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) 2011, 2012, Whamcloud, Inc.
  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 #define DEBUG_SUBSYSTEM S_LDLM
  38
  39 #ifdef __KERNEL__
  40 # include <libcfs/libcfs.h>
  41 #else
  42 # include <liblustre.h>
  43 #endif
  44 #include <obd.h>
  45 #include <lustre_mds.h>
  46 #include <lustre_dlm.h>
  47 #include <lustre_net.h>
  48 #include <lustre_sec.h>
  49 #include "ldlm_internal.h"
  50
  51 /* @priority: if non-zero, move the selected to the list head
  52  * @create: if zero, only search in existed connections
  53  */
  54 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  55                            int priority, int create)
  56 {
  57         struct ptlrpc_connection *ptlrpc_conn;
  58         struct obd_import_conn *imp_conn = NULL, *item;
  59         int rc = 0;
  60         ENTRY;
  61
  62         if (!create && !priority) {
  63                 CDEBUG(D_HA, "Nothing to do\n");
  64                 RETURN(-EINVAL);
  65         }
  66
  67         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  68         if (!ptlrpc_conn) {
  69                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  70                 RETURN (-ENOENT);
  71         }
  72
  73         if (create) {
  74                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  75                 if (!imp_conn) {
  76                         GOTO(out_put, rc = -ENOMEM);
  77                 }
  78         }
  79
  80         cfs_spin_lock(&imp->imp_lock);
  81         cfs_list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  82                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  83                         if (priority) {
  84                                 cfs_list_del(&item->oic_item);
  85                                 cfs_list_add(&item->oic_item,
  86                                              &imp->imp_conn_list);
  87                                 item->oic_last_attempt = 0;
  88                         }
  89                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  90                                imp, imp->imp_obd->obd_name, uuid->uuid,
  91                                (priority ? ", moved to head" : ""));
  92                         cfs_spin_unlock(&imp->imp_lock);
  93                         GOTO(out_free, rc = 0);
  94                 }
  95         }
  96         /* not found */
  97         if (create) {
  98                 imp_conn->oic_conn = ptlrpc_conn;
  99                 imp_conn->oic_uuid = *uuid;
 100                 imp_conn->oic_last_attempt = 0;
 101                 if (priority)
 102                         cfs_list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 103                 else
 104                         cfs_list_add_tail(&imp_conn->oic_item,
 105                                           &imp->imp_conn_list);
 106                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 107                        imp, imp->imp_obd->obd_name, uuid->uuid,
 108                        (priority ? "head" : "tail"));
 109         } else {
 110                 cfs_spin_unlock(&imp->imp_lock);
 111                 GOTO(out_free, rc = -ENOENT);
 112         }
 113
 114         cfs_spin_unlock(&imp->imp_lock);
 115         RETURN(0);
 116 out_free:
 117         if (imp_conn)
 118                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 119 out_put:
 120         ptlrpc_connection_put(ptlrpc_conn);
 121         RETURN(rc);
 122 }
 123
 124 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 125 {
 126         return import_set_conn(imp, uuid, 1, 0);
 127 }
 128
 129 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 130                            int priority)
 131 {
 132         return import_set_conn(imp, uuid, priority, 1);
 133 }
 134
 135 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 136 {
 137         struct obd_import_conn *imp_conn;
 138         struct obd_export *dlmexp;
 139         int rc = -ENOENT;
 140         ENTRY;
 141
 142         cfs_spin_lock(&imp->imp_lock);
 143         if (cfs_list_empty(&imp->imp_conn_list)) {
 144                 LASSERT(!imp->imp_connection);
 145                 GOTO(out, rc);
 146         }
 147
 148         cfs_list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 149                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 150                         continue;
 151                 LASSERT(imp_conn->oic_conn);
 152
 153                 /* is current conn? */
 154                 if (imp_conn == imp->imp_conn_current) {
 155                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 156
 157                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 158                             imp->imp_state != LUSTRE_IMP_DISCON) {
 159                                 CERROR("can't remove current connection\n");
 160                                 GOTO(out, rc = -EBUSY);
 161                         }
 162
 163                         ptlrpc_connection_put(imp->imp_connection);
 164                         imp->imp_connection = NULL;
 165
 166                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 167                         if (dlmexp && dlmexp->exp_connection) {
 168                                 LASSERT(dlmexp->exp_connection ==
 169                                         imp_conn->oic_conn);
 170                                 ptlrpc_connection_put(dlmexp->exp_connection);
 171                                 dlmexp->exp_connection = NULL;
 172                         }
 173                 }
 174
 175                 cfs_list_del(&imp_conn->oic_item);
 176                 ptlrpc_connection_put(imp_conn->oic_conn);
 177                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 178                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 179                        imp, imp->imp_obd->obd_name, uuid->uuid);
 180                 rc = 0;
 181                 break;
 182         }
 183 out:
 184         cfs_spin_unlock(&imp->imp_lock);
 185         if (rc == -ENOENT)
 186                 CERROR("connection %s not found\n", uuid->uuid);
 187         RETURN(rc);
 188 }
 189
 190 /**
 191  * Find conn uuid by peer nid. @peer is a server nid. This function is used
 192  * to find a conn uuid of @imp which can reach @peer.
 193  */
 194 int client_import_find_conn(struct obd_import *imp, lnet_nid_t peer,
 195                             struct obd_uuid *uuid)
 196 {
 197         struct obd_import_conn *conn;
 198         int rc = -ENOENT;
 199         ENTRY;
 200
 201         cfs_spin_lock(&imp->imp_lock);
 202         cfs_list_for_each_entry(conn, &imp->imp_conn_list, oic_item) {
 203                 /* check if conn uuid does have this peer nid */
 204                 if (class_check_uuid(&conn->oic_uuid, peer)) {
 205                         *uuid = conn->oic_uuid;
 206                         rc = 0;
 207                         break;
 208                 }
 209         }
 210         cfs_spin_unlock(&imp->imp_lock);
 211         RETURN(rc);
 212 }
 213 EXPORT_SYMBOL(client_import_find_conn);
 214
 215 void client_destroy_import(struct obd_import *imp)
 216 {
 217         /* drop security policy instance after all rpc finished/aborted
 218          * to let all busy contexts be released. */
 219         class_import_get(imp);
 220         class_destroy_import(imp);
 221         sptlrpc_import_sec_put(imp);
 222         class_import_put(imp);
 223 }
 224 EXPORT_SYMBOL(client_destroy_import);
 225
 226 /**
 227  * check whether the osc is on MDT or not
 228  * In the config log,
 229  * osc on MDT
 230  *      setup 0:{fsname}-OSTxxxx-osc[-MDTxxxx] 1:lustre-OST0000_UUID 2:NID
 231  * osc on client
 232  *      setup 0:{fsname}-OSTxxxx-osc 1:lustre-OST0000_UUID 2:NID
 233  *
 234  **/
 235 static int osc_on_mdt(char *obdname)
 236 {
 237         char *ptr;
 238
 239         ptr = strrchr(obdname, '-');
 240         if (ptr == NULL)
 241                 return 0;
 242
 243         if (strncmp(ptr + 1, "MDT", 3) == 0)
 244                 return 1;
 245
 246         return 0;
 247 }
 248
 249 /* configure an RPC client OBD device
 250  *
 251  * lcfg parameters:
 252  * 1 - client UUID
 253  * 2 - server UUID
 254  * 3 - inactive-on-startup
 255  */
 256 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 257 {
 258         struct client_obd *cli = &obddev->u.cli;
 259         struct obd_import *imp;
 260         struct obd_uuid server_uuid;
 261         int rq_portal, rp_portal, connect_op;
 262         char *name = obddev->obd_type->typ_name;
 263         ldlm_ns_type_t ns_type = LDLM_NS_TYPE_UNKNOWN;
 264         int rc;
 265         ENTRY;
 266
 267         /* In a more perfect world, we would hang a ptlrpc_client off of
 268          * obd_type and just use the values from there. */
 269         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 270                 rq_portal = OST_REQUEST_PORTAL;
 271                 rp_portal = OSC_REPLY_PORTAL;
 272                 connect_op = OST_CONNECT;
 273                 cli->cl_sp_me = LUSTRE_SP_CLI;
 274                 cli->cl_sp_to = LUSTRE_SP_OST;
 275                 ns_type = LDLM_NS_TYPE_OSC;
 276
 277         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 278                 rq_portal = MDS_REQUEST_PORTAL;
 279                 rp_portal = MDC_REPLY_PORTAL;
 280                 connect_op = MDS_CONNECT;
 281                 cli->cl_sp_me = LUSTRE_SP_CLI;
 282                 cli->cl_sp_to = LUSTRE_SP_MDT;
 283                 ns_type = LDLM_NS_TYPE_MDC;
 284
 285         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 286                 rq_portal = MGS_REQUEST_PORTAL;
 287                 rp_portal = MGC_REPLY_PORTAL;
 288                 connect_op = MGS_CONNECT;
 289                 cli->cl_sp_me = LUSTRE_SP_MGC;
 290                 cli->cl_sp_to = LUSTRE_SP_MGS;
 291                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 292                 ns_type = LDLM_NS_TYPE_MGC;
 293
 294         } else {
 295                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 296                        name);
 297                 RETURN(-EINVAL);
 298         }
 299
 300         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 301                 CERROR("requires a TARGET UUID\n");
 302                 RETURN(-EINVAL);
 303         }
 304
 305         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 306                 CERROR("client UUID must be less than 38 characters\n");
 307                 RETURN(-EINVAL);
 308         }
 309
 310         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 311                 CERROR("setup requires a SERVER UUID\n");
 312                 RETURN(-EINVAL);
 313         }
 314
 315         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 316                 CERROR("target UUID must be less than 38 characters\n");
 317                 RETURN(-EINVAL);
 318         }
 319
 320         cfs_init_rwsem(&cli->cl_sem);
 321         cfs_sema_init(&cli->cl_mgc_sem, 1);
 322         cli->cl_conn_count = 0;
 323         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 324                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 325                      sizeof(server_uuid)));
 326
 327         cli->cl_dirty = 0;
 328         cli->cl_avail_grant = 0;
 329         /* FIXME: should limit this for the sum of all cl_dirty_max */
 330         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 331         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > cfs_num_physpages / 8)
 332                 cli->cl_dirty_max = cfs_num_physpages << (CFS_PAGE_SHIFT - 3);
 333         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 334         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 335         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 336         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 337         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 338         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 339         cfs_atomic_set(&cli->cl_pending_w_pages, 0);
 340         cfs_atomic_set(&cli->cl_pending_r_pages, 0);
 341         cli->cl_r_in_flight = 0;
 342         cli->cl_w_in_flight = 0;
 343
 344         cfs_spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 345         cfs_spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 346         cfs_spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 347         cfs_spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 348         cfs_spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 349         cfs_spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 350         cfs_waitq_init(&cli->cl_destroy_waitq);
 351         cfs_atomic_set(&cli->cl_destroy_in_flight, 0);
 352 #ifdef ENABLE_CHECKSUM
 353         /* Turn on checksumming by default. */
 354         cli->cl_checksum = 1;
 355         /*
 356          * The supported checksum types will be worked out at connect time
 357          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 358          * through procfs.
