lustre/target/tgt_lastrcvd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2014, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * Lustre Unified Target
  37  * These are common function to work with last_received file
  38  *
  39  * Author: Mikhail Pershin <mike.pershin@intel.com>
  40  */
  41 #include <obd.h>
  42 #include <obd_class.h>
  43 #include <lustre_fid.h>
  44
  45 #include "tgt_internal.h"
  46
  47 static inline struct lu_buf *tti_buf_lsd(struct tgt_thread_info *tti)
  48 {
  49         tti->tti_buf.lb_buf = &tti->tti_lsd;
  50         tti->tti_buf.lb_len = sizeof(tti->tti_lsd);
  51         return &tti->tti_buf;
  52 }
  53
  54 static inline struct lu_buf *tti_buf_lcd(struct tgt_thread_info *tti)
  55 {
  56         tti->tti_buf.lb_buf = &tti->tti_lcd;
  57         tti->tti_buf.lb_len = sizeof(tti->tti_lcd);
  58         return &tti->tti_buf;
  59 }
  60
  61 /**
  62  * Allocate in-memory data for client slot related to export.
  63  */
  64 int tgt_client_alloc(struct obd_export *exp)
  65 {
  66         ENTRY;
  67         LASSERT(exp != exp->exp_obd->obd_self_export);
  68
  69         OBD_ALLOC_PTR(exp->exp_target_data.ted_lcd);
  70         if (exp->exp_target_data.ted_lcd == NULL)
  71                 RETURN(-ENOMEM);
  72         /* Mark that slot is not yet valid, 0 doesn't work here */
  73         exp->exp_target_data.ted_lr_idx = -1;
  74         RETURN(0);
  75 }
  76 EXPORT_SYMBOL(tgt_client_alloc);
  77
  78 /**
  79  * Free in-memory data for client slot related to export.
  80  */
  81 void tgt_client_free(struct obd_export *exp)
  82 {
  83         struct tg_export_data   *ted = &exp->exp_target_data;
  84         struct lu_target        *lut = class_exp2tgt(exp);
  85
  86         LASSERT(exp != exp->exp_obd->obd_self_export);
  87
  88         OBD_FREE_PTR(ted->ted_lcd);
  89         ted->ted_lcd = NULL;
  90
  91         /* Slot may be not yet assigned */
  92         if (ted->ted_lr_idx < 0)
  93                 return;
  94         /* Clear bit when lcd is freed */
  95         LASSERT(lut && lut->lut_client_bitmap);
  96         if (!test_and_clear_bit(ted->ted_lr_idx, lut->lut_client_bitmap)) {
  97                 CERROR("%s: client %u bit already clear in bitmap\n",
  98                        exp->exp_obd->obd_name, ted->ted_lr_idx);
  99                 LBUG();
 100         }
 101 }
 102 EXPORT_SYMBOL(tgt_client_free);
 103
 104 int tgt_client_data_read(const struct lu_env *env, struct lu_target *tgt,
 105                          struct lsd_client_data *lcd, loff_t *off, int index)
 106 {
 107         struct tgt_thread_info  *tti = tgt_th_info(env);
 108         int                      rc;
 109
 110         tti_buf_lcd(tti);
 111         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf, off);
 112         if (rc == 0) {
 113                 check_lcd(tgt->lut_obd->obd_name, index, &tti->tti_lcd);
 114                 lcd_le_to_cpu(&tti->tti_lcd, lcd);
 115                 lcd->lcd_last_result = ptlrpc_status_ntoh(lcd->lcd_last_result);
 116                 lcd->lcd_last_close_result =
 117                         ptlrpc_status_ntoh(lcd->lcd_last_close_result);
 118         }
 119
 120         CDEBUG(D_INFO, "%s: read lcd @%lld uuid = %s, last_transno = "LPU64
 121                ", last_xid = "LPU64", last_result = %u, last_data = %u, "
 122                "last_close_transno = "LPU64", last_close_xid = "LPU64", "
 123                "last_close_result = %u, rc = %d\n", tgt->lut_obd->obd_name,
 124                *off, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid,
 125                lcd->lcd_last_result, lcd->lcd_last_data,
 126                lcd->lcd_last_close_transno, lcd->lcd_last_close_xid,
 127                lcd->lcd_last_close_result, rc);
 128         return rc;
 129 }
 130
 131 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
 132                           struct lsd_client_data *lcd, loff_t *off,
 133                           struct thandle *th)
 134 {
 135         struct tgt_thread_info *tti = tgt_th_info(env);
 136         struct dt_object        *dto;
 137
 138         lcd->lcd_last_result = ptlrpc_status_hton(lcd->lcd_last_result);
 139         lcd->lcd_last_close_result =
 140                 ptlrpc_status_hton(lcd->lcd_last_close_result);
 141         lcd_cpu_to_le(lcd, &tti->tti_lcd);
 142         tti_buf_lcd(tti);
 143
 144         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 145         return dt_record_write(env, dto, &tti->tti_buf, off, th);
 146 }
 147
 148 /**
 149  * Update client data in last_rcvd
 150  */
 151 static int tgt_client_data_update(const struct lu_env *env,
 152                                   struct obd_export *exp)
 153 {
 154         struct tg_export_data   *ted = &exp->exp_target_data;
 155         struct lu_target        *tgt = class_exp2tgt(exp);
 156         struct tgt_thread_info  *tti = tgt_th_info(env);
 157         struct thandle          *th;
 158         int                      rc = 0;
 159
 160         ENTRY;
 161
 162         if (unlikely(tgt == NULL)) {
 163                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
 164                           class_exp2obd(exp)->obd_name);
 165                 RETURN(-EINVAL);
 166         }
 167
 168         th = dt_trans_create(env, tgt->lut_bottom);
 169         if (IS_ERR(th))
 170                 RETURN(PTR_ERR(th));
 171
 172         tti_buf_lcd(tti);
 173         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 174                                      &tti->tti_buf,
 175                                      ted->ted_lr_off, th);
 176         if (rc)
 177                 GOTO(out, rc);
 178
 179         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 180         if (rc)
 181                 GOTO(out, rc);
 182         /*
 183          * Until this operations will be committed the sync is needed
 184          * for this export. This should be done _after_ starting the
 185          * transaction so that many connecting clients will not bring
 186          * server down with lots of sync writes.
