lustre/target/update_trans.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.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2015, 2017, Intel Corporation.
  24  */
  25 /*
  26  * lustre/target/update_trans.c
  27  *
  28  * This file implements the update distribute transaction API.
  29  *
  30  * To manage the cross-MDT operation (distribute operation) transaction,
  31  * the transaction will also be separated two layers on MD stack, top
  32  * transaction and sub transaction.
  33  *
  34  * During the distribute operation, top transaction is created in the LOD
  35  * layer, and represent the operation. Sub transaction is created by
  36  * each OSD or OSP. Top transaction start/stop will trigger all of its sub
  37  * transaction start/stop. Top transaction (the whole operation) is committed
  38  * only all of its sub transaction are committed.
  39  *
  40  * there are three kinds of transactions
  41  * 1. local transaction: All updates are in a single local OSD.
  42  * 2. Remote transaction: All Updates are only in the remote OSD,
  43  *    i.e. locally all updates are in OSP.
  44  * 3. Mixed transaction: Updates are both in local OSD and remote
  45  *    OSD.
  46  *
  47  * Author: Di Wang <di.wang@intel.com>
  48  */
  49
  50 #define DEBUG_SUBSYSTEM S_CLASS
  51
  52 #include <linux/kthread.h>
  53 #include <lu_target.h>
  54 #include <lustre_log.h>
  55 #include <lustre_update.h>
  56 #include <obd.h>
  57 #include <obd_class.h>
  58 #include <tgt_internal.h>
  59
  60 #include <tgt_internal.h>
  61 /**
  62  * Dump top mulitple thandle
  63  *
  64  * Dump top multiple thandle and all of its sub thandle to the debug log.
  65  *
  66  * \param[in]mask       debug mask
  67  * \param[in]top_th     top_thandle to be dumped
  68  */
  69 static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
  70                                       __u32 mask)
  71 {
  72         struct sub_thandle      *st;
  73
  74         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
  75         CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d batchid %llu\n",
  76                tmt->tmt_master_sub_dt ?
  77                tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
  78                "NULL",
  79                tmt, atomic_read(&tmt->tmt_refcount), tmt->tmt_committed,
  80                tmt->tmt_result, tmt->tmt_batchid);
  81
  82         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
  83                 struct sub_thandle_cookie *stc;
  84
  85                 CDEBUG(mask, "st %p obd %s committed %d stopped %d sub_th %p\n",
  86                        st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
  87                        st->st_committed, st->st_stopped, st->st_sub_th);
  88
  89                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
  90                         CDEBUG(mask, " cookie "DFID".%u\n",
  91                                PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
  92                                stc->stc_cookie.lgc_index);
  93                 }
  94         }
  95 }
  96
  97 /**
  98  * Declare write update to sub device
  99  *
 100  * Declare Write updates llog records to the sub device during distribute
 101  * transaction.
 102  *
 103  * \param[in] env       execution environment
 104  * \param[in] record    update records being written
 105  * \param[in] sub_th    sub transaction handle
 106  * \param[in] record_size total update record size
 107  *
 108  * \retval              0 if writing succeeds
 109  * \retval              negative errno if writing fails
 110  */
 111 static int sub_declare_updates_write(const struct lu_env *env,
 112                                      struct llog_update_record *record,
 113                                      struct thandle *sub_th, size_t record_size)
 114 {
 115         struct llog_ctxt        *ctxt;
 116         struct dt_device        *dt = sub_th->th_dev;
 117         int                     left = record_size;
 118         int rc;
 119
 120         /* If ctxt is NULL, it means not need to write update,
 121          * for example if the the OSP is used to connect to OST */
 122         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 123                                 LLOG_UPDATELOG_ORIG_CTXT);
 124
 125         /* Not ready to record updates yet. */
 126         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 127                 llog_ctxt_put(ctxt);
 128                 return 0;
 129         }
 130
 131         rc = llog_declare_add(env, ctxt->loc_handle,
 132                               &record->lur_hdr, sub_th);
 133         if (rc < 0)
 134                 GOTO(out_put, rc);
 135
 136         while (left > ctxt->loc_chunk_size) {
 137                 rc = llog_declare_add(env, ctxt->loc_handle,
 138                                       &record->lur_hdr, sub_th);
 139                 if (rc < 0)
 140                         GOTO(out_put, rc);
 141
 142                 left -= ctxt->loc_chunk_size;
 143         }
 144
 145 out_put:
 146         llog_ctxt_put(ctxt);
 147
 148         return rc;
 149 }
 150
 151 /**
 152  * write update to sub device
 153  *
 154  * Write llog update record to the sub device during distribute
 155  * transaction. If it succeeds, llog cookie of the record will be
 156  * returned by @cookie.