 359          */
 360         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 361 #endif
 362         cfs_atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 363
 364         /* This value may be changed at connect time in
 365            ptlrpc_connect_interpret. */
 366         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 367                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 368
 369         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 370                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 371         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 372                 cli->cl_max_rpcs_in_flight = 2;
 373         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 374                 cli->cl_max_rpcs_in_flight = 3;
 375         } else if (cfs_num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 376                 cli->cl_max_rpcs_in_flight = 4;
 377         } else {
 378                 if (osc_on_mdt(obddev->obd_name))
 379                         cli->cl_max_rpcs_in_flight = MDS_OSC_MAX_RIF_DEFAULT;
 380                 else
 381                         cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 382         }
 383         rc = ldlm_get_ref();
 384         if (rc) {
 385                 CERROR("ldlm_get_ref failed: %d\n", rc);
 386                 GOTO(err, rc);
 387         }
 388
 389         ptlrpc_init_client(rq_portal, rp_portal, name,
 390                            &obddev->obd_ldlm_client);
 391
 392         imp = class_new_import(obddev);
 393         if (imp == NULL)
 394                 GOTO(err_ldlm, rc = -ENOENT);
 395         imp->imp_client = &obddev->obd_ldlm_client;
 396         imp->imp_connect_op = connect_op;
 397         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 398         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 399                LUSTRE_CFG_BUFLEN(lcfg, 1));
 400         class_import_put(imp);
 401
 402         rc = client_import_add_conn(imp, &server_uuid, 1);
 403         if (rc) {
 404                 CERROR("can't add initial connection\n");
 405                 GOTO(err_import, rc);
 406         }
 407
 408         cli->cl_import = imp;
 409         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 410         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 411         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 412
 413         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 414                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 415                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 416                                name, obddev->obd_name,
 417                                cli->cl_target_uuid.uuid);
 418                         cfs_spin_lock(&imp->imp_lock);
 419                         imp->imp_deactive = 1;
 420                         cfs_spin_unlock(&imp->imp_lock);
 421                 }
 422         }
 423
 424         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 425                                                    LDLM_NAMESPACE_CLIENT,
 426                                                    LDLM_NAMESPACE_GREEDY,
 427                                                    ns_type);
 428         if (obddev->obd_namespace == NULL) {
 429                 CERROR("Unable to create client namespace - %s\n",
 430                        obddev->obd_name);
 431                 GOTO(err_import, rc = -ENOMEM);
 432         }
 433
 434         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 435
 436         RETURN(rc);
 437
 438 err_import:
 439         class_destroy_import(imp);
 440 err_ldlm:
 441         ldlm_put_ref();
 442 err:
 443         RETURN(rc);
 444
 445 }
 446
 447 int client_obd_cleanup(struct obd_device *obddev)
 448 {
 449         ENTRY;
 450
 451         ldlm_namespace_free_post(obddev->obd_namespace);
 452         obddev->obd_namespace = NULL;
 453
 454         LASSERT(obddev->u.cli.cl_import == NULL);
 455
 456         ldlm_put_ref();
 457         RETURN(0);
 458 }
 459
 460 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 461 int client_connect_import(const struct lu_env *env,
 462                           struct obd_export **exp,
 463                           struct obd_device *obd, struct obd_uuid *cluuid,
 464                           struct obd_connect_data *data, void *localdata)
 465 {
 466         struct client_obd *cli = &obd->u.cli;
 467         struct obd_import *imp = cli->cl_import;
 468         struct obd_connect_data *ocd;
 469         struct lustre_handle conn = { 0 };
 470         int rc;
 471         ENTRY;
 472
 473         *exp = NULL;
 474         cfs_down_write(&cli->cl_sem);
 475         if (cli->cl_conn_count > 0 )
 476                 GOTO(out_sem, rc = -EALREADY);
 477
 478         rc = class_connect(&conn, obd, cluuid);
 479         if (rc)
 480                 GOTO(out_sem, rc);
 481
 482         cli->cl_conn_count++;
 483         *exp = class_conn2export(&conn);
 484
 485         LASSERT(obd->obd_namespace);
 486
 487         imp->imp_dlm_handle = conn;
 488         rc = ptlrpc_init_import(imp);
 489         if (rc != 0)
 490                 GOTO(out_ldlm, rc);
 491
 492         ocd = &imp->imp_connect_data;
 493         if (data) {
 494                 *ocd = *data;
 495                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 496         }
 497
 498         rc = ptlrpc_connect_import(imp);
 499         if (rc != 0) {
 500                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 501                 GOTO(out_ldlm, rc);
 502         }
 503         LASSERT((*exp)->exp_connection);
 504
 505         if (data) {
 506                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 507                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 508                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 509                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 510         }
 511
 512         ptlrpc_pinger_add_import(imp);
 513
 514         EXIT;
 515
 516         if (rc) {
 517 out_ldlm:
 518                 cli->cl_conn_count--;
 519                 class_disconnect(*exp);
 520                 *exp = NULL;
 521         }
 522 out_sem:
 523         cfs_up_write(&cli->cl_sem);
 524
 525         return rc;
 526 }
 527
 528 int client_disconnect_export(struct obd_export *exp)
 529 {
 530         struct obd_device *obd = class_exp2obd(exp);
 531         struct client_obd *cli;
 532         struct obd_import *imp;
 533         int rc = 0, err;
 534         ENTRY;
 535
 536         if (!obd) {
 537                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 538                        exp, exp ? exp->exp_handle.h_cookie : -1);
 539                 RETURN(-EINVAL);
 540         }
 541
 542         cli = &obd->u.cli;
 543         imp = cli->cl_import;
 544
 545         cfs_down_write(&cli->cl_sem);
 546         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 547                cli->cl_conn_count);
 548
 549         if (!cli->cl_conn_count) {
 550                 CERROR("disconnecting disconnected device (%s)\n",
 551                        obd->obd_name);
 552                 GOTO(out_disconnect, rc = -EINVAL);
 553         }
 554
 555         cli->cl_conn_count--;
 556         if (cli->cl_conn_count)
 557                 GOTO(out_disconnect, rc = 0);
 558
 559         /* Mark import deactivated now, so we don't try to reconnect if any
 560          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 561          * fully deactivate the import, or that would drop all requests. */
 562         cfs_spin_lock(&imp->imp_lock);
 563         imp->imp_deactive = 1;
 564         cfs_spin_unlock(&imp->imp_lock);
 565
 566         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 567          * delete it regardless.  (It's safe to delete an import that was
 568          * never added.) */
 569         (void)ptlrpc_pinger_del_import(imp);
 570
 571         if (obd->obd_namespace != NULL) {
 572                 /* obd_force == local only */
 573                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 574                                        obd->obd_force ? LCF_LOCAL : 0, NULL);
 575                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 576         }
 577
 578         /*
 579          * there's no need to hold sem during disconnecting an import,
 580          * and actually it may cause deadlock in gss.
 581          */
 582         cfs_up_write(&cli->cl_sem);
 583         rc = ptlrpc_disconnect_import(imp, 0);
 584         cfs_down_write(&cli->cl_sem);
 585
 586         ptlrpc_invalidate_import(imp);
 587
 588         EXIT;
 589
 590  out_disconnect:
 591         /* use server style - class_disconnect should be always called for
 592          * o_disconnect */
 593         err = class_disconnect(exp);
 594         if (!rc && err)
 595                 rc = err;
 596
 597         cfs_up_write(&cli->cl_sem);
 598
 599         RETURN(rc);
 600 }
 601
 602 #ifdef HAVE_SERVER_SUPPORT
 603 int server_disconnect_export(struct obd_export *exp)
 604 {
 605         int rc;
 606         ENTRY;
 607
 608         /* Disconnect early so that clients can't keep using export */
 609         rc = class_disconnect(exp);
 610         /* close import for avoid sending any requests */
 611         if (exp->exp_imp_reverse)
 612                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 613
 614         if (exp->exp_obd->obd_namespace != NULL)
 615                 ldlm_cancel_locks_for_export(exp);
 616
 617         /* complete all outstanding replies */
 618         cfs_spin_lock(&exp->exp_lock);
 619         while (!cfs_list_empty(&exp->exp_outstanding_replies)) {
 620                 struct ptlrpc_reply_state *rs =
 621                         cfs_list_entry(exp->exp_outstanding_replies.next,
 622                                        struct ptlrpc_reply_state, rs_exp_list);
 623                 struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
 624
 625                 cfs_spin_lock(&svcpt->scp_rep_lock);
 626
 627                 cfs_list_del_init(&rs->rs_exp_list);
 628                 cfs_spin_lock(&rs->rs_lock);
 629                 ptlrpc_schedule_difficult_reply(rs);
 630                 cfs_spin_unlock(&rs->rs_lock);
 631
 632                 cfs_spin_unlock(&svcpt->scp_rep_lock);
 633         }
 634         cfs_spin_unlock(&exp->exp_lock);
 635
 636         RETURN(rc);
 637 }
 638
 639 /* --------------------------------------------------------------------------
 640  * from old lib/target.c
 641  * -------------------------------------------------------------------------- */
 642
 643 static int target_handle_reconnect(struct lustre_handle *conn,
 644                                    struct obd_export *exp,
 645                                    struct obd_uuid *cluuid)
 646 {
 647         ENTRY;
 648
 649         if (exp->exp_connection && exp->exp_imp_reverse) {
 650                 struct lustre_handle *hdl;
 651                 struct obd_device *target;
 652
 653                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 654                 target = exp->exp_obd;
 655
 656                 /* Might be a re-connect after a partition. */
 657                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 658                         if (target->obd_recovering) {
 659                                 int timeout = cfs_duration_sec(cfs_time_sub(
 660                                         cfs_timer_deadline(
 661                                         &target->obd_recovery_timer),
 662                                         cfs_time_current()));
 663
 664                                 LCONSOLE_WARN("%s: Client %s (at %s) reconnect"
 665                                         "ing, waiting for %d clients in recov"
 666                                         "ery for %d:%.02d\n", target->obd_name,
 667                                         obd_uuid2str(&exp->exp_client_uuid),
 668                                         obd_export_nid2str(exp),
 669                                         target->obd_max_recoverable_clients,
 670                                         timeout / 60, timeout % 60);
 671                         } else {
 672                                 LCONSOLE_WARN("%s: Client %s (at %s) "
 673                                         "reconnecting\n", target->obd_name,
 674                                         obd_uuid2str(&exp->exp_client_uuid),
 675                                         obd_export_nid2str(exp));
 676                         }
 677
 678                         conn->cookie = exp->exp_handle.h_cookie;
 679                         /* target_handle_connect() treats EALREADY and
 680                          * -EALREADY differently.  EALREADY means we are
 681                          * doing a valid reconnect from the same client. */
 682                         RETURN(EALREADY);
 683                 } else {
 684                         LCONSOLE_WARN("%s: already connected client %s (at %s) "
 685                                       "with handle "LPX64". Rejecting client "
 686                                       "with the same UUID trying to reconnect "
 687                                       "with handle "LPX64"\n", target->obd_name,
 688                                       obd_uuid2str(&exp->exp_client_uuid),
 689                                       obd_export_nid2str(exp),
 690                                       hdl->cookie, conn->cookie);
 691                         memset(conn, 0, sizeof *conn);
 692                         /* target_handle_connect() treats EALREADY and
 693                          * -EALREADY differently.  -EALREADY is an error
 694                          * (same UUID, different handle). */
 695                         RETURN(-EALREADY);
 696                 }
 697         }
 698
 699         conn->cookie = exp->exp_handle.h_cookie;
 700         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 701                cluuid->uuid, exp, conn->cookie);
 702         RETURN(0);
 703 }
 704
 705 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 706                           int error)
 707 {
 708         struct obd_export *exp = cb_data;
 709
 710         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 711                obd->obd_name, exp->exp_client_uuid.uuid);
 712
 713         cfs_spin_lock(&exp->exp_lock);
 714         exp->exp_need_sync = 0;
 715         cfs_spin_unlock(&exp->exp_lock);
 716         class_export_cb_put(exp);
 717 }
 718 EXPORT_SYMBOL(target_client_add_cb);
 719
 720 #ifdef __KERNEL__
 721 static void
 722 check_and_start_recovery_timer(struct obd_device *obd,
 723                                struct ptlrpc_request *req, int new_client);
 724 #else
 725 static inline void
 726 check_and_start_recovery_timer(struct obd_device *obd,
 727                                struct ptlrpc_request *req, int new_client)
 728 {
 729 }
 730 #endif
 731
 732 int target_handle_connect(struct ptlrpc_request *req)
 733 {
 734         struct obd_device *target = NULL, *targref = NULL;
 735         struct obd_export *export = NULL;
 736         struct obd_import *revimp;
 737         struct lustre_handle conn;
 738         struct lustre_handle *tmp;
 739         struct obd_uuid tgtuuid;
 740         struct obd_uuid cluuid;
 741         struct obd_uuid remote_uuid;
 742         char *str;
 743         int rc = 0;
 744         char *target_start;
 745         int target_len;
 746         int mds_conn = 0;
 747         struct obd_connect_data *data, *tmpdata;
 748         int size, tmpsize;
 749         lnet_nid_t *client_nid = NULL;
 750         ENTRY;
 751
 752         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 753
 754         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 755         if (str == NULL) {
 756                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 757                 GOTO(out, rc = -EINVAL);
 758         }
 759
 760         obd_str2uuid(&tgtuuid, str);
 761         target = class_uuid2obd(&tgtuuid);
 762         if (!target)
 763                 target = class_name2obd(str);
 764
 765         if (!target) {
 766                 deuuidify(str, NULL, &target_start, &target_len);
 767                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available for "
 768                                    "connect (no target)\n", str);
 769                 GOTO(out, rc = -ENODEV);
 770         }
 771
 772         cfs_spin_lock(&target->obd_dev_lock);
 773         if (target->obd_stopping || !target->obd_set_up) {
 774                 cfs_spin_unlock(&target->obd_dev_lock);
 775
 776                 deuuidify(str, NULL, &target_start, &target_len);
 777                 LCONSOLE_ERROR_MSG(0x137, "%.*s: Not available for connect "
 778                                    "from %s (%s)\n", target_len, target_start,
 779                                    libcfs_nid2str(req->rq_peer.nid),
 780                                    (target->obd_stopping ?
