lustre/ldlm/ldlm_lib.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License version 2 only,
  10  * as published by the Free Software Foundation.
  11  *
  12  * This program is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  15  * General Public License version 2 for more details (a copy is included
  16  * in the LICENSE file that accompanied this code).
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * version 2 along with this program; If not, see
  20  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  21  *
  22  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  23  * CA 95054 USA or visit www.sun.com if you need additional information or
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  30  * Use is subject to license terms.
  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 #ifndef EXPORT_SYMTAB
  38 # define EXPORT_SYMTAB
  39 #endif
  40 #define DEBUG_SUBSYSTEM S_LDLM
  41
  42 #ifdef __KERNEL__
  43 # include <libcfs/libcfs.h>
  44 #else
  45 # include <liblustre.h>
  46 #endif
  47 #include <obd.h>
  48 #include <lustre_mds.h>
  49 #include <lustre_dlm.h>
  50 #include <lustre_net.h>
  51 #include "ldlm_internal.h"
  52
  53 /* @priority: if non-zero, move the selected to the list head
  54  * @create: if zero, only search in existed connections
  55  */
  56 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  57                            int priority, int create)
  58 {
  59         struct ptlrpc_connection *ptlrpc_conn;
  60         struct obd_import_conn *imp_conn = NULL, *item;
  61         int rc = 0;
  62         ENTRY;
  63
  64         if (!create && !priority) {
  65                 CDEBUG(D_HA, "Nothing to do\n");
  66                 RETURN(-EINVAL);
  67         }
  68
  69         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  70         if (!ptlrpc_conn) {
  71                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  72                 RETURN (-ENOENT);
  73         }
  74
  75         if (create) {
  76                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  77                 if (!imp_conn) {
  78                         GOTO(out_put, rc = -ENOMEM);
  79                 }
  80         }
  81
  82         spin_lock(&imp->imp_lock);
  83         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  84                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  85                         if (priority) {
  86                                 list_del(&item->oic_item);
  87                                 list_add(&item->oic_item, &imp->imp_conn_list);
  88                                 item->oic_last_attempt = 0;
  89                         }
  90                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  91                                imp, imp->imp_obd->obd_name, uuid->uuid,
  92                                (priority ? ", moved to head" : ""));
  93                         spin_unlock(&imp->imp_lock);
  94                         GOTO(out_free, rc = 0);
  95                 }
  96         }
  97         /* not found */
  98         if (create) {
  99                 imp_conn->oic_conn = ptlrpc_conn;
 100                 imp_conn->oic_uuid = *uuid;
 101                 imp_conn->oic_last_attempt = 0;
 102                 if (priority)
 103                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 104                 else
 105                         list_add_tail(&imp_conn->oic_item, &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                 spin_unlock(&imp->imp_lock);
 111                 GOTO(out_free, rc = -ENOENT);
 112
 113         }
 114
 115         spin_unlock(&imp->imp_lock);
 116         RETURN(0);
 117 out_free:
 118         if (imp_conn)
 119                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 120 out_put:
 121         ptlrpc_connection_put(ptlrpc_conn);
 122         RETURN(rc);
 123 }
 124
 125 int import_set_conn_priority(struct obd_import *imp, struct obd_uuid *uuid)
 126 {
 127         return import_set_conn(imp, uuid, 1, 0);
 128 }
 129
 130 int client_import_add_conn(struct obd_import *imp, struct obd_uuid *uuid,
 131                            int priority)
 132 {
 133         return import_set_conn(imp, uuid, priority, 1);
 134 }
 135
 136 int client_import_del_conn(struct obd_import *imp, struct obd_uuid *uuid)
 137 {
 138         struct obd_import_conn *imp_conn;
 139         struct obd_export *dlmexp;
 140         int rc = -ENOENT;
 141         ENTRY;
 142
 143         spin_lock(&imp->imp_lock);
 144         if (list_empty(&imp->imp_conn_list)) {
 145                 LASSERT(!imp->imp_connection);
 146                 GOTO(out, rc);
 147         }
 148
 149         list_for_each_entry(imp_conn, &imp->imp_conn_list, oic_item) {
 150                 if (!obd_uuid_equals(uuid, &imp_conn->oic_uuid))
 151                         continue;
 152                 LASSERT(imp_conn->oic_conn);
 153
 154                 /* is current conn? */
 155                 if (imp_conn == imp->imp_conn_current) {
 156                         LASSERT(imp_conn->oic_conn == imp->imp_connection);
 157
 158                         if (imp->imp_state != LUSTRE_IMP_CLOSED &&
 159                             imp->imp_state != LUSTRE_IMP_DISCON) {
 160                                 CERROR("can't remove current connection\n");
 161                                 GOTO(out, rc = -EBUSY);
 162                         }
 163
 164                         ptlrpc_connection_put(imp->imp_connection);
 165                         imp->imp_connection = NULL;
 166
 167                         dlmexp = class_conn2export(&imp->imp_dlm_handle);
 168                         if (dlmexp && dlmexp->exp_connection) {
 169                                 LASSERT(dlmexp->exp_connection ==
 170                                         imp_conn->oic_conn);
 171                                 ptlrpc_connection_put(dlmexp->exp_connection);
 172                                 dlmexp->exp_connection = NULL;
 173                         }
 174                 }
 175
 176                 list_del(&imp_conn->oic_item);
 177                 ptlrpc_connection_put(imp_conn->oic_conn);
 178                 OBD_FREE(imp_conn, sizeof(*imp_conn));
 179                 CDEBUG(D_HA, "imp %p@%s: remove connection %s\n",
 180                        imp, imp->imp_obd->obd_name, uuid->uuid);
 181                 rc = 0;
 182                 break;
 183         }
 184 out:
 185         spin_unlock(&imp->imp_lock);
 186         if (rc == -ENOENT)
 187                 CERROR("connection %s not found\n", uuid->uuid);
 188         RETURN(rc);
 189 }
 190
 191 /* configure an RPC client OBD device
 192  *
 193  * lcfg parameters:
 194  * 1 - client UUID
 195  * 2 - server UUID
 196  * 3 - inactive-on-startup
 197  */
 198 int client_obd_setup(struct obd_device *obddev, obd_count len, void *buf)
 199 {
 200         struct lustre_cfg* lcfg = buf;
 201         struct client_obd *cli = &obddev->u.cli;
 202         struct obd_import *imp;
 203         struct obd_uuid server_uuid;
 204         int rq_portal, rp_portal, connect_op;
 205         char *name = obddev->obd_type->typ_name;
 206         int rc;
 207         ENTRY;
 208
 209         /* In a more perfect world, we would hang a ptlrpc_client off of
 210          * obd_type and just use the values from there. */
 211         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 212                 rq_portal = OST_REQUEST_PORTAL;
 213                 rp_portal = OSC_REPLY_PORTAL;
 214                 connect_op = OST_CONNECT;
 215         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 216                 rq_portal = MDS_REQUEST_PORTAL;
 217                 rp_portal = MDC_REPLY_PORTAL;
 218                 connect_op = MDS_CONNECT;
 219         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 220                 rq_portal = MGS_REQUEST_PORTAL;
 221                 rp_portal = MGC_REPLY_PORTAL;
 222                 connect_op = MGS_CONNECT;
 223         } else {
 224                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 225                        name);
 226                 RETURN(-EINVAL);
 227         }
 228
 229         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 230                 CERROR("requires a TARGET UUID\n");
 231                 RETURN(-EINVAL);
 232         }
 233
 234         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 235                 CERROR("client UUID must be less than 38 characters\n");
 236                 RETURN(-EINVAL);
 237         }
 238
 239         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 240                 CERROR("setup requires a SERVER UUID\n");
 241                 RETURN(-EINVAL);
 242         }
 243
 244         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 245                 CERROR("target UUID must be less than 38 characters\n");
 246                 RETURN(-EINVAL);
 247         }
 248
 249         init_rwsem(&cli->cl_sem);
 250         sema_init(&cli->cl_mgc_sem, 1);
 251         cli->cl_conn_count = 0;
 252         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 253                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 254                      sizeof(server_uuid)));
 255
 256         cli->cl_dirty = 0;
 257         cli->cl_avail_grant = 0;
 258         /* FIXME: should limit this for the sum of all cl_dirty_max */
 259         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 260         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 261                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 262         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 263         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 264         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 265         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 266         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 267         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 268         cli->cl_r_in_flight = 0;
 269         cli->cl_w_in_flight = 0;
 270         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 271         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 272         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 273         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 274         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 275         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 276         cfs_waitq_init(&cli->cl_destroy_waitq);
 277         atomic_set(&cli->cl_destroy_in_flight, 0);
 278 #ifdef ENABLE_CHECKSUM
 279         /* Turn on checksumming by default. */
 280         cli->cl_checksum = 1;
 281         /*
 282          * The supported checksum types will be worked out at connect time
 283          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 284          * through procfs.