 187          */
 188         rc = tgt_new_client_cb_add(th, exp);
 189         if (rc) {
 190                 /* can't add callback, do sync now */
 191                 th->th_sync = 1;
 192         } else {
 193                 spin_lock(&exp->exp_lock);
 194                 exp->exp_need_sync = 1;
 195                 spin_unlock(&exp->exp_lock);
 196         }
 197
 198         tti->tti_off = ted->ted_lr_off;
 199         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 200         EXIT;
 201 out:
 202         dt_trans_stop(env, tgt->lut_bottom, th);
 203         CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
 204                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 205                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 206
 207         return rc;
 208 }
 209
 210 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
 211 {
 212         struct tgt_thread_info  *tti = tgt_th_info(env);
 213         int                      rc;
 214
 215         tti->tti_off = 0;
 216         tti_buf_lsd(tti);
 217         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
 218                             &tti->tti_off);
 219         if (rc == 0)
 220                 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
 221
 222         CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
 223                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 224                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 225         return rc;
 226 }
 227
 228 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
 229                           struct thandle *th)
 230 {
 231         struct tgt_thread_info  *tti = tgt_th_info(env);
 232         struct dt_object        *dto;
 233         int                      rc;
 234
 235         ENTRY;
 236
 237         tti->tti_off = 0;
 238         tti_buf_lsd(tti);
 239         lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
 240
 241         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
 242         rc = dt_record_write(env, dto, &tti->tti_buf, &tti->tti_off, th);
 243
 244         CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
 245                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 246                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 247
 248         RETURN(rc);
 249 }
 250
 251 /**
 252  * Update server data in last_rcvd
 253  */
 254 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
 255                            int sync)
 256 {
 257         struct tgt_thread_info  *tti = tgt_th_info(env);
 258         struct thandle          *th;
 259         int                      rc = 0;
 260
 261         ENTRY;
 262
 263         CDEBUG(D_SUPER,
 264                "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
 265                tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
 266                tgt->lut_last_transno);
 267
 268         /* Always save latest transno to keep it fresh */
 269         spin_lock(&tgt->lut_translock);
 270         tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
 271         spin_unlock(&tgt->lut_translock);
 272
 273         th = dt_trans_create(env, tgt->lut_bottom);
 274         if (IS_ERR(th))
 275                 RETURN(PTR_ERR(th));
 276
 277         th->th_sync = sync;
 278
 279         tti_buf_lsd(tti);
 280         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 281                                      &tti->tti_buf, tti->tti_off, th);
 282         if (rc)
 283                 GOTO(out, rc);
 284
 285         rc = dt_trans_start(env, tgt->lut_bottom, th);
 286         if (rc)
 287                 GOTO(out, rc);
 288
 289         rc = tgt_server_data_write(env, tgt, th);
 290 out:
 291         dt_trans_stop(env, tgt->lut_bottom, th);
 292
 293         CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
 294                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 295                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 296         RETURN(rc);
 297 }
 298 EXPORT_SYMBOL(tgt_server_data_update);
 299
 300 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
 301                            loff_t size)
 302 {
 303         struct dt_object *dt = tgt->lut_last_rcvd;
 304         struct thandle   *th;
 305         struct lu_attr    attr;
 306         int               rc;
 307
 308         ENTRY;
 309
 310         attr.la_size = size;
 311         attr.la_valid = LA_SIZE;
 312
 313         th = dt_trans_create(env, tgt->lut_bottom);
 314         if (IS_ERR(th))
 315                 RETURN(PTR_ERR(th));
 316         rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
 317         if (rc)
 318                 GOTO(cleanup, rc);
 319         rc = dt_declare_attr_set(env, dt, &attr, th);
 320         if (rc)
 321                 GOTO(cleanup, rc);
 322         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 323         if (rc)
 324                 GOTO(cleanup, rc);
 325
 326         rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th);
 327         if (rc == 0)
 328                 rc = dt_attr_set(env, dt, &attr, th);
 329
 330 cleanup:
 331         dt_trans_stop(env, tgt->lut_bottom, th);
 332
 333         RETURN(rc);
 334 }
 335
 336 static void tgt_client_epoch_update(const struct lu_env *env,
 337                                     struct obd_export *exp)
 338 {
 339         struct lsd_client_data  *lcd = exp->exp_target_data.ted_lcd;
 340         struct lu_target        *tgt = class_exp2tgt(exp);
 341
 342         LASSERT(tgt && tgt->lut_bottom);
 343         /** VBR: set client last_epoch to current epoch */
 344         if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
 345                 return;
 346         lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
 347         tgt_client_data_update(env, exp);
 348 }
 349
 350 /**
 351  * Update boot epoch when recovery ends
 352  */
 353 void tgt_boot_epoch_update(struct lu_target *tgt)
 354 {
 355         struct lu_env            env;
 356         struct ptlrpc_request   *req;
 357         __u32                    start_epoch;
 358         struct list_head         client_list;
 359         int                      rc;
 360
 361         if (tgt->lut_obd->obd_stopping)
 362                 return;
 363
 364         rc = lu_env_init(&env, LCT_DT_THREAD);
