lustre/target/update_trans.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 /*
   4  * Copyright (c) 2015, 2017, Intel Corporation.
   5  */
   6
   7 /*
   8  * This file implements the update distribute transaction API.
   9  *
  10  * To manage the cross-MDT operation (distribute operation) transaction,
  11  * the transaction will also be separated two layers on MD stack, top
  12  * transaction and sub transaction.
  13  *
  14  * During the distribute operation, top transaction is created in the LOD
  15  * layer, and represent the operation. Sub transaction is created by
  16  * each OSD or OSP. Top transaction start/stop will trigger all of its sub
  17  * transaction start/stop. Top transaction (the whole operation) is committed
  18  * only all of its sub transaction are committed.
  19  *
  20  * there are three kinds of transactions
  21  * 1. local transaction: All updates are in a single local OSD.
  22  * 2. Remote transaction: All Updates are only in the remote OSD,
  23  *    i.e. locally all updates are in OSP.
  24  * 3. Mixed transaction: Updates are both in local OSD and remote
  25  *    OSD.
  26  *
  27  * Author: Di Wang <di.wang@intel.com>
  28  */
  29
  30 #define DEBUG_SUBSYSTEM S_CLASS
  31
  32 #include <linux/kthread.h>
  33 #include <lu_target.h>
  34 #include <lustre_log.h>
  35 #include <lustre_update.h>
  36 #include <obd.h>
  37 #include <obd_class.h>
  38 #include <tgt_internal.h>
  39
  40 #include <tgt_internal.h>
  41 /**
  42  * Dump top mulitple thandle
  43  *
  44  * Dump top multiple thandle and all of its sub thandle to the debug log.
  45  *
  46  * \param[in]mask       debug mask
  47  * \param[in]top_th     top_thandle to be dumped
  48  */
  49 static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
  50                                       __u32 mask)
  51 {
  52         struct sub_thandle      *st;
  53
  54         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
  55         CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d batchid %llu\n",
  56                tmt->tmt_master_sub_dt ?
  57                tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
  58                "NULL",
  59                tmt, kref_read(&tmt->tmt_refcount), tmt->tmt_committed,
  60                tmt->tmt_result, tmt->tmt_batchid);
  61
  62         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
  63                 struct sub_thandle_cookie *stc;
  64
  65                 CDEBUG(mask, "st %p obd %s committed %d started %d stopped %d "
  66                        "result %d sub_th %p\n",
  67                        st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
  68                        st->st_committed, st->st_started, st->st_stopped,
  69                        st->st_result, st->st_sub_th);
  70
  71                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
  72                         CDEBUG(mask, " cookie "DFID".%u\n",
  73                                PLOGID(&stc->stc_cookie.lgc_lgl),
  74                                stc->stc_cookie.lgc_index);
  75                 }
  76         }
  77 }
  78
  79 /**
  80  * Declare write update to sub device
  81  *
  82  * Declare Write updates llog records to the sub device during distribute
  83  * transaction.
  84  *
  85  * \param[in] env       execution environment
  86  * \param[in] record    update records being written
  87  * \param[in] sub_th    sub transaction handle
  88  * \param[in] record_size total update record size
  89  *
  90  * \retval              0 if writing succeeds
  91  * \retval              negative errno if writing fails
  92  */
  93 static int sub_declare_updates_write(const struct lu_env *env,
  94                                      struct llog_update_record *record,
  95                                      struct thandle *sub_th, size_t record_size)
  96 {
  97         struct llog_ctxt        *ctxt;
  98         struct dt_device        *dt = sub_th->th_dev;
  99         int                     left = record_size;
 100         int rc;
 101
 102         /* If ctxt is NULL, it means not need to write update,
 103          * for example if the the OSP is used to connect to OST */
 104         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 105                                 LLOG_UPDATELOG_ORIG_CTXT);
 106
 107         /* Not ready to record updates yet. */
 108         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 109                 llog_ctxt_put(ctxt);
 110                 return 0;
 111         }
 112
 113         rc = llog_declare_add(env, ctxt->loc_handle,
 114                               &record->lur_hdr, sub_th);
 115         if (rc < 0)
 116                 GOTO(out_put, rc);
 117
 118         while (left > ctxt->loc_chunk_size) {
 119                 rc = llog_declare_add(env, ctxt->loc_handle,
 120                                       &record->lur_hdr, sub_th);
 121                 if (rc < 0)
 122                         GOTO(out_put, rc);
 123
 124                 left -= ctxt->loc_chunk_size;
 125         }
 126
 127 out_put:
 128         llog_ctxt_put(ctxt);
 129
 130         return rc;
 131 }
 132
 133 /**
 134  * write update to sub device
 135  *
 136  * Write llog update record to the sub device during distribute
 137  * transaction. If it succeeds, llog cookie of the record will be
 138  * returned by @cookie.
 139  *
 140  * \param[in] env       execution environment
 141  * \param[in] record    update records being written
 142  * \param[in] sub_th    sub transaction handle
 143  * \param[out] cookie   llog cookie of the update record.
 144  *
 145  * \retval              1 if writing succeeds
 146  * \retval              negative errno if writing fails
 147  */
 148 static int sub_updates_write(const struct lu_env *env,
 149                              struct llog_update_record *record,
 150                              struct sub_thandle *sub_th)
 151 {
 152         struct dt_device *dt = sub_th->st_dt;
 153         struct llog_ctxt *ctxt;
 154         struct llog_update_record *lur = NULL;
 155         __u32 update_count = 0;
 156         __u32 param_count = 0;
 157         __u32 last_update_count = 0;
 158         __u32 last_param_count = 0;
 159         char *start;
 160         char *cur;
 161         char *next;
 162         struct sub_thandle_cookie *stc;
 163         size_t reclen;
 164         bool eof = false;
 165         int rc;
 166         ENTRY;
 167
 168         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 169                                 LLOG_UPDATELOG_ORIG_CTXT);
 170         /* If ctxt == NULL, then it means updates on OST (only happens
 171          * during migration), and we do not track those updates for now */
 172         /* If ctxt->loc_handle == NULL, then it does not need to record
 173          * update, usually happens in error handler path */
 174         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 175                 llog_ctxt_put(ctxt);
 176                 RETURN(0);
 177         }
 178
 179         /* Since the cross-MDT updates will includes both local
 180          * and remote updates, the update ops count must > 1 */
 181         LASSERT(record->lur_update_rec.ur_update_count > 1);
 182         LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
 183                  "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
 184                  llog_update_record_size(record));
 185
 186         /*
 187          * If its size > llog chunk_size, then write current chunk to the update
 188          * llog, NB the padding should >= LLOG_MIN_REC_SIZE.
 189          *
 190          * So check padding length is either >= LLOG_MIN_REC_SIZE or is 0
 191          * (record length just matches the chunk size).
