lustre/ldlm/ldlm_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2010, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 /**
  38  * This file deals with various client/target related logic including recovery.
  39  *
  40  * TODO: This code more logically belongs in the ptlrpc module than in ldlm and
  41  * should be moved.
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_LDLM
  45
  46 #ifdef __KERNEL__
  47 # include <libcfs/libcfs.h>
  48 #else
  49 # include <liblustre.h>
  50 #endif
  51 #include <obd.h>
  52 #include <obd_class.h>
  53 #include <lustre_dlm.h>
  54 #include <lustre_net.h>
  55 #include <lustre_sec.h>
  56 #include "ldlm_internal.h"
  57
  58 /* @priority: If non-zero, move the selected connection to the list head.
  59  * @create: If zero, only search in existing connections.
  60  */
  61 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  62                            int priority, int create)
  63 {
  64         struct ptlrpc_connection *ptlrpc_conn;
  65         struct obd_import_conn *imp_conn = NULL, *item;
  66         int rc = 0;
  67         ENTRY;
  68
  69         if (!create && !priority) {
  70                 CDEBUG(D_HA, "Nothing to do\n");
  71                 RETURN(-EINVAL);
  72         }
  73
  74         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  75         if (!ptlrpc_conn) {
  76                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  77                 RETURN (-ENOENT);
  78         }
  79
  80         if (create) {
  81                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  82                 if (!imp_conn) {
  83                         GOTO(out_put, rc = -ENOMEM);
  84                 }
  85         }
  86
  87         spin_lock(&imp->imp_lock);
  88         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  89                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  90                         if (priority) {
  91                                 cfs_list_del(&item->oic_item);
  92                                 cfs_list_add(&item->oic_item,
  93                                              &imp->imp_conn_list);
  94                                 item->oic_last_attempt = 0;
  95                         }
  96                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  97                                imp, imp->imp_obd->obd_name, uuid->uuid,
  98                                (priority ? ", moved to head" : ""));
  99                         spin_unlock(&imp->imp_lock);
 100                         GOTO(out_free, rc = 0);
 101                 }
 102         }
 103         /* No existing import connection found for \a uuid. */
 104         if (create) {
 105                 imp_conn->oic_conn = ptlrpc_conn;
 106                 imp_conn->oic_uuid = *uuid;
 107                 imp_conn->oic_last_attempt = 0;
 108                 if (priority)
 109                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 110                 else
 111                         cfs_list_add_tail(&imp_conn->oic_item,
 112                                           &imp->imp_conn_list);
 113                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 114                        imp, imp->imp_obd->obd_name, uuid->uuid,
 115                        (priority ? "head" : "tail"));
 116         } else {
 117                 spin_unlock(&imp->imp_lock);
 118                 GOTO(out_free, rc = -ENOENT);
 119         }
 120
 121         spin_unlock(&imp->imp_lock);
 122         RETURN(0);
 123 out_free:
 124         if (imp_conn)
 125                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 126 out_put:
 127         ptlrpc_connection_put(ptlrpc_conn);
 128         RETURN(rc);
 129 }
 130
 131 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 132 {
 133         return import_set_conn(imp, uuid, 1, 0);
 134 }
 135
 136 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 137                            int priority)
 138 {
 139         return import_set_conn(imp, uuid, priority, 1);
 140 }
 141 EXPORT_SYMBOL(client_import_add_conn);
 142
 143 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 144 {
 145         struct obd_import_conn *imp_conn;
 146         struct obd_export *dlmexp;
 147         int rc = -ENOENT;
 148         ENTRY;
 149
 150         spin_lock(&imp->imp_lock);
 151         if (cfs_list_empty(&imp->imp_conn_list)) {
 152                 LASSERT(!imp->imp_connection);
 153                 GOTO(out, rc);
 154         }
 155
 156         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 157                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 158                         continue;
 159                 LASSERT(imp_conn->oic_conn);
 160
 161                 if (imp_conn == imp->imp_conn_current) {
 162                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 163
 164                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 165                             imp->imp_state != LUSTRE_IMP_DISCON) {
 166                                 CERROR("can't remove current connection\n");
 167                                 GOTO(out, rc = -EBUSY);
 168                         }
 169
 170                         ptlrpc_connection_put(imp->imp_connection);
 171                         imp->imp_connection = NULL;
 172
 173                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 174                         if (dlmexp && dlmexp->exp_connection) {
 175                                 LASSERT(dlmexp->exp_connection ==
 176                                         imp_conn->oic_conn);
 177                                 ptlrpc_connection_put(dlmexp->exp_connection);
 178                                 dlmexp->exp_connection = NULL;
 179                         }
 180                 }
 181
 182                 cfs_list_del(&imp_conn->oic_item);
 183                 ptlrpc_connection_put(imp_conn->oic_conn);
 184                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 185                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 186                        imp, imp->imp_obd->obd_name, uuid->uuid);
 187                 rc = 0;
 188                 break;
 189         }
 190 out:
 191         spin_unlock(&imp->imp_lock);
 192         if (rc == -ENOENT)
 193                 CERROR("connection %s not found\n", uuid->uuid);
 194         RETURN(rc);
 195 }
 196 EXPORT_SYMBOL(client_import_del_conn);
 197
 198 /**
 199  * Find conn UUID by peer NID. \a peer is a server NID. This function is used
 200  * to find a conn uuid of \a imp which can reach \a peer.
 201  */
 202 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 203                             struct obd_uuid *uuid)
 204 {
 205         struct obd_import_conn *conn;
 206         int rc = -ENOENT;
 207         ENTRY;
 208
 209         spin_lock(&imp->imp_lock);
 210         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 211                 /* Check if conn UUID does have this peer NID. */
 212                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 213                         *uuid = conn->oic_uuid;
 214                         rc = 0;
 215                         break;
 216                 }
 217         }
 218         spin_unlock(&imp->imp_lock);
 219         RETURN(rc);
 220 }
 221 EXPORT_SYMBOL(client_import_find_conn);
 222
 223 void client_destroy_import(struct obd_import *imp)
 224 {
 225         /* Drop security policy instance after all RPCs have finished/aborted
 226          * to let all busy contexts be released. */
 227         class_import_get(imp);
 228         class_destroy_import(imp);
 229         sptlrpc_import_sec_put(imp);
 230         class_import_put(imp);
 231 }
 232 EXPORT_SYMBOL(client_destroy_import);
 233
 234 /**
 235  * Check whether or not the OSC is on MDT.
 236  * In the config log,
 237  * osc on MDT
 238  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 239  * osc on client
 240  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 241  *
 242  **/
 243 static int osc_on_mdt(char *obdname)
 244 {
 245         char *ptr;
 246
 247         ptr = strrchr(obdname, '-');
 248         if (ptr == NULL)
 249                 return 0;
 250
 251         if (strncmp(ptr + 1, "MDT", 3) == 0)
 252                 return 1;
 253
 254         return 0;
 255 }
 256
 257 /* Configure an RPC client OBD device.
 258  *
 259  * lcfg parameters:
 260  * 1 - client UUID
 261  * 2 - server UUID
 262  * 3 - inactive-on-startup
 263  */
 264 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 265 {
 266         struct client_obd *cli = &obddev->u.cli;
 267         struct obd_import *imp;
 268         struct obd_uuid server_uuid;
 269         int rq_portal, rp_portal, connect_op;
 270         char *name = obddev->obd_type->typ_name;
 271         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 272         int rc;
 273         char    *cli_name = lustre_cfg_buf(lcfg, 0);
 274         ENTRY;
 275
 276         /* In a more perfect world, we would hang a ptlrpc_client off of
 277          * obd_type and just use the values from there. */
 278         if (!strcmp(name, LUSTRE_OSC_NAME) ||
 279             (!(strcmp(name, LUSTRE_OSP_NAME)) &&
 280              (is_osp_on_mdt(cli_name) &&
 281                strstr(lustre_cfg_buf(lcfg, 1), "OST") != NULL))) {
 282                 /* OSC or OSP_on_MDT for OSTs */
 283                 rq_portal = OST_REQUEST_PORTAL;
 284                 rp_portal = OSC_REPLY_PORTAL;
 285                 connect_op = OST_CONNECT;
 286                 cli->cl_sp_me = LUSTRE_SP_CLI;
 287                 cli->cl_sp_to = LUSTRE_SP_OST;
 288                 ns_type = LDLM_NS_TYPE_OSC;
 289         } else if (!strcmp(name, LUSTRE_MDC_NAME) ||
 290                    !strcmp(name, LUSTRE_LWP_NAME) ||
 291                    (!strcmp(name, LUSTRE_OSP_NAME) &&
 292                     (is_osp_on_mdt(cli_name) &&
 293                      strstr(lustre_cfg_buf(lcfg, 1), "OST") == NULL))) {
 294                 /* MDC or OSP_on_MDT for other MDTs */
 295                 rq_portal = MDS_REQUEST_PORTAL;
 296                 rp_portal = MDC_REPLY_PORTAL;
 297                 connect_op = MDS_CONNECT;
 298                 cli->cl_sp_me = LUSTRE_SP_CLI;
 299                 cli->cl_sp_to = LUSTRE_SP_MDT;
 300                 ns_type = LDLM_NS_TYPE_MDC;
 301         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 302                 rq_portal = MGS_REQUEST_PORTAL;
 303                 rp_portal = MGC_REPLY_PORTAL;
 304                 connect_op = MGS_CONNECT;
 305                 cli->cl_sp_me = LUSTRE_SP_MGC;
 306                 cli->cl_sp_to = LUSTRE_SP_MGS;
 307                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 308                 ns_type = LDLM_NS_TYPE_MGC;
 309         } else {
 310                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 311                        name);
 312                 RETURN(-EINVAL);
 313         }
 314
 315         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 316                 CERROR("requires a TARGET UUID\n");
 317                 RETURN(-EINVAL);
 318         }
 319
 320         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 321                 CERROR("client UUID must be less than 38 characters\n");
 322                 RETURN(-EINVAL);
 323         }
 324
 325         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 326                 CERROR("setup requires a SERVER UUID\n");
 327                 RETURN(-EINVAL);
 328         }
 329
 330         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 331                 CERROR("target UUID must be less than 38 characters\n");
 332                 RETURN(-EINVAL);
 333         }
 334
 335         init_rwsem(&cli->cl_sem);
 336         sema_init(&cli->cl_mgc_sem, 1);
 337         cli->cl_conn_count = 0;
 338         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 339                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 340                      sizeof(server_uuid)));
 341
 342         cli->cl_dirty = 0;
 343         cli->cl_avail_grant = 0;
 344         /* FIXME: Should limit this for the sum of all cl_dirty_max. */
 345         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 346         if (cli->cl_dirty_max >> PAGE_CACHE_SHIFT > totalram_pages / 8)
 347                 cli->cl_dirty_max = totalram_pages << (PAGE_CACHE_SHIFT - 3);
 348         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 349         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 350         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 351         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 352         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 353         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 354         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
 355         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
 356         cli->cl_r_in_flight = 0;
 357         cli->cl_w_in_flight = 0;
 358
 359         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 360         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 361         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 362         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 363         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 364         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 365
 366         /* lru for osc. */
 367         CFS_INIT_LIST_HEAD(&cli->cl_lru_osc);
 368         cfs_atomic_set(&cli->cl_lru_shrinkers, 0);
 369         cfs_atomic_set(&cli->cl_lru_busy, 0);
 370         cfs_atomic_set(&cli->cl_lru_in_list, 0);
 371         CFS_INIT_LIST_HEAD(&cli->cl_lru_list);
 372         client_obd_list_lock_init(&cli->cl_lru_list_lock);
 373
 374         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         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 396                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 397         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 398                 cli->cl_max_rpcs_in_flight = 2;
 399         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 400                 cli->cl_max_rpcs_in_flight = 3;
 401         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 402                 cli->cl_max_rpcs_in_flight = 4;
 403         } else {
 404                 if (osc_on_mdt(obddev->obd_name))
 405                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
 406                 else
 407                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 408         }
 409         rc = ldlm_get_ref();
 410         if (rc) {
 411                 CERROR("ldlm_get_ref failed: %d\n", rc);
 412                 GOTO(err, rc);
 413         }
 414
 415         ptlrpc_init_client(rq_portal, rp_portal, name,
 416                            &obddev->obd_ldlm_client);
 417
 418         imp = class_new_import(obddev);
 419         if (imp == NULL)
 420                 GOTO(err_ldlm, rc = -ENOENT);
 421         imp->imp_client = &obddev->obd_ldlm_client;
 422         imp->imp_connect_op = connect_op;
 423         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 424                LUSTRE_CFG_BUFLEN(lcfg, 1));
 425         class_import_put(imp);
 426
 427         rc = client_import_add_conn(imp, &server_uuid, 1);
 428         if (rc) {