 285          */
 286         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 287 #endif
 288         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 289
 290         /* This value may be changed at connect time in
 291            ptlrpc_connect_interpret. */
 292         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 293                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 294
 295         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 296                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 297         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 298                 cli->cl_max_rpcs_in_flight = 2;
 299         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 300                 cli->cl_max_rpcs_in_flight = 3;
 301         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 302                 cli->cl_max_rpcs_in_flight = 4;
 303         } else {
 304                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 305         }
 306         rc = ldlm_get_ref();
 307         if (rc) {
 308                 CERROR("ldlm_get_ref failed: %d\n", rc);
 309                 GOTO(err, rc);
 310         }
 311
 312         ptlrpc_init_client(rq_portal, rp_portal, name,
 313                            &obddev->obd_ldlm_client);
 314
 315         imp = class_new_import(obddev);
 316         if (imp == NULL)
 317                 GOTO(err_ldlm, rc = -ENOENT);
 318         imp->imp_client = &obddev->obd_ldlm_client;
 319         imp->imp_connect_op = connect_op;
 320         imp->imp_initial_recov = 1;
 321         imp->imp_initial_recov_bk = 0;
 322         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 323         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 324                LUSTRE_CFG_BUFLEN(lcfg, 1));
 325         class_import_put(imp);
 326
 327         rc = client_import_add_conn(imp, &server_uuid, 1);
 328         if (rc) {
 329                 CERROR("can't add initial connection\n");
 330                 GOTO(err_import, rc);
 331         }
 332
 333         cli->cl_import = imp;
 334         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 335         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 336         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 337
 338         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 339                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 340                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 341                                name, obddev->obd_name,
 342                                cli->cl_target_uuid.uuid);
 343                         spin_lock(&imp->imp_lock);
 344                         imp->imp_deactive = 1;
 345                         spin_unlock(&imp->imp_lock);
 346                 }
 347         }
 348
 349         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 350                                                    LDLM_NAMESPACE_CLIENT,
 351                                                    LDLM_NAMESPACE_GREEDY);
 352         if (obddev->obd_namespace == NULL) {
 353                 CERROR("Unable to create client namespace - %s\n",
 354                        obddev->obd_name);
 355                 GOTO(err_import, rc = -ENOMEM);
 356         }
 357
 358         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 359
 360         RETURN(rc);
 361
 362 err_import:
 363         class_destroy_import(imp);
 364 err_ldlm:
 365         ldlm_put_ref();
 366 err:
 367         RETURN(rc);
 368
 369 }
 370
 371 int client_obd_cleanup(struct obd_device *obddev)
 372 {
 373         ENTRY;
 374
 375         ldlm_namespace_free_post(obddev->obd_namespace);
 376         obddev->obd_namespace = NULL;
 377
 378         ldlm_put_ref();
 379         RETURN(0);
 380 }
 381
 382 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 383 int client_connect_import(struct lustre_handle *dlm_handle,
 384                           struct obd_device *obd, struct obd_uuid *cluuid,
 385                           struct obd_connect_data *data, void *localdata)
 386 {
 387         struct client_obd *cli = &obd->u.cli;
 388         struct obd_import *imp = cli->cl_import;
 389         struct obd_export **exp = localdata;
 390         struct obd_connect_data *ocd;
 391         int rc;
 392         ENTRY;
 393
 394         down_write(&cli->cl_sem);
 395         CDEBUG(D_INFO, "connect %s - %d\n", obd->obd_name,
 396                cli->cl_conn_count);
 397
 398         if (cli->cl_conn_count > 0)
 399                 GOTO(out_sem, rc = -EALREADY);
 400
 401         rc = class_connect(dlm_handle, obd, cluuid);
 402         if (rc)
 403                 GOTO(out_sem, rc);
 404
 405         cli->cl_conn_count++;
 406         *exp = class_conn2export(dlm_handle);
 407
 408         LASSERT(obd->obd_namespace);
 409
 410         imp->imp_dlm_handle = *dlm_handle;
 411         rc = ptlrpc_init_import(imp);
 412         if (rc != 0)
 413                 GOTO(out_ldlm, rc);
 414
 415         ocd = &imp->imp_connect_data;
 416         if (data) {
 417                 *ocd = *data;
 418                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 419         }
 420
 421         rc = ptlrpc_connect_import(imp, NULL);
 422         if (rc != 0) {
 423                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 424                 GOTO(out_ldlm, rc);
 425         }
 426         LASSERT((*exp)->exp_connection);
 427
 428         if (data) {
 429                 LASSERT((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 430                         ocd->ocd_connect_flags);
 431                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 432         }
 433
 434         ptlrpc_pinger_add_import(imp);
 435         EXIT;
 436
 437         if (rc) {
 438 out_ldlm:
 439                 cli->cl_conn_count--;
 440                 class_disconnect(*exp);
 441                 *exp = NULL;
 442         }
 443 out_sem:
 444         up_write(&cli->cl_sem);
 445         return rc;
 446 }
 447
 448 int client_disconnect_export(struct obd_export *exp)
 449 {
 450         struct obd_device *obd = class_exp2obd(exp);
 451         struct client_obd *cli;
 452         struct obd_import *imp;
 453         int rc = 0, err;
 454         ENTRY;
 455
 456         if (!obd) {
 457                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 458                        exp, exp ? exp->exp_handle.h_cookie : -1);
 459                 RETURN(-EINVAL);
 460         }
 461
 462         cli = &obd->u.cli;
 463         imp = cli->cl_import;
 464
 465         down_write(&cli->cl_sem);
 466         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 467                cli->cl_conn_count);
 468
 469         if (!cli->cl_conn_count) {
 470                 CERROR("disconnecting disconnected device (%s)\n",
 471                        obd->obd_name);
 472                 GOTO(out_disconnect, rc = -EINVAL);
 473         }
 474
 475         cli->cl_conn_count--;
 476         if (cli->cl_conn_count)
 477                 GOTO(out_disconnect, rc = 0);
 478
 479         /* Mark import deactivated now, so we don't try to reconnect if any
 480          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 481          * fully deactivate the import, or that would drop all requests. */
 482         spin_lock(&imp->imp_lock);
 483         imp->imp_deactive = 1;
 484         spin_unlock(&imp->imp_lock);
 485
 486         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 487          * delete it regardless.  (It's safe to delete an import that was
 488          * never added.) */
 489         (void)ptlrpc_pinger_del_import(imp);
 490
 491         if (obd->obd_namespace != NULL) {
 492                 /* obd_force == local only */
 493                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 494                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 495                                        NULL);
 496                 ldlm_namespace_free_prior(obd->obd_namespace, imp,
 497                                           obd->obd_force);
 498         }
 499
 500         rc = ptlrpc_disconnect_import(imp, 0);
 501
 502         ptlrpc_invalidate_import(imp);
 503
 504         if (imp->imp_rq_pool) {
 505                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 506                 imp->imp_rq_pool = NULL;
 507         }
 508         class_destroy_import(imp);
 509         cli->cl_import = NULL;
 510
 511         EXIT;
 512
 513  out_disconnect:
 514         /* use server style - class_disconnect should be always called for
 515          * o_disconnect */
 516         err = class_disconnect(exp);
 517         if (!rc && err)
 518                 rc = err;
 519         up_write(&cli->cl_sem);
 520
 521         RETURN(rc);
 522 }
 523
 524 int server_disconnect_export(struct obd_export *exp)
 525 {
 526         int rc;
 527         ENTRY;
 528
 529         /* Disconnect early so that clients can't keep using export */
 530         rc = class_disconnect(exp);
 531
 532         /* close import for avoid sending any requests */
 533         if (exp->exp_imp_reverse)
 534                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 535
 536         if (exp->exp_obd->obd_namespace != NULL)
 537                 ldlm_cancel_locks_for_export(exp);
 538
 539         /* complete all outstanding replies */
 540         spin_lock(&exp->exp_lock);
 541         while (!list_empty(&exp->exp_outstanding_replies)) {
 542                 struct ptlrpc_reply_state *rs =
 543                         list_entry(exp->exp_outstanding_replies.next,
 544                                    struct ptlrpc_reply_state, rs_exp_list);
 545                 struct ptlrpc_service *svc = rs->rs_service;
 546
 547                 spin_lock(&svc->srv_lock);
 548                 list_del_init(&rs->rs_exp_list);
 549                 ptlrpc_schedule_difficult_reply(rs);
 550                 spin_unlock(&svc->srv_lock);
 551         }
 552         spin_unlock(&exp->exp_lock);
 553
 554         RETURN(rc);
 555 }
 556
 557 /* --------------------------------------------------------------------------
 558  * from old lib/target.c
 559  * -------------------------------------------------------------------------- */
 560
 561 static int target_handle_reconnect(struct lustre_handle *conn,
 562                                    struct obd_export *exp,
 563                                    struct obd_uuid *cluuid)
 564 {
 565         ENTRY;
 566         if (exp->exp_connection && exp->exp_imp_reverse) {
 567                 struct lustre_handle *hdl;
 568                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 569                 /* Might be a re-connect after a partition. */
 570                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 571                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 572                               cluuid->uuid);
 573                         conn->cookie = exp->exp_handle.h_cookie;
 574                         /* target_handle_connect() treats EALREADY and
 575                          * -EALREADY differently.  EALREADY means we are
 576                          * doing a valid reconnect from the same client. */
 577                         RETURN(EALREADY);
 578                 } else {
 579                         CERROR("%s reconnecting from %s, "
 580                                "handle mismatch (ours "LPX64", theirs "
 581                                LPX64")\n", cluuid->uuid,
 582                                exp->exp_connection->c_remote_uuid.uuid,
 583                                hdl->cookie, conn->cookie);
 584                         memset(conn, 0, sizeof *conn);
 585                         /* target_handle_connect() treats EALREADY and
 586                          * -EALREADY differently.  -EALREADY is an error
 587                          * (same UUID, different handle). */
 588                         RETURN(-EALREADY);
 589                 }
 590         }
 591
 592         conn->cookie = exp->exp_handle.h_cookie;
 593         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 594                cluuid->uuid, exp, conn->cookie);
 595         RETURN(0);
 596 }
 597
 598 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 599                           int error)
 600 {
 601         struct obd_export *exp = cb_data;
 602
 603         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 604                obd->obd_name, exp->exp_client_uuid.uuid);
 605
 606         spin_lock(&exp->exp_lock);
 607         exp->exp_need_sync = 0;
 608         spin_unlock(&exp->exp_lock);
 609 }
 610 EXPORT_SYMBOL(target_client_add_cb);
 611
 612 static void
 613 target_start_and_reset_recovery_timer(struct obd_device *obd,
 614                                       svc_handler_t handler,
 615                                       struct ptlrpc_request *req,
 616                                       int new_client);
 617 void target_stop_recovery(void *, int);
 618 static void reset_recovery_timer(struct obd_device *obd, int duration,
 619                                  int extend);
 620 int target_recovery_check_and_stop(struct obd_device *obd)
 621 {
 622         int abort_recovery = 0;
 623