 365         if (rc) {
 366                 CERROR("%s: can't initialize environment: rc = %d\n",
 367                         tgt->lut_obd->obd_name, rc);
 368                 return;
 369         }
 370
 371         spin_lock(&tgt->lut_translock);
 372         start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
 373         tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
 374         tgt->lut_lsd.lsd_start_epoch = start_epoch;
 375         spin_unlock(&tgt->lut_translock);
 376
 377         INIT_LIST_HEAD(&client_list);
 378         /**
 379          * The recovery is not yet finished and final queue can still be updated
 380          * with resend requests. Move final list to separate one for processing
 381          */
 382         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 383         list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
 384         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 385
 386         /**
 387          * go through list of exports participated in recovery and
 388          * set new epoch for them
 389          */
 390         list_for_each_entry(req, &client_list, rq_list) {
 391                 LASSERT(!req->rq_export->exp_delayed);
 392                 if (!req->rq_export->exp_vbr_failed)
 393                         tgt_client_epoch_update(&env, req->rq_export);
 394         }
 395         /** return list back at once */
 396         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 397         list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
 398         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 399         /** update server epoch */
 400         tgt_server_data_update(&env, tgt, 1);
 401         lu_env_fini(&env);
 402 }
 403
 404 /**
 405  * commit callback, need to update last_commited value
 406  */
 407 struct tgt_last_committed_callback {
 408         struct dt_txn_commit_cb  llcc_cb;
 409         struct lu_target        *llcc_tgt;
 410         struct obd_export       *llcc_exp;
 411         __u64                    llcc_transno;
 412 };
 413
 414 static void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 415                                   struct dt_txn_commit_cb *cb, int err)
 416 {
 417         struct tgt_last_committed_callback *ccb;
 418
 419         ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
 420
 421         LASSERT(ccb->llcc_tgt != NULL);
 422         LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
 423
 424         spin_lock(&ccb->llcc_tgt->lut_translock);
 425         if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
 426                 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
 427
 428         LASSERT(ccb->llcc_exp);
 429         if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
 430                 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
 431                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 432                 ptlrpc_commit_replies(ccb->llcc_exp);
 433         } else {
 434                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 435         }
 436         class_export_cb_put(ccb->llcc_exp);
 437         if (ccb->llcc_transno)
 438                 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
 439                        ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
 440         OBD_FREE_PTR(ccb);
 441 }
 442
 443 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 444                            struct obd_export *exp, __u64 transno)
 445 {
 446         struct tgt_last_committed_callback      *ccb;
 447         struct dt_txn_commit_cb                 *dcb;
 448         int                                      rc;
 449
 450         OBD_ALLOC_PTR(ccb);
 451         if (ccb == NULL)
 452                 return -ENOMEM;
 453
 454         ccb->llcc_tgt = tgt;
 455         ccb->llcc_exp = class_export_cb_get(exp);
 456         ccb->llcc_transno = transno;
 457
 458         dcb = &ccb->llcc_cb;
 459         dcb->dcb_func = tgt_cb_last_committed;
 460         INIT_LIST_HEAD(&dcb->dcb_linkage);
 461         strlcpy(dcb->dcb_name, "tgt_cb_last_committed", sizeof(dcb->dcb_name));
 462
 463         rc = dt_trans_cb_add(th, dcb);
 464         if (rc) {
 465                 class_export_cb_put(exp);
 466                 OBD_FREE_PTR(ccb);
 467         }
 468
 469         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 470                 /* report failure to force synchronous operation */
 471                 return -EPERM;
 472
 473         return rc;
 474 }
 475
 476 struct tgt_new_client_callback {
 477         struct dt_txn_commit_cb  lncc_cb;
 478         struct obd_export       *lncc_exp;
 479 };
 480
 481 static void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
 482                               struct dt_txn_commit_cb *cb, int err)
 483 {
 484         struct tgt_new_client_callback *ccb;
 485
 486         ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
 487
 488         LASSERT(ccb->lncc_exp->exp_obd);
 489
 490         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 491                ccb->lncc_exp->exp_obd->obd_name,
 492                ccb->lncc_exp->exp_client_uuid.uuid);
 493
 494         spin_lock(&ccb->lncc_exp->exp_lock);
 495
 496         ccb->lncc_exp->exp_need_sync = 0;
 497
 498         spin_unlock(&ccb->lncc_exp->exp_lock);
 499         class_export_cb_put(ccb->lncc_exp);
 500
 501         OBD_FREE_PTR(ccb);
 502 }
 503
 504 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
 505 {
 506         struct tgt_new_client_callback  *ccb;
 507         struct dt_txn_commit_cb         *dcb;
 508         int                              rc;
 509
 510         OBD_ALLOC_PTR(ccb);
 511         if (ccb == NULL)
 512                 return -ENOMEM;
 513
 514         ccb->lncc_exp = class_export_cb_get(exp);
 515
 516         dcb = &ccb->lncc_cb;
 517         dcb->dcb_func = tgt_cb_new_client;
 518         INIT_LIST_HEAD(&dcb->dcb_linkage);
 519         strlcpy(dcb->dcb_name, "tgt_cb_new_client", sizeof(dcb->dcb_name));
 520
 521         rc = dt_trans_cb_add(th, dcb);
 522         if (rc) {
 523                 class_export_cb_put(exp);
 524                 OBD_FREE_PTR(ccb);
 525         }
 526         return rc;
 527 }
 528
 529 /**
 530  * Add new client to the last_rcvd upon new connection.
