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         mutex_init(&cli->cl_mgc_mutex);
 337         cli->cl_conn_count = 0;
 338         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 339                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 340                      sizeof(server_uuid)));
 341
 342         cli->cl_dirty_pages = 0;
 343         cli->cl_avail_grant = 0;
 344         /* FIXME: Should limit this for the sum of all cl_dirty_max_pages. */
 345         /* cl_dirty_max_pages may be changed at connect time in
 346          * ptlrpc_connect_interpret(). */
 347         client_adjust_max_dirty(cli);
 348         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 349         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 350         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 351         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 352         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 353         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 354         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
 355         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
 356         cli->cl_r_in_flight = 0;
 357         cli->cl_w_in_flight = 0;
 358
 359         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 360         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 361         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 362         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 363         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 364         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 365
 366         /* lru for osc. */
 367         CFS_INIT_LIST_HEAD(&cli->cl_lru_osc);
 368         cfs_atomic_set(&cli->cl_lru_shrinkers, 0);
 369         cfs_atomic_set(&cli->cl_lru_busy, 0);
 370         cfs_atomic_set(&cli->cl_lru_in_list, 0);
 371         CFS_INIT_LIST_HEAD(&cli->cl_lru_list);
 372         client_obd_list_lock_init(&cli->cl_lru_list_lock);
 373
 374         init_waitqueue_head(&cli->cl_destroy_waitq);
 375         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 376 #ifdef ENABLE_CHECKSUM
 377         /* Turn on checksumming by default. */
 378         cli->cl_checksum = 1;
 379         /*
 380          * The supported checksum types will be worked out at connect time
 381          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 382          * through procfs.
 383          */
 384         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 385 #endif
 386         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 387
 388         /* This value may be reduced at connect time in
 389          * ptlrpc_connect_interpret() . We initialize it to only
 390          * 1MB until we know what the performance looks like.
 391          * In the future this should likely be increased. LU-1431 */
 392         cli->cl_max_pages_per_rpc = min_t(int, PTLRPC_MAX_BRW_PAGES,
 393                                           LNET_MTU >> PAGE_CACHE_SHIFT);
 394
 395         /* set cl_chunkbits default value to PAGE_CACHE_SHIFT,
 396          * it will be updated at OSC connection time. */
 397         cli->cl_chunkbits = PAGE_CACHE_SHIFT;
 398
 399         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 400                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 401         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 128 /* MB */) {
 402                 cli->cl_max_rpcs_in_flight = 2;
 403         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 256 /* MB */) {
 404                 cli->cl_max_rpcs_in_flight = 3;
 405         } else if (totalram_pages >> (20 - PAGE_CACHE_SHIFT) <= 512 /* MB */) {
 406                 cli->cl_max_rpcs_in_flight = 4;
 407         } else {
 408                 if (osc_on_mdt(obddev->obd_name))
 409                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
 410                 else
 411                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 412         }
 413         rc = ldlm_get_ref();
 414         if (rc) {
 415                 CERROR("ldlm_get_ref failed: %d\n", rc);
 416                 GOTO(err, rc);
 417         }
 418
 419         ptlrpc_init_client(rq_portal, rp_portal, name,
 420                            &obddev->obd_ldlm_client);
 421
 422         imp = class_new_import(obddev);
 423         if (imp == NULL)
 424                 GOTO(err_ldlm, rc = -ENOENT);
 425         imp->imp_client = &obddev->obd_ldlm_client;
 426         imp->imp_connect_op = connect_op;
 427         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 428                LUSTRE_CFG_BUFLEN(lcfg, 1));
 429         class_import_put(imp);
 430
 431         rc = client_import_add_conn(imp, &server_uuid, 1);
 432         if (rc) {
 433                 CERROR("can't add initial connection\n");
 434                 GOTO(err_import, rc);
 435         }
 436
 437         cli->cl_import = imp;
 438         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 439         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 440         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 441
 442         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 443                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 444                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 445                                name, obddev->obd_name,
 446                                cli->cl_target_uuid.uuid);
 447                         spin_lock(&imp->imp_lock);
 448                         imp->imp_deactive = 1;
 449                         spin_unlock(&imp->imp_lock);
 450                 }
 451         }
 452
 453         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 454                                                    LDLM_NAMESPACE_CLIENT,
 455                                                    LDLM_NAMESPACE_GREEDY,
 456                                                    ns_type);
 457         if (obddev->obd_namespace == NULL) {
 458                 CERROR("Unable to create client namespace - %s\n",
 459                        obddev->obd_name);
 460                 GOTO(err_import, rc = -ENOMEM);
 461         }
 462
 463         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 464
 465         RETURN(rc);
 466
 467 err_import:
 468         class_destroy_import(imp);
 469 err_ldlm:
 470         ldlm_put_ref();
 471 err:
 472         RETURN(rc);
 473
 474 }
 475 EXPORT_SYMBOL(client_obd_setup);
 476
 477 int client_obd_cleanup(struct obd_device *obddev)
 478 {
 479         ENTRY;
 480
 481         ldlm_namespace_free_post(obddev->obd_namespace);
 482         obddev->obd_namespace = NULL;
 483
 484         LASSERT(obddev->u.cli.cl_import == NULL);
 485
 486         ldlm_put_ref();
 487         RETURN(0);
 488 }
 489 EXPORT_SYMBOL(client_obd_cleanup);
 490
 491 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 492 int client_connect_import(const struct lu_env *env,
 493                           struct obd_export **exp,
 494                           struct obd_device *obd, struct obd_uuid *cluuid,
 495                           struct obd_connect_data *data, void *localdata)
 496 {
 497         struct client_obd       *cli    = &obd->u.cli;
 498         struct obd_import       *imp    = cli->cl_import;
 499         struct obd_connect_data *ocd;
 500         struct lustre_handle    conn    = { 0 };
 501         int                     rc;
 502         ENTRY;
 503
 504         *exp = NULL;
 505         down_write(&cli->cl_sem);
 506         if (cli->cl_conn_count > 0 )
 507                 GOTO(out_sem, rc = -EALREADY);
 508
 509         rc = class_connect(&conn, obd, cluuid);
 510         if (rc)
 511                 GOTO(out_sem, rc);
 512
 513         cli->cl_conn_count++;
 514         *exp = class_conn2export(&conn);
 515
 516         LASSERT(obd->obd_namespace);
 517
 518         imp->imp_dlm_handle = conn;
 519         rc = ptlrpc_init_import(imp);
 520         if (rc != 0)
 521                 GOTO(out_ldlm, rc);
 522
 523         ocd = &imp->imp_connect_data;
 524         if (data) {
 525                 *ocd = *data;
 526                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 527         }
 528
 529         rc = ptlrpc_connect_import(imp);
 530         if (rc != 0) {
 531                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 532                 GOTO(out_ldlm, rc);
 533         }
 534         LASSERT((*exp)->exp_connection);
 535
 536         if (data) {
 537                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 538                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 539                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 540                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 541         }
 542
 543         ptlrpc_pinger_add_import(imp);
 544
 545         EXIT;
 546
 547         if (rc) {
 548 out_ldlm:
 549                 cli->cl_conn_count--;
 550                 class_disconnect(*exp);
 551                 *exp = NULL;
 552         }
 553 out_sem:
 554         up_write(&cli->cl_sem);
 555
 556         return rc;
 557 }
 558 EXPORT_SYMBOL(client_connect_import);
 559
 560 int client_disconnect_export(struct obd_export *exp)
 561 {
 562         struct obd_device *obd = class_exp2obd(exp);
 563         struct client_obd *cli;
 564         struct obd_import *imp;
 565         int rc = 0, err;
 566         ENTRY;
 567
 568         if (!obd) {
 569                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 570                        exp, exp ? exp->exp_handle.h_cookie : -1);
 571                 RETURN(-EINVAL);
 572         }
 573
 574         cli = &obd->u.cli;
 575         imp = cli->cl_import;
 576
 577         down_write(&cli->cl_sem);
 578         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 579                cli->cl_conn_count);
 580
 581         if (!cli->cl_conn_count) {
 582                 CERROR("disconnecting disconnected device (%s)\n",
 583                        obd->obd_name);
 584                 GOTO(out_disconnect, rc = -EINVAL);
 585         }
 586
 587         cli->cl_conn_count--;
 588         if (cli->cl_conn_count)
 589                 GOTO(out_disconnect, rc = 0);
 590
 591         /* Mark import deactivated now, so we don't try to reconnect if any
 592          * of the cleanup RPCs fails (e.g. LDLM cancel, etc).  We don't
 593          * fully deactivate the import, or that would drop all requests. */
 594         spin_lock(&imp->imp_lock);
 595         imp->imp_deactive = 1;
 596         spin_unlock(&imp->imp_lock);
 597
 598         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 599          * delete it regardless.  (It's safe to delete an import that was
 600          * never added.) */
 601         (void)ptlrpc_pinger_del_import(imp);
 602
 603         if (obd->obd_namespace != NULL) {
 604                 /* obd_force == local only */
 605                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 606                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 607                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 608         }
 609
 610         /* There's no need to hold sem while disconnecting an import,
 611          * and it may actually cause deadlock in GSS. */
 612         up_write(&cli->cl_sem);
 613         rc = ptlrpc_disconnect_import(imp, 0);
 614         down_write(&cli->cl_sem);
 615
 616         ptlrpc_invalidate_import(imp);
 617
 618         EXIT;
 619
 620 out_disconnect:
 621         /* Use server style - class_disconnect should be always called for
 622          * o_disconnect. */
 623         err = class_disconnect(exp);
 624         if (!rc && err)
 625                 rc = err;
 626
 627         up_write(&cli->cl_sem);
 628
 629         RETURN(rc);
 630 }
 631 EXPORT_SYMBOL(client_disconnect_export);
 632
 633 #ifdef HAVE_SERVER_SUPPORT
 634 int server_disconnect_export(struct obd_export *exp)
 635 {
 636         int rc;
 637         ENTRY;
 638
 639         /* Disconnect early so that clients can't keep using export. */
 640         rc = class_disconnect(exp);
 641         /* Close import to avoid sending any requests. */
 642         if (exp->exp_imp_reverse)
 643                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 644
 645         if (exp->exp_obd->obd_namespace != NULL)
 646                 ldlm_cancel_locks_for_export(exp);
 647
 648         /* complete all outstanding replies */
 649         spin_lock(&exp->exp_lock);
 650         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 651                 struct ptlrpc_reply_state *rs =
 652                         cfs_list_entry(exp->exp_outstanding_replies.next,
 653                                        struct ptlrpc_reply_state, rs_exp_list);
 654                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 655
 656                 spin_lock(&svcpt->scp_rep_lock);
 657
 658                 cfs_list_del_init(&rs->rs_exp_list);
 659                 spin_lock(&rs->rs_lock);
 660                 ptlrpc_schedule_difficult_reply(rs);
 661                 spin_unlock(&rs->rs_lock);
 662
 663                 spin_unlock(&svcpt->scp_rep_lock);
 664         }
 665         spin_unlock(&exp->exp_lock);
 666
 667         RETURN(rc);
 668 }
 669 EXPORT_SYMBOL(server_disconnect_export);
 670
 671 /* --------------------------------------------------------------------------
 672  * from old lib/target.c
 673  * -------------------------------------------------------------------------- */
 674
 675 static int target_handle_reconnect(struct lustre_handle *conn,
 676                                    struct obd_export *exp,
 677                                    struct obd_uuid *cluuid)
 678 {