 157  *
 158  * \param[in] env       execution environment
 159  * \param[in] record    update records being written
 160  * \param[in] sub_th    sub transaction handle
 161  * \param[out] cookie   llog cookie of the update record.
 162  *
 163  * \retval              1 if writing succeeds
 164  * \retval              negative errno if writing fails
 165  */
 166 static int sub_updates_write(const struct lu_env *env,
 167                              struct llog_update_record *record,
 168                              struct sub_thandle *sub_th)
 169 {
 170         struct dt_device *dt = sub_th->st_dt;
 171         struct llog_ctxt *ctxt;
 172         struct llog_update_record *lur = NULL;
 173         __u32 update_count = 0;
 174         __u32 param_count = 0;
 175         __u32 last_update_count = 0;
 176         __u32 last_param_count = 0;
 177         char *start;
 178         char *cur;
 179         char *next;
 180         struct sub_thandle_cookie *stc;
 181         size_t reclen;
 182         bool eof = false;
 183         int rc;
 184         ENTRY;
 185
 186         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 187                                 LLOG_UPDATELOG_ORIG_CTXT);
 188         /* If ctxt == NULL, then it means updates on OST (only happens
 189          * during migration), and we do not track those updates for now */
 190         /* If ctxt->loc_handle == NULL, then it does not need to record
 191          * update, usually happens in error handler path */
 192         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 193                 llog_ctxt_put(ctxt);
 194                 RETURN(0);
 195         }
 196
 197         /* Since the cross-MDT updates will includes both local
 198          * and remote updates, the update ops count must > 1 */
 199         LASSERT(record->lur_update_rec.ur_update_count > 1);
 200         LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
 201                  "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
 202                  llog_update_record_size(record));
 203
 204         /*
 205          * If its size > llog chunk_size, then write current chunk to the update
 206          * llog, NB the padding should >= LLOG_MIN_REC_SIZE.
 207          *
 208          * So check padding length is either >= LLOG_MIN_REC_SIZE or is 0
 209          * (record length just matches the chunk size).
 210          */
 211
 212         reclen = record->lur_hdr.lrh_len;
 213         if (reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 214             reclen == ctxt->loc_chunk_size) {
 215                 OBD_ALLOC_PTR(stc);
 216                 if (stc == NULL)
 217                         GOTO(llog_put, rc = -ENOMEM);
 218                 INIT_LIST_HEAD(&stc->stc_list);
 219
 220                 rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
 221                               &stc->stc_cookie, sub_th->st_sub_th);
 222
 223                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 224                        dt->dd_lu_dev.ld_obd->obd_name,
 225                        PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
 226                        stc->stc_cookie.lgc_index, rc);
 227
 228                 if (rc > 0) {
 229                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 230                         rc = 0;
 231                 } else {
 232                         OBD_FREE_PTR(stc);
 233                 }
 234
 235                 GOTO(llog_put, rc);
 236         }
 237
 238         /* Split the records into chunk_size update record */
 239         OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
 240         if (lur == NULL)
 241                 GOTO(llog_put, rc = -ENOMEM);
 242
 243         memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
 244         lur->lur_update_rec.ur_update_count = 0;
 245         lur->lur_update_rec.ur_param_count = 0;
 246         start = (char *)&record->lur_update_rec.ur_ops;
 247         cur = next = start;
 248         do {
 249                 if (update_count < record->lur_update_rec.ur_update_count)
 250                         next = (char *)update_op_next_op(
 251                                                 (struct update_op *)cur);
 252                 else if (param_count < record->lur_update_rec.ur_param_count)
 253                         next = (char *)update_param_next_param(
 254                                                 (struct update_param *)cur);
 255                 else
 256                         eof = true;
 257
 258                 reclen = __llog_update_record_size(
 259                                 __update_records_size(next - start));
 260                 if ((reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 261                      reclen == ctxt->loc_chunk_size) &&
 262                     !eof) {
 263                         cur = next;
 264
 265                         if (update_count <
 266                             record->lur_update_rec.ur_update_count)
 267                                 update_count++;
 268                         else if (param_count <
 269                                  record->lur_update_rec.ur_param_count)
 270                                 param_count++;
 271                         continue;
 272                 }
 273
 274                 lur->lur_update_rec.ur_update_count = update_count -