 192          */
 193
 194         reclen = record->lur_hdr.lrh_len;
 195         if (reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 196             reclen == ctxt->loc_chunk_size) {
 197                 OBD_ALLOC_PTR(stc);
 198                 if (stc == NULL)
 199                         GOTO(llog_put, rc = -ENOMEM);
 200                 INIT_LIST_HEAD(&stc->stc_list);
 201
 202                 rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
 203                               &stc->stc_cookie, sub_th->st_sub_th);
 204
 205                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 206                        dt->dd_lu_dev.ld_obd->obd_name,
 207                        PLOGID(&stc->stc_cookie.lgc_lgl),
 208                        stc->stc_cookie.lgc_index, rc);
 209
 210                 if (rc > 0) {
 211                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 212                         rc = 0;
 213                 } else {
 214                         OBD_FREE_PTR(stc);
 215                 }
 216
 217                 GOTO(llog_put, rc);
 218         }
 219
 220         /* Split the records into chunk_size update record */
 221         OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
 222         if (lur == NULL)
 223                 GOTO(llog_put, rc = -ENOMEM);
 224
 225         memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
 226         lur->lur_update_rec.ur_update_count = 0;
 227         lur->lur_update_rec.ur_param_count = 0;
 228         start = (char *)&record->lur_update_rec.ur_ops;
 229         cur = next = start;
 230         do {
 231                 if (update_count < record->lur_update_rec.ur_update_count)
 232                         next = (char *)update_op_next_op(
 233                                                 (struct update_op *)cur);
 234                 else if (param_count < record->lur_update_rec.ur_param_count)
 235                         next = (char *)update_param_next_param(
 236                                                 (struct update_param *)cur);
 237                 else
 238                         eof = true;
 239
 240                 reclen = __llog_update_record_size(
 241                                 __update_records_size(next - start));
 242                 if ((reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 243                      reclen == ctxt->loc_chunk_size) &&
 244                     !eof) {
 245                         cur = next;
 246
 247                         if (update_count <
 248                             record->lur_update_rec.ur_update_count)
 249                                 update_count++;
 250                         else if (param_count <
 251                                  record->lur_update_rec.ur_param_count)
 252                                 param_count++;
 253                         continue;
 254                 }
 255
 256                 lur->lur_update_rec.ur_update_count = update_count -
 257                                                       last_update_count;
 258                 lur->lur_update_rec.ur_param_count = param_count -
 259                                                      last_param_count;
 260                 memcpy(&lur->lur_update_rec.ur_ops, start, cur - start);
 261                 lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 262
 263                 LASSERT(lur->lur_hdr.lrh_len ==
 264                          __llog_update_record_size(
 265                                 __update_records_size(cur - start)));
 266                 LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
 267
 268                 update_records_dump(&lur->lur_update_rec, D_INFO, true);
 269
 270                 OBD_ALLOC_PTR(stc);
 271                 if (stc == NULL)
 272                         GOTO(llog_put, rc = -ENOMEM);
 273                 INIT_LIST_HEAD(&stc->stc_list);
 274
 275                 rc = llog_add(env, ctxt->loc_handle, &lur->lur_hdr,
 276                               &stc->stc_cookie, sub_th->st_sub_th);
 277
 278                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 279                         dt->dd_lu_dev.ld_obd->obd_name,
 280                         PLOGID(&stc->stc_cookie.lgc_lgl),
 281                         stc->stc_cookie.lgc_index, rc);
 282
 283                 if (rc > 0) {
 284                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 285                         rc = 0;
 286                 } else {
 287                         OBD_FREE_PTR(stc);
 288                         GOTO(llog_put, rc);
 289                 }
 290
 291                 last_update_count = update_count;
 292                 last_param_count = param_count;
 293                 start = cur;
 294                 lur->lur_update_rec.ur_update_count = 0;
 295                 lur->lur_update_rec.ur_param_count = 0;
 296                 lur->lur_update_rec.ur_flags |= UPDATE_RECORD_CONTINUE;
 297         } while (!eof);
 298
 299 llog_put:
 300         if (lur != NULL)
 301                 OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
 302         llog_ctxt_put(ctxt);
 303
 304         RETURN(rc);
 305 }
 306
 307 /**
 308  * Prepare the update records.
 309  *
 310  * Merge params and ops into the update records, then initializing
 311  * the update buffer.
 312  *
 313  * During transaction execution phase, parameters and update ops
 314  * are collected in two different buffers (see lod_updates_pack()),
 315  * during transaction stop, it needs to be merged in one buffer,
 316  * so it will be written in the update log.
 317  *
 318  * \param[in] env       execution environment
 319  * \param[in] tmt       top_multiple_thandle for distribute txn
 320  *
 321  * \retval              0 if merging succeeds.
 322  * \retval              negaitive errno if merging fails.
 323  */
 324 static int prepare_writing_updates(const struct lu_env *env,
 325                                    struct top_multiple_thandle *tmt)
 326 {
 327         struct thandle_update_records   *tur = tmt->tmt_update_records;
 328         struct llog_update_record       *lur;
 329         struct update_params *params;
 330         size_t params_size;
 331         size_t update_size;
 332
 333         if (tur == NULL || tur->tur_update_records == NULL ||
 334             tur->tur_update_params == NULL)
 335                 return 0;
 336
 337         lur = tur->tur_update_records;
 338         /* Extends the update records buffer if needed */
 339         params_size = update_params_size(tur->tur_update_params,
 340                                          tur->tur_update_param_count);
 341         LASSERT(lur->lur_update_rec.ur_param_count == 0);
 342         update_size = round_up(llog_update_record_size(lur) + params_size, 8);
 343         if (update_size > tur->tur_update_records_buf_size) {
 344                 int rc;
 345
 346                 rc = tur_update_records_extend(tur, update_size);
 347                 if (rc < 0)
 348                         return rc;
 349
 350                 lur = tur->tur_update_records;
 351         }
 352
 353         params = update_records_get_params(&lur->lur_update_rec);
 354         memcpy(params, tur->tur_update_params, params_size);
 355
 356         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
 357         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
 358         /* Init update record header */
 359         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 360         lur->lur_hdr.lrh_type = UPDATE_REC;
 361
 362         /* Dump updates for debugging purpose */
 363         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 364
 365         return 0;
 366 }
 367
 368 /**
 369  * Top thandle commit callback
 370  *
 371  * This callback will be called when all of sub transactions are committed.
 372  *
 373  * \param[in] th        top thandle to be committed.
 374  */
 375 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
 376 {
 377         struct lu_target *lut;
 378         ENTRY;
 379
 380         LASSERT(kref_read(&tmt->tmt_refcount) > 0);
 381
 382         top_multiple_thandle_dump(tmt, D_HA);
 383         tmt->tmt_committed = 1;
 384         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
 385         if (lut->lut_tdtd && lut->lut_tdtd->tdtd_commit_task)
 386                 wake_up_process(lut->lut_tdtd->tdtd_commit_task);
 387
 388         RETURN_EXIT;
 389 }
 390
 391 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
 392                                        struct dt_device *dt_dev)
 393 {
 394         struct sub_thandle *st;
 395
 396         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 397                 if (st->st_dt == dt_dev)
 398                         return st;
 399         }
 400         return NULL;
 401 }
 402 EXPORT_SYMBOL(lookup_sub_thandle);
 403
 404 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
 405                                        struct dt_device *dt_dev)
 406 {
 407         struct sub_thandle *st;
 408
 409         OBD_ALLOC_PTR(st);
 410         if (st == NULL)
 411                 RETURN(ERR_PTR(-ENOMEM));
 412
 413         INIT_LIST_HEAD(&st->st_sub_list);
 414         INIT_LIST_HEAD(&st->st_cookie_list);
 415         st->st_dt = dt_dev;
 416
 417         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
 418         return st;
 419 }
 420
 421 static void sub_trans_commit_cb_internal(struct top_multiple_thandle *tmt,
 422                                          struct thandle *sub_th, int err)
 423 {
 424         struct sub_thandle      *st;
 425         bool                    all_committed = true;
 426
 427         /* Check if all sub thandles are committed */
 428         spin_lock(&tmt->tmt_sub_lock);
 429         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 430                 if (st->st_sub_th == sub_th) {
 431                         st->st_committed = 1;
 432                         st->st_result = err;
 433                 }
 434                 if (!st->st_committed)
 435                         all_committed = false;
 436         }
 437         spin_unlock(&tmt->tmt_sub_lock);
 438
 439         if (tmt->tmt_result == 0)
 440                 tmt->tmt_result = err;
 441
 442         if (all_committed)
 443                 top_trans_committed_cb(tmt);
 444
 445         top_multiple_thandle_dump(tmt, D_INFO);
 446         top_multiple_thandle_put(tmt);
 447         RETURN_EXIT;
 448 }
 449
 450 /**
 451  * sub thandle commit callback
 452  *
 453  * Mark the sub thandle to be committed and if all sub thandle are committed
 454  * notify the top thandle.