 679         ENTRY;
 680
 681         if (exp->exp_connection && exp->exp_imp_reverse) {
 682                 struct lustre_handle *hdl;
 683                 struct obd_device *target;
 684
 685                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 686                 target = exp->exp_obd;
 687
 688                 /* Might be a re-connect after a partition. */
 689                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 690                         if (target->obd_recovering) {
 691                                 int timeout = cfs_duration_sec(cfs_time_sub(
 692                                         cfs_timer_deadline(
 693                                         &target->obd_recovery_timer),
 694                                         cfs_time_current()));
 695
 696                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 697                                         "ing, waiting for %d clients in recov"
 698                                         "ery for %d:%.02d\n", target->obd_name,
 699                                         obd_uuid2str(&exp->exp_client_uuid),
 700                                         obd_export_nid2str(exp),
 701                                         target->obd_max_recoverable_clients,
 702                                         timeout / 60, timeout % 60);
 703                         } else {
 704                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 705                                         "reconnecting\n", target->obd_name,
 706                                         obd_uuid2str(&exp->exp_client_uuid),
 707                                         obd_export_nid2str(exp));
 708                         }
 709
 710                         conn->cookie = exp->exp_handle.h_cookie;
 711                         /* target_handle_connect() treats EALREADY and
 712                          * -EALREADY differently.  EALREADY means we are
 713                          * doing a valid reconnect from the same client. */
 714                         RETURN(EALREADY);
 715                 } else {
 716                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 717                                       "with handle "LPX64". Rejecting client "
 718                                       "with the same UUID trying to reconnect "
 719                                       "with handle "LPX64"\n", target->obd_name,
 720                                       obd_uuid2str(&exp->exp_client_uuid),
 721                                       obd_export_nid2str(exp),
 722                                       hdl->cookie, conn->cookie);
 723                         memset(conn, 0, sizeof *conn);
 724                         /* target_handle_connect() treats EALREADY and
 725                          * -EALREADY differently.  -EALREADY is an error
 726                          * (same UUID, different handle). */
 727                         RETURN(-EALREADY);
 728                 }
 729         }
 730
 731         conn->cookie = exp->exp_handle.h_cookie;
 732         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 733                cluuid->uuid, exp, conn->cookie);
 734         RETURN(0);
 735 }
 736
 737 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 738                           int error)
 739 {
 740         struct obd_export *exp = cb_data;
 741
 742         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 743                obd->obd_name, exp->exp_client_uuid.uuid);
 744
 745         spin_lock(&exp->exp_lock);
 746         exp->exp_need_sync = 0;
 747         spin_unlock(&exp->exp_lock);
 748         class_export_cb_put(exp);
 749 }
 750 EXPORT_SYMBOL(target_client_add_cb);
 751
 752 #ifdef __KERNEL__
 753 static void
 754 check_and_start_recovery_timer(struct obd_device *obd,
 755                                struct ptlrpc_request *req, int new_client);
 756 #else
 757 static inline void
 758 check_and_start_recovery_timer(struct obd_device *obd,
 759                                struct ptlrpc_request *req, int new_client)
 760 {
 761 }
 762 #endif
 763
 764 int target_handle_connect(struct ptlrpc_request *req)
 765 {
 766         struct obd_device *target = NULL, *targref = NULL;
 767         struct obd_export *export = NULL;
 768         struct obd_import *revimp;
 769         struct obd_import *tmp_imp = NULL;
 770         struct lustre_handle conn;
 771         struct lustre_handle *tmp;
 772         struct obd_uuid tgtuuid;
 773         struct obd_uuid cluuid;
 774         struct obd_uuid remote_uuid;
 775         char *str;
 776         int rc = 0;
 777         char *target_start;
 778         int target_len;
 779         bool     mds_conn = false, lw_client = false;
 780         struct obd_connect_data *data, *tmpdata;
 781         int size, tmpsize;
 782         lnet_nid_t *client_nid = NULL;
 783         ENTRY;
 784
 785         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 786
 787         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 788         if (str == NULL) {
 789                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 790                 GOTO(out, rc = -EINVAL);
 791         }
 792
 793         obd_str2uuid(&tgtuuid, str);
 794         target = class_uuid2obd(&tgtuuid);
 795         if (!target)
 796                 target = class_name2obd(str);
 797
 798         if (!target) {
 799                 deuuidify(str, NULL, &target_start, &target_len);
 800                 LCONSOLE_ERROR_MSG(0x137, "%s: not available for connect "
 801                                    "from %s (no target). If you are running "
 802                                    "an HA pair check that the target is "
 803                                    "mounted on the other server.\n", str,
 804                                    libcfs_nid2str(req->rq_peer.nid));
 805                 GOTO(out, rc = -ENODEV);
 806         }
 807
 808         spin_lock(&target->obd_dev_lock);
 809         if (target->obd_stopping || !target->obd_set_up) {
 810                 spin_unlock(&target->obd_dev_lock);
 811
 812                 deuuidify(str, NULL, &target_start, &target_len);
 813                 LCONSOLE_INFO("%.*s: Not available for connect from %s (%s)\n",
 814                               target_len, target_start,
 815                               libcfs_nid2str(req->rq_peer.nid),
 816                               (target->obd_stopping ?
 817                                "stopping" : "not set up"));
 818                 GOTO(out, rc = -ENODEV);
 819         }
 820
 821         if (target->obd_no_conn) {
 822                 spin_unlock(&target->obd_dev_lock);
 823
 824                 CDEBUG(D_INFO, "%s: Temporarily refusing client connection "
 825                                "from %s\n", target->obd_name,
 826                                libcfs_nid2str(req->rq_peer.nid));
 827                 GOTO(out, rc = -EAGAIN);
 828         }
 829
 830         /* Make sure the target isn't cleaned up while we're here. Yes,
 831          * there's still a race between the above check and our incref here.
 832          * Really, class_uuid2obd should take the ref. */
 833         targref = class_incref(target, __FUNCTION__, current);
 834
 835         target->obd_conn_inprogress++;
 836         spin_unlock(&target->obd_dev_lock);
 837
 838         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 839         if (str == NULL) {
 840                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 841                 GOTO(out, rc = -EINVAL);
 842         }
 843
 844         obd_str2uuid(&cluuid, str);
 845
 846         /* XXX Extract a nettype and format accordingly. */
 847         switch (sizeof(lnet_nid_t)) {
 848         /* NB the casts only avoid compiler warnings. */
 849         case 8:
 850                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 851                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 852                 break;
 853         case 4:
 854                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 855                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 856                 break;
 857         default:
 858                 LBUG();
 859         }
 860
 861         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 862         if (tmp == NULL)
 863                 GOTO(out, rc = -EPROTO);
 864
 865         conn = *tmp;
 866
 867         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 868                                     RCL_CLIENT);
 869         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 870         if (!data)
 871                 GOTO(out, rc = -EPROTO);
 872
 873         rc = req_capsule_server_pack(&req->rq_pill);
 874         if (rc)
 875                 GOTO(out, rc);
 876
 877         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 878                 if (data->ocd_version < LUSTRE_VERSION_CODE -
 879                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 880                     data->ocd_version > LUSTRE_VERSION_CODE +
 881                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 882                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 883                                   "libclient connection attempt",
 884                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 885                                   "old" : "new",
 886                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 887                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 888                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 889                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 890                         data = req_capsule_server_sized_get(&req->rq_pill,
 891                                                             &RMF_CONNECT_DATA,
 892                                     offsetof(typeof(*data), ocd_version) +
 893                                              sizeof(data->ocd_version));
 894                         if (data) {
 895                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 896                                 data->ocd_version = LUSTRE_VERSION_CODE;
 897                         }
 898                         GOTO(out, rc = -EPROTO);
 899                 }
 900         }
 901
 902         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 903             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 904                 mds_conn = true;
 905
 906         if ((data->ocd_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0)
 907                 lw_client = true;
 908
 909         /* lctl gets a backstage, all-access pass. */
 910         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 911                 goto dont_check_exports;
 912
 913         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 914         if (!export)
 915                 goto no_export;
 916
 917         /* We've found an export in the hash. */
 918
 919         spin_lock(&export->exp_lock);
 920
 921         if (export->exp_connecting) { /* bug 9635, et. al. */
 922                 spin_unlock(&export->exp_lock);
 923                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 924                               export->exp_obd->obd_name, export,
 925                               libcfs_nid2str(req->rq_peer.nid));
 926                 class_export_put(export);
 927                 export = NULL;
 928                 rc = -EALREADY;
 929         } else if (mds_conn && export->exp_connection) {
 930                 spin_unlock(&export->exp_lock);
 931                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 932                         /* MDS reconnected after failover. */
 933                         LCONSOLE_WARN("%s: Received MDS connection from "
 934                             "%s, removing former export from %s\n",
 935                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 936                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 937                 else
 938                         /* New MDS connection from the same NID. */
 939                         LCONSOLE_WARN("%s: Received new MDS connection from "
 940                             "%s, removing former export from same NID\n",
 941                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 942                 class_fail_export(export);
 943                 class_export_put(export);
 944                 export = NULL;
 945                 rc = 0;
 946         } else if (export->exp_connection != NULL &&
 947                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 948                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 949                     MSG_CONNECT_INITIAL)) {
 950                 spin_unlock(&export->exp_lock);
 951                 /* In MDS failover we have static UUID but NID can change. */
 952                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 953                               "existing nid %s is already connected\n",
 954                               target->obd_name, cluuid.uuid,
 955                               libcfs_nid2str(req->rq_peer.nid),
 956                               libcfs_nid2str(
 957                                       export->exp_connection->c_peer.nid));
 958                 rc = -EALREADY;
 959                 class_export_put(export);
 960                 export = NULL;
 961         } else {
 962                 export->exp_connecting = 1;
 963                 spin_unlock(&export->exp_lock);
 964                 LASSERT(export->exp_obd == target);
 965
 966                 rc = target_handle_reconnect(&conn, export, &cluuid);
 967         }
 968
 969         /* If we found an export, we already unlocked. */