 531  *
 532  * We use a bitmap to locate a free space in the last_rcvd file and initialize
 533  * tg_export_data.
 534  */
 535 int tgt_client_new(const struct lu_env *env, struct obd_export *exp)
 536 {
 537         struct tg_export_data   *ted = &exp->exp_target_data;
 538         struct lu_target        *tgt = class_exp2tgt(exp);
 539         int                      rc = 0, idx;
 540
 541         ENTRY;
 542
 543         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 544         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid))
 545                 RETURN(0);
 546
 547         mutex_init(&ted->ted_lcd_lock);
 548
 549         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 550                 RETURN(0);
 551
 552         /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so
 553          * there's no need for extra complication here
 554          */
 555         idx = find_first_zero_bit(tgt->lut_client_bitmap, LR_MAX_CLIENTS);
 556 repeat:
 557         if (idx >= LR_MAX_CLIENTS ||
 558             OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) {
 559                 CERROR("%s: no room for %u clients - fix LR_MAX_CLIENTS\n",
 560                        tgt->lut_obd->obd_name,  idx);
 561                 RETURN(-EOVERFLOW);
 562         }
 563         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
 564                 idx = find_next_zero_bit(tgt->lut_client_bitmap,
 565                                              LR_MAX_CLIENTS, idx);
 566                 goto repeat;
 567         }
 568
 569         CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added\n",
 570                tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid);
 571
 572         ted->ted_lr_idx = idx;
 573         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
 574                           idx * tgt->lut_lsd.lsd_client_size;
 575
 576         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 577
 578         CDEBUG(D_INFO, "%s: new client at index %d (%llu) with UUID '%s'\n",
 579                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
 580                ted->ted_lcd->lcd_uuid);
 581
 582         if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD))
 583                 RETURN(-ENOSPC);
 584
 585         rc = tgt_client_data_update(env, exp);
 586         if (rc)
 587                 CERROR("%s: Failed to write client lcd at idx %d, rc %d\n",
 588                        tgt->lut_obd->obd_name, idx, rc);
 589
 590         RETURN(rc);
 591 }
 592 EXPORT_SYMBOL(tgt_client_new);
 593
 594 /* Add client data to the MDS.  We use a bitmap to locate a free space
 595  * in the last_rcvd file if cl_off is -1 (i.e. a new client).
 596  * Otherwise, we just have to read the data from the last_rcvd file and
 597  * we know its offset.
 598  *
 599  * It should not be possible to fail adding an existing client - otherwise
 600  * mdt_init_server_data() callsite needs to be fixed.
 601  */
 602 int tgt_client_add(const struct lu_env *env,  struct obd_export *exp, int idx)
 603 {
 604         struct tg_export_data   *ted = &exp->exp_target_data;
 605         struct lu_target        *tgt = class_exp2tgt(exp);
 606
 607         ENTRY;
 608
 609         LASSERT(tgt && tgt->lut_client_bitmap != NULL);
 610         LASSERTF(idx >= 0, "%d\n", idx);
 611
 612         if (!strcmp(ted->ted_lcd->lcd_uuid, tgt->lut_obd->obd_uuid.uuid) ||
 613             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 614                 RETURN(0);
 615
 616         if (test_and_set_bit(idx, tgt->lut_client_bitmap)) {
 617                 CERROR("%s: client %d: bit already set in bitmap!!\n",
 618                        tgt->lut_obd->obd_name,  idx);
 619                 LBUG();
 620         }
 621
 622         CDEBUG(D_INFO, "%s: client at idx %d with UUID '%s' added\n",
 623                tgt->lut_obd->obd_name, idx, ted->ted_lcd->lcd_uuid);
 624
 625         ted->ted_lr_idx = idx;
 626         ted->ted_lr_off = tgt->lut_lsd.lsd_client_start +
 627                           idx * tgt->lut_lsd.lsd_client_size;
 628
 629         mutex_init(&ted->ted_lcd_lock);
 630
 631         LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off);
 632
 633         RETURN(0);
 634 }
 635
 636 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
 637 {
 638         struct tg_export_data   *ted = &exp->exp_target_data;
 639         struct lu_target        *tgt = class_exp2tgt(exp);
 640         int                      rc;
 641
 642         ENTRY;
 643
 644         LASSERT(ted->ted_lcd);
 645
 646         if (unlikely(tgt == NULL)) {
 647                 CDEBUG(D_ERROR, "%s: No target for connected export\n",
 648                        class_exp2obd(exp)->obd_name);
 649                 RETURN(-EINVAL);
 650         }
 651
 652         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 653         if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
 654                     (char *)tgt->lut_obd->obd_uuid.uuid) ||
 655             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 656                 RETURN(0);
 657
 658         CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
 659                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
 660                ted->ted_lcd->lcd_uuid);
 661
 662         /* Clear the bit _after_ zeroing out the client so we don't
 663            race with filter_client_add and zero out new clients.*/
 664         if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
 665                 CERROR("%s: client %u: bit already clear in bitmap!!\n",
 666                        tgt->lut_obd->obd_name, ted->ted_lr_idx);
 667                 LBUG();
 668         }
 669
 670         /* Do not erase record for recoverable client. */
 671         if (exp->exp_flags & OBD_OPT_FAILOVER)
 672                 RETURN(0);
 673
 674         /* Make sure the server's last_transno is up to date.
