
 472 {
 473         ENTRY;
 474
 475         ldlm_namespace_free_post(obddev->obd_namespace);
 476         obddev->obd_namespace = NULL;
 477
 478         LASSERT(obddev->u.cli.cl_import == NULL);
 479
 480         ldlm_put_ref();
 481         RETURN(0);
 482 }
 483 EXPORT_SYMBOL(client_obd_cleanup);
 484
 485 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 486 int client_connect_import(const struct lu_env *env,
 487                           struct obd_export **exp,
 488                           struct obd_device *obd, struct obd_uuid *cluuid,
 489                           struct obd_connect_data *data, void *localdata)
 490 {
 491         struct client_obd *cli = &obd->u.cli;
 492         struct obd_import *imp = cli->cl_import;
 493         struct obd_connect_data *ocd;
 494         struct lustre_handle conn = { 0 };
 495         int rc;
 496         ENTRY;
 497
 498         *exp = NULL;
 499         down_write(&cli->cl_sem);
 500         if (cli->cl_conn_count > 0 )
 501                 GOTO(out_sem, rc = -EALREADY);
 502
 503         rc = class_connect(&conn, obd, cluuid);
 504         if (rc)
 505                 GOTO(out_sem, rc);
 506
 507         cli->cl_conn_count++;
 508         *exp = class_conn2export(&conn);
 509
 510         LASSERT(obd->obd_namespace);
 511
 512         imp->imp_dlm_handle = conn;
 513         rc = ptlrpc_init_import(imp);
 514         if (rc != 0)
 515                 GOTO(out_ldlm, rc);
 516
 517         ocd = &imp->imp_connect_data;
 518         if (data) {
 519                 *ocd = *data;
 520                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 521         }
 522
 523         rc = ptlrpc_connect_import(imp);
 524         if (rc != 0) {
 525                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 526                 GOTO(out_ldlm, rc);
 527         }
 528         LASSERT((*exp)->exp_connection);
 529
 530         if (data) {
 531                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 532                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 533                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 534                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 535         }
 536
 537         ptlrpc_pinger_add_import(imp);
 538
 539         EXIT;
 540
 541         if (rc) {
 542 out_ldlm:
 543                 cli->cl_conn_count--;
 544                 class_disconnect(*exp);
 545                 *exp = NULL;
 546         }
 547 out_sem:
 548         up_write(&cli->cl_sem);
 549
 550         return rc;
 551 }
 552 EXPORT_SYMBOL(client_connect_import);
 553
 554 int client_disconnect_export(struct obd_export *exp)
 555 {
 556         struct obd_device *obd = class_exp2obd(exp);
 557         struct client_obd *cli;
 558         struct obd_import *imp;
 559         int rc = 0, err;
 560         ENTRY;
 561
 562         if (!obd) {
 563                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 564                        exp, exp ? exp->exp_handle.h_cookie : -1);
 565                 RETURN(-EINVAL);
 566         }
 567
 568         cli = &obd->u.cli;
 569         imp = cli->cl_import;
 570
 571         down_write(&cli->cl_sem);
 572         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 573                cli->cl_conn_count);
 574
 575         if (!cli->cl_conn_count) {
 576                 CERROR("disconnecting disconnected device (%s)\n",
 577                        obd->obd_name);
 578                 GOTO(out_disconnect, rc = -EINVAL);
 579         }
 580
 581         cli->cl_conn_count--;
 582         if (cli->cl_conn_count)
 583                 GOTO(out_disconnect, rc = 0);
 584
 585         /* Mark import deactivated now, so we don't try to reconnect if any
 586          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 587          * fully deactivate the import, or that would drop all requests. */
 588         spin_lock(&imp->imp_lock);
 589         imp->imp_deactive = 1;
 590         spin_unlock(&imp->imp_lock);
 591
 592         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 593          * delete it regardless.  (It's safe to delete an import that was
 594          * never added.) */
 595         (void)ptlrpc_pinger_del_import(imp);
 596
 597         if (obd->obd_namespace != NULL) {
 598                 /* obd_force == local only */
 599                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 600                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 601                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 602         }
 603
 604         /*
 605          * there's no need to hold sem during disconnecting an import,
 606          * and actually it may cause deadlock in gss.
 607          */
 608         up_write(&cli->cl_sem);
 609         rc = ptlrpc_disconnect_import(imp, 0);
 610         down_write(&cli->cl_sem);
 611
 612         ptlrpc_invalidate_import(imp);
 613
 614         EXIT;
 615
 616  out_disconnect:
 617         /* use server style - class_disconnect should be always called for
 618          * o_disconnect */
 619         err = class_disconnect(exp);
 620         if (!rc && err)
 621                 rc = err;
 622
 623         up_write(&cli->cl_sem);
 624
 625         RETURN(rc);
 626 }
 627 EXPORT_SYMBOL(client_disconnect_export);
 628
 629 #ifdef HAVE_SERVER_SUPPORT
 630 int server_disconnect_export(struct obd_export *exp)
 631 {
 632         int rc;
 633         ENTRY;
 634
 635         /* Disconnect early so that clients can't keep using export */
 636         rc = class_disconnect(exp);
 637         /* close import for avoid sending any requests */
 638         if (exp->exp_imp_reverse)
 639                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 640
 641         if (exp->exp_obd->obd_namespace != NULL)
 642                 ldlm_cancel_locks_for_export(exp);
 643
 644         /* complete all outstanding replies */
 645         spin_lock(&exp->exp_lock);
 646         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 647                 struct ptlrpc_reply_state *rs =
 648                         cfs_list_entry(exp->exp_outstanding_replies.next,
 649                                        struct ptlrpc_reply_state, rs_exp_list);
 650                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 651
 652                 spin_lock(&svcpt->scp_rep_lock);
 653
 654                 cfs_list_del_init(&rs->rs_exp_list);
 655                 spin_lock(&rs->rs_lock);
 656                 ptlrpc_schedule_difficult_reply(rs);
 657                 spin_unlock(&rs->rs_lock);
 658
 659                 spin_unlock(&svcpt->scp_rep_lock);
 660         }
 661         spin_unlock(&exp->exp_lock);
 662
 663         RETURN(rc);
 664 }
 665 EXPORT_SYMBOL(server_disconnect_export);
 666
 667 /* --------------------------------------------------------------------------
 668  * from old lib/target.c
 669  * -------------------------------------------------------------------------- */
 670
 671 static int target_handle_reconnect(struct lustre_handle *conn,
 672                                    struct obd_export *exp,
 673                                    struct obd_uuid *cluuid)
 674 {
 675         ENTRY;
 676
 677         if (exp->exp_connection && exp->exp_imp_reverse) {
 678                 struct lustre_handle *hdl;
 679                 struct obd_device *target;
 680
 681                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 682                 target = exp->exp_obd;
 683
 684                 /* Might be a re-connect after a partition. */
 685                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 686                         if (target->obd_recovering) {
 687                                 int timeout = cfs_duration_sec(cfs_time_sub(
 688                                         cfs_timer_deadline(
 689                                         &target->obd_recovery_timer),
 690                                         cfs_time_current()));
 691
 692                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 693                                         "ing, waiting for %d clients in recov"
 694                                         "ery for %d:%.02d\n", target->obd_name,
 695                                         obd_uuid2str(&exp->exp_client_uuid),
 696                                         obd_export_nid2str(exp),
 697                                         target->obd_max_recoverable_clients,
 698                                         timeout / 60, timeout % 60);
 699                         } else {
 700                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 701                                         "reconnecting\n", target->obd_name,
 702                                         obd_uuid2str(&exp->exp_client_uuid),
 703                                         obd_export_nid2str(exp));
 704                         }
 705
 706                         conn->cookie = exp->exp_handle.h_cookie;
 707                         /* target_handle_connect() treats EALREADY and
 708                          * -EALREADY differently.  EALREADY means we are
 709                          * doing a valid reconnect from the same client. */
 710                         RETURN(EALREADY);
 711                 } else {
 712                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 713                                       "with handle "LPX64". Rejecting client "
 714                                       "with the same UUID trying to reconnect "
 715                                       "with handle "LPX64"\n", target->obd_name,
 716                                       obd_uuid2str(&exp->exp_client_uuid),
 717                                       obd_export_nid2str(exp),
 718                                       hdl->cookie, conn->cookie);
 719                         memset(conn, 0, sizeof *conn);
 720                         /* target_handle_connect() treats EALREADY and
 721                          * -EALREADY differently.  -EALREADY is an error
 722                          * (same UUID, different handle). */
 723                         RETURN(-EALREADY);
 724                 }
 725         }
 726
 727         conn->cookie = exp->exp_handle.h_cookie;
 728         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 729                cluuid->uuid, exp, conn->cookie);
 730         RETURN(0);
 731 }
 732
 733 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 734                           int error)
 735 {
 736         struct obd_export *exp = cb_data;
 737
 738         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 739                obd->obd_name, exp->exp_client_uuid.uuid);
 740
 741         spin_lock(&exp->exp_lock);
 742         exp->exp_need_sync = 0;
 743         spin_unlock(&exp->exp_lock);
 744         class_export_cb_put(exp);
 745 }
 746 EXPORT_SYMBOL(target_client_add_cb);
 747
 748 #ifdef __KERNEL__
 749 static void
 750 check_and_start_recovery_timer(struct obd_device *obd,
 751                                struct ptlrpc_request *req, int new_client);
 752 #else
 753 static inline void
 754 check_and_start_recovery_timer(struct obd_device *obd,
 755                                struct ptlrpc_request *req, int new_client)
 756 {
 757 }
 758 #endif
 759
 760 int target_handle_connect(struct ptlrpc_request *req)
 761 {
 762         struct obd_device *target = NULL, *targref = NULL;
 763         struct obd_export *export = NULL;
 764         struct obd_import *revimp;
 765         struct obd_import *tmp_imp = NULL;
 766         struct lustre_handle conn;
 767         struct lustre_handle *tmp;
 768         struct obd_uuid tgtuuid;
 769         struct obd_uuid cluuid;
 770         struct obd_uuid remote_uuid;
 771         char *str;
 772         int rc = 0;
 773         char *target_start;
 774         int target_len;
 775         bool     mds_conn = false, lw_client = false;
 776         struct obd_connect_data *data, *tmpdata;
 777         int size, tmpsize;
 778         lnet_nid_t *client_nid = NULL;
 779         ENTRY;
 780
 781         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 782
 783         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 784         if (str == NULL) {
 785                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 786                 GOTO(out, rc = -EINVAL);
 787         }
 788
 789         obd_str2uuid(&tgtuuid, str);
 790         target = class_uuid2obd(&tgtuuid);
 791         if (!target)
 792                 target = class_name2obd(str);
 793
 794         if (!target) {
 795                 deuuidify(str, NULL, &target_start, &target_len);
 796                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
 797                                    "connect (no target)\n", str);
 798                 GOTO(out, rc = -ENODEV);
 799         }
 800
 801         spin_lock(&target->obd_dev_lock);
 802         if (target->obd_stopping || !target->obd_set_up) {
 803                 spin_unlock(&target->obd_dev_lock);
 804
 805                 deuuidify(str, NULL, &target_start, &target_len);
 806                 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
 807                                    "from %s (%s)\n", target_len, target_start,
 808                                    libcfs_nid2str(req->rq_peer.nid),
 809                                    (target->obd_stopping ?