 970         if (!export) {
 971 no_export:
 972                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 973         } else if (req->rq_export == NULL &&
 974                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 975                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 976                               "still busy with %d references\n",
 977                               target->obd_name, cluuid.uuid,
 978                               libcfs_nid2str(req->rq_peer.nid),
 979                               cfs_atomic_read(&export->exp_refcount));
 980                 GOTO(out, rc = -EBUSY);
 981         } else if (req->rq_export != NULL &&
 982                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 983                 /* The current connect RPC has increased exp_rpc_count. */
 984                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
 985                               "still busy with %d active RPCs\n",
 986                               target->obd_name, cluuid.uuid,
 987                               libcfs_nid2str(req->rq_peer.nid),
 988                               cfs_atomic_read(&export->exp_rpc_count) - 1);
 989                 spin_lock(&export->exp_lock);
 990                 if (req->rq_export->exp_conn_cnt <
 991                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 992                         /* try to abort active requests */
 993                         req->rq_export->exp_abort_active_req = 1;
 994                 spin_unlock(&export->exp_lock);
 995                 GOTO(out, rc = -EBUSY);
 996         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 997                 if (!strstr(cluuid.uuid, "mdt"))
 998                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 999                                       "known client %s (at %s) because it "
1000                                       "is indicating it is a new client",
1001                                       target->obd_name, cluuid.uuid,
1002                                       libcfs_nid2str(req->rq_peer.nid));
1003                 GOTO(out, rc = -EALREADY);
1004         } else {
1005                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
1006         }
1007
1008         if (rc < 0) {
1009                 GOTO(out, rc);
1010         }
1011
1012         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
1013                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1014               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
1015               export, (long)cfs_time_current_sec(),
1016               export ? (long)export->exp_last_request_time : 0);
1017
1018         /* If this is the first time a client connects, reset the recovery
1019          * timer. Discard lightweight connections which might be local. */
1020         if (!lw_client && rc == 0 && target->obd_recovering)
1021                 check_and_start_recovery_timer(target, req, export == NULL);
1022
1023         /* We want to handle EALREADY but *not* -EALREADY from
1024          * target_handle_reconnect(), return reconnection state in a flag. */
1025         if (rc == EALREADY) {
1026                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
1027                 rc = 0;
1028         } else {
1029                 LASSERT(rc == 0);
1030         }
1031
1032         /* Tell the client if we support replayable requests. */
1033         if (target->obd_replayable)
1034                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
1035         client_nid = &req->rq_peer.nid;
1036
1037         if (export == NULL) {
1038                 /* allow lightweight connections during recovery */
1039                 if (target->obd_recovering && !lw_client) {
1040                         cfs_time_t t;
1041                         int     c; /* connected */
1042                         int     i; /* in progress */
1043                         int     k; /* known */
1044                         int     s; /* stale/evicted */
1045
1046                         c = cfs_atomic_read(&target->obd_connected_clients);
1047                         i = cfs_atomic_read(&target->obd_lock_replay_clients);
1048                         k = target->obd_max_recoverable_clients;
1049                         s = target->obd_stale_clients;
1050                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1051                         t = cfs_time_sub(t, cfs_time_current());
1052                         t = cfs_duration_sec(t);
1053                         LCONSOLE_WARN("%s: Denying connection for new client "
1054                                       "%s (at %s), waiting for all %d known "
1055                                       "clients (%d recovered, %d in progress, "
1056                                       "and %d evicted) to recover in %d:%.02d\n",
1057                                       target->obd_name, cluuid.uuid,
1058                                       libcfs_nid2str(req->rq_peer.nid), k,
1059                                       c - i, i, s, (int)t / 60,
1060                                       (int)t % 60);
1061                         rc = -EBUSY;
1062                 } else {
1063 dont_check_exports:
1064                         rc = obd_connect(req->rq_svc_thread->t_env,
1065                                          &export, target, &cluuid, data,
1066                                          client_nid);
1067                         if (mds_conn && OBD_FAIL_CHECK(OBD_FAIL_TGT_RCVG_FLAG))
1068                                 lustre_msg_add_op_flags(req->rq_repmsg,
1069                                                 MSG_CONNECT_RECOVERING);
1070                         if (rc == 0)
1071                                 conn.cookie = export->exp_handle.h_cookie;
1072                 }
1073         } else {
1074                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1075                                    export, target, &cluuid, data, client_nid);
1076         }
1077         if (rc)
1078                 GOTO(out, rc);
1079
1080 #if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(2, 6, 50, 0)
1081         /* 2.2.0 clients always swab nidtbl entries due to a bug, so server
1082          * will do the swabbing for if the client is using the same endianness.
1083          *
1084          * This fixup is version-limited, because we don't want to carry the
1085          * OBD_CONNECT_MNE_SWAB flag around forever, just so long as we need
1086          * interop with unpatched 2.2 clients.  For newer clients, servers
1087          * will never do MNE swabbing, let the client handle that.  LU-1644 */
1088         spin_lock(&export->exp_lock);
1089         export->exp_need_mne_swab = !ptlrpc_req_need_swab(req) &&
1090                         !(data->ocd_connect_flags & OBD_CONNECT_MNE_SWAB);
1091         spin_unlock(&export->exp_lock);
1092 #else
1093 #warning "LU-1644: Remove old OBD_CONNECT_MNE_SWAB fixup and exp_need_mne_swab"
1094 #endif
1095
1096         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1097         data->ocd_instance = target->u.obt.obt_instance;
1098
1099         /* Return only the parts of obd_connect_data that we understand, so the
1100          * client knows that we don't understand the rest. */
1101         if (data) {
1102                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1103                                                RCL_SERVER);
1104                 tmpdata = req_capsule_server_get(&req->rq_pill,
1105                                                  &RMF_CONNECT_DATA);
1106                 /* Don't use struct assignment here, because the client reply
1107                  * buffer may be smaller/larger than the local struct
1108                  * obd_connect_data. */
1109                 memcpy(tmpdata, data, min(tmpsize, size));
1110         }
1111
1112         /* If all else goes well, this is our RPC return code. */
1113         req->rq_status = 0;
1114
1115         lustre_msg_set_handle(req->rq_repmsg, &conn);
1116
1117         /* If the client and the server are the same node, we will already
1118          * have an export that really points to the client's DLM export,
1119          * because we have a shared handles table.
1120          *
1121          * XXX this will go away when shaver stops sending the "connect" handle
1122          * in the real "remote handle" field of the request --phik 24 Apr 2003
1123          */
1124         ptlrpc_request_change_export(req, export);
1125
1126         spin_lock(&export->exp_lock);
1127         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1128                 spin_unlock(&export->exp_lock);
1129                 CDEBUG(D_RPCTRACE, "%s: %s already connected at greater "
1130                        "or equal conn_cnt: %d >= %d\n",
1131                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1132                        export->exp_conn_cnt,
1133                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1134
1135                 GOTO(out, rc = -EALREADY);
1136         }
1137         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1138         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1139         export->exp_abort_active_req = 0;
1140
1141         /* Don't evict liblustre clients for not pinging. */
1142         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1143                 export->exp_libclient = 1;
1144                 spin_unlock(&export->exp_lock);
1145
1146                 spin_lock(&target->obd_dev_lock);
1147                 cfs_list_del_init(&export->exp_obd_chain_timed);
1148                 spin_unlock(&target->obd_dev_lock);
1149         } else {
1150                 spin_unlock(&export->exp_lock);
1151         }
1152
1153         if (export->exp_connection != NULL) {
1154                 /* Check to see if connection came from another NID. */
1155                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1156                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1157                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1158                                      &export->exp_connection->c_peer.nid,
1159                                      &export->exp_nid_hash);
1160
1161                 ptlrpc_connection_put(export->exp_connection);
1162         }
1163
1164         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1165                                                        req->rq_self,
1166                                                        &remote_uuid);
1167         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1168                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1169                              &export->exp_connection->c_peer.nid,
1170                              &export->exp_nid_hash);
1171         }
1172
1173         if (target->obd_recovering && !export->exp_in_recovery && !lw_client) {
1174                 int has_transno;
1175                 __u64 transno = data->ocd_transno;
1176
1177                 spin_lock(&export->exp_lock);
1178                 /* possible race with class_disconnect_stale_exports,
1179                  * export may be already in the eviction process */
1180                 if (export->exp_failed) {
1181                         spin_unlock(&export->exp_lock);
1182                         GOTO(out, rc = -ENODEV);
1183                 }
1184                 export->exp_in_recovery = 1;
1185                 export->exp_req_replay_needed = 1;
1186                 export->exp_lock_replay_needed = 1;
1187                 spin_unlock(&export->exp_lock);
1188
1189                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1190                                  MSG_CONNECT_TRANSNO);
1191                 if (has_transno && transno == 0)
1192                         CWARN("Connect with zero transno!\n");
1193
1194                 if (has_transno && transno > 0 &&
1195                     transno < target->obd_next_recovery_transno &&
1196                     transno > target->obd_last_committed) {
1197                         /* Another way is to use cmpxchg() to be lock-free. */
1198                         spin_lock(&target->obd_recovery_task_lock);
1199                         if (transno < target->obd_next_recovery_transno)
1200                                 target->obd_next_recovery_transno = transno;
1201                         spin_unlock(&target->obd_recovery_task_lock);
1202                 }
1203
1204                 cfs_atomic_inc(&target->obd_req_replay_clients);
1205                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1206                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1207                     target->obd_max_recoverable_clients)
1208                         wake_up(&target->obd_next_transno_waitq);
1209         }
1210
1211         /* Tell the client we're in recovery, when client is involved in it. */
1212         if (target->obd_recovering && !lw_client)
1213                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1214
1215         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1216         conn = *tmp;
1217
1218         /* Return -ENOTCONN in case of errors to let client reconnect. */
1219         revimp = class_new_import(target);
1220         if (revimp == NULL) {
1221                 CERROR("fail to alloc new reverse import.\n");
1222                 GOTO(out, rc = -ENOTCONN);
1223         }
1224
1225         spin_lock(&export->exp_lock);
1226         if (export->exp_imp_reverse != NULL)
1227                 /* destroyed import can be still referenced in ctxt */
1228                 tmp_imp = export->exp_imp_reverse;
1229         export->exp_imp_reverse = revimp;
