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