 810                                    "stopping" : "not set up"));
 811                 GOTO(out, rc = -ENODEV);
 812         }
 813
 814         if (target->obd_no_conn) {
 815                 spin_unlock(&target->obd_dev_lock);
 816
 817                 LCONSOLE_WARN("%s: Temporarily refusing client connection "
 818                               "from %s\n", target->obd_name,
 819                               libcfs_nid2str(req->rq_peer.nid));
 820                 GOTO(out, rc = -EAGAIN);
 821         }
 822
 823         /* Make sure the target isn't cleaned up while we're here. Yes,
 824            there's still a race between the above check and our incref here.
 825            Really, class_uuid2obd should take the ref. */
 826         targref = class_incref(target, __FUNCTION__, cfs_current());
 827
 828         target->obd_conn_inprogress++;
 829         spin_unlock(&target->obd_dev_lock);
 830
 831         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 832         if (str == NULL) {
 833                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 834                 GOTO(out, rc = -EINVAL);
 835         }
 836
 837         obd_str2uuid(&cluuid, str);
 838
 839         /* XXX extract a nettype and format accordingly */
 840         switch (sizeof(lnet_nid_t)) {
 841                 /* NB the casts only avoid compiler warnings */
 842         case 8:
 843                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 844                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 845                 break;
 846         case 4:
 847                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 848                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 849                 break;
 850         default:
 851                 LBUG();
 852         }
 853
 854         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 855         if (tmp == NULL)
 856                 GOTO(out, rc = -EPROTO);
 857
 858         conn = *tmp;
 859
 860         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 861                                     RCL_CLIENT);
 862         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 863         if (!data)
 864                 GOTO(out, rc = -EPROTO);
 865
 866         rc = req_capsule_server_pack(&req->rq_pill);
 867         if (rc)
 868                 GOTO(out, rc);
 869
 870         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 871                 if (!data) {
 872                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 873                                   "libclient connection attempt");
 874                         GOTO(out, rc = -EPROTO);
 875                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 876                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 877                            data->ocd_version > LUSTRE_VERSION_CODE +
 878                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 879                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 880                                   "libclient connection attempt",
 881                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 882                                   "old" : "new",
 883                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 884                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 885                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 886                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 887                         data = req_capsule_server_sized_get(&req->rq_pill,
 888                                                         &RMF_CONNECT_DATA,
 889                                     offsetof(typeof(*data), ocd_version) +
 890                                              sizeof(data->ocd_version));
 891                         if (data) {
 892                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 893                                 data->ocd_version = LUSTRE_VERSION_CODE;
 894                         }
 895                         GOTO(out, rc = -EPROTO);
 896                 }
 897         }
 898
 899         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 900             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 901                 mds_conn = true;
 902
 903         if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
 904                 lw_client = true;
 905
 906         /* lctl gets a backstage, all-access pass. */
 907         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 908                 goto dont_check_exports;
 909
 910         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 911         if (!export)
 912                 goto no_export;
 913
 914         /* we've found an export in the hash */
 915
 916         spin_lock(&export->exp_lock);
 917
 918         if (export->exp_connecting) { /* bug 9635, et. al. */
 919                 spin_unlock(&export->exp_lock);
 920                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 921                               export->exp_obd->obd_name, export,
 922                               libcfs_nid2str(req->rq_peer.nid));
 923                 class_export_put(export);
 924                 export = NULL;
 925                 rc = -EALREADY;
 926         } else if (mds_conn && export->exp_connection) {
 927                 spin_unlock(&export->exp_lock);
 928                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 929                         /* mds reconnected after failover */
 930                         LCONSOLE_WARN("%s: Received MDS connection from "
 931                             "%s, removing former export from %s\n",
 932                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 933                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 934                 else
 935                         /* new mds connection from the same nid */
 936                         LCONSOLE_WARN("%s: Received new MDS connection from "
 937                             "%s, removing former export from same NID\n",
 938                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 939                 class_fail_export(export);
 940                 class_export_put(export);
 941                 export = NULL;
 942                 rc = 0;
 943         } else if (export->exp_connection != NULL &&
 944                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 945                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 946                     MSG_CONNECT_INITIAL)) {
 947                 spin_unlock(&export->exp_lock);
 948                 /* in mds failover we have static uuid but nid can be
 949                  * changed*/
 950                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 951                               "existing nid %s is already connected\n",
 952                               target->obd_name, cluuid.uuid,
 953                               libcfs_nid2str(req->rq_peer.nid),
 954                               libcfs_nid2str(
 955                                       export->exp_connection->c_peer.nid));
 956                 rc = -EALREADY;
 957                 class_export_put(export);
 958                 export = NULL;
 959         } else {
 960                 export->exp_connecting = 1;
 961                 spin_unlock(&export->exp_lock);
 962                 LASSERT(export->exp_obd == target);
 963
 964                 rc = target_handle_reconnect(&conn, export, &cluuid);
 965         }
 966
 967         /* If we found an export, we already unlocked. */
 968         if (!export) {
 969 no_export:
 970                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 971         } else if (req->rq_export == NULL &&
 972                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 973                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 974                               "still busy with %d references\n",
 975                               target->obd_name, cluuid.uuid,
 976                               libcfs_nid2str(req->rq_peer.nid),
 977                               cfs_atomic_read(&export->exp_refcount));
 978                 GOTO(out, rc = -EBUSY);
 979         } else if (req->rq_export != NULL &&
 980                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 981                 /* the current connect rpc has increased exp_rpc_count */
 982                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
 983                               "still busy with %d active RPCs\n",
 984                               target->obd_name, cluuid.uuid,
 985                               libcfs_nid2str(req->rq_peer.nid),
 986                               cfs_atomic_read(&export->exp_rpc_count) - 1);
 987                 spin_lock(&export->exp_lock);
 988                 if (req->rq_export->exp_conn_cnt <
 989                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 990                         /* try to abort active requests */
 991                         req->rq_export->exp_abort_active_req = 1;
 992                 spin_unlock(&export->exp_lock);
 993                 GOTO(out, rc = -EBUSY);
 994         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 995                 if (!strstr(cluuid.uuid, "mdt"))
 996                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 997                                       "known client %s (at %s) because it "
 998                                       "is indicating it is a new client",
 999                                       target->obd_name, cluuid.uuid,
1000                                       libcfs_nid2str(req->rq_peer.nid));
1001                 GOTO(out, rc = -EALREADY);
1002         } else {
1003                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1004         }
1005
1006         if (rc < 0) {
1007                 GOTO(out, rc);
1008         }
1009
1010         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1011                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1012               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1013               export, (long)cfs_time_current_sec(),
1014               export ? (long)export->exp_last_request_time : 0);
1015
1016         /* If this is the first time a client connects, reset the recovery
1017          * timer. Discard lightweight connections which might be local */
1018         if (!lw_client && rc == 0 && target->obd_recovering)
1019                 check_and_start_recovery_timer(target, req, export == NULL);
1020
1021         /* We want to handle EALREADY but *not* -EALREADY from
1022          * target_handle_reconnect(), return reconnection state in a flag */
1023         if (rc == EALREADY) {
1024                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1025                 rc = 0;
1026         } else {
1027                 LASSERT(rc == 0);
1028         }
1029
1030         /* Tell the client if we support replayable requests */
1031         if (target->obd_replayable)
1032                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1033         client_nid = &req->rq_peer.nid;
1034
1035         if (export == NULL) {
1036                 /* allow lightweight connections during recovery */
1037                 if (target->obd_recovering && !lw_client) {
1038                         cfs_time_t t;
1039                         int        c; /* connected */
1040                         int        i; /* in progress */
1041                         int        k; /* known */
1042
1043                         c = cfs_atomic_read(&target->obd_connected_clients);
1044                         i = cfs_atomic_read(&target->obd_lock_replay_clients);
1045                         k = target->obd_max_recoverable_clients;
1046                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1047                         t = cfs_time_sub(t, cfs_time_current());
1048                         t = cfs_duration_sec(t);
1049                         LCONSOLE_WARN("%s: Denying connection for new client "
1050                                       "%s (at %s), waiting for all %d known "
1051                                       "clients (%d recovered, %d in progress, "
1052                                       "and %d unseen) to recover in %d:%.02d\n",
1053                                       target->obd_name, cluuid.uuid,
1054                                       libcfs_nid2str(req->rq_peer.nid), k,
1055                                       c - i, i, k - c, (int)t / 60,
1056                                       (int)t % 60);
1057                         rc = -EBUSY;
1058                 } else {
1059 dont_check_exports:
1060                         rc = obd_connect(req->rq_svc_thread->t_env,
1061                                          &export, target, &cluuid, data,
1062                                          client_nid);
1063                         if (rc == 0)
1064                                 conn.cookie = export->exp_handle.h_cookie;
1065                 }
1066         } else {
1067                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1068                                    export, target, &cluuid, data, client_nid);
1069         }
1070         if (rc)
1071                 GOTO(out, rc);
1072
1073 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1074         /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1075          * will do the swabbing for if the client is using the same endianness.
