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, 2012, 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_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 EXPORT_SYMBOL(tgt_client_data_read);
 131
 132 int tgt_client_data_write(const struct lu_env *env, struct lu_target *tgt,
 133                           struct lsd_client_data *lcd, loff_t *off,
 134                           struct thandle *th)
 135 {
 136         struct tgt_thread_info *tti = tgt_th_info(env);
 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         return dt_record_write(env, tgt->lut_last_rcvd, &tti->tti_buf, off, th);
 145 }
 146 EXPORT_SYMBOL(tgt_client_data_write);
 147
 148 /**
 149  * Update client data in last_rcvd
 150  */
 151 int tgt_client_data_update(const struct lu_env *env, struct obd_export *exp)
 152 {
 153         struct tg_export_data   *ted = &exp->exp_target_data;
 154         struct lu_target        *tgt = class_exp2tgt(exp);
 155         struct tgt_thread_info  *tti = tgt_th_info(env);
 156         struct thandle          *th;
 157         int                      rc = 0;
 158
 159         ENTRY;
 160
 161         th = dt_trans_create(env, tgt->lut_bottom);
 162         if (IS_ERR(th))
 163                 RETURN(PTR_ERR(th));
 164
 165         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 166                                      sizeof(struct lsd_client_data),
 167                                      ted->ted_lr_off, th);
 168         if (rc)
 169                 GOTO(out, rc);
 170
 171         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 172         if (rc)
 173                 GOTO(out, rc);
 174         /*
 175          * Until this operations will be committed the sync is needed
 176          * for this export. This should be done _after_ starting the
 177          * transaction so that many connecting clients will not bring
 178          * server down with lots of sync writes.
 179          */
 180         rc = tgt_new_client_cb_add(th, exp);
 181         if (rc) {
 182                 /* can't add callback, do sync now */
 183                 th->th_sync = 1;
 184         } else {
 185                 spin_lock(&exp->exp_lock);
 186                 exp->exp_need_sync = 1;
 187                 spin_unlock(&exp->exp_lock);
 188         }
 189
 190         tti->tti_off = ted->ted_lr_off;
 191         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 192         EXIT;
 193 out:
 194         dt_trans_stop(env, tgt->lut_bottom, th);
 195         CDEBUG(D_INFO, "%s: update last_rcvd client data for UUID = %s, "
 196                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 197                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 198
 199         return rc;
 200 }
 201
 202 int tgt_server_data_read(const struct lu_env *env, struct lu_target *tgt)
 203 {
 204         struct tgt_thread_info  *tti = tgt_th_info(env);
 205         int                      rc;
 206
 207         tti->tti_off = 0;
 208         tti_buf_lsd(tti);
 209         rc = dt_record_read(env, tgt->lut_last_rcvd, &tti->tti_buf,
 210                             &tti->tti_off);
 211         if (rc == 0)
 212                 lsd_le_to_cpu(&tti->tti_lsd, &tgt->lut_lsd);
 213
 214         CDEBUG(D_INFO, "%s: read last_rcvd server data for UUID = %s, "
 215                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 216                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 217         return rc;
 218 }
 219 EXPORT_SYMBOL(tgt_server_data_read);
 220
 221 int tgt_server_data_write(const struct lu_env *env, struct lu_target *tgt,
 222                           struct thandle *th)
 223 {
 224         struct tgt_thread_info  *tti = tgt_th_info(env);
 225         int                      rc;
 226
 227         ENTRY;
 228
 229         tti->tti_off = 0;
 230         tti_buf_lsd(tti);
 231         lsd_cpu_to_le(&tgt->lut_lsd, &tti->tti_lsd);
 232
 233         rc = dt_record_write(env, tgt->lut_last_rcvd, &tti->tti_buf,
 234                              &tti->tti_off, th);
 235
 236         CDEBUG(D_INFO, "%s: write last_rcvd server data for UUID = %s, "
 237                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 238                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 239
 240         RETURN(rc);
 241 }
 242 EXPORT_SYMBOL(tgt_server_data_write);
 243
 244 /**
 245  * Update server data in last_rcvd
 246  */
 247 int tgt_server_data_update(const struct lu_env *env, struct lu_target *tgt,
 248                            int sync)
 249 {
 250         struct tgt_thread_info  *tti = tgt_th_info(env);
 251         struct thandle          *th;
 252         int                      rc = 0;
 253
 254         ENTRY;
 255
 256         CDEBUG(D_SUPER,
 257                "%s: mount_count is "LPU64", last_transno is "LPU64"\n",
 258                tgt->lut_lsd.lsd_uuid, tgt->lut_obd->u.obt.obt_mount_count,
 259                tgt->lut_last_transno);
 260
 261         /* Always save latest transno to keep it fresh */
 262         spin_lock(&tgt->lut_translock);
 263         tgt->lut_lsd.lsd_last_transno = tgt->lut_last_transno;
 264         spin_unlock(&tgt->lut_translock);
 265
 266         th = dt_trans_create(env, tgt->lut_bottom);
 267         if (IS_ERR(th))
 268                 RETURN(PTR_ERR(th));
 269
 270         th->th_sync = sync;
 271
 272         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
 273                                      sizeof(struct lr_server_data),
 274                                      tti->tti_off, th);
 275         if (rc)