 675          * This should be done before zeroing client slot so last_transno will
 676          * be in server data or in client data in case of failure */
 677         rc = tgt_server_data_update(env, tgt, 0);
 678         if (rc != 0) {
 679                 CERROR("%s: failed to update server data, skip client %s "
 680                        "zeroing, rc %d\n", tgt->lut_obd->obd_name,
 681                        ted->ted_lcd->lcd_uuid, rc);
 682                 RETURN(rc);
 683         }
 684
 685         mutex_lock(&ted->ted_lcd_lock);
 686         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
 687         rc = tgt_client_data_update(env, exp);
 688         mutex_unlock(&ted->ted_lcd_lock);
 689
 690         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 691                "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
 692                tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
 693                ted->ted_lr_idx, ted->ted_lr_off, rc);
 694         RETURN(rc);
 695 }
 696 EXPORT_SYMBOL(tgt_client_del);
 697
 698 /*
 699  * last_rcvd & last_committed update callbacks
 700  */
 701 static int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 702                                 struct dt_object *obj, __u64 opdata,
 703                                 struct thandle *th, struct ptlrpc_request *req)
 704 {
 705         struct tgt_thread_info  *tti = tgt_th_info(env);
 706         struct tg_export_data   *ted;
 707         __u64                   *transno_p;
 708         int                      rc = 0;
 709         bool                     lw_client, update = false;
 710
 711         ENTRY;
 712
 713         ted = &req->rq_export->exp_target_data;
 714
 715         lw_client = exp_connect_flags(req->rq_export) & OBD_CONNECT_LIGHTWEIGHT;
 716         if (ted->ted_lr_idx < 0 && !lw_client)
 717                 /* ofd connect may cause transaction before export has
 718                  * last_rcvd slot */
 719                 RETURN(0);
 720
 721         tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
 722
 723         spin_lock(&tgt->lut_translock);
 724         if (th->th_result != 0) {
 725                 if (tti->tti_transno != 0) {
 726                         CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
 727                                tgt_name(tgt), tti->tti_transno, th->th_result);
 728                 }
 729         } else if (tti->tti_transno == 0) {
 730                 tti->tti_transno = ++tgt->lut_last_transno;
 731         } else {
 732                 /* should be replay */
 733                 if (tti->tti_transno > tgt->lut_last_transno)
 734                         tgt->lut_last_transno = tti->tti_transno;
 735         }
 736         spin_unlock(&tgt->lut_translock);
 737
 738         /** VBR: set new versions */
 739         if (th->th_result == 0 && obj != NULL) {
 740                 struct dt_object *dto = dt_object_locate(obj, th->th_dev);
 741                 dt_version_set(env, dto, tti->tti_transno, th);
 742         }
 743
 744         /* filling reply data */
 745         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 746                tti->tti_transno, tgt->lut_obd->obd_last_committed);
 747
 748         req->rq_transno = tti->tti_transno;
 749         lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
 750
 751         /* if can't add callback, do sync write */
 752         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, req->rq_export,
 753                                                 tti->tti_transno);
 754
 755         if (lw_client) {
 756                 /* All operations performed by LW clients are synchronous and
 757                  * we store the committed transno in the last_rcvd header */
 758                 spin_lock(&tgt->lut_translock);
 759                 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
 760                         tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
 761                         update = true;
 762                 }
 763                 spin_unlock(&tgt->lut_translock);
 764                 /* Although lightweight (LW) connections have no slot in
 765                  * last_rcvd, we still want to maintain the in-memory
 766                  * lsd_client_data structure in order to properly handle reply
 767                  * reconstruction. */
 768         } else if (ted->ted_lr_off == 0) {
 769                 CERROR("%s: client idx %d has offset %lld\n",
 770                        tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
 771                 RETURN(-EINVAL);
 772         }
 773
 774         /* if the export has already been disconnected, we have no last_rcvd
 775          * slot, update server data with latest transno then */
 776         if (ted->ted_lcd == NULL) {
 777                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 778                       req->rq_export->exp_client_uuid.uuid, rc);
 779                 GOTO(srv_update, rc = 0);
 780         }
 781
 782         mutex_lock(&ted->ted_lcd_lock);
 783         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
 784         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) {
 785                 transno_p = &ted->ted_lcd->lcd_last_close_transno;
 786                 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
 787                 ted->ted_lcd->lcd_last_close_result = th->th_result;
 788         } else {
 789                 /* VBR: save versions in last_rcvd for reconstruct. */
 790                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 791
 792                 if (pre_versions) {
 793                         ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
 794                         ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
 795                         ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
 796                         ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
 797                 }
 798                 transno_p = &ted->ted_lcd->lcd_last_transno;