 624         if (obd->obd_stopping || !obd->obd_recovering)
 625                 return 1;
 626
 627         spin_lock_bh(&obd->obd_processing_task_lock);
 628         abort_recovery = obd->obd_abort_recovery;
 629         obd->obd_abort_recovery = 0;
 630         spin_unlock_bh(&obd->obd_processing_task_lock);
 631         if (!abort_recovery)
 632                 return 0;
 633         /** check if fs version-capable */
 634         if (target_fs_version_capable(obd)) {
 635                 class_handle_stale_exports(obd);
 636         } else {
 637                 CWARN("Versions are not supported by ldiskfs, VBR is OFF\n");
 638                 class_disconnect_stale_exports(obd, exp_flags_from_obd(obd));
 639         }
 640         /* VBR: no clients are remained to replay, stop recovery */
 641         spin_lock_bh(&obd->obd_processing_task_lock);
 642         if (obd->obd_recovering && obd->obd_recoverable_clients == 0) {
 643                 spin_unlock_bh(&obd->obd_processing_task_lock);
 644                 target_stop_recovery(obd, 0);
 645                 return 1;
 646         }
 647         /* always check versions now */
 648         obd->obd_version_recov = 1;
 649         cfs_waitq_signal(&obd->obd_next_transno_waitq);
 650         spin_unlock_bh(&obd->obd_processing_task_lock);
 651         /* reset timer, recovery will proceed with versions now */
 652         reset_recovery_timer(obd, OBD_RECOVERY_TIME_SOFT, 1);
 653         return 0;
 654 }
 655 EXPORT_SYMBOL(target_recovery_check_and_stop);
 656
 657 int target_handle_connect(struct ptlrpc_request *req, svc_handler_t handler)
 658 {
 659         struct obd_device *target, *targref = NULL;
 660         struct obd_export *export = NULL;
 661         struct obd_import *revimp;
 662         struct lustre_handle conn;
 663         struct obd_uuid tgtuuid;
 664         struct obd_uuid cluuid;
 665         struct obd_uuid remote_uuid;
 666         char *str, *tmp;
 667         int rc = 0;
 668         struct obd_connect_data *data;
 669         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*data) };
 670         lnet_nid_t *client_nid = NULL;
 671         int mds_conn = 0;
 672         ENTRY;
 673
 674         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 675
 676         lustre_set_req_swabbed(req, REQ_REC_OFF);
 677         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF, sizeof(tgtuuid)-1);
 678         if (str == NULL) {
 679                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 680                 GOTO(out, rc = -EINVAL);
 681         }
 682
 683         obd_str2uuid (&tgtuuid, str);
 684         target = class_uuid2obd(&tgtuuid);
 685         /* COMPAT_146 */
 686         /* old (pre 1.6) lustre_process_log tries to connect to mdsname
 687            (eg. mdsA) instead of uuid. */
 688         if (!target) {
 689                 snprintf((char *)tgtuuid.uuid, sizeof(tgtuuid), "%s_UUID", str);
 690                 target = class_uuid2obd(&tgtuuid);
 691         }
 692         if (!target)
 693                 target = class_name2obd(str);
 694         /* end COMPAT_146 */
 695
 696         if (!target || target->obd_stopping || !target->obd_set_up) {
 697                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 698                                    " for connect (%s)\n", str,
 699                                    !target ? "no target" :
 700                                    (target->obd_stopping ? "stopping" :
 701                                    "not set up"));
 702                 GOTO(out, rc = -ENODEV);
 703         }
 704
 705         if (target->obd_no_conn) {
 706                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 707                               "from %s\n", target->obd_name,
 708                               libcfs_nid2str(req->rq_peer.nid));
 709                 GOTO(out, rc = -EAGAIN);
 710         }
 711
 712         /* Make sure the target isn't cleaned up while we're here. Yes,
 713            there's still a race between the above check and our incref here.
 714            Really, class_uuid2obd should take the ref. */
 715         targref = class_incref(target);
 716
 717         lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
 718         str = lustre_msg_string(req->rq_reqmsg, REQ_REC_OFF + 1,
 719                                 sizeof(cluuid) - 1);
 720         if (str == NULL) {
 721                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 722                 GOTO(out, rc = -EINVAL);
 723         }
 724
 725         obd_str2uuid (&cluuid, str);
 726
 727         /* XXX extract a nettype and format accordingly */
 728         switch (sizeof(lnet_nid_t)) {
 729                 /* NB the casts only avoid compiler warnings */
 730         case 8:
 731                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 732                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 733                 break;
 734         case 4:
 735                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 736                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 737                 break;
 738         default:
 739                 LBUG();
 740         }
 741
 742         target_recovery_check_and_stop(target);
 743
 744         tmp = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn);
 745         if (tmp == NULL)
 746                 GOTO(out, rc = -EPROTO);
 747
 748         memcpy(&conn, tmp, sizeof conn);
 749
 750         data = lustre_swab_reqbuf(req, REQ_REC_OFF + 3, sizeof(*data),
 751                                   lustre_swab_connect);
 752         rc = lustre_pack_reply(req, 2, size, NULL);
 753         if (rc)
 754                 GOTO(out, rc);
 755
 756         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 757                 if (!data) {
 758                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 759                                   "libclient connection attempt");
 760                         GOTO(out, rc = -EPROTO);
 761                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 762                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 763                            data->ocd_version > LUSTRE_VERSION_CODE +
 764                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 765                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 766                                   "libclient connection attempt",
 767                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 768                                   "old" : "new",
 769                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 770                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 771                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 772                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 773                         data = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 774                                               offsetof(typeof(*data),
 775                                                        ocd_version) +
 776                                               sizeof(data->ocd_version));
 777                         if (data) {
 778                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 779                                 data->ocd_version = LUSTRE_VERSION_CODE;
 780                         }
 781                         GOTO(out, rc = -EPROTO);
 782                 }
 783         }
 784
 785         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 786             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 787                 mds_conn = 1;
 788
 789         /* lctl gets a backstage, all-access pass. */
 790         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 791                 goto dont_check_exports;
 792
 793         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 794         if (!export)
 795                 goto no_export;
 796
 797         /* we've found an export in the hash */
 798         if (export->exp_connecting) {
 799                 /* bug 9635, et. al. */
 800                 CWARN("%s: exp %p already connecting\n",
 801                       export->exp_obd->obd_name, export);
 802                 class_export_put(export);
 803                 export = NULL;
 804                 rc = -EALREADY;
 805         } else if (mds_conn && export->exp_connection) {
 806                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 807                         /* mds reconnected after failover */
 808                         CWARN("%s: received MDS connection from NID %s,"
 809                               " removing former export from NID %s\n",
 810                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 811                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 812                 else
 813                         /* new mds connection from the same nid */
 814                         CWARN("%s: received new MDS connection from NID %s,"
 815                               " removing former export from same NID\n",
 816                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 817                 class_fail_export(export);
 818                 class_export_put(export);
 819                 export = NULL;
 820                 rc = 0;
 821         } else if (export->exp_connection &&
 822                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 823                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 824                     MSG_CONNECT_INITIAL)) {
 825                 CWARN("%s: cookie %s seen on new NID %s when "
 826                       "existing NID %s is already connected\n",
 827                       target->obd_name, cluuid.uuid,
 828                       libcfs_nid2str(req->rq_peer.nid),
 829                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 830                 rc = -EALREADY;
 831                 class_export_put(export);
 832                 export = NULL;
 833         } else if (export->exp_failed) { /* bug 11327 */
 834                 CDEBUG(D_HA, "%s: exp %p evict in progress - new cookie needed "
 835                       "for connect\n", export->exp_obd->obd_name, export);
 836                 class_export_put(export);
 837                 export = NULL;
 838                 rc = -ENODEV;
 839         } else if (export->exp_delayed &&
 840                    !(data && data->ocd_connect_flags & OBD_CONNECT_VBR)) {
 841                 class_fail_export(export);
 842                 class_export_put(export);
 843                 export = NULL;
 844                 GOTO(out, rc = -ENODEV);
 845         } else {
 846                 spin_lock(&export->exp_lock);
 847                 export->exp_connecting = 1;
 848                 spin_unlock(&export->exp_lock);
 849                 class_export_put(export);
 850                 LASSERT(export->exp_obd == target);
 851
 852                 rc = target_handle_reconnect(&conn, export, &cluuid);
 853         }
 854
 855         /* If we found an export, we already unlocked. */
 856         if (!export) {
 857 no_export:
 858                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 859         } else if (req->rq_export == NULL &&
 860                    atomic_read(&export->exp_rpc_count) > 0) {
 861                 CWARN("%s: refuse connection from %s/%s to 0x%p; still busy "
 862                       "with %d references\n", target->obd_name, cluuid.uuid,
 863                       libcfs_nid2str(req->rq_peer.nid),
 864                       export, atomic_read(&export->exp_refcount));
 865                 GOTO(out, rc = -EBUSY);
 866         } else if (req->rq_export != NULL &&
 867                    atomic_read(&export->exp_rpc_count) > 1) {
 868                 /* the current connect rpc has increased exp_rpc_count */
 869                 CWARN("%s: refuse reconnection from %s@%s to 0x%p; still busy "
 870                       "with %d active RPCs\n", target->obd_name, cluuid.uuid,
 871                       libcfs_nid2str(req->rq_peer.nid),
 872                       export, atomic_read(&export->exp_rpc_count) - 1);
 873                 spin_lock(&export->exp_lock);
 874                 if (req->rq_export->exp_conn_cnt <
 875                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 876                         /* try to abort active requests */
 877                         req->rq_export->exp_abort_active_req = 1;
 878                 spin_unlock(&export->exp_lock);
 879                 GOTO(out, rc = -EBUSY);
 880         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 881                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 882                        "cookies not random?\n", target->obd_name,
 883                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 884                 GOTO(out, rc = -EALREADY);
 885         } else if (export->exp_delayed && target->obd_recovering) {
 886                 /* VBR: don't allow delayed connection during recovery */
 887                 CWARN("%s: NID %s (%s) export was already marked as delayed "
 888                       "and will wait for end of recovery\n", target->obd_name,