 455  *
 456  * \param[in] env       execution environment
 457  * \param[in] sub_th    sub thandle being committed
 458  * \param[in] cb        commit callback
 459  * \param[in] err       trans result
 460  */
 461 static void sub_trans_commit_cb(struct lu_env *env,
 462                                 struct thandle *sub_th,
 463                                 struct dt_txn_commit_cb *cb, int err)
 464 {
 465         struct top_multiple_thandle *tmt = cb->dcb_data;
 466
 467         sub_trans_commit_cb_internal(tmt, sub_th, err);
 468 }
 469
 470 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
 471                                     struct top_multiple_thandle *tmt)
 472 {
 473         LASSERT(st->st_sub_th != NULL);
 474         top_multiple_thandle_get(tmt);
 475         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
 476         st->st_commit_dcb.dcb_data = tmt;
 477         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
 478         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
 479 }
 480
 481 /**
 482  * Sub thandle stop call back
 483  *
 484  * After sub thandle is stopped, it will call this callback to notify
 485  * the top thandle.
 486  *
 487  * \param[in] th        sub thandle to be stopped
 488  * \param[in] rc        result of sub trans
 489  */
 490 static void sub_trans_stop_cb(struct lu_env *env,
 491                               struct thandle *sub_th,
 492                               struct dt_txn_commit_cb *cb, int err)
 493 {
 494         struct sub_thandle              *st;
 495         struct top_multiple_thandle     *tmt = cb->dcb_data;
 496         ENTRY;
 497
 498         spin_lock(&tmt->tmt_sub_lock);
 499         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 500                 if (st->st_stopped)
 501                         continue;
 502
 503                 if (st->st_dt == sub_th->th_dev) {
 504                         st->st_stopped = 1;
 505                         st->st_result = err;
 506                         break;
 507                 }
 508         }
 509         spin_unlock(&tmt->tmt_sub_lock);
 510
 511         wake_up(&tmt->tmt_stop_waitq);
 512         RETURN_EXIT;
 513 }
 514
 515 static void sub_thandle_register_stop_cb(struct sub_thandle *st,
 516                                          struct top_multiple_thandle *tmt)
 517 {
 518         st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
 519         st->st_stop_dcb.dcb_data = tmt;
 520         st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
 521         INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
 522         dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
 523 }
 524
 525 /**
 526  * Create sub thandle
 527  *
 528  * Create transaction handle for sub_thandle
 529  *
 530  * \param[in] env       execution environment
 531  * \param[in] th        top thandle
 532  * \param[in] st        sub_thandle
 533  *
 534  * \retval              0 if creation succeeds.
 535  * \retval              negative errno if creation fails.
 536  */
 537 int sub_thandle_trans_create(const struct lu_env *env,
 538                              struct top_thandle *top_th,
 539                              struct sub_thandle *st)
 540 {
 541         struct thandle *sub_th;
 542
 543         sub_th = dt_trans_create(env, st->st_dt);
 544         if (IS_ERR(sub_th))
 545                 return PTR_ERR(sub_th);
 546
 547         sub_th->th_top = &top_th->tt_super;
 548         st->st_sub_th = sub_th;
 549
 550         sub_th->th_wait_submit = 1;
 551         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
 552         return 0;
 553 }
 554
 555 /**
 556  * Create the top transaction.
 557  *
 558  * Create the top transaction on the master device. It will create a top
 559  * thandle and a sub thandle on the master device.
 560  *
 561  * \param[in] env               execution environment
 562  * \param[in] master_dev        master_dev the top thandle will be created
 563  *
 564  * \retval                      pointer to the created thandle.
 565  * \retval                      ERR_PTR(errno) if creation failed.
 566  */
 567 struct thandle *
 568 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
 569 {
 570         struct top_thandle      *top_th;
 571         struct thandle          *child_th;
 572
 573         OBD_ALLOC_GFP(top_th, sizeof(*top_th), GFP_NOFS);
 574         if (top_th == NULL)
 575                 return ERR_PTR(-ENOMEM);
 576
 577         top_th->tt_super.th_top = &top_th->tt_super;
 578
 579         if (master_dev != NULL) {
 580                 child_th = dt_trans_create(env, master_dev);
 581                 if (IS_ERR(child_th)) {
 582                         OBD_FREE_PTR(top_th);
 583                         return child_th;
 584                 }
 585
 586                 child_th->th_top = &top_th->tt_super;
 587                 child_th->th_wait_submit = 1;
 588                 top_th->tt_master_sub_thandle = child_th;
 589         }
 590         return &top_th->tt_super;
 591 }
 592 EXPORT_SYMBOL(top_trans_create);
 593
 594 /**
 595  * Declare write update transaction
 596  *
 597  * Check if there are updates being recorded in this transaction,
 598  * it will write the record into the disk.
 599  *
 600  * \param[in] env       execution environment
 601  * \param[in] tmt       top multiple transaction handle
 602  *
 603  * \retval              0 if writing succeeds
 604  * \retval              negative errno if writing fails
 605  */
 606 static int declare_updates_write(const struct lu_env *env,
 607                                  struct top_multiple_thandle *tmt)
 608 {
 609         struct llog_update_record *record;
 610         struct sub_thandle *st;
 611         int rc = 0;
 612
 613         record = tmt->tmt_update_records->tur_update_records;
 614         /* Declare update write for all other target */
 615         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 616                 if (st->st_sub_th == NULL)
 617                         continue;
 618
 619                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
 620                                                tmt->tmt_record_size);
 621                 if (rc < 0)
 622                         break;
 623         }
 624
 625         return rc;
 626 }
 627
 628 /**
 629  * Assign batchid to the distribute transaction.
 630  *
 631  * Assign batchid to the distribute transaction
 632  *
 633  * \param[in] tmt       distribute transaction
 634  */
 635 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 636 {
 637         struct target_distribute_txn_data *tdtd;
 638         struct dt_device *dt = new->tmt_master_sub_dt;
 639         struct sub_thandle *st;
 640
 641         LASSERT(dt != NULL);
 642         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 643         spin_lock(&tdtd->tdtd_batchid_lock);
 644         new->tmt_batchid = tdtd->tdtd_batchid++;
 645         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
 646         spin_unlock(&tdtd->tdtd_batchid_lock);
 647         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 648                 if (st->st_sub_th != NULL)
 649                         sub_thandle_register_commit_cb(st, new);
 650         }
 651         top_multiple_thandle_get(new);
 652         top_multiple_thandle_dump(new, D_INFO);
 653 }
 654
 655 /**
 656  * Insert distribute transaction to the distribute txn list.
 657  *
 658  * Insert distribute transaction to the distribute txn list.
 659  *
 660  * \param[in] new       the distribute txn to be inserted.
