lustre/ldlm/ldlm_lib.c

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