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  2008 Sun Microsystems, Inc. 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 <lustre_sec.h>
  52 #include "ldlm_internal.h"
  53
  54 /* @priority: if non-zero, move the selected to the list head
  55  * @create: if zero, only search in existed connections
  56  */
  57 static int import_set_conn(struct obd_import *imp, struct obd_uuid *uuid,
  58                            int priority, int create)
  59 {
  60         struct ptlrpc_connection *ptlrpc_conn;
  61         struct obd_import_conn *imp_conn = NULL, *item;
  62         int rc = 0;
  63         ENTRY;
  64
  65         if (!create && !priority) {
  66                 CDEBUG(D_HA, "Nothing to do\n");
  67                 RETURN(-EINVAL);
  68         }
  69
  70         ptlrpc_conn = ptlrpc_uuid_to_connection(uuid);
  71         if (!ptlrpc_conn) {
  72                 CDEBUG(D_HA, "can't find connection %s\n", uuid->uuid);
  73                 RETURN (-ENOENT);
  74         }
  75
  76         if (create) {
  77                 OBD_ALLOC(imp_conn, sizeof(*imp_conn));
  78                 if (!imp_conn) {
  79                         GOTO(out_put, rc = -ENOMEM);
  80                 }
  81         }
  82
  83         spin_lock(&imp->imp_lock);
  84         list_for_each_entry(item, &imp->imp_conn_list, oic_item) {
  85                 if (obd_uuid_equals(uuid, &item->oic_uuid)) {
  86                         if (priority) {
  87                                 list_del(&item->oic_item);
  88                                 list_add(&item->oic_item, &imp->imp_conn_list);
  89                                 item->oic_last_attempt = 0;
  90                         }
  91                         CDEBUG(D_HA, "imp %p@%s: found existing conn %s%s\n",
  92                                imp, imp->imp_obd->obd_name, uuid->uuid,
  93                                (priority ? ", moved to head" : ""));
  94                         spin_unlock(&imp->imp_lock);
  95                         GOTO(out_free, rc = 0);
  96                 }
  97         }
  98         /* not found */
  99         if (create) {
 100                 imp_conn->oic_conn = ptlrpc_conn;
 101                 imp_conn->oic_uuid = *uuid;
 102                 imp_conn->oic_last_attempt = 0;
 103                 if (priority)
 104                         list_add(&imp_conn->oic_item, &imp->imp_conn_list);
 105                 else
 106                         list_add_tail(&imp_conn->oic_item, &imp->imp_conn_list);
 107                 CDEBUG(D_HA, "imp %p@%s: add connection %s at %s\n",
 108                        imp, imp->imp_obd->obd_name, uuid->uuid,
 109                        (priority ? "head" : "tail"));
 110         } else {
 111                 spin_unlock(&imp->imp_lock);
 112                 GOTO(out_free, rc = -ENOENT);
 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 void client_destroy_import(struct obd_import *imp)
 192 {
 193         /* drop security policy instance after all rpc finished/aborted
 194          * to let all busy contexts be released. */
 195         class_import_get(imp);
 196         class_destroy_import(imp);
 197         sptlrpc_import_sec_put(imp);
 198         class_import_put(imp);
 199 }
 200
 201 /* configure an RPC client OBD device
 202  *
 203  * lcfg parameters:
 204  * 1 - client UUID
 205  * 2 - server UUID
 206  * 3 - inactive-on-startup
 207  */
 208 int client_obd_setup(struct obd_device *obddev, struct lustre_cfg *lcfg)
 209 {
 210         struct client_obd *cli = &obddev->u.cli;
 211         struct obd_import *imp;
 212         struct obd_uuid server_uuid;
 213         int rq_portal, rp_portal, connect_op;
 214         char *name = obddev->obd_type->typ_name;
 215         int rc;
 216         ENTRY;
 217
 218         /* In a more perfect world, we would hang a ptlrpc_client off of
 219          * obd_type and just use the values from there. */
 220         if (!strcmp(name, LUSTRE_OSC_NAME)) {
 221                 rq_portal = OST_REQUEST_PORTAL;
 222                 rp_portal = OSC_REPLY_PORTAL;
 223                 connect_op = OST_CONNECT;
 224                 cli->cl_sp_me = LUSTRE_SP_CLI;
 225                 cli->cl_sp_to = LUSTRE_SP_OST;
 226         } else if (!strcmp(name, LUSTRE_MDC_NAME)) {
 227                 rq_portal = MDS_REQUEST_PORTAL;
 228                 rp_portal = MDC_REPLY_PORTAL;
 229                 connect_op = MDS_CONNECT;
 230                 cli->cl_sp_me = LUSTRE_SP_CLI;
 231                 cli->cl_sp_to = LUSTRE_SP_MDT;
 232         } else if (!strcmp(name, LUSTRE_MGC_NAME)) {
 233                 rq_portal = MGS_REQUEST_PORTAL;
 234                 rp_portal = MGC_REPLY_PORTAL;
 235                 connect_op = MGS_CONNECT;
 236                 cli->cl_sp_me = LUSTRE_SP_MGC;
 237                 cli->cl_sp_to = LUSTRE_SP_MGS;
 238                 cli->cl_flvr_mgc.sf_rpc = SPTLRPC_FLVR_INVALID;
 239         } else {
 240                 CERROR("unknown client OBD type \"%s\", can't setup\n",
 241                        name);
 242                 RETURN(-EINVAL);
 243         }
 244
 245         if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
 246                 CERROR("requires a TARGET UUID\n");
 247                 RETURN(-EINVAL);
 248         }
 249
 250         if (LUSTRE_CFG_BUFLEN(lcfg, 1) > 37) {
 251                 CERROR("client UUID must be less than 38 characters\n");
 252                 RETURN(-EINVAL);
 253         }
 254
 255         if (LUSTRE_CFG_BUFLEN(lcfg, 2) < 1) {
 256                 CERROR("setup requires a SERVER UUID\n");
 257                 RETURN(-EINVAL);
 258         }
 259
 260         if (LUSTRE_CFG_BUFLEN(lcfg, 2) > 37) {
 261                 CERROR("target UUID must be less than 38 characters\n");
 262                 RETURN(-EINVAL);
 263         }
 264
 265         init_rwsem(&cli->cl_sem);
 266         sema_init(&cli->cl_mgc_sem, 1);
 267         cli->cl_conn_count = 0;
 268         memcpy(server_uuid.uuid, lustre_cfg_buf(lcfg, 2),
 269                min_t(unsigned int, LUSTRE_CFG_BUFLEN(lcfg, 2),
 270                      sizeof(server_uuid)));
 271
 272         cli->cl_dirty = 0;
 273         cli->cl_avail_grant = 0;
 274         /* FIXME: should limit this for the sum of all cl_dirty_max */
 275         cli->cl_dirty_max = OSC_MAX_DIRTY_DEFAULT * 1024 * 1024;
 276         if (cli->cl_dirty_max >> CFS_PAGE_SHIFT > num_physpages / 8)
 277                 cli->cl_dirty_max = num_physpages << (CFS_PAGE_SHIFT - 3);
 278         CFS_INIT_LIST_HEAD(&cli->cl_cache_waiters);
 279         CFS_INIT_LIST_HEAD(&cli->cl_loi_ready_list);
 280         CFS_INIT_LIST_HEAD(&cli->cl_loi_hp_ready_list);
 281         CFS_INIT_LIST_HEAD(&cli->cl_loi_write_list);
 282         CFS_INIT_LIST_HEAD(&cli->cl_loi_read_list);
 283         client_obd_list_lock_init(&cli->cl_loi_list_lock);
 284         cli->cl_r_in_flight = 0;
 285         cli->cl_w_in_flight = 0;
 286
 287         spin_lock_init(&cli->cl_read_rpc_hist.oh_lock);
 288         spin_lock_init(&cli->cl_write_rpc_hist.oh_lock);
 289         spin_lock_init(&cli->cl_read_page_hist.oh_lock);
 290         spin_lock_init(&cli->cl_write_page_hist.oh_lock);
 291         spin_lock_init(&cli->cl_read_offset_hist.oh_lock);
 292         spin_lock_init(&cli->cl_write_offset_hist.oh_lock);
 293         cfs_waitq_init(&cli->cl_destroy_waitq);
 294         atomic_set(&cli->cl_destroy_in_flight, 0);
 295 #ifdef ENABLE_CHECKSUM
 296         /* Turn on checksumming by default. */
 297         cli->cl_checksum = 1;
 298         /*
 299          * The supported checksum types will be worked out at connect time
 300          * Set cl_chksum* to CRC32 for now to avoid returning screwed info
 301          * through procfs.
 302          */
 303         cli->cl_cksum_type = cli->cl_supp_cksum_types = OBD_CKSUM_CRC32;
 304 #endif
 305         atomic_set(&cli->cl_resends, OSC_DEFAULT_RESENDS);
 306         atomic_set(&cli->cl_quota_resends, CLIENT_QUOTA_DEFAULT_RESENDS);
 307
 308         /* This value may be changed at connect time in
 309            ptlrpc_connect_interpret. */
 310         cli->cl_max_pages_per_rpc = min((int)PTLRPC_MAX_BRW_PAGES,
 311                                         (int)(1024 * 1024 >> CFS_PAGE_SHIFT));
 312
 313         if (!strcmp(name, LUSTRE_MDC_NAME)) {
 314                 cli->cl_max_rpcs_in_flight = MDC_MAX_RIF_DEFAULT;
 315         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 128 /* MB */) {
 316                 cli->cl_max_rpcs_in_flight = 2;
 317         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 256 /* MB */) {
 318                 cli->cl_max_rpcs_in_flight = 3;
 319         } else if (num_physpages >> (20 - CFS_PAGE_SHIFT) <= 512 /* MB */) {
 320                 cli->cl_max_rpcs_in_flight = 4;
 321         } else {
 322                 cli->cl_max_rpcs_in_flight = OSC_MAX_RIF_DEFAULT;
 323         }
 324
 325         rc = ldlm_get_ref();
 326         if (rc) {
 327                 CERROR("ldlm_get_ref failed: %d\n", rc);
 328                 GOTO(err, rc);
 329         }
 330
 331         ptlrpc_init_client(rq_portal, rp_portal, name,
 332                            &obddev->obd_ldlm_client);
 333
 334         imp = class_new_import(obddev);
 335         if (imp == NULL)
 336                 GOTO(err_ldlm, rc = -ENOENT);
 337         imp->imp_client = &obddev->obd_ldlm_client;
 338         imp->imp_connect_op = connect_op;
 339         imp->imp_initial_recov = 1;
 340         imp->imp_initial_recov_bk = 0;
 341         CFS_INIT_LIST_HEAD(&imp->imp_pinger_chain);
 342         memcpy(cli->cl_target_uuid.uuid, lustre_cfg_buf(lcfg, 1),
 343                LUSTRE_CFG_BUFLEN(lcfg, 1));
 344         class_import_put(imp);
 345
 346         rc = client_import_add_conn(imp, &server_uuid, 1);
 347         if (rc) {
 348                 CERROR("can't add initial connection\n");
 349                 GOTO(err_import, rc);
 350         }
 351
 352         cli->cl_import = imp;
 353         /* cli->cl_max_mds_{easize,cookiesize} updated by mdc_init_ea_size() */
 354         cli->cl_max_mds_easize = sizeof(struct lov_mds_md_v3);
 355         cli->cl_max_mds_cookiesize = sizeof(struct llog_cookie);
 356
 357         if (LUSTRE_CFG_BUFLEN(lcfg, 3) > 0) {
 358                 if (!strcmp(lustre_cfg_string(lcfg, 3), "inactive")) {
 359                         CDEBUG(D_HA, "marking %s %s->%s as inactive\n",
 360                                name, obddev->obd_name,
 361                                cli->cl_target_uuid.uuid);
 362                         spin_lock(&imp->imp_lock);
 363                         imp->imp_deactive = 1;
 364                         spin_unlock(&imp->imp_lock);
 365                 }
 366         }
 367
 368         obddev->obd_namespace = ldlm_namespace_new(obddev, obddev->obd_name,
 369                                                    LDLM_NAMESPACE_CLIENT,
 370                                                    LDLM_NAMESPACE_GREEDY);
 371         if (obddev->obd_namespace == NULL) {
 372                 CERROR("Unable to create client namespace - %s\n",
 373                        obddev->obd_name);
 374                 GOTO(err_import, rc = -ENOMEM);
 375         }
 376
 377         cli->cl_qchk_stat = CL_NOT_QUOTACHECKED;
 378
 379         RETURN(rc);
 380
 381 err_import:
 382         class_destroy_import(imp);
 383 err_ldlm:
 384         ldlm_put_ref();
 385 err:
 386         RETURN(rc);
 387
 388 }
 389
 390 int client_obd_cleanup(struct obd_device *obddev)
 391 {
 392         ENTRY;
 393
 394         ldlm_namespace_free_post(obddev->obd_namespace);
 395         obddev->obd_namespace = NULL;
 396
 397         ldlm_put_ref();
 398         RETURN(0);
 399 }