 781                                    "stopping" : "not set up"));
 782                 GOTO(out, rc = -ENODEV);
 783         }
 784
 785         if (target->obd_no_conn) {
 786                 cfs_spin_unlock(&target->obd_dev_lock);
 787
 788                 LCONSOLE_WARN("%s: Temporarily refusing client connection "
 789                               "from %s\n", target->obd_name,
 790                               libcfs_nid2str(req->rq_peer.nid));
 791                 GOTO(out, rc = -EAGAIN);
 792         }
 793
 794         /* Make sure the target isn't cleaned up while we're here. Yes,
 795            there's still a race between the above check and our incref here.
 796            Really, class_uuid2obd should take the ref. */
 797         targref = class_incref(target, __FUNCTION__, cfs_current());
 798
 799         target->obd_conn_inprogress++;
 800         cfs_spin_unlock(&target->obd_dev_lock);
 801
 802         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 803         if (str == NULL) {
 804                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 805                 GOTO(out, rc = -EINVAL);
 806         }
 807
 808         obd_str2uuid(&cluuid, str);
 809
 810         /* XXX extract a nettype and format accordingly */
 811         switch (sizeof(lnet_nid_t)) {
 812                 /* NB the casts only avoid compiler warnings */
 813         case 8:
 814                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 815                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 816                 break;
 817         case 4:
 818                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 819                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 820                 break;
 821         default:
 822                 LBUG();
 823         }
 824
 825         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 826         if (tmp == NULL)
 827                 GOTO(out, rc = -EPROTO);
 828
 829         conn = *tmp;
 830
 831         size = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
 832                                     RCL_CLIENT);
 833         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 834         if (!data)
 835                 GOTO(out, rc = -EPROTO);
 836
 837         rc = req_capsule_server_pack(&req->rq_pill);
 838         if (rc)
 839                 GOTO(out, rc);
 840
 841         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 842                 if (!data) {
 843                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 844                                   "libclient connection attempt");
 845                         GOTO(out, rc = -EPROTO);
 846                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 847                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 848                            data->ocd_version > LUSTRE_VERSION_CODE +
 849                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 850                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 851                                   "libclient connection attempt",
 852                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 853                                   "old" : "new",
 854                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 855                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 856                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 857                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 858                         data = req_capsule_server_sized_get(&req->rq_pill,
 859                                                         &RMF_CONNECT_DATA,
 860                                     offsetof(typeof(*data), ocd_version) +
 861                                              sizeof(data->ocd_version));
 862                         if (data) {
 863                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 864                                 data->ocd_version = LUSTRE_VERSION_CODE;
 865                         }
 866                         GOTO(out, rc = -EPROTO);
 867                 }
 868         }
 869
 870         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 871             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 872                 mds_conn = 1;
 873
 874         /* lctl gets a backstage, all-access pass. */
 875         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 876                 goto dont_check_exports;
 877
 878         export = cfs_hash_lookup(target->obd_uuid_hash, &cluuid);
 879         if (!export)
 880                 goto no_export;
 881
 882         /* we've found an export in the hash */
 883         if (export->exp_connecting) { /* bug 9635, et. al. */
 884                 LCONSOLE_WARN("%s: Export %p already connecting from %s\n",
 885                               export->exp_obd->obd_name, export,
 886                               libcfs_nid2str(req->rq_peer.nid));
 887                 class_export_put(export);
 888                 export = NULL;
 889                 rc = -EALREADY;
 890         } else if (mds_conn && export->exp_connection) {
 891                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 892                         /* mds reconnected after failover */
 893                         LCONSOLE_WARN("%s: Received MDS connection from "
 894                             "%s, removing former export from %s\n",
 895                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 896                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 897                 else
 898                         /* new mds connection from the same nid */
 899                         LCONSOLE_WARN("%s: Received new MDS connection from "
 900                             "%s, removing former export from same NID\n",
 901                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 902                 class_fail_export(export);
 903                 class_export_put(export);
 904                 export = NULL;
 905                 rc = 0;
 906         } else if (export->exp_connection != NULL &&
 907                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 908                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 909                     MSG_CONNECT_INITIAL)) {
 910                 /* in mds failover we have static uuid but nid can be
 911                  * changed*/
 912                 LCONSOLE_WARN("%s: Client %s seen on new nid %s when "
 913                               "existing nid %s is already connected\n",
 914                               target->obd_name, cluuid.uuid,
 915                               libcfs_nid2str(req->rq_peer.nid),
 916                               libcfs_nid2str(
 917                                       export->exp_connection->c_peer.nid));
 918                 rc = -EALREADY;
 919                 class_export_put(export);
 920                 export = NULL;
 921         } else {
 922                 cfs_spin_lock(&export->exp_lock);
 923                 export->exp_connecting = 1;
 924                 cfs_spin_unlock(&export->exp_lock);
 925                 LASSERT(export->exp_obd == target);
 926
 927                 rc = target_handle_reconnect(&conn, export, &cluuid);
 928         }
 929
 930         /* If we found an export, we already unlocked. */
 931         if (!export) {
 932 no_export:
 933                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 934         } else if (req->rq_export == NULL &&
 935                    cfs_atomic_read(&export->exp_rpc_count) > 0) {
 936                 LCONSOLE_WARN("%s: Client %s (at %s) refused connection, "
 937                               "still busy with %d references\n",
 938                               target->obd_name, cluuid.uuid,
 939                               libcfs_nid2str(req->rq_peer.nid),
 940                               cfs_atomic_read(&export->exp_refcount));
 941                 GOTO(out, rc = -EBUSY);
 942         } else if (req->rq_export != NULL &&
 943                    (cfs_atomic_read(&export->exp_rpc_count) > 1)) {
 944                 /* the current connect rpc has increased exp_rpc_count */
 945                 LCONSOLE_WARN("%s: Client %s (at %s) refused reconnection, "
 946                               "still busy with %d active RPCs\n",
 947                               target->obd_name, cluuid.uuid,
 948                               libcfs_nid2str(req->rq_peer.nid),
 949                               cfs_atomic_read(&export->exp_rpc_count) - 1);
 950                 cfs_spin_lock(&export->exp_lock);
 951                 if (req->rq_export->exp_conn_cnt <
 952                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 953                         /* try to abort active requests */
 954                         req->rq_export->exp_abort_active_req = 1;
 955                 cfs_spin_unlock(&export->exp_lock);
 956                 GOTO(out, rc = -EBUSY);
 957         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 958                 if (!strstr(cluuid.uuid, "mdt"))
 959                         LCONSOLE_WARN("%s: Rejecting reconnect from the "
 960                                       "known client %s (at %s) because it "
 961                                       "is indicating it is a new client",
 962                                       target->obd_name, cluuid.uuid,
 963                                       libcfs_nid2str(req->rq_peer.nid));
 964                 GOTO(out, rc = -EALREADY);
 965         } else {
 966                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 967         }
 968
 969         if (rc < 0) {
 970                 GOTO(out, rc);
 971         }
 972
 973         CDEBUG(D_HA, "%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 974                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 975               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 976               export, (long)cfs_time_current_sec(),
 977               export ? (long)export->exp_last_request_time : 0);
 978
 979         /* If this is the first time a client connects, reset the recovery
 980          * timer */
 981         if (rc == 0 && target->obd_recovering)
 982                 check_and_start_recovery_timer(target, req, export == NULL);
 983
 984         /* We want to handle EALREADY but *not* -EALREADY from
 985          * target_handle_reconnect(), return reconnection state in a flag */
 986         if (rc == EALREADY) {
 987                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 988                 rc = 0;
 989         } else {
 990                 LASSERT(rc == 0);
 991         }
 992
 993         /* Tell the client if we support replayable requests */
 994         if (target->obd_replayable)
 995                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 996         client_nid = &req->rq_peer.nid;
 997
 998         if (export == NULL) {
 999                 if (target->obd_recovering) {
1000                         cfs_time_t t;
1001
1002                         t = cfs_timer_deadline(&target->obd_recovery_timer);
1003                         t = cfs_time_sub(t, cfs_time_current());
1004                         t = cfs_duration_sec(t);
1005                         LCONSOLE_WARN("%s: Denying connection for new client "
1006                                       "%s (at %s), waiting for %d clients in "
1007                                       "recovery for %d:%.02d\n",
1008                                       target->obd_name,
1009                                       libcfs_nid2str(req->rq_peer.nid),
1010                                       cluuid.uuid,
1011                                       cfs_atomic_read(&target-> \
1012                                                       obd_lock_replay_clients),
1013                                       (int)t / 60, (int)t % 60);
1014                         rc = -EBUSY;
1015                 } else {
1016 dont_check_exports:
1017                         rc = obd_connect(req->rq_svc_thread->t_env,
1018                                          &export, target, &cluuid, data,
1019                                          client_nid);
1020                         if (rc == 0)
1021                                 conn.cookie = export->exp_handle.h_cookie;
1022                 }
1023         } else {
1024                 rc = obd_reconnect(req->rq_svc_thread->t_env,
1025                                    export, target, &cluuid, data, client_nid);
1026         }
1027         if (rc)
1028                 GOTO(out, rc);
1029
1030         LASSERT(target->u.obt.obt_magic == OBT_MAGIC);
1031         data->ocd_instance = target->u.obt.obt_instance;
1032
1033         /* Return only the parts of obd_connect_data that we understand, so the
1034          * client knows that we don't understand the rest. */
1035         if (data) {
1036                 tmpsize = req_capsule_get_size(&req->rq_pill, &RMF_CONNECT_DATA,
1037                                                RCL_SERVER);
1038                 tmpdata = req_capsule_server_get(&req->rq_pill,
1039                                                  &RMF_CONNECT_DATA);
1040                 /* Don't use struct assignment here, because the client reply
1041                  * buffer may be smaller/larger than the local struct
1042                  * obd_connect_data. */
1043                 memcpy(tmpdata, data, min(tmpsize, size));
1044         }
1045
1046         /* If all else goes well, this is our RPC return code. */
1047         req->rq_status = 0;
1048
1049         lustre_msg_set_handle(req->rq_repmsg, &conn);
1050
1051         /* If the client and the server are the same node, we will already
1052          * have an export that really points to the client's DLM export,
1053          * because we have a shared handles table.