1076          *
1077          * This fixup is version-limited, because we don't want to carry the
1078          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1079          * interop with unpatched 2.2 clients.  For newer clients, servers
1080          * will never do MNE swabbing, let the client handle that.  LU-1644 */
1081         export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1082                         !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1083 #else
1084 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1085 #endif
1086
1087         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1088         data->ocd_instance = target->u.obt.obt_instance;
1089
1090         /* Return only the parts of obd_connect_data that we understand, so the
1091          * client knows that we don't understand the rest. */
1092         if (data) {
1093                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1094                                                RCL_SERVER);
1095                 tmpdata = req_capsule_server_get(&req->rq_pill,
1096                                                  &RMF_CONNECT_DATA);
1097                 /* Don't use struct assignment here, because the client reply
1098                  * buffer may be smaller/larger than the local struct
1099                  * obd_connect_data. */
1100                 memcpy(tmpdata, data, min(tmpsize, size));
1101         }
1102
1103         /* If all else goes well, this is our RPC return code. */
1104         req->rq_status = 0;
1105
1106         lustre_msg_set_handle(req->rq_repmsg, &conn);
1107
1108         /* If the client and the server are the same node, we will already
1109          * have an export that really points to the client's DLM export,
1110          * because we have a shared handles table.
1111          *
1112          * XXX this will go away when shaver stops sending the "connect" handle
1113          * in the real "remote handle" field of the request --phik 24 Apr 2003
1114          */
1115         if (req->rq_export != NULL)
1116                 class_export_put(req->rq_export);
1117
1118         /* request takes one export refcount */
1119         req->rq_export = class_export_get(export);
1120
1121         spin_lock(&export->exp_lock);
1122         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1123                 spin_unlock(&export->exp_lock);
1124                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1125                        "or equal conn_cnt: %d >= %d\n",
1126                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1127                        export->exp_conn_cnt,
1128                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1129
1130                 GOTO(out, rc = -EALREADY);
1131         }
1132         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1133         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1134         export->exp_abort_active_req = 0;
1135
1136         /* request from liblustre?  Don't evict it for not pinging. */
1137         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1138                 export->exp_libclient = 1;
1139                 spin_unlock(&export->exp_lock);
1140
1141                 spin_lock(&target->obd_dev_lock);
1142                 cfs_list_del_init(&export->exp_obd_chain_timed);
1143                 spin_unlock(&target->obd_dev_lock);
1144         } else {
1145                 spin_unlock(&export->exp_lock);
1146         }
1147
1148         if (export->exp_connection != NULL) {
1149                 /* Check to see if connection came from another NID */
1150                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1151                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1152                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1153                                      &export->exp_connection->c_peer.nid,
1154                                      &export->exp_nid_hash);
1155
1156                 ptlrpc_connection_put(export->exp_connection);
1157         }
1158
1159         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1160                                                        req->rq_self,
1161                                                        &remote_uuid);
1162         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1163                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1164                              &export->exp_connection->c_peer.nid,
1165                              &export->exp_nid_hash);
1166         }
1167
1168         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1169                 int has_transno;
1170                 __u64 transno = data->ocd_transno;
1171
1172                 spin_lock(&export->exp_lock);
1173                 /* possible race with class_disconnect_stale_exports,
1174                  * export may be already in the eviction process */
1175                 if (export->exp_failed) {
1176                         spin_unlock(&export->exp_lock);
1177                         GOTO(out, rc = -ENODEV);
1178                 }
1179                 export->exp_in_recovery = 1;
1180                 export->exp_req_replay_needed = 1;
1181                 export->exp_lock_replay_needed = 1;
1182                 spin_unlock(&export->exp_lock);
1183
1184                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1185                                  MSG_CONNECT_TRANSNO);
1186                 if (has_transno && transno == 0)
1187                         CWARN("Connect with zero transno!\n");
1188
1189                 if (has_transno && transno > 0 &&
1190                     transno < target->obd_next_recovery_transno &&
1191                     transno > target->obd_last_committed) {
1192                         /* another way is to use cmpxchg() so it will be
1193                          * lock free */
1194                         spin_lock(&target->obd_recovery_task_lock);
1195                         if (transno < target->obd_next_recovery_transno)
1196                                 target->obd_next_recovery_transno = transno;
1197                         spin_unlock(&target->obd_recovery_task_lock);
1198                 }
1199
1200                 cfs_atomic_inc(&target->obd_req_replay_clients);
1201                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1202                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1203                     target->obd_max_recoverable_clients)
1204                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1205         }
1206
1207         /* Tell the client we're in recovery, when client is involved in it. */
1208         if (target->obd_recovering && !lw_client)
1209                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1210
1211         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1212         conn = *tmp;
1213
1214         /* for the rest part, we return -ENOTCONN in case of errors
1215          * in order to let client initialize connection again.
1216          */
1217         revimp = class_new_import(target);
1218         if (revimp == NULL) {
1219                 CERROR("fail to alloc new reverse import.\n");
1220                 GOTO(out, rc = -ENOTCONN);
1221         }
1222
1223         spin_lock(&export->exp_lock);
1224         if (export->exp_imp_reverse != NULL)
1225                 /* destroyed import can be still referenced in ctxt */
1226                 tmp_imp = export->exp_imp_reverse;
1227         export->exp_imp_reverse = revimp;
1228         spin_unlock(&export->exp_lock);
1229
1230         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1231         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1232         revimp->imp_remote_handle = conn;
1233         revimp->imp_dlm_fake = 1;
1234         revimp->imp_state = LUSTRE_IMP_FULL;
1235
1236         /* unknown versions will be caught in
1237          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1238         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1239
1240         if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1241             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1242                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1243         else
1244                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1245
1246         if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1247             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1248                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1249         else
1250                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1251
1252         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1253         if (rc) {
1254                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1255                 spin_lock(&export->exp_lock);
1256                 export->exp_imp_reverse = NULL;
1257                 spin_unlock(&export->exp_lock);
1258                 class_destroy_import(revimp);
1259         }
1260
1261         class_import_put(revimp);
1262
1263 out:
1264         if (tmp_imp != NULL)
1265                 client_destroy_import(tmp_imp);
1266         if (export) {
1267                 spin_lock(&export->exp_lock);
1268                 export->exp_connecting = 0;
1269                 spin_unlock(&export->exp_lock);
1270
1271                 class_export_put(export);
1272         }
1273         if (targref) {
1274                 spin_lock(&target->obd_dev_lock);
1275                 target->obd_conn_inprogress--;
1276                 spin_unlock(&target->obd_dev_lock);
1277
1278                 class_decref(targref, __func__, cfs_current());
1279         }
1280         if (rc)
1281                 req->rq_status = rc;
1282         RETURN(rc);
1283 }
1284 EXPORT_SYMBOL(target_handle_connect);
1285
1286 int target_handle_disconnect(struct ptlrpc_request *req)
1287 {
1288         int rc;
1289         ENTRY;
1290
1291         rc = req_capsule_server_pack(&req->rq_pill);
1292         if (rc)
1293                 RETURN(rc);
1294
1295         /* keep the rq_export around so we can send the reply */
1296         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1297
1298         RETURN(0);
1299 }
1300 EXPORT_SYMBOL(target_handle_disconnect);
1301
1302 void target_destroy_export(struct obd_export *exp)
1303 {
1304         struct obd_import       *imp = NULL;
1305         /* exports created from last_rcvd data, and "fake"
1306            exports created by lctl don't have an import */
1307         spin_lock(&exp->exp_lock);
1308         if (exp->exp_imp_reverse != NULL) {
1309                 imp = exp->exp_imp_reverse;
1310                 exp->exp_imp_reverse = NULL;
1311         }
1312         spin_unlock(&exp->exp_lock);
1313         if (imp != NULL)
1314                 client_destroy_import(imp);
1315
1316         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1317         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1318         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1319         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1320 }
1321 EXPORT_SYMBOL(target_destroy_export);
1322
1323 /*
1324  * Recovery functions
1325  */
1326 static void target_request_copy_get(struct ptlrpc_request *req)
1327 {
1328         class_export_rpc_get(req->rq_export);
1329         LASSERT(cfs_list_empty(&req->rq_list));
1330         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1331
1332         /* increase refcount to keep request in queue */
1333         cfs_atomic_inc(&req->rq_refcount);
1334         /** let export know it has replays to be handled */
1335         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1336 }
1337