 400
 401 /* ->o_connect() method for client side (OSC and MDC and MGC) */
 402 int client_connect_import(const struct lu_env *env,
 403                           struct obd_export **exp,
 404                           struct obd_device *obd, struct obd_uuid *cluuid,
 405                           struct obd_connect_data *data, void *localdata)
 406 {
 407         struct client_obd *cli = &obd->u.cli;
 408         struct obd_import *imp = cli->cl_import;
 409         struct obd_connect_data *ocd;
 410         struct lustre_handle conn = { 0 };
 411         int rc;
 412         ENTRY;
 413
 414         *exp = NULL;
 415         down_write(&cli->cl_sem);
 416         if (cli->cl_conn_count > 0 )
 417                 GOTO(out_sem, rc = -EALREADY);
 418
 419         rc = class_connect(&conn, obd, cluuid);
 420         if (rc)
 421                 GOTO(out_sem, rc);
 422
 423         cli->cl_conn_count++;
 424         *exp = class_conn2export(&conn);
 425
 426         LASSERT(obd->obd_namespace);
 427
 428         imp->imp_dlm_handle = conn;
 429         rc = ptlrpc_init_import(imp);
 430         if (rc != 0)
 431                 GOTO(out_ldlm, rc);
 432
 433         ocd = &imp->imp_connect_data;
 434         if (data) {
 435                 *ocd = *data;
 436                 imp->imp_connect_flags_orig = data->ocd_connect_flags;
 437         }
 438
 439         rc = ptlrpc_connect_import(imp, NULL);
 440         if (rc != 0) {
 441                 LASSERT (imp->imp_state == LUSTRE_IMP_DISCON);
 442                 GOTO(out_ldlm, rc);
 443         }
 444         LASSERT((*exp)->exp_connection);
 445
 446         if (data) {
 447                 LASSERTF((ocd->ocd_connect_flags & data->ocd_connect_flags) ==
 448                          ocd->ocd_connect_flags, "old "LPX64", new "LPX64"\n",
 449                          data->ocd_connect_flags, ocd->ocd_connect_flags);
 450                 data->ocd_connect_flags = ocd->ocd_connect_flags;
 451         }
 452
 453         ptlrpc_pinger_add_import(imp);
 454
 455         EXIT;
 456
 457         if (rc) {
 458 out_ldlm:
 459                 cli->cl_conn_count--;
 460                 class_disconnect(*exp);
 461                 *exp = NULL;
 462         }
 463 out_sem:
 464         up_write(&cli->cl_sem);
 465
 466         return rc;
 467 }
 468
 469 int client_disconnect_export(struct obd_export *exp)
 470 {
 471         struct obd_device *obd = class_exp2obd(exp);
 472         struct client_obd *cli;
 473         struct obd_import *imp;
 474         int rc = 0, err;
 475         ENTRY;
 476
 477         if (!obd) {
 478                 CERROR("invalid export for disconnect: exp %p cookie "LPX64"\n",
 479                        exp, exp ? exp->exp_handle.h_cookie : -1);
 480                 RETURN(-EINVAL);
 481         }
 482
 483         cli = &obd->u.cli;
 484         imp = cli->cl_import;
 485
 486         down_write(&cli->cl_sem);
 487         CDEBUG(D_INFO, "disconnect %s - %d\n", obd->obd_name,
 488                cli->cl_conn_count);
 489
 490         if (!cli->cl_conn_count) {
 491                 CERROR("disconnecting disconnected device (%s)\n",
 492                        obd->obd_name);
 493                 GOTO(out_disconnect, rc = -EINVAL);
 494         }
 495
 496         cli->cl_conn_count--;
 497         if (cli->cl_conn_count)
 498                 GOTO(out_disconnect, rc = 0);
 499
 500         /* Mark import deactivated now, so we don't try to reconnect if any
 501          * of the cleanup RPCs fails (e.g. ldlm cancel, etc).  We don't
 502          * fully deactivate the import, or that would drop all requests. */
 503         spin_lock(&imp->imp_lock);
 504         imp->imp_deactive = 1;
 505         spin_unlock(&imp->imp_lock);
 506
 507         /* Some non-replayable imports (MDS's OSCs) are pinged, so just
 508          * delete it regardless.  (It's safe to delete an import that was
 509          * never added.) */
 510         (void)ptlrpc_pinger_del_import(imp);
 511
 512         if (obd->obd_namespace != NULL) {
 513                 /* obd_force == local only */
 514                 ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
 515                                        obd->obd_force ? LDLM_FL_LOCAL_ONLY:0,
 516                                        NULL);
 517                 ldlm_namespace_free_prior(obd->obd_namespace, imp, obd->obd_force);
 518         }
 519
 520         /*
 521          * there's no need to hold sem during disconnecting an import,
 522          * and actually it may cause deadlock in gss.
 523          */
 524         up_write(&cli->cl_sem);
 525         rc = ptlrpc_disconnect_import(imp, 0);
 526         down_write(&cli->cl_sem);
 527
 528         ptlrpc_invalidate_import(imp);
 529
 530         if (imp->imp_rq_pool) {
 531                 ptlrpc_free_rq_pool(imp->imp_rq_pool);
 532                 imp->imp_rq_pool = NULL;
 533         }
 534         client_destroy_import(imp);
 535         cli->cl_import = NULL;
 536
 537         EXIT;
 538
 539  out_disconnect:
 540         /* use server style - class_disconnect should be always called for
 541          * o_disconnect */
 542         err = class_disconnect(exp);
 543         if (!rc && err)
 544                 rc = err;
 545
 546         up_write(&cli->cl_sem);
 547
 548         RETURN(rc);
 549 }
 550
 551 int server_disconnect_export(struct obd_export *exp)
 552 {
 553         int rc;
 554         ENTRY;
 555
 556         /* Disconnect early so that clients can't keep using export */
 557         rc = class_disconnect(exp);
 558         /* close import for avoid sending any requests */
 559         if (exp->exp_imp_reverse)
 560                 ptlrpc_cleanup_imp(exp->exp_imp_reverse);
 561
 562         if (exp->exp_obd->obd_namespace != NULL)
 563                 ldlm_cancel_locks_for_export(exp);
 564
 565         /* complete all outstanding replies */
 566         spin_lock(&exp->exp_lock);
 567         while (!list_empty(&exp->exp_outstanding_replies)) {
 568                 struct ptlrpc_reply_state *rs =
 569                         list_entry(exp->exp_outstanding_replies.next,
 570                                    struct ptlrpc_reply_state, rs_exp_list);
 571                 struct ptlrpc_service *svc = rs->rs_service;
 572
 573                 spin_lock(&svc->srv_lock);
 574                 list_del_init(&rs->rs_exp_list);
 575                 spin_lock(&rs->rs_lock);
 576                 ptlrpc_schedule_difficult_reply(rs);
 577                 spin_unlock(&rs->rs_lock);
 578                 spin_unlock(&svc->srv_lock);
 579         }
 580         spin_unlock(&exp->exp_lock);
 581
 582         /* release nid stat refererence */
 583         lprocfs_exp_cleanup(exp);
 584
 585         RETURN(rc);
 586 }
 587
 588 /* --------------------------------------------------------------------------
 589  * from old lib/target.c
 590  * -------------------------------------------------------------------------- */
 591
 592 static int target_handle_reconnect(struct lustre_handle *conn,
 593                                    struct obd_export *exp,
 594                                    struct obd_uuid *cluuid)
 595 {
 596         ENTRY;
 597         if (exp->exp_connection && exp->exp_imp_reverse) {
 598                 struct lustre_handle *hdl;
 599                 hdl = &exp->exp_imp_reverse->imp_remote_handle;
 600                 /* Might be a re-connect after a partition. */
 601                 if (!memcmp(&conn->cookie, &hdl->cookie, sizeof conn->cookie)) {
 602                         CWARN("%s: %s reconnecting\n", exp->exp_obd->obd_name,
 603                               cluuid->uuid);
 604                         conn->cookie = exp->exp_handle.h_cookie;
 605                         /* target_handle_connect() treats EALREADY and
 606                          * -EALREADY differently.  EALREADY means we are
 607                          * doing a valid reconnect from the same client. */
 608                         RETURN(EALREADY);
 609                 } else {
 610                         CERROR("%s reconnecting from %s, "
 611                                "handle mismatch (ours "LPX64", theirs "
 612                                LPX64")\n", cluuid->uuid,
 613                                exp->exp_connection->c_remote_uuid.uuid,
 614                                hdl->cookie, conn->cookie);
 615                         memset(conn, 0, sizeof *conn);
 616                         /* target_handle_connect() treats EALREADY and
 617                          * -EALREADY differently.  -EALREADY is an error
 618                          * (same UUID, different handle). */
 619                         RETURN(-EALREADY);
 620                 }
 621         }
 622
 623         conn->cookie = exp->exp_handle.h_cookie;
 624         CDEBUG(D_HA, "connect export for UUID '%s' at %p, cookie "LPX64"\n",
 625                cluuid->uuid, exp, conn->cookie);
 626         RETURN(0);
 627 }
 628
 629 void target_client_add_cb(struct obd_device *obd, __u64 transno, void *cb_data,
 630                           int error)
 631 {
 632         struct obd_export *exp = cb_data;
 633
 634         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 635                obd->obd_name, exp->exp_client_uuid.uuid);
 636
 637         spin_lock(&exp->exp_lock);
 638         exp->exp_need_sync = 0;
 639         spin_unlock(&exp->exp_lock);
 640 }
 641 EXPORT_SYMBOL(target_client_add_cb);
 642 static void
 643 target_start_and_reset_recovery_timer(struct obd_device *obd,
 644                                       struct ptlrpc_request *req,
 645                                       int new_client);
 646
 647 int target_handle_connect(struct ptlrpc_request *req)
 648 {
 649         struct obd_device *target, *targref = NULL;
 650         struct obd_export *export = NULL;
 651         struct obd_import *revimp;
 652         struct lustre_handle conn;
 653         struct lustre_handle *tmp;
 654         struct obd_uuid tgtuuid;
 655         struct obd_uuid cluuid;
 656         struct obd_uuid remote_uuid;
 657         char *str;
 658         int rc = 0;
 659         int mds_conn = 0;
 660         struct obd_connect_data *data, *tmpdata;
 661         lnet_nid_t *client_nid = NULL;
 662         ENTRY;
 663
 664         OBD_RACE(OBD_FAIL_TGT_CONN_RACE);
 665
 666         str = req_capsule_client_get(&req->rq_pill, &RMF_TGTUUID);
 667         if (str == NULL) {
 668                 DEBUG_REQ(D_ERROR, req, "bad target UUID for connect");
 669                 GOTO(out, rc = -EINVAL);
 670         }
 671
 672         obd_str2uuid(&tgtuuid, str);
 673         target = class_uuid2obd(&tgtuuid);
 674         if (!target)
 675                 target = class_name2obd(str);
 676
 677         if (!target || target->obd_stopping || !target->obd_set_up) {
 678                 LCONSOLE_ERROR_MSG(0x137, "UUID '%s' is not available "
 679                                    " for connect (%s)\n", str,
 680                                    !target ? "no target" :
 681                                    (target->obd_stopping ? "stopping" :
 682                                    "not set up"));
 683                 GOTO(out, rc = -ENODEV);
 684         }
 685
 686         if (target->obd_no_conn) {
 687                 LCONSOLE_WARN("%s: temporarily refusing client connection "
 688                               "from %s\n", target->obd_name,
 689                               libcfs_nid2str(req->rq_peer.nid));
 690                 GOTO(out, rc = -EAGAIN);
 691         }
 692
 693         /* Make sure the target isn't cleaned up while we're here. Yes,
 694            there's still a race between the above check and our incref here.