 889                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 890                 GOTO(out, rc = -EBUSY);
 891         } else {
 892                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 893         }
 894
 895         if (rc < 0)
 896                 GOTO(out, rc);
 897
 898         /* Tell the client if we're in recovery. */
 899         if (target->obd_recovering) {
 900                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 901                 /* If this is the first time a client connects,
 902                    reset the recovery timer */
 903                 if (rc == 0)
 904                         target_start_and_reset_recovery_timer(target, handler,
 905                                                               req, !export);
 906         }
 907
 908         /* We want to handle EALREADY but *not* -EALREADY from
 909          * target_handle_reconnect(), return reconnection state in a flag */
 910         if (rc == EALREADY) {
 911                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 912                 rc = 0;
 913         } else {
 914                 LASSERT(rc == 0);
 915         }
 916
 917         /* Tell the client if we support replayable requests */
 918         if (target->obd_replayable)
 919                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 920         client_nid = &req->rq_peer.nid;
 921
 922         /* VBR: for delayed connections we start recovery */
 923         if (export && export->exp_delayed && !export->exp_in_recovery) {
 924                 LASSERT(!target->obd_recovering);
 925                 LASSERT(data && data->ocd_connect_flags & OBD_CONNECT_VBR);
 926                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_DELAYED |
 927                                         MSG_CONNECT_RECOVERING);
 928                 spin_lock_bh(&target->obd_processing_task_lock);
 929                 target->obd_version_recov = 1;
 930                 spin_unlock_bh(&target->obd_processing_task_lock);
 931                 target_start_and_reset_recovery_timer(target, handler, req, 1);
 932         }
 933
 934         if (export == NULL) {
 935                 if (target->obd_recovering) {
 936                         CERROR("%s: denying connection for new client %s (%s): "
 937                                "%d clients in recovery for %lds\n",
 938                                target->obd_name,
 939                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 940                                target->obd_recoverable_clients,
 941                                cfs_duration_sec(cfs_time_sub(cfs_timer_deadline(&target->obd_recovery_timer),
 942                                                              cfs_time_current())));
 943                         rc = -EBUSY;
 944                 } else {
 945  dont_check_exports:
 946                         rc = obd_connect(&conn, target, &cluuid, data,
 947                                          client_nid);
 948                 }
 949         } else {
 950                 rc = obd_reconnect(export, target, &cluuid, data, client_nid);
 951         }
 952
 953         if (rc)
 954                 GOTO(out, rc);
 955
 956         /* Return only the parts of obd_connect_data that we understand, so the
 957          * client knows that we don't understand the rest. */
 958         if (data)
 959                 memcpy(lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 960                                       sizeof(*data)),
 961                        data, sizeof(*data));
 962
 963         /* If all else goes well, this is our RPC return code. */
 964         req->rq_status = 0;
 965
 966         lustre_msg_set_handle(req->rq_repmsg, &conn);
 967
 968         /* ownership of this export ref transfers to the request AFTER we
 969          * drop any previous reference the request had, but we don't want
 970          * that to go to zero before we get our new export reference. */
 971         export = class_conn2export(&conn);
 972         if (!export) {
 973                 DEBUG_REQ(D_ERROR, req, "Missing export!");
 974                 GOTO(out, rc = -ENODEV);
 975         }
 976
 977         /* If the client and the server are the same node, we will already
 978          * have an export that really points to the client's DLM export,
 979          * because we have a shared handles table.
 980          *
 981          * XXX this will go away when shaver stops sending the "connect" handle
 982          * in the real "remote handle" field of the request --phik 24 Apr 2003
 983          */
 984         if (req->rq_export != NULL)
 985                 class_export_put(req->rq_export);
 986
 987         req->rq_export = export;
 988
 989         spin_lock(&export->exp_lock);
 990         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 991                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 992                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 993                        export->exp_conn_cnt,
 994                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 995
 996                 spin_unlock(&export->exp_lock);
 997                 GOTO(out, rc = -EALREADY);
 998         }
 999         LASSERT(lustre_msg_get_conn_cnt(req->rq_reqmsg) > 0);
1000         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
1001         export->exp_abort_active_req = 0;
1002
1003         /* request from liblustre?  Don't evict it for not pinging. */
1004         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
1005                 export->exp_libclient = 1;
1006                 spin_unlock(&export->exp_lock);
1007
1008                 spin_lock(&target->obd_dev_lock);
1009                 list_del_init(&export->exp_obd_chain_timed);
1010                 spin_unlock(&target->obd_dev_lock);
1011         } else {
1012                 spin_unlock(&export->exp_lock);
1013         }
1014
1015         if (export->exp_connection != NULL) {
1016                 /* Check to see if connection came from another NID */
1017                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
1018                     !hlist_unhashed(&export->exp_nid_hash))
1019                         lustre_hash_del(export->exp_obd->obd_nid_hash,
1020                                         &export->exp_connection->c_peer.nid,
1021                                         &export->exp_nid_hash);
1022
1023                 ptlrpc_connection_put(export->exp_connection);
1024         }
1025
1026         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
1027                                                        req->rq_self,
1028                                                        &remote_uuid);
1029
1030         if (hlist_unhashed(&export->exp_nid_hash)) {
1031                 lustre_hash_add(export->exp_obd->obd_nid_hash,
1032                                 &export->exp_connection->c_peer.nid,
1033                                 &export->exp_nid_hash);
1034         }
1035
1036         if (lustre_msg_get_op_flags(req->rq_repmsg) & MSG_CONNECT_RECONNECT) {
1037                 revimp = class_import_get(export->exp_imp_reverse);
1038                 ptlrpc_connection_put(revimp->imp_connection);
1039                 revimp->imp_connection = NULL;
1040                 GOTO(set_flags, rc = 0);
1041         }
1042
1043         if (target->obd_recovering && !export->exp_in_recovery) {
1044                 spin_lock(&export->exp_lock);
1045                 export->exp_in_recovery = 1;
1046                 spin_unlock(&export->exp_lock);
1047                 target->obd_connected_clients++;
1048         }
1049         memcpy(&conn,
1050                lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 2, sizeof conn),
1051                sizeof conn);
1052
1053         if (export->exp_imp_reverse != NULL)
1054                 class_destroy_import(export->exp_imp_reverse);
1055         revimp = export->exp_imp_reverse = class_new_import(target);
1056         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1057         revimp->imp_remote_handle = conn;
1058         revimp->imp_dlm_fake = 1;
1059         revimp->imp_state = LUSTRE_IMP_FULL;
1060
1061 set_flags:
1062         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1063         if (req->rq_reqmsg->lm_magic == LUSTRE_MSG_MAGIC_V1 &&
1064             lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_NEXT_VER) {
1065                 revimp->imp_msg_magic = LUSTRE_MSG_MAGIC_V2;
1066                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_NEXT_VER);
1067         } else {
1068                 /* unknown versions will be caught in
1069                  * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1070                 revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1071         }
1072
1073         if (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1) {
1074                 if (export->exp_connect_flags & OBD_CONNECT_AT)
1075                         revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1076                 else
1077                         revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1078         }
1079
1080         class_import_put(revimp);
1081 out:
1082         if (export) {
1083                 spin_lock(&export->exp_lock);
1084                 export->exp_connecting = 0;
1085                 spin_unlock(&export->exp_lock);
1086         }
1087         if (targref)
1088                 class_decref(targref);
1089         if (rc)
1090                 req->rq_status = rc;
1091         RETURN(rc);
1092 }
1093
1094 int target_handle_disconnect(struct ptlrpc_request *req)
1095 {
1096         int rc;
1097         ENTRY;
1098
1099         rc = lustre_pack_reply(req, 1, NULL, NULL);
1100         if (rc)
1101                 RETURN(rc);
1102
1103         /* keep the rq_export around so we can send the reply */
1104         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1105         RETURN(0);
1106 }
1107
1108 void target_destroy_export(struct obd_export *exp)
1109 {
1110         /* exports created from last_rcvd data, and "fake"
1111            exports created by lctl don't have an import */
1112         if (exp->exp_imp_reverse != NULL)
1113                 class_destroy_import(exp->exp_imp_reverse);
1114
1115         /* We cancel locks at disconnect time, but this will catch any locks
1116          * granted in a race with recovery-induced disconnect. */
1117         if (exp->exp_obd->obd_namespace != NULL)
1118                 ldlm_cancel_locks_for_export(exp);
1119 }
1120
1121 /*
1122  * Recovery functions
1123  */
1124
1125 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1126 {
1127         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1128         struct obd_export     *exp = req->rq_export;
1129         struct ptlrpc_request *reqiter;
1130         int                    dup = 0;
1131
1132         LASSERT(exp);
1133
1134         spin_lock(&exp->exp_lock);
1135         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1136                             rq_replay_list) {
1137                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1138                         dup = 1;
1139                         break;
1140                 }
1141         }
1142
1143         if (dup) {
1144                 /* we expect it with RESENT and REPLAY flags */
1145                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1146                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1147                         CERROR("invalid flags %x of resent replay\n",
1148                                lustre_msg_get_flags(req->rq_reqmsg));
1149         } else {
1150                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1151         }
1152
1153         spin_unlock(&exp->exp_lock);
1154         return dup;
1155 }
1156
1157 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1158 {
1159         LASSERT(!list_empty(&req->rq_replay_list));
1160         LASSERT(req->rq_export);
1161
1162         spin_lock(&req->rq_export->exp_lock);
1163         list_del_init(&req->rq_replay_list);
1164         spin_unlock(&req->rq_export->exp_lock);
1165 }
1166
1167 static void target_request_copy_get(struct ptlrpc_request *req)
1168 {
1169         /* mark that request is in recovery queue, so request handler will not
1170          * drop rpc count in export, bug 19870*/
1171         LASSERT(!req->rq_copy_queued);
1172         spin_lock(&req->rq_lock);
1173         req->rq_copy_queued = 1;
1174         spin_unlock(&req->rq_lock);
1175         /* increase refcount to keep request in queue */
1176         atomic_inc(&req->rq_refcount);
1177         /* release service thread while request is queued
1178          * we are moving the request from active processing
1179          * to waiting on the replay queue */
1180         ptlrpc_server_active_request_dec(req);
1181 }
1182
1183 static void target_request_copy_put(struct ptlrpc_request *req)
1184 {
1185         LASSERTF(list_empty(&req->rq_replay_list), "next: %p, prev: %p\n",
1186                  req->rq_replay_list.next, req->rq_replay_list.prev);
1187         /* class_export_rpc_get was done before handling request,
1188          * drop it early to allow new requests, see bug 19870.