1338 static void target_request_copy_put(struct ptlrpc_request *req)
1339 {
1340         LASSERT(cfs_list_empty(&req->rq_replay_list));
1341         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1342
1343         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1344         class_export_rpc_put(req->rq_export);
1345         ptlrpc_server_drop_request(req);
1346 }
1347
1348 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1349 {
1350         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1351         struct obd_export     *exp = req->rq_export;
1352         struct ptlrpc_request *reqiter;
1353         int                    dup = 0;
1354
1355         LASSERT(exp);
1356
1357         spin_lock(&exp->exp_lock);
1358         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1359                                 rq_replay_list) {
1360                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1361                         dup = 1;
1362                         break;
1363                 }
1364         }
1365
1366         if (dup) {
1367                 /* we expect it with RESENT and REPLAY flags */
1368                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1369                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1370                         CERROR("invalid flags %x of resent replay\n",
1371                                lustre_msg_get_flags(req->rq_reqmsg));
1372         } else {
1373                 cfs_list_add_tail(&req->rq_replay_list,
1374                                   &exp->exp_req_replay_queue);
1375         }
1376
1377         spin_unlock(&exp->exp_lock);
1378         return dup;
1379 }
1380
1381 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1382 {
1383         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1384         LASSERT(req->rq_export);
1385
1386         spin_lock(&req->rq_export->exp_lock);
1387         cfs_list_del_init(&req->rq_replay_list);
1388         spin_unlock(&req->rq_export->exp_lock);
1389 }
1390
1391 #ifdef __KERNEL__
1392 static void target_finish_recovery(struct obd_device *obd)
1393 {
1394         ENTRY;
1395
1396         /* only log a recovery message when recovery has occurred */
1397         if (obd->obd_recovery_start) {
1398                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1399                                         obd->obd_recovery_start);
1400                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1401                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1402                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1403                         obd->obd_max_recoverable_clients,
1404                         cfs_atomic_read(&obd->obd_connected_clients),
1405                         obd->obd_stale_clients,
1406                         obd->obd_stale_clients == 1 ? "was" : "were");
1407         }
1408
1409         ldlm_reprocess_all_ns(obd->obd_namespace);
1410         spin_lock(&obd->obd_recovery_task_lock);
1411         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1412             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1413             !cfs_list_empty(&obd->obd_final_req_queue)) {
1414                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1415                        obd->obd_name,
1416                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1417                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1418                                "" : "lock ",
1419                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1420                                "" : "final ");
1421                 spin_unlock(&obd->obd_recovery_task_lock);
1422                 LBUG();
1423         }
1424         spin_unlock(&obd->obd_recovery_task_lock);
1425
1426         obd->obd_recovery_end = cfs_time_current_sec();
1427
1428         /* when recovery finished, cleanup orphans on mds and ost */
1429         if (OBT(obd) && OBP(obd, postrecov)) {
1430                 int rc = OBP(obd, postrecov)(obd);
1431                 if (rc < 0)
1432                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1433                                       obd->obd_name, rc);
1434         }
1435         EXIT;
1436 }
1437
1438 static void abort_req_replay_queue(struct obd_device *obd)
1439 {
1440         struct ptlrpc_request *req, *n;
1441         cfs_list_t abort_list;
1442
1443         CFS_INIT_LIST_HEAD(&abort_list);
1444         spin_lock(&obd->obd_recovery_task_lock);
1445         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1446         spin_unlock(&obd->obd_recovery_task_lock);
1447         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1448                 DEBUG_REQ(D_WARNING, req, "aborted:");
1449                 req->rq_status = -ENOTCONN;
1450                 if (ptlrpc_error(req)) {
1451                         DEBUG_REQ(D_ERROR, req,
1452                                   "failed abort_req_reply; skipping");
1453                 }
1454                 target_exp_dequeue_req_replay(req);
1455                 target_request_copy_put(req);
1456         }
1457 }
1458
1459 static void abort_lock_replay_queue(struct obd_device *obd)
1460 {
1461         struct ptlrpc_request *req, *n;
1462         cfs_list_t abort_list;
1463
1464         CFS_INIT_LIST_HEAD(&abort_list);
1465         spin_lock(&obd->obd_recovery_task_lock);
1466         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1467         spin_unlock(&obd->obd_recovery_task_lock);
1468         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1469                 DEBUG_REQ(D_ERROR, req, "aborted:");
1470                 req->rq_status = -ENOTCONN;
1471                 if (ptlrpc_error(req)) {
1472                         DEBUG_REQ(D_ERROR, req,
1473                                   "failed abort_lock_reply; skipping");
1474                 }
1475                 target_request_copy_put(req);
1476         }
1477 }
1478
1479 /* Called from a cleanup function if the device is being cleaned up
1480    forcefully.  The exports should all have been disconnected already,
1481    the only thing left to do is
1482      - clear the recovery flags
1483      - cancel the timer
1484      - free queued requests and replies, but don't send replies
1485    Because the obd_stopping flag is set, no new requests should be received.
1486
1487 */
1488 void target_cleanup_recovery(struct obd_device *obd)
1489 {
1490         struct ptlrpc_request *req, *n;
1491         cfs_list_t clean_list;
1492         ENTRY;
1493
1494         CFS_INIT_LIST_HEAD(&clean_list);
1495         spin_lock(&obd->obd_dev_lock);
1496         if (!obd->obd_recovering) {
1497                 spin_unlock(&obd->obd_dev_lock);
1498                 EXIT;
1499                 return;
1500         }
1501         obd->obd_recovering = obd->obd_abort_recovery = 0;
1502         spin_unlock(&obd->obd_dev_lock);
1503
1504         spin_lock(&obd->obd_recovery_task_lock);
1505         target_cancel_recovery_timer(obd);
1506         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1507         spin_unlock(&obd->obd_recovery_task_lock);
1508
1509         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1510                 LASSERT(req->rq_reply_state == 0);
1511                 target_exp_dequeue_req_replay(req);
1512                 target_request_copy_put(req);
1513         }
1514
1515         spin_lock(&obd->obd_recovery_task_lock);
1516         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1517         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1518         spin_unlock(&obd->obd_recovery_task_lock);
1519
1520         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1521                 LASSERT(req->rq_reply_state == 0);
1522                 target_request_copy_put(req);
1523         }
1524
1525         EXIT;
1526 }
1527 EXPORT_SYMBOL(target_cleanup_recovery);
1528
1529 /* obd_recovery_task_lock should be held */
1530 void target_cancel_recovery_timer(struct obd_device *obd)
1531 {
1532         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1533         cfs_timer_disarm(&obd->obd_recovery_timer);
1534 }
1535 EXPORT_SYMBOL(target_cancel_recovery_timer);
1536
1537 static void target_start_recovery_timer(struct obd_device *obd)
1538 {
1539         if (obd->obd_recovery_start != 0)
1540                 return;
1541
1542         spin_lock(&obd->obd_dev_lock);
1543         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1544                 spin_unlock(&obd->obd_dev_lock);
1545                 return;
1546         }
1547
1548         LASSERT(obd->obd_recovery_timeout != 0);
1549
1550         if (obd->obd_recovery_start != 0) {
1551                 spin_unlock(&obd->obd_dev_lock);
1552                 return;
1553         }
1554
1555         cfs_timer_arm(&obd->obd_recovery_timer,
1556                       cfs_time_shift(obd->obd_recovery_timeout));
1557         obd->obd_recovery_start = cfs_time_current_sec();
1558         spin_unlock(&obd->obd_dev_lock);
1559
1560         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1561                       "or until %d client%s reconnect%s\n",
1562                       obd->obd_name,
1563                       obd->obd_recovery_timeout / 60,
1564                       obd->obd_recovery_timeout % 60,
1565                       obd->obd_max_recoverable_clients,
1566                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1567                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1568 }
1569
1570 /**
1571  * extend recovery window.
1572  *
1573  * if @extend is true, extend recovery window to have @drt remaining at least;
1574  * otherwise, make sure the recovery timeout value is not less than @drt.
1575  */
1576 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1577 {
1578         cfs_time_t now;
1579         cfs_time_t end;
1580         cfs_duration_t left;
1581         int to;
1582
1583         spin_lock(&obd->obd_dev_lock);
1584         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1585                 spin_unlock(&obd->obd_dev_lock);
1586                 return;
1587         }
1588         LASSERT(obd->obd_recovery_start != 0);
1589
1590         now  = cfs_time_current_sec();
1591         to   = obd->obd_recovery_timeout;
1592         end  = obd->obd_recovery_start + to;
1593         left = cfs_time_sub(end, now);
1594
1595         if (extend && (drt > left)) {
1596                 to += drt - left;
1597         } else if (!extend && (drt > to)) {
1598                 to = drt;
1599                 /* reduce drt by already passed time */
1600                 drt -= obd->obd_recovery_timeout - left;
1601         }
1602
1603         if (to > obd->obd_recovery_time_hard)
1604                 to = obd->obd_recovery_time_hard;
1605         if (obd->obd_recovery_timeout < to) {
1606                 obd->obd_recovery_timeout = to;
1607                 cfs_timer_arm(&obd->obd_recovery_timer,
1608                               cfs_time_shift(drt));
1609         }
1610         spin_unlock(&obd->obd_dev_lock);
1611
1612         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1613                obd->obd_name, (unsigned)drt);
1614 }
1615
1616 /* Reset the timer with each new client connection */
1617 /*
1618  * This timer is actually reconnect_timer, which is for making sure
1619  * the total recovery window is at least as big as my reconnect
1620  * attempt timing. So the initial recovery time_out will be set to
1621  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1622  * from client is bigger than this, then the recovery time_out will
1623  * be extended to make sure the client could be reconnected, in the
1624  * process, the timeout from the new client should be ignored.