 695            Really, class_uuid2obd should take the ref. */
 696         targref = class_incref(target, __FUNCTION__, cfs_current());
 697
 698
 699         str = req_capsule_client_get(&req->rq_pill, &RMF_CLUUID);
 700         if (str == NULL) {
 701                 DEBUG_REQ(D_ERROR, req, "bad client UUID for connect");
 702                 GOTO(out, rc = -EINVAL);
 703         }
 704
 705         obd_str2uuid(&cluuid, str);
 706
 707         /* XXX extract a nettype and format accordingly */
 708         switch (sizeof(lnet_nid_t)) {
 709                 /* NB the casts only avoid compiler warnings */
 710         case 8:
 711                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 712                          "NET_"LPX64"_UUID", (__u64)req->rq_peer.nid);
 713                 break;
 714         case 4:
 715                 snprintf(remote_uuid.uuid, sizeof remote_uuid,
 716                          "NET_%x_UUID", (__u32)req->rq_peer.nid);
 717                 break;
 718         default:
 719                 LBUG();
 720         }
 721
 722         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
 723         if (tmp == NULL)
 724                 GOTO(out, rc = -EPROTO);
 725
 726         conn = *tmp;
 727
 728         data = req_capsule_client_get(&req->rq_pill, &RMF_CONNECT_DATA);
 729         if (!data)
 730                 GOTO(out, rc = -EPROTO);
 731
 732         rc = req_capsule_server_pack(&req->rq_pill);
 733         if (rc)
 734                 GOTO(out, rc);
 735
 736         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 737                 if (!data) {
 738                         DEBUG_REQ(D_WARNING, req, "Refusing old (unversioned) "
 739                                   "libclient connection attempt");
 740                         GOTO(out, rc = -EPROTO);
 741                 } else if (data->ocd_version < LUSTRE_VERSION_CODE -
 742                                                LUSTRE_VERSION_ALLOWED_OFFSET ||
 743                            data->ocd_version > LUSTRE_VERSION_CODE +
 744                                                LUSTRE_VERSION_ALLOWED_OFFSET) {
 745                         DEBUG_REQ(D_WARNING, req, "Refusing %s (%d.%d.%d.%d) "
 746                                   "libclient connection attempt",
 747                                   data->ocd_version < LUSTRE_VERSION_CODE ?
 748                                   "old" : "new",
 749                                   OBD_OCD_VERSION_MAJOR(data->ocd_version),
 750                                   OBD_OCD_VERSION_MINOR(data->ocd_version),
 751                                   OBD_OCD_VERSION_PATCH(data->ocd_version),
 752                                   OBD_OCD_VERSION_FIX(data->ocd_version));
 753                         data = req_capsule_server_sized_get(&req->rq_pill,
 754                                                         &RMF_CONNECT_DATA,
 755                                     offsetof(typeof(*data), ocd_version) +
 756                                              sizeof(data->ocd_version));
 757                         if (data) {
 758                                 data->ocd_connect_flags = OBD_CONNECT_VERSION;
 759                                 data->ocd_version = LUSTRE_VERSION_CODE;
 760                         }
 761                         GOTO(out, rc = -EPROTO);
 762                 }
 763         }
 764
 765         if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_INITIAL) &&
 766             (data->ocd_connect_flags & OBD_CONNECT_MDS))
 767                 mds_conn = 1;
 768
 769         /* lctl gets a backstage, all-access pass. */
 770         if (obd_uuid_equals(&cluuid, &target->obd_uuid))
 771                 goto dont_check_exports;
 772
 773         export = lustre_hash_lookup(target->obd_uuid_hash, &cluuid);
 774         if (!export)
 775                 goto no_export;
 776
 777         /* we've found an export in the hash */
 778         if (export->exp_connecting) { /* bug 9635, et. al. */
 779                 CWARN("%s: exp %p already connecting\n",
 780                       export->exp_obd->obd_name, export);
 781                 class_export_put(export);
 782                 export = NULL;
 783                 rc = -EALREADY;
 784         } else if (mds_conn && export->exp_connection) {
 785                 if (req->rq_peer.nid != export->exp_connection->c_peer.nid)
 786                         /* mds reconnected after failover */
 787                         CWARN("%s: received MDS connection from NID %s,"
 788                               " removing former export from NID %s\n",
 789                             target->obd_name, libcfs_nid2str(req->rq_peer.nid),
 790                             libcfs_nid2str(export->exp_connection->c_peer.nid));
 791                 else
 792                         /* new mds connection from the same nid */
 793                         CWARN("%s: received new MDS connection from NID %s,"
 794                               " removing former export from same NID\n",
 795                             target->obd_name, libcfs_nid2str(req->rq_peer.nid));
 796                 class_fail_export(export);
 797                 class_export_put(export);
 798                 export = NULL;
 799                 rc = 0;
 800         } else if (export->exp_connection != NULL &&
 801                    req->rq_peer.nid != export->exp_connection->c_peer.nid &&
 802                    (lustre_msg_get_op_flags(req->rq_reqmsg) &
 803                     MSG_CONNECT_INITIAL)) {
 804                 /* in mds failover we have static uuid but nid can be
 805                  * changed*/
 806                 CWARN("%s: cookie %s seen on new NID %s when "
 807                       "existing NID %s is already connected\n",
 808                       target->obd_name, cluuid.uuid,
 809                       libcfs_nid2str(req->rq_peer.nid),
 810                       libcfs_nid2str(export->exp_connection->c_peer.nid));
 811                 rc = -EALREADY;
 812                 class_export_put(export);
 813                 export = NULL;
 814         } else {
 815                 spin_lock(&export->exp_lock);
 816                 export->exp_connecting = 1;
 817                 spin_unlock(&export->exp_lock);
 818                 class_export_put(export);
 819                 LASSERT(export->exp_obd == target);
 820
 821                 rc = target_handle_reconnect(&conn, export, &cluuid);
 822         }
 823
 824         /* If we found an export, we already unlocked. */
 825         if (!export) {
 826 no_export:
 827                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_CONNECT, 2 * obd_timeout);
 828         } else if (req->rq_export == NULL &&
 829                    atomic_read(&export->exp_rpc_count) > 0) {
 830                 CWARN("%s: refuse connection from %s/%s to 0x%p/%d\n",
 831                       target->obd_name, cluuid.uuid,
 832                       libcfs_nid2str(req->rq_peer.nid),
 833                       export, atomic_read(&export->exp_refcount));
 834                 GOTO(out, rc = -EBUSY);
 835         } else if (req->rq_export != NULL &&
 836                    (atomic_read(&export->exp_rpc_count) > 1)) {
 837                 /* the current connect rpc has increased exp_rpc_count */
 838                 CWARN("%s: refuse reconnection from %s@%s to 0x%p/%d\n",
 839                       target->obd_name, cluuid.uuid,
 840                       libcfs_nid2str(req->rq_peer.nid),
 841                       export, atomic_read(&export->exp_rpc_count) - 1);
 842                 spin_lock(&export->exp_lock);
 843                 if (req->rq_export->exp_conn_cnt <
 844                     lustre_msg_get_conn_cnt(req->rq_reqmsg))
 845                         /* try to abort active requests */
 846                         req->rq_export->exp_abort_active_req = 1;
 847                 spin_unlock(&export->exp_lock);
 848                 GOTO(out, rc = -EBUSY);
 849         } else if (lustre_msg_get_conn_cnt(req->rq_reqmsg) == 1) {
 850                 CERROR("%s: NID %s (%s) reconnected with 1 conn_cnt; "
 851                        "cookies not random?\n", target->obd_name,
 852                        libcfs_nid2str(req->rq_peer.nid), cluuid.uuid);
 853                 GOTO(out, rc = -EALREADY);
 854         } else {
 855                 OBD_FAIL_TIMEOUT(OBD_FAIL_TGT_DELAY_RECONNECT, 2 * obd_timeout);
 856         }
 857
 858         if (rc < 0) {
 859                 GOTO(out, rc);
 860         }
 861
 862         CWARN("%s: connection from %s@%s %st"LPU64" exp %p cur %ld last %ld\n",
 863                target->obd_name, cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 864               target->obd_recovering ? "recovering/" : "", data->ocd_transno,
 865               export, (long)cfs_time_current_sec(),
 866               export ? (long)export->exp_last_request_time : 0);
 867
 868         /* Tell the client if we're in recovery. */
 869         if (target->obd_recovering) {
 870                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECOVERING);
 871                 /* If this is the first time a client connects,
 872                    reset the recovery timer */
 873                 if (rc == 0)
 874                         target_start_and_reset_recovery_timer(target, req,
 875                                                               !export);
 876         }
 877
 878         /* We want to handle EALREADY but *not* -EALREADY from
 879          * target_handle_reconnect(), return reconnection state in a flag */
 880         if (rc == EALREADY) {
 881                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_RECONNECT);
 882                 rc = 0;
 883         } else {
 884                 LASSERT(rc == 0);
 885         }
 886
 887         /* Tell the client if we support replayable requests */
 888         if (target->obd_replayable)
 889                 lustre_msg_add_op_flags(req->rq_repmsg, MSG_CONNECT_REPLAYABLE);
 890         client_nid = &req->rq_peer.nid;
 891
 892         if (export == NULL) {
 893                 if (target->obd_recovering) {
 894                         cfs_time_t t;
 895
 896                         t = cfs_timer_deadline(&target->obd_recovery_timer);
 897                         t = cfs_time_sub(t, cfs_time_current());
 898                         CERROR("%s: denying connection for new client %s (%s): "
 899                                "%d clients in recovery for "CFS_TIME_T"s\n",
 900                                target->obd_name,
 901                                libcfs_nid2str(req->rq_peer.nid), cluuid.uuid,
 902                                atomic_read(&target->obd_lock_replay_clients),
 903                                cfs_duration_sec(t));
 904                         rc = -EBUSY;
 905                 } else {
 906 dont_check_exports:
 907                         rc = obd_connect(req->rq_svc_thread->t_env,
 908                                          &export, target, &cluuid, data,
 909                                          client_nid);
 910                         if (rc == 0)
 911                                 conn.cookie = export->exp_handle.h_cookie;
 912                 }
 913         } else {
 914                 rc = obd_reconnect(req->rq_svc_thread->t_env,
 915                                    export, target, &cluuid, data, client_nid);
 916                 if (rc == 0)
 917                         /* prevous done via class_conn2export */
 918                         class_export_get(export);
 919         }
 920         if (rc)
 921                 GOTO(out, rc);
 922         /* Return only the parts of obd_connect_data that we understand, so the
 923          * client knows that we don't understand the rest. */
 924         if (data) {
 925                  tmpdata = req_capsule_server_get(&req->rq_pill,
 926                                                   &RMF_CONNECT_DATA);
 927                   //data->ocd_connect_flags &= OBD_CONNECT_SUPPORTED;
 928                  *tmpdata = *data;
 929         }
 930
 931         /* If all else goes well, this is our RPC return code. */
 932         req->rq_status = 0;
 933
 934         lustre_msg_set_handle(req->rq_repmsg, &conn);
 935
 936         /* If the client and the server are the same node, we will already
 937          * have an export that really points to the client's DLM export,
 938          * because we have a shared handles table.
 939          *
 940          * XXX this will go away when shaver stops sending the "connect" handle
 941          * in the real "remote handle" field of the request --phik 24 Apr 2003
 942          */
 943         if (req->rq_export != NULL)
 944                 class_export_put(req->rq_export);
 945
 946         req->rq_export = export;
 947
 948         spin_lock(&export->exp_lock);
 949         if (export->exp_conn_cnt >= lustre_msg_get_conn_cnt(req->rq_reqmsg)) {
 950                 spin_unlock(&export->exp_lock);
 951                 CERROR("%s: %s already connected at higher conn_cnt: %d > %d\n",
 952                        cluuid.uuid, libcfs_nid2str(req->rq_peer.nid),
 953                        export->exp_conn_cnt,
 954                        lustre_msg_get_conn_cnt(req->rq_reqmsg));
 955
 956                 GOTO(out, rc = -EALREADY);
 957         }
 958         export->exp_conn_cnt = lustre_msg_get_conn_cnt(req->rq_reqmsg);
 959         export->exp_abort_active_req = 0;
 960
 961         /* request from liblustre?  Don't evict it for not pinging. */
 962         if (lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_LIBCLIENT) {
 963                 export->exp_libclient = 1;
 964                 spin_unlock(&export->exp_lock);
 965
 966                 spin_lock(&target->obd_dev_lock);
 967                 list_del_init(&export->exp_obd_chain_timed);
 968                 spin_unlock(&target->obd_dev_lock);
 969         } else {
 970                 spin_unlock(&export->exp_lock);
 971         }
 972
 973         if (export->exp_connection != NULL) {
 974                 /* Check to see if connection came from another NID */
 975                 if ((export->exp_connection->c_peer.nid != req->rq_peer.nid) &&
 976                     !hlist_unhashed(&export->exp_nid_hash))
 977                         lustre_hash_del(export->exp_obd->obd_nid_hash,
 978                                         &export->exp_connection->c_peer.nid,
 979                                         &export->exp_nid_hash);
 980
 981                 ptlrpc_connection_put(export->exp_connection);
 982         }
 983
 984         export->exp_connection = ptlrpc_connection_get(req->rq_peer,
 985                                                        req->rq_self,
 986                                                        &remote_uuid);
 987         if (hlist_unhashed(&export->exp_nid_hash)) {
 988                 lustre_hash_add_unique(export->exp_obd->obd_nid_hash,
 989                                        &export->exp_connection->c_peer.nid,
 990                                        &export->exp_nid_hash);
 991         }
 992
 993         spin_lock_bh(&target->obd_processing_task_lock);
 994         if (target->obd_recovering && !export->exp_in_recovery) {
 995                 spin_lock(&export->exp_lock);
 996                 export->exp_in_recovery = 1;
 997                 export->exp_req_replay_needed = 1;
 998                 export->exp_lock_replay_needed = 1;
 999                 spin_unlock(&export->exp_lock);
1000                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1001                      && (data->ocd_transno == 0))
1002                         CWARN("Connect with zero transno!\n");
1003
1004                 if ((lustre_msg_get_op_flags(req->rq_reqmsg) & MSG_CONNECT_TRANSNO)
1005                      && data->ocd_transno < target->obd_next_recovery_transno)
1006                         target->obd_next_recovery_transno = data->ocd_transno;
1007                 target->obd_connected_clients++;
1008                 atomic_inc(&target->obd_req_replay_clients);
1009                 atomic_inc(&target->obd_lock_replay_clients);
1010                 if (target->obd_connected_clients ==
1011                     target->obd_max_recoverable_clients)
1012                         cfs_waitq_signal(&target->obd_next_transno_waitq);
1013         }
1014         spin_unlock_bh(&target->obd_processing_task_lock);
1015         tmp = req_capsule_client_get(&req->rq_pill, &RMF_CONN);
1016         conn = *tmp;
1017
1018         if (export->exp_imp_reverse != NULL) {
1019                 /* destroyed import can be still referenced in ctxt */
1020                 obd_set_info_async(export, sizeof(KEY_REVIMP_UPD),
1021                                    KEY_REVIMP_UPD, 0, NULL, NULL);
1022
1023                 client_destroy_import(export->exp_imp_reverse);
1024         }
1025
1026         /* for the rest part, we return -ENOTCONN in case of errors
1027          * in order to let client initialize connection again.