 799                 ted->ted_lcd->lcd_last_xid = req->rq_xid;
 800                 ted->ted_lcd->lcd_last_result = th->th_result;
 801                 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
 802                  * see struct ldlm_reply->lock_policy_res1; */
 803                 ted->ted_lcd->lcd_last_data = opdata;
 804         }
 805
 806         /* Update transno in slot only if non-zero number, i.e. no errors */
 807         if (likely(tti->tti_transno != 0)) {
 808                 if (*transno_p > tti->tti_transno &&
 809                     !tgt->lut_no_reconstruct) {
 810                         CERROR("%s: trying to overwrite bigger transno:"
 811                                "on-disk: "LPU64", new: "LPU64" replay: %d. "
 812                                "see LU-617.\n", tgt_name(tgt), *transno_p,
 813                                tti->tti_transno, req_is_replay(req));
 814                         if (req_is_replay(req)) {
 815                                 spin_lock(&req->rq_export->exp_lock);
 816                                 req->rq_export->exp_vbr_failed = 1;
 817                                 spin_unlock(&req->rq_export->exp_lock);
 818                         }
 819                         mutex_unlock(&ted->ted_lcd_lock);
 820                         RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 821                 }
 822                 *transno_p = tti->tti_transno;
 823         }
 824
 825         if (!lw_client) {
 826                 tti->tti_off = ted->ted_lr_off;
 827                 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 828                 if (rc < 0) {
 829                         mutex_unlock(&ted->ted_lcd_lock);
 830                         RETURN(rc);
 831                 }
 832         }
 833         mutex_unlock(&ted->ted_lcd_lock);
 834         EXIT;
 835 srv_update:
 836         if (update)
 837                 rc = tgt_server_data_write(env, tgt, th);
 838         return rc;
 839 }
 840
 841 /*
 842  * last_rcvd update for echo client simulation.
 843  * It updates last_rcvd client slot and version of object in
 844  * simple way but with all locks to simulate all drawbacks
 845  */
 846 static int tgt_last_rcvd_update_echo(const struct lu_env *env,
 847                                      struct lu_target *tgt,
 848                                      struct dt_object *obj,
 849                                      struct thandle *th,
 850                                      struct obd_export *exp)
 851 {
 852         struct tgt_thread_info  *tti = tgt_th_info(env);
 853         struct tg_export_data   *ted = &exp->exp_target_data;
 854         int                      rc = 0;
 855
 856         ENTRY;
 857
 858         tti->tti_transno = 0;
 859
 860         spin_lock(&tgt->lut_translock);
 861         if (th->th_result == 0)
 862                 tti->tti_transno = ++tgt->lut_last_transno;
 863         spin_unlock(&tgt->lut_translock);
 864
 865         /** VBR: set new versions */
 866         if (th->th_result == 0 && obj != NULL)
 867                 dt_version_set(env, obj, tti->tti_transno, th);
 868
 869         /* if can't add callback, do sync write */
 870         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
 871                                                 tti->tti_transno);
 872
 873         LASSERT(ted->ted_lr_off > 0);
 874
 875         mutex_lock(&ted->ted_lcd_lock);
 876         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
 877         ted->ted_lcd->lcd_last_transno = tti->tti_transno;
 878         ted->ted_lcd->lcd_last_result = th->th_result;
 879
 880         tti->tti_off = ted->ted_lr_off;
 881         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 882         mutex_unlock(&ted->ted_lcd_lock);
 883         RETURN(rc);
 884 }
 885
 886 static int tgt_clients_data_init(const struct lu_env *env,
 887                                  struct lu_target *tgt,
 888                                  unsigned long last_size)
 889 {
 890         struct obd_device       *obd = tgt->lut_obd;
 891         struct lr_server_data   *lsd = &tgt->lut_lsd;
 892         struct lsd_client_data  *lcd = NULL;
 893         struct tg_export_data   *ted;
 894         int                      cl_idx;
 895         int                      rc = 0;
 896         loff_t                   off = lsd->lsd_client_start;
 897
 898         ENTRY;
 899
 900         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
 901                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
 902
 903         OBD_ALLOC_PTR(lcd);
 904         if (lcd == NULL)
 905                 RETURN(-ENOMEM);
 906
 907         for (cl_idx = 0; off < last_size; cl_idx++) {
 908                 struct obd_export       *exp;
 909                 __u64                    last_transno;
 910
 911                 /* Don't assume off is incremented properly by
 912                  * read_record(), in case sizeof(*lcd)
 913                  * isn't the same as fsd->lsd_client_size.  */
 914                 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
 915                 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
 916                 if (rc) {
 917                         CERROR("%s: error reading last_rcvd %s idx %d off "
 918                                "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
 919                                cl_idx, off, rc);
 920                         rc = 0;
 921                         break; /* read error shouldn't cause startup to fail */
 922                 }
 923
 924                 if (lcd->lcd_uuid[0] == '\0') {
 925                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 926                                cl_idx);
 927                         continue;
 928                 }
 929
 930                 last_transno = lcd_last_transno(lcd);
 931
 932                 /* These exports are cleaned up by disconnect, so they
 933                  * need to be set up like real exports as connect does.