1230         spin_unlock(&export->exp_lock);
1231
1232         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1233         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1234         revimp->imp_remote_handle = conn;
1235         revimp->imp_dlm_fake = 1;
1236         revimp->imp_state = LUSTRE_IMP_FULL;
1237
1238         /* Unknown versions will be caught in
1239          * ptlrpc_handle_server_req_in->lustre_unpack_msg(). */
1240         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1241
1242         if ((data->ocd_connect_flags & OBD_CONNECT_AT) &&
1243             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1244                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1245         else
1246                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1247
1248         if ((data->ocd_connect_flags & OBD_CONNECT_FULL20) &&
1249             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1250                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1251         else
1252                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1253
1254         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1255         if (rc) {
1256                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1257                 spin_lock(&export->exp_lock);
1258                 export->exp_imp_reverse = NULL;
1259                 spin_unlock(&export->exp_lock);
1260                 class_destroy_import(revimp);
1261         }
1262
1263         class_import_put(revimp);
1264
1265 out:
1266         if (tmp_imp != NULL)
1267                 client_destroy_import(tmp_imp);
1268         if (export) {
1269                 spin_lock(&export->exp_lock);
1270                 export->exp_connecting = 0;
1271                 spin_unlock(&export->exp_lock);
1272
1273                 class_export_put(export);
1274         }
1275         if (targref) {
1276                 spin_lock(&target->obd_dev_lock);
1277                 target->obd_conn_inprogress--;
1278                 spin_unlock(&target->obd_dev_lock);
1279
1280                 class_decref(targref, __func__, current);
1281         }
1282         if (rc)
1283                 req->rq_status = rc;
1284         RETURN(rc);
1285 }
1286 EXPORT_SYMBOL(target_handle_connect);
1287
1288 int target_handle_disconnect(struct ptlrpc_request *req)
1289 {
1290         int rc;
1291         ENTRY;
1292
1293         rc = req_capsule_server_pack(&req->rq_pill);
1294         if (rc)
1295                 RETURN(rc);
1296
1297         /* Keep the rq_export around so we can send the reply. */
1298         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1299
1300         RETURN(0);
1301 }
1302 EXPORT_SYMBOL(target_handle_disconnect);
1303
1304 void target_destroy_export(struct obd_export *exp)
1305 {
1306         struct obd_import       *imp = NULL;
1307         /* exports created from last_rcvd data, and "fake"
1308            exports created by lctl don't have an import */
1309         spin_lock(&exp->exp_lock);
1310         if (exp->exp_imp_reverse != NULL) {
1311                 imp = exp->exp_imp_reverse;
1312                 exp->exp_imp_reverse = NULL;
1313         }
1314         spin_unlock(&exp->exp_lock);
1315         if (imp != NULL)
1316                 client_destroy_import(imp);
1317
1318         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1319         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1320         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1321         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1322 }
1323 EXPORT_SYMBOL(target_destroy_export);
1324
1325 /*
1326  * Recovery functions
1327  */
1328 static void target_request_copy_get(struct ptlrpc_request *req)
1329 {
1330         class_export_rpc_inc(req->rq_export);
1331         LASSERT(cfs_list_empty(&req->rq_list));
1332         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1333
1334         /* Increase refcount to keep request in queue. */
1335         cfs_atomic_inc(&req->rq_refcount);
1336         /* Let export know it has replays to be handled. */
1337         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1338 }
1339
1340 static void target_request_copy_put(struct ptlrpc_request *req)
1341 {
1342         LASSERT(cfs_list_empty(&req->rq_replay_list));
1343         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1344
1345         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1346         class_export_rpc_dec(req->rq_export);
1347         ptlrpc_server_drop_request(req);
1348 }
1349
1350 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1351 {
1352         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1353         struct obd_export     *exp = req->rq_export;
1354         struct ptlrpc_request *reqiter;
1355         int                    dup = 0;
1356
1357         LASSERT(exp);
1358
1359         spin_lock(&exp->exp_lock);
1360         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1361                                 rq_replay_list) {
1362                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1363                         dup = 1;
1364                         break;
1365                 }
1366         }
1367
1368         if (dup) {
1369                 /* We expect it with RESENT and REPLAY flags. */
1370                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1371                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1372                         CERROR("invalid flags %x of resent replay\n",
1373                                lustre_msg_get_flags(req->rq_reqmsg));
1374         } else {
1375                 cfs_list_add_tail(&req->rq_replay_list,
1376                                   &exp->exp_req_replay_queue);
1377         }
1378
1379         spin_unlock(&exp->exp_lock);
1380         return dup;
1381 }
1382
1383 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1384 {
1385         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1386         LASSERT(req->rq_export);
1387
1388         spin_lock(&req->rq_export->exp_lock);
1389         cfs_list_del_init(&req->rq_replay_list);
1390         spin_unlock(&req->rq_export->exp_lock);
1391 }
1392
1393 #ifdef __KERNEL__
1394 static void target_finish_recovery(struct obd_device *obd)
1395 {
1396         ENTRY;
1397
1398         /* Only log a recovery message when recovery has occurred. */
1399         if (obd->obd_recovery_start) {
1400                 time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1401                                         obd->obd_recovery_start);
1402                 LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1403                         "%d recovered and %d %s evicted.\n", obd->obd_name,
1404                         (int)elapsed_time / 60, (int)elapsed_time % 60,
1405                         obd->obd_max_recoverable_clients,
1406                         cfs_atomic_read(&obd->obd_connected_clients),
1407                         obd->obd_stale_clients,
1408                         obd->obd_stale_clients == 1 ? "was" : "were");
1409         }
1410
1411         ldlm_reprocess_all_ns(obd->obd_namespace);
1412         spin_lock(&obd->obd_recovery_task_lock);
1413         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1414             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1415             !cfs_list_empty(&obd->obd_final_req_queue)) {
1416                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1417                        obd->obd_name,
1418                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1419                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1420                                "" : "lock ",
1421                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1422                                "" : "final ");
1423                 spin_unlock(&obd->obd_recovery_task_lock);
1424                 LBUG();
1425         }
1426         spin_unlock(&obd->obd_recovery_task_lock);
1427
1428         obd->obd_recovery_end = cfs_time_current_sec();
1429
1430         /* When recovery finished, cleanup orphans on MDS and OST. */
1431         if (OBT(obd) && OBP(obd, postrecov)) {
1432                 int rc = OBP(obd, postrecov)(obd);
1433                 if (rc < 0)
1434                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1435                                       obd->obd_name, rc);
1436         }
1437         EXIT;
1438 }
1439
1440 static void abort_req_replay_queue(struct obd_device *obd)
1441 {
1442         struct ptlrpc_request *req, *n;
1443         cfs_list_t abort_list;
1444
1445         CFS_INIT_LIST_HEAD(&abort_list);
1446         spin_lock(&obd->obd_recovery_task_lock);
1447         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1448         spin_unlock(&obd->obd_recovery_task_lock);
1449         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1450                 DEBUG_REQ(D_WARNING, req, "aborted:");
1451                 req->rq_status = -ENOTCONN;
1452                 if (ptlrpc_error(req)) {
1453                         DEBUG_REQ(D_ERROR, req,
1454                                   "failed abort_req_reply; skipping");
1455                 }
1456                 target_exp_dequeue_req_replay(req);
1457                 target_request_copy_put(req);
1458         }
1459 }
1460
1461 static void abort_lock_replay_queue(struct obd_device *obd)
1462 {
1463         struct ptlrpc_request *req, *n;
1464         cfs_list_t abort_list;
1465
1466         CFS_INIT_LIST_HEAD(&abort_list);
1467         spin_lock(&obd->obd_recovery_task_lock);
1468         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1469         spin_unlock(&obd->obd_recovery_task_lock);
1470         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1471                 DEBUG_REQ(D_ERROR, req, "aborted:");
1472                 req->rq_status = -ENOTCONN;
1473                 if (ptlrpc_error(req)) {
1474                         DEBUG_REQ(D_ERROR, req,
1475                                   "failed abort_lock_reply; skipping");
1476                 }
1477                 target_request_copy_put(req);
1478         }
1479 }
1480
1481 /* Called from a cleanup function if the device is being cleaned up
1482    forcefully.  The exports should all have been disconnected already,
1483    the only thing left to do is
1484      - clear the recovery flags
1485      - cancel the timer
1486      - free queued requests and replies, but don't send replies
1487    Because the obd_stopping flag is set, no new requests should be received.
1488
1489 */
1490 void target_cleanup_recovery(struct obd_device *obd)
1491 {
1492         struct ptlrpc_request *req, *n;
1493         cfs_list_t clean_list;
1494         ENTRY;
1495
1496         CFS_INIT_LIST_HEAD(&clean_list);
1497         spin_lock(&obd->obd_dev_lock);
1498         if (!obd->obd_recovering) {
1499                 spin_unlock(&obd->obd_dev_lock);
1500                 EXIT;
1501                 return;
1502         }
1503         obd->obd_recovering = obd->obd_abort_recovery = 0;
1504         spin_unlock(&obd->obd_dev_lock);
1505
1506         spin_lock(&obd->obd_recovery_task_lock);
1507         target_cancel_recovery_timer(obd);
1508         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1509         spin_unlock(&obd->obd_recovery_task_lock);
1510
1511         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1512                 LASSERT(req->rq_reply_state == 0);
1513                 target_exp_dequeue_req_replay(req);
1514                 target_request_copy_put(req);
1515         }
1516
1517         spin_lock(&obd->obd_recovery_task_lock);
1518         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1519         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1520         spin_unlock(&obd->obd_recovery_task_lock);
1521
1522         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1523                 LASSERT(req->rq_reply_state == 0);
1524                 target_request_copy_put(req);
1525         }
1526
1527         EXIT;
1528 }
1529 EXPORT_SYMBOL(target_cleanup_recovery);
1530
1531 /* obd_recovery_task_lock should be held */
1532 void target_cancel_recovery_timer(struct obd_device *obd)
1533 {
1534         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1535         cfs_timer_disarm(&obd->obd_recovery_timer);
1536 }
1537 EXPORT_SYMBOL(target_cancel_recovery_timer);
1538
1539 static void target_start_recovery_timer(struct obd_device *obd)
1540 {
1541         if (obd->obd_recovery_start != 0)
1542                 return;
1543
1544         spin_lock(&obd->obd_dev_lock);
1545         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1546                 spin_unlock(&obd->obd_dev_lock);
1547                 return;
1548         }
1549
1550         LASSERT(obd->obd_recovery_timeout != 0);
1551
1552         if (obd->obd_recovery_start != 0) {
1553                 spin_unlock(&obd->obd_dev_lock);
1554                 return;
1555         }
1556
1557         cfs_timer_arm(&obd->obd_recovery_timer,
1558                       cfs_time_shift(obd->obd_recovery_timeout));
1559         obd->obd_recovery_start = cfs_time_current_sec();
1560         spin_unlock(&obd->obd_dev_lock);
1561
1562         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1563                       "or until %d client%s reconnect%s\n",
1564                       obd->obd_name,
1565                       obd->obd_recovery_timeout / 60,
1566                       obd->obd_recovery_timeout % 60,
1567                       obd->obd_max_recoverable_clients,
1568                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1569                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1570 }
1571
1572 /**
1573  * extend recovery window.