1189          */
1190         LASSERT(req->rq_copy_queued);
1191         class_export_rpc_put(req->rq_export);
1192         /* ptlrpc_server_drop_request() assumes the request is active */
1193         ptlrpc_server_active_request_inc(req);
1194         ptlrpc_server_drop_request(req);
1195 }
1196
1197 static void target_send_delayed_replies(struct obd_device *obd)
1198 {
1199         int max_clients = obd->obd_max_recoverable_clients;
1200         struct ptlrpc_request *req, *tmp;
1201         time_t elapsed_time = max_t(time_t, 1, cfs_time_current_sec() -
1202                                     obd->obd_recovery_start);
1203
1204         LCONSOLE_INFO("%s: Recovery period over after %d:%.02d, of %d clients "
1205                       "%d recovered and %d %s evicted.\n", obd->obd_name,
1206                       (int)elapsed_time/60, (int)elapsed_time%60, max_clients,
1207                       obd->obd_connected_clients,
1208                       obd->obd_stale_clients,
1209                       obd->obd_stale_clients == 1 ? "was" : "were");
1210
1211         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1212                       obd->obd_name);
1213
1214         list_for_each_entry_safe(req, tmp, &obd->obd_delayed_reply_queue,
1215                                  rq_list) {
1216                 list_del_init(&req->rq_list);
1217                 DEBUG_REQ(D_HA, req, "delayed:");
1218                 ptlrpc_reply(req);
1219                 target_request_copy_put(req);
1220         }
1221         obd->obd_recovery_end = cfs_time_current_sec();
1222 }
1223
1224 static void target_finish_recovery(struct obd_device *obd)
1225 {
1226         OBD_RACE(OBD_FAIL_TGT_REPLAY_DELAY);
1227
1228         ldlm_reprocess_all_ns(obd->obd_namespace);
1229         spin_lock_bh(&obd->obd_processing_task_lock);
1230         if (list_empty(&obd->obd_recovery_queue)) {
1231                 obd->obd_recovery_thread = NULL;
1232                 obd->obd_processing_task = 0;
1233         } else {
1234                 spin_unlock_bh(&obd->obd_processing_task_lock);
1235                 CERROR("%s: Recovery queue isn't empty\n", obd->obd_name);
1236                 LBUG();
1237         }
1238         spin_unlock_bh(&obd->obd_processing_task_lock);
1239                 ;
1240         /* when recovery finished, cleanup orphans on mds and ost */
1241         if (OBT(obd) && OBP(obd, postrecov)) {
1242                 int rc = OBP(obd, postrecov)(obd);
1243                 if (rc < 0)
1244                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1245                                       obd->obd_name, rc);
1246         }
1247         target_send_delayed_replies(obd);
1248 }
1249
1250 static void abort_recovery_queue(struct obd_device *obd)
1251 {
1252         struct ptlrpc_request *req, *n;
1253         struct list_head abort_list;
1254         int rc;
1255
1256         CFS_INIT_LIST_HEAD(&abort_list);
1257         spin_lock_bh(&obd->obd_processing_task_lock);
1258         list_splice_init(&obd->obd_recovery_queue, &abort_list);
1259         spin_unlock_bh(&obd->obd_processing_task_lock);
1260         /* process abort list unlocked */
1261         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1262                 target_exp_dequeue_req_replay(req);
1263                 list_del_init(&req->rq_list);
1264                 DEBUG_REQ(D_ERROR, req, "%s: aborted:", obd->obd_name);
1265                 req->rq_status = -ENOTCONN;
1266                 req->rq_type = PTL_RPC_MSG_ERR;
1267                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1268                 if (rc == 0)
1269                         ptlrpc_reply(req);
1270                 else
1271                         DEBUG_REQ(D_ERROR, req,
1272                                   "packing failed for abort-reply; skipping");
1273                 target_request_copy_put(req);
1274         }
1275 }
1276
1277 /* Called from a cleanup function if the device is being cleaned up
1278    forcefully.  The exports should all have been disconnected already,
1279    the only thing left to do is
1280      - clear the recovery flags
1281      - cancel the timer
1282      - free queued requests and replies, but don't send replies
1283    Because the obd_stopping flag is set, no new requests should be received.
1284
1285 */
1286 void target_cleanup_recovery(struct obd_device *obd)
1287 {
1288         struct list_head *tmp, *n;
1289         struct ptlrpc_request *req;
1290         struct list_head clean_list;
1291         ENTRY;
1292
1293         LASSERT(obd->obd_stopping);
1294
1295         spin_lock_bh(&obd->obd_processing_task_lock);
1296         if (!obd->obd_recovering) {
1297                 spin_unlock_bh(&obd->obd_processing_task_lock);
1298                 EXIT;
1299                 return;
1300         }
1301         obd->obd_recovering = obd->obd_abort_recovery = 0;
1302         target_cancel_recovery_timer(obd);
1303         spin_unlock_bh(&obd->obd_processing_task_lock);
1304
1305         list_for_each_safe(tmp, n, &obd->obd_delayed_reply_queue) {
1306                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1307                 list_del(&req->rq_list);
1308                 target_request_copy_put(req);
1309         }
1310
1311         CFS_INIT_LIST_HEAD(&clean_list);
1312         spin_lock_bh(&obd->obd_processing_task_lock);
1313         list_splice_init(&obd->obd_recovery_queue, &clean_list);
1314         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1315         spin_unlock_bh(&obd->obd_processing_task_lock);
1316         list_for_each_safe(tmp, n, &clean_list) {
1317                 req = list_entry(tmp, struct ptlrpc_request, rq_list);
1318                 target_exp_dequeue_req_replay(req);
1319                 list_del_init(&req->rq_list);
1320                 target_request_copy_put(req);
1321         }
1322         EXIT;
1323 }
1324
1325 void target_stop_recovery(void *data, int abort)
1326 {
1327         struct obd_device *obd = data;
1328         enum obd_option flags;
1329         ENTRY;
1330
1331         spin_lock_bh(&obd->obd_processing_task_lock);
1332         if (!obd->obd_recovering) {
1333                 spin_unlock_bh(&obd->obd_processing_task_lock);
1334                 EXIT;
1335                 return;
1336         }
1337         flags = exp_flags_from_obd(obd) | OBD_OPT_ABORT_RECOV;
1338         obd->obd_recovering = 0;
1339         obd->obd_abort_recovery = 0;
1340         obd->obd_processing_task = 0;
1341         if (abort == 0)
1342                 LASSERT(obd->obd_recoverable_clients == 0);
1343
1344         target_cancel_recovery_timer(obd);
1345         spin_unlock_bh(&obd->obd_processing_task_lock);
1346
1347         if (abort) {
1348                 LCONSOLE_WARN("%s: recovery is aborted by administrative "
1349                               "request; %d clients are not recovered "
1350                               "(%d clients did)\n", obd->obd_name,
1351                               obd->obd_recoverable_clients,
1352                               obd->obd_connected_clients);
1353                 class_disconnect_stale_exports(obd, flags);
1354         }
1355         abort_recovery_queue(obd);
1356         target_finish_recovery(obd);
1357         CDEBUG(D_HA, "%s: recovery complete\n", obd_uuid2str(&obd->obd_uuid));
1358         EXIT;
1359 }
1360
1361 void target_abort_recovery(void *data)
1362 {
1363         target_stop_recovery(data, 1);
1364 }
1365
1366 static void reset_recovery_timer(struct obd_device *, int, int);
1367 static void target_recovery_expired(unsigned long castmeharder)
1368 {
1369         struct obd_device *obd = (struct obd_device *)castmeharder;
1370         CDEBUG(D_HA, "%s: recovery period over; %d clients never reconnected "
1371                "after %lds (%d clients did)\n", obd->obd_name,
1372                obd->obd_recoverable_clients,
1373                cfs_time_current_sec() - obd->obd_recovery_start,
1374                obd->obd_connected_clients);
1375
1376         spin_lock_bh(&obd->obd_processing_task_lock);
1377         obd->obd_abort_recovery = 1;
1378         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1379         spin_unlock_bh(&obd->obd_processing_task_lock);
1380
1381         /* bug 18948:
1382          * The recovery timer expired and target_check_and_stop_recovery()
1383          * must be called.  We cannot call it directly because we are in
1384          * interrupt context, so we need to wake up another thread to call it.
1385          * This may happen if there are obd->obd_next_transno_waitq waiters,
1386          * or if we happen to handle a connect request.  However, we cannot
1387          * count on either of those things so we wake up the ping evictor
1388          * and leverage it's context to complete recovery.
1389          *
1390          * Note: HEAD has a separate recovery thread and handle this.