 934                  */
 935                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 936                        " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
 937                        last_transno, lsd->lsd_last_transno, lcd_last_xid(lcd));
 938
 939                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
 940                 if (IS_ERR(exp)) {
 941                         if (PTR_ERR(exp) == -EALREADY) {
 942                                 /* export already exists, zero out this one */
 943                                 CERROR("%s: Duplicate export %s!\n",
 944                                        tgt_name(tgt), lcd->lcd_uuid);
 945                                 continue;
 946                         }
 947                         GOTO(err_out, rc = PTR_ERR(exp));
 948                 }
 949
 950                 ted = &exp->exp_target_data;
 951                 *ted->ted_lcd = *lcd;
 952
 953                 rc = tgt_client_add(env, exp, cl_idx);
 954                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
 955                 /* VBR: set export last committed version */
 956                 exp->exp_last_committed = last_transno;
 957                 spin_lock(&exp->exp_lock);
 958                 exp->exp_connecting = 0;
 959                 exp->exp_in_recovery = 0;
 960                 spin_unlock(&exp->exp_lock);
 961                 obd->obd_max_recoverable_clients++;
 962                 class_export_put(exp);
 963
 964                 /* Need to check last_rcvd even for duplicated exports. */
 965                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 966                        cl_idx, last_transno);
 967
 968                 spin_lock(&tgt->lut_translock);
 969                 tgt->lut_last_transno = max(last_transno,
 970                                             tgt->lut_last_transno);
 971                 spin_unlock(&tgt->lut_translock);
 972         }
 973
 974 err_out:
 975         OBD_FREE_PTR(lcd);
 976         RETURN(rc);
 977 }
 978
 979 struct server_compat_data {
 980         __u32 rocompat;
 981         __u32 incompat;
 982         __u32 rocinit;
 983         __u32 incinit;
 984 };
 985
 986 static struct server_compat_data tgt_scd[] = {
 987         [LDD_F_SV_TYPE_MDT] = {
 988                 .rocompat = OBD_ROCOMPAT_LOVOBJID,
 989                 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
 990                             OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
 991                             OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI,
 992                 .rocinit = OBD_ROCOMPAT_LOVOBJID,
 993                 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
 994                            OBD_INCOMPAT_MULTI_OI,
 995         },
 996         [LDD_F_SV_TYPE_OST] = {
 997                 .rocompat = OBD_ROCOMPAT_IDX_IN_IDIF,
 998                 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
 999                             OBD_INCOMPAT_FID,
1000                 .rocinit = OBD_ROCOMPAT_IDX_IN_IDIF,
1001                 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
1002         }
1003 };
1004
1005 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
1006 {
1007         struct tgt_thread_info          *tti = tgt_th_info(env);
1008         struct lr_server_data           *lsd = &tgt->lut_lsd;
1009         unsigned long                    last_rcvd_size;
1010         __u32                            index;
1011         int                              rc, type;
1012
1013         rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr);
1014         if (rc)
1015                 RETURN(rc);
1016
1017         last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1018
1019         /* ensure padding in the struct is the correct size */
1020         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1021                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1022
1023         rc = server_name2index(tgt_name(tgt), &index, NULL);
1024         if (rc < 0) {
1025                 CERROR("%s: Can not get index from name: rc = %d\n",
1026                        tgt_name(tgt), rc);
1027                 RETURN(rc);
1028         }
1029         /* server_name2index() returns type */
1030         type = rc;
1031         if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1032                 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1033                 RETURN(-EINVAL);
1034         }
1035
1036         /* last_rcvd on OST doesn't provide reconstruct support because there
1037          * may be up to 8 in-flight write requests per single slot in
1038          * last_rcvd client data
1039          */
1040         tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1041
1042         if (last_rcvd_size == 0) {
1043                 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1044
1045                 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1046                        sizeof(lsd->lsd_uuid));
1047                 lsd->lsd_last_transno = 0;
1048                 lsd->lsd_mount_count = 0;
1049                 lsd->lsd_server_size = LR_SERVER_SIZE;
1050                 lsd->lsd_client_start = LR_CLIENT_START;
1051                 lsd->lsd_client_size = LR_CLIENT_SIZE;
1052                 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1053                 lsd->lsd_osd_index = index;
1054                 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1055                 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1056         } else {
1057                 rc = tgt_server_data_read(env, tgt);
1058                 if (rc) {
1059                         CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1060                                tgt_name(tgt), rc);
1061                         RETURN(rc);
1062                 }
1063                 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1064                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1065                                            "using the wrong disk %s. Were the"
1066                                            " /dev/ assignments rearranged?\n",
1067                                            tgt->lut_obd->obd_uuid.uuid,
1068                                            lsd->lsd_uuid);
1069                         RETURN(-EINVAL);
1070                 }
1071
1072                 if (lsd->lsd_osd_index != index) {
1073                         LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1074                                            "is different with the index %d in"
1075                                            "config log, It might be disk"
1076                                            "corruption!\n", tgt_name(tgt),
1077                                            lsd->lsd_osd_index, index);
1078                         RETURN(-EINVAL);
1079                 }
1080         }
1081
1082         if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1083                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1084                        tgt_name(tgt),
1085                        lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1086                 RETURN(-EINVAL);
1087         }
1088
1089         if (type == LDD_F_SV_TYPE_MDT)
1090                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1091