 276                 GOTO(out, rc);
 277
 278         rc = dt_trans_start(env, tgt->lut_bottom, th);
 279         if (rc)
 280                 GOTO(out, rc);
 281
 282         rc = tgt_server_data_write(env, tgt, th);
 283 out:
 284         dt_trans_stop(env, tgt->lut_bottom, th);
 285
 286         CDEBUG(D_INFO, "%s: update last_rcvd server data for UUID = %s, "
 287                "last_transno = "LPU64": rc = %d\n", tgt->lut_obd->obd_name,
 288                tgt->lut_lsd.lsd_uuid, tgt->lut_lsd.lsd_last_transno, rc);
 289         RETURN(rc);
 290 }
 291 EXPORT_SYMBOL(tgt_server_data_update);
 292
 293 int tgt_truncate_last_rcvd(const struct lu_env *env, struct lu_target *tgt,
 294                            loff_t size)
 295 {
 296         struct dt_object *dt = tgt->lut_last_rcvd;
 297         struct thandle   *th;
 298         struct lu_attr    attr;
 299         int               rc;
 300
 301         ENTRY;
 302
 303         attr.la_size = size;
 304         attr.la_valid = LA_SIZE;
 305
 306         th = dt_trans_create(env, tgt->lut_bottom);
 307         if (IS_ERR(th))
 308                 RETURN(PTR_ERR(th));
 309         rc = dt_declare_punch(env, dt, size, OBD_OBJECT_EOF, th);
 310         if (rc)
 311                 GOTO(cleanup, rc);
 312         rc = dt_declare_attr_set(env, dt, &attr, th);
 313         if (rc)
 314                 GOTO(cleanup, rc);
 315         rc = dt_trans_start_local(env, tgt->lut_bottom, th);
 316         if (rc)
 317                 GOTO(cleanup, rc);
 318
 319         rc = dt_punch(env, dt, size, OBD_OBJECT_EOF, th, BYPASS_CAPA);
 320         if (rc == 0)
 321                 rc = dt_attr_set(env, dt, &attr, th, BYPASS_CAPA);
 322
 323 cleanup:
 324         dt_trans_stop(env, tgt->lut_bottom, th);
 325
 326         RETURN(rc);
 327 }
 328 EXPORT_SYMBOL(tgt_truncate_last_rcvd);
 329
 330 void tgt_client_epoch_update(const struct lu_env *env, struct obd_export *exp)
 331 {
 332         struct lsd_client_data  *lcd = exp->exp_target_data.ted_lcd;
 333         struct lu_target        *tgt = class_exp2tgt(exp);
 334
 335         LASSERT(tgt->lut_bottom);
 336         /** VBR: set client last_epoch to current epoch */
 337         if (lcd->lcd_last_epoch >= tgt->lut_lsd.lsd_start_epoch)
 338                 return;
 339         lcd->lcd_last_epoch = tgt->lut_lsd.lsd_start_epoch;
 340         tgt_client_data_update(env, exp);
 341 }
 342
 343 /**
 344  * Update boot epoch when recovery ends
 345  */
 346 void tgt_boot_epoch_update(struct lu_target *tgt)
 347 {
 348         struct lu_env            env;
 349         struct ptlrpc_request   *req;
 350         __u32                    start_epoch;
 351         cfs_list_t               client_list;
 352         int                      rc;
 353
 354         if (tgt->lut_obd->obd_stopping)
 355                 return;
 356
 357         rc = lu_env_init(&env, LCT_DT_THREAD);
 358         if (rc) {
 359                 CERROR("%s: can't initialize environment: rc = %d\n",
 360                         tgt->lut_obd->obd_name, rc);
 361                 return;
 362         }
 363
 364         spin_lock(&tgt->lut_translock);
 365         start_epoch = lr_epoch(tgt->lut_last_transno) + 1;
 366         tgt->lut_last_transno = (__u64)start_epoch << LR_EPOCH_BITS;
 367         tgt->lut_lsd.lsd_start_epoch = start_epoch;
 368         spin_unlock(&tgt->lut_translock);
 369
 370         CFS_INIT_LIST_HEAD(&client_list);
 371         /**
 372          * The recovery is not yet finished and final queue can still be updated
 373          * with resend requests. Move final list to separate one for processing
 374          */
 375         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 376         cfs_list_splice_init(&tgt->lut_obd->obd_final_req_queue, &client_list);
 377         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 378
 379         /**
 380          * go through list of exports participated in recovery and
 381          * set new epoch for them
 382          */
 383         cfs_list_for_each_entry(req, &client_list, rq_list) {
 384                 LASSERT(!req->rq_export->exp_delayed);
 385                 if (!req->rq_export->exp_vbr_failed)
 386                         tgt_client_epoch_update(&env, req->rq_export);
 387         }
 388         /** return list back at once */
 389         spin_lock(&tgt->lut_obd->obd_recovery_task_lock);
 390         cfs_list_splice_init(&client_list, &tgt->lut_obd->obd_final_req_queue);
 391         spin_unlock(&tgt->lut_obd->obd_recovery_task_lock);
 392         /** update server epoch */
 393         tgt_server_data_update(&env, tgt, 1);
 394         lu_env_fini(&env);
 395 }
 396 EXPORT_SYMBOL(tgt_boot_epoch_update);
 397
 398 /**
 399  * commit callback, need to update last_commited value
 400  */
 401 struct tgt_last_committed_callback {
 402         struct dt_txn_commit_cb  llcc_cb;
 403         struct lu_target        *llcc_tgt;
 404         struct obd_export       *llcc_exp;
 405         __u64                    llcc_transno;
 406 };
 407
 408 void tgt_cb_last_committed(struct lu_env *env, struct thandle *th,
 409                            struct dt_txn_commit_cb *cb, int err)
 410 {
 411         struct tgt_last_committed_callback *ccb;
 412
 413         ccb = container_of0(cb, struct tgt_last_committed_callback, llcc_cb);
 414
 415         LASSERT(ccb->llcc_tgt != NULL);
 416         LASSERT(ccb->llcc_exp->exp_obd == ccb->llcc_tgt->lut_obd);
 417
 418         spin_lock(&ccb->llcc_tgt->lut_translock);
 419         if (ccb->llcc_transno > ccb->llcc_tgt->lut_obd->obd_last_committed)