1054          *
1055          * XXX this will go away when shaver stops sending the "connect" handle
1056          * in the real "remote handle" field of the request --phik 24 Apr 2003
1057          */
1058         if (req->rq_export != NULL)
1059                 class_export_put(req->rq_export);
1060
1061         /* request takes one export refcount */
1062         req->rq_export = class_export_get(export);
1063
1064         cfs_spin_lock(&export->exp_lock);
1065         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
1066                 cfs_spin_unlock(&export->exp_lock);
1067                 CDEBUG(D_RPCTRACE, "%s: %s already connected at higher "
1068                        "conn_cnt: %d > %d\n",
1069                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
1070                        export->exp_conn_cnt,
1071                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
1072
1073                 GOTO(out, rc = -EALREADY);
1074         }
1075         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1076         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1077         export->exp_abort_active_req = 0;
1078
1079         /* request from liblustre?  Don't evict it for not pinging. */
1080         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1081                 export->exp_libclient = 1;
1082                 cfs_spin_unlock(&export->exp_lock);
1083
1084                 cfs_spin_lock(&target->obd_dev_lock);
1085                 cfs_list_del_init(&export->exp_obd_chain_timed);
1086                 cfs_spin_unlock(&target->obd_dev_lock);
1087         } else {
1088                 cfs_spin_unlock(&export->exp_lock);
1089         }
1090
1091         if (export->exp_connection != NULL) {
1092                 /* Check to see if connection came from another NID */
1093                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1094                     !cfs_hlist_unhashed(&export->exp_nid_hash))
1095                         cfs_hash_del(export->exp_obd->obd_nid_hash,
1096                                      &export->exp_connection->c_peer.nid,
1097                                      &export->exp_nid_hash);
1098
1099                 ptlrpc_connection_put(export->exp_connection);
1100         }
1101
1102         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1103                                                        req->rq_self,
1104                                                        &remote_uuid);
1105         if (cfs_hlist_unhashed(&export->exp_nid_hash)) {
1106                 cfs_hash_add(export->exp_obd->obd_nid_hash,
1107                              &export->exp_connection->c_peer.nid,
1108                              &export->exp_nid_hash);
1109         }
1110
1111         if (target->obd_recovering && !export->exp_in_recovery) {
1112                 int has_transno;
1113                 __u64 transno = data->ocd_transno;
1114
1115                 cfs_spin_lock(&export->exp_lock);
1116                 /* possible race with class_disconnect_stale_exports,
1117                  * export may be already in the eviction process */
1118                 if (export->exp_failed) {
1119                         cfs_spin_unlock(&export->exp_lock);
1120                         GOTO(out, rc = -ENODEV);
1121                 }
1122                 export->exp_in_recovery = 1;
1123                 export->exp_req_replay_needed = 1;
1124                 export->exp_lock_replay_needed = 1;
1125                 cfs_spin_unlock(&export->exp_lock);
1126
1127                 has_transno = !!(lustre_msg_get_op_flags(req->rq_reqmsg) &
1128                                  MSG_CONNECT_TRANSNO);
1129                 if (has_transno && transno == 0)
1130                         CWARN("Connect with zero transno!\n");
1131
1132                 if (has_transno && transno > 0 &&
1133                     transno < target->obd_next_recovery_transno &&
1134                     transno > target->obd_last_committed) {
1135                         /* another way is to use cmpxchg() so it will be
1136                          * lock free */
1137                         cfs_spin_lock(&target->obd_recovery_task_lock);
1138                         if (transno < target->obd_next_recovery_transno)
1139                                 target->obd_next_recovery_transno = transno;
1140                         cfs_spin_unlock(&target->obd_recovery_task_lock);
1141                 }
1142
1143                 cfs_atomic_inc(&target->obd_req_replay_clients);
1144                 cfs_atomic_inc(&target->obd_lock_replay_clients);
1145                 if (cfs_atomic_inc_return(&target->obd_connected_clients) ==
1146                     target->obd_max_recoverable_clients)
1147                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1148         }
1149
1150         /* Tell the client we're in recovery, when client is involved in it. */
1151         if (target->obd_recovering)
1152                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
1153
1154         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1155         conn = *tmp;
1156
1157         if (export->exp_imp_reverse != NULL) {
1158                 /* destroyed import can be still referenced in ctxt */
1159                 obd_set_info_async(req->rq_svc_thread->t_env, export,
1160                                    sizeof(KEY_REVIMP_UPD), KEY_REVIMP_UPD,
1161                                    0, NULL, NULL);
1162
1163                 client_destroy_import(export->exp_imp_reverse);
1164         }
1165
1166         /* for the rest part, we return -ENOTCONN in case of errors
1167          * in order to let client initialize connection again.
1168          */
1169         revimp = export->exp_imp_reverse = class_new_import(target);
1170         if (!revimp) {
1171                 CERROR("fail to alloc new reverse import.\n");
1172                 GOTO(out, rc = -ENOTCONN);
1173         }
1174
1175         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1176         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1177         revimp->imp_remote_handle = conn;
1178         revimp->imp_dlm_fake = 1;
1179         revimp->imp_state = LUSTRE_IMP_FULL;
1180
1181         /* unknown versions will be caught in
1182          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1183         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1184
1185         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1186             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1187                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1188         else
1189                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1190
1191         if ((export->exp_connect_flags & OBD_CONNECT_FULL20) &&
1192             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1193                 revimp->imp_msghdr_flags |= MSGHDR_CKSUM_INCOMPAT18;
1194         else
1195                 revimp->imp_msghdr_flags &= ~MSGHDR_CKSUM_INCOMPAT18;
1196
1197         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1198         if (rc) {
1199                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1200                 export->exp_imp_reverse = NULL;
1201                 class_destroy_import(revimp);
1202         }
1203
1204         class_import_put(revimp);
1205 out:
1206         if (export) {
1207                 cfs_spin_lock(&export->exp_lock);
1208                 export->exp_connecting = 0;
1209                 cfs_spin_unlock(&export->exp_lock);
1210
1211                 class_export_put(export);
1212         }
1213         if (targref) {
1214                 cfs_spin_lock(&target->obd_dev_lock);
1215                 target->obd_conn_inprogress--;
1216                 cfs_spin_unlock(&target->obd_dev_lock);
1217
1218                 class_decref(targref, __FUNCTION__, cfs_current());
1219         }
1220         if (rc)
1221                 req->rq_status = rc;
1222         RETURN(rc);
1223 }
1224
1225 int target_handle_disconnect(struct ptlrpc_request *req)
1226 {
1227         int rc;
1228         ENTRY;
1229
1230         rc = req_capsule_server_pack(&req->rq_pill);
1231         if (rc)
1232                 RETURN(rc);
1233
1234         /* keep the rq_export around so we can send the reply */
1235         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1236
1237         RETURN(0);
1238 }
1239
1240 void target_destroy_export(struct obd_export *exp)
1241 {
1242         /* exports created from last_rcvd data, and "fake"
1243            exports created by lctl don't have an import */
1244         if (exp->exp_imp_reverse != NULL)
1245                 client_destroy_import(exp->exp_imp_reverse);
1246
1247         LASSERT_ATOMIC_ZERO(&exp->exp_locks_count);
1248         LASSERT_ATOMIC_ZERO(&exp->exp_rpc_count);
1249         LASSERT_ATOMIC_ZERO(&exp->exp_cb_count);
1250         LASSERT_ATOMIC_ZERO(&exp->exp_replay_count);
1251 }
1252
1253 /*
1254  * Recovery functions
1255  */
1256 static void target_request_copy_get(struct ptlrpc_request *req)
1257 {
1258         class_export_rpc_get(req->rq_export);
1259         LASSERT(cfs_list_empty(&req->rq_list));
1260         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1261
1262         /* increase refcount to keep request in queue */
1263         cfs_atomic_inc(&req->rq_refcount);
1264         /** let export know it has replays to be handled */
1265         cfs_atomic_inc(&req->rq_export->exp_replay_count);
1266 }
1267
1268 static void target_request_copy_put(struct ptlrpc_request *req)
1269 {
1270         LASSERT(cfs_list_empty(&req->rq_replay_list));
1271         LASSERT_ATOMIC_POS(&req->rq_export->exp_replay_count);
1272
1273         cfs_atomic_dec(&req->rq_export->exp_replay_count);
1274         class_export_rpc_put(req->rq_export);
1275         ptlrpc_server_drop_request(req);
1276 }
1277
1278 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1279 {
1280         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1281         struct obd_export     *exp = req->rq_export;
1282         struct ptlrpc_request *reqiter;
1283         int                    dup = 0;
1284
1285         LASSERT(exp);
1286
1287         cfs_spin_lock(&exp->exp_lock);
1288         cfs_list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1289                                 rq_replay_list) {
1290                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1291                         dup = 1;
1292                         break;
1293                 }
1294         }
1295
1296         if (dup) {
1297                 /* we expect it with RESENT and REPLAY flags */
1298                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1299                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1300                         CERROR("invalid flags %x of resent replay\n",
1301                                lustre_msg_get_flags(req->rq_reqmsg));
1302         } else {
1303                 cfs_list_add_tail(&req->rq_replay_list,
1304                                   &exp->exp_req_replay_queue);
1305         }
1306
1307         cfs_spin_unlock(&exp->exp_lock);
1308         return dup;
1309 }
1310
1311 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1312 {
1313         LASSERT(!cfs_list_empty(&req->rq_replay_list));
1314         LASSERT(req->rq_export);
1315
1316         cfs_spin_lock(&req->rq_export->exp_lock);
1317         cfs_list_del_init(&req->rq_replay_list);
1318         cfs_spin_unlock(&req->rq_export->exp_lock);
1319 }
1320
1321 #ifdef __KERNEL__
1322 static void target_finish_recovery(struct obd_device *obd)
1323 {
1324         time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1325                                     obd->obd_recovery_start);
1326         ENTRY;
1327
1328         LCONSOLE_INFO("%s: Recovery over after %d:%.02d, of %d clients "
1329                       "%d recovered and %d %s evicted.\n", obd->obd_name,
1330                       (int)elapsed_time / 60, (int)elapsed_time % 60,
1331                       obd->obd_max_recoverable_clients,
1332                       cfs_atomic_read(&obd->obd_connected_clients),
1333                       obd->obd_stale_clients,
1334                       obd->obd_stale_clients == 1 ? "was" : "were");
1335
1336         ldlm_reprocess_all_ns(obd->obd_namespace);
1337         cfs_spin_lock(&obd->obd_recovery_task_lock);
1338         if (!cfs_list_empty(&obd->obd_req_replay_queue) ||
1339             !cfs_list_empty(&obd->obd_lock_replay_queue) ||
1340             !cfs_list_empty(&obd->obd_final_req_queue)) {
1341                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1342                        obd->obd_name,
1343                        cfs_list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1344                        cfs_list_empty(&obd->obd_lock_replay_queue) ? \
1345                                "" : "lock ",
1346                        cfs_list_empty(&obd->obd_final_req_queue) ? \
1347                                "" : "final ");
1348                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1349                 LBUG();
1350         }
1351         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1352
1353         obd->obd_recovery_end = cfs_time_current_sec();
1354
1355         /* when recovery finished, cleanup orphans on mds and ost */
1356         if (OBT(obd) && OBP(obd, postrecov)) {
1357                 int rc = OBP(obd, postrecov)(obd);
1358                 if (rc < 0)
1359                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1360                                       obd->obd_name, rc);
1361         }
1362         EXIT;
1363 }
1364
1365 static void abort_req_replay_queue(struct obd_device *obd)
1366 {
1367         struct ptlrpc_request *req, *n;
1368         cfs_list_t abort_list;
1369
1370         CFS_INIT_LIST_HEAD(&abort_list);
1371         cfs_spin_lock(&obd->obd_recovery_task_lock);
1372         cfs_list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1373         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1374         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1375                 DEBUG_REQ(D_WARNING, req, "aborted:");
1376                 req->rq_status = -ENOTCONN;
1377                 if (ptlrpc_error(req)) {
1378                         DEBUG_REQ(D_ERROR, req,
1379                                   "failed abort_req_reply; skipping");
1380                 }
1381                 target_exp_dequeue_req_replay(req);
1382                 target_request_copy_put(req);
1383         }
1384 }
1385
1386 static void abort_lock_replay_queue(struct obd_device *obd)
1387 {
1388         struct ptlrpc_request *req, *n;
1389         cfs_list_t abort_list;
1390
1391         CFS_INIT_LIST_HEAD(&abort_list);
1392         cfs_spin_lock(&obd->obd_recovery_task_lock);
1393         cfs_list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1394         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1395         cfs_list_for_each_entry_safe(req, n, &abort_list, rq_list){
1396                 DEBUG_REQ(D_ERROR, req, "aborted:");
1397                 req->rq_status = -ENOTCONN;
1398                 if (ptlrpc_error(req)) {
1399                         DEBUG_REQ(D_ERROR, req,
1400                                   "failed abort_lock_reply; skipping");
1401                 }
1402                 target_request_copy_put(req);