1391          */
1392         spin_lock(&obd->obd_dev_lock);
1393         ping_evictor_wake(obd->obd_self_export);
1394         spin_unlock(&obd->obd_dev_lock);
1395 }
1396
1397 /* obd_processing_task_lock should be held */
1398 void target_cancel_recovery_timer(struct obd_device *obd)
1399 {
1400         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1401         cfs_timer_disarm(&obd->obd_recovery_timer);
1402 }
1403
1404 /* extend = 1 means require at least "duration" seconds left in the timer,
1405    extend = 0 means set the total duration (start_recovery_timer) */
1406 static void reset_recovery_timer(struct obd_device *obd, int duration,
1407                                  int extend)
1408 {
1409         cfs_time_t now = cfs_time_current_sec();
1410         cfs_duration_t left;
1411
1412         spin_lock_bh(&obd->obd_processing_task_lock);
1413         if (!obd->obd_recovering) {
1414                 spin_unlock_bh(&obd->obd_processing_task_lock);
1415                 return;
1416         }
1417
1418         left = cfs_time_sub(obd->obd_recovery_end, now);
1419
1420         if (extend && (duration > left))
1421                 obd->obd_recovery_timeout += duration - left;
1422         else if (!extend && (duration > obd->obd_recovery_timeout))
1423                 /* Track the client's largest expected replay time */
1424                 obd->obd_recovery_timeout = duration;
1425
1426         /* Hard limit of obd_recovery_time_hard which should not happen */
1427         if(obd->obd_recovery_timeout > obd->obd_recovery_time_hard)
1428                 obd->obd_recovery_timeout = obd->obd_recovery_time_hard;
1429
1430         obd->obd_recovery_end = obd->obd_recovery_start +
1431                                 obd->obd_recovery_timeout;
1432         if (cfs_time_before(now, obd->obd_recovery_end)) {
1433                 left = cfs_time_sub(obd->obd_recovery_end, now);
1434                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1435         }
1436         spin_unlock_bh(&obd->obd_processing_task_lock);
1437         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1438                obd->obd_name, (unsigned)left);
1439 }
1440
1441 static void check_and_start_recovery_timer(struct obd_device *obd,
1442                                            svc_handler_t handler)
1443 {
1444         spin_lock_bh(&obd->obd_processing_task_lock);
1445         if (obd->obd_recovery_handler) {
1446                 spin_unlock_bh(&obd->obd_processing_task_lock);
1447                 return;
1448         }
1449         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1450         obd->obd_recovery_start = cfs_time_current_sec();
1451         obd->obd_recovery_handler = handler;
1452         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1453         spin_unlock_bh(&obd->obd_processing_task_lock);
1454
1455         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1456 }
1457
1458 /* Reset the timer with each new client connection */
1459 /*
1460  * This timer is actually reconnect_timer, which is for making sure
1461  * the total recovery window is at least as big as my reconnect
1462  * attempt timing. So the initial recovery time_out will be set to
1463  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1464  * from client is bigger than this, then the recovery time_out will
1465  * be extend to make sure the client could be reconnected, in the
1466  * process, the timeout from the new client should be ignored.
1467  */
1468
1469 static void
1470 target_start_and_reset_recovery_timer(struct obd_device *obd,
1471                                       svc_handler_t handler,
1472                                       struct ptlrpc_request *req,
1473                                       int new_client)
1474 {
1475         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1476
1477         if (!new_client && service_time)
1478                 /* Teach server about old server's estimates, as first guess
1479                    at how long new requests will take. */
1480                 at_measured(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1481                             service_time);
1482
1483         check_and_start_recovery_timer(obd, handler);
1484
1485         /* convert the service time to rpc timeout,
1486          * reuse service_time to limit stack usage */
1487         service_time = at_est2timeout(service_time);
1488
1489         /* We expect other clients to timeout within service_time, then try
1490          * to reconnect, then try the failover server.  The max delay between
1491          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1492         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1493                              INITIAL_CONNECT_TIMEOUT);
1494         if (service_time > obd->obd_recovery_timeout && !new_client)
1495                 reset_recovery_timer(obd, service_time, 0);
1496 }
1497
1498 static int check_for_next_transno(struct obd_device *obd)
1499 {
1500         struct ptlrpc_request *req;
1501         int wake_up = 0, connected, completed, queue_len, max;
1502         __u64 next_transno, req_transno;
1503
1504         if (obd->obd_stopping) {
1505                 CDEBUG(D_HA, "waking for stopping device\n");
1506                 return 1;
1507         }
1508
1509         spin_lock_bh(&obd->obd_processing_task_lock);
1510         if (obd->obd_abort_recovery) {
1511                 CDEBUG(D_HA, "waking for aborted recovery\n");
1512                 spin_unlock_bh(&obd->obd_processing_task_lock);
1513                 return 1;
1514         } else if (!obd->obd_recovering) {
1515                 CDEBUG(D_HA, "waking for completed recovery (?)\n");
1516                 spin_unlock_bh(&obd->obd_processing_task_lock);
1517                 return 1;
1518         }
1519
1520         LASSERT(!list_empty(&obd->obd_recovery_queue));
1521         req = list_entry(obd->obd_recovery_queue.next,
1522                          struct ptlrpc_request, rq_list);
1523         max = obd->obd_max_recoverable_clients;
1524         req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1525         connected = obd->obd_connected_clients;
1526         completed = max - obd->obd_recoverable_clients -
1527                     obd->obd_delayed_clients;
1528         queue_len = obd->obd_requests_queued_for_recovery;
1529         next_transno = obd->obd_next_recovery_transno;
1530
1531         CDEBUG(D_HA,"max: %d, connected: %d, delayed %d, completed: %d, "
1532                "queue_len: %d, req_transno: "LPU64", next_transno: "LPU64"\n",
1533                max, connected, obd->obd_delayed_clients, completed, queue_len,
1534                req_transno, next_transno);
1535         if (req_transno == next_transno) {
1536                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1537                 wake_up = 1;
1538         } else if (queue_len == obd->obd_recoverable_clients) {
1539                 CDEBUG(D_ERROR,
1540                        "%s: waking for skipped transno (skip: "LPD64
1541                        ", ql: %d, comp: %d, conn: %d, next: "LPD64")\n",
1542                        obd->obd_name, next_transno, queue_len, completed, max,
1543                        req_transno);
1544                 obd->obd_next_recovery_transno = req_transno;
1545                 wake_up = 1;
1546         }
1547         spin_unlock_bh(&obd->obd_processing_task_lock);
1548         LASSERT(lustre_msg_get_transno(req->rq_reqmsg) >= next_transno);
1549         return wake_up;
1550 }
1551
1552 static void process_recovery_queue(struct obd_device *obd)
1553 {
1554         struct ptlrpc_request *req;
1555         struct l_wait_info lwi = { 0 };
1556         ENTRY;
1557
1558         for (;;) {
1559                 spin_lock_bh(&obd->obd_processing_task_lock);
1560
1561                 if (!obd->obd_recovering) {
1562                         spin_unlock_bh(&obd->obd_processing_task_lock);
1563                         EXIT;
1564                         return;
1565                 }
1566
1567                 LASSERTF(obd->obd_processing_task == cfs_curproc_pid(),
1568                          "%s: invalid pid in obd_processing_task (%d != %d)\n",
1569                          obd->obd_name, obd->obd_processing_task,
1570                          cfs_curproc_pid());
1571                 req = list_entry(obd->obd_recovery_queue.next,
1572                                  struct ptlrpc_request, rq_list);
1573
1574                 if (lustre_msg_get_transno(req->rq_reqmsg) !=
1575                     obd->obd_next_recovery_transno) {
1576                         spin_unlock_bh(&obd->obd_processing_task_lock);
1577                         CDEBUG(D_HA, "%s: waiting for transno "LPD64" (1st is "
1578                                LPD64", x"LPU64")\n", obd->obd_name,
1579                                obd->obd_next_recovery_transno,
1580                                lustre_msg_get_transno(req->rq_reqmsg),
1581                                req->rq_xid);
1582                         l_wait_event(obd->obd_next_transno_waitq,
1583                                      check_for_next_transno(obd), &lwi);
1584                         if (target_recovery_check_and_stop(obd)) {
1585                                 EXIT;
1586                                 return;
1587                         }
1588                         continue;
1589                 }
1590                 list_del_init(&req->rq_list);
1591                 LASSERT(obd->obd_recovery_thread);
1592                 /* replace request initial thread with current one, bug #18221 */
1593                 req->rq_svc_thread = obd->obd_recovery_thread;
1594                 obd->obd_requests_queued_for_recovery--;
1595                 spin_unlock_bh(&obd->obd_processing_task_lock);
1596
1597                 DEBUG_REQ(D_HA, req, "processing: ");
1598                 (void)obd->obd_recovery_handler(req);
1599                 obd->obd_replayed_requests++;
1600                 /* Extend the recovery timer enough to complete the next
1601                  * replayed rpc */
1602                 reset_recovery_timer(obd, AT_OFF ? obd_timeout :
1603                        at_get(&req->rq_rqbd->rqbd_service->srv_at_estimate), 1);
1604                 /* bug 1580: decide how to properly sync() in recovery */
1605                 //mds_fsync_super(obd->u.obt.obt_sb);
1606                 spin_lock_bh(&obd->obd_processing_task_lock);
1607                 obd->obd_next_recovery_transno++;
1608                 spin_unlock_bh(&obd->obd_processing_task_lock);
1609                 target_exp_dequeue_req_replay(req);
1610                 target_request_copy_put(req);
1611                 OBD_RACE(OBD_FAIL_TGT_REPLAY_DELAY);
1612                 spin_lock_bh(&obd->obd_processing_task_lock);
1613                 if (list_empty(&obd->obd_recovery_queue)) {
1614                         obd->obd_processing_task = 0;
1615                         obd->obd_recovery_thread = NULL;
1616                         spin_unlock_bh(&obd->obd_processing_task_lock);
1617                         break;
1618                 }
1619                 spin_unlock_bh(&obd->obd_processing_task_lock);
1620         }
1621         EXIT;
1622 }
1623
1624 int target_queue_recovery_request(struct ptlrpc_request *req,
1625                                   struct obd_device *obd)
1626 {
1627         struct list_head *tmp;
1628         int inserted = 0;
1629         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1630         ENTRY;
1631         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1632          * (i.e. buflens etc are in my own byte order), but type-dependent
1633          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1634
1635         if (!transno) {
1636                 CFS_INIT_LIST_HEAD(&req->rq_list);
1637                 DEBUG_REQ(D_HA, req, "not queueing");
1638                 RETURN(1);
1639         }
1640
1641         spin_lock_bh(&obd->obd_processing_task_lock);
1642
1643         if (!obd->obd_recovering) {
1644                 spin_unlock_bh(&obd->obd_processing_task_lock);
1645                 RETURN(0);
1646         }
1647
1648         /* If we're processing the queue, we want don't want to queue this
1649          * message.