 420                 ccb->llcc_tgt->lut_obd->obd_last_committed = ccb->llcc_transno;
 421
 422         LASSERT(ccb->llcc_exp);
 423         if (ccb->llcc_transno > ccb->llcc_exp->exp_last_committed) {
 424                 ccb->llcc_exp->exp_last_committed = ccb->llcc_transno;
 425                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 426                 ptlrpc_commit_replies(ccb->llcc_exp);
 427         } else {
 428                 spin_unlock(&ccb->llcc_tgt->lut_translock);
 429         }
 430         class_export_cb_put(ccb->llcc_exp);
 431         if (ccb->llcc_transno)
 432                 CDEBUG(D_HA, "%s: transno "LPD64" is committed\n",
 433                        ccb->llcc_tgt->lut_obd->obd_name, ccb->llcc_transno);
 434         OBD_FREE_PTR(ccb);
 435 }
 436
 437 int tgt_last_commit_cb_add(struct thandle *th, struct lu_target *tgt,
 438                            struct obd_export *exp, __u64 transno)
 439 {
 440         struct tgt_last_committed_callback      *ccb;
 441         struct dt_txn_commit_cb                 *dcb;
 442         int                                      rc;
 443
 444         OBD_ALLOC_PTR(ccb);
 445         if (ccb == NULL)
 446                 return -ENOMEM;
 447
 448         ccb->llcc_tgt = tgt;
 449         ccb->llcc_exp = class_export_cb_get(exp);
 450         ccb->llcc_transno = transno;
 451
 452         dcb = &ccb->llcc_cb;
 453         dcb->dcb_func = tgt_cb_last_committed;
 454         CFS_INIT_LIST_HEAD(&dcb->dcb_linkage);
 455         strncpy(dcb->dcb_name, "tgt_cb_last_committed", MAX_COMMIT_CB_STR_LEN);
 456         dcb->dcb_name[MAX_COMMIT_CB_STR_LEN - 1] = '\0';
 457
 458         rc = dt_trans_cb_add(th, dcb);
 459         if (rc) {
 460                 class_export_cb_put(exp);
 461                 OBD_FREE_PTR(ccb);
 462         }
 463
 464         if (exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 465                 /* report failure to force synchronous operation */
 466                 return -EPERM;
 467
 468         return rc;
 469 }
 470 EXPORT_SYMBOL(tgt_last_commit_cb_add);
 471
 472 struct tgt_new_client_callback {
 473         struct dt_txn_commit_cb  lncc_cb;
 474         struct obd_export       *lncc_exp;
 475 };
 476
 477 void tgt_cb_new_client(struct lu_env *env, struct thandle *th,
 478                        struct dt_txn_commit_cb *cb, int err)
 479 {
 480         struct tgt_new_client_callback *ccb;
 481
 482         ccb = container_of0(cb, struct tgt_new_client_callback, lncc_cb);
 483
 484         LASSERT(ccb->lncc_exp->exp_obd);
 485
 486         CDEBUG(D_RPCTRACE, "%s: committing for initial connect of %s\n",
 487                ccb->lncc_exp->exp_obd->obd_name,
 488                ccb->lncc_exp->exp_client_uuid.uuid);
 489
 490         spin_lock(&ccb->lncc_exp->exp_lock);
 491         /* XXX: Currently, we use per-export based sync/async policy for
 492          *      the update via OUT RPC, it is coarse-grained policy, and
 493          *      will be changed as per-request based by DNE II patches. */
 494         if (!ccb->lncc_exp->exp_keep_sync)
 495                 ccb->lncc_exp->exp_need_sync = 0;
 496
 497         spin_unlock(&ccb->lncc_exp->exp_lock);
 498         class_export_cb_put(ccb->lncc_exp);
 499
 500         OBD_FREE_PTR(ccb);
 501 }
 502
 503 int tgt_new_client_cb_add(struct thandle *th, struct obd_export *exp)
 504 {
 505         struct tgt_new_client_callback  *ccb;
 506         struct dt_txn_commit_cb         *dcb;
 507         int                              rc;
 508
 509         OBD_ALLOC_PTR(ccb);
 510         if (ccb == NULL)
 511                 return -ENOMEM;
 512
 513         ccb->lncc_exp = class_export_cb_get(exp);
 514
 515         dcb = &ccb->lncc_cb;
 516         dcb->dcb_func = tgt_cb_new_client;
 517         CFS_INIT_LIST_HEAD(&dcb->dcb_linkage);
 518         strncpy(dcb->dcb_name, "tgt_cb_new_client", MAX_COMMIT_CB_STR_LEN);
 519         dcb->dcb_name[MAX_COMMIT_CB_STR_LEN - 1] = '\0';
 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->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->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 EXPORT_SYMBOL(tgt_client_add);
 636
 637 int tgt_client_del(const struct lu_env *env, struct obd_export *exp)
 638 {
 639         struct tg_export_data   *ted = &exp->exp_target_data;
 640         struct lu_target        *tgt = class_exp2tgt(exp);
 641         int                      rc;
 642
 643         ENTRY;
 644
 645         LASSERT(ted->ted_lcd);
 646
 647         /* XXX if lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */
 648         if (!strcmp((char *)ted->ted_lcd->lcd_uuid,
 649                     (char *)tgt->lut_obd->obd_uuid.uuid) ||
 650             exp_connect_flags(exp) & OBD_CONNECT_LIGHTWEIGHT)
 651                 RETURN(0);
 652
 653         CDEBUG(D_INFO, "%s: del client at idx %u, off %lld, UUID '%s'\n",
 654                tgt->lut_obd->obd_name, ted->ted_lr_idx, ted->ted_lr_off,
 655                ted->ted_lcd->lcd_uuid);
 656
 657         /* Clear the bit _after_ zeroing out the client so we don't
 658            race with filter_client_add and zero out new clients.*/
 659         if (!test_bit(ted->ted_lr_idx, tgt->lut_client_bitmap)) {
 660                 CERROR("%s: client %u: bit already clear in bitmap!!\n",
 661                        tgt->lut_obd->obd_name, ted->ted_lr_idx);
 662                 LBUG();
 663         }
 664
 665         /* Do not erase record for recoverable client. */
 666         if (exp->exp_flags & OBD_OPT_FAILOVER)
 667                 RETURN(0);
 668
 669         /* Make sure the server's last_transno is up to date.