1625  */
1626
1627 static void
1628 check_and_start_recovery_timer(struct obd_device *obd,
1629                                struct ptlrpc_request *req,
1630                                int new_client)
1631 {
1632         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1633         struct obd_device_target *obt = &obd->u.obt;
1634         struct lustre_sb_info *lsi;
1635
1636         if (!new_client && service_time)
1637                 /* Teach server about old server's estimates, as first guess
1638                  * at how long new requests will take. */
1639                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1640                             service_time);
1641
1642         target_start_recovery_timer(obd);
1643
1644         /* convert the service time to rpc timeout,
1645          * reuse service_time to limit stack usage */
1646         service_time = at_est2timeout(service_time);
1647
1648         /* We expect other clients to timeout within service_time, then try
1649          * to reconnect, then try the failover server.  The max delay between
1650          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1651         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1652
1653         LASSERT(obt->obt_magic == OBT_MAGIC);
1654         lsi = s2lsi(obt->obt_sb);
1655         if (!(lsi->lsi_flags | LDD_F_IR_CAPABLE))
1656                 service_time += 2 * (CONNECTION_SWITCH_MAX +
1657                                      CONNECTION_SWITCH_INC);
1658         if (service_time > obd->obd_recovery_timeout && !new_client)
1659                 extend_recovery_timer(obd, service_time, false);
1660 }
1661
1662 /** Health checking routines */
1663 static inline int exp_connect_healthy(struct obd_export *exp)
1664 {
1665         return (exp->exp_in_recovery);
1666 }
1667
1668 /** if export done req_replay or has replay in queue */
1669 static inline int exp_req_replay_healthy(struct obd_export *exp)
1670 {
1671         return (!exp->exp_req_replay_needed ||
1672                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1673 }
1674 /** if export done lock_replay or has replay in queue */
1675 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1676 {
1677         return (!exp->exp_lock_replay_needed ||
1678                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1679 }
1680
1681 static inline int exp_vbr_healthy(struct obd_export *exp)
1682 {
1683         return (!exp->exp_vbr_failed);
1684 }
1685
1686 static inline int exp_finished(struct obd_export *exp)
1687 {
1688         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1689 }
1690
1691 /** Checking routines for recovery */
1692 static int check_for_clients(struct obd_device *obd)
1693 {
1694         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1695
1696         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1697                 return 1;
1698         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1699         if (obd->obd_no_conn == 0 &&
1700             clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1701                 return 1;
1702         return 0;
1703 }
1704
1705 static int check_for_next_transno(struct obd_device *obd)
1706 {
1707         struct ptlrpc_request *req = NULL;
1708         int wake_up = 0, connected, completed, queue_len;
1709         __u64 next_transno, req_transno;
1710         ENTRY;
1711
1712         spin_lock(&obd->obd_recovery_task_lock);
1713         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1714                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1715                                      struct ptlrpc_request, rq_list);
1716                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1717         } else {
1718                 req_transno = 0;
1719         }
1720
1721         connected = cfs_atomic_read(&obd->obd_connected_clients);
1722         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1723         queue_len = obd->obd_requests_queued_for_recovery;
1724         next_transno = obd->obd_next_recovery_transno;
1725
1726         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1727                "req_transno: "LPU64", next_transno: "LPU64"\n",
1728                obd->obd_max_recoverable_clients, connected, completed,
1729                queue_len, req_transno, next_transno);
1730
1731         if (obd->obd_abort_recovery) {
1732                 CDEBUG(D_HA, "waking for aborted recovery\n");
1733                 wake_up = 1;
1734         } else if (obd->obd_recovery_expired) {
1735                 CDEBUG(D_HA, "waking for expired recovery\n");
1736                 wake_up = 1;
1737         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1738                 CDEBUG(D_HA, "waking for completed recovery\n");
1739                 wake_up = 1;
1740         } else if (req_transno == next_transno) {
1741                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1742                 wake_up = 1;
1743         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1744                 int d_lvl = D_HA;
1745                 /** handle gaps occured due to lost reply or VBR */
1746                 LASSERTF(req_transno >= next_transno,
1747                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1748                          req_transno, next_transno);
1749                 if (req_transno > obd->obd_last_committed &&
1750                     !obd->obd_version_recov)
1751                         d_lvl = D_ERROR;
1752                 CDEBUG(d_lvl,
1753                        "%s: waking for gap in transno, VBR is %s (skip: "
1754                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1755                        ", last_committed: "LPD64")\n",
1756                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1757                        next_transno, queue_len, completed, connected,
1758                        req_transno, obd->obd_last_committed);
1759                 obd->obd_next_recovery_transno = req_transno;
1760                 wake_up = 1;
1761         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1762                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1763                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1764                 obd->obd_next_recovery_transno = req_transno;
1765                 wake_up = 1;
1766         }
1767         spin_unlock(&obd->obd_recovery_task_lock);
1768         return wake_up;
1769 }
1770
1771 static int check_for_next_lock(struct obd_device *obd)
1772 {
1773         int wake_up = 0;
1774
1775         spin_lock(&obd->obd_recovery_task_lock);
1776         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1777                 CDEBUG(D_HA, "waking for next lock\n");
1778                 wake_up = 1;
1779         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1780                 CDEBUG(D_HA, "waking for completed lock replay\n");
1781                 wake_up = 1;
1782         } else if (obd->obd_abort_recovery) {
1783                 CDEBUG(D_HA, "waking for aborted recovery\n");
1784                 wake_up = 1;
1785         } else if (obd->obd_recovery_expired) {
1786                 CDEBUG(D_HA, "waking for expired recovery\n");
1787                 wake_up = 1;
1788         }
1789         spin_unlock(&obd->obd_recovery_task_lock);
1790
1791         return wake_up;
1792 }
1793
1794 /**
1795  * wait for recovery events,
1796  * check its status with help of check_routine
1797  * evict dead clients via health_check
1798  */
1799 static int target_recovery_overseer(struct obd_device *obd,
1800                                     int (*check_routine)(struct obd_device *),
1801                                     int (*health_check)(struct obd_export *))
1802 {
1803 repeat:
1804         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1805         if (obd->obd_abort_recovery) {
1806                 CWARN("recovery is aborted, evict exports in recovery\n");
1807                 /** evict exports which didn't finish recovery yet */
1808                 class_disconnect_stale_exports(obd, exp_finished);
1809                 return 1;
1810         } else if (obd->obd_recovery_expired) {
1811                 obd->obd_recovery_expired = 0;
1812                 /** If some clients died being recovered, evict them */
1813                 LCONSOLE_WARN("%s: recovery is timed out, "
1814                               "evict stale exports\n", obd->obd_name);
1815                 /** evict cexports with no replay in queue, they are stalled */
1816                 class_disconnect_stale_exports(obd, health_check);
1817                 /** continue with VBR */
1818                 spin_lock(&obd->obd_dev_lock);
1819                 obd->obd_version_recov = 1;
1820                 spin_unlock(&obd->obd_dev_lock);
1821                 /**
1822                  * reset timer, recovery will proceed with versions now,
1823                  * timeout is set just to handle reconnection delays
1824                  */
1825                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1826                 /** Wait for recovery events again, after evicting bad clients */
1827                 goto repeat;
1828         }
1829         return 0;
1830 }
1831
1832 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1833 {
1834         struct ptlrpc_request *req = NULL;
1835         ENTRY;
1836
1837         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1838                obd->obd_next_recovery_transno);
1839
1840         if (target_recovery_overseer(obd, check_for_next_transno,
1841                                      exp_req_replay_healthy)) {
1842                 abort_req_replay_queue(obd);
1843                 abort_lock_replay_queue(obd);
1844         }
1845
1846         spin_lock(&obd->obd_recovery_task_lock);
1847         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1848                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1849                                      struct ptlrpc_request, rq_list);
1850                 cfs_list_del_init(&req->rq_list);
1851                 obd->obd_requests_queued_for_recovery--;
1852                 spin_unlock(&obd->obd_recovery_task_lock);
1853         } else {
1854                 spin_unlock(&obd->obd_recovery_task_lock);
1855                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1856                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1857                 /** evict exports failed VBR */
1858                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1859         }
1860         RETURN(req);
1861 }
1862
1863 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1864 {
1865         struct ptlrpc_request *req = NULL;
1866
1867         CDEBUG(D_HA, "Waiting for lock\n");
1868         if (target_recovery_overseer(obd, check_for_next_lock,
1869                                      exp_lock_replay_healthy))
1870                 abort_lock_replay_queue(obd);
1871
1872         spin_lock(&obd->obd_recovery_task_lock);
1873         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1874                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1875                                      struct ptlrpc_request, rq_list);
1876                 cfs_list_del_init(&req->rq_list);
1877                 spin_unlock(&obd->obd_recovery_task_lock);
1878         } else {
1879                 spin_unlock(&obd->obd_recovery_task_lock);