1574  *
1575  * if @extend is true, extend recovery window to have @drt remaining at least;
1576  * otherwise, make sure the recovery timeout value is not less than @drt.
1577  */
1578 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1579 {
1580         cfs_time_t now;
1581         cfs_time_t end;
1582         cfs_duration_t left;
1583         int to;
1584
1585         spin_lock(&obd->obd_dev_lock);
1586         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1587                 spin_unlock(&obd->obd_dev_lock);
1588                 return;
1589         }
1590         LASSERT(obd->obd_recovery_start != 0);
1591
1592         now  = cfs_time_current_sec();
1593         to   = obd->obd_recovery_timeout;
1594         end  = obd->obd_recovery_start + to;
1595         left = cfs_time_sub(end, now);
1596
1597         if (extend && (drt > left)) {
1598                 to += drt - left;
1599         } else if (!extend && (drt > to)) {
1600                 to = drt;
1601                 /* reduce drt by already passed time */
1602                 drt -= obd->obd_recovery_timeout - left;
1603         }
1604
1605         if (to > obd->obd_recovery_time_hard)
1606                 to = obd->obd_recovery_time_hard;
1607         if (obd->obd_recovery_timeout < to ||
1608             obd->obd_recovery_timeout == obd->obd_recovery_time_hard) {
1609                 obd->obd_recovery_timeout = to;
1610                 cfs_timer_arm(&obd->obd_recovery_timer,
1611                               cfs_time_shift(drt));
1612         }
1613         spin_unlock(&obd->obd_dev_lock);
1614
1615         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1616                obd->obd_name, (unsigned)drt);
1617 }
1618
1619 /* Reset the timer with each new client connection */
1620 /*
1621  * This timer is actually reconnect_timer, which is for making sure
1622  * the total recovery window is at least as big as my reconnect
1623  * attempt timing. So the initial recovery time_out will be set to
1624  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1625  * from client is bigger than this, then the recovery time_out will
1626  * be extended to make sure the client could be reconnected, in the
1627  * process, the timeout from the new client should be ignored.
1628  */
1629
1630 static void
1631 check_and_start_recovery_timer(struct obd_device *obd,
1632                                struct ptlrpc_request *req,
1633                                int new_client)
1634 {
1635         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1636         struct obd_device_target *obt = &obd->u.obt;
1637
1638         if (!new_client && service_time)
1639                 /* Teach server about old server's estimates, as first guess
1640                  * at how long new requests will take. */
1641                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1642                             service_time);
1643
1644         target_start_recovery_timer(obd);
1645
1646         /* Convert the service time to RPC timeout,
1647          * and reuse service_time to limit stack usage. */
1648         service_time = at_est2timeout(service_time);
1649
1650         /* We expect other clients to timeout within service_time, then try
1651          * to reconnect, then try the failover server.  The max delay between
1652          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL. */
1653         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1654
1655         LASSERT(obt->obt_magic == OBT_MAGIC);
1656         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC);
1657         if (service_time > obd->obd_recovery_timeout && !new_client)
1658                 extend_recovery_timer(obd, service_time, false);
1659 }
1660
1661 /** Health checking routines */
1662 static inline int exp_connect_healthy(struct obd_export *exp)
1663 {
1664         return (exp->exp_in_recovery);
1665 }
1666
1667 /** if export done req_replay or has replay in queue */
1668 static inline int exp_req_replay_healthy(struct obd_export *exp)
1669 {
1670         return (!exp->exp_req_replay_needed ||
1671                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1672 }
1673 /** if export done lock_replay or has replay in queue */
1674 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1675 {
1676         return (!exp->exp_lock_replay_needed ||
1677                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1678 }
1679
1680 static inline int exp_vbr_healthy(struct obd_export *exp)
1681 {
1682         return (!exp->exp_vbr_failed);
1683 }
1684
1685 static inline int exp_finished(struct obd_export *exp)
1686 {
1687         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1688 }
1689
1690 /** Checking routines for recovery */
1691 static int check_for_clients(struct obd_device *obd)
1692 {
1693         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1694
1695         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1696                 return 1;
1697         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1698         return (clnts + obd->obd_stale_clients ==
1699                 obd->obd_max_recoverable_clients);
1700 }
1701
1702 static int check_for_next_transno(struct obd_device *obd)
1703 {
1704         struct ptlrpc_request *req = NULL;
1705         int wake_up = 0, connected, completed, queue_len;
1706         __u64 next_transno, req_transno;
1707         ENTRY;
1708
1709         spin_lock(&obd->obd_recovery_task_lock);
1710         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1711                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1712                                      struct ptlrpc_request, rq_list);
1713                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1714         } else {
1715                 req_transno = 0;
1716         }
1717
1718         connected = cfs_atomic_read(&obd->obd_connected_clients);
1719         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1720         queue_len = obd->obd_requests_queued_for_recovery;
1721         next_transno = obd->obd_next_recovery_transno;
1722
1723         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1724                "req_transno: "LPU64", next_transno: "LPU64"\n",
1725                obd->obd_max_recoverable_clients, connected, completed,
1726                queue_len, req_transno, next_transno);
1727
1728         if (obd->obd_abort_recovery) {
1729                 CDEBUG(D_HA, "waking for aborted recovery\n");
1730                 wake_up = 1;
1731         } else if (obd->obd_recovery_expired) {
1732                 CDEBUG(D_HA, "waking for expired recovery\n");
1733                 wake_up = 1;
1734         } else if (req_transno == next_transno) {
1735                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1736                 wake_up = 1;
1737         } else if (queue_len > 0 &&
1738                    queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1739                 int d_lvl = D_HA;
1740                 /** handle gaps occured due to lost reply or VBR */
1741                 LASSERTF(req_transno >= next_transno,
1742                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1743                          req_transno, next_transno);
1744                 if (req_transno > obd->obd_last_committed &&
1745                     !obd->obd_version_recov)
1746                         d_lvl = D_ERROR;
1747                 CDEBUG(d_lvl,
1748                        "%s: waking for gap in transno, VBR is %s (skip: "
1749                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1750                        ", last_committed: "LPD64")\n",
1751                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1752                        next_transno, queue_len, completed, connected,
1753                        req_transno, obd->obd_last_committed);
1754                 obd->obd_next_recovery_transno = req_transno;
1755                 wake_up = 1;
1756         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1757                 CDEBUG(D_HA, "waking for completed recovery\n");
1758                 wake_up = 1;
1759         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1760                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1761                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1762                 obd->obd_next_recovery_transno = req_transno;
1763                 wake_up = 1;
1764         }
1765         spin_unlock(&obd->obd_recovery_task_lock);
1766         return wake_up;
1767 }
1768
1769 static int check_for_next_lock(struct obd_device *obd)
1770 {
1771         int wake_up = 0;
1772
1773         spin_lock(&obd->obd_recovery_task_lock);
1774         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1775                 CDEBUG(D_HA, "waking for next lock\n");
1776                 wake_up = 1;
1777         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1778                 CDEBUG(D_HA, "waking for completed lock replay\n");
1779                 wake_up = 1;
1780         } else if (obd->obd_abort_recovery) {
1781                 CDEBUG(D_HA, "waking for aborted recovery\n");
1782                 wake_up = 1;
1783         } else if (obd->obd_recovery_expired) {
1784                 CDEBUG(D_HA, "waking for expired recovery\n");
1785                 wake_up = 1;
1786         }
1787         spin_unlock(&obd->obd_recovery_task_lock);
1788
1789         return wake_up;
1790 }
1791
1792 /**
1793  * wait for recovery events,
1794  * check its status with help of check_routine
1795  * evict dead clients via health_check
1796  */
1797 static int target_recovery_overseer(struct obd_device *obd,
1798                                     int (*check_routine)(struct obd_device *),
1799                                     int (*health_check)(struct obd_export *))
1800 {
1801 repeat:
1802         wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1803         if (obd->obd_abort_recovery) {
1804                 CWARN("recovery is aborted, evict exports in recovery\n");
1805                 /** evict exports which didn't finish recovery yet */
1806                 class_disconnect_stale_exports(obd, exp_finished);
1807                 return 1;
1808         } else if (obd->obd_recovery_expired) {
1809                 obd->obd_recovery_expired = 0;
1810                 /** If some clients died being recovered, evict them */
1811                 LCONSOLE_WARN("%s: recovery is timed out, "
1812                               "evict stale exports\n", obd->obd_name);
1813                 /** evict cexports with no replay in queue, they are stalled */
1814                 class_disconnect_stale_exports(obd, health_check);
1815                 /** continue with VBR */
1816                 spin_lock(&obd->obd_dev_lock);
1817                 obd->obd_version_recov = 1;
1818                 spin_unlock(&obd->obd_dev_lock);
1819                 /**
1820                  * reset timer, recovery will proceed with versions now,
1821                  * timeout is set just to handle reconnection delays
1822                  */
1823                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1824                 /** Wait for recovery events again, after evicting bad clients */
1825                 goto repeat;
1826         }
1827         return 0;
1828 }
1829
1830 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1831 {
1832         struct ptlrpc_request *req = NULL;
1833         ENTRY;
1834
1835         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1836                 obd->obd_next_recovery_transno);
1837
1838         CFS_FAIL_TIMEOUT(OBD_FAIL_TGT_REPLAY_DELAY2, cfs_fail_val);
1839
1840         if (target_recovery_overseer(obd, check_for_next_transno,
1841                                      exp_req_replay_healthy)) {
1842                 abort_req_replay_queue(obd);
1843                 abort_lock_replay_queue(obd);
