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