1028          */
1029         revimp = export->exp_imp_reverse = class_new_import(target);
1030         if (!revimp) {
1031                 CERROR("fail to alloc new reverse import.\n");
1032                 GOTO(out, rc = -ENOTCONN);
1033         }
1034
1035         revimp->imp_connection = ptlrpc_connection_addref(export->exp_connection);
1036         revimp->imp_client = &export->exp_obd->obd_ldlm_client;
1037         revimp->imp_remote_handle = conn;
1038         revimp->imp_dlm_fake = 1;
1039         revimp->imp_state = LUSTRE_IMP_FULL;
1040
1041         /* unknown versions will be caught in
1042          * ptlrpc_handle_server_req_in->lustre_unpack_msg() */
1043         revimp->imp_msg_magic = req->rq_reqmsg->lm_magic;
1044
1045         if ((export->exp_connect_flags & OBD_CONNECT_AT) &&
1046             (revimp->imp_msg_magic != LUSTRE_MSG_MAGIC_V1))
1047                 revimp->imp_msghdr_flags |= MSGHDR_AT_SUPPORT;
1048         else
1049                 revimp->imp_msghdr_flags &= ~MSGHDR_AT_SUPPORT;
1050
1051         rc = sptlrpc_import_sec_adapt(revimp, req->rq_svc_ctx, &req->rq_flvr);
1052         if (rc) {
1053                 CERROR("Failed to get sec for reverse import: %d\n", rc);
1054                 export->exp_imp_reverse = NULL;
1055                 class_destroy_import(revimp);
1056         }
1057
1058         class_import_put(revimp);
1059 out:
1060         if (export) {
1061                 spin_lock(&export->exp_lock);
1062                 export->exp_connecting = 0;
1063                 spin_unlock(&export->exp_lock);
1064         }
1065         if (targref)
1066                 class_decref(targref, __FUNCTION__, cfs_current());
1067         if (rc)
1068                 req->rq_status = rc;
1069         RETURN(rc);
1070 }
1071
1072 int target_handle_disconnect(struct ptlrpc_request *req)
1073 {
1074         int rc;
1075         ENTRY;
1076
1077         rc = req_capsule_server_pack(&req->rq_pill);
1078         if (rc)
1079                 RETURN(rc);
1080
1081         /* keep the rq_export around so we can send the reply */
1082         req->rq_status = obd_disconnect(class_export_get(req->rq_export));
1083
1084         RETURN(0);
1085 }
1086
1087 void target_destroy_export(struct obd_export *exp)
1088 {
1089         /* exports created from last_rcvd data, and "fake"
1090            exports created by lctl don't have an import */
1091         if (exp->exp_imp_reverse != NULL)
1092                 client_destroy_import(exp->exp_imp_reverse);
1093
1094         LASSERT(atomic_read(&exp->exp_locks_count) == 0);
1095         LASSERT(atomic_read(&exp->exp_rpc_count) == 0);
1096         LASSERT(atomic_read(&exp->exp_cb_count) == 0);
1097         LASSERT(atomic_read(&exp->exp_replay_count) == 0);
1098 }
1099
1100 /*
1101  * Recovery functions
1102  */
1103 static void target_request_copy_get(struct ptlrpc_request *req)
1104 {
1105         class_export_rpc_get(req->rq_export);
1106         LASSERT(list_empty(&req->rq_list));
1107         CFS_INIT_LIST_HEAD(&req->rq_replay_list);
1108         /* increase refcount to keep request in queue */
1109         LASSERT(atomic_read(&req->rq_refcount));
1110         atomic_inc(&req->rq_refcount);
1111         /** let export know it has replays to be handled */
1112         atomic_inc(&req->rq_export->exp_replay_count);
1113 }
1114
1115 static void target_request_copy_put(struct ptlrpc_request *req)
1116 {
1117         LASSERT(list_empty(&req->rq_replay_list));
1118         LASSERT(atomic_read(&req->rq_export->exp_replay_count) > 0);
1119         atomic_dec(&req->rq_export->exp_replay_count);
1120         class_export_rpc_put(req->rq_export);
1121         ptlrpc_server_drop_request(req);
1122 }
1123
1124 static int target_exp_enqueue_req_replay(struct ptlrpc_request *req)
1125 {
1126         __u64                  transno = lustre_msg_get_transno(req->rq_reqmsg);
1127         struct obd_export     *exp = req->rq_export;
1128         struct ptlrpc_request *reqiter;
1129         int                    dup = 0;
1130
1131         LASSERT(exp);
1132
1133         spin_lock(&exp->exp_lock);
1134         list_for_each_entry(reqiter, &exp->exp_req_replay_queue,
1135                             rq_replay_list) {
1136                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) == transno) {
1137                         dup = 1;
1138                         break;
1139                 }
1140         }
1141
1142         if (dup) {
1143                 /* we expect it with RESENT and REPLAY flags */
1144                 if ((lustre_msg_get_flags(req->rq_reqmsg) &
1145                      (MSG_RESENT | MSG_REPLAY)) != (MSG_RESENT | MSG_REPLAY))
1146                         CERROR("invalid flags %x of resent replay\n",
1147                                lustre_msg_get_flags(req->rq_reqmsg));
1148         } else {
1149                 list_add_tail(&req->rq_replay_list, &exp->exp_req_replay_queue);
1150         }
1151
1152         spin_unlock(&exp->exp_lock);
1153         return dup;
1154 }
1155
1156 static void target_exp_dequeue_req_replay(struct ptlrpc_request *req)
1157 {
1158         LASSERT(!list_empty(&req->rq_replay_list));
1159         LASSERT(req->rq_export);
1160
1161         spin_lock(&req->rq_export->exp_lock);
1162         list_del_init(&req->rq_replay_list);
1163         spin_unlock(&req->rq_export->exp_lock);
1164 }
1165
1166 #ifdef __KERNEL__
1167 static void target_finish_recovery(struct obd_device *obd)
1168 {
1169         ENTRY;
1170         LCONSOLE_INFO("%s: sending delayed replies to recovered clients\n",
1171                       obd->obd_name);
1172
1173         ldlm_reprocess_all_ns(obd->obd_namespace);
1174         spin_lock_bh(&obd->obd_processing_task_lock);
1175         if (!list_empty(&obd->obd_req_replay_queue) ||
1176             !list_empty(&obd->obd_lock_replay_queue) ||
1177             !list_empty(&obd->obd_final_req_queue)) {
1178                 CERROR("%s: Recovery queues ( %s%s%s) are not empty\n",
1179                        obd->obd_name,
1180                        list_empty(&obd->obd_req_replay_queue) ? "" : "req ",
1181                        list_empty(&obd->obd_lock_replay_queue) ? "" : "lock ",
1182                        list_empty(&obd->obd_final_req_queue) ? "" : "final ");
1183                 spin_unlock_bh(&obd->obd_processing_task_lock);
1184                 LBUG();
1185         }
1186         spin_unlock_bh(&obd->obd_processing_task_lock);
1187
1188         obd->obd_recovery_end = cfs_time_current_sec();
1189
1190         /* when recovery finished, cleanup orphans on mds and ost */
1191         if (OBT(obd) && OBP(obd, postrecov)) {
1192                 int rc = OBP(obd, postrecov)(obd);
1193                 if (rc < 0)
1194                         LCONSOLE_WARN("%s: Post recovery failed, rc %d\n",
1195                                       obd->obd_name, rc);
1196         }
1197         EXIT;
1198 }
1199
1200 static void abort_req_replay_queue(struct obd_device *obd)
1201 {
1202         struct ptlrpc_request *req, *n;
1203         struct list_head abort_list;
1204
1205         CFS_INIT_LIST_HEAD(&abort_list);
1206         spin_lock_bh(&obd->obd_processing_task_lock);
1207         list_splice_init(&obd->obd_req_replay_queue, &abort_list);
1208         spin_unlock_bh(&obd->obd_processing_task_lock);
1209         list_for_each_entry_safe(req, n, &abort_list, rq_list) {
1210                 DEBUG_REQ(D_WARNING, req, "aborted:");
1211                 req->rq_status = -ENOTCONN;
1212                 if (ptlrpc_error(req)) {
1213                         DEBUG_REQ(D_ERROR, req,
1214                                   "failed abort_req_reply; skipping");
1215                 }
1216                 target_exp_dequeue_req_replay(req);
1217                 target_request_copy_put(req);
1218         }
1219 }
1220
1221 static void abort_lock_replay_queue(struct obd_device *obd)
1222 {
1223         struct ptlrpc_request *req, *n;
1224         struct list_head abort_list;
1225
1226         CFS_INIT_LIST_HEAD(&abort_list);
1227         spin_lock_bh(&obd->obd_processing_task_lock);
1228         list_splice_init(&obd->obd_lock_replay_queue, &abort_list);
1229         spin_unlock_bh(&obd->obd_processing_task_lock);
1230         list_for_each_entry_safe(req, n, &abort_list, rq_list){
1231                 DEBUG_REQ(D_ERROR, req, "aborted:");
1232                 req->rq_status = -ENOTCONN;
1233                 if (ptlrpc_error(req)) {
1234                         DEBUG_REQ(D_ERROR, req,
1235                                   "failed abort_lock_reply; skipping");
1236                 }
1237                 target_request_copy_put(req);