1880                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1881                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1882                 /** evict exports failed VBR */
1883                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1884         }
1885         return req;
1886 }
1887
1888 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1889 {
1890         struct ptlrpc_request *req = NULL;
1891
1892         spin_lock(&obd->obd_recovery_task_lock);
1893         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1894                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1895                                      struct ptlrpc_request, rq_list);
1896                 cfs_list_del_init(&req->rq_list);
1897                 spin_unlock(&obd->obd_recovery_task_lock);
1898                 if (req->rq_export->exp_in_recovery) {
1899                         spin_lock(&req->rq_export->exp_lock);
1900                         req->rq_export->exp_in_recovery = 0;
1901                         spin_unlock(&req->rq_export->exp_lock);
1902                 }
1903         } else {
1904                 spin_unlock(&obd->obd_recovery_task_lock);
1905         }
1906         return req;
1907 }
1908
1909 static int handle_recovery_req(struct ptlrpc_thread *thread,
1910                                struct ptlrpc_request *req,
1911                                svc_handler_t handler)
1912 {
1913         int rc;
1914         ENTRY;
1915
1916         /**
1917          * export can be evicted during recovery, no need to handle replays for
1918          * it after that, discard such request silently
1919          */
1920         if (req->rq_export->exp_disconnected)
1921                 GOTO(reqcopy_put, rc = 0);
1922
1923         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1924         if (rc) {
1925                 CERROR("Failure to initialize session: %d\n", rc);
1926                 GOTO(reqcopy_put, rc);
1927         }
1928
1929         req->rq_recov_session.lc_thread = thread;
1930         lu_context_enter(&req->rq_recov_session);
1931         req->rq_svc_thread = thread;
1932         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1933
1934         /* thread context */
1935         lu_context_enter(&thread->t_env->le_ctx);
1936         (void)handler(req);
1937         lu_context_exit(&thread->t_env->le_ctx);
1938
1939         lu_context_exit(&req->rq_recov_session);
1940         lu_context_fini(&req->rq_recov_session);
1941         /* don't reset timer for final stage */
1942         if (!exp_finished(req->rq_export)) {
1943                 int to = obd_timeout;
1944
1945                 /**
1946                  * Add request timeout to the recovery time so next request from
1947                  * this client may come in recovery time
1948                  */
1949                 if (!AT_OFF) {
1950                         struct ptlrpc_service_part *svcpt;
1951
1952                         svcpt = req->rq_rqbd->rqbd_svcpt;
1953                         /* If the server sent early reply for this request,
1954                          * the client will recalculate the timeout according to
1955                          * current server estimate service time, so we will
1956                          * use the maxium timeout here for waiting the client
1957                          * sending the next req */
1958                         to = max((int)at_est2timeout(
1959                                  at_get(&svcpt->scp_at_estimate)),
1960                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1961                         /* Add 2 net_latency, one for balance rq_deadline
1962                          * (see ptl_send_rpc), one for resend the req to server,
1963                          * Note: client will pack net_latency in replay req
1964                          * (see ptlrpc_replay_req) */
1965                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1966                 }
1967                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1968         }
1969 reqcopy_put:
1970         RETURN(rc);
1971 }
1972
1973 static int target_recovery_thread(void *arg)
1974 {
1975         struct lu_target *lut = arg;
1976         struct obd_device *obd = lut->lut_obd;
1977         struct ptlrpc_request *req;
1978         struct target_recovery_data *trd = &obd->obd_recovery_data;
1979         unsigned long delta;
1980         unsigned long flags;
1981         struct lu_env *env;
1982         struct ptlrpc_thread *thread = NULL;
1983         int rc = 0;
1984         ENTRY;
1985
1986         cfs_daemonize_ctxt("tgt_recov");
1987
1988         SIGNAL_MASK_LOCK(current, flags);
1989         sigfillset(&current->blocked);
1990         RECALC_SIGPENDING;
1991         SIGNAL_MASK_UNLOCK(current, flags);
1992
1993         OBD_ALLOC_PTR(thread);
1994         if (thread == NULL)
1995                 RETURN(-ENOMEM);
1996
1997         OBD_ALLOC_PTR(env);
1998         if (env == NULL) {
1999                 OBD_FREE_PTR(thread);
2000                 RETURN(-ENOMEM);
2001         }
2002
2003         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
2004         if (rc) {
2005                 OBD_FREE_PTR(thread);
2006                 OBD_FREE_PTR(env);
2007                 RETURN(rc);
2008         }
2009
2010         thread->t_env = env;
2011         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2012         env->le_ctx.lc_thread = thread;
2013         thread->t_data = NULL;
2014         thread->t_watchdog = NULL;
2015
2016         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2017                cfs_curproc_pid());
2018         trd->trd_processing_task = cfs_curproc_pid();
2019
2020         spin_lock(&obd->obd_dev_lock);
2021         obd->obd_recovering = 1;
2022         spin_unlock(&obd->obd_dev_lock);
2023         complete(&trd->trd_starting);
2024
2025         /* first of all, we have to know the first transno to replay */
2026         if (target_recovery_overseer(obd, check_for_clients,
2027                                      exp_connect_healthy)) {
2028                 abort_req_replay_queue(obd);
2029                 abort_lock_replay_queue(obd);
2030         }
2031
2032         /* next stage: replay requests */
2033         delta = jiffies;
2034         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2035                cfs_atomic_read(&obd->obd_req_replay_clients),
2036                obd->obd_next_recovery_transno);
2037         while ((req = target_next_replay_req(obd))) {
2038                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2039                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2040                           lustre_msg_get_transno(req->rq_reqmsg),
2041                           libcfs_nid2str(req->rq_peer.nid));
2042                 handle_recovery_req(thread, req,
2043                                     trd->trd_recovery_handler);
2044                 /**
2045                  * bz18031: increase next_recovery_transno before
2046                  * target_request_copy_put() will drop exp_rpc reference
2047                  */
2048                 spin_lock(&obd->obd_recovery_task_lock);
2049                 obd->obd_next_recovery_transno++;
2050                 spin_unlock(&obd->obd_recovery_task_lock);
2051                 target_exp_dequeue_req_replay(req);
2052                 target_request_copy_put(req);
2053                 obd->obd_replayed_requests++;
2054         }
2055
2056         /**
2057          * The second stage: replay locks
2058          */
2059         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2060                cfs_atomic_read(&obd->obd_lock_replay_clients));
2061         while ((req = target_next_replay_lock(obd))) {
2062                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2063                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2064                           libcfs_nid2str(req->rq_peer.nid));
2065                 handle_recovery_req(thread, req,
2066                                     trd->trd_recovery_handler);
2067                 target_request_copy_put(req);
2068                 obd->obd_replayed_locks++;
2069         }
2070
2071         /**
2072          * The third stage: reply on final pings, at this moment all clients
2073          * must have request in final queue
2074          */
2075         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2076         /** Update server last boot epoch */
2077         tgt_boot_epoch_update(lut);
2078         /* We drop recoverying flag to forward all new requests
2079          * to regular mds_handle() since now */
2080         spin_lock(&obd->obd_dev_lock);
2081         obd->obd_recovering = obd->obd_abort_recovery = 0;
2082         spin_unlock(&obd->obd_dev_lock);
2083         spin_lock(&obd->obd_recovery_task_lock);
2084         target_cancel_recovery_timer(obd);
2085         spin_unlock(&obd->obd_recovery_task_lock);
2086         while ((req = target_next_final_ping(obd))) {
2087                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2088                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2089                           libcfs_nid2str(req->rq_peer.nid));
2090                 handle_recovery_req(thread, req,
2091                                     trd->trd_recovery_handler);
2092                 target_request_copy_put(req);
2093         }
2094
2095         delta = (jiffies - delta) / CFS_HZ;
2096         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2097               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2098         if (delta > OBD_RECOVERY_TIME_SOFT) {
2099                 CWARN("too long recovery - read logs\n");
2100                 libcfs_debug_dumplog();
2101         }
2102
2103         target_finish_recovery(obd);
2104
2105         lu_context_fini(&env->le_ctx);
2106         trd->trd_processing_task = 0;
2107         complete(&trd->trd_finishing);
2108
2109         OBD_FREE_PTR(thread);
2110         OBD_FREE_PTR(env);
2111         RETURN(rc);
2112 }
2113
2114 static int target_start_recovery_thread(struct lu_target *lut,
2115                                         svc_handler_t handler)
2116 {
2117         struct obd_device *obd = lut->lut_obd;
2118         int rc = 0;
2119         struct target_recovery_data *trd = &obd->obd_recovery_data;
2120
2121         memset(trd, 0, sizeof(*trd));
2122         init_completion(&trd->trd_starting);
2123         init_completion(&trd->trd_finishing);
2124         trd->trd_recovery_handler = handler;
2125
2126         if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2127                 wait_for_completion(&trd->trd_starting);
2128                 LASSERT(obd->obd_recovering != 0);
2129         } else
2130                 rc = -ECHILD;
2131
2132         return rc;
2133 }
2134
2135 void target_stop_recovery_thread(struct obd_device *obd)
2136 {
2137         if (obd->obd_recovery_data.trd_processing_task > 0) {
2138                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2139                 /** recovery can be done but postrecovery is not yet */
2140                 spin_lock(&obd->obd_dev_lock);
2141                 if (obd->obd_recovering) {
2142                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2143                         obd->obd_abort_recovery = 1;
2144                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2145                 }
2146                 spin_unlock(&obd->obd_dev_lock);