 275                                                       last_update_count;
 276                 lur->lur_update_rec.ur_param_count = param_count -
 277                                                      last_param_count;
 278                 memcpy(&lur->lur_update_rec.ur_ops, start, cur - start);
 279                 lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 280
 281                 LASSERT(lur->lur_hdr.lrh_len ==
 282                          __llog_update_record_size(
 283                                 __update_records_size(cur - start)));
 284                 LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
 285
 286                 update_records_dump(&lur->lur_update_rec, D_INFO, true);
 287
 288                 OBD_ALLOC_PTR(stc);
 289                 if (stc == NULL)
 290                         GOTO(llog_put, rc = -ENOMEM);
 291                 INIT_LIST_HEAD(&stc->stc_list);
 292
 293                 rc = llog_add(env, ctxt->loc_handle, &lur->lur_hdr,
 294                               &stc->stc_cookie, sub_th->st_sub_th);
 295
 296                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 297                         dt->dd_lu_dev.ld_obd->obd_name,
 298                         PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
 299                         stc->stc_cookie.lgc_index, rc);
 300
 301                 if (rc > 0) {
 302                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 303                         rc = 0;
 304                 } else {
 305                         OBD_FREE_PTR(stc);
 306                         GOTO(llog_put, rc);
 307                 }
 308
 309                 last_update_count = update_count;
 310                 last_param_count = param_count;
 311                 start = cur;
 312                 lur->lur_update_rec.ur_update_count = 0;
 313                 lur->lur_update_rec.ur_param_count = 0;
 314                 lur->lur_update_rec.ur_flags |= UPDATE_RECORD_CONTINUE;
 315         } while (!eof);
 316
 317 llog_put:
 318         if (lur != NULL)
 319                 OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
 320         llog_ctxt_put(ctxt);
 321
 322         RETURN(rc);
 323 }
 324
 325 /**
 326  * Prepare the update records.
 327  *
 328  * Merge params and ops into the update records, then initializing
 329  * the update buffer.
 330  *
 331  * During transaction execution phase, parameters and update ops
 332  * are collected in two different buffers (see lod_updates_pack()),
 333  * during transaction stop, it needs to be merged in one buffer,
 334  * so it will be written in the update log.
 335  *
 336  * \param[in] env       execution environment
 337  * \param[in] tmt       top_multiple_thandle for distribute txn
 338  *
 339  * \retval              0 if merging succeeds.
 340  * \retval              negaitive errno if merging fails.
 341  */
 342 static int prepare_writing_updates(const struct lu_env *env,
 343                                    struct top_multiple_thandle *tmt)
 344 {
 345         struct thandle_update_records   *tur = tmt->tmt_update_records;
 346         struct llog_update_record       *lur;
 347         struct update_params *params;
 348         size_t params_size;
 349         size_t update_size;
 350
 351         if (tur == NULL || tur->tur_update_records == NULL ||
 352             tur->tur_update_params == NULL)
 353                 return 0;
 354
 355         lur = tur->tur_update_records;
 356         /* Extends the update records buffer if needed */
 357         params_size = update_params_size(tur->tur_update_params,
 358                                          tur->tur_update_param_count);
 359         LASSERT(lur->lur_update_rec.ur_param_count == 0);
 360         update_size = llog_update_record_size(lur);
 361         if (cfs_size_round(update_size + params_size) >
 362             tur->tur_update_records_buf_size) {
 363                 int rc;
 364
 365                 rc = tur_update_records_extend(tur,
 366                         cfs_size_round(update_size + params_size));
 367                 if (rc < 0)
 368                         return rc;
 369
 370                 lur = tur->tur_update_records;
 371         }
 372
 373         params = update_records_get_params(&lur->lur_update_rec);
 374         memcpy(params, tur->tur_update_params, params_size);
 375
 376         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
 377         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
 378         /* Init update record header */
 379         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 380         lur->lur_hdr.lrh_type = UPDATE_REC;
 381
 382         /* Dump updates for debugging purpose */
 383         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 384
 385         return 0;
 386 }
 387
 388 static inline int
 389 distribute_txn_commit_thread_running(struct lu_target *lut)
 390 {
 391         return lut->lut_tdtd_commit_thread.t_flags & SVC_RUNNING;
 392 }
 393
 394 static inline int
 395 distribute_txn_commit_thread_stopped(struct lu_target *lut)
 396 {
 397         return lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED;
 398 }
 399
 400 /**
 401  * Top thandle commit callback
 402  *
 403  * This callback will be called when all of sub transactions are committed.
 404  *
 405  * \param[in] th        top thandle to be committed.