 670          * This should be done before zeroing client slot so last_transno will
 671          * be in server data or in client data in case of failure */
 672         rc = tgt_server_data_update(env, tgt, 0);
 673         if (rc != 0) {
 674                 CERROR("%s: failed to update server data, skip client %s "
 675                        "zeroing, rc %d\n", tgt->lut_obd->obd_name,
 676                        ted->ted_lcd->lcd_uuid, rc);
 677                 RETURN(rc);
 678         }
 679
 680         mutex_lock(&ted->ted_lcd_lock);
 681         memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid);
 682         rc = tgt_client_data_update(env, exp);
 683         mutex_unlock(&ted->ted_lcd_lock);
 684
 685         CDEBUG(rc == 0 ? D_INFO : D_ERROR,
 686                "%s: zeroing out client %s at idx %u (%llu), rc %d\n",
 687                tgt->lut_obd->obd_name, ted->ted_lcd->lcd_uuid,
 688                ted->ted_lr_idx, ted->ted_lr_off, rc);
 689         RETURN(rc);
 690 }
 691 EXPORT_SYMBOL(tgt_client_del);
 692
 693 /*
 694  * last_rcvd & last_committed update callbacks
 695  */
 696 int tgt_last_rcvd_update(const struct lu_env *env, struct lu_target *tgt,
 697                          struct dt_object *obj, __u64 opdata,
 698                          struct thandle *th, struct ptlrpc_request *req)
 699 {
 700         struct tgt_thread_info  *tti = tgt_th_info(env);
 701         struct tg_export_data   *ted;
 702         __u64                   *transno_p;
 703         int                      rc = 0;
 704         bool                     lw_client, update = false;
 705
 706         ENTRY;
 707
 708         ted = &req->rq_export->exp_target_data;
 709
 710         lw_client = exp_connect_flags(req->rq_export) & OBD_CONNECT_LIGHTWEIGHT;
 711         if (ted->ted_lr_idx < 0 && !lw_client)
 712                 /* ofd connect may cause transaction before export has
 713                  * last_rcvd slot */
 714                 RETURN(0);
 715
 716         tti->tti_transno = lustre_msg_get_transno(req->rq_reqmsg);
 717
 718         spin_lock(&tgt->lut_translock);
 719         if (th->th_result != 0) {
 720                 if (tti->tti_transno != 0) {
 721                         CERROR("%s: replay transno "LPU64" failed: rc = %d\n",
 722                                tgt_name(tgt), tti->tti_transno, th->th_result);
 723                 }
 724         } else if (tti->tti_transno == 0) {
 725                 tti->tti_transno = ++tgt->lut_last_transno;
 726         } else {
 727                 /* should be replay */
 728                 if (tti->tti_transno > tgt->lut_last_transno)
 729                         tgt->lut_last_transno = tti->tti_transno;
 730         }
 731         spin_unlock(&tgt->lut_translock);
 732
 733         /** VBR: set new versions */
 734         if (th->th_result == 0 && obj != NULL)
 735                 dt_version_set(env, obj, tti->tti_transno, th);
 736
 737         /* filling reply data */
 738         CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n",
 739                tti->tti_transno, tgt->lut_obd->obd_last_committed);
 740
 741         req->rq_transno = tti->tti_transno;
 742         lustre_msg_set_transno(req->rq_repmsg, tti->tti_transno);
 743
 744         /* if can't add callback, do sync write */
 745         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, req->rq_export,
 746                                                 tti->tti_transno);
 747
 748         if (lw_client) {
 749                 /* All operations performed by LW clients are synchronous and
 750                  * we store the committed transno in the last_rcvd header */
 751                 spin_lock(&tgt->lut_translock);
 752                 if (tti->tti_transno > tgt->lut_lsd.lsd_last_transno) {
 753                         tgt->lut_lsd.lsd_last_transno = tti->tti_transno;
 754                         update = true;
 755                 }
 756                 spin_unlock(&tgt->lut_translock);
 757                 /* Although lightweight (LW) connections have no slot in
 758                  * last_rcvd, we still want to maintain the in-memory
 759                  * lsd_client_data structure in order to properly handle reply
 760                  * reconstruction. */
 761         } else if (ted->ted_lr_off == 0) {
 762                 CERROR("%s: client idx %d has offset %lld\n",
 763                        tgt_name(tgt), ted->ted_lr_idx, ted->ted_lr_off);
 764                 RETURN(-EINVAL);
 765         }
 766
 767         /* if the export has already been disconnected, we have no last_rcvd
 768          * slot, update server data with latest transno then */
 769         if (ted->ted_lcd == NULL) {
 770                 CWARN("commit transaction for disconnected client %s: rc %d\n",
 771                       req->rq_export->exp_client_uuid.uuid, rc);
 772                 GOTO(srv_update, rc = 0);
 773         }
 774
 775         mutex_lock(&ted->ted_lcd_lock);
 776         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
 777         if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE ||
 778             lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) {
 779                 transno_p = &ted->ted_lcd->lcd_last_close_transno;
 780                 ted->ted_lcd->lcd_last_close_xid = req->rq_xid;
 781                 ted->ted_lcd->lcd_last_close_result = th->th_result;