1403         }
1404 }
1405
1406 /* Called from a cleanup function if the device is being cleaned up
1407    forcefully.  The exports should all have been disconnected already,
1408    the only thing left to do is
1409      - clear the recovery flags
1410      - cancel the timer
1411      - free queued requests and replies, but don't send replies
1412    Because the obd_stopping flag is set, no new requests should be received.
1413
1414 */
1415 void target_cleanup_recovery(struct obd_device *obd)
1416 {
1417         struct ptlrpc_request *req, *n;
1418         cfs_list_t clean_list;
1419         ENTRY;
1420
1421         CFS_INIT_LIST_HEAD(&clean_list);
1422         cfs_spin_lock(&obd->obd_dev_lock);
1423         if (!obd->obd_recovering) {
1424                 cfs_spin_unlock(&obd->obd_dev_lock);
1425                 EXIT;
1426                 return;
1427         }
1428         obd->obd_recovering = obd->obd_abort_recovery = 0;
1429         cfs_spin_unlock(&obd->obd_dev_lock);
1430
1431         cfs_spin_lock(&obd->obd_recovery_task_lock);
1432         target_cancel_recovery_timer(obd);
1433         cfs_list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1434         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1435
1436         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1437                 LASSERT(req->rq_reply_state == 0);
1438                 target_exp_dequeue_req_replay(req);
1439                 target_request_copy_put(req);
1440         }
1441
1442         cfs_spin_lock(&obd->obd_recovery_task_lock);
1443         cfs_list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1444         cfs_list_splice_init(&obd->obd_final_req_queue, &clean_list);
1445         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1446
1447         cfs_list_for_each_entry_safe(req, n, &clean_list, rq_list){
1448                 LASSERT(req->rq_reply_state == 0);
1449                 target_request_copy_put(req);
1450         }
1451
1452         EXIT;
1453 }
1454
1455 /* obd_recovery_task_lock should be held */
1456 void target_cancel_recovery_timer(struct obd_device *obd)
1457 {
1458         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1459         cfs_timer_disarm(&obd->obd_recovery_timer);
1460 }
1461
1462 static void target_start_recovery_timer(struct obd_device *obd)
1463 {
1464         if (obd->obd_recovery_start != 0)
1465                 return;
1466
1467         cfs_spin_lock(&obd->obd_dev_lock);
1468         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1469                 cfs_spin_unlock(&obd->obd_dev_lock);
1470                 return;
1471         }
1472
1473         LASSERT(obd->obd_recovery_timeout != 0);
1474
1475         if (obd->obd_recovery_start != 0) {
1476                 cfs_spin_unlock(&obd->obd_dev_lock);
1477                 return;
1478         }
1479
1480         cfs_timer_arm(&obd->obd_recovery_timer,
1481                       cfs_time_shift(obd->obd_recovery_timeout));
1482         obd->obd_recovery_start = cfs_time_current_sec();
1483         cfs_spin_unlock(&obd->obd_dev_lock);
1484
1485         LCONSOLE_WARN("%s: Will be in recovery for at least %d:%.02d, "
1486                       "or until %d client%s reconnect%s\n",
1487                       obd->obd_name,
1488                       obd->obd_recovery_timeout / 60,
1489                       obd->obd_recovery_timeout % 60,
1490                       obd->obd_max_recoverable_clients,
1491                       (obd->obd_max_recoverable_clients == 1) ? "" : "s",
1492                       (obd->obd_max_recoverable_clients == 1) ? "s": "");
1493 }
1494
1495 /**
1496  * extend recovery window.
1497  *
1498  * if @extend is true, extend recovery window to have @drt remaining at least;
1499  * otherwise, make sure the recovery timeout value is not less than @drt.
1500  */
1501 static void extend_recovery_timer(struct obd_device *obd, int drt, bool extend)
1502 {
1503         cfs_time_t now;
1504         cfs_time_t end;
1505         cfs_duration_t left;
1506         int to;
1507
1508         cfs_spin_lock(&obd->obd_dev_lock);
1509         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1510                 cfs_spin_unlock(&obd->obd_dev_lock);
1511                 return;
1512         }
1513         LASSERT(obd->obd_recovery_start != 0);
1514
1515         now  = cfs_time_current_sec();
1516         to   = obd->obd_recovery_timeout;
1517         end  = obd->obd_recovery_start + to;
1518         left = cfs_time_sub(end, now);
1519
1520         if (extend && (drt > left)) {
1521                 to += drt - left;
1522         } else if (!extend && (drt > to)) {
1523                 to = drt;
1524                 /* reduce drt by already passed time */
1525                 drt -= obd->obd_recovery_timeout - left;
1526         }
1527
1528         if (to > obd->obd_recovery_time_hard)
1529                 to = obd->obd_recovery_time_hard;
1530         if (obd->obd_recovery_timeout < to) {
1531                 obd->obd_recovery_timeout = to;
1532                 cfs_timer_arm(&obd->obd_recovery_timer,
1533                               cfs_time_shift(drt));
1534         }
1535         cfs_spin_unlock(&obd->obd_dev_lock);
1536
1537         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1538                obd->obd_name, (unsigned)drt);
1539 }
1540
1541 /* Reset the timer with each new client connection */
1542 /*
1543  * This timer is actually reconnect_timer, which is for making sure
1544  * the total recovery window is at least as big as my reconnect
1545  * attempt timing. So the initial recovery time_out will be set to
1546  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1547  * from client is bigger than this, then the recovery time_out will
1548  * be extended to make sure the client could be reconnected, in the
1549  * process, the timeout from the new client should be ignored.
1550  */
1551
1552 static void
1553 check_and_start_recovery_timer(struct obd_device *obd,
1554                                struct ptlrpc_request *req,
1555                                int new_client)
1556 {
1557         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1558         struct obd_device_target *obt = &obd->u.obt;
1559         struct lustre_sb_info *lsi;
1560
1561         if (!new_client && service_time)
1562                 /* Teach server about old server's estimates, as first guess
1563                  * at how long new requests will take. */
1564                 at_measured(&req->rq_rqbd->rqbd_svcpt->scp_at_estimate,
1565                             service_time);
1566
1567         target_start_recovery_timer(obd);
1568
1569         /* convert the service time to rpc timeout,
1570          * reuse service_time to limit stack usage */
1571         service_time = at_est2timeout(service_time);
1572
1573         /* We expect other clients to timeout within service_time, then try
1574          * to reconnect, then try the failover server.  The max delay between
1575          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1576         service_time += 2 * INITIAL_CONNECT_TIMEOUT;
1577
1578         LASSERT(obt->obt_magic == OBT_MAGIC);
1579         lsi = s2lsi(obt->obt_sb);
1580         if (!(lsi->lsi_flags | LSI_IR_CAPABLE))
1581                 service_time += 2 * (CONNECTION_SWITCH_MAX +
1582                                      CONNECTION_SWITCH_INC);
1583         if (service_time > obd->obd_recovery_timeout && !new_client)
1584                 extend_recovery_timer(obd, service_time, false);
1585 }
1586
1587 /** Health checking routines */
1588 static inline int exp_connect_healthy(struct obd_export *exp)
1589 {
1590         return (exp->exp_in_recovery);
1591 }
1592
1593 /** if export done req_replay or has replay in queue */
1594 static inline int exp_req_replay_healthy(struct obd_export *exp)
1595 {
1596         return (!exp->exp_req_replay_needed ||
1597                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1598 }
1599 /** if export done lock_replay or has replay in queue */
1600 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1601 {
1602         return (!exp->exp_lock_replay_needed ||
1603                 cfs_atomic_read(&exp->exp_replay_count) > 0);
1604 }
1605
1606 static inline int exp_vbr_healthy(struct obd_export *exp)
1607 {
1608         return (!exp->exp_vbr_failed);
1609 }
1610
1611 static inline int exp_finished(struct obd_export *exp)
1612 {
1613         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1614 }
1615
1616 /** Checking routines for recovery */
1617 static int check_for_clients(struct obd_device *obd)
1618 {
1619         unsigned int clnts = cfs_atomic_read(&obd->obd_connected_clients);
1620
1621         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1622                 return 1;
1623         LASSERT(clnts <= obd->obd_max_recoverable_clients);
1624         if (obd->obd_no_conn == 0 &&
1625             clnts + obd->obd_stale_clients == obd->obd_max_recoverable_clients)
1626                 return 1;
1627         return 0;
1628 }
1629
1630 static int check_for_next_transno(struct obd_device *obd)
1631 {
1632         struct ptlrpc_request *req = NULL;
1633         int wake_up = 0, connected, completed, queue_len;
1634         __u64 next_transno, req_transno;
1635         ENTRY;
1636
1637         cfs_spin_lock(&obd->obd_recovery_task_lock);
1638         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1639                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1640                                      struct ptlrpc_request, rq_list);
1641                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1642         } else {
1643                 req_transno = 0;
1644         }
1645
1646         connected = cfs_atomic_read(&obd->obd_connected_clients);
1647         completed = connected - cfs_atomic_read(&obd->obd_req_replay_clients);
1648         queue_len = obd->obd_requests_queued_for_recovery;
1649         next_transno = obd->obd_next_recovery_transno;
1650
1651         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1652                "req_transno: "LPU64", next_transno: "LPU64"\n",
1653                obd->obd_max_recoverable_clients, connected, completed,
1654                queue_len, req_transno, next_transno);
1655
1656         if (obd->obd_abort_recovery) {
1657                 CDEBUG(D_HA, "waking for aborted recovery\n");
1658                 wake_up = 1;
1659         } else if (obd->obd_recovery_expired) {
1660                 CDEBUG(D_HA, "waking for expired recovery\n");
1661                 wake_up = 1;
1662         } else if (cfs_atomic_read(&obd->obd_req_replay_clients) == 0) {
1663                 CDEBUG(D_HA, "waking for completed recovery\n");
1664                 wake_up = 1;
1665         } else if (req_transno == next_transno) {
1666                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1667                 wake_up = 1;
1668         } else if (queue_len == cfs_atomic_read(&obd->obd_req_replay_clients)) {
1669                 int d_lvl = D_HA;
1670                 /** handle gaps occured due to lost reply or VBR */
1671                 LASSERTF(req_transno >= next_transno,
1672                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1673                          req_transno, next_transno);
1674                 if (req_transno > obd->obd_last_committed &&
1675                     !obd->obd_version_recov)
1676                         d_lvl = D_ERROR;
1677                 CDEBUG(d_lvl,
1678                        "%s: waking for gap in transno, VBR is %s (skip: "
1679                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1680                        ", last_committed: "LPD64")\n",
1681                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1682                        next_transno, queue_len, completed, connected,
1683                        req_transno, obd->obd_last_committed);
1684                 obd->obd_next_recovery_transno = req_transno;
1685                 wake_up = 1;
1686         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1687                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1688                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1689                 obd->obd_next_recovery_transno = req_transno;
1690                 wake_up = 1;
1691         }
1692         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1693         return wake_up;
1694 }
1695
1696 static int check_for_next_lock(struct obd_device *obd)
1697 {
1698         int wake_up = 0;
1699
1700         cfs_spin_lock(&obd->obd_recovery_task_lock);
1701         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1702                 CDEBUG(D_HA, "waking for next lock\n");
1703                 wake_up = 1;
1704         } else if (cfs_atomic_read(&obd->obd_lock_replay_clients) == 0) {
1705                 CDEBUG(D_HA, "waking for completed lock replay\n");
1706                 wake_up = 1;
1707         } else if (obd->obd_abort_recovery) {
1708                 CDEBUG(D_HA, "waking for aborted recovery\n");
1709                 wake_up = 1;
1710         } else if (obd->obd_recovery_expired) {
1711                 CDEBUG(D_HA, "waking for expired recovery\n");
1712                 wake_up = 1;
1713         }
1714         cfs_spin_unlock(&obd->obd_recovery_task_lock);
1715
1716         return wake_up;
1717 }
1718
1719 /**
1720  * wait for recovery events,
1721  * check its status with help of check_routine
1722  * evict dead clients via health_check
1723  */
1724 static int target_recovery_overseer(struct obd_device *obd,
1725                                     int (*check_routine)(struct obd_device *),
1726                                     int (*health_check)(struct obd_export *))
1727 {
1728 repeat:
1729         cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1730         if (obd->obd_abort_recovery) {
1731                 CWARN("recovery is aborted, evict exports in recovery\n");
1732                 /** evict exports which didn't finish recovery yet */
1733                 class_disconnect_stale_exports(obd, exp_finished);
1734                 return 1;
1735         } else if (obd->obd_recovery_expired) {
1736                 obd->obd_recovery_expired = 0;
1737                 /** If some clients died being recovered, evict them */
1738                 LCONSOLE_WARN("%s: recovery is timed out, "
1739                               "evict stale exports\n", obd->obd_name);
1740                 /** evict cexports with no replay in queue, they are stalled */