 406  */
 407 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
 408 {
 409         struct lu_target *lut;
 410         ENTRY;
 411
 412         LASSERT(atomic_read(&tmt->tmt_refcount) > 0);
 413
 414         top_multiple_thandle_dump(tmt, D_HA);
 415         tmt->tmt_committed = 1;
 416         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
 417         if (distribute_txn_commit_thread_running(lut))
 418                 wake_up(&lut->lut_tdtd->tdtd_commit_thread_waitq);
 419         RETURN_EXIT;
 420 }
 421
 422 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
 423                                        struct dt_device *dt_dev)
 424 {
 425         struct sub_thandle *st;
 426
 427         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 428                 if (st->st_dt == dt_dev)
 429                         return st;
 430         }
 431         return NULL;
 432 }
 433 EXPORT_SYMBOL(lookup_sub_thandle);
 434
 435 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
 436                                        struct dt_device *dt_dev)
 437 {
 438         struct sub_thandle *st;
 439
 440         OBD_ALLOC_PTR(st);
 441         if (st == NULL)
 442                 RETURN(ERR_PTR(-ENOMEM));
 443
 444         INIT_LIST_HEAD(&st->st_sub_list);
 445         INIT_LIST_HEAD(&st->st_cookie_list);
 446         st->st_dt = dt_dev;
 447
 448         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
 449         return st;
 450 }
 451
 452 static void sub_trans_commit_cb_internal(struct top_multiple_thandle *tmt,
 453                                          struct thandle *sub_th, int err)
 454 {
 455         struct sub_thandle      *st;
 456         bool                    all_committed = true;
 457
 458         /* Check if all sub thandles are committed */
 459         spin_lock(&tmt->tmt_sub_lock);
 460         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 461                 if (st->st_sub_th == sub_th) {
 462                         st->st_committed = 1;
 463                         st->st_result = err;
 464                 }
 465                 if (!st->st_committed)
 466                         all_committed = false;
 467         }
 468         spin_unlock(&tmt->tmt_sub_lock);
 469
 470         if (tmt->tmt_result == 0)
 471                 tmt->tmt_result = err;
 472
 473         if (all_committed)
 474                 top_trans_committed_cb(tmt);
 475
 476         top_multiple_thandle_dump(tmt, D_INFO);
 477         top_multiple_thandle_put(tmt);
 478         RETURN_EXIT;
 479 }
 480
 481 /**
 482  * sub thandle commit callback
 483  *
 484  * Mark the sub thandle to be committed and if all sub thandle are committed
 485  * notify the top thandle.
 486  *
 487  * \param[in] env       execution environment
 488  * \param[in] sub_th    sub thandle being committed
 489  * \param[in] cb        commit callback
 490  * \param[in] err       trans result
 491  */
 492 static void sub_trans_commit_cb(struct lu_env *env,
 493                                 struct thandle *sub_th,
 494                                 struct dt_txn_commit_cb *cb, int err)
 495 {
 496         struct top_multiple_thandle *tmt = cb->dcb_data;
 497
 498         sub_trans_commit_cb_internal(tmt, sub_th, err);
 499 }
 500
 501 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
 502                                     struct top_multiple_thandle *tmt)
 503 {
 504         LASSERT(st->st_sub_th != NULL);
 505         top_multiple_thandle_get(tmt);
 506         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
 507         st->st_commit_dcb.dcb_data = tmt;
 508         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
 509         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
 510 }
 511
 512 /**
 513  * Sub thandle stop call back
 514  *
 515  * After sub thandle is stopped, it will call this callback to notify
 516  * the top thandle.
 517  *
 518  * \param[in] th        sub thandle to be stopped
 519  * \param[in] rc        result of sub trans
 520  */
 521 static void sub_trans_stop_cb(struct lu_env *env,
 522                               struct thandle *sub_th,
 523                               struct dt_txn_commit_cb *cb, int err)
 524 {
 525         struct sub_thandle              *st;
 526         struct top_multiple_thandle     *tmt = cb->dcb_data;
 527         ENTRY;
 528
 529         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 530                 if (st->st_stopped)
 531                         continue;
 532
 533                 if (st->st_dt == sub_th->th_dev) {
 534                         st->st_stopped = 1;
 535                         st->st_result = err;
 536                         break;
 537                 }
 538         }
 539
 540         wake_up(&tmt->tmt_stop_waitq);
 541         RETURN_EXIT;
 542 }
 543
 544 static void sub_thandle_register_stop_cb(struct sub_thandle *st,
 545                                          struct top_multiple_thandle *tmt)
 546 {
 547         st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
 548         st->st_stop_dcb.dcb_data = tmt;
 549         st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
 550         INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
 551         dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
 552 }
 553
 554 /**
 555  * Create sub thandle
 556  *
 557  * Create transaction handle for sub_thandle
 558  *
 559  * \param[in] env       execution environment
 560  * \param[in] th        top thandle
 561  * \param[in] st        sub_thandle
 562  *
 563  * \retval              0 if creation succeeds.