 429                 CERROR("can't add initial connection\n");
 430                 GOTO(err_import, rc);
 431         }
 432
 433         cli->cl_import = imp;
 434         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 435         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 436         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 437
 438         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 439                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 440                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 441                                name, obddev->obd_name,
 442                                cli->cl_target_uuid.uuid);
 443                         spin_lock(&imp->imp_lock);
 444                         imp->imp_deactive = 1;
 445                         spin_unlock(&imp->imp_lock);
 446                 }
 447         }
 448
 449         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 450                                                    LDLM_NAMESPACE_CLIENT,
 451                                                    LDLM_NAMESPACE_GREEDY,
 452                                                    ns_type);
 453         if (obddev->obd_namespace == NULL) {
 454                 CERROR("Unable to create client namespace - %s\n",
 455                        obddev->obd_name);
 456                 GOTO(err_import, rc = -ENOMEM);
 457         }
 458
 459         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 460
 461         RETURN(rc);
 462
 463 err_import:
 464         class_destroy_import(imp);
 465 err_ldlm:
 466         ldlm_put_ref();
 467 err:
 468         RETURN(rc);
 469
 470 }
 471 EXPORT_SYMBOL(client_obd_setup);
 472
 473 int client_obd_cleanup(struct obd_device *obddev)
 474 {
 475         ENTRY;
 476
 477         ldlm_namespace_free_post(obddev->obd_namespace);
 478         obddev->obd_namespace = NULL;
 479
 480         LASSERT(obddev->u.cli.cl_import == NULL);
 481
 482         ldlm_put_ref();
 483         RETURN(0);
 484 }
 485 EXPORT_SYMBOL(client_obd_cleanup);
 486
 487 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 488 int client_connect_import(const struct lu_env *env,
 489                           struct obd_export **exp,
 490                           struct obd_device *obd, struct obd_uuid *cluuid,
 491                           struct obd_connect_data *data, void *localdata)
 492 {
 493         struct client_obd       *cli    = &obd->u.cli;
 494         struct obd_import       *imp    = cli->cl_import;
 495         struct obd_connect_data *ocd;
 496         struct lustre_handle    conn    = { 0 };
 497         int                     rc;
 498         ENTRY;
 499
 500         *exp = NULL;
 501         down_write(&cli->cl_sem);
 502         if (cli->cl_conn_count > 0 )
 503                 GOTO(out_sem, rc = -EALREADY);
 504
 505         rc = class_connect(&conn, obd, cluuid);
 506         if (rc)
 507                 GOTO(out_sem, rc);
 508
 509         cli->cl_conn_count++;
 510         *exp = class_conn2export(&conn);
 511
 512         LASSERT(obd->obd_namespace);
 513
 514         imp->imp_dlm_handle = conn;
 515         rc = ptlrpc_init_import(imp);
 516         if (rc != 0)
 517                 GOTO(out_ldlm, rc);
 518
 519         ocd = &imp->imp_connect_data;
 520         if (data) {
 521                 *ocd = *data;
 522                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 523         }
 524
 525         rc = ptlrpc_connect_import(imp);
 526         if (rc != 0) {
 527                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 528                 GOTO(out_ldlm, rc);
 529         }
 530         LASSERT((*exp)->exp_connection);
 531
 532         if (data) {
 533                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 534                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 535                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 536                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 537         }
 538
 539         ptlrpc_pinger_add_import(imp);
 540
 541         EXIT;
 542
 543         if (rc) {
 544 out_ldlm:
 545                 cli->cl_conn_count--;
 546                 class_disconnect(*exp);
 547                 *exp = NULL;
 548         }
 549 out_sem:
 550         up_write(&cli->cl_sem);
 551
 552         return rc;
 553 }
 554 EXPORT_SYMBOL(client_connect_import);
 555
 556 int client_disconnect_export(struct obd_export *exp)
 557 {
 558         struct obd_device *obd = class_exp2obd(exp);
 559         struct client_obd *cli;
 560         struct obd_import *imp;
 561         int rc = 0, err;
 562         ENTRY;
 563
 564         if (!obd) {
 565                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 566                        exp, exp ? exp->exp_handle.h_cookie : -1);
 567                 RETURN(-EINVAL);
 568         }
 569
 570         cli = &obd->u.cli;
 571         imp = cli->cl_import;
 572
 573         down_write(&cli->cl_sem);
 574         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 575                cli->cl_conn_count);
 576
 577         if (!cli->cl_conn_count) {
 578                 CERROR("disconnecting disconnected device (%s)\n",
 579                        obd->obd_name);
 580                 GOTO(out_disconnect, rc = -EINVAL);
 581         }
 582
 583         cli->cl_conn_count--;
 584         if (cli->cl_conn_count)
 585                 GOTO(out_disconnect, rc = 0);
 586
 587         /* Mark import deactivated now, so we don't try to reconnect if any
 588          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 589          * fully deactivate the import, or that would drop all requests. */
 590         spin_lock(&imp->imp_lock);
 591         imp->imp_deactive = 1;
 592         spin_unlock(&imp->imp_lock);
 593
 594         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 595          * delete it regardless.  (It's safe to delete an import that was
 596          * never added.) */
 597         (void)ptlrpc_pinger_del_import(imp);
 598
 599         if (obd->obd_namespace != NULL) {
 600                 /* obd_force == local only */
 601                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 602                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 603                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 604         }
 605
 606         /* There's no need to hold sem while disconnecting an import,
 607          * and it may actually cause deadlock in GSS. */
 608         up_write(&cli->cl_sem);
 609         rc = ptlrpc_disconnect_import(imp, 0);
 610         down_write(&cli->cl_sem);
 611
 612         ptlrpc_invalidate_import(imp);
 613
 614         EXIT;
 615
 616 out_disconnect:
 617         /* Use server style - class_disconnect should be always called for
 618          * o_disconnect. */
 619         err = class_disconnect(exp);
 620         if (!rc && err)
 621                 rc = err;
 622
 623         up_write(&cli->cl_sem);
 624
 625         RETURN(rc);
 626 }
 627 EXPORT_SYMBOL(client_disconnect_export);
 628
 629 #ifdef HAVE_SERVER_SUPPORT
 630 int server_disconnect_export(struct obd_export *exp)
 631 {
 632         int rc;
 633         ENTRY;
 634
 635         /* Disconnect early so that clients can't keep using export. */
 636         rc = class_disconnect(exp);
 637         /* Close import to avoid sending any requests. */
 638         if (exp->exp_imp_reverse)
 639                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 640
 641         if (exp->exp_obd->obd_namespace != NULL)
 642                 ldlm_cancel_locks_for_export(exp);
 643
 644         /* complete all outstanding replies */
 645         spin_lock(&exp->exp_lock);
 646         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 647                 struct ptlrpc_reply_state *rs =
 648                         cfs_list_entry(exp->exp_outstanding_replies.next,
 649                                        struct ptlrpc_reply_state, rs_exp_list);
 650                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 651
 652                 spin_lock(&svcpt->scp_rep_lock);
 653
 654                 cfs_list_del_init(&rs->rs_exp_list);
 655                 spin_lock(&rs->rs_lock);
 656                 ptlrpc_schedule_difficult_reply(rs);
 657                 spin_unlock(&rs->rs_lock);
 658
 659                 spin_unlock(&svcpt->scp_rep_lock);
 660         }
 661         spin_unlock(&exp->exp_lock);
 662
 663         RETURN(rc);
 664 }
 665 EXPORT_SYMBOL(server_disconnect_export);
 666
 667 /* --------------------------------------------------------------------------
 668  * from old lib/target.c
 669  * -------------------------------------------------------------------------- */
 670
 671 static int target_handle_reconnect(struct lustre_handle *conn,
 672                                    struct obd_export *exp,
 673                                    struct obd_uuid *cluuid)
 674 {
 675         ENTRY;
 676
 677         if (exp->exp_connection && exp->exp_imp_reverse) {
 678                 struct lustre_handle *hdl;
 679                 struct obd_device *target;
 680
 681                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 682                 target = exp->exp_obd;
 683
 684                 /* Might be a re-connect after a partition. */
 685                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 686                         if (target->obd_recovering) {
 687                                 int timeout = cfs_duration_sec(cfs_time_sub(
 688                                         cfs_timer_deadline(
 689                                         &target->obd_recovery_timer),
 690                                         cfs_time_current()));
 691
 692                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 693                                         "ing, waiting for %d clients in recov"
 694                                         "ery for %d:%.02d\n", target->obd_name,
 695                                         obd_uuid2str(&exp->exp_client_uuid),
 696                                         obd_export_nid2str(exp),
 697                                         target->obd_max_recoverable_clients,
 698                                         timeout / 60, timeout % 60);
 699                         } else {
 700                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 701                                         "reconnecting\n", target->obd_name,
 702                                         obd_uuid2str(&exp->exp_client_uuid),
 703                                         obd_export_nid2str(exp));
 704                         }
 705
 706                         conn->cookie = exp->exp_handle.h_cookie;
 707                         /* target_handle_connect() treats EALREADY and
 708                          * -EALREADY differently.  EALREADY means we are
 709                          * doing a valid reconnect from the same client. */
 710                         RETURN(EALREADY);
 711                 } else {
 712                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 713                                       "with handle "LPX64". Rejecting client "
 714                                       "with the same UUID trying to reconnect "
 715                                       "with handle "LPX64"\n", target->obd_name,
 716                                       obd_uuid2str(&exp->exp_client_uuid),
 717                                       obd_export_nid2str(exp),
 718                                       hdl->cookie, conn->cookie);
 719                         memset(conn, 0, sizeof *conn);
 720                         /* target_handle_connect() treats EALREADY and
 721                          * -EALREADY differently.  -EALREADY is an error
 722                          * (same UUID, different handle). */
 723                         RETURN(-EALREADY);
 724                 }
 725         }
 726
 727         conn->cookie = exp->exp_handle.h_cookie;
 728         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 729                cluuid->uuid, exp, conn->cookie);
 730         RETURN(0);
 731 }
 732
 733 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 734                           int error)
 735 {
 736         struct obd_export *exp = cb_data;
 737
 738         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 739                obd->obd_name, exp->exp_client_uuid.uuid);
 740
 741         spin_lock(&exp->exp_lock);
 742         exp->exp_need_sync = 0;
 743         spin_unlock(&exp->exp_lock);
 744         class_export_cb_put(exp);
 745 }
 746 EXPORT_SYMBOL(target_client_add_cb);
 747
 748 #ifdef __KERNEL__
 749 static void
 750 check_and_start_recovery_timer(struct obd_device *obd,
 751                                struct ptlrpc_request *req, int new_client);
 752 #else
 753 static inline void
 754 check_and_start_recovery_timer(struct obd_device *obd,
 755                                struct ptlrpc_request *req, int new_client)
 756 {
 757 }
 758 #endif
 759
 760 int target_handle_connect(struct ptlrpc_request *req)
 761 {
 762         struct obd_device *target = NULL, *targref = NULL;
 763         struct obd_export *export = NULL;
 764         struct obd_import *revimp;
 765         struct obd_import *tmp_imp = NULL;
 766         struct lustre_handle conn;
 767         struct lustre_handle *tmp;
 768         struct obd_uuid tgtuuid;
 769         struct obd_uuid cluuid;
 770         struct obd_uuid remote_uuid;
 771         char *str;
 772         int rc = 0;
 773         char *target_start;
 774         int target_len;
 775         bool     mds_conn = false, lw_client = false;
 776         struct obd_connect_data *data, *tmpdata;
 777         int size, tmpsize;
 778         lnet_nid_t *client_nid = NULL;
 779         ENTRY;
 780
 781         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 782
 783         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 784         if (str == NULL) {
 785                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 786                 GOTO(out, rc = -EINVAL);
 787         }
 788
 789         obd_str2uuid(&tgtuuid, str);
 790         target = class_uuid2obd(&tgtuuid);
 791         if (!target)
 792                 target = class_name2obd(str);
 793
 794         if (!target) {
 795                 deuuidify(str, NULL, &target_start, &target_len);
 796                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 797                                    "from %s (no target). If you are running "
 798                                    "an HA pair check that the target is "
 799                                    "mounted on the other server.\n", str,
 800                                    libcfs_nid2str(req->rq_peer.nid));
 801                 GOTO(out, rc = -ENODEV);
 802         }
 803
 804         spin_lock(&target->obd_dev_lock);
 805         if (target->obd_stopping || !target->obd_set_up) {
 806                 spin_unlock(&target->obd_dev_lock);
 807
 808                 deuuidify(str, NULL, &target_start, &target_len);
 809                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 810                               target_len, target_start,
 811                               libcfs_nid2str(req->rq_peer.nid),
 812                               (target->obd_stopping ?