1092         if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1093                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1094                        tgt_name(tgt),
1095                        lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1096                 RETURN(-EINVAL);
1097         }
1098         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1099         if (type == LDD_F_SV_TYPE_MDT &&
1100             !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1101                 if (last_rcvd_size > lsd->lsd_client_start) {
1102                         LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1103                                       "remove all clients for interop needs\n",
1104                                       tgt_name(tgt));
1105                         rc = tgt_truncate_last_rcvd(env, tgt,
1106                                                     lsd->lsd_client_start);
1107                         if (rc)
1108                                 RETURN(rc);
1109                         last_rcvd_size = lsd->lsd_client_start;
1110                 }
1111                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1112                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1113         }
1114
1115         spin_lock(&tgt->lut_translock);
1116         tgt->lut_last_transno = lsd->lsd_last_transno;
1117         spin_unlock(&tgt->lut_translock);
1118
1119         lsd->lsd_mount_count++;
1120
1121         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1122         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1123                tgt_name(tgt), tgt->lut_last_transno);
1124         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1125                tgt_name(tgt), lsd->lsd_mount_count);
1126         CDEBUG(D_INODE, "%s: server data size: %u\n",
1127                tgt_name(tgt), lsd->lsd_server_size);
1128         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1129                tgt_name(tgt), lsd->lsd_client_start);
1130         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1131                tgt_name(tgt), lsd->lsd_client_size);
1132         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1133                tgt_name(tgt), last_rcvd_size);
1134         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1135                tgt_name(tgt), lsd->lsd_subdir_count);
1136         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1137                last_rcvd_size <= lsd->lsd_client_start ? 0 :
1138                (last_rcvd_size - lsd->lsd_client_start) /
1139                 lsd->lsd_client_size);
1140         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1141
1142         if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1143             lsd->lsd_client_size == 0) {
1144                 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1145                 RETURN(-EINVAL);
1146         }
1147
1148         if (!tgt->lut_obd->obd_replayable)
1149                 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1150
1151         rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1152         if (rc < 0)
1153                 GOTO(err_client, rc);
1154
1155         spin_lock(&tgt->lut_translock);
1156         /* obd_last_committed is used for compatibility
1157          * with other lustre recovery code */
1158         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1159         spin_unlock(&tgt->lut_translock);
1160
1161         tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1162         tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1163
1164         /* save it, so mount count and last_transno is current */
1165         rc = tgt_server_data_update(env, tgt, 0);
1166         if (rc < 0)
1167                 GOTO(err_client, rc);
1168
1169         RETURN(0);
1170
1171 err_client:
1172         class_disconnect_exports(tgt->lut_obd);
1173         return rc;
1174 }
1175
1176 /* add credits for last_rcvd update */
1177 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1178                      void *cookie)
1179 {
1180         struct lu_target        *tgt = cookie;
1181         struct tgt_session_info *tsi;
1182         struct tgt_thread_info  *tti = tgt_th_info(env);
1183         struct dt_object        *dto;
1184         int                      rc;
1185
1186         /* if there is no session, then this transaction is not result of
1187          * request processing but some local operation */
1188         if (env->le_ses == NULL)
1189                 return 0;
1190
1191         LASSERT(tgt->lut_last_rcvd);
1192         tsi = tgt_ses_info(env);
1193         /* OFD may start transaction without export assigned */
1194         if (tsi->tsi_exp == NULL)
1195                 return 0;
1196
1197         dto = dt_object_locate(tgt->lut_last_rcvd, th->th_dev);
1198         tti_buf_lcd(tti);
1199
1200         rc = dt_declare_record_write(env, dto, &tti->tti_buf,
1201                                      tsi->tsi_exp->exp_target_data.ted_lr_off,
1202                                      th);
1203         if (rc)
1204                 return rc;
1205
1206         tti_buf_lsd(tti);
1207         rc = dt_declare_record_write(env, dto, &tti->tti_buf, 0, th);
1208         if (rc)
1209                 return rc;
1210
1211         if (tsi->tsi_vbr_obj != NULL &&
1212             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1213                 dto = dt_object_locate(tsi->tsi_vbr_obj, th->th_dev);
1214                 rc = dt_declare_version_set(env, dto, th);
1215         }
1216
1217         return rc;
1218 }
1219
1220 /* Update last_rcvd records with latests transaction data */
1221 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1222                     void *cookie)
1223 {
1224         struct lu_target        *tgt = cookie;
1225         struct tgt_session_info *tsi;
1226         struct tgt_thread_info  *tti = tgt_th_info(env);
1227         struct dt_object        *obj = NULL;
1228         int                      rc;
1229         bool                     echo_client;
1230
1231         if (env->le_ses == NULL)
1232                 return 0;
1233
1234         tsi = tgt_ses_info(env);
1235         /* OFD may start transaction without export assigned */
1236         if (tsi->tsi_exp == NULL)
1237                 return 0;
1238
1239         echo_client = (tgt_ses_req(tsi) == NULL);
1240
1241         if (tti->tti_has_trans && !echo_client) {
1242                 if (tti->tti_mult_trans == 0) {
1243                         CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1244                                tti->tti_transno);
1245                         RETURN(0);
1246                 }
1247                 /* we need another transno to be assigned */
1248                 tti->tti_transno = 0;
1249         } else if (th->th_result == 0) {
1250                 tti->tti_has_trans = 1;
1251         }
1252
1253         if (tsi->tsi_vbr_obj != NULL &&
1254             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1255                 obj = tsi->tsi_vbr_obj;
1256         }
1257
1258         if (unlikely(echo_client)) /* echo client special case */
1259                 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1260                                                tsi->tsi_exp);
1261         else
1262                 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1263                                           tgt_ses_req(tsi));
1264         return rc;
1265 }