1238         }
1239 }
1240 #endif
1241
1242 /* Called from a cleanup function if the device is being cleaned up
1243    forcefully.  The exports should all have been disconnected already,
1244    the only thing left to do is
1245      - clear the recovery flags
1246      - cancel the timer
1247      - free queued requests and replies, but don't send replies
1248    Because the obd_stopping flag is set, no new requests should be received.
1249
1250 */
1251 void target_cleanup_recovery(struct obd_device *obd)
1252 {
1253         struct ptlrpc_request *req, *n;
1254         struct list_head clean_list;
1255         ENTRY;
1256
1257         CFS_INIT_LIST_HEAD(&clean_list);
1258         spin_lock_bh(&obd->obd_processing_task_lock);
1259         if (!obd->obd_recovering) {
1260                 spin_unlock_bh(&obd->obd_processing_task_lock);
1261                 EXIT;
1262                 return;
1263         }
1264         obd->obd_recovering = obd->obd_abort_recovery = 0;
1265         target_cancel_recovery_timer(obd);
1266
1267         list_splice_init(&obd->obd_req_replay_queue, &clean_list);
1268         spin_unlock_bh(&obd->obd_processing_task_lock);
1269
1270         list_for_each_entry_safe(req, n, &clean_list, rq_list) {
1271                 LASSERT(req->rq_reply_state == 0);
1272                 target_exp_dequeue_req_replay(req);
1273                 target_request_copy_put(req);
1274         }
1275
1276         spin_lock_bh(&obd->obd_processing_task_lock);
1277         list_splice_init(&obd->obd_lock_replay_queue, &clean_list);
1278         list_splice_init(&obd->obd_final_req_queue, &clean_list);
1279         spin_unlock_bh(&obd->obd_processing_task_lock);
1280
1281         list_for_each_entry_safe(req, n, &clean_list, rq_list){
1282                 LASSERT(req->rq_reply_state == 0);
1283                 target_request_copy_put(req);
1284         }
1285
1286         EXIT;
1287 }
1288
1289 /* obd_processing_task_lock should be held */
1290 void target_cancel_recovery_timer(struct obd_device *obd)
1291 {
1292         CDEBUG(D_HA, "%s: cancel recovery timer\n", obd->obd_name);
1293         cfs_timer_disarm(&obd->obd_recovery_timer);
1294 }
1295
1296 /* extend = 1 means require at least "duration" seconds left in the timer,
1297    extend = 0 means set the total duration (start_recovery_timer) */
1298 static void reset_recovery_timer(struct obd_device *obd, int duration,
1299                                  int extend)
1300 {
1301         cfs_time_t now = cfs_time_current_sec();
1302         cfs_duration_t left;
1303
1304         spin_lock_bh(&obd->obd_processing_task_lock);
1305         if (!obd->obd_recovering || obd->obd_abort_recovery) {
1306                 spin_unlock_bh(&obd->obd_processing_task_lock);
1307                 return;
1308         }
1309
1310         left = cfs_time_sub(obd->obd_recovery_end, now);
1311
1312         if (extend && (duration > left))
1313                 obd->obd_recovery_timeout += duration - left;
1314         else if (!extend && (duration > obd->obd_recovery_timeout))
1315                 /* Track the client's largest expected replay time */
1316                 obd->obd_recovery_timeout = duration;
1317 #ifdef CRAY_XT3
1318         /*
1319          * If total recovery time already exceed the
1320          * obd_recovery_max_time, then CRAY XT3 will
1321          * abort the recovery
1322          */
1323         if(obd->obd_recovery_timeout > obd->obd_recovery_max_time)
1324                 obd->obd_recovery_timeout = obd->obd_recovery_max_time;
1325 #endif
1326         obd->obd_recovery_end = obd->obd_recovery_start +
1327                                 obd->obd_recovery_timeout;
1328         if (!cfs_timer_is_armed(&obd->obd_recovery_timer) ||
1329             cfs_time_before(now, obd->obd_recovery_end)) {
1330                 left = cfs_time_sub(obd->obd_recovery_end, now);
1331                 cfs_timer_arm(&obd->obd_recovery_timer, cfs_time_shift(left));
1332         }
1333         spin_unlock_bh(&obd->obd_processing_task_lock);
1334         CDEBUG(D_HA, "%s: recovery timer will expire in %u seconds\n",
1335                obd->obd_name, (unsigned)left);
1336 }
1337
1338 static void check_and_start_recovery_timer(struct obd_device *obd)
1339 {
1340         spin_lock_bh(&obd->obd_processing_task_lock);
1341         if (cfs_timer_is_armed(&obd->obd_recovery_timer)) {
1342                 spin_unlock_bh(&obd->obd_processing_task_lock);
1343                 return;
1344         }
1345         CDEBUG(D_HA, "%s: starting recovery timer\n", obd->obd_name);
1346         obd->obd_recovery_start = cfs_time_current_sec();
1347         /* minimum */
1348         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1349         spin_unlock_bh(&obd->obd_processing_task_lock);
1350
1351         reset_recovery_timer(obd, obd->obd_recovery_timeout, 0);
1352 }
1353
1354 /* Reset the timer with each new client connection */
1355 /*
1356  * This timer is actually reconnect_timer, which is for making sure
1357  * the total recovery window is at least as big as my reconnect
1358  * attempt timing. So the initial recovery time_out will be set to
1359  * OBD_RECOVERY_FACTOR * obd_timeout. If the timeout coming
1360  * from client is bigger than this, then the recovery time_out will
1361  * be extended to make sure the client could be reconnected, in the
1362  * process, the timeout from the new client should be ignored.
1363  */
1364
1365 static void
1366 target_start_and_reset_recovery_timer(struct obd_device *obd,
1367                                       struct ptlrpc_request *req,
1368                                       int new_client)
1369 {
1370         int service_time = lustre_msg_get_service_time(req->rq_reqmsg);
1371
1372         if (!new_client && service_time)
1373                 /* Teach server about old server's estimates, as first guess
1374                  * at how long new requests will take. */
1375                 at_add(&req->rq_rqbd->rqbd_service->srv_at_estimate,
1376                        service_time);
1377
1378         check_and_start_recovery_timer(obd);
1379
1380         /* convert the service time to rpc timeout,
1381          * reuse service_time to limit stack usage */
1382         service_time = at_est2timeout(service_time);
1383
1384         /* We expect other clients to timeout within service_time, then try
1385          * to reconnect, then try the failover server.  The max delay between
1386          * connect attempts is SWITCH_MAX + SWITCH_INC + INITIAL */
1387         service_time += 2 * (CONNECTION_SWITCH_MAX + CONNECTION_SWITCH_INC +
1388                              INITIAL_CONNECT_TIMEOUT);
1389         if (service_time > obd->obd_recovery_timeout && !new_client)
1390                 reset_recovery_timer(obd, service_time, 0);
1391 }
1392
1393 #ifdef __KERNEL__
1394
1395 /** Health checking routines */
1396 static inline int exp_connect_healthy(struct obd_export *exp)
1397 {
1398         return (exp->exp_in_recovery);
1399 }
1400
1401 /** if export done req_replay or has replay in queue */
1402 static inline int exp_req_replay_healthy(struct obd_export *exp)
1403 {
1404         return (!exp->exp_req_replay_needed ||
1405                 atomic_read(&exp->exp_replay_count) > 0);
1406 }
1407 /** if export done lock_replay or has replay in queue */
1408 static inline int exp_lock_replay_healthy(struct obd_export *exp)
1409 {
1410         return (!exp->exp_lock_replay_needed ||
1411                 atomic_read(&exp->exp_replay_count) > 0);
1412 }
1413
1414 static inline int exp_vbr_healthy(struct obd_export *exp)
1415 {
1416         return (!exp->exp_vbr_failed);
1417 }
1418
1419 static inline int exp_finished(struct obd_export *exp)
1420 {
1421         return (exp->exp_in_recovery && !exp->exp_lock_replay_needed);
1422 }
1423
1424 /** Checking routines for recovery */
1425 static int check_for_clients(struct obd_device *obd)
1426 {
1427         if (obd->obd_abort_recovery || obd->obd_recovery_expired)
1428                 return 1;
1429         LASSERT(obd->obd_connected_clients <= obd->obd_max_recoverable_clients);
1430         if (obd->obd_no_conn == 0 &&
1431             obd->obd_connected_clients + obd->obd_stale_clients ==
1432             obd->obd_max_recoverable_clients)
1433                 return 1;
1434         return 0;
1435 }
1436
1437 static int check_for_next_transno(struct obd_device *obd)
1438 {
1439         struct ptlrpc_request *req = NULL;
1440         int wake_up = 0, connected, completed, queue_len;
1441         __u64 next_transno, req_transno;
1442         ENTRY;
1443         spin_lock_bh(&obd->obd_processing_task_lock);
1444
1445         if (!list_empty(&obd->obd_req_replay_queue)) {
1446                 req = list_entry(obd->obd_req_replay_queue.next,
1447                                  struct ptlrpc_request, rq_list);
1448                 req_transno = lustre_msg_get_transno(req->rq_reqmsg);
1449         } else {
1450                 req_transno = 0;
1451         }
1452
1453         connected = obd->obd_connected_clients;
1454         completed = connected - atomic_read(&obd->obd_req_replay_clients);
1455         queue_len = obd->obd_requests_queued_for_recovery;
1456         next_transno = obd->obd_next_recovery_transno;
1457
1458         CDEBUG(D_HA, "max: %d, connected: %d, completed: %d, queue_len: %d, "
1459                "req_transno: "LPU64", next_transno: "LPU64"\n",
1460                obd->obd_max_recoverable_clients, connected, completed,
1461                queue_len, req_transno, next_transno);
1462
1463         if (obd->obd_abort_recovery) {
1464                 CDEBUG(D_HA, "waking for aborted recovery\n");
1465                 wake_up = 1;
1466         } else if (obd->obd_recovery_expired) {
1467                 CDEBUG(D_HA, "waking for expired recovery\n");
1468                 wake_up = 1;
1469         } else if (atomic_read(&obd->obd_req_replay_clients) == 0) {
1470                 CDEBUG(D_HA, "waking for completed recovery\n");
1471                 wake_up = 1;
1472         } else if (req_transno == next_transno) {
1473                 CDEBUG(D_HA, "waking for next ("LPD64")\n", next_transno);
1474                 wake_up = 1;
1475         } else if (queue_len == atomic_read(&obd->obd_req_replay_clients)) {
1476                 int d_lvl = D_HA;
1477                 /** handle gaps occured due to lost reply or VBR */
1478                 LASSERTF(req_transno >= next_transno,
1479                          "req_transno: "LPU64", next_transno: "LPU64"\n",
1480                          req_transno, next_transno);
1481                 if (req_transno > obd->obd_last_committed &&
1482                     !obd->obd_version_recov)
1483                         d_lvl = D_ERROR;
1484                 CDEBUG(d_lvl,
1485                        "%s: waking for gap in transno, VBR is %s (skip: "
1486                        LPD64", ql: %d, comp: %d, conn: %d, next: "LPD64
1487                        ", last_committed: "LPD64")\n",
1488                        obd->obd_name, obd->obd_version_recov ? "ON" : "OFF",
1489                        next_transno, queue_len, completed, connected,
1490                        req_transno, obd->obd_last_committed);
1491                 obd->obd_next_recovery_transno = req_transno;
1492                 wake_up = 1;
1493         } else if (OBD_FAIL_CHECK(OBD_FAIL_MDS_RECOVERY_ACCEPTS_GAPS)) {
1494                 CDEBUG(D_HA, "accepting transno gaps is explicitly allowed"
1495                        " by fail_lock, waking up ("LPD64")\n", next_transno);
1496                 obd->obd_next_recovery_transno = req_transno;
1497                 wake_up = 1;
1498         }
1499         spin_unlock_bh(&obd->obd_processing_task_lock);
1500         return wake_up;
1501 }
1502
1503 static int check_for_next_lock(struct obd_device *obd)
1504 {
1505         int wake_up = 0;
1506
1507         spin_lock_bh(&obd->obd_processing_task_lock);
1508         if (!list_empty(&obd->obd_lock_replay_queue)) {
1509                 CDEBUG(D_HA, "waking for next lock\n");
1510                 wake_up = 1;
1511         } else if (atomic_read(&obd->obd_lock_replay_clients) == 0) {
1512                 CDEBUG(D_HA, "waking for completed lock replay\n");
1513                 wake_up = 1;
1514         } else if (obd->obd_abort_recovery) {
1515                 CDEBUG(D_HA, "waking for aborted recovery\n");
1516                 wake_up = 1;
1517         } else if (obd->obd_recovery_expired) {
1518                 CDEBUG(D_HA, "waking for expired recovery\n");
1519                 wake_up = 1;
1520         }
1521         spin_unlock_bh(&obd->obd_processing_task_lock);
1522
1523         return wake_up;
1524 }
1525
1526 /**
1527  * wait for recovery events,
1528  * check its status with help of check_routine