 564  * \retval              negative errno if creation fails.
 565  */
 566 int sub_thandle_trans_create(const struct lu_env *env,
 567                              struct top_thandle *top_th,
 568                              struct sub_thandle *st)
 569 {
 570         struct thandle *sub_th;
 571
 572         sub_th = dt_trans_create(env, st->st_dt);
 573         if (IS_ERR(sub_th))
 574                 return PTR_ERR(sub_th);
 575
 576         sub_th->th_top = &top_th->tt_super;
 577         st->st_sub_th = sub_th;
 578
 579         sub_th->th_wait_submit = 1;
 580         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
 581         return 0;
 582 }
 583
 584 /**
 585  * Create the top transaction.
 586  *
 587  * Create the top transaction on the master device. It will create a top
 588  * thandle and a sub thandle on the master device.
 589  *
 590  * \param[in] env               execution environment
 591  * \param[in] master_dev        master_dev the top thandle will be created
 592  *
 593  * \retval                      pointer to the created thandle.
 594  * \retval                      ERR_PTR(errno) if creation failed.
 595  */
 596 struct thandle *
 597 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
 598 {
 599         struct top_thandle      *top_th;
 600         struct thandle          *child_th;
 601
 602         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
 603         if (top_th == NULL)
 604                 return ERR_PTR(-ENOMEM);
 605
 606         top_th->tt_super.th_top = &top_th->tt_super;
 607
 608         if (master_dev != NULL) {
 609                 child_th = dt_trans_create(env, master_dev);
 610                 if (IS_ERR(child_th)) {
 611                         OBD_FREE_PTR(top_th);
 612                         return child_th;
 613                 }
 614
 615                 child_th->th_top = &top_th->tt_super;
 616                 child_th->th_wait_submit = 1;
 617                 top_th->tt_master_sub_thandle = child_th;
 618         }
 619         return &top_th->tt_super;
 620 }
 621 EXPORT_SYMBOL(top_trans_create);
 622
 623 /**
 624  * Declare write update transaction
 625  *
 626  * Check if there are updates being recorded in this transaction,
 627  * it will write the record into the disk.
 628  *
 629  * \param[in] env       execution environment
 630  * \param[in] tmt       top multiple transaction handle
 631  *
 632  * \retval              0 if writing succeeds
 633  * \retval              negative errno if writing fails
 634  */
 635 static int declare_updates_write(const struct lu_env *env,
 636                                  struct top_multiple_thandle *tmt)
 637 {
 638         struct llog_update_record *record;
 639         struct sub_thandle *st;
 640         int rc = 0;
 641
 642         record = tmt->tmt_update_records->tur_update_records;
 643         /* Declare update write for all other target */
 644         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 645                 if (st->st_sub_th == NULL)
 646                         continue;
 647
 648                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
 649                                                tmt->tmt_record_size);
 650                 if (rc < 0)
 651                         break;
 652         }
 653
 654         return rc;
 655 }
 656
 657 /**
 658  * Assign batchid to the distribute transaction.
 659  *
 660  * Assign batchid to the distribute transaction
 661  *
 662  * \param[in] tmt       distribute transaction
 663  */
 664 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 665 {
 666         struct target_distribute_txn_data *tdtd;
 667         struct dt_device *dt = new->tmt_master_sub_dt;
 668         struct sub_thandle *st;
 669
 670         LASSERT(dt != NULL);
 671         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 672         spin_lock(&tdtd->tdtd_batchid_lock);
 673         new->tmt_batchid = tdtd->tdtd_batchid++;
 674         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
 675         spin_unlock(&tdtd->tdtd_batchid_lock);
 676         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 677                 if (st->st_sub_th != NULL)
 678                         sub_thandle_register_commit_cb(st, new);
 679         }
 680         top_multiple_thandle_get(new);
 681         top_multiple_thandle_dump(new, D_INFO);
 682 }
 683
 684 /**
 685  * Insert distribute transaction to the distribute txn list.
 686  *
 687  * Insert distribute transaction to the distribute txn list.
 688  *
 689  * \param[in] new       the distribute txn to be inserted.