1650          *
1651          * Also, if this request has a transno less than the one we're waiting
1652          * for, we should process it now.  It could (and currently always will)
1653          * be an open request for a descriptor that was opened some time ago.
1654          *
1655          * Also, a resent, replayed request that has already been
1656          * handled will pass through here and be processed immediately.
1657          */
1658         if (obd->obd_processing_task == cfs_curproc_pid() ||
1659             transno < obd->obd_next_recovery_transno) {
1660                 /* Processing the queue right now, don't re-add. */
1661                 LASSERT(list_empty(&req->rq_list));
1662                 spin_unlock_bh(&obd->obd_processing_task_lock);
1663                 RETURN(1);
1664         }
1665
1666         if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))) {
1667                 spin_unlock_bh(&obd->obd_processing_task_lock);
1668                 RETURN(0);
1669         }
1670
1671         if (target_exp_enqueue_req_replay(req)) {
1672                 spin_unlock_bh(&obd->obd_processing_task_lock);
1673                 DEBUG_REQ(D_ERROR, req, "%s: dropping resent queued req",
1674                                         obd->obd_name);
1675                 RETURN(0);
1676         }
1677
1678         /* XXX O(n^2) */
1679         list_for_each(tmp, &obd->obd_recovery_queue) {
1680                 struct ptlrpc_request *reqiter =
1681                         list_entry(tmp, struct ptlrpc_request, rq_list);
1682
1683                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
1684                         list_add_tail(&req->rq_list, &reqiter->rq_list);
1685                         inserted = 1;
1686                         break;
1687                 }
1688
1689                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
1690                              transno)) {
1691                         spin_unlock_bh(&obd->obd_processing_task_lock);
1692                         DEBUG_REQ(D_ERROR, req, "%s: dropping replay: transno "
1693                                   "has been claimed by another client",
1694                                   obd->obd_name);
1695                         target_exp_dequeue_req_replay(req);
1696                         RETURN(0);
1697                 }
1698         }
1699
1700         if (!inserted) {
1701                 list_add_tail(&req->rq_list, &obd->obd_recovery_queue);
1702         }
1703
1704         target_request_copy_get(req);
1705         obd->obd_requests_queued_for_recovery++;
1706
1707         if (obd->obd_processing_task != 0) {
1708                 /* Someone else is processing this queue, we'll leave it to
1709                  * them.
1710                  */
1711                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1712                 spin_unlock_bh(&obd->obd_processing_task_lock);
1713                 RETURN(0);
1714         }
1715
1716         /* Nobody is processing, and we know there's (at least) one to process
1717          * now, so we'll do the honours.
1718          */
1719         obd->obd_processing_task = cfs_curproc_pid();
1720         /* save thread that handle recovery queue */
1721         obd->obd_recovery_thread = req->rq_svc_thread;
1722         spin_unlock_bh(&obd->obd_processing_task_lock);
1723
1724         process_recovery_queue(obd);
1725         RETURN(0);
1726 }
1727
1728 struct obd_device * target_req2obd(struct ptlrpc_request *req)
1729 {
1730         return req->rq_export->exp_obd;
1731 }
1732
1733 int target_queue_last_replay_reply(struct ptlrpc_request *req, int rc)
1734 {
1735         struct obd_device *obd = target_req2obd(req);
1736         struct obd_export *exp = req->rq_export;
1737         int recovery_done = 0, delayed_done = 0;
1738
1739         LASSERT ((rc == 0) == req->rq_packed_final);
1740
1741         if (!req->rq_packed_final) {
1742                 /* Just like ptlrpc_error, but without the sending. */
1743                 rc = lustre_pack_reply(req, 1, NULL, NULL);
1744                 if (rc)
1745                         return rc;
1746                 req->rq_type = PTL_RPC_MSG_ERR;
1747         }
1748
1749         LASSERT(!req->rq_reply_state->rs_difficult);
1750         LASSERT(list_empty(&req->rq_list));
1751
1752         /* Don't race cleanup */
1753         spin_lock_bh(&obd->obd_processing_task_lock);
1754         if (obd->obd_stopping) {
1755                 spin_unlock_bh(&obd->obd_processing_task_lock);
1756                 goto out_noconn;
1757         }
1758
1759         if (!exp->exp_vbr_failed) {
1760                 target_request_copy_get(req);
1761                 list_add(&req->rq_list, &obd->obd_delayed_reply_queue);
1762         }
1763
1764         /* only count the first "replay over" request from each
1765            export */
1766         if (exp->exp_replay_needed) {
1767                 spin_lock(&exp->exp_lock);
1768                 exp->exp_replay_needed = 0;
1769                 spin_unlock(&exp->exp_lock);
1770
1771                 if (!exp->exp_delayed) {
1772                         --obd->obd_recoverable_clients;
1773                 } else {
1774                         spin_lock(&exp->exp_lock);
1775                         exp->exp_delayed = 0;
1776                         spin_unlock(&exp->exp_lock);
1777                         delayed_done = 1;
1778                         if (obd->obd_delayed_clients == 0) {
1779                                 spin_unlock_bh(&obd->obd_processing_task_lock);
1780                                 LBUG();
1781                         }
1782                         --obd->obd_delayed_clients;
1783                 }
1784         }
1785         recovery_done = (obd->obd_recoverable_clients == 0);
1786         spin_unlock_bh(&obd->obd_processing_task_lock);
1787
1788         if (delayed_done) {
1789                 /* start pinging export */
1790                 spin_lock(&obd->obd_dev_lock);
1791                 list_add_tail(&exp->exp_obd_chain_timed,
1792                               &obd->obd_exports_timed);
1793                 list_move_tail(&exp->exp_obd_chain, &obd->obd_exports);
1794                 spin_unlock(&obd->obd_dev_lock);
1795                 target_send_delayed_replies(obd);
1796         }
1797
1798         OBD_RACE(OBD_FAIL_LDLM_RECOV_CLIENTS);
1799         if (recovery_done) {
1800                 spin_lock_bh(&obd->obd_processing_task_lock);
1801                 obd->obd_recovering = 0;
1802                 obd->obd_version_recov = 0;
1803                 obd->obd_abort_recovery = 0;
1804                 target_cancel_recovery_timer(obd);
1805                 spin_unlock_bh(&obd->obd_processing_task_lock);
1806
1807                 if (!delayed_done)
1808                         target_finish_recovery(obd);
1809                 CDEBUG(D_HA, "%s: recovery complete\n",
1810                        obd_uuid2str(&obd->obd_uuid));
1811         } else {
1812                 CWARN("%s: %d recoverable clients remain\n",
1813                       obd->obd_name, obd->obd_recoverable_clients);
1814                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1815         }
1816
1817         /* VBR: disconnect export with failed recovery */
1818         if (exp->exp_vbr_failed) {
1819                 CWARN("%s: disconnect export %s\n", obd->obd_name,
1820                       exp->exp_client_uuid.uuid);
1821                 class_fail_export(exp);
1822                 req->rq_status = 0;
1823                 ptlrpc_send_reply(req, 0);
1824         }
1825
1826         return 1;
1827
1828 out_noconn:
1829         req->rq_status = -ENOTCONN;
1830         /* rv is ignored anyhow */
1831         return -ENOTCONN;
1832 }
1833
1834 int target_handle_reply(struct ptlrpc_request *req, int rc, int fail)
1835 {
1836         struct obd_device *obd = NULL;
1837
1838         if (req->rq_export)
1839                 obd = target_req2obd(req);
1840
1841         /* handle replay reply for version recovery */
1842         if (obd && obd->obd_version_recov &&
1843             (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1844                 LASSERT(req->rq_repmsg);
1845                 lustre_msg_add_flags(req->rq_repmsg, MSG_VERSION_REPLAY);
1846         }
1847
1848         /* handle last replay */
1849         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY) {
1850                 if (obd &&
1851                     lustre_msg_get_flags(req->rq_reqmsg) & MSG_DELAY_REPLAY) {
1852                         DEBUG_REQ(D_HA, req,
1853                                   "delayed LAST_REPLAY, queuing reply");
1854                         rc = target_queue_last_replay_reply(req, rc);
1855                         LASSERT(req->rq_export->exp_delayed == 0);
1856                         return rc;
1857                 }
1858
1859                 if (obd && obd->obd_recovering) { /* normal recovery */
1860                         DEBUG_REQ(D_HA, req, "LAST_REPLAY, queuing reply");
1861                         rc = target_queue_last_replay_reply(req, rc);
1862                         return rc;
1863                 }
1864
1865                 /* Lost a race with recovery; let the error path DTRT. */
1866                 rc = req->rq_status = -ENOTCONN;
1867         }
1868         target_send_reply(req, rc, fail);
1869         return 0;
1870 }
1871
1872 static inline struct ldlm_pool *ldlm_exp2pl(struct obd_export *exp)
1873 {
1874         LASSERT(exp != NULL);
1875         return &exp->exp_obd->obd_namespace->ns_pool;
1876 }
1877
1878 int target_pack_pool_reply(struct ptlrpc_request *req)
1879 {
1880         struct obd_device *obd;
1881         ENTRY;
1882
1883         /*
1884          * Check that we still have all structures alive as this may
1885          * be some late rpc in shutdown time.
1886          */
1887         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
1888                      !exp_connect_lru_resize(req->rq_export))) {
1889                 lustre_msg_set_slv(req->rq_repmsg, 0);
1890                 lustre_msg_set_limit(req->rq_repmsg, 0);
1891                 RETURN(0);
1892         }
1893
1894         /*
1895          * OBD is alive here as export is alive, which we checked above.