1529  * evict dead clients via health_check
1530  */
1531 static int target_recovery_overseer(struct obd_device *obd,
1532                                     int (*check_routine)(struct obd_device *),
1533                                     int (*health_check)(struct obd_export *))
1534 {
1535         int abort = 0, expired = 1;
1536
1537         do {
1538                 cfs_wait_event(obd->obd_next_transno_waitq, check_routine(obd));
1539                 spin_lock_bh(&obd->obd_processing_task_lock);
1540                 abort = obd->obd_abort_recovery;
1541                 expired = obd->obd_recovery_expired;
1542                 obd->obd_recovery_expired = 0;
1543                 spin_unlock_bh(&obd->obd_processing_task_lock);
1544                 if (abort) {
1545                         CWARN("recovery is aborted, evict exports in recovery\n");
1546                         /** evict exports which didn't finish recovery yet */
1547                         class_disconnect_stale_exports(obd, exp_finished);
1548                 } else if (expired) {
1549                         /** If some clients died being recovered, evict them */
1550                         CDEBUG(D_WARNING, "recovery is timed out, evict stale exports\n");
1551                         /** evict cexports with no replay in queue, they are stalled */
1552                         class_disconnect_stale_exports(obd, health_check);
1553                         /** continue with VBR */
1554                         spin_lock_bh(&obd->obd_processing_task_lock);
1555                         obd->obd_version_recov = 1;
1556                         spin_unlock_bh(&obd->obd_processing_task_lock);
1557                         /**
1558                          * reset timer, recovery will proceed with versions now,
1559                          * timeout is set just to handle reconnection delays
1560                          */
1561                         reset_recovery_timer(obd, RECONNECT_DELAY_MAX, 1);
1562                         /** Wait for recovery events again, after evicting bad clients */
1563                 }
1564         } while (!abort && expired);
1565
1566         return abort;
1567 }
1568
1569 static struct ptlrpc_request *target_next_replay_req(struct obd_device *obd)
1570 {
1571         struct ptlrpc_request *req = NULL;
1572         ENTRY;
1573
1574         CDEBUG(D_HA, "Waiting for transno "LPD64"\n",
1575                obd->obd_next_recovery_transno);
1576
1577         if (target_recovery_overseer(obd, check_for_next_transno,
1578                                      exp_req_replay_healthy)) {
1579                 abort_req_replay_queue(obd);
1580                 abort_lock_replay_queue(obd);
1581         }
1582
1583         spin_lock_bh(&obd->obd_processing_task_lock);
1584         if (!list_empty(&obd->obd_req_replay_queue)) {
1585                 req = list_entry(obd->obd_req_replay_queue.next,
1586                                  struct ptlrpc_request, rq_list);
1587                 list_del_init(&req->rq_list);
1588                 obd->obd_requests_queued_for_recovery--;
1589                 spin_unlock_bh(&obd->obd_processing_task_lock);
1590         } else {
1591                 spin_unlock_bh(&obd->obd_processing_task_lock);
1592                 LASSERT(list_empty(&obd->obd_req_replay_queue));
1593                 LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0);
1594                 /** evict exports failed VBR */
1595                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1596         }
1597         RETURN(req);
1598 }
1599
1600 static struct ptlrpc_request *target_next_replay_lock(struct obd_device *obd)
1601 {
1602         struct ptlrpc_request *req = NULL;
1603
1604         CDEBUG(D_HA, "Waiting for lock\n");
1605         if (target_recovery_overseer(obd, check_for_next_lock,
1606                                      exp_lock_replay_healthy))
1607                 abort_lock_replay_queue(obd);
1608
1609         spin_lock_bh(&obd->obd_processing_task_lock);
1610         if (!list_empty(&obd->obd_lock_replay_queue)) {
1611                 req = list_entry(obd->obd_lock_replay_queue.next,
1612                                  struct ptlrpc_request, rq_list);
1613                 list_del_init(&req->rq_list);
1614                 spin_unlock_bh(&obd->obd_processing_task_lock);
1615         } else {
1616                 spin_unlock_bh(&obd->obd_processing_task_lock);
1617                 LASSERT(list_empty(&obd->obd_lock_replay_queue));
1618                 LASSERT(atomic_read(&obd->obd_lock_replay_clients) == 0);
1619                 /** evict exports failed VBR */
1620                 class_disconnect_stale_exports(obd, exp_vbr_healthy);
1621         }
1622         return req;
1623 }
1624
1625 static struct ptlrpc_request *target_next_final_ping(struct obd_device *obd)
1626 {
1627         struct ptlrpc_request *req = NULL;
1628
1629         spin_lock_bh(&obd->obd_processing_task_lock);
1630         if (!list_empty(&obd->obd_final_req_queue)) {
1631                 req = list_entry(obd->obd_final_req_queue.next,
1632                                  struct ptlrpc_request, rq_list);
1633                 list_del_init(&req->rq_list);
1634                 if (req->rq_export->exp_in_recovery) {
1635                         spin_lock(&req->rq_export->exp_lock);
1636                         req->rq_export->exp_in_recovery = 0;
1637                         spin_unlock(&req->rq_export->exp_lock);
1638                 }
1639         }
1640         spin_unlock_bh(&obd->obd_processing_task_lock);
1641         return req;
1642 }
1643
1644 static int handle_recovery_req(struct ptlrpc_thread *thread,
1645                                struct ptlrpc_request *req,
1646                                svc_handler_t handler)
1647 {
1648         int rc;
1649         ENTRY;
1650
1651         rc = lu_context_init(&req->rq_recov_session, LCT_SESSION);
1652         if (rc) {
1653                 CERROR("Failure to initialize session: %d\n", rc);
1654                 GOTO(reqcopy_put, rc);
1655         }
1656         /**
1657          * export can be evicted during recovery, no need to handle replays for
1658          * it after that, discard such request silently
1659          */
1660         if (req->rq_export->exp_disconnected)
1661                 GOTO(reqcopy_put, rc);
1662
1663         req->rq_recov_session.lc_thread = thread;
1664         lu_context_enter(&req->rq_recov_session);
1665         req->rq_svc_thread = thread;
1666         req->rq_svc_thread->t_env->le_ses = &req->rq_recov_session;
1667
1668         /* thread context */
1669         lu_context_enter(&thread->t_env->le_ctx);
1670         (void)handler(req);
1671         lu_context_exit(&thread->t_env->le_ctx);
1672
1673         lu_context_exit(&req->rq_recov_session);
1674         lu_context_fini(&req->rq_recov_session);
1675         /* don't reset timer for final stage */
1676         if (!exp_finished(req->rq_export)) {
1677                 /**
1678                  * Add request timeout to the recovery time so next request from
1679                  * this client may come in recovery time
1680                  */
1681                  reset_recovery_timer(class_exp2obd(req->rq_export),
1682                                       AT_OFF ? obd_timeout :
1683                                       lustre_msg_get_timeout(req->rq_reqmsg), 1);
1684         }
1685         /**
1686          * bz18031: increase next_recovery_transno before target_request_copy_put()
1687          * will drop exp_rpc reference
1688          */
1689         if (req->rq_export->exp_req_replay_needed) {
1690                 spin_lock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1691                 req->rq_export->exp_obd->obd_next_recovery_transno++;
1692                 spin_unlock_bh(&req->rq_export->exp_obd->obd_processing_task_lock);
1693         }
1694 reqcopy_put:
1695         if (req->rq_export->exp_req_replay_needed)
1696                 target_exp_dequeue_req_replay(req);
1697         target_request_copy_put(req);
1698         RETURN(0);
1699 }
1700
1701 static int target_recovery_thread(void *arg)
1702 {
1703         struct lu_target *lut = arg;
1704         struct obd_device *obd = lut->lut_obd;
1705         struct ptlrpc_request *req;
1706         struct target_recovery_data *trd = &obd->obd_recovery_data;
1707         unsigned long delta;
1708         unsigned long flags;
1709         struct lu_env env;
1710         struct ptlrpc_thread fake_svc_thread, *thread = &fake_svc_thread;
1711         int rc = 0;
1712         ENTRY;
1713
1714         cfs_daemonize("tgt_recov");
1715
1716         SIGNAL_MASK_LOCK(current, flags);
1717         sigfillset(&current->blocked);
1718         RECALC_SIGPENDING;
1719         SIGNAL_MASK_UNLOCK(current, flags);
1720
1721         rc = lu_context_init(&env.le_ctx, LCT_MD_THREAD);
1722         if (rc)
1723                 RETURN(rc);
1724
1725         thread->t_env = &env;
1726         env.le_ctx.lc_thread = thread;
1727
1728         CERROR("%s: started recovery thread pid %d\n", obd->obd_name,
1729                cfs_curproc_pid());
1730         trd->trd_processing_task = cfs_curproc_pid();
1731
1732         obd->obd_recovering = 1;
1733         complete(&trd->trd_starting);
1734
1735         /* first of all, we have to know the first transno to replay */
1736         if (target_recovery_overseer(obd, check_for_clients,
1737                                      exp_connect_healthy)) {
1738                 abort_req_replay_queue(obd);
1739                 abort_lock_replay_queue(obd);
1740         }
1741
1742         /* next stage: replay requests */
1743         delta = jiffies;
1744         obd->obd_req_replaying = 1;
1745         CDEBUG(D_INFO, "1: request replay stage - %d clients from t"LPU64"\n",
1746                atomic_read(&obd->obd_req_replay_clients),
1747                obd->obd_next_recovery_transno);
1748         while ((req = target_next_replay_req(obd))) {
1749                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1750                 DEBUG_REQ(D_HA, req, "processing t"LPD64" from %s",
1751                           lustre_msg_get_transno(req->rq_reqmsg),
1752                           libcfs_nid2str(req->rq_peer.nid));
1753                 handle_recovery_req(thread, req,
1754                                     trd->trd_recovery_handler);
1755                 obd->obd_replayed_requests++;
1756         }
1757
1758         /**
1759          * The second stage: replay locks
1760          */
1761         CDEBUG(D_INFO, "2: lock replay stage - %d clients\n",
1762                atomic_read(&obd->obd_lock_replay_clients));
1763         while ((req = target_next_replay_lock(obd))) {
1764                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1765                 DEBUG_REQ(D_HA, req, "processing lock from %s: ",
1766                           libcfs_nid2str(req->rq_peer.nid));
1767                 handle_recovery_req(thread, req,
1768                                     trd->trd_recovery_handler);
1769                 obd->obd_replayed_locks++;
1770         }
1771
1772         /**
1773          * The third stage: reply on final pings, at this moment all clients
1774          * must have request in final queue
1775          */
1776         CDEBUG(D_INFO, "3: final stage - process recovery completion pings\n");
1777         /** Update server last boot epoch */
1778         lut_boot_epoch_update(lut);
1779         /* We drop recoverying flag to forward all new requests
1780          * to regular mds_handle() since now */
1781         spin_lock_bh(&obd->obd_processing_task_lock);
1782         obd->obd_recovering = obd->obd_abort_recovery = 0;
1783         target_cancel_recovery_timer(obd);
1784         spin_unlock_bh(&obd->obd_processing_task_lock);
1785         while ((req = target_next_final_ping(obd))) {
1786                 LASSERT(trd->trd_processing_task == cfs_curproc_pid());
1787                 DEBUG_REQ(D_HA, req, "processing final ping from %s: ",
1788                           libcfs_nid2str(req->rq_peer.nid));
1789                 handle_recovery_req(thread, req,
1790                                     trd->trd_recovery_handler);
1791         }
1792
1793         delta = (jiffies - delta) / HZ;
1794         CDEBUG(D_INFO,"4: recovery completed in %lus - %d/%d reqs/locks\n",
1795               delta, obd->obd_replayed_requests, obd->obd_replayed_locks);
1796         if (delta > obd_timeout * OBD_RECOVERY_FACTOR) {
1797                 CWARN("too long recovery - read logs\n");
1798                 libcfs_debug_dumplog();
1799         }
1800
1801         target_finish_recovery(obd);
1802
1803         lu_context_fini(&env.le_ctx);
1804         trd->trd_processing_task = 0;
1805         complete(&trd->trd_finishing);