1741                 class_disconnect_stale_exports(obd, health_check);
1742                 /** continue with VBR */
1743                 cfs_spin_lock(&obd->obd_dev_lock);
1744                 obd->obd_version_recov = 1;
1745                 cfs_spin_unlock(&obd->obd_dev_lock);
1746                 /**
1747                  * reset timer, recovery will proceed with versions now,
1748                  * timeout is set just to handle reconnection delays
1749                  */
1750                 extend_recovery_timer(obd, RECONNECT_DELAY_MAX, true);
1751                 /** Wait for recovery events again, after evicting bad clients */
1752                 goto repeat;
1753         }
1754         return 0;
1755 }
1756
1757 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1758 {
1759         struct ptlrpc_request *req = NULL;
1760         ENTRY;
1761
1762         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1763                obd->obd_next_recovery_transno);
1764
1765         if (target_recovery_overseer(obd, check_for_next_transno,
1766                                      exp_req_replay_healthy)) {
1767                 abort_req_replay_queue(obd);
1768                 abort_lock_replay_queue(obd);
1769         }
1770
1771         cfs_spin_lock(&obd->obd_recovery_task_lock);
1772         if (!cfs_list_empty(&obd->obd_req_replay_queue)) {
1773                 req = cfs_list_entry(obd->obd_req_replay_queue.next,
1774                                      struct ptlrpc_request, rq_list);
1775                 cfs_list_del_init(&req->rq_list);
1776                 obd->obd_requests_queued_for_recovery--;
1777                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1778         } else {
1779                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1780                 LASSERT(cfs_list_empty(&obd->obd_req_replay_queue));
1781                 LASSERT(cfs_atomic_read(&obd->obd_req_replay_clients) == 0);
1782                 /** evict exports failed VBR */
1783                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1784         }
1785         RETURN(req);
1786 }
1787
1788 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1789 {
1790         struct ptlrpc_request *req = NULL;
1791
1792         CDEBUG(D_HA, "Waiting for lock\n");
1793         if (target_recovery_overseer(obd, check_for_next_lock,
1794                                      exp_lock_replay_healthy))
1795                 abort_lock_replay_queue(obd);
1796
1797         cfs_spin_lock(&obd->obd_recovery_task_lock);
1798         if (!cfs_list_empty(&obd->obd_lock_replay_queue)) {
1799                 req = cfs_list_entry(obd->obd_lock_replay_queue.next,
1800                                      struct ptlrpc_request, rq_list);
1801                 cfs_list_del_init(&req->rq_list);
1802                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1803         } else {
1804                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1805                 LASSERT(cfs_list_empty(&obd->obd_lock_replay_queue));
1806                 LASSERT(cfs_atomic_read(&obd->obd_lock_replay_clients) == 0);
1807                 /** evict exports failed VBR */
1808                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1809         }
1810         return req;
1811 }
1812
1813 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1814 {
1815         struct ptlrpc_request *req = NULL;
1816
1817         cfs_spin_lock(&obd->obd_recovery_task_lock);
1818         if (!cfs_list_empty(&obd->obd_final_req_queue)) {
1819                 req = cfs_list_entry(obd->obd_final_req_queue.next,
1820                                      struct ptlrpc_request, rq_list);
1821                 cfs_list_del_init(&req->rq_list);
1822                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1823                 if (req->rq_export->exp_in_recovery) {
1824                         cfs_spin_lock(&req->rq_export->exp_lock);
1825                         req->rq_export->exp_in_recovery = 0;
1826                         cfs_spin_unlock(&req->rq_export->exp_lock);
1827                 }
1828         } else {
1829                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1830         }
1831         return req;
1832 }
1833
1834 static int handle_recovery_req(struct ptlrpc_thread *thread,
1835                                struct ptlrpc_request *req,
1836                                svc_handler_t handler)
1837 {
1838         int rc;
1839         ENTRY;
1840
1841         /**
1842          * export can be evicted during recovery, no need to handle replays for
1843          * it after that, discard such request silently
1844          */
1845         if (req->rq_export->exp_disconnected)
1846                 GOTO(reqcopy_put, rc = 0);
1847
1848         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1849         if (rc) {
1850                 CERROR("Failure to initialize session: %d\n", rc);
1851                 GOTO(reqcopy_put, rc);
1852         }
1853
1854         req->rq_recov_session.lc_thread = thread;
1855         lu_context_enter(&req->rq_recov_session);
1856         req->rq_svc_thread = thread;
1857         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1858
1859         /* thread context */
1860         lu_context_enter(&thread->t_env->le_ctx);
1861         (void)handler(req);
1862         lu_context_exit(&thread->t_env->le_ctx);
1863
1864         lu_context_exit(&req->rq_recov_session);
1865         lu_context_fini(&req->rq_recov_session);
1866         /* don't reset timer for final stage */
1867         if (!exp_finished(req->rq_export)) {
1868                 int to = obd_timeout;
1869
1870                 /**
1871                  * Add request timeout to the recovery time so next request from
1872                  * this client may come in recovery time
1873                  */
1874                 if (!AT_OFF) {
1875                         struct ptlrpc_service_part *svcpt;
1876
1877                         svcpt = req->rq_rqbd->rqbd_svcpt;
1878                         /* If the server sent early reply for this request,
1879                          * the client will recalculate the timeout according to
1880                          * current server estimate service time, so we will
1881                          * use the maxium timeout here for waiting the client
1882                          * sending the next req */
1883                         to = max((int)at_est2timeout(
1884                                  at_get(&svcpt->scp_at_estimate)),
1885                                  (int)lustre_msg_get_timeout(req->rq_reqmsg));
1886                         /* Add net_latency (see ptlrpc_replay_req) */
1887                         to += lustre_msg_get_service_time(req->rq_reqmsg);
1888                 }
1889                 extend_recovery_timer(class_exp2obd(req->rq_export), to, true);
1890         }
1891 reqcopy_put:
1892         RETURN(rc);
1893 }
1894
1895 static int target_recovery_thread(void *arg)
1896 {
1897         struct lu_target *lut = arg;
1898         struct obd_device *obd = lut->lut_obd;
1899         struct ptlrpc_request *req;
1900         struct target_recovery_data *trd = &obd->obd_recovery_data;
1901         unsigned long delta;
1902         unsigned long flags;
1903         struct lu_env *env;
1904         struct ptlrpc_thread *thread = NULL;
1905         int rc = 0;
1906         ENTRY;
1907
1908         cfs_daemonize_ctxt("tgt_recov");
1909
1910         SIGNAL_MASK_LOCK(current, flags);
1911         sigfillset(&current->blocked);
1912         RECALC_SIGPENDING;
1913         SIGNAL_MASK_UNLOCK(current, flags);
1914
1915         OBD_ALLOC_PTR(thread);
1916         if (thread == NULL)
1917                 RETURN(-ENOMEM);
1918
1919         OBD_ALLOC_PTR(env);
1920         if (env == NULL) {
1921                 OBD_FREE_PTR(thread);
1922                 RETURN(-ENOMEM);
1923         }
1924
1925         rc = lu_context_init(&env->le_ctx, LCT_MD_THREAD | LCT_DT_THREAD);
1926         if (rc) {
1927                 OBD_FREE_PTR(thread);
1928                 OBD_FREE_PTR(env);
1929                 RETURN(rc);
1930         }
1931
1932         thread->t_env = env;
1933         thread->t_id = -1; /* force filter_iobuf_get/put to use local buffers */
1934         env->le_ctx.lc_thread = thread;
1935         thread->t_data = NULL;
1936         thread->t_watchdog = NULL;
1937
1938         CDEBUG(D_HA, "%s: started recovery thread pid %d\n", obd->obd_name,
1939                cfs_curproc_pid());
1940         trd->trd_processing_task = cfs_curproc_pid();
1941
1942         cfs_spin_lock(&obd->obd_dev_lock);
1943         obd->obd_recovering = 1;
1944         cfs_spin_unlock(&obd->obd_dev_lock);
1945         cfs_complete(&trd->trd_starting);
1946
1947         /* first of all, we have to know the first transno to replay */
1948         if (target_recovery_overseer(obd, check_for_clients,
1949                                      exp_connect_healthy)) {
1950                 abort_req_replay_queue(obd);
1951                 abort_lock_replay_queue(obd);
1952         }
1953
1954         /* next stage: replay requests */
1955         delta = jiffies;
1956         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1957                cfs_atomic_read(&obd->obd_req_replay_clients),
1958                obd->obd_next_recovery_transno);
1959         while ((req = target_next_replay_req(obd))) {
1960                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1961                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1962                           lustre_msg_get_transno(req->rq_reqmsg),
1963                           libcfs_nid2str(req->rq_peer.nid));
1964                 handle_recovery_req(thread, req,
1965                                     trd->trd_recovery_handler);
1966                 /**
1967                  * bz18031: increase next_recovery_transno before
1968                  * target_request_copy_put() will drop exp_rpc reference
1969                  */
1970                 cfs_spin_lock(&obd->obd_recovery_task_lock);
1971                 obd->obd_next_recovery_transno++;
1972                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
1973                 target_exp_dequeue_req_replay(req);
1974                 target_request_copy_put(req);
1975                 obd->obd_replayed_requests++;
1976         }
1977
1978         /**
1979          * The second stage: replay locks
1980          */
1981         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1982                cfs_atomic_read(&obd->obd_lock_replay_clients));
1983         while ((req = target_next_replay_lock(obd))) {
1984                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1985                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1986                           libcfs_nid2str(req->rq_peer.nid));
1987                 handle_recovery_req(thread, req,
1988                                     trd->trd_recovery_handler);
1989                 target_request_copy_put(req);
1990                 obd->obd_replayed_locks++;
1991         }
1992
1993         /**
1994          * The third stage: reply on final pings, at this moment all clients
1995          * must have request in final queue
1996          */
1997         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1998         /** Update server last boot epoch */
1999         lut_boot_epoch_update(lut);
2000         /* We drop recoverying flag to forward all new requests
2001          * to regular mds_handle() since now */
2002         cfs_spin_lock(&obd->obd_dev_lock);
2003         obd->obd_recovering = obd->obd_abort_recovery = 0;
2004         cfs_spin_unlock(&obd->obd_dev_lock);
2005         cfs_spin_lock(&obd->obd_recovery_task_lock);
2006         target_cancel_recovery_timer(obd);
2007         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2008         while ((req = target_next_final_ping(obd))) {
2009                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
2010                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
2011                           libcfs_nid2str(req->rq_peer.nid));
2012                 handle_recovery_req(thread, req,
2013                                     trd->trd_recovery_handler);
2014                 target_request_copy_put(req);
2015         }
2016
2017         delta = (jiffies - delta) / CFS_HZ;
2018         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
2019               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
2020         if (delta > OBD_RECOVERY_TIME_SOFT) {
2021                 CWARN("too long recovery - read logs\n");
2022                 libcfs_debug_dumplog();
2023         }
2024
2025         target_finish_recovery(obd);
2026
2027         lu_context_fini(&env->le_ctx);
2028         trd->trd_processing_task = 0;
2029         cfs_complete(&trd->trd_finishing);
2030
2031         OBD_FREE_PTR(thread);
2032         OBD_FREE_PTR(env);
2033         RETURN(rc);
2034 }
2035
2036 static int target_start_recovery_thread(struct lu_target *lut,
2037                                         svc_handler_t handler)
2038 {
2039         struct obd_device *obd = lut->lut_obd;
2040         int rc = 0;
2041         struct target_recovery_data *trd = &obd->obd_recovery_data;
2042
2043         memset(trd, 0, sizeof(*trd));
2044         cfs_init_completion(&trd->trd_starting);
2045         cfs_init_completion(&trd->trd_finishing);
2046         trd->trd_recovery_handler = handler;
2047
2048         if (cfs_create_thread(target_recovery_thread, lut, 0) > 0) {
2049                 cfs_wait_for_completion(&trd->trd_starting);
2050                 LASSERT(obd->obd_recovering != 0);
2051         } else
2052                 rc = -ECHILD;
2053
2054         return rc;
2055 }
2056
2057 void target_stop_recovery_thread(struct obd_device *obd)
2058 {
2059         if (obd->obd_recovery_data.trd_processing_task > 0) {
2060                 struct target_recovery_data *trd = &obd->obd_recovery_data;
2061                 /** recovery can be done but postrecovery is not yet */
2062                 cfs_spin_lock(&obd->obd_dev_lock);
2063                 if (obd->obd_recovering) {
2064                         CERROR("%s: Aborting recovery\n", obd->obd_name);
2065                         obd->obd_abort_recovery = 1;
2066                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2067                 }
2068                 cfs_spin_unlock(&obd->obd_dev_lock);