1896          */
1897         obd = req->rq_export->exp_obd;
1898
1899         read_lock(&obd->obd_pool_lock);
1900         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
1901         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
1902         read_unlock(&obd->obd_pool_lock);
1903
1904         RETURN(0);
1905 }
1906
1907 int
1908 target_send_reply_msg (struct ptlrpc_request *req, int rc, int fail_id)
1909 {
1910         if (OBD_FAIL_CHECK(fail_id | OBD_FAIL_ONCE)) {
1911                 obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
1912                 DEBUG_REQ(D_ERROR, req, "dropping reply");
1913                 return (-ECOMM);
1914         }
1915
1916         if (rc) {
1917                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
1918                 req->rq_status = rc;
1919                 return (ptlrpc_send_error(req, 1));
1920         } else {
1921                 DEBUG_REQ(D_NET, req, "sending reply");
1922         }
1923
1924         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
1925 }
1926
1927 void
1928 target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
1929 {
1930         int                        netrc;
1931         struct ptlrpc_reply_state *rs;
1932         struct obd_device         *obd;
1933         struct obd_export         *exp;
1934         struct ptlrpc_service     *svc;
1935
1936         svc = req->rq_rqbd->rqbd_service;
1937         rs = req->rq_reply_state;
1938         if (rs == NULL || !rs->rs_difficult) {
1939                 /* no notifiers */
1940                 target_send_reply_msg (req, rc, fail_id);
1941                 return;
1942         }
1943
1944         /* must be an export if locks saved */
1945         LASSERT (req->rq_export != NULL);
1946         /* req/reply consistent */
1947         LASSERT (rs->rs_service == svc);
1948
1949         /* "fresh" reply */
1950         LASSERT (!rs->rs_scheduled);
1951         LASSERT (!rs->rs_scheduled_ever);
1952         LASSERT (!rs->rs_handled);
1953         LASSERT (!rs->rs_on_net);
1954         LASSERT (rs->rs_export == NULL);
1955         LASSERT (list_empty(&rs->rs_obd_list));
1956         LASSERT (list_empty(&rs->rs_exp_list));
1957
1958         exp = class_export_get(req->rq_export);
1959         obd = exp->exp_obd;
1960
1961         /* disable reply scheduling onto srv_reply_queue while I'm setting up */
1962         rs->rs_scheduled = 1;
1963         rs->rs_on_net    = 1;
1964         rs->rs_xid       = req->rq_xid;
1965         rs->rs_transno   = req->rq_transno;
1966         rs->rs_export    = exp;
1967
1968         spin_lock(&exp->exp_uncommitted_replies_lock);
1969
1970         /* VBR: use exp_last_committed */
1971         if (rs->rs_transno > exp->exp_last_committed) {
1972                 /* not committed already */
1973                 list_add_tail (&rs->rs_obd_list,
1974                                &exp->exp_uncommitted_replies);
1975         }
1976
1977         spin_unlock (&exp->exp_uncommitted_replies_lock);
1978         spin_lock (&exp->exp_lock);
1979
1980         list_add_tail (&rs->rs_exp_list, &exp->exp_outstanding_replies);
1981
1982         spin_unlock(&exp->exp_lock);
1983
1984         netrc = target_send_reply_msg (req, rc, fail_id);
1985
1986         spin_lock(&svc->srv_lock);
1987
1988         svc->srv_n_difficult_replies++;
1989
1990         if (netrc != 0) {
1991                 /* error sending: reply is off the net.  Also we need +1
1992                  * reply ref until ptlrpc_server_handle_reply() is done
1993                  * with the reply state (if the send was successful, there
1994                  * would have been +1 ref for the net, which
1995                  * reply_out_callback leaves alone) */
1996                 rs->rs_on_net = 0;
1997                 ptlrpc_rs_addref(rs);
1998                 atomic_inc (&svc->srv_outstanding_replies);
1999         }
2000
2001         if (!rs->rs_on_net ||                   /* some notifier */
2002             list_empty(&rs->rs_exp_list) ||     /* completed already */
2003             list_empty(&rs->rs_obd_list)) {
2004                 list_add_tail (&rs->rs_list, &svc->srv_reply_queue);
2005                 cfs_waitq_signal (&svc->srv_waitq);
2006         } else {
2007                 list_add (&rs->rs_list, &svc->srv_active_replies);
2008                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2009         }
2010
2011         spin_unlock(&svc->srv_lock);
2012 }
2013
2014 int target_handle_ping(struct ptlrpc_request *req)
2015 {
2016         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LAST_REPLAY &&
2017             req->rq_export->exp_in_recovery) {
2018                 spin_lock(&req->rq_export->exp_lock);
2019                 req->rq_export->exp_in_recovery = 0;
2020                 spin_unlock(&req->rq_export->exp_lock);
2021         }
2022         obd_ping(req->rq_export);
2023         return lustre_pack_reply(req, 1, NULL, NULL);
2024 }
2025
2026 void target_committed_to_req(struct ptlrpc_request *req)
2027 {
2028         struct obd_export *exp = req->rq_export;
2029         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL) {
2030                 lustre_msg_set_last_committed(req->rq_repmsg,
2031                                               exp->exp_last_committed);
2032         } else {
2033                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2034                           "%d)", exp->exp_obd->obd_no_transno,
2035                           req->rq_repmsg == NULL);
2036         }
2037         CDEBUG(D_INFO, "last_committed x"LPU64", this req x"LPU64"\n",
2038                exp->exp_obd->obd_last_committed, req->rq_xid);
2039 }
2040
2041 EXPORT_SYMBOL(target_committed_to_req);
2042
2043 int target_handle_qc_callback(struct ptlrpc_request *req)
2044 {
2045         struct obd_quotactl *oqctl;
2046         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2047
2048         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
2049                                    lustre_swab_obd_quotactl);
2050         if (oqctl == NULL) {
2051                 CERROR("Can't unpack obd_quatactl\n");
2052                 RETURN(-EPROTO);
2053         }
2054
2055         cli->cl_qchk_stat = oqctl->qc_stat;
2056
2057         return 0;
2058 }
2059
2060 #ifdef HAVE_QUOTA_SUPPORT
2061 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2062 {
2063 #ifdef __KERNEL__
2064         struct obd_device *obd = req->rq_export->exp_obd;
2065         struct obd_device *master_obd = NULL, *lov_obd = NULL;
2066         struct lustre_quota_ctxt *qctxt;
2067         struct qunit_data *qdata = NULL;
2068         int rc = 0;
2069         int repsize[2] = { sizeof(struct ptlrpc_body), 0 };
2070         ENTRY;
2071
2072         if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2073                 RETURN(rc);
2074
2075         repsize[1] = quota_get_qunit_data_size(req->rq_export->
2076                                                exp_connect_flags);
2077
2078         rc = lustre_pack_reply(req, 2, repsize, NULL);
2079         if (rc)
2080                 RETURN(rc);
2081
2082         LASSERT(req->rq_export);
2083
2084         /* there are three forms of qunit(historic causes), so we need to
2085          * adjust qunits from slaves to the same form here */
2086         OBD_ALLOC(qdata, sizeof(struct qunit_data));
2087         if (!qdata)
2088                 RETURN(-ENOMEM);
2089         rc = quota_get_qdata(req, qdata, QUOTA_REQUEST, QUOTA_EXPORT);
2090         if (rc < 0) {
2091                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2092                 GOTO(out, rc);
2093         }
2094
2095         /* we use the observer */
2096         if (obd_pin_observer(obd, &lov_obd) ||
2097             obd_pin_observer(lov_obd, &master_obd)) {
2098                 CERROR("Can't find the observer, it is recovering\n");
2099                 req->rq_status = -EAGAIN;
2100                 GOTO(send_reply, rc = -EAGAIN);
2101         }
2102
2103         qctxt = &master_obd->u.obt.obt_qctxt;
2104
2105         if (!qctxt->lqc_setup) {
2106                 /* quota_type has not been processed yet, return EAGAIN
2107                  * until we know whether or not quotas are supposed to
2108                  * be enabled */
2109                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2110                                 "-EAGAIN\n");
2111                 req->rq_status = -EAGAIN;
2112                 rc = ptlrpc_reply(req);
2113                 GOTO(out, rc);
2114         }
2115
2116         LASSERT(qctxt->lqc_handler);
2117         rc = qctxt->lqc_handler(master_obd, qdata,
2118                                 lustre_msg_get_opc(req->rq_reqmsg));
2119         if (rc && rc != -EDQUOT)
2120                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2121                        "dqacq failed! (rc:%d)\n", rc);
2122         req->rq_status = rc;
2123
2124         /* there are three forms of qunit(historic causes), so we need to
2125          * adjust the same form to different forms slaves needed */
2126         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2127         if (rc < 0) {
2128                 CDEBUG(D_ERROR, "Can't pack qunit_data(rc: %d)\n", rc);
2129                 GOTO(out, rc);
2130         }
2131
2132         /* Block the quota req. b=14840 */
2133         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2134  send_reply:
2135         rc = ptlrpc_reply(req);
2136 out:
2137         if (master_obd)
2138                 obd_unpin_observer(lov_obd);
2139         if (lov_obd)
2140                 obd_unpin_observer(obd);
2141         OBD_FREE(qdata, sizeof(struct qunit_data));
2142         RETURN(rc);
2143 #else
2144         return 0;
2145 #endif /* !__KERNEL__ */
2146 }
2147 #endif /* HAVE_QUOTA_SUPPORT */
2148
2149 ldlm_mode_t lck_compat_array[] = {
2150         [LCK_EX] LCK_COMPAT_EX,
2151         [LCK_PW] LCK_COMPAT_PW,
2152         [LCK_PR] LCK_COMPAT_PR,
2153         [LCK_CW] LCK_COMPAT_CW,
2154         [LCK_CR] LCK_COMPAT_CR,
2155         [LCK_NL] LCK_COMPAT_NL,
2156         [LCK_GROUP] LCK_COMPAT_GROUP
2157 };