1806         RETURN(rc);
1807 }
1808
1809 static int target_start_recovery_thread(struct lu_target *lut,
1810                                         svc_handler_t handler)
1811 {
1812         struct obd_device *obd = lut->lut_obd;
1813         int rc = 0;
1814         struct target_recovery_data *trd = &obd->obd_recovery_data;
1815
1816         memset(trd, 0, sizeof(*trd));
1817         init_completion(&trd->trd_starting);
1818         init_completion(&trd->trd_finishing);
1819         trd->trd_recovery_handler = handler;
1820
1821         if (kernel_thread(target_recovery_thread, lut, 0) > 0) {
1822                 wait_for_completion(&trd->trd_starting);
1823                 LASSERT(obd->obd_recovering != 0);
1824         } else
1825                 rc = -ECHILD;
1826
1827         return rc;
1828 }
1829
1830 void target_stop_recovery_thread(struct obd_device *obd)
1831 {
1832         spin_lock_bh(&obd->obd_processing_task_lock);
1833         if (obd->obd_recovery_data.trd_processing_task > 0) {
1834                 struct target_recovery_data *trd = &obd->obd_recovery_data;
1835                 /** recovery can be done but postrecovery is not yet */
1836                 if (obd->obd_recovering) {
1837                         CERROR("%s: Aborting recovery\n", obd->obd_name);
1838                         obd->obd_abort_recovery = 1;
1839                         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1840                 }
1841                 spin_unlock_bh(&obd->obd_processing_task_lock);
1842                 wait_for_completion(&trd->trd_finishing);
1843         } else {
1844                 spin_unlock_bh(&obd->obd_processing_task_lock);
1845         }
1846 }
1847
1848 void target_recovery_fini(struct obd_device *obd)
1849 {
1850         class_disconnect_exports(obd);
1851         target_stop_recovery_thread(obd);
1852         target_cleanup_recovery(obd);
1853 }
1854 EXPORT_SYMBOL(target_recovery_fini);
1855
1856 static void target_recovery_expired(unsigned long castmeharder)
1857 {
1858         struct obd_device *obd = (struct obd_device *)castmeharder;
1859         CDEBUG(D_HA, "%s: recovery timed out; %d clients are still in recovery"
1860                " after %lds (%d clients connected)\n",
1861                obd->obd_name, atomic_read(&obd->obd_lock_replay_clients),
1862                cfs_time_current_sec()- obd->obd_recovery_start,
1863                obd->obd_connected_clients);
1864
1865         spin_lock_bh(&obd->obd_processing_task_lock);
1866         obd->obd_recovery_expired = 1;
1867         cfs_waitq_signal(&obd->obd_next_transno_waitq);
1868         spin_unlock_bh(&obd->obd_processing_task_lock);
1869 }
1870
1871 void target_recovery_init(struct lu_target *lut, svc_handler_t handler)
1872 {
1873         struct obd_device *obd = lut->lut_obd;
1874         if (obd->obd_max_recoverable_clients == 0) {
1875                 /** Update server last boot epoch */
1876                 lut_boot_epoch_update(lut);
1877                 return;
1878         }
1879
1880         CWARN("RECOVERY: service %s, %d recoverable clients, "
1881               "last_transno "LPU64"\n", obd->obd_name,
1882               obd->obd_max_recoverable_clients, obd->obd_last_committed);
1883         LASSERT(obd->obd_stopping == 0);
1884         obd->obd_next_recovery_transno = obd->obd_last_committed + 1;
1885         obd->obd_recovery_start = 0;
1886         obd->obd_recovery_end = 0;
1887         obd->obd_recovery_timeout = OBD_RECOVERY_FACTOR * obd_timeout;
1888         /* bz13079: this should be set to desired value for ost but not for mds */
1889         obd->obd_recovery_max_time = OBD_RECOVERY_MAX_TIME;
1890         cfs_timer_init(&obd->obd_recovery_timer, target_recovery_expired, obd);
1891         target_start_recovery_thread(lut, handler);
1892 }
1893 EXPORT_SYMBOL(target_recovery_init);
1894
1895 #endif
1896
1897 static int target_process_req_flags(struct obd_device *obd,
1898                                     struct ptlrpc_request *req)
1899 {
1900         struct obd_export *exp = req->rq_export;
1901         LASSERT(exp != NULL);
1902         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1903                 /* client declares he's ready to replay locks */
1904                 spin_lock_bh(&obd->obd_processing_task_lock);
1905                 if (exp->exp_req_replay_needed) {
1906                         LASSERT(atomic_read(&obd->obd_req_replay_clients) > 0);
1907                         spin_lock(&exp->exp_lock);
1908                         exp->exp_req_replay_needed = 0;
1909                         spin_unlock(&exp->exp_lock);
1910                         atomic_dec(&obd->obd_req_replay_clients);
1911                 }
1912                 spin_unlock_bh(&obd->obd_processing_task_lock);
1913         }
1914         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1915                 /* client declares he's ready to complete recovery
1916                  * so, we put the request on th final queue */
1917                 spin_lock_bh(&obd->obd_processing_task_lock);
1918                 if (exp->exp_lock_replay_needed) {
1919                         LASSERT(atomic_read(&obd->obd_lock_replay_clients) > 0);
1920                         spin_lock(&exp->exp_lock);
1921                         exp->exp_lock_replay_needed = 0;
1922                         spin_unlock(&exp->exp_lock);
1923                         atomic_dec(&obd->obd_lock_replay_clients);
1924                 }
1925                 spin_unlock_bh(&obd->obd_processing_task_lock);
1926         }
1927
1928         return 0;
1929 }
1930
1931 int target_queue_recovery_request(struct ptlrpc_request *req,
1932                                   struct obd_device *obd)
1933 {
1934         struct list_head *tmp;
1935         int inserted = 0;
1936         __u64 transno = lustre_msg_get_transno(req->rq_reqmsg);
1937         ENTRY;
1938
1939         if (obd->obd_recovery_data.trd_processing_task == cfs_curproc_pid()) {
1940                 /* Processing the queue right now, don't re-add. */
1941                 RETURN(1);
1942         }
1943
1944         target_process_req_flags(obd, req);
1945
1946         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_LOCK_REPLAY_DONE) {
1947                 /* client declares he's ready to complete recovery
1948                  * so, we put the request on th final queue */
1949                 target_request_copy_get(req);
1950                 DEBUG_REQ(D_HA, req, "queue final req");
1951                 spin_lock_bh(&obd->obd_processing_task_lock);
1952                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1953                 if (obd->obd_recovering) {
1954                         list_add_tail(&req->rq_list, &obd->obd_final_req_queue);
1955                 } else {
1956                         spin_unlock_bh(&obd->obd_processing_task_lock);
1957                         target_request_copy_put(req);
1958                         RETURN(obd->obd_stopping ? -ENOTCONN : 1);
1959                 }
1960                 spin_unlock_bh(&obd->obd_processing_task_lock);
1961                 RETURN(0);
1962         }
1963         if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REQ_REPLAY_DONE) {
1964                 /* client declares he's ready to replay locks */
1965                 target_request_copy_get(req);
1966                 DEBUG_REQ(D_HA, req, "queue lock replay req");
1967                 spin_lock_bh(&obd->obd_processing_task_lock);
1968                 cfs_waitq_signal(&obd->obd_next_transno_waitq);
1969                 LASSERT(obd->obd_recovering);
1970                 /* usually due to recovery abort */
1971                 if (!req->rq_export->exp_in_recovery) {
1972                         spin_unlock_bh(&obd->obd_processing_task_lock);
1973                         target_request_copy_put(req);
1974                         RETURN(-ENOTCONN);
1975                 }
1976                 LASSERT(req->rq_export->exp_lock_replay_needed);
1977                 list_add_tail(&req->rq_list, &obd->obd_lock_replay_queue);
1978                 spin_unlock_bh(&obd->obd_processing_task_lock);
1979                 RETURN(0);
1980         }
1981
1982         /* CAVEAT EMPTOR: The incoming request message has been swabbed
1983          * (i.e. buflens etc are in my own byte order), but type-dependent
1984          * buffers (eg mds_body, ost_body etc) have NOT been swabbed. */
1985
1986         if (!transno) {
1987                 CFS_INIT_LIST_HEAD(&req->rq_list);
1988                 DEBUG_REQ(D_HA, req, "not queueing");
1989                 RETURN(1);
1990         }
1991
1992         spin_lock_bh(&obd->obd_processing_task_lock);
1993
1994         /* If we're processing the queue, we want don't want to queue this
1995          * message.
1996          *
1997          * Also, if this request has a transno less than the one we're waiting
1998          * for, we should process it now.  It could (and currently always will)
1999          * be an open request for a descriptor that was opened some time ago.
2000          *
2001          * Also, a resent, replayed request that has already been
2002          * handled will pass through here and be processed immediately.
2003          */
2004         CWARN("Next recovery transno: "LPU64", current: "LPU64", replaying: %i\n",
2005               obd->obd_next_recovery_transno, transno, obd->obd_req_replaying);
2006         if (transno < obd->obd_next_recovery_transno && obd->obd_req_replaying) {
2007                 /* Processing the queue right now, don't re-add. */
2008                 LASSERT(list_empty(&req->rq_list));
2009                 spin_unlock_bh(&obd->obd_processing_task_lock);
2010                 RETURN(1);
2011         }
2012         spin_unlock_bh(&obd->obd_processing_task_lock);
2013
2014         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_REPLAY_DROP))
2015                 RETURN(0);
2016
2017         target_request_copy_get(req);
2018         spin_lock_bh(&obd->obd_processing_task_lock);
2019         LASSERT(obd->obd_recovering);
2020         if (!req->rq_export->exp_in_recovery) {
2021                 spin_unlock_bh(&obd->obd_processing_task_lock);
2022                 target_request_copy_put(req);
2023                 RETURN(-ENOTCONN);
2024         }
2025         LASSERT(req->rq_export->exp_req_replay_needed);
2026
2027         if (target_exp_enqueue_req_replay(req)) {
2028                 spin_unlock_bh(&obd->obd_processing_task_lock);
2029                 DEBUG_REQ(D_ERROR, req, "dropping resent queued req");
2030                 target_request_copy_put(req);
2031                 RETURN(0);
2032         }
2033
2034         /* XXX O(n^2) */
2035         list_for_each(tmp, &obd->obd_req_replay_queue) {
2036                 struct ptlrpc_request *reqiter =
2037                         list_entry(tmp, struct ptlrpc_request, rq_list);
2038
2039                 if (lustre_msg_get_transno(reqiter->rq_reqmsg) > transno) {
2040                         list_add_tail(&req->rq_list, &reqiter->rq_list);
2041                         inserted = 1;
2042                         break;
2043                 }
2044
2045                 if (unlikely(lustre_msg_get_transno(reqiter->rq_reqmsg) ==
2046                              transno)) {
2047                         DEBUG_REQ(D_ERROR, req, "dropping replay: transno "
2048                                   "has been claimed by another client");
2049                         spin_unlock_bh(&obd->obd_processing_task_lock);
2050                         target_exp_dequeue_req_replay(req);
2051                         target_request_copy_put(req);
2052                         RETURN(0);
2053                 }
2054         }
2055
2056         if (!inserted)
2057                 list_add_tail(&req->rq_list, &obd->obd_req_replay_queue);
2058
2059         obd->obd_requests_queued_for_recovery++;
2060         cfs_waitq_signal(&obd->obd_next_transno_waitq);
2061         spin_unlock_bh(&obd->obd_processing_task_lock);
2062         RETURN(0);
2063 }
2064
2065 /**
2066  * Packs current SLV and Limit into \a req.
2067  */
2068 int target_pack_pool_reply(struct ptlrpc_request *req)
2069 {
2070         struct obd_device *obd;
2071         ENTRY;
2072
2073         /*
2074          * Check that we still have all structures alive as this may
2075          * be some late rpc in shutdown time.
2076          */
2077         if (unlikely(!req->rq_export || !req->rq_export->exp_obd ||
2078                      !exp_connect_lru_resize(req->rq_export))) {
2079                 lustre_msg_set_slv(req->rq_repmsg, 0);
2080                 lustre_msg_set_limit(req->rq_repmsg, 0);
2081                 RETURN(0);
2082         }
2083
2084         /*
2085          * OBD is alive here as export is alive, which we checked above.