 661  */
 662 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
 663 {
 664         struct dt_device *dt = new->tmt_master_sub_dt;
 665         struct top_multiple_thandle *tmt;
 666         struct target_distribute_txn_data *tdtd;
 667         struct sub_thandle *st;
 668         bool    at_head = false;
 669
 670         LASSERT(dt != NULL);
 671         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 672
 673         /* have a reference so the competing commit thread can't release it*/
 674         top_multiple_thandle_get(new);
 675
 676         spin_lock(&tdtd->tdtd_batchid_lock);
 677         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
 678                 if (new->tmt_batchid > tmt->tmt_batchid) {
 679                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
 680                         break;
 681                 }
 682         }
 683         if (list_empty(&new->tmt_commit_list)) {
 684                 at_head = true;
 685                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
 686         }
 687         spin_unlock(&tdtd->tdtd_batchid_lock);
 688
 689         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 690                 if (st->st_sub_th != NULL)
 691                         sub_thandle_register_commit_cb(st, new);
 692         }
 693
 694         top_multiple_thandle_dump(new, D_INFO);
 695         if (new->tmt_committed && at_head && tdtd->tdtd_commit_task)
 696                 wake_up_process(tdtd->tdtd_commit_task);
 697 }
 698
 699 /**
 700  * Prepare cross-MDT operation.
 701  *
 702  * Create the update record buffer to record updates for cross-MDT operation,
 703  * add master sub transaction to tt_sub_trans_list, and declare the update
 704  * writes.
 705  *
 706  * During updates packing, all of parameters will be packed in
 707  * tur_update_params, and updates will be packed in tur_update_records.
 708  * Then in transaction stop, parameters and updates will be merged
 709  * into one updates buffer.
 710  *
 711  * And also master thandle will be added to the sub_th list, so it will be
 712  * easy to track the commit status.
 713  *
 714  * \param[in] env       execution environment
 715  * \param[in] th        top transaction handle
 716  *
 717  * \retval              0 if preparation succeeds.
 718  * \retval              negative errno if preparation fails.
 719  */
 720 static int prepare_multiple_node_trans(const struct lu_env *env,
 721                                        struct top_multiple_thandle *tmt)
 722 {
 723         struct thandle_update_records   *tur;
 724         int                             rc;
 725         ENTRY;
 726
 727         if (tmt->tmt_update_records == NULL) {
 728                 tur = &update_env_info(env)->uti_tur;
 729                 rc = check_and_prepare_update_record(env, tur);
 730                 if (rc < 0)
 731                         RETURN(rc);
 732
 733                 tmt->tmt_update_records = tur;
 734                 distribute_txn_assign_batchid(tmt);
 735         }
 736
 737         rc = declare_updates_write(env, tmt);
 738
 739         RETURN(rc);
 740 }
 741
 742 /**
 743  * start the top transaction.
 744  *
 745  * Start all of its sub transactions, then start master sub transaction.
 746  *
 747  * \param[in] env               execution environment
 748  * \param[in] master_dev        master_dev the top thandle will be start
 749  * \param[in] th                top thandle
 750  *
 751  * \retval                      0 if transaction start succeeds.
 752  * \retval                      negative errno if start fails.
 753  */
 754 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
 755                     struct thandle *th)
 756 {
 757         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 758                                                        tt_super);
 759         struct sub_thandle              *st;
 760         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
 761         int                             rc = 0;
 762         ENTRY;
 763
 764         if (tmt == NULL) {
 765                 if (th->th_sync)
 766                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 767                 if (th->th_local)
 768                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 769                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
 770                                     top_th->tt_master_sub_thandle);
 771                 RETURN(rc);
 772         }
 773
 774         tmt = top_th->tt_multiple_thandle;
 775         rc = prepare_multiple_node_trans(env, tmt);
 776         if (rc < 0)
 777                 RETURN(rc);
 778
 779         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 780                 if (st->st_sub_th == NULL)
 781                         continue;
 782                 if (th->th_sync)
 783                         st->st_sub_th->th_sync = th->th_sync;
 784                 if (th->th_local)
 785                         st->st_sub_th->th_local = th->th_local;
 786                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
 787                                     st->st_sub_th);
 788                 if (rc != 0)
 789                         GOTO(out, rc);
 790
 791                 LASSERT(st->st_started == 0);
 792                 st->st_started = 1;
 793         }
 794 out:
 795         th->th_result = rc;
 796         RETURN(rc);
 797 }
 798 EXPORT_SYMBOL(top_trans_start);
 799
 800 /**
 801  * Check whether we need write updates record
 802  *
 803  * Check if the updates for the top_thandle needs to be writen
 804  * to all targets. Only if the transaction succeeds and the updates
 805  * number > 2, it will write the updates,
 806  *
 807  * \params [in] top_th  top thandle.
 808  *
 809  * \retval              true if it needs to write updates
 810  * \retval              false if it does not need to write updates
 811  **/
 812 static bool top_check_write_updates(struct top_thandle *top_th)
 813 {
 814         struct top_multiple_thandle     *tmt;
 815         struct thandle_update_records   *tur;
 816
 817         /* Do not write updates to records if the transaction fails */
 818         if (top_th->tt_super.th_result != 0)
 819                 return false;
 820
 821         tmt = top_th->tt_multiple_thandle;
 822         if (tmt == NULL)
 823                 return false;
 824
 825         tur = tmt->tmt_update_records;
 826         if (tur == NULL)
 827                 return false;
 828
 829         /* Hmm, false update records, since the cross-MDT operation
 830          * should includes both local and remote updates, so the
 831          * updates count should >= 2 */
 832         if (tur->tur_update_records == NULL ||
 833             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
 834                 return false;
 835
 836         return true;
 837 }
 838
 839 /**
 840  * Check if top transaction is stopped
 841  *
 842  * Check if top transaction is stopped, only if all sub transaction
 843  * is stopped, then the top transaction is stopped.
 844  *
 845  * \param [in] top_th   top thandle
 846  *
 847  * \retval              true if the top transaction is stopped.
 848  * \retval              false if the top transaction is not stopped.
 849  */
 850 static bool top_trans_is_stopped(struct top_thandle *top_th)
 851 {
 852         struct top_multiple_thandle     *tmt;
 853         struct sub_thandle              *st;
 854         bool                    all_stopped = true;
 855
 856         tmt = top_th->tt_multiple_thandle;
 857         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 858                 if (!st->st_stopped && st->st_sub_th != NULL) {
 859                         all_stopped = false;
 860                         break;
 861                 }
 862
 863                 if (st->st_result != 0 &&
 864                     top_th->tt_super.th_result == 0)
 865                         top_th->tt_super.th_result = st->st_result;
 866         }
 867
 868         return all_stopped;
 869 }
 870
 871 /**
 872  * Wait result of top transaction
 873  *
 874  * Wait until all sub transaction get its result.
 875  *
 876  * \param [in] top_th   top thandle.
 877  *
 878  * \retval              the result of top thandle.
 879  */
 880 static int top_trans_wait_result(struct top_thandle *top_th)
 881 {
 882         wait_event_idle(top_th->tt_multiple_thandle->tmt_stop_waitq,
 883                         top_trans_is_stopped(top_th));
 884
 885         RETURN(top_th->tt_super.th_result);
 886 }
 887
 888 /**
 889  * Stop the top transaction.
 890  *
 891  * Stop the transaction on the master device first, then stop transactions
 892  * on other sub devices.
 893  *
 894  * \param[in] env               execution environment
 895  * \param[in] master_dev        master_dev the top thandle will be created
 896  * \param[in] th                top thandle
 897  *
 898  * \retval                      0 if stop transaction succeeds.