 690  */
 691 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
 692 {
 693         struct dt_device *dt = new->tmt_master_sub_dt;
 694         struct top_multiple_thandle *tmt;
 695         struct target_distribute_txn_data *tdtd;
 696         struct sub_thandle *st;
 697         bool    at_head = false;
 698
 699         LASSERT(dt != NULL);
 700         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 701
 702         spin_lock(&tdtd->tdtd_batchid_lock);
 703         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
 704                 if (new->tmt_batchid > tmt->tmt_batchid) {
 705                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
 706                         break;
 707                 }
 708         }
 709         if (list_empty(&new->tmt_commit_list)) {
 710                 at_head = true;
 711                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
 712         }
 713         spin_unlock(&tdtd->tdtd_batchid_lock);
 714
 715         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 716                 if (st->st_sub_th != NULL)
 717                         sub_thandle_register_commit_cb(st, new);
 718         }
 719
 720         top_multiple_thandle_get(new);
 721         top_multiple_thandle_dump(new, D_INFO);
 722         if (new->tmt_committed && at_head)
 723                 wake_up(&tdtd->tdtd_commit_thread_waitq);
 724 }
 725
 726 /**
 727  * Prepare cross-MDT operation.
 728  *
 729  * Create the update record buffer to record updates for cross-MDT operation,
 730  * add master sub transaction to tt_sub_trans_list, and declare the update
 731  * writes.
 732  *
 733  * During updates packing, all of parameters will be packed in
 734  * tur_update_params, and updates will be packed in tur_update_records.
 735  * Then in transaction stop, parameters and updates will be merged
 736  * into one updates buffer.
 737  *
 738  * And also master thandle will be added to the sub_th list, so it will be
 739  * easy to track the commit status.
 740  *
 741  * \param[in] env       execution environment
 742  * \param[in] th        top transaction handle
 743  *
 744  * \retval              0 if preparation succeeds.
 745  * \retval              negative errno if preparation fails.
 746  */
 747 static int prepare_multiple_node_trans(const struct lu_env *env,
 748                                        struct top_multiple_thandle *tmt)
 749 {
 750         struct thandle_update_records   *tur;
 751         int                             rc;
 752         ENTRY;
 753
 754         if (tmt->tmt_update_records == NULL) {
 755                 tur = &update_env_info(env)->uti_tur;
 756                 rc = check_and_prepare_update_record(env, tur);
 757                 if (rc < 0)
 758                         RETURN(rc);
 759
 760                 tmt->tmt_update_records = tur;
 761                 distribute_txn_assign_batchid(tmt);
 762         }
 763
 764         rc = declare_updates_write(env, tmt);
 765
 766         RETURN(rc);
 767 }
 768
 769 /**
 770  * start the top transaction.
 771  *
 772  * Start all of its sub transactions, then start master sub transaction.
 773  *
 774  * \param[in] env               execution environment
 775  * \param[in] master_dev        master_dev the top thandle will be start
 776  * \param[in] th                top thandle
 777  *
 778  * \retval                      0 if transaction start succeeds.
 779  * \retval                      negative errno if start fails.
 780  */
 781 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
 782                     struct thandle *th)
 783 {
 784         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 785                                                        tt_super);
 786         struct sub_thandle              *st;
 787         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
 788         int                             rc = 0;
 789         ENTRY;
 790
 791         if (tmt == NULL) {
 792                 if (th->th_sync)
 793                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 794                 if (th->th_local)
 795                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 796                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
 797                                     top_th->tt_master_sub_thandle);
 798                 RETURN(rc);
 799         }
 800
 801         tmt = top_th->tt_multiple_thandle;
 802         rc = prepare_multiple_node_trans(env, tmt);
 803         if (rc < 0)
 804                 RETURN(rc);
 805
 806         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 807                 if (st->st_sub_th == NULL)
 808                         continue;
 809                 if (th->th_sync)
 810                         st->st_sub_th->th_sync = th->th_sync;
 811                 if (th->th_local)
 812                         st->st_sub_th->th_local = th->th_local;
 813                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
 814                                     st->st_sub_th);
 815                 if (rc != 0)
 816                         GOTO(out, rc);
 817
 818                 LASSERT(st->st_started == 0);
 819                 st->st_started = 1;
 820         }
 821 out:
 822         th->th_result = rc;
 823         RETURN(rc);
 824 }
 825 EXPORT_SYMBOL(top_trans_start);
 826
 827 /**
 828  * Check whether we need write updates record
 829  *
 830  * Check if the updates for the top_thandle needs to be writen
 831  * to all targets. Only if the transaction succeeds and the updates
 832  * number > 2, it will write the updates,
 833  *
 834  * \params [in] top_th  top thandle.