 782         } else {
 783                 /* VBR: save versions in last_rcvd for reconstruct. */
 784                 __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg);
 785
 786                 if (pre_versions) {
 787                         ted->ted_lcd->lcd_pre_versions[0] = pre_versions[0];
 788                         ted->ted_lcd->lcd_pre_versions[1] = pre_versions[1];
 789                         ted->ted_lcd->lcd_pre_versions[2] = pre_versions[2];
 790                         ted->ted_lcd->lcd_pre_versions[3] = pre_versions[3];
 791                 }
 792                 transno_p = &ted->ted_lcd->lcd_last_transno;
 793                 ted->ted_lcd->lcd_last_xid = req->rq_xid;
 794                 ted->ted_lcd->lcd_last_result = th->th_result;
 795                 /* XXX: lcd_last_data is __u32 but intent_dispostion is __u64,
 796                  * see struct ldlm_reply->lock_policy_res1; */
 797                 ted->ted_lcd->lcd_last_data = opdata;
 798         }
 799
 800         /* Update transno in slot only if non-zero number, i.e. no errors */
 801         if (likely(tti->tti_transno != 0)) {
 802                 if (*transno_p > tti->tti_transno &&
 803                     !tgt->lut_no_reconstruct) {
 804                         CERROR("%s: trying to overwrite bigger transno:"
 805                                "on-disk: "LPU64", new: "LPU64" replay: %d. "
 806                                "see LU-617.\n", tgt_name(tgt), *transno_p,
 807                                tti->tti_transno, req_is_replay(req));
 808                         if (req_is_replay(req)) {
 809                                 spin_lock(&req->rq_export->exp_lock);
 810                                 req->rq_export->exp_vbr_failed = 1;
 811                                 spin_unlock(&req->rq_export->exp_lock);
 812                         }
 813                         mutex_unlock(&ted->ted_lcd_lock);
 814                         RETURN(req_is_replay(req) ? -EOVERFLOW : 0);
 815                 }
 816                 *transno_p = tti->tti_transno;
 817         }
 818
 819         if (!lw_client) {
 820                 tti->tti_off = ted->ted_lr_off;
 821                 rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 822                 if (rc < 0) {
 823                         mutex_unlock(&ted->ted_lcd_lock);
 824                         RETURN(rc);
 825                 }
 826         }
 827         mutex_unlock(&ted->ted_lcd_lock);
 828         EXIT;
 829 srv_update:
 830         if (update)
 831                 rc = tgt_server_data_write(env, tgt, th);
 832         return rc;
 833 }
 834
 835 /*
 836  * last_rcvd update for echo client simulation.
 837  * It updates last_rcvd client slot and version of object in
 838  * simple way but with all locks to simulate all drawbacks
 839  */
 840 int tgt_last_rcvd_update_echo(const struct lu_env *env, struct lu_target *tgt,
 841                               struct dt_object *obj, struct thandle *th,
 842                               struct obd_export *exp)
 843 {
 844         struct tgt_thread_info  *tti = tgt_th_info(env);
 845         struct tg_export_data   *ted = &exp->exp_target_data;
 846         int                      rc = 0;
 847
 848         ENTRY;
 849
 850         tti->tti_transno = 0;
 851
 852         spin_lock(&tgt->lut_translock);
 853         if (th->th_result == 0)
 854                 tti->tti_transno = ++tgt->lut_last_transno;
 855         spin_unlock(&tgt->lut_translock);
 856
 857         /** VBR: set new versions */
 858         if (th->th_result == 0 && obj != NULL)
 859                 dt_version_set(env, obj, tti->tti_transno, th);
 860
 861         /* if can't add callback, do sync write */
 862         th->th_sync |= !!tgt_last_commit_cb_add(th, tgt, exp,
 863                                                 tti->tti_transno);
 864
 865         LASSERT(ted->ted_lr_off > 0);
 866
 867         mutex_lock(&ted->ted_lcd_lock);
 868         LASSERT(ergo(tti->tti_transno == 0, th->th_result != 0));
 869         ted->ted_lcd->lcd_last_transno = tti->tti_transno;
 870         ted->ted_lcd->lcd_last_result = th->th_result;
 871
 872         tti->tti_off = ted->ted_lr_off;
 873         rc = tgt_client_data_write(env, tgt, ted->ted_lcd, &tti->tti_off, th);
 874         mutex_unlock(&ted->ted_lcd_lock);
 875         RETURN(rc);
 876 }
 877
 878 int tgt_clients_data_init(const struct lu_env *env, struct lu_target *tgt,
 879                           unsigned long last_size)
 880 {
 881         struct obd_device       *obd = tgt->lut_obd;
 882         struct lr_server_data   *lsd = &tgt->lut_lsd;
 883         struct lsd_client_data  *lcd = NULL;
 884         struct tg_export_data   *ted;
 885         int                      cl_idx;
 886         int                      rc = 0;
 887         loff_t                   off = lsd->lsd_client_start;
 888
 889         ENTRY;
 890
 891         CLASSERT(offsetof(struct lsd_client_data, lcd_padding) +
 892                  sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE);
 893
 894         OBD_ALLOC_PTR(lcd);
 895         if (lcd == NULL)
 896                 RETURN(-ENOMEM);
 897
 898         for (cl_idx = 0; off < last_size; cl_idx++) {
 899                 struct obd_export       *exp;
 900                 __u64                    last_transno;
 901