 813                                "stopping" : "not set up"));
 814                 GOTO(out, rc = -ENODEV);
 815         }
 816
 817         if (target->obd_no_conn) {
 818                 spin_unlock(&target->obd_dev_lock);
 819
 820                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 821                                "from %s\n", target->obd_name,
 822                                libcfs_nid2str(req->rq_peer.nid));
 823                 GOTO(out, rc = -EAGAIN);
 824         }
 825
 826         /* Make sure the target isn't cleaned up while we're here. Yes,
 827          * there's still a race between the above check and our incref here.
 828          * Really, class_uuid2obd should take the ref. */
 829         targref = class_incref(target, __FUNCTION__, current);
 830
 831         target->obd_conn_inprogress++;
 832         spin_unlock(&target->obd_dev_lock);
 833
 834         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 835         if (str == NULL) {
 836                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 837                 GOTO(out, rc = -EINVAL);
 838         }
 839
 840         obd_str2uuid(&cluuid, str);
 841
 842         /* XXX Extract a nettype and format accordingly. */
 843         switch (sizeof(lnet_nid_t)) {
 844         /* NB the casts only avoid compiler warnings. */
 845         case 8:
 846                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 847                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 848                 break;
 849         case 4:
 850                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 851                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 852                 break;
 853         default:
 854                 LBUG();
 855         }
 856
 857         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 858         if (tmp == NULL)
 859                 GOTO(out, rc = -EPROTO);
 860
 861         conn = *tmp;
 862
 863         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 864                                     RCL_CLIENT);
 865         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 866         if (!data)
 867                 GOTO(out, rc = -EPROTO);
 868
 869         rc = req_capsule_server_pack(&req->rq_pill);
 870         if (rc)
 871                 GOTO(out, rc);
 872
 873         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 874                 if (data->ocd_version < LUSTRE_VERSION_CODE -
 875                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 876                     data->ocd_version > LUSTRE_VERSION_CODE +
 877                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 878                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 879                                   "libclient connection attempt",
 880                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 881                                   "old" : "new",
 882                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 883                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 884                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 885                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 886                         data = req_capsule_server_sized_get(&req->rq_pill,
 887                                                             &RMF_CONNECT_DATA,
 888                                     offsetof(typeof(*data), ocd_version) +
 889                                              sizeof(data->ocd_version));
 890                         if (data) {
 891                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 892                                 data->ocd_version = LUSTRE_VERSION_CODE;
 893                         }
 894                         GOTO(out, rc = -EPROTO);
 895                 }
 896         }
 897
 898         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 899             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 900                 mds_conn = true;
 901
 902         if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
 903                 lw_client = true;
 904
 905         /* lctl gets a backstage, all-access pass. */
 906         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 907                 goto dont_check_exports;
 908
 909         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 910         if (!export)
 911                 goto no_export;
 912
 913         /* We've found an export in the hash. */
 914
 915         spin_lock(&export->exp_lock);
 916
 917         if (export->exp_connecting) { /* bug 9635, et. al. */
 918                 spin_unlock(&export->exp_lock);
 919                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 920                               export->exp_obd->obd_name, export,
 921                               libcfs_nid2str(req->rq_peer.nid));
 922                 class_export_put(export);
 923                 export = NULL;
 924                 rc = -EALREADY;
 925         } else if (mds_conn && export->exp_connection) {
 926                 spin_unlock(&export->exp_lock);
 927                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 928                         /* MDS reconnected after failover. */
 929                         LCONSOLE_WARN("%s: Received MDS connection from "
 930                             "%s, removing former export from %s\n",
 931                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 932                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 933                 else
 934                         /* New MDS connection from the same NID. */
 935                         LCONSOLE_WARN("%s: Received new MDS connection from "
 936                             "%s, removing former export from same NID\n",
 937                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 938                 class_fail_export(export);
 939                 class_export_put(export);
 940                 export = NULL;
 941                 rc = 0;
 942         } else if (export->exp_connection != NULL &&
 943                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 944                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 945                     MSG_CONNECT_INITIAL)) {
 946                 spin_unlock(&export->exp_lock);
 947                 /* In MDS failover we have static UUID but NID can change. */
 948                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 949                               "existing nid %s is already connected\n",
 950                               target->obd_name, cluuid.uuid,
 951                               libcfs_nid2str(req->rq_peer.nid),
 952                               libcfs_nid2str(
 953                                       export->exp_connection->c_peer.nid));
 954                 rc = -EALREADY;
 955                 class_export_put(export);
 956                 export = NULL;
 957         } else {
 958                 export->exp_connecting = 1;
 959                 spin_unlock(&export->exp_lock);
 960                 LASSERT(export->exp_obd == target);
 961
 962                 rc = target_handle_reconnect(&conn, export, &cluuid);
 963         }
 964
 965         /* If we found an export, we already unlocked. */
 966         if (!export) {
 967 no_export:
 968                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 969         } else if (req->rq_export == NULL &&
 970                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 971                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 972                               "still busy with %d references\n",
 973                               target->obd_name, cluuid.uuid,
 974                               libcfs_nid2str(req->rq_peer.nid),
 975                               cfs_atomic_read(&export->exp_refcount));
 976                 GOTO(out, rc = -EBUSY);
 977         } else if (req->rq_export != NULL &&
 978                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 979                 /* The current connect RPC has increased exp_rpc_count. */
 980                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
 981                               "still busy with %d active RPCs\n",
 982                               target->obd_name, cluuid.uuid,
 983                               libcfs_nid2str(req->rq_peer.nid),
 984                               cfs_atomic_read(&export->exp_rpc_count) - 1);
 985                 spin_lock(&export->exp_lock);
 986                 if (req->rq_export->exp_conn_cnt <
 987                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 988                         /* try to abort active requests */
 989                         req->rq_export->exp_abort_active_req = 1;
 990                 spin_unlock(&export->exp_lock);
 991                 GOTO(out, rc = -EBUSY);
 992         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 993                 if (!strstr(cluuid.uuid, "mdt"))
 994                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 995                                       "known client %s (at %s) because it "
 996                                       "is indicating it is a new client",
 997                                       target->obd_name, cluuid.uuid,
 998                                       libcfs_nid2str(req->rq_peer.nid));
 999                 GOTO(out, rc = -EALREADY);
1000         } else {
1001                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1002         }
1003
1004         if (rc < 0) {
1005                 GOTO(out, rc);
1006         }
1007
1008         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1009                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1010               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1011               export, (long)cfs_time_current_sec(),
1012               export ? (long)export->exp_last_request_time : 0);
1013
1014         /* If this is the first time a client connects, reset the recovery
1015          * timer. Discard lightweight connections which might be local. */
1016         if (!lw_client && rc == 0 && target->obd_recovering)
1017                 check_and_start_recovery_timer(target, req, export == NULL);
1018
1019         /* We want to handle EALREADY but *not* -EALREADY from
1020          * target_handle_reconnect(), return reconnection state in a flag. */
1021         if (rc == EALREADY) {
1022                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1023                 rc = 0;
1024         } else {
1025                 LASSERT(rc == 0);
1026         }
1027
1028         /* Tell the client if we support replayable requests. */
1029         if (target->obd_replayable)
1030                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1031         client_nid = &req->rq_peer.nid;
1032
1033         if (export == NULL) {
1034                 /* allow lightweight connections during recovery */
1035                 if (target->obd_recovering && !lw_client) {
1036                         cfs_time_t t;
1037                         int     c; /* connected */
1038                         int     i; /* in progress */
1039                         int     k; /* known */
1040                         int     s; /* stale/evicted */
1041
1042                         c = cfs_atomic_read(&target->obd_connected_clients);
1043                         i = cfs_atomic_read(&target->obd_lock_replay_clients);
1044                         k = target->obd_max_recoverable_clients;
1045                         s = target->obd_stale_clients;
1046                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1047                         t = cfs_time_sub(t, cfs_time_current());
1048                         t = cfs_duration_sec(t);
1049                         LCONSOLE_WARN("%s: Denying connection for new client "
1050                                       "%s (at %s), waiting for all %d known "
1051                                       "clients (%d recovered, %d in progress, "
1052                                       "and %d evicted) to recover in %d:%.02d\n",
1053                                       target->obd_name, cluuid.uuid,
1054                                       libcfs_nid2str(req->rq_peer.nid), k,
1055                                       c - i, i, s, (int)t / 60,
1056                                       (int)t % 60);
1057                         rc = -EBUSY;
1058                 } else {
1059 dont_check_exports:
1060                         rc = obd_connect(req->rq_svc_thread->t_env,
1061                                          &export, target, &cluuid, data,
1062                                          client_nid);
1063                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1064                                 lustre_msg_add_op_flags(req->rq_repmsg,
1065                                                 MSG_CONNECT_RECOVERING);
1066                         if (rc == 0)
1067                                 conn.cookie = export->exp_handle.h_cookie;
1068                 }
1069         } else {
1070                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1071                                    export, target, &cluuid, data, client_nid);
1072         }
1073         if (rc)
1074                 GOTO(out, rc);
1075
1076 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1077         /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1078          * will do the swabbing for if the client is using the same endianness.