 899  * \retval                      negative errno if stop transaction fails.
 900  */
 901 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
 902                    struct thandle *th)
 903 {
 904         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 905                                                        tt_super);
 906         struct sub_thandle              *st;
 907         struct sub_thandle              *master_st;
 908         struct top_multiple_thandle     *tmt;
 909         struct thandle_update_records   *tur;
 910         bool                            write_updates = false;
 911         int                     rc = 0;
 912         ENTRY;
 913
 914         if (likely(top_th->tt_multiple_thandle == NULL)) {
 915                 LASSERT(master_dev != NULL);
 916
 917                 if (th->th_sync)
 918                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 919                 if (th->th_local)
 920                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 921                 rc = dt_trans_stop(env, master_dev,
 922                                    top_th->tt_master_sub_thandle);
 923                 OBD_FREE_PTR(top_th);
 924                 RETURN(rc);
 925         }
 926
 927         tmt = top_th->tt_multiple_thandle;
 928         tur = tmt->tmt_update_records;
 929
 930         /* Note: we need stop the master thandle first, then the stop
 931          * callback will fill the master transno in the update logs,
 932          * then these update logs will be sent to other MDTs */
 933         /* get the master sub thandle */
 934         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
 935         write_updates = top_check_write_updates(top_th);
 936
 937         /* Step 1: write the updates log on Master MDT */
 938         if (master_st != NULL && master_st->st_sub_th != NULL &&
 939             write_updates) {
 940                 struct llog_update_record *lur;
 941
 942                 /* Merge the parameters and updates into one buffer */
 943                 rc = prepare_writing_updates(env, tmt);
 944                 if (rc < 0) {
 945                         CERROR("%s: cannot prepare updates: rc = %d\n",
 946                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 947                         th->th_result = rc;
 948                         write_updates = false;
 949                         GOTO(stop_master_trans, rc);
 950                 }
 951
 952                 lur = tur->tur_update_records;
 953                 /* Write updates to the master MDT */
 954                 rc = sub_updates_write(env, lur, master_st);
 955
 956                 /* Cleanup the common parameters in the update records,
 957                  * master transno callback might add more parameters.
 958                  * and we need merge the update records again in the
 959                  * following */
 960                 if (tur->tur_update_params != NULL)
 961                         lur->lur_update_rec.ur_param_count = 0;
 962
 963                 if (rc < 0) {
 964                         CERROR("%s: write updates failed: rc = %d\n",
 965                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 966                         th->th_result = rc;
 967                         write_updates = false;
 968                         GOTO(stop_master_trans, rc);
 969                 }
 970         }
 971
 972 stop_master_trans:
 973         /* Step 2: Stop the transaction on the master MDT, and fill the
 974          * master transno in the update logs to other MDT. */
 975         if (master_st != NULL && master_st->st_sub_th != NULL) {
 976                 if (th->th_local)
 977                         master_st->st_sub_th->th_local = th->th_local;
 978                 if (th->th_sync)
 979                         master_st->st_sub_th->th_sync = th->th_sync;
 980                 master_st->st_sub_th->th_result = th->th_result;
 981                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
 982                 /* If it does not write_updates, then we call submit callback
 983                  * here, otherwise callback is done through
 984                  * osd(osp)_trans_commit_cb() */
 985                 if (!master_st->st_started &&
 986                     !list_empty(&tmt->tmt_commit_list))
 987                         sub_trans_commit_cb_internal(tmt,
 988                                                 master_st->st_sub_th, rc);
 989                 if (rc < 0) {
 990                         CERROR("%s: stop trans failed: rc = %d\n",
 991                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 992                         th->th_result = rc;
 993                         GOTO(stop_other_trans, rc);
 994                 } else if (tur != NULL && tur->tur_update_records != NULL) {
 995                         struct llog_update_record *lur;
 996
 997                         lur = tur->tur_update_records;
 998                         if (lur->lur_update_rec.ur_master_transno == 0)
 999                                 /* Update master transno after master stop
1000                                  * callback */
1001                                 lur->lur_update_rec.ur_master_transno =
1002                                                 tgt_th_info(env)->tti_transno;
1003                 }
1004         }
1005
1006         /* Step 3: write updates to other MDTs */
1007         if (write_updates) {
1008                 struct llog_update_record *lur;
1009                 if (CFS_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
1010                         if (cfs_fail_val == 1) {
1011                                 long timeout = cfs_time_seconds(1) / 10;
1012
1013                                 CFS_RACE(OBD_FAIL_OUT_OBJECT_MISS);
1014                                 set_current_state(TASK_UNINTERRUPTIBLE);
1015                                 schedule_timeout(schedule_timeout(timeout));
1016                                 cfs_fail_loc = 0;
1017                         }
1018                         cfs_fail_val++;
1019                 }
1020
1021                 /* Stop callback of master will add more updates and also update
1022                  * master transno, so merge the parameters and updates into one
1023                  * buffer again */
1024                 rc = prepare_writing_updates(env, tmt);
1025                 if (rc < 0) {
1026                         CERROR("%s: prepare updates failed: rc = %d\n",
1027                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1028                         th->th_result = rc;
1029                         GOTO(stop_other_trans, rc);
1030                 }
1031                 lur = tur->tur_update_records;
1032                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1033                                     st_sub_list) {
1034                         if (st->st_sub_th == NULL || st == master_st ||
1035                             st->st_sub_th->th_result < 0)
1036                                 continue;
1037
1038                         rc = sub_updates_write(env, lur, st);
1039                         if (rc < 0) {
1040                                 CERROR("%s: write updates failed: rc = %d\n",
1041                                        st->st_dt->dd_lu_dev.ld_obd->obd_name,
1042                                        rc);
1043                                 th->th_result = rc;
1044                                 break;
1045                         }
1046                 }
1047         }
1048
1049 stop_other_trans:
1050         /* Step 4: Stop the transaction on other MDTs */
1051         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1052                 if (st == master_st || st->st_sub_th == NULL)
1053                         continue;
1054
1055                 if (th->th_sync)
1056                         st->st_sub_th->th_sync = th->th_sync;
1057                 if (th->th_local)
1058                         st->st_sub_th->th_local = th->th_local;
1059                 st->st_sub_th->th_result = th->th_result;
1060                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1061                                    st->st_sub_th);
1062                 if (rc < 0) {
1063                         CERROR("%s: stop trans failed: rc = %d\n",
1064                                st->st_dt->dd_lu_dev.ld_obd->obd_name, rc);
1065                         if (th->th_result == 0)
1066                                 th->th_result = rc;
1067                 }
1068         }
1069
1070         rc = top_trans_wait_result(top_th);
1071
1072         tmt->tmt_result = rc;
1073
1074         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1075          * multiple node transaction, the top transaction will be destroyed. */
1076         top_multiple_thandle_put(tmt);
1077         OBD_FREE_PTR(top_th);
1078         RETURN(rc);
1079 }
1080 EXPORT_SYMBOL(top_trans_stop);
1081
1082 /**
1083  * Create top_multiple_thandle for top_thandle
1084  *
1085  * Create top_mutilple_thandle to manage the mutiple node transaction
1086  * for top_thandle, and it also needs to add master sub thandle to the
1087  * sub trans list now.