 902                 /* Don't assume off is incremented properly by
 903                  * read_record(), in case sizeof(*lcd)
 904                  * isn't the same as fsd->lsd_client_size.  */
 905                 off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size;
 906                 rc = tgt_client_data_read(env, tgt, lcd, &off, cl_idx);
 907                 if (rc) {
 908                         CERROR("%s: error reading last_rcvd %s idx %d off "
 909                                "%llu: rc = %d\n", tgt_name(tgt), LAST_RCVD,
 910                                cl_idx, off, rc);
 911                         rc = 0;
 912                         break; /* read error shouldn't cause startup to fail */
 913                 }
 914
 915                 if (lcd->lcd_uuid[0] == '\0') {
 916                         CDEBUG(D_INFO, "skipping zeroed client at offset %d\n",
 917                                cl_idx);
 918                         continue;
 919                 }
 920
 921                 last_transno = lcd_last_transno(lcd);
 922
 923                 /* These exports are cleaned up by disconnect, so they
 924                  * need to be set up like real exports as connect does.
 925                  */
 926                 CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64
 927                        " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx,
 928                        last_transno, lsd->lsd_last_transno, lcd_last_xid(lcd));
 929
 930                 exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid);
 931                 if (IS_ERR(exp)) {
 932                         if (PTR_ERR(exp) == -EALREADY) {
 933                                 /* export already exists, zero out this one */
 934                                 CERROR("%s: Duplicate export %s!\n",
 935                                        tgt_name(tgt), lcd->lcd_uuid);
 936                                 continue;
 937                         }
 938                         GOTO(err_out, rc = PTR_ERR(exp));
 939                 }
 940
 941                 ted = &exp->exp_target_data;
 942                 *ted->ted_lcd = *lcd;
 943
 944                 rc = tgt_client_add(env, exp, cl_idx);
 945                 LASSERTF(rc == 0, "rc = %d\n", rc); /* can't fail existing */
 946                 /* VBR: set export last committed version */
 947                 exp->exp_last_committed = last_transno;
 948                 spin_lock(&exp->exp_lock);
 949                 exp->exp_connecting = 0;
 950                 exp->exp_in_recovery = 0;
 951                 spin_unlock(&exp->exp_lock);
 952                 obd->obd_max_recoverable_clients++;
 953                 class_export_put(exp);
 954
 955                 /* Need to check last_rcvd even for duplicated exports. */
 956                 CDEBUG(D_OTHER, "client at idx %d has last_transno = "LPU64"\n",
 957                        cl_idx, last_transno);
 958
 959                 spin_lock(&tgt->lut_translock);
 960                 tgt->lut_last_transno = max(last_transno,
 961                                             tgt->lut_last_transno);
 962                 spin_unlock(&tgt->lut_translock);
 963         }
 964
 965 err_out:
 966         OBD_FREE_PTR(lcd);
 967         RETURN(rc);
 968 }
 969
 970 struct server_compat_data {
 971         __u32 rocompat;
 972         __u32 incompat;
 973         __u32 rocinit;
 974         __u32 incinit;
 975 };
 976
 977 static struct server_compat_data tgt_scd[] = {
 978         [LDD_F_SV_TYPE_MDT] = {
 979                 .rocompat = OBD_ROCOMPAT_LOVOBJID,
 980                 .incompat = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
 981                             OBD_INCOMPAT_FID | OBD_INCOMPAT_IAM_DIR |
 982                             OBD_INCOMPAT_LMM_VER | OBD_INCOMPAT_MULTI_OI,
 983                 .rocinit = OBD_ROCOMPAT_LOVOBJID,
 984                 .incinit = OBD_INCOMPAT_MDT | OBD_INCOMPAT_COMMON_LR |
 985                            OBD_INCOMPAT_MULTI_OI,
 986         },
 987         [LDD_F_SV_TYPE_OST] = {
 988                 .rocompat = 0,
 989                 .incompat = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR |
 990                             OBD_INCOMPAT_FID,
 991                 .rocinit = 0,
 992                 .incinit = OBD_INCOMPAT_OST | OBD_INCOMPAT_COMMON_LR,
 993         }
 994 };
 995
 996 int tgt_server_data_init(const struct lu_env *env, struct lu_target *tgt)
 997 {
 998         struct tgt_thread_info          *tti = tgt_th_info(env);
 999         struct lr_server_data           *lsd = &tgt->lut_lsd;
1000         unsigned long                    last_rcvd_size;
1001         __u32                            index;
1002         int                              rc, type;
1003
1004         rc = dt_attr_get(env, tgt->lut_last_rcvd, &tti->tti_attr, BYPASS_CAPA);
1005         if (rc)
1006                 RETURN(rc);
1007
1008         last_rcvd_size = (unsigned long)tti->tti_attr.la_size;
1009
1010         /* ensure padding in the struct is the correct size */
1011         CLASSERT(offsetof(struct lr_server_data, lsd_padding) +
1012                  sizeof(lsd->lsd_padding) == LR_SERVER_SIZE);
1013
1014         rc = server_name2index(tgt_name(tgt), &index, NULL);
1015         if (rc < 0) {
1016                 CERROR("%s: Can not get index from name: rc = %d\n",
1017                        tgt_name(tgt), rc);