2069                 cfs_wait_for_completion(&trd->trd_finishing);
2070         }
2071 }
2072
2073 void target_recovery_fini(struct obd_device *obd)
2074 {
2075         class_disconnect_exports(obd);
2076         target_stop_recovery_thread(obd);
2077         target_cleanup_recovery(obd);
2078 }
2079 EXPORT_SYMBOL(target_recovery_fini);
2080
2081 static void target_recovery_expired(unsigned long castmeharder)
2082 {
2083         struct obd_device *obd = (struct obd_device *)castmeharder;
2084         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
2085                " after %lds (%d clients connected)\n",
2086                obd->obd_name, cfs_atomic_read(&obd->obd_lock_replay_clients),
2087                cfs_time_current_sec()- obd->obd_recovery_start,
2088                cfs_atomic_read(&obd->obd_connected_clients));
2089
2090         obd->obd_recovery_expired = 1;
2091         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2092 }
2093
2094 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
2095 {
2096         struct obd_device *obd = lut->lut_obd;
2097         if (obd->obd_max_recoverable_clients == 0) {
2098                 /** Update server last boot epoch */
2099                 lut_boot_epoch_update(lut);
2100                 return;
2101         }
2102
2103         CWARN("RECOVERY: service %s, %d recoverable clients, "
2104               "last_transno "LPU64"\n", obd->obd_name,
2105               obd->obd_max_recoverable_clients, obd->obd_last_committed);
2106         LASSERT(obd->obd_stopping == 0);
2107         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
2108         obd->obd_recovery_start = 0;
2109         obd->obd_recovery_end = 0;
2110
2111         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
2112         target_start_recovery_thread(lut, handler);
2113 }
2114 EXPORT_SYMBOL(target_recovery_init);
2115
2116 #endif /* __KERNEL__ */
2117
2118 static int target_process_req_flags(struct obd_device *obd,
2119                                     struct ptlrpc_request *req)
2120 {
2121         struct obd_export *exp = req->rq_export;
2122         LASSERT(exp != NULL);
2123         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2124                 /* client declares he's ready to replay locks */
2125                 cfs_spin_lock(&exp->exp_lock);
2126                 if (exp->exp_req_replay_needed) {
2127                         exp->exp_req_replay_needed = 0;
2128                         cfs_spin_unlock(&exp->exp_lock);
2129
2130                         LASSERT_ATOMIC_POS(&obd->obd_req_replay_clients);
2131                         cfs_atomic_dec(&obd->obd_req_replay_clients);
2132                 } else {
2133                         cfs_spin_unlock(&exp->exp_lock);
2134                 }
2135         }
2136         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2137                 /* client declares he's ready to complete recovery
2138                  * so, we put the request on th final queue */
2139                 cfs_spin_lock(&exp->exp_lock);
2140                 if (exp->exp_lock_replay_needed) {
2141                         exp->exp_lock_replay_needed = 0;
2142                         cfs_spin_unlock(&exp->exp_lock);
2143
2144                         LASSERT_ATOMIC_POS(&obd->obd_lock_replay_clients);
2145                         cfs_atomic_dec(&obd->obd_lock_replay_clients);
2146                 } else {
2147                         cfs_spin_unlock(&exp->exp_lock);
2148                 }
2149         }
2150         return 0;
2151 }
2152
2153 int target_queue_recovery_request(struct ptlrpc_request *req,
2154                                   struct obd_device *obd)
2155 {
2156         cfs_list_t *tmp;
2157         int inserted = 0;
2158         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
2159         ENTRY;
2160
2161         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
2162                 /* Processing the queue right now, don't re-add. */
2163                 RETURN(1);
2164         }
2165
2166         target_process_req_flags(obd, req);
2167
2168         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
2169                 /* client declares he's ready to complete recovery
2170                  * so, we put the request on th final queue */
2171                 target_request_copy_get(req);
2172                 DEBUG_REQ(D_HA, req, "queue final req");
2173                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2174                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2175                 if (obd->obd_recovering) {
2176                         cfs_list_add_tail(&req->rq_list,
2177                                           &obd->obd_final_req_queue);
2178                 } else {
2179                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2180                         target_request_copy_put(req);
2181                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
2182                 }
2183                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2184                 RETURN(0);
2185         }
2186         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
2187                 /* client declares he's ready to replay locks */
2188                 target_request_copy_get(req);
2189                 DEBUG_REQ(D_HA, req, "queue lock replay req");
2190                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
2191                 cfs_spin_lock(&obd->obd_recovery_task_lock);
2192                 LASSERT(obd->obd_recovering);
2193                 /* usually due to recovery abort */
2194                 if (!req->rq_export->exp_in_recovery) {
2195                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2196                         target_request_copy_put(req);
2197                         RETURN(-ENOTCONN);
2198                 }
2199                 LASSERT(req->rq_export->exp_lock_replay_needed);
2200                 cfs_list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
2201                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2202                 RETURN(0);
2203         }
2204
2205         /* CAVEAT EMPTOR: The incoming request message has been swabbed
2206          * (i.e. buflens etc are in my own byte order), but type-dependent
2207          * buffers (eg mdt_body, ost_body etc) have NOT been swabbed. */
2208
2209         if (!transno) {
2210                 CFS_INIT_LIST_HEAD(&req->rq_list);
2211                 DEBUG_REQ(D_HA, req, "not queueing");
2212                 RETURN(1);
2213         }
2214
2215         /* If we're processing the queue, we want don't want to queue this
2216          * message.
2217          *
2218          * Also, if this request has a transno less than the one we're waiting
2219          * for, we should process it now.  It could (and currently always will)
2220          * be an open request for a descriptor that was opened some time ago.
2221          *
2222          * Also, a resent, replayed request that has already been
2223          * handled will pass through here and be processed immediately.
2224          */
2225         CDEBUG(D_HA, "Next recovery transno: "LPU64
2226                ", current: "LPU64", replaying\n",
2227                obd->obd_next_recovery_transno, transno);
2228         cfs_spin_lock(&obd->obd_recovery_task_lock);
2229         if (transno < obd->obd_next_recovery_transno) {
2230                 /* Processing the queue right now, don't re-add. */
2231                 LASSERT(cfs_list_empty(&req->rq_list));
2232                 cfs_spin_unlock(&obd->obd_recovery_task_lock);
2233                 RETURN(1);
2234         }
2235         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2236
2237         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2238                 RETURN(0);
2239
2240         target_request_copy_get(req);
2241         if (!req->rq_export->exp_in_recovery) {
2242                 target_request_copy_put(req);
2243                 RETURN(-ENOTCONN);
2244         }
2245         LASSERT(req->rq_export->exp_req_replay_needed);
2246
2247         if (target_exp_enqueue_req_replay(req)) {
2248                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2249                 target_request_copy_put(req);
2250                 RETURN(0);
2251         }
2252
2253         /* XXX O(n^2) */
2254         cfs_spin_lock(&obd->obd_recovery_task_lock);
2255         LASSERT(obd->obd_recovering);
2256         cfs_list_for_each(tmp, &obd->obd_req_replay_queue) {
2257                 struct ptlrpc_request *reqiter =
2258                         cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2259
2260                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2261                         cfs_list_add_tail(&req->rq_list, &reqiter->rq_list);
2262                         inserted = 1;
2263                         break;
2264                 }
2265
2266                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2267                              transno)) {
2268                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2269                                   "has been claimed by another client");
2270                         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2271                         target_exp_dequeue_req_replay(req);
2272                         target_request_copy_put(req);
2273                         RETURN(0);
2274                 }
2275         }
2276
2277         if (!inserted)
2278                 cfs_list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2279
2280         obd->obd_requests_queued_for_recovery++;
2281         cfs_spin_unlock(&obd->obd_recovery_task_lock);
2282         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2283         RETURN(0);
2284 }
2285
2286 int target_handle_ping(struct ptlrpc_request *req)
2287 {
2288         obd_ping(req->rq_svc_thread->t_env, req->rq_export);
2289         return req_capsule_server_pack(&req->rq_pill);
2290 }
2291
2292 void target_committed_to_req(struct ptlrpc_request *req)
2293 {
2294         struct obd_export *exp = req->rq_export;
2295
2296         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2297                 lustre_msg_set_last_committed(req->rq_repmsg,
2298                                               exp->exp_last_committed);
2299         else
2300                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2301                           "%d)", exp->exp_obd->obd_no_transno,
2302                           req->rq_repmsg == NULL);
2303
2304         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2305                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2306 }
2307 EXPORT_SYMBOL(target_committed_to_req);
2308
2309 #endif /* HAVE_SERVER_SUPPORT */
2310
2311 /**
2312  * Packs current SLV and Limit into \a req.
2313  */
2314 int target_pack_pool_reply(struct ptlrpc_request *req)
2315 {
2316         struct obd_device *obd;
2317         ENTRY;
2318
2319         /*
2320          * Check that we still have all structures alive as this may
2321          * be some late rpc in shutdown time.
2322          */
2323         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2324                      !exp_connect_lru_resize(req->rq_export))) {
2325                 lustre_msg_set_slv(req->rq_repmsg, 0);
2326                 lustre_msg_set_limit(req->rq_repmsg, 0);
2327                 RETURN(0);
2328         }
2329
2330         /*
2331          * OBD is alive here as export is alive, which we checked above.
2332          */
2333         obd = req->rq_export->exp_obd;
2334
2335         cfs_read_lock(&obd->obd_pool_lock);
2336         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2337         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2338         cfs_read_unlock(&obd->obd_pool_lock);
2339
2340         RETURN(0);
2341 }
2342
2343 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2344 {
2345         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2346                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2347                 return (-ECOMM);
2348         }
2349
2350         if (unlikely(rc)) {
2351                 DEBUG_REQ(D_NET, req, "processing error (%d)", rc);
2352                 req->rq_status = rc;
2353                 return (ptlrpc_send_error(req, 1));
2354         } else {
2355                 DEBUG_REQ(D_NET, req, "sending reply");
2356         }
2357
2358         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2359 }
2360
2361 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2362 {
2363         struct ptlrpc_service_part *svcpt;
2364         int                        netrc;
2365         struct ptlrpc_reply_state *rs;
2366         struct obd_export         *exp;
2367         ENTRY;
2368
2369         if (req->rq_no_reply) {
2370                 EXIT;
2371                 return;
2372         }
2373
2374         svcpt = req->rq_rqbd->rqbd_svcpt;
2375         rs = req->rq_reply_state;
2376         if (rs == NULL || !rs->rs_difficult) {
2377                 /* no notifiers */
2378                 target_send_reply_msg (req, rc, fail_id);
2379                 EXIT;
2380                 return;
2381         }
2382
2383         /* must be an export if locks saved */
2384         LASSERT (req->rq_export != NULL);
2385         /* req/reply consistent */
2386         LASSERT(rs->rs_svcpt == svcpt);
2387
2388         /* "fresh" reply */
2389         LASSERT (!rs->rs_scheduled);
2390         LASSERT (!rs->rs_scheduled_ever);
2391         LASSERT (!rs->rs_handled);
2392         LASSERT (!rs->rs_on_net);
2393         LASSERT (rs->rs_export == NULL);
2394         LASSERT (cfs_list_empty(&rs->rs_obd_list));
2395         LASSERT (cfs_list_empty(&rs->rs_exp_list));
2396
2397         exp = class_export_get (req->rq_export);
2398
2399         /* disable reply scheduling while I'm setting up */
2400         rs->rs_scheduled = 1;
2401         rs->rs_on_net    = 1;
2402         rs->rs_xid       = req->rq_xid;
2403         rs->rs_transno   = req->rq_transno;
2404         rs->rs_export    = exp;
2405         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2406
2407         cfs_spin_lock(&exp->exp_uncommitted_replies_lock);
2408         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2409                rs->rs_transno, exp->exp_last_committed);
2410         if (rs->rs_transno > exp->exp_last_committed) {
2411                 /* not committed already */