1088  *
1089  * \param[in] env       execution environment
1090  * \param[in] top_th    the top thandle
1091  *
1092  * \retval      0 if creation succeeds
1093  * \retval      negative errno if creation fails
1094  */
1095 int top_trans_create_tmt(const struct lu_env *env,
1096                          struct top_thandle *top_th)
1097 {
1098         struct top_multiple_thandle *tmt;
1099
1100         OBD_ALLOC_PTR(tmt);
1101         if (tmt == NULL)
1102                 return -ENOMEM;
1103
1104         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1105         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1106         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1107         kref_init(&tmt->tmt_refcount);
1108         spin_lock_init(&tmt->tmt_sub_lock);
1109         init_waitqueue_head(&tmt->tmt_stop_waitq);
1110
1111         top_th->tt_multiple_thandle = tmt;
1112
1113         return 0;
1114 }
1115
1116 static struct sub_thandle *
1117 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1118                                 struct thandle *sub_th)
1119 {
1120         struct sub_thandle *st;
1121
1122         /* create and init sub th to the top trans list */
1123         st = create_sub_thandle(top_th->tt_multiple_thandle,
1124                                 sub_th->th_dev);
1125         if (IS_ERR(st))
1126                 return st;
1127
1128         st->st_sub_th = sub_th;
1129
1130         sub_th->th_top = &top_th->tt_super;
1131         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
1132         return st;
1133 }
1134
1135 /**
1136  * Get sub thandle.
1137  *
1138  * Get sub thandle from the top thandle according to the sub dt_device.
1139  *
1140  * \param[in] env       execution environment
1141  * \param[in] th        thandle on the top layer.
1142  * \param[in] sub_dt    sub dt_device used to get sub transaction
1143  *
1144  * \retval              thandle of sub transaction if succeed
1145  * \retval              PTR_ERR(errno) if failed
1146  */
1147 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1148                                       struct thandle *th,
1149                                       struct dt_device *sub_dt)
1150 {
1151         struct sub_thandle      *st = NULL;
1152         struct sub_thandle      *master_st = NULL;
1153         struct top_thandle      *top_th;
1154         struct thandle          *sub_th = NULL;
1155         int                     rc = 0;
1156         ENTRY;
1157
1158         top_th = container_of(th, struct top_thandle, tt_super);
1159
1160         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1161                 RETURN(top_th->tt_master_sub_thandle);
1162
1163         if (top_th->tt_multiple_thandle != NULL) {
1164                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1165                 if (st != NULL)
1166                         RETURN(st->st_sub_th);
1167         }
1168
1169         sub_th = dt_trans_create(env, sub_dt);
1170         if (IS_ERR(sub_th))
1171                 RETURN(sub_th);
1172
1173         /* Create top_multiple_thandle if necessary */
1174         if (top_th->tt_multiple_thandle == NULL) {
1175                 struct top_multiple_thandle *tmt;
1176
1177                 rc = top_trans_create_tmt(env, top_th);
1178                 if (rc < 0)
1179                         GOTO(stop_trans, rc);
1180
1181                 tmt = top_th->tt_multiple_thandle;
1182
1183                 /* Add master sub th to the top trans list */
1184                 tmt->tmt_master_sub_dt =
1185                         top_th->tt_master_sub_thandle->th_dev;
1186                 master_st = create_sub_thandle_with_thandle(top_th,
1187                                         top_th->tt_master_sub_thandle);
1188                 if (IS_ERR(master_st)) {
1189                         rc = PTR_ERR(master_st);
1190                         master_st = NULL;
1191                         GOTO(stop_trans, rc);
1192                 }
1193         }
1194
1195         /* create and init sub th to the top trans list */
1196         st = create_sub_thandle_with_thandle(top_th, sub_th);
1197         if (IS_ERR(st)) {
1198                 rc = PTR_ERR(st);
1199                 st = NULL;
1200                 GOTO(stop_trans, rc);
1201         }
1202         st->st_sub_th->th_wait_submit = 1;
1203 stop_trans:
1204         if (rc < 0) {
1205                 if (master_st != NULL) {
1206                         list_del(&master_st->st_sub_list);
1207                         OBD_FREE_PTR(master_st);
1208                 }
1209                 sub_th->th_result = rc;
1210                 dt_trans_stop(env, sub_dt, sub_th);
1211                 sub_th = ERR_PTR(rc);
1212         }
1213
1214         RETURN(sub_th);
1215 }
1216 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1217
1218 /**
1219  * Top multiple thandle destroy
1220  *
1221  * Destroy multiple thandle and all its sub thandle.
1222  *
1223  * \param[in] kref      Pointer to struct kref
1224  */
1225 void top_multiple_thandle_destroy(struct kref *kref)
1226 {
1227         struct top_multiple_thandle *tmt;
1228         struct sub_thandle *st;
1229         struct sub_thandle *tmp;
1230
1231         tmt = container_of(kref, struct top_multiple_thandle, tmt_refcount);
1232
1233         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1234         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1235                                  st_sub_list) {
1236                 struct sub_thandle_cookie *stc;
1237                 struct sub_thandle_cookie *tmp;
1238
1239                 list_del(&st->st_sub_list);
1240                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1241                                          stc_list) {
1242                         list_del(&stc->stc_list);
1243                         OBD_FREE_PTR(stc);
1244                 }
1245                 OBD_FREE_PTR(st);
1246         }
1247         OBD_FREE_PTR(tmt);
1248 }
1249 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1250
1251 /**
1252  * Cancel the update log on MDTs
1253  *
1254  * Cancel the update log on MDTs then destroy the thandle.
1255  *
1256  * \param[in] env       execution environment
1257  * \param[in] tmt       the top multiple thandle whose updates records
1258  *                      will be cancelled.
1259  *
1260  * \retval              0 if cancellation succeeds.
1261  * \retval              negative errno if cancellation fails.
1262  */
1263 static int distribute_txn_cancel_records(const struct lu_env *env,
1264                                          struct top_multiple_thandle *tmt)
1265 {
1266         struct sub_thandle *st;
1267         ENTRY;
1268
1269         if (CFS_FAIL_CHECK(OBD_FAIL_TGT_TXN_NO_CANCEL))
1270                 RETURN(0);
1271
1272         top_multiple_thandle_dump(tmt, D_INFO);
1273         /* Cancel update logs on other MDTs */
1274         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1275                 struct llog_ctxt        *ctxt;
1276                 struct obd_device       *obd;
1277                 struct llog_cookie      *cookie;
1278                 struct sub_thandle_cookie *stc;
1279                 int rc;
1280
1281                 obd = st->st_dt->dd_lu_dev.ld_obd;
1282                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1283                 if (ctxt == NULL)
1284                         continue;
1285                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1286                         cookie = &stc->stc_cookie;
1287                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1288                                 continue;
1289
1290                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1291                                                      cookie);
1292                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1293                                DFID".%u: rc = %d\n", obd->obd_name,
1294                                tmt->tmt_batchid, PLOGID(&cookie->lgc_lgl),
1295                                cookie->lgc_index, rc);
1296                 }
1297
1298                 llog_ctxt_put(ctxt);
1299         }
1300
1301         RETURN(0);
1302 }
1303
1304 struct distribute_txn_bid_data {
1305         struct dt_txn_commit_cb  dtbd_cb;
1306         struct target_distribute_txn_data      *dtbd_tdtd;
1307         __u64                    dtbd_batchid;
1308 };
1309
1310 /**
1311  * callback of updating commit batchid
1312  *
1313  * Updating commit batchid then wake up the commit thread to cancel the
1314  * records.