1079          *
1080          * This fixup is version-limited, because we don't want to carry the
1081          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1082          * interop with unpatched 2.2 clients.  For newer clients, servers
1083          * will never do MNE swabbing, let the client handle that.  LU-1644 */
1084         export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1085                         !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1086 #else
1087 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1088 #endif
1089
1090         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1091         data->ocd_instance = target->u.obt.obt_instance;
1092
1093         /* Return only the parts of obd_connect_data that we understand, so the
1094          * client knows that we don't understand the rest. */
1095         if (data) {
1096                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1097                                                RCL_SERVER);
1098                 tmpdata = req_capsule_server_get(&req->rq_pill,
1099                                                  &RMF_CONNECT_DATA);
1100                 /* Don't use struct assignment here, because the client reply
1101                  * buffer may be smaller/larger than the local struct
1102                  * obd_connect_data. */
1103                 memcpy(tmpdata, data, min(tmpsize, size));
1104         }
1105
1106         /* If all else goes well, this is our RPC return code. */
1107         req->rq_status = 0;
1108
1109         lustre_msg_set_handle(req->rq_repmsg, &conn);
1110
1111         /* If the client and the server are the same node, we will already
1112          * have an export that really points to the client's DLM export,
1113          * because we have a shared handles table.
1114          *
1115          * XXX this will go away when shaver stops sending the "connect" handle
1116          * in the real "remote handle" field of the request --phik 24 Apr 2003
1117          */
1118         ptlrpc_request_change_export(req, export);
1119
1120         spin_lock(&export->exp_lock);
1121         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1122                 spin_unlock(&export->exp_lock);
1123                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1124                        "or equal conn_cnt: %d >= %d\n",
1125                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1126                        export->exp_conn_cnt,
1127                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1128
1129                 GOTO(out, rc = -EALREADY);
1130         }
1131         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1132         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1133         export->exp_abort_active_req = 0;
1134
1135         /* Don't evict liblustre clients for not pinging. */
1136         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1137                 export->exp_libclient = 1;
1138                 spin_unlock(&export->exp_lock);
1139
1140                 spin_lock(&target->obd_dev_lock);
1141                 cfs_list_del_init(&export->exp_obd_chain_timed);
1142                 spin_unlock(&target->obd_dev_lock);
1143         } else {
1144                 spin_unlock(&export->exp_lock);
1145         }
1146
1147         if (export->exp_connection != NULL) {
1148                 /* Check to see if connection came from another NID. */
1149                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1150                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1151                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1152                                      &export->exp_connection->c_peer.nid,
1153                                      &export->exp_nid_hash);
1154
1155                 ptlrpc_connection_put(export->exp_connection);
1156         }
1157
1158         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1159                                                        req->rq_self,
1160                                                        &remote_uuid);
1161         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1162                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1163                              &export->exp_connection->c_peer.nid,
1164                              &export->exp_nid_hash);
1165         }
1166
1167         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1168                 int has_transno;
1169                 __u64 transno = data->ocd_transno;
1170
1171                 spin_lock(&export->exp_lock);
1172                 /* possible race with class_disconnect_stale_exports,
1173                  * export may be already in the eviction process */
1174                 if (export->exp_failed) {
1175                         spin_unlock(&export->exp_lock);
1176                         GOTO(out, rc = -ENODEV);
1177                 }
1178                 export->exp_in_recovery = 1;
1179                 export->exp_req_replay_needed = 1;
1180                 export->exp_lock_replay_needed = 1;
1181                 spin_unlock(&export->exp_lock);
1182
1183                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1184                                  MSG_CONNECT_TRANSNO);
1185                 if (has_transno && transno == 0)
1186                         CWARN("Connect with zero transno!\n");
1187
1188                 if (has_transno && transno > 0 &&
1189                     transno < target->obd_next_recovery_transno &&
1190                     transno > target->obd_last_committed) {
1191                         /* Another way is to use cmpxchg() to be lock-free. */
1192                         spin_lock(&target->obd_recovery_task_lock);
1193                         if (transno < target->obd_next_recovery_transno)
1194                                 target->obd_next_recovery_transno = transno;
1195                         spin_unlock(&target->obd_recovery_task_lock);
1196                 }
1197
1198                 cfs_atomic_inc(&target->obd_req_replay_clients);
1199                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1200                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1201                     target->obd_max_recoverable_clients)
1202                         wake_up(&target->obd_next_transno_waitq);
1203         }
1204
1205         /* Tell the client we're in recovery, when client is involved in it. */
1206         if (target->obd_recovering && !lw_client)
1207                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1208
1209         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1210         conn = *tmp;
1211
1212         /* Return -ENOTCONN in case of errors to let client reconnect. */
1213         revimp = class_new_import(target);
1214         if (revimp == NULL) {
1215                 CERROR("fail to alloc new reverse import.\n");
1216                 GOTO(out, rc = -ENOTCONN);
1217         }
1218
1219         spin_lock(&export->exp_lock);
1220         if (export->exp_imp_reverse != NULL)
1221                 /* destroyed import can be still referenced in ctxt */
1222                 tmp_imp = export->exp_imp_reverse;
1223         export->exp_imp_reverse = revimp;
1224         spin_unlock(&export->exp_lock);
1225
1226         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1227         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1228         revimp->imp_remote_handle = conn;
1229         revimp->imp_dlm_fake = 1;
1230         revimp->imp_state = LUSTRE_IMP_FULL;
1231
1232         /* Unknown versions will be caught in
1233          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1234         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1235
1236         if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1237             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1238                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1239         else
1240                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1241
1242         if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1243             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1244                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1245         else
1246                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1247
1248         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1249         if (rc) {
1250                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1251                 spin_lock(&export->exp_lock);
1252                 export->exp_imp_reverse = NULL;
1253                 spin_unlock(&export->exp_lock);
1254                 class_destroy_import(revimp);
1255         }
1256
1257         class_import_put(revimp);
1258
1259 out:
1260         if (tmp_imp != NULL)
1261                 client_destroy_import(tmp_imp);
1262         if (export) {
1263                 spin_lock(&export->exp_lock);
1264                 export->exp_connecting = 0;
1265                 spin_unlock(&export->exp_lock);
1266
1267                 class_export_put(export);
1268         }
1269         if (targref) {
1270                 spin_lock(&target->obd_dev_lock);
1271                 target->obd_conn_inprogress--;
1272                 spin_unlock(&target->obd_dev_lock);
1273
1274                 class_decref(targref, __func__, current);
1275         }
1276         if (rc)
1277                 req->rq_status = rc;
1278         RETURN(rc);
1279 }
1280 EXPORT_SYMBOL(target_handle_connect);
1281
1282 int target_handle_disconnect(struct ptlrpc_request *req)
1283 {
1284         int rc;
1285         ENTRY;
1286
1287         rc = req_capsule_server_pack(&req->rq_pill);
1288         if (rc)
1289                 RETURN(rc);
1290
1291         /* Keep the rq_export around so we can send the reply. */
1292         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1293
1294         RETURN(0);
1295 }
1296 EXPORT_SYMBOL(target_handle_disconnect);
1297
1298 void target_destroy_export(struct obd_export *exp)
1299 {
1300         struct obd_import       *imp = NULL;
1301         /* exports created from last_rcvd data, and "fake"
1302            exports created by lctl don't have an import */
1303         spin_lock(&exp->exp_lock);
1304         if (exp->exp_imp_reverse != NULL) {
1305                 imp = exp->exp_imp_reverse;
1306                 exp->exp_imp_reverse = NULL;
1307         }
1308         spin_unlock(&exp->exp_lock);
1309         if (imp != NULL)
1310                 client_destroy_import(imp);
1311
1312         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1313         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1314         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1315         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1316 }
1317 EXPORT_SYMBOL(target_destroy_export);
1318
1319 /*
1320  * Recovery functions
1321  */
1322 static void target_request_copy_get(struct ptlrpc_request *req)
1323 {
1324         class_export_rpc_inc(req->rq_export);
1325         LASSERT(cfs_list_empty(&req->rq_list));
1326         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1327
1328         /* Increase refcount to keep request in queue. */
1329         cfs_atomic_inc(&req->rq_refcount);
1330         /* Let export know it has replays to be handled. */
1331         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1332 }
1333
1334 static void target_request_copy_put(struct ptlrpc_request *req)
1335 {
1336         LASSERT(cfs_list_empty(&req->rq_replay_list));
1337         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1338
1339         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1340         class_export_rpc_dec(req->rq_export);
1341         ptlrpc_server_drop_request(req);
1342 }
1343
1344 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1345 {
1346         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1347         struct obd_export     *exp = req->rq_export;
1348         struct ptlrpc_request *reqiter;
1349         int                    dup = 0;
1350
1351         LASSERT(exp);
1352
1353         spin_lock(&exp->exp_lock);
1354         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1355                                 rq_replay_list) {
1356                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1357                         dup = 1;
1358                         break;
1359                 }
1360         }
1361
1362         if (dup) {
1363                 /* We expect it with RESENT and REPLAY flags. */
1364                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1365                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1366                         CERROR("invalid flags %x of resent replay\n",
1367                                lustre_msg_get_flags(req->rq_reqmsg));
1368         } else {
1369                 cfs_list_add_tail(&req->rq_replay_list,
1370                                   &exp->exp_req_replay_queue);
1371         }
1372
1373         spin_unlock(&exp->exp_lock);
1374         return dup;
1375 }
1376
1377 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1378 {
1379         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1380         LASSERT(req->rq_export);
1381
1382         spin_lock(&req->rq_export->exp_lock);
1383         cfs_list_del_init(&req->rq_replay_list);
1384         spin_unlock(&req->rq_export->exp_lock);
1385 }
1386
1387 #ifdef __KERNEL__
1388 static void target_finish_recovery(struct obd_device *obd)
1389 {
1390         ENTRY;
1391
1392         /* Only log a recovery message when recovery has occurred. */
1393         if (obd->obd_recovery_start) {
1394                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1395                                         obd->obd_recovery_start);
1396                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1397                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1398                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1399                         obd->obd_max_recoverable_clients,
1400                         cfs_atomic_read(&obd->obd_connected_clients),
1401                         obd->obd_stale_clients,
1402                         obd->obd_stale_clients == 1 ? "was" : "were");
1403         }
1404
1405         ldlm_reprocess_all_ns(obd->obd_namespace);
1406         spin_lock(&obd->obd_recovery_task_lock);
1407         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1408             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1409             !cfs_list_empty(&obd->obd_final_req_queue)) {
1410                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1411                        obd->obd_name,
1412                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1413                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1414                                "" : "lock ",
1415                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1416                                "" : "final ");
1417                 spin_unlock(&obd->obd_recovery_task_lock);
1418                 LBUG();
1419         }
1420         spin_unlock(&obd->obd_recovery_task_lock);
1421
1422         obd->obd_recovery_end = cfs_time_current_sec();
1423
1424         /* When recovery finished, cleanup orphans on MDS and OST. */
1425         if (OBT(obd) && OBP(obd, postrecov)) {
1426                 int rc = OBP(obd, postrecov)(obd);
1427                 if (rc < 0)
1428                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1429                                       obd->obd_name, rc);
1430         }
1431         EXIT;
1432 }
1433
1434 static void abort_req_replay_queue(struct obd_device *obd)
1435 {
1436         struct ptlrpc_request *req, *n;
1437         cfs_list_t abort_list;
1438
1439         CFS_INIT_LIST_HEAD(&abort_list);
1440         spin_lock(&obd->obd_recovery_task_lock);
1441         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1442         spin_unlock(&obd->obd_recovery_task_lock);
1443         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1444                 DEBUG_REQ(D_WARNING, req, "aborted:");
1445                 req->rq_status = -ENOTCONN;
1446                 if (ptlrpc_error(req)) {
1447                         DEBUG_REQ(D_ERROR, req,
1448                                   "failed abort_req_reply; skipping");
1449                 }
1450                 target_exp_dequeue_req_replay(req);
1451                 target_request_copy_put(req);
1452         }
1453 }
1454
1455 static void abort_lock_replay_queue(struct obd_device *obd)
1456 {
1457         struct ptlrpc_request *req, *n;
1458         cfs_list_t abort_list;
1459
1460         CFS_INIT_LIST_HEAD(&abort_list);
1461         spin_lock(&obd->obd_recovery_task_lock);
1462         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1463         spin_unlock(&obd->obd_recovery_task_lock);
1464         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1465                 DEBUG_REQ(D_ERROR, req, "aborted:");
1466                 req->rq_status = -ENOTCONN;
1467                 if (ptlrpc_error(req)) {
1468                         DEBUG_REQ(D_ERROR, req,
1469                                   "failed abort_lock_reply; skipping");
1470                 }
1471                 target_request_copy_put(req);
