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