1844         }
1845
1846         spin_lock(&obd->obd_recovery_task_lock);
1847         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1848                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1849                                      struct ptlrpc_request, rq_list);
1850                 cfs_list_del_init(&req->rq_list);
1851                 obd->obd_requests_queued_for_recovery--;
1852                 spin_unlock(&obd->obd_recovery_task_lock);
1853         } else {
1854                 spin_unlock(&obd->obd_recovery_task_lock);
1855                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1856                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1857                 /** evict exports failed VBR */
1858                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1859         }
1860         RETURN(req);
1861 }
1862
1863 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1864 {
1865         struct ptlrpc_request *req = NULL;
1866
1867         CDEBUG(D_HA, "Waiting for lock\n");
1868         if (target_recovery_overseer(obd, check_for_next_lock,
1869                                      exp_lock_replay_healthy))
1870                 abort_lock_replay_queue(obd);
1871
1872         spin_lock(&obd->obd_recovery_task_lock);
1873         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1874                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1875                                      struct ptlrpc_request, rq_list);
1876                 cfs_list_del_init(&req->rq_list);
1877                 spin_unlock(&obd->obd_recovery_task_lock);
1878         } else {
1879                 spin_unlock(&obd->obd_recovery_task_lock);
1880                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1881                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1882                 /** evict exports failed VBR */
1883                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1884         }
1885         return req;
1886 }
1887
1888 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1889 {
1890         struct ptlrpc_request *req = NULL;
1891
1892         spin_lock(&obd->obd_recovery_task_lock);
1893         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1894                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1895                                      struct ptlrpc_request, rq_list);
1896                 cfs_list_del_init(&req->rq_list);
1897                 spin_unlock(&obd->obd_recovery_task_lock);
1898                 if (req->rq_export->exp_in_recovery) {
1899                         spin_lock(&req->rq_export->exp_lock);
1900                         req->rq_export->exp_in_recovery = 0;
1901                         spin_unlock(&req->rq_export->exp_lock);
1902                 }
1903         } else {
1904                 spin_unlock(&obd->obd_recovery_task_lock);
1905         }
1906         return req;
1907 }
1908
1909 static int handle_recovery_req(struct ptlrpc_thread *thread,
1910                                struct ptlrpc_request *req,
1911                                svc_handler_t handler)
1912 {
1913         int rc;
1914         ENTRY;
1915
1916         /**
1917          * export can be evicted during recovery, no need to handle replays for
1918          * it after that, discard such request silently
1919          */
1920         if (req->rq_export->exp_disconnected)
1921                 GOTO(reqcopy_put, rc = 0);
1922
1923         rc = lu_context_init(&req->rq_recov_session, LCT_SERVER_SESSION);
1924         if (rc) {
1925                 CERROR("Failure to initialize session: %d\n", rc);
1926                 GOTO(reqcopy_put, rc);
1927         }
1928
1929         req->rq_recov_session.lc_thread = thread;
1930         lu_context_enter(&req->rq_recov_session);
1931         req->rq_svc_thread = thread;
1932         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1933
1934         /* thread context */
1935         lu_context_enter(&thread->t_env->le_ctx);
1936         (void)handler(req);
1937         lu_context_exit(&thread->t_env->le_ctx);
1938
1939         lu_context_exit(&req->rq_recov_session);
1940         lu_context_fini(&req->rq_recov_session);
1941         /* don't reset timer for final stage */
1942         if (!exp_finished(req->rq_export)) {
1943                 int to = obd_timeout;
1944
1945                 /**
1946                  * Add request timeout to the recovery time so next request from
1947                  * this client may come in recovery time
1948                  */
1949                 if (!AT_OFF) {
1950                         struct ptlrpc_service_part *svcpt;
1951
1952                         svcpt = req->rq_rqbd->rqbd_svcpt;
1953                         /* If the server sent early reply for this request,
1954                          * the client will recalculate the timeout according to
1955                          * current server estimate service time, so we will
1956                          * use the maxium timeout here for waiting the client
1957                          * sending the next req */
1958                         to = max((int)at_est2timeout(
1959                                  at_get(&svcpt->scp_at_estimate)),
1960                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1961                         /* Add 2 net_latency, one for balance rq_deadline
1962                          * (see ptl_send_rpc), one for resend the req to server,
1963                          * Note: client will pack net_latency in replay req
1964                          * (see ptlrpc_replay_req) */
1965                         to += 2 * lustre_msg_get_service_time(req->rq_reqmsg);
1966                 }
1967                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1968         }
1969 reqcopy_put:
1970         RETURN(rc);
1971 }
1972
1973 static int target_recovery_thread(void *arg)
1974 {
1975         struct lu_target *lut = arg;
1976         struct obd_device *obd = lut->lut_obd;
1977         struct ptlrpc_request *req;
1978         struct target_recovery_data *trd = &obd->obd_recovery_data;
1979         unsigned long delta;
1980         struct lu_env *env;
1981         struct ptlrpc_thread *thread = NULL;
1982         int rc = 0;
1983         ENTRY;
1984
1985         unshare_fs_struct();
1986         OBD_ALLOC_PTR(thread);
1987         if (thread == NULL)
1988                 RETURN(-ENOMEM);
1989
1990         OBD_ALLOC_PTR(env);
1991         if (env == NULL) {
1992                 OBD_FREE_PTR(thread);
1993                 RETURN(-ENOMEM);
1994         }
1995
1996         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1997         if (rc) {
1998                 OBD_FREE_PTR(thread);
1999                 OBD_FREE_PTR(env);
2000                 RETURN(rc);
2001         }
2002
2003         thread->t_env = env;
2004         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
2005         env->le_ctx.lc_thread = thread;
2006         thread->t_data = NULL;
2007         thread->t_watchdog = NULL;
2008
2009         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
2010                current_pid());
2011         trd->trd_processing_task = current_pid();
2012
2013         spin_lock(&obd->obd_dev_lock);
2014         obd->obd_recovering = 1;
2015         spin_unlock(&obd->obd_dev_lock);
2016         complete(&trd->trd_starting);
2017
2018         /* first of all, we have to know the first transno to replay */
2019         if (target_recovery_overseer(obd, check_for_clients,
2020                                      exp_connect_healthy)) {
2021                 abort_req_replay_queue(obd);
2022                 abort_lock_replay_queue(obd);
2023         }
2024
2025         /* next stage: replay requests */
2026         delta = jiffies;
2027         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
2028                cfs_atomic_read(&obd->obd_req_replay_clients),
2029                obd->obd_next_recovery_transno);
2030         while ((req = target_next_replay_req(obd))) {
2031                 LASSERT(trd->trd_processing_task == current_pid());
2032                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
2033                           lustre_msg_get_transno(req->rq_reqmsg),
2034                           libcfs_nid2str(req->rq_peer.nid));
2035                 handle_recovery_req(thread, req,
2036                                     trd->trd_recovery_handler);
2037                 /**
2038                  * bz18031: increase next_recovery_transno before
2039                  * target_request_copy_put() will drop exp_rpc reference
2040                  */
2041                 spin_lock(&obd->obd_recovery_task_lock);
2042                 obd->obd_next_recovery_transno++;
2043                 spin_unlock(&obd->obd_recovery_task_lock);
2044                 target_exp_dequeue_req_replay(req);
2045                 target_request_copy_put(req);
2046                 obd->obd_replayed_requests++;
2047         }
2048
2049         /**
2050          * The second stage: replay locks
2051          */
2052         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
2053                cfs_atomic_read(&obd->obd_lock_replay_clients));
2054         while ((req = target_next_replay_lock(obd))) {
2055                 LASSERT(trd->trd_processing_task == current_pid());
2056                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
2057                           libcfs_nid2str(req->rq_peer.nid));
2058                 handle_recovery_req(thread, req,
2059                                     trd->trd_recovery_handler);
2060                 target_request_copy_put(req);
2061                 obd->obd_replayed_locks++;
2062         }
2063
2064         /**
2065          * The third stage: reply on final pings, at this moment all clients
2066          * must have request in final queue
2067          */
2068         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
2069         /** Update server last boot epoch */
2070         tgt_boot_epoch_update(lut);
2071         /* We drop recoverying flag to forward all new requests
2072          * to regular mds_handle() since now */
2073         spin_lock(&obd->obd_dev_lock);
2074         obd->obd_recovering = obd->obd_abort_recovery = 0;
2075         spin_unlock(&obd->obd_dev_lock);
2076         spin_lock(&obd->obd_recovery_task_lock);
2077         target_cancel_recovery_timer(obd);
2078         spin_unlock(&obd->obd_recovery_task_lock);
2079         while ((req = target_next_final_ping(obd))) {
2080                 LASSERT(trd->trd_processing_task == current_pid());
2081                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2082                           libcfs_nid2str(req->rq_peer.nid));
2083                 handle_recovery_req(thread, req,
2084                                     trd->trd_recovery_handler);
2085                 target_request_copy_put(req);
2086         }
2087
2088         delta = (jiffies - delta) / HZ;
2089         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2090               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2091         if (delta > OBD_RECOVERY_TIME_SOFT) {
2092                 CWARN("too long recovery - read logs\n");
2093                 libcfs_debug_dumplog();
2094         }
2095
2096         target_finish_recovery(obd);
2097
2098         lu_context_fini(&env->le_ctx);
2099         trd->trd_processing_task = 0;
2100         complete(&trd->trd_finishing);
2101
2102         OBD_FREE_PTR(thread);
2103         OBD_FREE_PTR(env);
2104         RETURN(rc);
2105 }
2106
2107 static int target_start_recovery_thread(struct lu_target *lut,
2108                                         svc_handler_t handler)
2109 {
2110         struct obd_device *obd = lut->lut_obd;
2111         int rc = 0;
2112         struct target_recovery_data *trd = &obd->obd_recovery_data;
2113
2114         memset(trd, 0, sizeof(*trd));