1472         }
1473 }
1474
1475 /* Called from a cleanup function if the device is being cleaned up
1476    forcefully.  The exports should all have been disconnected already,
1477    the only thing left to do is
1478      - clear the recovery flags
1479      - cancel the timer
1480      - free queued requests and replies, but don't send replies
1481    Because the obd_stopping flag is set, no new requests should be received.
1482
1483 */
1484 void target_cleanup_recovery(struct obd_device *obd)
1485 {
1486         struct ptlrpc_request *req, *n;
1487         cfs_list_t clean_list;
1488         ENTRY;
1489
1490         CFS_INIT_LIST_HEAD(&clean_list);
1491         spin_lock(&obd->obd_dev_lock);
1492         if (!obd->obd_recovering) {
1493                 spin_unlock(&obd->obd_dev_lock);
1494                 EXIT;
1495                 return;
1496         }
1497         obd->obd_recovering = obd->obd_abort_recovery = 0;
1498         spin_unlock(&obd->obd_dev_lock);
1499
1500         spin_lock(&obd->obd_recovery_task_lock);
1501         target_cancel_recovery_timer(obd);
1502         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1503         spin_unlock(&obd->obd_recovery_task_lock);
1504
1505         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1506                 LASSERT(req->rq_reply_state == 0);
1507                 target_exp_dequeue_req_replay(req);
1508                 target_request_copy_put(req);
1509         }
1510
1511         spin_lock(&obd->obd_recovery_task_lock);
1512         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1513         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1514         spin_unlock(&obd->obd_recovery_task_lock);
1515
1516         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1517                 LASSERT(req->rq_reply_state == 0);
1518                 target_request_copy_put(req);
1519         }
1520
1521         EXIT;
1522 }
1523 EXPORT_SYMBOL(target_cleanup_recovery);
1524
1525 /* obd_recovery_task_lock should be held */
1526 void target_cancel_recovery_timer(struct obd_device *obd)
1527 {
1528         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1529         cfs_timer_disarm(&obd->obd_recovery_timer);
1530 }
1531 EXPORT_SYMBOL(target_cancel_recovery_timer);
1532
1533 static void target_start_recovery_timer(struct obd_device *obd)
1534 {
1535         if (obd->obd_recovery_start != 0)
1536                 return;
1537
1538         spin_lock(&obd->obd_dev_lock);
1539         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1540                 spin_unlock(&obd->obd_dev_lock);
1541                 return;
1542         }
1543
1544         LASSERT(obd->obd_recovery_timeout != 0);
1545
1546         if (obd->obd_recovery_start != 0) {
1547                 spin_unlock(&obd->obd_dev_lock);
1548                 return;
1549         }
1550
1551         cfs_timer_arm(&obd->obd_recovery_timer,
1552                       cfs_time_shift(obd->obd_recovery_timeout));
1553         obd->obd_recovery_start = cfs_time_current_sec();
1554         spin_unlock(&obd->obd_dev_lock);
1555
1556         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1557                       "or until %d client%s reconnect%s\n",
1558                       obd->obd_name,
1559                       obd->obd_recovery_timeout / 60,
1560                       obd->obd_recovery_timeout % 60,
1561                       obd->obd_max_recoverable_clients,
1562                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1563                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1564 }
1565
1566 /**
1567  * extend recovery window.
1568  *
1569  * if @extend is true, extend recovery window to have @drt remaining at least;
1570  * otherwise, make sure the recovery timeout value is not less than @drt.
1571  */
1572 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1573 {
1574         cfs_time_t now;
1575         cfs_time_t end;
1576         cfs_duration_t left;
1577         int to;
1578
1579         spin_lock(&obd->obd_dev_lock);
1580         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1581                 spin_unlock(&obd->obd_dev_lock);
1582                 return;
1583         }
1584         LASSERT(obd->obd_recovery_start != 0);
1585
1586         now  = cfs_time_current_sec();
1587         to   = obd->obd_recovery_timeout;
1588         end  = obd->obd_recovery_start + to;
1589         left = cfs_time_sub(end, now);
1590
1591         if (extend && (drt > left)) {
1592                 to += drt - left;
1593         } else if (!extend && (drt > to)) {
1594                 to = drt;
1595                 /* reduce drt by already passed time */
1596                 drt -= obd->obd_recovery_timeout - left;
1597         }
1598
1599         if (to > obd->obd_recovery_time_hard)
1600                 to = obd->obd_recovery_time_hard;
1601         if (obd->obd_recovery_timeout < to ||
1602             obd->obd_recovery_timeout == obd->obd_recovery_time_hard) {
1603                 obd->obd_recovery_timeout = to;
1604                 cfs_timer_arm(&obd->obd_recovery_timer,
1605                               cfs_time_shift(drt));
1606         }
1607         spin_unlock(&obd->obd_dev_lock);
1608
1609         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1610                obd->obd_name, (unsigned)drt);
1611 }
1612
1613 /* Reset the timer with each new client connection */
1614 /*
1615  * This timer is actually reconnect_timer, which is for making sure
1616  * the total recovery window is at least as big as my reconnect
1617  * attempt timing. So the initial recovery time_out will be set to
1618  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1619  * from client is bigger than this, then the recovery time_out will
1620  * be extended to make sure the client could be reconnected, in the
1621  * process, the timeout from the new client should be ignored.
1622  */
1623
1624 static void
1625 check_and_start_recovery_timer(struct obd_device *obd,
1626                                struct ptlrpc_request *req,
1627                                int new_client)
1628 {
1629         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1630         struct obd_device_target *obt = &obd->u.obt;
1631
1632         if (!new_client && service_time)
1633                 /* Teach server about old server's estimates, as first guess
1634                  * at how long new requests will take. */
1635                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1636                             service_time);
1637
1638         target_start_recovery_timer(obd);
1639
1640         /* Convert the service time to RPC timeout,
1641          * and reuse service_time to limit stack usage. */
1642         service_time = at_est2timeout(service_time);
1643
1644         /* We expect other clients to timeout within service_time, then try
1645          * to reconnect, then try the failover server.  The max delay between
1646          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1647         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1648
1649         LASSERT(obt->obt_magic == OBT_MAGIC);
1650         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1651         if (service_time > obd->obd_recovery_timeout && !new_client)
1652                 extend_recovery_timer(obd, service_time, false);
1653 }
1654
1655 /** Health checking routines */
1656 static inline int exp_connect_healthy(struct obd_export *exp)
1657 {
1658         return (exp->exp_in_recovery);
1659 }
1660
1661 /** if export done req_replay or has replay in queue */
1662 static inline int exp_req_replay_healthy(struct obd_export *exp)
1663 {
1664         return (!exp->exp_req_replay_needed ||
1665                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1666 }
1667 /** if export done lock_replay or has replay in queue */
1668 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1669 {
1670         return (!exp->exp_lock_replay_needed ||
1671                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1672 }
1673
1674 static inline int exp_vbr_healthy(struct obd_export *exp)
1675 {
1676         return (!exp->exp_vbr_failed);
1677 }
1678
1679 static inline int exp_finished(struct obd_export *exp)
1680 {
1681         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1682 }
1683
1684 /** Checking routines for recovery */
1685 static int check_for_clients(struct obd_device *obd)
1686 {
1687         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1688
1689         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1690                 return 1;
1691         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1692         return (clnts + obd->obd_stale_clients ==
1693                 obd->obd_max_recoverable_clients);
1694 }
1695
1696 static int check_for_next_transno(struct obd_device *obd)
1697 {
1698         struct ptlrpc_request *req = NULL;
1699         int wake_up = 0, connected, completed, queue_len;
1700         __u64 next_transno, req_transno;
1701         ENTRY;
1702
1703         spin_lock(&obd->obd_recovery_task_lock);
1704         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1705                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1706                                      struct ptlrpc_request, rq_list);
1707                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1708         } else {
1709                 req_transno = 0;
1710         }
1711
1712         connected = cfs_atomic_read(&obd->obd_connected_clients);
1713         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1714         queue_len = obd->obd_requests_queued_for_recovery;
1715         next_transno = obd->obd_next_recovery_transno;
1716
1717         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1718                "req_transno: "LPU64", next_transno: "LPU64"\n",
1719                obd->obd_max_recoverable_clients, connected, completed,
1720                queue_len, req_transno, next_transno);
1721
1722         if (obd->obd_abort_recovery) {
1723                 CDEBUG(D_HA, "waking for aborted recovery\n");
1724                 wake_up = 1;
1725         } else if (obd->obd_recovery_expired) {
1726                 CDEBUG(D_HA, "waking for expired recovery\n");
1727                 wake_up = 1;
1728         } else if (req_transno == next_transno) {
1729                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1730                 wake_up = 1;
1731         } else if (queue_len > 0 &&
1732                    queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1733                 int d_lvl = D_HA;
1734                 /** handle gaps occured due to lost reply or VBR */
1735                 LASSERTF(req_transno >= next_transno,
1736                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1737                          req_transno, next_transno);
1738                 if (req_transno > obd->obd_last_committed &&
1739                     !obd->obd_version_recov)
1740                         d_lvl = D_ERROR;
1741                 CDEBUG(d_lvl,
1742                        "%s: waking for gap in transno, VBR is %s (skip: "
1743                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1744                        ", last_committed: "LPD64")\n",
1745                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1746                        next_transno, queue_len, completed, connected,
1747                        req_transno, obd->obd_last_committed);
1748                 obd->obd_next_recovery_transno = req_transno;
1749                 wake_up = 1;
1750         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1751                 CDEBUG(D_HA, "waking for completed recovery\n");
1752                 wake_up = 1;
1753         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1754                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1755                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1756                 obd->obd_next_recovery_transno = req_transno;
1757                 wake_up = 1;
1758         }
1759         spin_unlock(&obd->obd_recovery_task_lock);
1760         return wake_up;
1761 }
1762
1763 static int check_for_next_lock(struct obd_device *obd)
1764 {
1765         int wake_up = 0;
1766
1767         spin_lock(&obd->obd_recovery_task_lock);
1768         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1769                 CDEBUG(D_HA, "waking for next lock\n");
1770                 wake_up = 1;
1771         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1772                 CDEBUG(D_HA, "waking for completed lock replay\n");
1773                 wake_up = 1;
1774         } else if (obd->obd_abort_recovery) {
1775                 CDEBUG(D_HA, "waking for aborted recovery\n");
1776                 wake_up = 1;
1777         } else if (obd->obd_recovery_expired) {
1778                 CDEBUG(D_HA, "waking for expired recovery\n");
1779                 wake_up = 1;
1780         }
1781         spin_unlock(&obd->obd_recovery_task_lock);
1782
1783         return wake_up;
1784 }
1785