1018                 RETURN(rc);
1019         }
1020         /* server_name2index() returns type */
1021         type = rc;
1022         if (type != LDD_F_SV_TYPE_MDT && type != LDD_F_SV_TYPE_OST) {
1023                 CERROR("%s: unknown target type %x\n", tgt_name(tgt), type);
1024                 RETURN(-EINVAL);
1025         }
1026
1027         /* last_rcvd on OST doesn't provide reconstruct support because there
1028          * may be up to 8 in-flight write requests per single slot in
1029          * last_rcvd client data
1030          */
1031         tgt->lut_no_reconstruct = (type == LDD_F_SV_TYPE_OST);
1032
1033         if (last_rcvd_size == 0) {
1034                 LCONSOLE_WARN("%s: new disk, initializing\n", tgt_name(tgt));
1035
1036                 memcpy(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid,
1037                        sizeof(lsd->lsd_uuid));
1038                 lsd->lsd_last_transno = 0;
1039                 lsd->lsd_mount_count = 0;
1040                 lsd->lsd_server_size = LR_SERVER_SIZE;
1041                 lsd->lsd_client_start = LR_CLIENT_START;
1042                 lsd->lsd_client_size = LR_CLIENT_SIZE;
1043                 lsd->lsd_subdir_count = OBJ_SUBDIR_COUNT;
1044                 lsd->lsd_osd_index = index;
1045                 lsd->lsd_feature_rocompat = tgt_scd[type].rocinit;
1046                 lsd->lsd_feature_incompat = tgt_scd[type].incinit;
1047         } else {
1048                 rc = tgt_server_data_read(env, tgt);
1049                 if (rc) {
1050                         CERROR("%s: error reading LAST_RCVD: rc= %d\n",
1051                                tgt_name(tgt), rc);
1052                         RETURN(rc);
1053                 }
1054                 if (strcmp(lsd->lsd_uuid, tgt->lut_obd->obd_uuid.uuid)) {
1055                         LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s "
1056                                            "using the wrong disk %s. Were the"
1057                                            " /dev/ assignments rearranged?\n",
1058                                            tgt->lut_obd->obd_uuid.uuid,
1059                                            lsd->lsd_uuid);
1060                         RETURN(-EINVAL);
1061                 }
1062
1063                 if (lsd->lsd_osd_index != index) {
1064                         LCONSOLE_ERROR_MSG(0x157, "%s: index %d in last rcvd "
1065                                            "is different with the index %d in"
1066                                            "config log, It might be disk"
1067                                            "corruption!\n", tgt_name(tgt),
1068                                            lsd->lsd_osd_index, index);
1069                         RETURN(-EINVAL);
1070                 }
1071         }
1072
1073         if (lsd->lsd_feature_incompat & ~tgt_scd[type].incompat) {
1074                 CERROR("%s: unsupported incompat filesystem feature(s) %x\n",
1075                        tgt_name(tgt),
1076                        lsd->lsd_feature_incompat & ~tgt_scd[type].incompat);
1077                 RETURN(-EINVAL);
1078         }
1079
1080         if (type == LDD_F_SV_TYPE_MDT)
1081                 lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID;
1082
1083         if (lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat) {
1084                 CERROR("%s: unsupported read-only filesystem feature(s) %x\n",
1085                        tgt_name(tgt),
1086                        lsd->lsd_feature_rocompat & ~tgt_scd[type].rocompat);
1087                 RETURN(-EINVAL);
1088         }
1089         /** Interop: evict all clients at first boot with 1.8 last_rcvd */
1090         if (type == LDD_F_SV_TYPE_MDT &&
1091             !(lsd->lsd_feature_compat & OBD_COMPAT_20)) {
1092                 if (last_rcvd_size > lsd->lsd_client_start) {
1093                         LCONSOLE_WARN("%s: mounting at first time on 1.8 FS, "
1094                                       "remove all clients for interop needs\n",
1095                                       tgt_name(tgt));
1096                         rc = tgt_truncate_last_rcvd(env, tgt,
1097                                                     lsd->lsd_client_start);
1098                         if (rc)
1099                                 RETURN(rc);
1100                         last_rcvd_size = lsd->lsd_client_start;
1101                 }
1102                 /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */
1103                 lsd->lsd_feature_compat |= OBD_COMPAT_20;
1104         }
1105
1106         spin_lock(&tgt->lut_translock);
1107         tgt->lut_last_transno = lsd->lsd_last_transno;
1108         spin_unlock(&tgt->lut_translock);
1109
1110         lsd->lsd_mount_count++;
1111
1112         CDEBUG(D_INODE, "=======,=BEGIN DUMPING LAST_RCVD========\n");
1113         CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n",
1114                tgt_name(tgt), tgt->lut_last_transno);
1115         CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n",
1116                tgt_name(tgt), lsd->lsd_mount_count);
1117         CDEBUG(D_INODE, "%s: server data size: %u\n",
1118                tgt_name(tgt), lsd->lsd_server_size);
1119         CDEBUG(D_INODE, "%s: per-client data start: %u\n",
1120                tgt_name(tgt), lsd->lsd_client_start);