2412                 cfs_list_add_tail(&rs->rs_obd_list,
2413                                   &exp->exp_uncommitted_replies);
2414         }
2415         cfs_spin_unlock (&exp->exp_uncommitted_replies_lock);
2416
2417         cfs_spin_lock(&exp->exp_lock);
2418         cfs_list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2419         cfs_spin_unlock(&exp->exp_lock);
2420
2421         netrc = target_send_reply_msg (req, rc, fail_id);
2422
2423         cfs_spin_lock(&svcpt->scp_rep_lock);
2424
2425         cfs_atomic_inc(&svcpt->scp_nreps_difficult);
2426
2427         if (netrc != 0) {
2428                 /* error sending: reply is off the net.  Also we need +1
2429                  * reply ref until ptlrpc_handle_rs() is done
2430                  * with the reply state (if the send was successful, there
2431                  * would have been +1 ref for the net, which
2432                  * reply_out_callback leaves alone) */
2433                 rs->rs_on_net = 0;
2434                 ptlrpc_rs_addref(rs);
2435         }
2436
2437         cfs_spin_lock(&rs->rs_lock);
2438         if (rs->rs_transno <= exp->exp_last_committed ||
2439             (!rs->rs_on_net && !rs->rs_no_ack) ||
2440              cfs_list_empty(&rs->rs_exp_list) ||     /* completed already */
2441              cfs_list_empty(&rs->rs_obd_list)) {
2442                 CDEBUG(D_HA, "Schedule reply immediately\n");
2443                 ptlrpc_dispatch_difficult_reply(rs);
2444         } else {
2445                 cfs_list_add(&rs->rs_list, &svcpt->scp_rep_active);
2446                 rs->rs_scheduled = 0;   /* allow notifier to schedule */
2447         }
2448         cfs_spin_unlock(&rs->rs_lock);
2449         cfs_spin_unlock(&svcpt->scp_rep_lock);
2450         EXIT;
2451 }
2452
2453 int target_handle_qc_callback(struct ptlrpc_request *req)
2454 {
2455         struct obd_quotactl *oqctl;
2456         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2457
2458         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2459         if (oqctl == NULL) {
2460                 CERROR("Can't unpack obd_quotactl\n");
2461                 RETURN(-EPROTO);
2462         }
2463
2464         cli->cl_qchk_stat = oqctl->qc_stat;
2465
2466         return 0;
2467 }
2468
2469 #ifdef HAVE_QUOTA_SUPPORT
2470 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2471 {
2472 #ifdef __KERNEL__
2473         struct obd_device *obd = req->rq_export->exp_obd;
2474         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2475         struct obd_device_target *obt;
2476         struct lustre_quota_ctxt *qctxt;
2477         struct qunit_data *qdata = NULL;
2478         int rc = 0;
2479         ENTRY;
2480
2481         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2482                 RETURN(rc);
2483
2484         rc = req_capsule_server_pack(&req->rq_pill);
2485         if (rc) {
2486                 CERROR("packing reply failed!: rc = %d\n", rc);
2487                 RETURN(rc);
2488         }
2489
2490         LASSERT(req->rq_export);
2491
2492         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2493         if (IS_ERR(qdata)) {
2494                 rc = PTR_ERR(qdata);
2495                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2496                 req->rq_status = rc;
2497                 GOTO(out, rc);
2498         }
2499
2500         /* we use the observer */
2501         if (obd_pin_observer(obd, &lov_obd) ||
2502             obd_pin_observer(lov_obd, &master_obd)) {
2503                 CERROR("Can't find the observer, it is recovering\n");
2504                 req->rq_status = -EAGAIN;
2505                 GOTO(out, rc);
2506         }
2507
2508         obt = &master_obd->u.obt;
2509         qctxt = &obt->obt_qctxt;
2510
2511         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2512                 /* quota_type has not been processed yet, return EAGAIN
2513                  * until we know whether or not quotas are supposed to
2514                  * be enabled */
2515                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2516                        "-EAGAIN\n");
2517                 req->rq_status = -EAGAIN;
2518                 GOTO(out, rc);
2519         }
2520
2521         cfs_down_read(&obt->obt_rwsem);
2522         if (qctxt->lqc_lqs_hash == NULL) {
2523                 cfs_up_read(&obt->obt_rwsem);
2524                 /* quota_type has not been processed yet, return EAGAIN
2525                  * until we know whether or not quotas are supposed to
2526                  * be enabled */
2527                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2528                        "-EAGAIN\n");
2529                 req->rq_status = -EAGAIN;
2530                 GOTO(out, rc);
2531         }
2532
2533         LASSERT(qctxt->lqc_handler);
2534         rc = qctxt->lqc_handler(master_obd, qdata,
2535                                 lustre_msg_get_opc(req->rq_reqmsg));
2536         cfs_up_read(&obt->obt_rwsem);
2537         if (rc && rc != -EDQUOT)
2538                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2539                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2540         req->rq_status = rc;
2541
2542         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2543         if (rc < 0) {
2544                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2545                 GOTO(out, rc);
2546         }
2547
2548         /* Block the quota req. b=14840 */
2549         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2550         EXIT;
2551
2552 out:
2553         if (master_obd)
2554                 obd_unpin_observer(lov_obd);
2555         if (lov_obd)
2556                 obd_unpin_observer(obd);
2557
2558         rc = ptlrpc_reply(req);
2559         return rc;
2560 #else
2561         return 0;
2562 #endif /* !__KERNEL__ */
2563 }
2564 #endif /* HAVE_QUOTA_SUPPORT */
2565
2566 ldlm_mode_t lck_compat_array[] = {
2567         [LCK_EX] LCK_COMPAT_EX,
2568         [LCK_PW] LCK_COMPAT_PW,
2569         [LCK_PR] LCK_COMPAT_PR,
2570         [LCK_CW] LCK_COMPAT_CW,
2571         [LCK_CR] LCK_COMPAT_CR,
2572         [LCK_NL] LCK_COMPAT_NL,
2573         [LCK_GROUP] LCK_COMPAT_GROUP,
2574         [LCK_COS] LCK_COMPAT_COS,
2575 };
2576
2577 /**
2578  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2579  * not escape to the user level.
2580  */
2581 int ldlm_error2errno(ldlm_error_t error)
2582 {
2583         int result;
2584
2585         switch (error) {
2586         case ELDLM_OK:
2587                 result = 0;
2588                 break;
2589         case ELDLM_LOCK_CHANGED:
2590                 result = -ESTALE;
2591                 break;
2592         case ELDLM_LOCK_ABORTED:
2593                 result = -ENAVAIL;
2594                 break;
2595         case ELDLM_LOCK_REPLACED:
2596                 result = -ESRCH;
2597                 break;
2598         case ELDLM_NO_LOCK_DATA:
2599                 result = -ENOENT;
2600                 break;
2601         case ELDLM_NAMESPACE_EXISTS:
2602                 result = -EEXIST;
2603                 break;
2604         case ELDLM_BAD_NAMESPACE:
2605                 result = -EBADF;
2606                 break;
2607         default:
2608                 if (((int)error) < 0)  /* cast to signed type */
2609                         result = error; /* as ldlm_error_t can be unsigned */
2610                 else {
2611                         CERROR("Invalid DLM result code: %d\n", error);
2612                         result = -EPROTO;
2613                 }
2614         }
2615         return result;
2616 }
2617 EXPORT_SYMBOL(ldlm_error2errno);
2618
2619 /**
2620  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2621  */
2622 ldlm_error_t ldlm_errno2error(int err_no)
2623 {
2624         int error;
2625
2626         switch (err_no) {
2627         case 0:
2628                 error = ELDLM_OK;
2629                 break;
2630         case -ESTALE:
2631                 error = ELDLM_LOCK_CHANGED;
2632                 break;
2633         case -ENAVAIL:
2634                 error = ELDLM_LOCK_ABORTED;
2635                 break;
2636         case -ESRCH:
2637                 error = ELDLM_LOCK_REPLACED;
2638                 break;
2639         case -ENOENT:
2640                 error = ELDLM_NO_LOCK_DATA;
2641                 break;
2642         case -EEXIST:
2643                 error = ELDLM_NAMESPACE_EXISTS;
2644                 break;
2645         case -EBADF:
2646                 error = ELDLM_BAD_NAMESPACE;
2647                 break;
2648         default:
2649                 error = err_no;
2650         }
2651         return error;
2652 }
2653 EXPORT_SYMBOL(ldlm_errno2error);
2654
2655 #if LUSTRE_TRACKS_LOCK_EXP_REFS
2656 void ldlm_dump_export_locks(struct obd_export *exp)
2657 {
2658         cfs_spin_lock(&exp->exp_locks_list_guard);
2659         if (!cfs_list_empty(&exp->exp_locks_list)) {
2660             struct ldlm_lock *lock;
2661
2662             CERROR("dumping locks for export %p,"
2663                    "ignore if the unmount doesn't hang\n", exp);
2664             cfs_list_for_each_entry(lock, &exp->exp_locks_list, l_exp_refs_link)
2665                 LDLM_ERROR(lock, "lock:");
2666         }
2667         cfs_spin_unlock(&exp->exp_locks_list_guard);
2668 }
2669 #endif
2670
2671 #ifdef HAVE_SERVER_SUPPORT
2672 static int target_bulk_timeout(void *data)
2673 {
2674         ENTRY;
2675         /* We don't fail the connection here, because having the export
2676          * killed makes the (vital) call to commitrw very sad.
2677          */
2678         RETURN(1);
2679 }
2680
2681 static inline char *bulk2type(struct ptlrpc_bulk_desc *desc)
2682 {
2683         return desc->bd_type == BULK_GET_SINK ? "GET" : "PUT";
2684 }
2685
2686 int target_bulk_io(struct obd_export *exp, struct ptlrpc_bulk_desc *desc,
2687                    struct l_wait_info *lwi)
2688 {
2689         struct ptlrpc_request *req = desc->bd_req;
2690         int rc = 0;
2691         ENTRY;
2692
2693         /* Check if there is eviction in progress, and if so, wait for
2694          * it to finish */
2695         if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
2696                 *lwi = LWI_INTR(NULL, NULL);
2697                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
2698                                   !cfs_atomic_read(&exp->exp_obd->
2699                                                    obd_evict_inprogress),
2700                                   lwi);
2701         }
2702
2703         /* Check if client was evicted or tried to reconnect already */
2704         if (exp->exp_failed || exp->exp_abort_active_req) {
2705                 rc = -ENOTCONN;
2706         } else {
2707                 if (desc->bd_type == BULK_PUT_SINK)
2708                         rc = sptlrpc_svc_wrap_bulk(req, desc);
2709                 if (rc == 0)
2710                         rc = ptlrpc_start_bulk_transfer(desc);
2711         }
2712
2713         if (rc == 0 && OBD_FAIL_CHECK(OBD_FAIL_MDS_SENDPAGE)) {
2714                 ptlrpc_abort_bulk(desc);
2715         } else if (rc == 0) {
2716                 time_t start = cfs_time_current_sec();
2717                 do {
2718                         long timeoutl = req->rq_deadline - cfs_time_current_sec();
2719                         cfs_duration_t timeout = timeoutl <= 0 ?
2720                                 CFS_TICK : cfs_time_seconds(timeoutl);
2721                         *lwi = LWI_TIMEOUT_INTERVAL(timeout,
2722                                                     cfs_time_seconds(1),
2723                                                    target_bulk_timeout,
2724                                                    desc);
2725                         rc = l_wait_event(desc->bd_waitq,
2726                                           !ptlrpc_server_bulk_active(desc) ||
2727                                           exp->exp_failed ||
2728                                           exp->exp_abort_active_req,
2729                                           lwi);
2730                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
2731                         /* Wait again if we changed deadline */
2732                 } while ((rc == -ETIMEDOUT) &&
2733                          (req->rq_deadline > cfs_time_current_sec()));
2734
2735                 if (rc == -ETIMEDOUT) {
2736                         DEBUG_REQ(D_ERROR, req,
2737                                   "timeout on bulk %s after %ld%+lds",
2738                                   bulk2type(desc),
2739                                   req->rq_deadline - start,
2740                                   cfs_time_current_sec() -
2741                                   req->rq_deadline);
2742                         ptlrpc_abort_bulk(desc);
2743                 } else if (exp->exp_failed) {
2744                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk %s",
2745                                   bulk2type(desc));
2746                         rc = -ENOTCONN;
2747                         ptlrpc_abort_bulk(desc);
2748                 } else if (exp->exp_abort_active_req) {
2749                         DEBUG_REQ(D_ERROR, req, "Reconnect on bulk %s",
2750                                   bulk2type(desc));
2751                         /* we don't reply anyway */
2752                         rc = -ETIMEDOUT;
2753                         ptlrpc_abort_bulk(desc);
2754                 } else if (!desc->bd_success ||
2755                            desc->bd_nob_transferred != desc->bd_nob) {
2756                         DEBUG_REQ(D_ERROR, req, "%s bulk %s %d(%d)",
2757                                   desc->bd_success ?
2758                                   "truncated" : "network error on",
2759                                   bulk2type(desc),
2760                                   desc->bd_nob_transferred,
2761                                   desc->bd_nob);
2762                         /* XXX should this be a different errno? */
2763                         rc = -ETIMEDOUT;
2764                 } else if (desc->bd_type == BULK_GET_SINK) {
2765                         rc = sptlrpc_svc_unwrap_bulk(req, desc);
2766                 }
2767         } else {
2768                 DEBUG_REQ(D_ERROR, req, "bulk %s failed: rc %d",
2769                           bulk2type(desc), rc);
2770         }
2771
2772         RETURN(rc);
2773 }
2774 EXPORT_SYMBOL(target_bulk_io);
2775
2776 #endif /* HAVE_SERVER_SUPPORT */