1786 /**
1787  * wait for recovery events,
1788  * check its status with help of check_routine
1789  * evict dead clients via health_check
1790  */
1791 static int target_recovery_overseer(struct obd_device *obd,
1792                                     int (*check_routine)(struct obd_device *),
1793                                     int (*health_check)(struct obd_export *))
1794 {
1795 repeat:
1796         wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1797         if (obd->obd_abort_recovery) {
1798                 CWARN("recovery is aborted, evict exports in recovery\n");
1799                 /** evict exports which didn't finish recovery yet */
1800                 class_disconnect_stale_exports(obd, exp_finished);
1801                 return 1;
1802         } else if (obd->obd_recovery_expired) {
1803                 obd->obd_recovery_expired = 0;
1804                 /** If some clients died being recovered, evict them */
1805                 LCONSOLE_WARN("%s: recovery is timed out, "
1806                               "evict stale exports\n", obd->obd_name);
1807                 /** evict cexports with no replay in queue, they are stalled */
1808                 class_disconnect_stale_exports(obd, health_check);
1809                 /** continue with VBR */
1810                 spin_lock(&obd->obd_dev_lock);
1811                 obd->obd_version_recov = 1;
1812                 spin_unlock(&obd->obd_dev_lock);
1813                 /**
1814                  * reset timer, recovery will proceed with versions now,
1815                  * timeout is set just to handle reconnection delays
1816                  */
1817                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1818                 /** Wait for recovery events again, after evicting bad clients */
1819                 goto repeat;
1820         }
1821         return 0;
1822 }
1823
1824 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1825 {
1826         struct ptlrpc_request *req = NULL;
1827         ENTRY;
1828
1829         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1830                obd->obd_next_recovery_transno);
1831
1832         if (target_recovery_overseer(obd, check_for_next_transno,
1833                                      exp_req_replay_healthy)) {
1834                 abort_req_replay_queue(obd);
1835                 abort_lock_replay_queue(obd);
1836         }
1837
1838         spin_lock(&obd->obd_recovery_task_lock);
1839         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1840                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1841                                      struct ptlrpc_request, rq_list);
1842                 cfs_list_del_init(&req->rq_list);
1843                 obd->obd_requests_queued_for_recovery--;
1844                 spin_unlock(&obd->obd_recovery_task_lock);
1845         } else {
1846                 spin_unlock(&obd->obd_recovery_task_lock);
1847                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1848                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1849                 /** evict exports failed VBR */
1850                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1851         }
1852         RETURN(req);
1853 }
1854
1855 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1856 {
1857         struct ptlrpc_request *req = NULL;
1858
1859         CDEBUG(D_HA, "Waiting for lock\n");
1860         if (target_recovery_overseer(obd, check_for_next_lock,
1861                                      exp_lock_replay_healthy))
1862                 abort_lock_replay_queue(obd);
1863
1864         spin_lock(&obd->obd_recovery_task_lock);
1865         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1866                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1867                                      struct ptlrpc_request, rq_list);
1868                 cfs_list_del_init(&req->rq_list);
1869                 spin_unlock(&obd->obd_recovery_task_lock);
1870         } else {
1871                 spin_unlock(&obd->obd_recovery_task_lock);
1872                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1873                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1874                 /** evict exports failed VBR */
1875                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1876         }
1877         return req;
1878 }
1879
1880 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1881 {
1882         struct ptlrpc_request *req = NULL;
1883
1884         spin_lock(&obd->obd_recovery_task_lock);
1885         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1886                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1887                                      struct ptlrpc_request, rq_list);
1888                 cfs_list_del_init(&req->rq_list);
1889                 spin_unlock(&obd->obd_recovery_task_lock);
1890                 if (req->rq_export->exp_in_recovery) {
1891                         spin_lock(&req->rq_export->exp_lock);
1892                         req->rq_export->exp_in_recovery = 0;
1893                         spin_unlock(&req->rq_export->exp_lock);
1894                 }
1895         } else {
1896                 spin_unlock(&obd->obd_recovery_task_lock);
1897         }
1898         return req;
1899 }
1900
1901 static int handle_recovery_req(struct ptlrpc_thread *thread,
1902                                struct ptlrpc_request *req,
1903                                svc_handler_t handler)
1904 {
1905         int rc;
1906         ENTRY;
1907
1908         /**
1909          * export can be evicted during recovery, no need to handle replays for
1910          * it after that, discard such request silently
1911          */
1912         if (req->rq_export->exp_disconnected)
1913                 GOTO(reqcopy_put, rc = 0);
1914
1915         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1916         if (rc) {
1917                 CERROR("Failure to initialize session: %d\n", rc);
1918                 GOTO(reqcopy_put, rc);
1919         }
1920
1921         req->rq_recov_session.lc_thread = thread;
1922         lu_context_enter(&req->rq_recov_session);
1923         req->rq_svc_thread = thread;
1924         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1925
1926         /* thread context */
1927         lu_context_enter(&thread->t_env->le_ctx);
1928         (void)handler(req);
1929         lu_context_exit(&thread->t_env->le_ctx);
1930
1931         lu_context_exit(&req->rq_recov_session);
1932         lu_context_fini(&req->rq_recov_session);
1933         /* don't reset timer for final stage */
1934         if (!exp_finished(req->rq_export)) {
1935                 int to = obd_timeout;
1936
1937                 /**
1938                  * Add request timeout to the recovery time so next request from
1939                  * this client may come in recovery time
1940                  */
1941                 if (!AT_OFF) {
1942                         struct ptlrpc_service_part *svcpt;
1943
1944                         svcpt = req->rq_rqbd->rqbd_svcpt;
1945                         /* If the server sent early reply for this request,
1946                          * the client will recalculate the timeout according to
1947                          * current server estimate service time, so we will
1948                          * use the maxium timeout here for waiting the client
1949                          * sending the next req */
1950                         to = max((int)at_est2timeout(
1951                                  at_get(&svcpt->scp_at_estimate)),
1952                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1953                         /* Add 2 net_latency, one for balance rq_deadline
1954                          * (see ptl_send_rpc), one for resend the req to server,
1955                          * Note: client will pack net_latency in replay req
1956                          * (see ptlrpc_replay_req) */
1957                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1958                 }
1959                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1960         }
1961 reqcopy_put:
1962         RETURN(rc);
1963 }
1964
1965 static int target_recovery_thread(void *arg)
1966 {
1967         struct lu_target *lut = arg;
1968         struct obd_device *obd = lut->lut_obd;
1969         struct ptlrpc_request *req;
1970         struct target_recovery_data *trd = &obd->obd_recovery_data;
1971         unsigned long delta;
1972         struct lu_env *env;
1973         struct ptlrpc_thread *thread = NULL;
1974         int rc = 0;
1975         ENTRY;
1976
1977         unshare_fs_struct();
1978         OBD_ALLOC_PTR(thread);
1979         if (thread == NULL)
1980                 RETURN(-ENOMEM);
1981
1982         OBD_ALLOC_PTR(env);
1983         if (env == NULL) {
1984                 OBD_FREE_PTR(thread);
1985                 RETURN(-ENOMEM);
1986         }
1987
1988         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1989         if (rc) {
1990                 OBD_FREE_PTR(thread);
1991                 OBD_FREE_PTR(env);
1992                 RETURN(rc);
1993         }
1994
1995         thread->t_env = env;
1996         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1997         env->le_ctx.lc_thread = thread;
1998         thread->t_data = NULL;
1999         thread->t_watchdog = NULL;
2000
2001         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2002                current_pid());
2003         trd->trd_processing_task = current_pid();
2004
2005         spin_lock(&obd->obd_dev_lock);
2006         obd->obd_recovering = 1;
2007         spin_unlock(&obd->obd_dev_lock);
2008         complete(&trd->trd_starting);
2009
2010         /* first of all, we have to know the first transno to replay */
2011         if (target_recovery_overseer(obd, check_for_clients,
2012                                      exp_connect_healthy)) {
2013                 abort_req_replay_queue(obd);
2014                 abort_lock_replay_queue(obd);
2015         }
2016
2017         /* next stage: replay requests */
2018         delta = jiffies;
2019         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2020                cfs_atomic_read(&obd->obd_req_replay_clients),
2021                obd->obd_next_recovery_transno);
2022         while ((req = target_next_replay_req(obd))) {
2023                 LASSERT(trd->trd_processing_task == current_pid());
2024                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2025                           lustre_msg_get_transno(req->rq_reqmsg),
2026                           libcfs_nid2str(req->rq_peer.nid));
2027                 handle_recovery_req(thread, req,
2028                                     trd->trd_recovery_handler);
2029                 /**
2030                  * bz18031: increase next_recovery_transno before
2031                  * target_request_copy_put() will drop exp_rpc reference
2032                  */
2033                 spin_lock(&obd->obd_recovery_task_lock);
2034                 obd->obd_next_recovery_transno++;
2035                 spin_unlock(&obd->obd_recovery_task_lock);
2036                 target_exp_dequeue_req_replay(req);
2037                 target_request_copy_put(req);
2038                 obd->obd_replayed_requests++;
2039         }
2040
2041         /**
2042          * The second stage: replay locks
2043          */
2044         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2045                cfs_atomic_read(&obd->obd_lock_replay_clients));
2046         while ((req = target_next_replay_lock(obd))) {
2047                 LASSERT(trd->trd_processing_task == current_pid());
2048                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2049                           libcfs_nid2str(req->rq_peer.nid));
2050                 handle_recovery_req(thread, req,
2051                                     trd->trd_recovery_handler);
2052                 target_request_copy_put(req);
2053                 obd->obd_replayed_locks++;
2054         }
2055
2056         /**
2057          * The third stage: reply on final pings, at this moment all clients
2058          * must have request in final queue
2059          */
2060         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2061         /** Update server last boot epoch */
2062         tgt_boot_epoch_update(lut);
2063         /* We drop recoverying flag to forward all new requests
2064          * to regular mds_handle() since now */
2065         spin_lock(&obd->obd_dev_lock);
2066         obd->obd_recovering = obd->obd_abort_recovery = 0;
2067         spin_unlock(&obd->obd_dev_lock);
2068         spin_lock(&obd->obd_recovery_task_lock);
2069         target_cancel_recovery_timer(obd);
2070         spin_unlock(&obd->obd_recovery_task_lock);
2071         while ((req = target_next_final_ping(obd))) {
2072                 LASSERT(trd->trd_processing_task == current_pid());
2073                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2074                           libcfs_nid2str(req->rq_peer.nid));