 835  *
 836  * \retval              true if it needs to write updates
 837  * \retval              false if it does not need to write updates
 838  **/
 839 static bool top_check_write_updates(struct top_thandle *top_th)
 840 {
 841         struct top_multiple_thandle     *tmt;
 842         struct thandle_update_records   *tur;
 843
 844         /* Do not write updates to records if the transaction fails */
 845         if (top_th->tt_super.th_result != 0)
 846                 return false;
 847
 848         tmt = top_th->tt_multiple_thandle;
 849         if (tmt == NULL)
 850                 return false;
 851
 852         tur = tmt->tmt_update_records;
 853         if (tur == NULL)
 854                 return false;
 855
 856         /* Hmm, false update records, since the cross-MDT operation
 857          * should includes both local and remote updates, so the
 858          * updates count should >= 2 */
 859         if (tur->tur_update_records == NULL ||
 860             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
 861                 return false;
 862
 863         return true;
 864 }
 865
 866 /**
 867  * Check if top transaction is stopped
 868  *
 869  * Check if top transaction is stopped, only if all sub transaction
 870  * is stopped, then the top transaction is stopped.
 871  *
 872  * \param [in] top_th   top thandle
 873  *
 874  * \retval              true if the top transaction is stopped.
 875  * \retval              false if the top transaction is not stopped.
 876  */
 877 static bool top_trans_is_stopped(struct top_thandle *top_th)
 878 {
 879         struct top_multiple_thandle     *tmt;
 880         struct sub_thandle              *st;
 881         bool                    all_stopped = true;
 882
 883         tmt = top_th->tt_multiple_thandle;
 884         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 885                 if (!st->st_stopped && st->st_sub_th != NULL) {
 886                         all_stopped = false;
 887                         break;
 888                 }
 889
 890                 if (st->st_result != 0 &&
 891                     top_th->tt_super.th_result == 0)
 892                         top_th->tt_super.th_result = st->st_result;
 893         }
 894
 895         return all_stopped;
 896 }
 897
 898 /**
 899  * Wait result of top transaction
 900  *
 901  * Wait until all sub transaction get its result.
 902  *
 903  * \param [in] top_th   top thandle.
 904  *
 905  * \retval              the result of top thandle.
 906  */
 907 static int top_trans_wait_result(struct top_thandle *top_th)
 908 {
 909         struct l_wait_info      lwi = {0};
 910
 911         l_wait_event(top_th->tt_multiple_thandle->tmt_stop_waitq,
 912                      top_trans_is_stopped(top_th), &lwi);
 913
 914         RETURN(top_th->tt_super.th_result);
 915 }
 916
 917 /**
 918  * Stop the top transaction.
 919  *
 920  * Stop the transaction on the master device first, then stop transactions
 921  * on other sub devices.
 922  *
 923  * \param[in] env               execution environment
 924  * \param[in] master_dev        master_dev the top thandle will be created
 925  * \param[in] th                top thandle
 926  *
 927  * \retval                      0 if stop transaction succeeds.
 928  * \retval                      negative errno if stop transaction fails.
 929  */
 930 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
 931                    struct thandle *th)
 932 {
 933         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 934                                                        tt_super);
 935         struct sub_thandle              *st;
 936         struct sub_thandle              *master_st;
 937         struct top_multiple_thandle     *tmt;
 938         struct thandle_update_records   *tur;
 939         bool                            write_updates = false;
 940         int                     rc = 0;
 941         ENTRY;
 942
 943         if (likely(top_th->tt_multiple_thandle == NULL)) {
 944                 LASSERT(master_dev != NULL);
 945
 946                 if (th->th_sync)
 947                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 948                 if (th->th_local)
 949                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 950                 rc = dt_trans_stop(env, master_dev,
 951                                    top_th->tt_master_sub_thandle);
 952                 OBD_FREE_PTR(top_th);
 953                 RETURN(rc);
 954         }
 955
 956         tmt = top_th->tt_multiple_thandle;
 957         tur = tmt->tmt_update_records;
 958
 959         /* Note: we need stop the master thandle first, then the stop
 960          * callback will fill the master transno in the update logs,
 961          * then these update logs will be sent to other MDTs */
 962         /* get the master sub thandle */
 963         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
 964         write_updates = top_check_write_updates(top_th);
 965
 966         /* Step 1: write the updates log on Master MDT */
 967         if (master_st != NULL && master_st->st_sub_th != NULL &&
 968             write_updates) {
 969                 struct llog_update_record *lur;
 970
 971                 /* Merge the parameters and updates into one buffer */
 972                 rc = prepare_writing_updates(env, tmt);
 973                 if (rc < 0) {
 974                         CERROR("%s: cannot prepare updates: rc = %d\n",
 975                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 976                         th->th_result = rc;
 977                         write_updates = false;
 978                         GOTO(stop_master_trans, rc);
 979                 }
 980
 981                 lur = tur->tur_update_records;
 982                 /* Write updates to the master MDT */
 983                 rc = sub_updates_write(env, lur, master_st);
 984
 985                 /* Cleanup the common parameters in the update records,
 986                  * master transno callback might add more parameters.