1315  *
1316  * \param[in]env        execution environment
1317  * \param[in]th         thandle to updating commit batchid
1318  * \param[in]cb         commit callback
1319  * \param[in]err        result of thandle
1320  */
1321 static void distribute_txn_batchid_cb(struct lu_env *env,
1322                                       struct thandle *th,
1323                                       struct dt_txn_commit_cb *cb,
1324                                       int err)
1325 {
1326         struct distribute_txn_bid_data          *dtbd = NULL;
1327         struct target_distribute_txn_data       *tdtd;
1328
1329         dtbd = container_of(cb, struct distribute_txn_bid_data, dtbd_cb);
1330         tdtd = dtbd->dtbd_tdtd;
1331
1332         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1333               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1334         spin_lock(&tdtd->tdtd_batchid_lock);
1335         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1336             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1337                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1338         spin_unlock(&tdtd->tdtd_batchid_lock);
1339         if (atomic_dec_and_test(&tdtd->tdtd_refcount))
1340                 wake_up_process(tdtd->tdtd_commit_task);
1341
1342         OBD_FREE_PTR(dtbd);
1343 }
1344
1345 /**
1346  * Update the commit batchid in disk
1347  *
1348  * Update commit batchid in the disk, after this is committed, it can start
1349  * to cancel the update records.
1350  *
1351  * \param[in] env       execution environment
1352  * \param[in] tdtd      distribute transaction structure
1353  * \param[in] batchid   commit batchid to be updated
1354  *
1355  * \retval              0 if update succeeds.
1356  * \retval              negative errno if update fails.
1357  */
1358 static int
1359 distribute_txn_commit_batchid_update(const struct lu_env *env,
1360                               struct target_distribute_txn_data *tdtd,
1361                               __u64 batchid)
1362 {
1363         struct distribute_txn_bid_data  *dtbd = NULL;
1364         struct thandle          *th;
1365         struct lu_buf            buf;
1366         __u64                    tmp;
1367         __u64                    off;
1368         int                      rc;
1369         ENTRY;
1370
1371         OBD_ALLOC_PTR(dtbd);
1372         if (dtbd == NULL)
1373                 RETURN(-ENOMEM);
1374         dtbd->dtbd_batchid = batchid;
1375         dtbd->dtbd_tdtd = tdtd;
1376         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1377         atomic_inc(&tdtd->tdtd_refcount);
1378
1379         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1380         if (IS_ERR(th)) {
1381                 atomic_dec(&tdtd->tdtd_refcount);
1382                 OBD_FREE_PTR(dtbd);
1383                 RETURN(PTR_ERR(th));
1384         }
1385
1386         tmp = cpu_to_le64(batchid);
1387         buf.lb_buf = &tmp;
1388         buf.lb_len = sizeof(tmp);
1389         off = 0;
1390
1391         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1392                                      th);
1393         if (rc < 0)
1394                 GOTO(stop, rc);
1395
1396         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1397         if (rc < 0)
1398                 GOTO(stop, rc);
1399
1400         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1401         if (rc < 0)
1402                 GOTO(stop, rc);
1403
1404         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1405                              &off, th);
1406
1407         CDEBUG(D_INFO, "%s: update batchid %llu: rc = %d\n",
1408                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1409
1410 stop:
1411         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1412         if (rc < 0) {
1413                 atomic_dec(&tdtd->tdtd_refcount);
1414                 OBD_FREE_PTR(dtbd);
1415         }
1416         RETURN(rc);
1417 }
1418
1419 /**
1420  * Init commit batchid for distribute transaction.
1421  *
1422  * Initialize the batchid object and get commit batchid from the object.
1423  *
1424  * \param[in] env       execution environment
1425  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1426  *
1427  * \retval              0 if initialization succeeds.
1428  * \retval              negative errno if initialization fails.
1429  **/
1430 static int
1431 distribute_txn_commit_batchid_init(const struct lu_env *env,
1432                                    struct target_distribute_txn_data *tdtd)
1433 {
1434         struct tgt_thread_info  *tti = tgt_th_info(env);
1435         struct lu_target        *lut = tdtd->tdtd_lut;
1436         struct lu_attr          *attr = &tti->tti_attr;
1437         struct lu_fid           *fid = &tti->tti_fid1;
1438         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1439         struct dt_object        *dt_obj = NULL;
1440         struct lu_buf           buf;
1441         __u64                   tmp;
1442         __u64                   off;
1443         int                     rc;
1444         ENTRY;
1445
1446         memset(attr, 0, sizeof(*attr));
1447         attr->la_valid = LA_MODE;
1448         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1449         dof->dof_type = dt_mode_to_dft(S_IFREG);
1450
1451         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1452
1453         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1454                                    attr);
1455         if (IS_ERR(dt_obj)) {
1456                 rc = PTR_ERR(dt_obj);
1457                 dt_obj = NULL;
1458                 GOTO(out_put, rc);
1459         }
1460
1461         tdtd->tdtd_batchid_obj = dt_obj;
1462
1463         buf.lb_buf = &tmp;
1464         buf.lb_len = sizeof(tmp);
1465         off = 0;
1466         rc = dt_read(env, dt_obj, &buf, &off);
1467         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1468                 CERROR("%s can't read last committed batchid: rc = %d\n",
1469                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1470                 if (rc > 0)
1471                         rc = -EINVAL;
1472                 GOTO(out_put, rc);
1473         } else if (rc == buf.lb_len) {
1474                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1475                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1476                        tdtd->tdtd_lut->lut_obd->obd_name,
1477                        tdtd->tdtd_committed_batchid);
1478                 rc = 0;
1479         }
1480
1481 out_put:
1482         if (rc < 0 && dt_obj != NULL) {
1483                 dt_object_put(env, dt_obj);
1484                 tdtd->tdtd_batchid_obj = NULL;
1485         }
1486         return rc;
1487 }
1488
1489 #ifndef TASK_IDLE
1490 #define TASK_IDLE TASK_INTERRUPTIBLE
1491 #endif
1492
1493 /**
1494  * manage the distribute transaction thread
1495  *
1496  * Distribute transaction are linked to the list, and once the distribute
1497  * transaction is committed, it will update the last committed batchid first,
1498  * after it is committed, it will cancel the records.
1499  *
1500  * \param[in] _arg      argument for commit thread
1501  *
1502  * \retval              0 if thread is running successfully
1503  * \retval              negative errno if the thread can not be run.
1504  */
1505 static int distribute_txn_commit_thread(void *_arg)
1506 {
1507         struct target_distribute_txn_data *tdtd = _arg;
1508         struct lu_env           *env = &tdtd->tdtd_env;
1509         LIST_HEAD(list);
1510         int                      rc;
1511         struct top_multiple_thandle *tmt;
1512         struct top_multiple_thandle *tmp;
1513         __u64                    batchid = 0, committed;
1514
1515         ENTRY;
1516
1517
1518         CDEBUG(D_HA, "%s: start commit thread committed batchid %llu\n",
1519                tdtd->tdtd_lut->lut_obd->obd_name,
1520                tdtd->tdtd_committed_batchid);
1521
1522         while (({set_current_state(TASK_IDLE);
1523                  !kthread_should_stop(); })) {
1524                 spin_lock(&tdtd->tdtd_batchid_lock);
1525                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1526                                          tmt_commit_list) {
1527                         if (tmt->tmt_committed == 0)
1528                                 break;
1529
1530                         /* Note: right now, replay is based on master MDT
1531                          * transno, but cancellation is based on batchid.