2147                 wait_for_completion(&trd->trd_finishing);
2148         }
2149 }
2150 EXPORT_SYMBOL(target_stop_recovery_thread);
2151
2152 void target_recovery_fini(struct obd_device *obd)
2153 {
2154         class_disconnect_exports(obd);
2155         target_stop_recovery_thread(obd);
2156         target_cleanup_recovery(obd);
2157 }
2158 EXPORT_SYMBOL(target_recovery_fini);
2159
2160 static void target_recovery_expired(unsigned long castmeharder)
2161 {
2162         struct obd_device *obd = (struct obd_device *)castmeharder;
2163         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2164                " after %lds (%d clients connected)\n",
2165                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2166                cfs_time_current_sec()- obd->obd_recovery_start,
2167                cfs_atomic_read(&obd->obd_connected_clients));
2168
2169         obd->obd_recovery_expired = 1;
2170         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2171 }
2172
2173 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2174 {
2175         struct obd_device *obd = lut->lut_obd;
2176         if (obd->obd_max_recoverable_clients == 0) {
2177                 /** Update server last boot epoch */
2178                 tgt_boot_epoch_update(lut);
2179                 return;
2180         }
2181
2182         CWARN("RECOVERY: service %s, %d recoverable clients, "
2183               "last_transno "LPU64"\n", obd->obd_name,
2184               obd->obd_max_recoverable_clients, obd->obd_last_committed);
2185         LASSERT(obd->obd_stopping == 0);
2186         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2187         obd->obd_recovery_start = 0;
2188         obd->obd_recovery_end = 0;
2189
2190         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2191         target_start_recovery_thread(lut, handler);
2192 }
2193 EXPORT_SYMBOL(target_recovery_init);
2194
2195 #endif /* __KERNEL__ */
2196
2197 static int target_process_req_flags(struct obd_device *obd,
2198                                     struct ptlrpc_request *req)
2199 {
2200         struct obd_export *exp = req->rq_export;
2201         LASSERT(exp != NULL);
2202         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2203                 /* client declares he's ready to replay locks */
2204                 spin_lock(&exp->exp_lock);
2205                 if (exp->exp_req_replay_needed) {
2206                         exp->exp_req_replay_needed = 0;
2207                         spin_unlock(&exp->exp_lock);
2208
2209                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2210                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2211                 } else {
2212                         spin_unlock(&exp->exp_lock);
2213                 }
2214         }
2215         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2216                 /* client declares he's ready to complete recovery
2217                  * so, we put the request on th final queue */
2218                 spin_lock(&exp->exp_lock);
2219                 if (exp->exp_lock_replay_needed) {
2220                         exp->exp_lock_replay_needed = 0;
2221                         spin_unlock(&exp->exp_lock);
2222
2223                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2224                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2225                 } else {
2226                         spin_unlock(&exp->exp_lock);
2227                 }
2228         }
2229         return 0;
2230 }
2231
2232 int target_queue_recovery_request(struct ptlrpc_request *req,
2233                                   struct obd_device *obd)
2234 {
2235         cfs_list_t *tmp;
2236         int inserted = 0;
2237         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2238         ENTRY;
2239
2240         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2241                 /* Processing the queue right now, don't re-add. */
2242                 RETURN(1);
2243         }
2244
2245         target_process_req_flags(obd, req);
2246
2247         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2248                 /* client declares he's ready to complete recovery
2249                  * so, we put the request on th final queue */
2250                 target_request_copy_get(req);
2251                 DEBUG_REQ(D_HA, req, "queue final req");
2252                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2253                 spin_lock(&obd->obd_recovery_task_lock);
2254                 if (obd->obd_recovering) {
2255                         cfs_list_add_tail(&req->rq_list,
2256                                           &obd->obd_final_req_queue);
2257                 } else {
2258                         spin_unlock(&obd->obd_recovery_task_lock);
2259                         target_request_copy_put(req);
2260                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2261                 }
2262                 spin_unlock(&obd->obd_recovery_task_lock);
2263                 RETURN(0);
2264         }
2265         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2266                 /* client declares he's ready to replay locks */
2267                 target_request_copy_get(req);
2268                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2269                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2270                 spin_lock(&obd->obd_recovery_task_lock);
2271                 LASSERT(obd->obd_recovering);
2272                 /* usually due to recovery abort */
2273                 if (!req->rq_export->exp_in_recovery) {
2274                         spin_unlock(&obd->obd_recovery_task_lock);
2275                         target_request_copy_put(req);
2276                         RETURN(-ENOTCONN);
2277                 }
2278                 LASSERT(req->rq_export->exp_lock_replay_needed);
2279                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2280                 spin_unlock(&obd->obd_recovery_task_lock);
2281                 RETURN(0);
2282         }
2283
2284         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2285          * (i.e. buflens etc are in my own byte order), but type-dependent
2286          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2287
2288         if (!transno) {
2289                 CFS_INIT_LIST_HEAD(&req->rq_list);
2290                 DEBUG_REQ(D_HA, req, "not queueing");
2291                 RETURN(1);
2292         }
2293
2294         /* If we're processing the queue, we want don't want to queue this
2295          * message.
2296          *
2297          * Also, if this request has a transno less than the one we're waiting
2298          * for, we should process it now.  It could (and currently always will)
2299          * be an open request for a descriptor that was opened some time ago.
2300          *
2301          * Also, a resent, replayed request that has already been
2302          * handled will pass through here and be processed immediately.
2303          */
2304         CDEBUG(D_HA, "Next recovery transno: "LPU64
2305                ", current: "LPU64", replaying\n",
2306                obd->obd_next_recovery_transno, transno);
2307         spin_lock(&obd->obd_recovery_task_lock);
2308         if (transno < obd->obd_next_recovery_transno) {
2309                 /* Processing the queue right now, don't re-add. */
2310                 LASSERT(cfs_list_empty(&req->rq_list));
2311                 spin_unlock(&obd->obd_recovery_task_lock);
2312                 RETURN(1);
2313         }
2314         spin_unlock(&obd->obd_recovery_task_lock);
2315
2316         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2317                 RETURN(0);
2318
2319         target_request_copy_get(req);
2320         if (!req->rq_export->exp_in_recovery) {
2321                 target_request_copy_put(req);
2322                 RETURN(-ENOTCONN);
2323         }
2324         LASSERT(req->rq_export->exp_req_replay_needed);
2325
2326         if (target_exp_enqueue_req_replay(req)) {
2327                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2328                 target_request_copy_put(req);
2329                 RETURN(0);
2330         }
2331
2332         /* XXX O(n^2) */
2333         spin_lock(&obd->obd_recovery_task_lock);
2334         LASSERT(obd->obd_recovering);
2335         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2336                 struct ptlrpc_request *reqiter =
2337                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2338
2339                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2340                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2341                         inserted = 1;
2342                         break;
2343                 }
2344
2345                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2346                              transno)) {
2347                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2348                                   "has been claimed by another client");
2349                         spin_unlock(&obd->obd_recovery_task_lock);
2350                         target_exp_dequeue_req_replay(req);
2351                         target_request_copy_put(req);
2352                         RETURN(0);
2353                 }
2354         }
2355
2356         if (!inserted)
2357                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2358
2359         obd->obd_requests_queued_for_recovery++;
2360         spin_unlock(&obd->obd_recovery_task_lock);
2361         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2362         RETURN(0);
2363 }
2364 EXPORT_SYMBOL(target_queue_recovery_request);
2365
2366 int target_handle_ping(struct ptlrpc_request *req)
2367 {
2368         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2369         return req_capsule_server_pack(&req->rq_pill);
2370 }
2371 EXPORT_SYMBOL(target_handle_ping);
2372
2373 void target_committed_to_req(struct ptlrpc_request *req)
2374 {
2375         struct obd_export *exp = req->rq_export;
2376
2377         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2378                 lustre_msg_set_last_committed(req->rq_repmsg,
2379                                               exp->exp_last_committed);
2380         else
2381                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2382                           "%d)", exp->exp_obd->obd_no_transno,
2383                           req->rq_repmsg == NULL);
2384
2385         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2386                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2387 }
2388 EXPORT_SYMBOL(target_committed_to_req);
2389
2390 #endif /* HAVE_SERVER_SUPPORT */
2391
2392 /**
2393  * Packs current SLV and Limit into \a req.
2394  */
2395 int target_pack_pool_reply(struct ptlrpc_request *req)
2396 {
2397         struct obd_device *obd;
2398         ENTRY;
2399
2400         /*
2401          * Check that we still have all structures alive as this may
2402          * be some late rpc in shutdown time.
2403          */
2404         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2405                      !exp_connect_lru_resize(req->rq_export))) {
2406                 lustre_msg_set_slv(req->rq_repmsg, 0);
2407                 lustre_msg_set_limit(req->rq_repmsg, 0);
2408                 RETURN(0);
2409         }
2410
2411         /*
2412          * OBD is alive here as export is alive, which we checked above.