2115         init_completion(&trd->trd_starting);
2116         init_completion(&trd->trd_finishing);
2117         trd->trd_recovery_handler = handler;
2118
2119         if (!IS_ERR(kthread_run(target_recovery_thread,
2120                                 lut, "tgt_recov"))) {
2121                 wait_for_completion(&trd->trd_starting);
2122                 LASSERT(obd->obd_recovering != 0);
2123         } else {
2124                 rc = -ECHILD;
2125         }
2126
2127         return rc;
2128 }
2129
2130 void target_stop_recovery_thread(struct obd_device *obd)
2131 {
2132         if (obd->obd_recovery_data.trd_processing_task > 0) {
2133                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2134                 /** recovery can be done but postrecovery is not yet */
2135                 spin_lock(&obd->obd_dev_lock);
2136                 if (obd->obd_recovering) {
2137                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2138                         obd->obd_abort_recovery = 1;
2139                         wake_up(&obd->obd_next_transno_waitq);
2140                 }
2141                 spin_unlock(&obd->obd_dev_lock);
2142                 wait_for_completion(&trd->trd_finishing);
2143         }
2144 }
2145 EXPORT_SYMBOL(target_stop_recovery_thread);
2146
2147 void target_recovery_fini(struct obd_device *obd)
2148 {
2149         class_disconnect_exports(obd);
2150         target_stop_recovery_thread(obd);
2151         target_cleanup_recovery(obd);
2152 }
2153 EXPORT_SYMBOL(target_recovery_fini);
2154
2155 static void target_recovery_expired(unsigned long castmeharder)
2156 {
2157         struct obd_device *obd = (struct obd_device *)castmeharder;
2158         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2159                " after %lds (%d clients connected)\n",
2160                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2161                cfs_time_current_sec()- obd->obd_recovery_start,
2162                cfs_atomic_read(&obd->obd_connected_clients));
2163
2164         obd->obd_recovery_expired = 1;
2165         wake_up(&obd->obd_next_transno_waitq);
2166 }
2167
2168 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2169 {
2170         struct obd_device *obd = lut->lut_obd;
2171         if (obd->obd_max_recoverable_clients == 0) {
2172                 /** Update server last boot epoch */
2173                 tgt_boot_epoch_update(lut);
2174                 return;
2175         }
2176
2177         CDEBUG(D_HA, "RECOVERY: service %s, %d recoverable clients, "
2178                "last_transno "LPU64"\n", obd->obd_name,
2179                obd->obd_max_recoverable_clients, obd->obd_last_committed);
2180         LASSERT(obd->obd_stopping == 0);
2181         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2182         obd->obd_recovery_start = 0;
2183         obd->obd_recovery_end = 0;
2184
2185         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2186         target_start_recovery_thread(lut, handler);
2187 }
2188 EXPORT_SYMBOL(target_recovery_init);
2189
2190 #endif /* __KERNEL__ */
2191
2192 static int target_process_req_flags(struct obd_device *obd,
2193                                     struct ptlrpc_request *req)
2194 {
2195         struct obd_export *exp = req->rq_export;
2196         LASSERT(exp != NULL);
2197         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2198                 /* client declares he's ready to replay locks */
2199                 spin_lock(&exp->exp_lock);
2200                 if (exp->exp_req_replay_needed) {
2201                         exp->exp_req_replay_needed = 0;
2202                         spin_unlock(&exp->exp_lock);
2203
2204                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2205                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2206                 } else {
2207                         spin_unlock(&exp->exp_lock);
2208                 }
2209         }
2210         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2211                 /* client declares he's ready to complete recovery
2212                  * so, we put the request on th final queue */
2213                 spin_lock(&exp->exp_lock);
2214                 if (exp->exp_lock_replay_needed) {
2215                         exp->exp_lock_replay_needed = 0;
2216                         spin_unlock(&exp->exp_lock);
2217
2218                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2219                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2220                 } else {
2221                         spin_unlock(&exp->exp_lock);
2222                 }
2223         }
2224         return 0;
2225 }
2226
2227 int target_queue_recovery_request(struct ptlrpc_request *req,
2228                                   struct obd_device *obd)
2229 {
2230         cfs_list_t *tmp;
2231         int inserted = 0;
2232         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2233         ENTRY;
2234
2235         if (obd->obd_recovery_data.trd_processing_task == current_pid()) {
2236                 /* Processing the queue right now, don't re-add. */
2237                 RETURN(1);
2238         }
2239
2240         target_process_req_flags(obd, req);
2241
2242         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2243                 /* client declares he's ready to complete recovery
2244                  * so, we put the request on th final queue */
2245                 target_request_copy_get(req);
2246                 DEBUG_REQ(D_HA, req, "queue final req");
2247                 wake_up(&obd->obd_next_transno_waitq);
2248                 spin_lock(&obd->obd_recovery_task_lock);
2249                 if (obd->obd_recovering) {
2250                         cfs_list_add_tail(&req->rq_list,
2251                                           &obd->obd_final_req_queue);
2252                 } else {
2253                         spin_unlock(&obd->obd_recovery_task_lock);
2254                         target_request_copy_put(req);
2255                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2256                 }
2257                 spin_unlock(&obd->obd_recovery_task_lock);
2258                 RETURN(0);
2259         }
2260         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2261                 /* client declares he's ready to replay locks */
2262                 target_request_copy_get(req);
2263                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2264                 wake_up(&obd->obd_next_transno_waitq);
2265                 spin_lock(&obd->obd_recovery_task_lock);
2266                 LASSERT(obd->obd_recovering);
2267                 /* usually due to recovery abort */
2268                 if (!req->rq_export->exp_in_recovery) {
2269                         spin_unlock(&obd->obd_recovery_task_lock);
2270                         target_request_copy_put(req);
2271                         RETURN(-ENOTCONN);
2272                 }
2273                 LASSERT(req->rq_export->exp_lock_replay_needed);
2274                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2275                 spin_unlock(&obd->obd_recovery_task_lock);
2276                 RETURN(0);
2277         }
2278
2279         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2280          * (i.e. buflens etc are in my own byte order), but type-dependent
2281          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2282
2283         if (!transno) {
2284                 CFS_INIT_LIST_HEAD(&req->rq_list);
2285                 DEBUG_REQ(D_HA, req, "not queueing");
2286                 RETURN(1);
2287         }
2288
2289         /* If we're processing the queue, we want don't want to queue this
2290          * message.
2291          *
2292          * Also, if this request has a transno less than the one we're waiting
2293          * for, we should process it now.  It could (and currently always will)
2294          * be an open request for a descriptor that was opened some time ago.
2295          *
2296          * Also, a resent, replayed request that has already been
2297          * handled will pass through here and be processed immediately.
2298          */
2299         CDEBUG(D_HA, "Next recovery transno: "LPU64
2300                ", current: "LPU64", replaying\n",
2301                obd->obd_next_recovery_transno, transno);
2302         spin_lock(&obd->obd_recovery_task_lock);
2303         if (transno < obd->obd_next_recovery_transno) {
2304                 /* Processing the queue right now, don't re-add. */
2305                 LASSERT(cfs_list_empty(&req->rq_list));
2306                 spin_unlock(&obd->obd_recovery_task_lock);
2307                 RETURN(1);
2308         }
2309         spin_unlock(&obd->obd_recovery_task_lock);
2310
2311         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2312                 RETURN(0);
2313
2314         target_request_copy_get(req);
2315         if (!req->rq_export->exp_in_recovery) {
2316                 target_request_copy_put(req);
2317                 RETURN(-ENOTCONN);
2318         }
2319         LASSERT(req->rq_export->exp_req_replay_needed);
2320
2321         if (target_exp_enqueue_req_replay(req)) {
2322                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2323                 target_request_copy_put(req);
2324                 RETURN(0);
2325         }
2326
2327         /* XXX O(n^2) */
2328         spin_lock(&obd->obd_recovery_task_lock);
2329         LASSERT(obd->obd_recovering);
2330         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2331                 struct ptlrpc_request *reqiter =
2332                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2333
2334                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2335                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2336                         inserted = 1;
2337                         break;
2338                 }
2339
2340                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2341                              transno)) {
2342                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2343                                   "has been claimed by another client");
2344                         spin_unlock(&obd->obd_recovery_task_lock);
2345                         target_exp_dequeue_req_replay(req);
2346                         target_request_copy_put(req);
2347                         RETURN(0);
2348                 }
2349         }
2350
2351         if (!inserted)
2352                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2353
2354         obd->obd_requests_queued_for_recovery++;
2355         spin_unlock(&obd->obd_recovery_task_lock);
2356         wake_up(&obd->obd_next_transno_waitq);
2357         RETURN(0);
2358 }
2359 EXPORT_SYMBOL(target_queue_recovery_request);
2360
2361 int target_handle_ping(struct ptlrpc_request *req)
2362 {
2363         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2364         return req_capsule_server_pack(&req->rq_pill);
2365 }
2366 EXPORT_SYMBOL(target_handle_ping);
2367
2368 void target_committed_to_req(struct ptlrpc_request *req)
2369 {
2370         struct obd_export *exp = req->rq_export;
2371
2372         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2373                 lustre_msg_set_last_committed(req->rq_repmsg,
2374                                               exp->exp_last_committed);
2375         else
2376                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2377                           "%d)", exp->exp_obd->obd_no_transno,
2378                           req->rq_repmsg == NULL);
2379
2380         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2381                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2382 }
2383 EXPORT_SYMBOL(target_committed_to_req);
2384
2385 #endif /* HAVE_SERVER_SUPPORT */
2386
2387 /**
2388  * Packs current SLV and Limit into \a req.