1532                          * so we do not try to cancel the update log until
1533                          * the recoverying is done, unless the update records
1534                          * batchid < committed_batchid. */
1535                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1536                                 __set_current_state(TASK_RUNNING);
1537                                 list_move_tail(&tmt->tmt_commit_list, &list);
1538                         } else if (!test_bit(OBDF_RECOVERING, tdtd->tdtd_lut->lut_obd->obd_flags)) {
1539                                 __set_current_state(TASK_RUNNING);
1540                                 LASSERTF(tmt->tmt_batchid >= batchid,
1541                                          "tmt %px tmt_batchid: %llu, batchid %llu\n",
1542                                          tmt, tmt->tmt_batchid, batchid);
1543                                 /* There are three types of distribution
1544                                  * transaction result
1545                                  *
1546                                  * 1. If tmt_result < 0, it means the
1547                                  * distribution transaction fails, which should
1548                                  * be rare, because once declare phase succeeds,
1549                                  * the operation should succeeds anyway. Note in
1550                                  * this case, we will still update batchid so
1551                                  * cancellation would be stopped.
1552                                  *
1553                                  * 2. If tmt_result == 0, it means the
1554                                  * distribution transaction succeeds, and we
1555                                  * will update batchid.
1556                                  *
1557                                  * 3. If tmt_result > 0, it means distribute
1558                                  * transaction is not yet committed on every
1559                                  * node, but we need release this tmt before
1560                                  * that, which usuually happens during umount.
1561                                  */
1562                                 if (tmt->tmt_result <= 0)
1563                                         batchid = tmt->tmt_batchid;
1564                                 list_move_tail(&tmt->tmt_commit_list, &list);
1565                         }
1566                 }
1567                 spin_unlock(&tdtd->tdtd_batchid_lock);
1568
1569                 CDEBUG(D_HA, "%s: batchid: %llu committed batchid "
1570                        "%llu\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1571                        tdtd->tdtd_committed_batchid);
1572                 /* update globally committed on a storage */
1573                 if (batchid > tdtd->tdtd_committed_batchid) {
1574                         rc = distribute_txn_commit_batchid_update(env, tdtd,
1575                                                              batchid);
1576                         if (rc == 0)
1577                                 batchid = 0;
1578                 }
1579                 /* cancel the records for committed batchid's */
1580                 /* XXX: should we postpone cancel's till the end of recovery? */
1581                 committed = tdtd->tdtd_committed_batchid;
1582                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1583                         if (tmt->tmt_batchid > committed)
1584                                 break;
1585                         __set_current_state(TASK_RUNNING);
1586                         list_del_init(&tmt->tmt_commit_list);
1587                         if (tmt->tmt_result <= 0)
1588                                 distribute_txn_cancel_records(env, tmt);
1589                         top_multiple_thandle_put(tmt);
1590                 }
1591
1592                 if (!task_is_running(current))
1593                         schedule();
1594
1595                 if (CFS_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
1596                         set_current_state(TASK_UNINTERRUPTIBLE);
1597                         schedule_timeout(cfs_time_seconds(5));
1598                 }
1599         }
1600
1601         while (({set_current_state(TASK_IDLE);
1602                  atomic_read(&tdtd->tdtd_refcount) != 0; }))
1603                 schedule();
1604         __set_current_state(TASK_RUNNING);
1605
1606         spin_lock(&tdtd->tdtd_batchid_lock);
1607         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1608                                  tmt_commit_list)
1609                 list_move_tail(&tmt->tmt_commit_list, &list);
1610         spin_unlock(&tdtd->tdtd_batchid_lock);
1611
1612         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1613                tdtd->tdtd_lut->lut_obd->obd_name);
1614         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1615                 list_del_init(&tmt->tmt_commit_list);
1616                 top_multiple_thandle_dump(tmt, D_HA);
1617                 top_multiple_thandle_put(tmt);
1618         }
1619         RETURN(0);
1620 }
1621
1622 /**
1623  * Start llog cancel thread
1624  *
1625  * Start llog cancel(master/slave) thread on LOD
1626  *
1627  * \param[in]lclt       cancel log thread to be started.
1628  *
1629  * \retval              0 if the thread is started successfully.
1630  * \retval              negative errno if the thread is not being
1631  *                      started.
1632  */
1633 int distribute_txn_init(const struct lu_env *env,
1634                         struct lu_target *lut,
1635                         struct target_distribute_txn_data *tdtd,
1636                         __u32 index)
1637 {
1638         struct task_struct      *task;
1639         int                     rc;
1640         ENTRY;
1641
1642         INIT_LIST_HEAD(&tdtd->tdtd_list);
1643         INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
1644         INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
1645         spin_lock_init(&tdtd->tdtd_batchid_lock);
1646         spin_lock_init(&tdtd->tdtd_replay_list_lock);
1647         tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
1648         tdtd->tdtd_replay_ready = 0;
1649
1650         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1651
1652         init_waitqueue_head(&tdtd->tdtd_recovery_threads_waitq);
1653         atomic_set(&tdtd->tdtd_refcount, 0);
1654         atomic_set(&tdtd->tdtd_recovery_threads_count, 0);
1655
1656         tdtd->tdtd_lut = lut;
1657         if (lut->lut_bottom->dd_rdonly)
1658                 RETURN(0);
1659
1660         rc = distribute_txn_commit_batchid_init(env, tdtd);
1661         if (rc != 0)
1662                 RETURN(rc);
1663
1664         rc = lu_env_init(&tdtd->tdtd_env, LCT_LOCAL | LCT_MD_THREAD);
1665         if (rc)
1666                 RETURN(rc);
1667
1668         task = kthread_create(distribute_txn_commit_thread, tdtd, "dist_txn-%u",
1669                               index);
1670         if (IS_ERR(task)) {
1671                 lu_env_fini(&tdtd->tdtd_env);
1672                 RETURN(PTR_ERR(task));
1673         }
1674         tdtd->tdtd_commit_task = task;
1675         wake_up_process(task);
1676
1677         RETURN(0);
1678 }
1679 EXPORT_SYMBOL(distribute_txn_init);
1680
1681 /**
1682  * Stop llog cancel thread
1683  *
1684  * Stop llog cancel(master/slave) thread on LOD and also destory
1685  * all of transaction in the list.
1686  *
1687  * \param[in]lclt       cancel log thread to be stopped.
1688  */
1689 void distribute_txn_fini(const struct lu_env *env,
1690                          struct target_distribute_txn_data *tdtd)
1691 {
1692         struct top_multiple_thandle *tmt;
1693         LIST_HEAD(list);
1694
1695         /* Stop cancel thread */
1696         if (!tdtd->tdtd_commit_task)
1697                 return;
1698
1699         kthread_stop(tdtd->tdtd_commit_task);
1700         tdtd->tdtd_commit_task = NULL;
1701
1702         spin_lock(&tdtd->tdtd_batchid_lock);
1703         list_splice_init(&tdtd->tdtd_list, &list);
1704         spin_unlock(&tdtd->tdtd_batchid_lock);
1705
1706         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1707                tdtd->tdtd_lut->lut_obd->obd_name);
1708         while ((tmt = list_first_entry_or_null(&list,
1709                                                struct top_multiple_thandle,
1710                                                tmt_commit_list)) != NULL) {
1711                 list_del_init(&tmt->tmt_commit_list);
1712                 top_multiple_thandle_dump(tmt, D_HA);
1713                 top_multiple_thandle_put(tmt);
1714         }
1715
1716         lu_env_fini(&tdtd->tdtd_env);
1717
1718         dtrq_list_destroy(tdtd);
1719         if (tdtd->tdtd_batchid_obj != NULL) {
1720                 dt_object_put(env, tdtd->tdtd_batchid_obj);
1721                 tdtd->tdtd_batchid_obj = NULL;
1722         }
1723 }
1724 EXPORT_SYMBOL(distribute_txn_fini);