2413          */
2414         obd = req->rq_export->exp_obd;
2415
2416         read_lock(&obd->obd_pool_lock);
2417         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2418         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2419         read_unlock(&obd->obd_pool_lock);
2420
2421         RETURN(0);
2422 }
2423 EXPORT_SYMBOL(target_pack_pool_reply);
2424
2425 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2426 {
2427         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2428                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2429                 return (-ECOMM);
2430         }
2431
2432         if (unlikely(rc)) {
2433                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2434                 req->rq_status = rc;
2435                 return (ptlrpc_send_error(req, 1));
2436         } else {
2437                 DEBUG_REQ(D_NET, req, "sending reply");
2438         }
2439
2440         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2441 }
2442
2443 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2444 {
2445         struct ptlrpc_service_part *svcpt;
2446         int                        netrc;
2447         struct ptlrpc_reply_state *rs;
2448         struct obd_export         *exp;
2449         ENTRY;
2450
2451         if (req->rq_no_reply) {
2452                 EXIT;
2453                 return;
2454         }
2455
2456         svcpt = req->rq_rqbd->rqbd_svcpt;
2457         rs = req->rq_reply_state;
2458         if (rs == NULL || !rs->rs_difficult) {
2459                 /* no notifiers */
2460                 target_send_reply_msg (req, rc, fail_id);
2461                 EXIT;
2462                 return;
2463         }
2464
2465         /* must be an export if locks saved */
2466         LASSERT (req->rq_export != NULL);
2467         /* req/reply consistent */
2468         LASSERT(rs->rs_svcpt == svcpt);
2469
2470         /* "fresh" reply */
2471         LASSERT (!rs->rs_scheduled);
2472         LASSERT (!rs->rs_scheduled_ever);
2473         LASSERT (!rs->rs_handled);
2474         LASSERT (!rs->rs_on_net);
2475         LASSERT (rs->rs_export == NULL);
2476         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2477         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2478
2479         exp = class_export_get (req->rq_export);
2480
2481         /* disable reply scheduling while I'm setting up */
2482         rs->rs_scheduled = 1;
2483         rs->rs_on_net    = 1;
2484         rs->rs_xid       = req->rq_xid;
2485         rs->rs_transno   = req->rq_transno;
2486         rs->rs_export    = exp;
2487         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2488
2489         spin_lock(&exp->exp_uncommitted_replies_lock);
2490         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2491                rs->rs_transno, exp->exp_last_committed);
2492         if (rs->rs_transno > exp->exp_last_committed) {
2493                 /* not committed already */
2494                 cfs_list_add_tail(&rs->rs_obd_list,
2495                                   &exp->exp_uncommitted_replies);
2496         }
2497         spin_unlock(&exp->exp_uncommitted_replies_lock);
2498
2499         spin_lock(&exp->exp_lock);
2500         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2501         spin_unlock(&exp->exp_lock);
2502
2503         netrc = target_send_reply_msg(req, rc, fail_id);
2504
2505         spin_lock(&svcpt->scp_rep_lock);
2506
2507         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2508
2509         if (netrc != 0) {
2510                 /* error sending: reply is off the net.  Also we need +1
2511                  * reply ref until ptlrpc_handle_rs() is done
2512                  * with the reply state (if the send was successful, there
2513                  * would have been +1 ref for the net, which
2514                  * reply_out_callback leaves alone) */
2515                 rs->rs_on_net = 0;
2516                 ptlrpc_rs_addref(rs);
2517         }
2518
2519         spin_lock(&rs->rs_lock);
2520         if (rs->rs_transno <= exp->exp_last_committed ||
2521             (!rs->rs_on_net && !rs->rs_no_ack) ||
2522             cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2523             cfs_list_empty(&rs->rs_obd_list)) {
2524                 CDEBUG(D_HA, "Schedule reply immediately\n");
2525                 ptlrpc_dispatch_difficult_reply(rs);
2526         } else {
2527                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2528                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2529         }
2530         spin_unlock(&rs->rs_lock);
2531         spin_unlock(&svcpt->scp_rep_lock);
2532         EXIT;
2533 }
2534 EXPORT_SYMBOL(target_send_reply);
2535
2536 ldlm_mode_t lck_compat_array[] = {
2537         [LCK_EX] LCK_COMPAT_EX,
2538         [LCK_PW] LCK_COMPAT_PW,
2539         [LCK_PR] LCK_COMPAT_PR,
2540         [LCK_CW] LCK_COMPAT_CW,
2541         [LCK_CR] LCK_COMPAT_CR,
2542         [LCK_NL] LCK_COMPAT_NL,
2543         [LCK_GROUP] LCK_COMPAT_GROUP,
2544         [LCK_COS] LCK_COMPAT_COS,
2545 };
2546
2547 /**
2548  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2549  * not escape to the user level.
2550  */
2551 int ldlm_error2errno(ldlm_error_t error)
2552 {
2553         int result;
2554
2555         switch (error) {
2556         case ELDLM_OK:
2557                 result = 0;
2558                 break;
2559         case ELDLM_LOCK_CHANGED:
2560                 result = -ESTALE;
2561                 break;
2562         case ELDLM_LOCK_ABORTED:
2563                 result = -ENAVAIL;
2564                 break;
2565         case ELDLM_LOCK_REPLACED:
2566                 result = -ESRCH;
2567                 break;
2568         case ELDLM_NO_LOCK_DATA:
2569                 result = -ENOENT;
2570                 break;
2571         case ELDLM_NAMESPACE_EXISTS:
2572                 result = -EEXIST;
2573                 break;
2574         case ELDLM_BAD_NAMESPACE:
2575                 result = -EBADF;
2576                 break;
2577         default:
2578                 if (((int)error) < 0)  /* cast to signed type */
2579                         result = error; /* as ldlm_error_t can be unsigned */
2580                 else {
2581                         CERROR("Invalid DLM result code: %d\n", error);
2582                         result = -EPROTO;
2583                 }
2584         }
2585         return result;
2586 }
2587 EXPORT_SYMBOL(ldlm_error2errno);
2588
2589 /**
2590  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2591  */
2592 ldlm_error_t ldlm_errno2error(int err_no)
2593 {
2594         int error;
2595
2596         switch (err_no) {
2597         case 0:
2598                 error = ELDLM_OK;
2599                 break;
2600         case -ESTALE:
2601                 error = ELDLM_LOCK_CHANGED;
2602                 break;
2603         case -ENAVAIL:
2604                 error = ELDLM_LOCK_ABORTED;
2605                 break;
2606         case -ESRCH:
2607                 error = ELDLM_LOCK_REPLACED;
2608                 break;
2609         case -ENOENT:
2610                 error = ELDLM_NO_LOCK_DATA;
2611                 break;
2612         case -EEXIST:
2613                 error = ELDLM_NAMESPACE_EXISTS;
2614                 break;
2615         case -EBADF:
2616                 error = ELDLM_BAD_NAMESPACE;
2617                 break;
2618         default:
2619                 error = err_no;
2620         }
2621         return error;
2622 }
2623 EXPORT_SYMBOL(ldlm_errno2error);
2624
2625 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2626 void ldlm_dump_export_locks(struct obd_export *exp)
2627 {
2628         spin_lock(&exp->exp_locks_list_guard);
2629         if (!cfs_list_empty(&exp->exp_locks_list)) {
2630                 struct ldlm_lock *lock;
2631
2632                 CERROR("dumping locks for export %p,"
2633                        "ignore if the unmount doesn't hang\n", exp);
2634                 cfs_list_for_each_entry(lock, &exp->exp_locks_list,
2635                                         l_exp_refs_link)
2636                         LDLM_ERROR(lock, "lock:");
2637         }
2638         spin_unlock(&exp->exp_locks_list_guard);
2639 }
2640 #endif
2641
2642 #ifdef HAVE_SERVER_SUPPORT
2643 static int target_bulk_timeout(void *data)
2644 {
2645         ENTRY;
2646         /* We don't fail the connection here, because having the export
2647          * killed makes the (vital) call to commitrw very sad.
2648          */
2649         RETURN(1);
2650 }
2651
2652 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2653 {
2654         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2655 }
2656
2657 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2658                    struct l_wait_info *lwi)
2659 {
2660         struct ptlrpc_request *req = desc->bd_req;
2661         int rc = 0;
2662         ENTRY;
2663
2664         /* Check if there is eviction in progress, and if so, wait for
2665          * it to finish */
2666         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2667                 *lwi = LWI_INTR(NULL, NULL);
2668                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2669                                   !cfs_atomic_read(&exp->exp_obd->
2670                                                    obd_evict_inprogress),
2671                                   lwi);
2672         }
2673
2674         /* Check if client was evicted or tried to reconnect already */
2675         if (exp->exp_failed || exp->exp_abort_active_req) {
2676                 rc = -ENOTCONN;
2677         } else {
2678                 if (desc->bd_type == BULK_PUT_SINK)
2679                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2680                 if (rc == 0)
2681                         rc = ptlrpc_start_bulk_transfer(desc);
2682         }
2683
2684         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2685                 ptlrpc_abort_bulk(desc);
2686         } else if (rc == 0) {
2687                 time_t start = cfs_time_current_sec();
2688                 do {
2689                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2690                         cfs_duration_t timeout = timeoutl <= 0 ?
2691                                 CFS_TICK : cfs_time_seconds(timeoutl);
2692                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2693                                                     cfs_time_seconds(1),
2694                                                    target_bulk_timeout,
2695                                                    desc);
2696                         rc = l_wait_event(desc->bd_waitq,
2697                                           !ptlrpc_server_bulk_active(desc) ||
2698                                           exp->exp_failed ||
2699                                           exp->exp_abort_active_req,
2700                                           lwi);
2701                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2702                         /* Wait again if we changed deadline */
2703                 } while ((rc == -ETIMEDOUT) &&
2704                          (req->rq_deadline > cfs_time_current_sec()));
2705
2706                 if (rc == -ETIMEDOUT) {
2707                         DEBUG_REQ(D_ERROR, req,
2708                                   "timeout on bulk %s after %ld%+lds",
2709                                   bulk2type(desc),
2710                                   req->rq_deadline - start,
2711                                   cfs_time_current_sec() -
2712                                   req->rq_deadline);
2713                         ptlrpc_abort_bulk(desc);
2714                 } else if (exp->exp_failed) {
2715                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2716                                   bulk2type(desc));
2717                         rc = -ENOTCONN;
2718                         ptlrpc_abort_bulk(desc);
2719                 } else if (exp->exp_abort_active_req) {
2720                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2721                                   bulk2type(desc));
2722                         /* we don't reply anyway */
2723                         rc = -ETIMEDOUT;
2724                         ptlrpc_abort_bulk(desc);
2725                 } else if (!desc->bd_success ||
2726                            desc->bd_nob_transferred != desc->bd_nob) {
2727                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2728                                   desc->bd_success ?
2729                                   "truncated" : "network error on",
2730                                   bulk2type(desc),
2731                                   desc->bd_nob_transferred,
2732                                   desc->bd_nob);
2733                         /* XXX should this be a different errno? */
2734                         rc = -ETIMEDOUT;
2735                 } else if (desc->bd_type == BULK_GET_SINK) {
2736                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2737                 }
2738         } else {
2739                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2740                           bulk2type(desc), rc);
2741         }
2742
2743         RETURN(rc);
2744 }
2745 EXPORT_SYMBOL(target_bulk_io);
2746
2747 #endif /* HAVE_SERVER_SUPPORT */