2389  */
2390 int target_pack_pool_reply(struct ptlrpc_request *req)
2391 {
2392         struct obd_device *obd;
2393         ENTRY;
2394
2395         /* Check that we still have all structures alive as this may
2396          * be some late RPC at shutdown time. */
2397         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2398                      !exp_connect_lru_resize(req->rq_export))) {
2399                 lustre_msg_set_slv(req->rq_repmsg, 0);
2400                 lustre_msg_set_limit(req->rq_repmsg, 0);
2401                 RETURN(0);
2402         }
2403
2404         /* OBD is alive here as export is alive, which we checked above. */
2405         obd = req->rq_export->exp_obd;
2406
2407         read_lock(&obd->obd_pool_lock);
2408         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2409         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2410         read_unlock(&obd->obd_pool_lock);
2411
2412         RETURN(0);
2413 }
2414 EXPORT_SYMBOL(target_pack_pool_reply);
2415
2416 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2417 {
2418         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2419                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2420                 return (-ECOMM);
2421         }
2422
2423         if (unlikely(rc)) {
2424                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2425                 req->rq_status = rc;
2426                 return (ptlrpc_send_error(req, 1));
2427         } else {
2428                 DEBUG_REQ(D_NET, req, "sending reply");
2429         }
2430
2431         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2432 }
2433
2434 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2435 {
2436         struct ptlrpc_service_part *svcpt;
2437         int                        netrc;
2438         struct ptlrpc_reply_state *rs;
2439         struct obd_export         *exp;
2440         ENTRY;
2441
2442         if (req->rq_no_reply) {
2443                 EXIT;
2444                 return;
2445         }
2446
2447         svcpt = req->rq_rqbd->rqbd_svcpt;
2448         rs = req->rq_reply_state;
2449         if (rs == NULL || !rs->rs_difficult) {
2450                 /* no notifiers */
2451                 target_send_reply_msg (req, rc, fail_id);
2452                 EXIT;
2453                 return;
2454         }
2455
2456         /* must be an export if locks saved */
2457         LASSERT (req->rq_export != NULL);
2458         /* req/reply consistent */
2459         LASSERT(rs->rs_svcpt == svcpt);
2460
2461         /* "fresh" reply */
2462         LASSERT (!rs->rs_scheduled);
2463         LASSERT (!rs->rs_scheduled_ever);
2464         LASSERT (!rs->rs_handled);
2465         LASSERT (!rs->rs_on_net);
2466         LASSERT (rs->rs_export == NULL);
2467         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2468         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2469
2470         exp = class_export_get (req->rq_export);
2471
2472         /* disable reply scheduling while I'm setting up */
2473         rs->rs_scheduled = 1;
2474         rs->rs_on_net    = 1;
2475         rs->rs_xid       = req->rq_xid;
2476         rs->rs_transno   = req->rq_transno;
2477         rs->rs_export    = exp;
2478         rs->rs_opc       = lustre_msg_get_opc(req->rq_reqmsg);
2479
2480         spin_lock(&exp->exp_uncommitted_replies_lock);
2481         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2482                rs->rs_transno, exp->exp_last_committed);
2483         if (rs->rs_transno > exp->exp_last_committed) {
2484                 /* not committed already */
2485                 cfs_list_add_tail(&rs->rs_obd_list,
2486                                   &exp->exp_uncommitted_replies);
2487         }
2488         spin_unlock(&exp->exp_uncommitted_replies_lock);
2489
2490         spin_lock(&exp->exp_lock);
2491         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2492         spin_unlock(&exp->exp_lock);
2493
2494         netrc = target_send_reply_msg(req, rc, fail_id);
2495
2496         spin_lock(&svcpt->scp_rep_lock);
2497
2498         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2499
2500         if (netrc != 0) {
2501                 /* error sending: reply is off the net.  Also we need +1
2502                  * reply ref until ptlrpc_handle_rs() is done
2503                  * with the reply state (if the send was successful, there
2504                  * would have been +1 ref for the net, which
2505                  * reply_out_callback leaves alone) */
2506                 rs->rs_on_net = 0;
2507                 ptlrpc_rs_addref(rs);
2508         }
2509
2510         spin_lock(&rs->rs_lock);
2511         if (rs->rs_transno <= exp->exp_last_committed ||
2512             (!rs->rs_on_net && !rs->rs_no_ack) ||
2513             cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2514             cfs_list_empty(&rs->rs_obd_list)) {
2515                 CDEBUG(D_HA, "Schedule reply immediately\n");
2516                 ptlrpc_dispatch_difficult_reply(rs);
2517         } else {
2518                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2519                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2520         }
2521         spin_unlock(&rs->rs_lock);
2522         spin_unlock(&svcpt->scp_rep_lock);
2523         EXIT;
2524 }
2525 EXPORT_SYMBOL(target_send_reply);
2526
2527 ldlm_mode_t lck_compat_array[] = {
2528         [LCK_EX] LCK_COMPAT_EX,
2529         [LCK_PW] LCK_COMPAT_PW,
2530         [LCK_PR] LCK_COMPAT_PR,
2531         [LCK_CW] LCK_COMPAT_CW,
2532         [LCK_CR] LCK_COMPAT_CR,
2533         [LCK_NL] LCK_COMPAT_NL,
2534         [LCK_GROUP] LCK_COMPAT_GROUP,
2535         [LCK_COS] LCK_COMPAT_COS,
2536 };
2537
2538 /**
2539  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2540  * not escape to the user level.
2541  */
2542 int ldlm_error2errno(ldlm_error_t error)
2543 {
2544         int result;
2545
2546         switch (error) {
2547         case ELDLM_OK:
2548                 result = 0;
2549                 break;
2550         case ELDLM_LOCK_CHANGED:
2551                 result = -ESTALE;
2552                 break;
2553         case ELDLM_LOCK_ABORTED:
2554                 result = -ENAVAIL;
2555                 break;
2556         case ELDLM_LOCK_REPLACED:
2557                 result = -ESRCH;
2558                 break;
2559         case ELDLM_NO_LOCK_DATA:
2560                 result = -ENOENT;
2561                 break;
2562         case ELDLM_NAMESPACE_EXISTS:
2563                 result = -EEXIST;
2564                 break;
2565         case ELDLM_BAD_NAMESPACE:
2566                 result = -EBADF;
2567                 break;
2568         default:
2569                 if (((int)error) < 0)  /* cast to signed type */
2570                         result = error; /* as ldlm_error_t can be unsigned */
2571                 else {
2572                         CERROR("Invalid DLM result code: %d\n", error);
2573                         result = -EPROTO;
2574                 }
2575         }
2576         return result;
2577 }
2578 EXPORT_SYMBOL(ldlm_error2errno);
2579
2580 /**
2581  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2582  */
2583 ldlm_error_t ldlm_errno2error(int err_no)
2584 {
2585         int error;
2586
2587         switch (err_no) {
2588         case 0:
2589                 error = ELDLM_OK;
2590                 break;
2591         case -ESTALE:
2592                 error = ELDLM_LOCK_CHANGED;
2593                 break;
2594         case -ENAVAIL:
2595                 error = ELDLM_LOCK_ABORTED;
2596                 break;
2597         case -ESRCH:
2598                 error = ELDLM_LOCK_REPLACED;
2599                 break;
2600         case -ENOENT:
2601                 error = ELDLM_NO_LOCK_DATA;
2602                 break;
2603         case -EEXIST:
2604                 error = ELDLM_NAMESPACE_EXISTS;
2605                 break;
2606         case -EBADF:
2607                 error = ELDLM_BAD_NAMESPACE;
2608                 break;
2609         default:
2610                 error = err_no;
2611         }
2612         return error;
2613 }
2614 EXPORT_SYMBOL(ldlm_errno2error);
2615
2616 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2617 void ldlm_dump_export_locks(struct obd_export *exp)
2618 {
2619         spin_lock(&exp->exp_locks_list_guard);
2620         if (!cfs_list_empty(&exp->exp_locks_list)) {
2621                 struct ldlm_lock *lock;
2622
2623                 CERROR("dumping locks for export %p,"
2624                        "ignore if the unmount doesn't hang\n", exp);
2625                 cfs_list_for_each_entry(lock, &exp->exp_locks_list,
2626                                         l_exp_refs_link)
2627                         LDLM_ERROR(lock, "lock:");
2628         }
2629         spin_unlock(&exp->exp_locks_list_guard);
2630 }
2631 #endif
2632
2633 #ifdef HAVE_SERVER_SUPPORT
2634 static int target_bulk_timeout(void *data)
2635 {
2636         ENTRY;
2637         /* We don't fail the connection here, because having the export
2638          * killed makes the (vital) call to commitrw very sad.
2639          */
2640         RETURN(1);
2641 }
2642
2643 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2644 {
2645         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2646 }
2647
2648 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2649                    struct l_wait_info *lwi)
2650 {
2651         struct ptlrpc_request *req = desc->bd_req;
2652         time_t start = cfs_time_current_sec();
2653         time_t deadline;
2654         int rc = 0;
2655
2656         ENTRY;
2657
2658         /* If there is eviction in progress, wait for it to finish. */
2659         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2660                 *lwi = LWI_INTR(NULL, NULL);
2661                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2662                                   !cfs_atomic_read(&exp->exp_obd->
2663                                                    obd_evict_inprogress),
2664                                   lwi);
2665         }
2666
2667         /* Check if client was evicted or tried to reconnect already. */
2668         if (exp->exp_failed || exp->exp_abort_active_req) {
2669                 rc = -ENOTCONN;
2670         } else {
2671                 if (desc->bd_type == BULK_PUT_SINK)
2672                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2673                 if (rc == 0)
2674                         rc = ptlrpc_start_bulk_transfer(desc);
2675         }
2676
2677         if (rc < 0) {
2678                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2679                           bulk2type(desc), rc);
2680                 RETURN(rc);
2681         }
2682
2683         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2684                 ptlrpc_abort_bulk(desc);
2685                 RETURN(0);
2686         }
2687
2688         /* limit actual bulk transfer to bulk_timeout seconds */
2689         deadline = start + bulk_timeout;
2690         if (deadline > req->rq_deadline)
2691                 deadline = req->rq_deadline;
2692
2693         do {
2694                 long timeoutl = deadline - cfs_time_current_sec();
2695                 cfs_duration_t timeout = timeoutl <= 0 ?
2696                                          CFS_TICK : cfs_time_seconds(timeoutl);
2697
2698                 *lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
2699                                             target_bulk_timeout, desc);
2700                 rc = l_wait_event(desc->bd_waitq,
2701                                   !ptlrpc_server_bulk_active(desc) ||
2702                                   exp->exp_failed ||
2703                                   exp->exp_abort_active_req, lwi);
2704                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
2705                 /* Wait again if we changed rq_deadline. */
2706                 deadline = start + bulk_timeout;
2707                 if (deadline > req->rq_deadline)
2708                         deadline = req->rq_deadline;
2709         } while ((rc == -ETIMEDOUT) &&
2710                  (deadline > cfs_time_current_sec()));
2711
2712         if (rc == -ETIMEDOUT) {
2713                 DEBUG_REQ(D_ERROR, req, "timeout on bulk %s after %ld%+lds",
2714                           bulk2type(desc), deadline - start,
2715                           cfs_time_current_sec() - deadline);
2716                 ptlrpc_abort_bulk(desc);
2717         } else if (exp->exp_failed) {
2718                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2719                           bulk2type(desc));
2720                 rc = -ENOTCONN;
2721                 ptlrpc_abort_bulk(desc);
2722         } else if (exp->exp_abort_active_req) {
2723                 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2724                           bulk2type(desc));
2725                 /* We don't reply anyway. */
2726                 rc = -ETIMEDOUT;
2727                 ptlrpc_abort_bulk(desc);
2728         } else if (desc->bd_failure ||
2729                    desc->bd_nob_transferred != desc->bd_nob) {
2730                 DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2731                           desc->bd_failure ? "network error on" : "truncated",
2732                           bulk2type(desc), desc->bd_nob_transferred,
2733                           desc->bd_nob);
2734                 /* XXX Should this be a different errno? */
2735                 rc = -ETIMEDOUT;
2736         } else if (desc->bd_type == BULK_GET_SINK) {
2737                 rc = sptlrpc_svc_unwrap_bulk(req, desc);
2738         }
2739
2740         RETURN(rc);
2741 }
2742 EXPORT_SYMBOL(target_bulk_io);
2743
2744 #endif /* HAVE_SERVER_SUPPORT */