2086          */
2087         obd = req->rq_export->exp_obd;
2088
2089         read_lock(&obd->obd_pool_lock);
2090         lustre_msg_set_slv(req->rq_repmsg, obd->obd_pool_slv);
2091         lustre_msg_set_limit(req->rq_repmsg, obd->obd_pool_limit);
2092         read_unlock(&obd->obd_pool_lock);
2093
2094         RETURN(0);
2095 }
2096
2097 int target_send_reply_msg(struct ptlrpc_request *req, int rc, int fail_id)
2098 {
2099         if (OBD_FAIL_CHECK_ORSET(fail_id & ~OBD_FAIL_ONCE, OBD_FAIL_ONCE)) {
2100                 DEBUG_REQ(D_ERROR, req, "dropping reply");
2101                 return (-ECOMM);
2102         }
2103
2104         if (unlikely(rc)) {
2105                 DEBUG_REQ(D_ERROR, req, "processing error (%d)", rc);
2106                 req->rq_status = rc;
2107                 return (ptlrpc_send_error(req, 1));
2108         } else {
2109                 DEBUG_REQ(D_NET, req, "sending reply");
2110         }
2111
2112         return (ptlrpc_send_reply(req, PTLRPC_REPLY_MAYBE_DIFFICULT));
2113 }
2114
2115 void target_send_reply(struct ptlrpc_request *req, int rc, int fail_id)
2116 {
2117         int                        netrc;
2118         struct ptlrpc_reply_state *rs;
2119         struct obd_device         *obd;
2120         struct obd_export         *exp;
2121         struct ptlrpc_service     *svc;
2122         ENTRY;
2123
2124         if (req->rq_no_reply) {
2125                 EXIT;
2126                 return;
2127         }
2128
2129         svc = req->rq_rqbd->rqbd_service;
2130         rs = req->rq_reply_state;
2131         if (rs == NULL || !rs->rs_difficult) {
2132                 /* no notifiers */
2133                 target_send_reply_msg (req, rc, fail_id);
2134                 EXIT;
2135                 return;
2136         }
2137
2138         /* must be an export if locks saved */
2139         LASSERT (req->rq_export != NULL);
2140         /* req/reply consistent */
2141         LASSERT (rs->rs_service == svc);
2142
2143         /* "fresh" reply */
2144         LASSERT (!rs->rs_scheduled);
2145         LASSERT (!rs->rs_scheduled_ever);
2146         LASSERT (!rs->rs_handled);
2147         LASSERT (!rs->rs_on_net);
2148         LASSERT (rs->rs_export == NULL);
2149         LASSERT (list_empty(&rs->rs_obd_list));
2150         LASSERT (list_empty(&rs->rs_exp_list));
2151
2152         exp = class_export_get (req->rq_export);
2153         obd = exp->exp_obd;
2154
2155         /* disable reply scheduling while I'm setting up */
2156         rs->rs_scheduled = 1;
2157         rs->rs_on_net    = 1;
2158         rs->rs_xid       = req->rq_xid;
2159         rs->rs_transno   = req->rq_transno;
2160         rs->rs_export    = exp;
2161         rs->rs_opc       = lustre_msg_get_opc(rs->rs_msg);
2162
2163         spin_lock(&exp->exp_uncommitted_replies_lock);
2164         CDEBUG(D_NET, "rs transno = "LPU64", last committed = "LPU64"\n",
2165                rs->rs_transno, exp->exp_last_committed);
2166         if (rs->rs_transno > exp->exp_last_committed) {
2167                 /* not committed already */
2168                 list_add_tail(&rs->rs_obd_list,
2169                               &exp->exp_uncommitted_replies);
2170         }
2171         spin_unlock (&exp->exp_uncommitted_replies_lock);
2172
2173         spin_lock(&exp->exp_lock);
2174         list_add_tail(&rs->rs_exp_list, &exp->exp_outstanding_replies);
2175         spin_unlock(&exp->exp_lock);
2176
2177         netrc = target_send_reply_msg (req, rc, fail_id);
2178
2179         spin_lock(&svc->srv_lock);
2180
2181         atomic_inc(&svc->srv_n_difficult_replies);
2182
2183         if (netrc != 0) {
2184                 /* error sending: reply is off the net.  Also we need +1
2185                  * reply ref until ptlrpc_handle_rs() is done
2186                  * with the reply state (if the send was successful, there
2187                  * would have been +1 ref for the net, which
2188                  * reply_out_callback leaves alone) */
2189                 rs->rs_on_net = 0;
2190                 ptlrpc_rs_addref(rs);
2191                 atomic_inc (&svc->srv_outstanding_replies);
2192         }
2193
2194         spin_lock(&rs->rs_lock);
2195         if (rs->rs_transno <= exp->exp_last_committed ||
2196             (!rs->rs_on_net && !rs->rs_no_ack) ||
2197              list_empty(&rs->rs_exp_list) ||     /* completed already */
2198              list_empty(&rs->rs_obd_list)) {
2199                 CDEBUG(D_HA, "Schedule reply immediately\n");
2200                 ptlrpc_dispatch_difficult_reply(rs);
2201         } else {
2202                 list_add (&rs->rs_list, &svc->srv_active_replies);
2203                 rs->rs_scheduled = 0;           /* allow notifier to schedule */
2204         }
2205         spin_unlock(&rs->rs_lock);
2206         spin_unlock(&svc->srv_lock);
2207         EXIT;
2208 }
2209
2210 int target_handle_ping(struct ptlrpc_request *req)
2211 {
2212         obd_ping(req->rq_export);
2213         return req_capsule_server_pack(&req->rq_pill);
2214 }
2215
2216 void target_committed_to_req(struct ptlrpc_request *req)
2217 {
2218         struct obd_export *exp = req->rq_export;
2219
2220         if (!exp->exp_obd->obd_no_transno && req->rq_repmsg != NULL)
2221                 lustre_msg_set_last_committed(req->rq_repmsg,
2222                                               exp->exp_last_committed);
2223         else
2224                 DEBUG_REQ(D_IOCTL, req, "not sending last_committed update (%d/"
2225                           "%d)", exp->exp_obd->obd_no_transno,
2226                           req->rq_repmsg == NULL);
2227
2228         CDEBUG(D_INFO, "last_committed "LPU64", transno "LPU64", xid "LPU64"\n",
2229                exp->exp_last_committed, req->rq_transno, req->rq_xid);
2230 }
2231 EXPORT_SYMBOL(target_committed_to_req);
2232
2233 int target_handle_qc_callback(struct ptlrpc_request *req)
2234 {
2235         struct obd_quotactl *oqctl;
2236         struct client_obd *cli = &req->rq_export->exp_obd->u.cli;
2237
2238         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
2239         if (oqctl == NULL) {
2240                 CERROR("Can't unpack obd_quotactl\n");
2241                 RETURN(-EPROTO);
2242         }
2243
2244         cli->cl_qchk_stat = oqctl->qc_stat;
2245
2246         return 0;
2247 }
2248
2249 #ifdef HAVE_QUOTA_SUPPORT
2250 int target_handle_dqacq_callback(struct ptlrpc_request *req)
2251 {
2252 #ifdef __KERNEL__
2253         struct obd_device *obd = req->rq_export->exp_obd;
2254         struct obd_device *master_obd;
2255         struct obd_device_target *obt;
2256         struct lustre_quota_ctxt *qctxt;
2257         struct qunit_data *qdata = NULL;
2258         int rc = 0;
2259         ENTRY;
2260
2261         if (OBD_FAIL_CHECK(OBD_FAIL_MDS_DROP_QUOTA_REQ))
2262                 RETURN(rc);
2263
2264         rc = req_capsule_server_pack(&req->rq_pill);
2265         if (rc) {
2266                 CERROR("packing reply failed!: rc = %d\n", rc);
2267                 RETURN(rc);
2268         }
2269
2270         LASSERT(req->rq_export);
2271
2272         qdata = quota_get_qdata(req, QUOTA_REQUEST, QUOTA_EXPORT);
2273         if (IS_ERR(qdata)) {
2274                 rc = PTR_ERR(qdata);
2275                 CDEBUG(D_ERROR, "Can't unpack qunit_data(rc: %d)\n", rc);
2276                 req->rq_status = rc;
2277                 GOTO(out, rc);
2278         }
2279
2280         /* we use the observer */
2281         if (!obd->obd_observer || !obd->obd_observer->obd_observer) {
2282                 CERROR("Can't find the observer, it is recovering\n");
2283                 req->rq_status = -EAGAIN;
2284                 GOTO(out, rc);
2285         }
2286
2287         master_obd = obd->obd_observer->obd_observer;
2288         obt = &master_obd->u.obt;
2289         qctxt = &obt->obt_qctxt;
2290
2291         if (!qctxt->lqc_setup || !qctxt->lqc_valid) {
2292                 /* quota_type has not been processed yet, return EAGAIN
2293                  * until we know whether or not quotas are supposed to
2294                  * be enabled */
2295                 CDEBUG(D_QUOTA, "quota_type not processed yet, return "
2296                        "-EAGAIN\n");
2297                 req->rq_status = -EAGAIN;
2298                 GOTO(out, rc);
2299         }
2300
2301         down_read(&obt->obt_rwsem);
2302         if (qctxt->lqc_lqs_hash == NULL) {
2303                 up_read(&obt->obt_rwsem);
2304                 /* quota_type has not been processed yet, return EAGAIN
2305                  * until we know whether or not quotas are supposed to
2306                  * be enabled */
2307                 CDEBUG(D_QUOTA, "quota_ctxt is not ready yet, return "
2308                        "-EAGAIN\n");
2309                 req->rq_status = -EAGAIN;
2310                 GOTO(out, rc);
2311         }
2312
2313         LASSERT(qctxt->lqc_handler);
2314         rc = qctxt->lqc_handler(master_obd, qdata,
2315                                 lustre_msg_get_opc(req->rq_reqmsg));
2316         up_read(&obt->obt_rwsem);
2317         if (rc && rc != -EDQUOT)
2318                 CDEBUG(rc == -EBUSY  ? D_QUOTA : D_ERROR,
2319                        "dqacq/dqrel failed! (rc:%d)\n", rc);
2320         req->rq_status = rc;
2321
2322         rc = quota_copy_qdata(req, qdata, QUOTA_REPLY, QUOTA_EXPORT);
2323         if (rc < 0) {
2324                 CERROR("Can't pack qunit_data(rc: %d)\n", rc);
2325                 GOTO(out, rc);
2326         }
2327
2328         /* Block the quota req. b=14840 */
2329         OBD_FAIL_TIMEOUT(OBD_FAIL_MDS_BLOCK_QUOTA_REQ, obd_timeout);
2330         EXIT;
2331
2332 out:
2333         rc = ptlrpc_reply(req);
2334         return rc;
2335 #else
2336         return 0;
2337 #endif /* !__KERNEL__ */
2338 }
2339 #endif /* HAVE_QUOTA_SUPPORT */
2340
2341 /* Send a remote set_info_async.
2342  * This may go from client to server or server to client
2343  */
2344 int target_set_info_rpc(struct obd_import *imp, int opcode,
2345                         obd_count keylen, void *key,
2346                         obd_count vallen, void *val,
2347                         struct ptlrpc_request_set *set)
2348 {
2349         struct ptlrpc_request *req;
2350         char                  *tmp;
2351         int                    rc;
2352         ENTRY;
2353
2354         req = ptlrpc_request_alloc(imp, &RQF_OBD_SET_INFO);
2355         if (req == NULL)
2356                 RETURN(-ENOMEM);
2357
2358         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_KEY,
2359                              RCL_CLIENT, keylen);
2360         req_capsule_set_size(&req->rq_pill, &RMF_SETINFO_VAL,
2361                              RCL_CLIENT, vallen);
2362         rc = ptlrpc_request_pack(req, LUSTRE_OBD_VERSION, opcode);
2363         if (rc) {
2364                 ptlrpc_request_free(req);
2365                 RETURN(rc);
2366         }
2367
2368         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
2369         memcpy(tmp, key, keylen);
2370         tmp = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_VAL);
2371         memcpy(tmp, val, vallen);
2372
2373         ptlrpc_request_set_replen(req);
2374
2375         if (set) {
2376                 ptlrpc_set_add_req(set, req);
2377                 ptlrpc_check_set(NULL, set);
2378         } else {
2379                 rc = ptlrpc_queue_wait(req);
2380                 ptlrpc_req_finished(req);
2381         }
2382
2383         RETURN(rc);
2384 }
2385 EXPORT_SYMBOL(target_set_info_rpc);
2386
2387
2388 ldlm_mode_t lck_compat_array[] = {
2389         [LCK_EX] LCK_COMPAT_EX,
2390         [LCK_PW] LCK_COMPAT_PW,
2391         [LCK_PR] LCK_COMPAT_PR,
2392         [LCK_CW] LCK_COMPAT_CW,
2393         [LCK_CR] LCK_COMPAT_CR,
2394         [LCK_NL] LCK_COMPAT_NL,
2395         [LCK_GROUP] LCK_COMPAT_GROUP,
2396         [LCK_COS] LCK_COMPAT_COS,
2397 };
2398
2399 /**
2400  * Rather arbitrary mapping from LDLM error codes to errno values. This should
2401  * not escape to the user level.
2402  */
2403 int ldlm_error2errno(ldlm_error_t error)
2404 {
2405         int result;
2406
2407         switch (error) {
2408         case ELDLM_OK:
2409                 result = 0;
2410                 break;
2411         case ELDLM_LOCK_CHANGED:
2412                 result = -ESTALE;
2413                 break;
2414         case ELDLM_LOCK_ABORTED:
2415                 result = -ENAVAIL;
2416                 break;
2417         case ELDLM_LOCK_REPLACED:
2418                 result = -ESRCH;
2419                 break;
2420         case ELDLM_NO_LOCK_DATA:
2421                 result = -ENOENT;
2422                 break;
2423         case ELDLM_NAMESPACE_EXISTS:
2424                 result = -EEXIST;
2425                 break;
2426         case ELDLM_BAD_NAMESPACE:
2427                 result = -EBADF;
2428                 break;
2429         default:
2430                 if (((int)error) < 0)  /* cast to signed type */
2431                         result = error; /* as ldlm_error_t can be unsigned */
2432                 else {
2433                         CERROR("Invalid DLM result code: %i\n", error);
2434                         result = -EPROTO;
2435                 }
2436         }
2437         return result;
2438 }
2439 EXPORT_SYMBOL(ldlm_error2errno);
2440
2441 /**
2442  * Dual to ldlm_error2errno(): maps errno values back to ldlm_error_t.
2443  */
2444 ldlm_error_t ldlm_errno2error(int err_no)
2445 {
2446         int error;
2447
2448         switch (err_no) {
2449         case 0:
2450                 error = ELDLM_OK;
2451                 break;
2452         case -ESTALE:
2453                 error = ELDLM_LOCK_CHANGED;
2454                 break;
2455         case -ENAVAIL:
2456                 error = ELDLM_LOCK_ABORTED;
2457                 break;
2458         case -ESRCH:
2459                 error = ELDLM_LOCK_REPLACED;
2460                 break;
2461         case -ENOENT:
2462                 error = ELDLM_NO_LOCK_DATA;
2463                 break;
2464         case -EEXIST:
2465                 error = ELDLM_NAMESPACE_EXISTS;
2466                 break;
2467         case -EBADF:
2468                 error = ELDLM_BAD_NAMESPACE;
2469                 break;
2470         default:
2471                 error = err_no;
2472         }
2473         return error;
2474 }
2475 EXPORT_SYMBOL(ldlm_errno2error);
2476