1121         CDEBUG(D_INODE, "%s: per-client data size: %u\n",
1122                tgt_name(tgt), lsd->lsd_client_size);
1123         CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n",
1124                tgt_name(tgt), last_rcvd_size);
1125         CDEBUG(D_INODE, "%s: server subdir_count: %u\n",
1126                tgt_name(tgt), lsd->lsd_subdir_count);
1127         CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", tgt_name(tgt),
1128                last_rcvd_size <= lsd->lsd_client_start ? 0 :
1129                (last_rcvd_size - lsd->lsd_client_start) /
1130                 lsd->lsd_client_size);
1131         CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n");
1132
1133         if (lsd->lsd_server_size == 0 || lsd->lsd_client_start == 0 ||
1134             lsd->lsd_client_size == 0) {
1135                 CERROR("%s: bad last_rcvd contents!\n", tgt_name(tgt));
1136                 RETURN(-EINVAL);
1137         }
1138
1139         if (!tgt->lut_obd->obd_replayable)
1140                 CWARN("%s: recovery support OFF\n", tgt_name(tgt));
1141
1142         rc = tgt_clients_data_init(env, tgt, last_rcvd_size);
1143         if (rc < 0)
1144                 GOTO(err_client, rc);
1145
1146         spin_lock(&tgt->lut_translock);
1147         /* obd_last_committed is used for compatibility
1148          * with other lustre recovery code */
1149         tgt->lut_obd->obd_last_committed = tgt->lut_last_transno;
1150         spin_unlock(&tgt->lut_translock);
1151
1152         tgt->lut_obd->u.obt.obt_mount_count = lsd->lsd_mount_count;
1153         tgt->lut_obd->u.obt.obt_instance = (__u32)lsd->lsd_mount_count;
1154
1155         /* save it, so mount count and last_transno is current */
1156         rc = tgt_server_data_update(env, tgt, 0);
1157         if (rc < 0)
1158                 GOTO(err_client, rc);
1159
1160         RETURN(0);
1161
1162 err_client:
1163         class_disconnect_exports(tgt->lut_obd);
1164         return rc;
1165 }
1166
1167 /* add credits for last_rcvd update */
1168 int tgt_txn_start_cb(const struct lu_env *env, struct thandle *th,
1169                      void *cookie)
1170 {
1171         struct lu_target        *tgt = cookie;
1172         struct tgt_session_info *tsi;
1173         int                      rc;
1174
1175         /* if there is no session, then this transaction is not result of
1176          * request processing but some local operation */
1177         if (env->le_ses == NULL)
1178                 return 0;
1179
1180         LASSERT(tgt->lut_last_rcvd);
1181         tsi = tgt_ses_info(env);
1182         /* OFD may start transaction without export assigned */
1183         if (tsi->tsi_exp == NULL)
1184                 return 0;
1185
1186         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
1187                                      sizeof(struct lsd_client_data),
1188                                      tsi->tsi_exp->exp_target_data.ted_lr_off,
1189                                      th);
1190         if (rc)
1191                 return rc;
1192
1193         rc = dt_declare_record_write(env, tgt->lut_last_rcvd,
1194                                      sizeof(struct lr_server_data), 0, th);
1195         if (rc)
1196                 return rc;
1197
1198         if (tsi->tsi_vbr_obj != NULL &&
1199             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu))
1200                 rc = dt_declare_version_set(env, tsi->tsi_vbr_obj, th);
1201
1202         return rc;
1203 }
1204
1205 /* Update last_rcvd records with latests transaction data */
1206 int tgt_txn_stop_cb(const struct lu_env *env, struct thandle *th,
1207                     void *cookie)
1208 {
1209         struct lu_target        *tgt = cookie;
1210         struct tgt_session_info *tsi;
1211         struct tgt_thread_info  *tti = tgt_th_info(env);
1212         struct dt_object        *obj = NULL;
1213         int                      rc;
1214         bool                     echo_client;
1215
1216         if (env->le_ses == NULL)
1217                 return 0;
1218
1219         tsi = tgt_ses_info(env);
1220         /* OFD may start transaction without export assigned */
1221         if (tsi->tsi_exp == NULL)
1222                 return 0;
1223
1224         echo_client = (tgt_ses_req(tsi) == NULL);
1225
1226         if (tti->tti_has_trans && !echo_client) {
1227                 if (tti->tti_mult_trans == 0) {
1228                         CDEBUG(D_HA, "More than one transaction "LPU64"\n",
1229                                tti->tti_transno);
1230                         RETURN(0);
1231                 }
1232                 /* we need another transno to be assigned */
1233                 tti->tti_transno = 0;
1234         } else if (th->th_result == 0) {
1235                 tti->tti_has_trans = 1;
1236         }
1237
1238         if (tsi->tsi_vbr_obj != NULL &&
1239             !lu_object_remote(&tsi->tsi_vbr_obj->do_lu)) {
1240                 obj = tsi->tsi_vbr_obj;
1241         }
1242
1243         if (unlikely(echo_client)) /* echo client special case */
1244                 rc = tgt_last_rcvd_update_echo(env, tgt, obj, th,
1245                                                tsi->tsi_exp);
1246         else
1247                 rc = tgt_last_rcvd_update(env, tgt, obj, tsi->tsi_opdata, th,
1248                                           tgt_ses_req(tsi));
1249         return rc;
1250 }