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