2075                 handle_recovery_req(thread, req,
2076                                     trd->trd_recovery_handler);
2077                 target_request_copy_put(req);
2078         }
2079
2080         delta = (jiffies - delta) / HZ;
2081         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2082               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2083         if (delta > OBD_RECOVERY_TIME_SOFT) {
2084                 CWARN("too long recovery - read logs\n");
2085                 libcfs_debug_dumplog();
2086         }
2087
2088         target_finish_recovery(obd);
2089
2090         lu_context_fini(&env->le_ctx);
2091         trd->trd_processing_task = 0;
2092         complete(&trd->trd_finishing);
2093
2094         OBD_FREE_PTR(thread);
2095         OBD_FREE_PTR(env);
2096         RETURN(rc);
2097 }
2098
2099 static int target_start_recovery_thread(struct lu_target *lut,
2100                                         svc_handler_t handler)
2101 {
2102         struct obd_device *obd = lut->lut_obd;
2103         int rc = 0;
2104         struct target_recovery_data *trd = &obd->obd_recovery_data;
2105
2106         memset(trd, 0, sizeof(*trd));
2107         init_completion(&trd->trd_starting);
2108         init_completion(&trd->trd_finishing);
2109         trd->trd_recovery_handler = handler;
2110
2111         if (!IS_ERR(kthread_run(target_recovery_thread,
2112                                 lut, "tgt_recov"))) {
2113                 wait_for_completion(&trd->trd_starting);
2114                 LASSERT(obd->obd_recovering != 0);
2115         } else {
2116                 rc = -ECHILD;
2117         }
2118
2119         return rc;
2120 }
2121
2122 void target_stop_recovery_thread(struct obd_device *obd)
2123 {
2124         if (obd->obd_recovery_data.trd_processing_task > 0) {
2125                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2126                 /** recovery can be done but postrecovery is not yet */
2127                 spin_lock(&obd->obd_dev_lock);
2128                 if (obd->obd_recovering) {
2129                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2130                         obd->obd_abort_recovery = 1;
2131                         wake_up(&obd->obd_next_transno_waitq);
2132                 }
2133                 spin_unlock(&obd->obd_dev_lock);
2134                 wait_for_completion(&trd->trd_finishing);
2135         }
2136 }
2137 EXPORT_SYMBOL(target_stop_recovery_thread);
2138
2139 void target_recovery_fini(struct obd_device *obd)
2140 {
2141         class_disconnect_exports(obd);
2142         target_stop_recovery_thread(obd);
2143         target_cleanup_recovery(obd);
2144 }
2145 EXPORT_SYMBOL(target_recovery_fini);
2146
2147 static void target_recovery_expired(unsigned long castmeharder)
2148 {
2149         struct obd_device *obd = (struct obd_device *)castmeharder;
2150         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2151                " after %lds (%d clients connected)\n",
2152                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2153                cfs_time_current_sec()- obd->obd_recovery_start,
2154                cfs_atomic_read(&obd->obd_connected_clients));
2155
2156         obd->obd_recovery_expired = 1;
2157         wake_up(&obd->obd_next_transno_waitq);
2158 }
2159
2160 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2161 {
2162         struct obd_device *obd = lut->lut_obd;
2163         if (obd->obd_max_recoverable_clients == 0) {
2164                 /** Update server last boot epoch */
2165                 tgt_boot_epoch_update(lut);
2166                 return;
2167         }
2168
2169         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2170                "last_transno "LPU64"\n", obd->obd_name,
2171                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2172         LASSERT(obd->obd_stopping == 0);
2173         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2174         obd->obd_recovery_start = 0;
2175         obd->obd_recovery_end = 0;
2176
2177         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2178         target_start_recovery_thread(lut, handler);
2179 }
2180 EXPORT_SYMBOL(target_recovery_init);
2181
2182 #endif /* __KERNEL__ */
2183
2184 static int target_process_req_flags(struct obd_device *obd,
2185                                     struct ptlrpc_request *req)
2186 {
2187         struct obd_export *exp = req->rq_export;
2188         LASSERT(exp != NULL);
2189         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2190                 /* client declares he's ready to replay locks */
2191                 spin_lock(&exp->exp_lock);
2192                 if (exp->exp_req_replay_needed) {
2193                         exp->exp_req_replay_needed = 0;
2194                         spin_unlock(&exp->exp_lock);
2195
2196                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2197                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2198                 } else {
2199                         spin_unlock(&exp->exp_lock);
2200                 }
2201         }
2202         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2203                 /* client declares he's ready to complete recovery
2204                  * so, we put the request on th final queue */
2205                 spin_lock(&exp->exp_lock);
2206                 if (exp->exp_lock_replay_needed) {
2207                         exp->exp_lock_replay_needed = 0;
2208                         spin_unlock(&exp->exp_lock);
2209
2210                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2211                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2212                 } else {
2213                         spin_unlock(&exp->exp_lock);
2214                 }
2215         }
2216         return 0;
2217 }
2218
2219 int target_queue_recovery_request(struct ptlrpc_request *req,
2220                                   struct obd_device *obd)
2221 {
2222         cfs_list_t *tmp;
2223         int inserted = 0;
2224         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2225         ENTRY;
2226
2227         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2228                 /* Processing the queue right now, don't re-add. */
2229                 RETURN(1);
2230         }
2231
2232         target_process_req_flags(obd, req);
2233
2234         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2235                 /* client declares he's ready to complete recovery
2236                  * so, we put the request on th final queue */
2237                 target_request_copy_get(req);
2238                 DEBUG_REQ(D_HA, req, "queue final req");
2239                 wake_up(&obd->obd_next_transno_waitq);
2240                 spin_lock(&obd->obd_recovery_task_lock);
2241                 if (obd->obd_recovering) {
2242                         cfs_list_add_tail(&req->rq_list,
2243                                           &obd->obd_final_req_queue);
2244                 } else {
2245                         spin_unlock(&obd->obd_recovery_task_lock);
2246                         target_request_copy_put(req);
2247                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2248                 }
2249                 spin_unlock(&obd->obd_recovery_task_lock);
2250                 RETURN(0);
2251         }
2252         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2253                 /* client declares he's ready to replay locks */
2254                 target_request_copy_get(req);
2255                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2256                 wake_up(&obd->obd_next_transno_waitq);
2257                 spin_lock(&obd->obd_recovery_task_lock);
2258                 LASSERT(obd->obd_recovering);
2259                 /* usually due to recovery abort */
2260                 if (!req->rq_export->exp_in_recovery) {
2261                         spin_unlock(&obd->obd_recovery_task_lock);
2262                         target_request_copy_put(req);
2263                         RETURN(-ENOTCONN);
2264                 }
2265                 LASSERT(req->rq_export->exp_lock_replay_needed);
2266                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2267                 spin_unlock(&obd->obd_recovery_task_lock);
2268                 RETURN(0);
2269         }
2270
2271         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2272          * (i.e. buflens etc are in my own byte order), but type-dependent
2273          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2274
2275         if (!transno) {
2276                 CFS_INIT_LIST_HEAD(&req->rq_list);
2277                 DEBUG_REQ(D_HA, req, "not queueing");
2278                 RETURN(1);
2279         }
2280
2281         /* If we're processing the queue, we want don't want to queue this
2282          * message.
2283          *
2284          * Also, if this request has a transno less than the one we're waiting
2285          * for, we should process it now.  It could (and currently always will)
2286          * be an open request for a descriptor that was opened some time ago.
2287          *
2288          * Also, a resent, replayed request that has already been
2289          * handled will pass through here and be processed immediately.
2290          */
2291         CDEBUG(D_HA, "Next recovery transno: "LPU64
2292                ", current: "LPU64", replaying\n",
2293                obd->obd_next_recovery_transno, transno);
2294         spin_lock(&obd->obd_recovery_task_lock);
2295         if (transno < obd->obd_next_recovery_transno) {
2296                 /* Processing the queue right now, don't re-add. */
2297                 LASSERT(cfs_list_empty(&req->rq_list));
2298                 spin_unlock(&obd->obd_recovery_task_lock);
2299                 RETURN(1);
2300         }
2301         spin_unlock(&obd->obd_recovery_task_lock);
2302
2303         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2304                 RETURN(0);
2305
2306         target_request_copy_get(req);
2307         if (!req->rq_export->exp_in_recovery) {
2308                 target_request_copy_put(req);
2309                 RETURN(-ENOTCONN);
2310         }
2311         LASSERT(req->rq_export->exp_req_replay_needed);
2312
2313         if (target_exp_enqueue_req_replay(req)) {
2314                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2315                 target_request_copy_put(req);
2316                 RETURN(0);
2317         }
2318
2319         /* XXX O(n^2) */
2320         spin_lock(&obd->obd_recovery_task_lock);
2321         LASSERT(obd->obd_recovering);
2322         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2323                 struct ptlrpc_request *reqiter =
2324                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2325
2326                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2327                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2328                         inserted = 1;
2329                         break;
2330                 }
2331
2332                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2333                              transno)) {
2334                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2335                                   "has been claimed by another client");
2336                         spin_unlock(&obd->obd_recovery_task_lock);
2337                         target_exp_dequeue_req_replay(req);
2338                         target_request_copy_put(req);
2339                         RETURN(0);
2340                 }
2341         }
2342
2343         if (!inserted)
2344                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2345
2346         obd->obd_requests_queued_for_recovery++;
2347         spin_unlock(&obd->obd_recovery_task_lock);
2348         wake_up(&obd->obd_next_transno_waitq);
2349         RETURN(0);
2350 }
2351 EXPORT_SYMBOL(target_queue_recovery_request);
2352
2353 int target_handle_ping(struct ptlrpc_request *req)
2354 {
2355         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2356         return req_capsule_server_pack(&req->rq_pill);
2357 }
2358 EXPORT_SYMBOL(target_handle_ping);
2359
2360 void target_committed_to_req(struct ptlrpc_request *req)
2361 {
2362         struct obd_export *exp = req->rq_export;
2363
2364         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2365                 lustre_msg_set_last_committed(req->rq_repmsg,
2366                                               exp->exp_last_committed);
2367         else
2368                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2369                           "%d)", exp->exp_obd->obd_no_transno,
2370                           req->rq_repmsg == NULL);
2371
2372         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2373                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2374 }
2375 EXPORT_SYMBOL(target_committed_to_req);
2376
2377 #endif /* HAVE_SERVER_SUPPORT */
2378
2379 /**
2380  * Packs current SLV and Limit into \a req.
2381  */
2382 int target_pack_pool_reply(struct ptlrpc_request *req)
2383 {
2384         struct obd_device *obd;
2385         ENTRY;
2386
2387         /* Check that we still have all structures alive as this may
2388          * be some late RPC at shutdown time. */
2389         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2390                      !exp_connect_lru_resize(req->rq_export))) {
2391                 lustre_msg_set_slv(req->rq_repmsg, 0);
2392                 lustre_msg_set_limit(req->rq_repmsg, 0);
2393                 RETURN(0);
2394         }
2395
2396         /* OBD is alive here as export is alive, which we checked above. */
2397         obd = req->rq_export->exp_obd;
2398
2399         read_lock(&obd->obd_pool_lock);
2400         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2401         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2402         read_unlock(&obd->obd_pool_lock);
2403
2404         RETURN(0);
2405 }
2406 EXPORT_SYMBOL(target_pack_pool_reply);
2407
2408 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2409 {
2410         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2411                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2412                 return (-ECOMM);
2413         }
2414
2415         if (unlikely(rc)) {
2416                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2417                 req->rq_status = rc;
2418                 return (ptlrpc_send_error(req, 1));
2419         } else {
2420                 DEBUG_REQ(D_NET, req, "sending reply");
2421         }
2422
2423         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2424 }
2425
2426 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2427 {
2428         struct ptlrpc_service_part *svcpt;
2429         int                        netrc;
2430         struct ptlrpc_reply_state *rs;
2431         struct obd_export         *exp;
2432         ENTRY;
2433
2434         if (req->rq_no_reply) {
2435                 EXIT;
2436                 return;
2437         }
2438
2439         svcpt = req->rq_rqbd->rqbd_svcpt;
2440         rs = req->rq_reply_state;
2441         if (rs == NULL || !rs->rs_difficult) {
2442                 /* no notifiers */
2443                 target_send_reply_msg (req, rc, fail_id);
2444                 EXIT;
2445                 return;
2446         }
2447
2448         /* must be an export if locks saved */
2449         LASSERT (req->rq_export != NULL);
2450         /* req/reply consistent */
2451         LASSERT(rs->rs_svcpt == svcpt);
2452
2453         /* "fresh" reply */
2454         LASSERT (!rs->rs_scheduled);
2455         LASSERT (!rs->rs_scheduled_ever);
2456         LASSERT (!rs->rs_handled);
2457         LASSERT (!rs->rs_on_net);
2458         LASSERT (rs->rs_export == NULL);
2459         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2460         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2461
2462         exp = class_export_get (req->rq_export);
2463
2464         /* disable reply scheduling while I'm setting up */
2465         rs->rs_scheduled = 1;
2466         rs->rs_on_net    = 1;
2467         rs->rs_xid       = req->rq_xid;
2468         rs->rs_transno   = req->rq_transno;
2469         rs->rs_export    = exp;
2470         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2471
2472         spin_lock(&exp->exp_uncommitted_replies_lock);
2473         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2474                rs->rs_transno, exp->exp_last_committed);
2475         if (rs->rs_transno > exp->exp_last_committed) {
2476                 /* not committed already */
2477                 cfs_list_add_tail(&rs->rs_obd_list,
2478                                   &exp->exp_uncommitted_replies);
2479         }
2480         spin_unlock(&exp->exp_uncommitted_replies_lock);
2481
2482         spin_lock(&exp->exp_lock);
2483         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2484         spin_unlock(&exp->exp_lock);
2485
2486         netrc = target_send_reply_msg(req, rc, fail_id);
2487
2488         spin_lock(&svcpt->scp_rep_lock);
2489
2490         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2491
2492         if (netrc != 0) {
2493                 /* error sending: reply is off the net.  Also we need +1
2494                  * reply ref until ptlrpc_handle_rs() is done
2495                  * with the reply state (if the send was successful, there
2496                  * would have been +1 ref for the net, which
2497                  * reply_out_callback leaves alone) */
2498                 rs->rs_on_net = 0;
2499                 ptlrpc_rs_addref(rs);
2500         }
2501
2502         spin_lock(&rs->rs_lock);
2503         if (rs->rs_transno <= exp->exp_last_committed ||
2504             (!rs->rs_on_net && !rs->rs_no_ack) ||
2505             cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2506             cfs_list_empty(&rs->rs_obd_list)) {
2507                 CDEBUG(D_HA, "Schedule reply immediately\n");
2508                 ptlrpc_dispatch_difficult_reply(rs);
2509         } else {
2510                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2511                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2512         }
2513         spin_unlock(&rs->rs_lock);
2514         spin_unlock(&svcpt->scp_rep_lock);
2515         EXIT;
2516 }
2517 EXPORT_SYMBOL(target_send_reply);
2518
2519 ldlm_mode_t lck_compat_array[] = {
2520         [LCK_EX] LCK_COMPAT_EX,
2521         [LCK_PW] LCK_COMPAT_PW,
2522         [LCK_PR] LCK_COMPAT_PR,
2523         [LCK_CW] LCK_COMPAT_CW,
2524         [LCK_CR] LCK_COMPAT_CR,
2525         [LCK_NL] LCK_COMPAT_NL,
2526         [LCK_GROUP] LCK_COMPAT_GROUP,
2527         [LCK_COS] LCK_COMPAT_COS,
2528 };
2529
2530 /**
2531  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2532  * not escape to the user level.
2533  */
2534 int ldlm_error2errno(ldlm_error_t error)
2535 {
2536         int result;
2537
2538         switch (error) {
2539         case ELDLM_OK:
2540                 result = 0;
2541                 break;
2542         case ELDLM_LOCK_CHANGED:
2543                 result = -ESTALE;
2544                 break;
2545         case ELDLM_LOCK_ABORTED:
2546                 result = -ENAVAIL;
2547                 break;
2548         case ELDLM_LOCK_REPLACED:
2549                 result = -ESRCH;
2550                 break;
2551         case ELDLM_NO_LOCK_DATA:
2552                 result = -ENOENT;
2553                 break;
2554         case ELDLM_NAMESPACE_EXISTS:
2555                 result = -EEXIST;
2556                 break;
2557         case ELDLM_BAD_NAMESPACE:
2558                 result = -EBADF;
2559                 break;
2560         default:
2561                 if (((int)error) < 0)  /* cast to signed type */
2562                         result = error; /* as ldlm_error_t can be unsigned */
2563                 else {
2564                         CERROR("Invalid DLM result code: %d\n", error);
2565                         result = -EPROTO;
2566                 }
2567         }
2568         return result;
2569 }
2570 EXPORT_SYMBOL(ldlm_error2errno);
2571
2572 /**
2573  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2574  */
2575 ldlm_error_t ldlm_errno2error(int err_no)
2576 {
2577         int error;
2578
2579         switch (err_no) {
2580         case 0:
2581                 error = ELDLM_OK;
2582                 break;
2583         case -ESTALE:
2584                 error = ELDLM_LOCK_CHANGED;
2585                 break;
2586         case -ENAVAIL:
2587                 error = ELDLM_LOCK_ABORTED;
2588                 break;
2589         case -ESRCH:
2590                 error = ELDLM_LOCK_REPLACED;
2591                 break;
2592         case -ENOENT:
2593                 error = ELDLM_NO_LOCK_DATA;
2594                 break;
2595         case -EEXIST:
2596                 error = ELDLM_NAMESPACE_EXISTS;
2597                 break;
2598         case -EBADF:
2599                 error = ELDLM_BAD_NAMESPACE;
2600                 break;
2601         default:
2602                 error = err_no;
2603         }
2604         return error;
2605 }
2606 EXPORT_SYMBOL(ldlm_errno2error);
2607
2608 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2609 void ldlm_dump_export_locks(struct obd_export *exp)
2610 {
2611         spin_lock(&exp->exp_locks_list_guard);
2612         if (!cfs_list_empty(&exp->exp_locks_list)) {
2613                 struct ldlm_lock *lock;
2614
2615                 CERROR("dumping locks for export %p,"
2616                        "ignore if the unmount doesn't hang\n", exp);
2617                 cfs_list_for_each_entry(lock, &exp->exp_locks_list,
2618                                         l_exp_refs_link)
2619                         LDLM_ERROR(lock, "lock:");
2620         }
2621         spin_unlock(&exp->exp_locks_list_guard);
2622 }
2623 #endif
2624
2625 #ifdef HAVE_SERVER_SUPPORT
2626 static int target_bulk_timeout(void *data)
2627 {
2628         ENTRY;
2629         /* We don't fail the connection here, because having the export
2630          * killed makes the (vital) call to commitrw very sad.
2631          */
2632         RETURN(1);
2633 }
2634
2635 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2636 {
2637         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2638 }
2639
2640 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2641                    struct l_wait_info *lwi)
2642 {
2643         struct ptlrpc_request *req = desc->bd_req;
2644         int rc = 0;
2645         ENTRY;
2646
2647         /* If there is eviction in progress, wait for it to finish. */
2648         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2649                 *lwi = LWI_INTR(NULL, NULL);
2650                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2651                                   !cfs_atomic_read(&exp->exp_obd->
2652                                                    obd_evict_inprogress),
2653                                   lwi);
2654         }
2655
2656         /* Check if client was evicted or tried to reconnect already. */
2657         if (exp->exp_failed || exp->exp_abort_active_req) {
2658                 rc = -ENOTCONN;
2659         } else {
2660                 if (desc->bd_type == BULK_PUT_SINK)
2661                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2662                 if (rc == 0)
2663                         rc = ptlrpc_start_bulk_transfer(desc);
2664         }
2665
2666         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2667                 ptlrpc_abort_bulk(desc);
2668         } else if (rc == 0) {
2669                 time_t start = cfs_time_current_sec();
2670                 do {
2671                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2672                         cfs_duration_t timeout = timeoutl <= 0 ?
2673                                 CFS_TICK : cfs_time_seconds(timeoutl);
2674                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2675                                                     cfs_time_seconds(1),
2676                                                    target_bulk_timeout,
2677                                                    desc);
2678                         rc = l_wait_event(desc->bd_waitq,
2679                                           !ptlrpc_server_bulk_active(desc) ||
2680                                           exp->exp_failed ||
2681                                           exp->exp_abort_active_req,
2682                                           lwi);
2683                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2684                         /* Wait again if we changed deadline. */
2685                 } while ((rc == -ETIMEDOUT) &&
2686                          (req->rq_deadline > cfs_time_current_sec()));
2687
2688                 if (rc == -ETIMEDOUT) {
2689                         DEBUG_REQ(D_ERROR, req,
2690                                   "timeout on bulk %s after %ld%+lds",
2691                                   bulk2type(desc),
2692                                   req->rq_deadline - start,
2693                                   cfs_time_current_sec() -
2694                                   req->rq_deadline);
2695                         ptlrpc_abort_bulk(desc);
2696                 } else if (exp->exp_failed) {
2697                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2698                                   bulk2type(desc));
2699                         rc = -ENOTCONN;
2700                         ptlrpc_abort_bulk(desc);
2701                 } else if (exp->exp_abort_active_req) {
2702                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2703                                   bulk2type(desc));
2704                         /* We don't reply anyway. */
2705                         rc = -ETIMEDOUT;
2706                         ptlrpc_abort_bulk(desc);
2707                 } else if (desc->bd_failure ||
2708                            desc->bd_nob_transferred != desc->bd_nob) {
2709                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2710                                   desc->bd_failure ?
2711                                   "network error on" : "truncated",
2712                                   bulk2type(desc),
2713                                   desc->bd_nob_transferred,
2714                                   desc->bd_nob);
2715                         /* XXX Should this be a different errno? */
2716                         rc = -ETIMEDOUT;
2717                 } else if (desc->bd_type == BULK_GET_SINK) {
2718                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2719                 }
2720         } else {
2721                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2722                           bulk2type(desc), rc);
2723         }
2724
2725         RETURN(rc);
2726 }
2727 EXPORT_SYMBOL(target_bulk_io);
2728
2729 #endif /* HAVE_SERVER_SUPPORT */