 987                  * and we need merge the update records again in the
 988                  * following */
 989                 if (tur->tur_update_params != NULL)
 990                         lur->lur_update_rec.ur_param_count = 0;
 991
 992                 if (rc < 0) {
 993                         CERROR("%s: write updates failed: rc = %d\n",
 994                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 995                         th->th_result = rc;
 996                         write_updates = false;
 997                         GOTO(stop_master_trans, rc);
 998                 }
 999         }
1000
1001 stop_master_trans:
1002         /* Step 2: Stop the transaction on the master MDT, and fill the
1003          * master transno in the update logs to other MDT. */
1004         if (master_st != NULL && master_st->st_sub_th != NULL) {
1005                 if (th->th_local)
1006                         master_st->st_sub_th->th_local = th->th_local;
1007                 if (th->th_sync)
1008                         master_st->st_sub_th->th_sync = th->th_sync;
1009                 master_st->st_sub_th->th_result = th->th_result;
1010                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
1011                 /* If it does not write_updates, then we call submit callback
1012                  * here, otherwise callback is done through
1013                  * osd(osp)_trans_commit_cb() */
1014                 if (!master_st->st_started &&
1015                     !list_empty(&tmt->tmt_commit_list))
1016                         sub_trans_commit_cb_internal(tmt,
1017                                                 master_st->st_sub_th, rc);
1018                 if (rc < 0) {
1019                         th->th_result = rc;
1020                         GOTO(stop_other_trans, rc);
1021                 } else if (tur != NULL && tur->tur_update_records != NULL) {
1022                         struct llog_update_record *lur;
1023
1024                         lur = tur->tur_update_records;
1025                         if (lur->lur_update_rec.ur_master_transno == 0)
1026                                 /* Update master transno after master stop
1027                                  * callback */
1028                                 lur->lur_update_rec.ur_master_transno =
1029                                                 tgt_th_info(env)->tti_transno;
1030                 }
1031         }
1032
1033         /* Step 3: write updates to other MDTs */
1034         if (write_updates) {
1035                 struct llog_update_record *lur;
1036
1037                 /* Stop callback of master will add more updates and also update
1038                  * master transno, so merge the parameters and updates into one
1039                  * buffer again */
1040                 rc = prepare_writing_updates(env, tmt);
1041                 if (rc < 0) {
1042                         CERROR("%s: prepare updates failed: rc = %d\n",
1043                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1044                         th->th_result = rc;
1045                         GOTO(stop_other_trans, rc);
1046                 }
1047                 lur = tur->tur_update_records;
1048                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1049                                     st_sub_list) {
1050                         if (st->st_sub_th == NULL || st == master_st ||
1051                             st->st_sub_th->th_result < 0)
1052                                 continue;
1053
1054                         rc = sub_updates_write(env, lur, st);
1055                         if (rc < 0) {
1056                                 th->th_result = rc;
1057                                 break;
1058                         }
1059                 }
1060         }
1061
1062 stop_other_trans:
1063         /* Step 4: Stop the transaction on other MDTs */
1064         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1065                 if (st == master_st || st->st_sub_th == NULL)
1066                         continue;
1067
1068                 if (th->th_sync)
1069                         st->st_sub_th->th_sync = th->th_sync;
1070                 if (th->th_local)
1071                         st->st_sub_th->th_local = th->th_local;
1072                 st->st_sub_th->th_result = th->th_result;
1073                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1074                                    st->st_sub_th);
1075                 if (unlikely(rc < 0 && th->th_result == 0))
1076                         th->th_result = rc;
1077         }
1078
1079         rc = top_trans_wait_result(top_th);
1080
1081         tmt->tmt_result = rc;
1082
1083         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1084          * multiple node transaction, the top transaction will be destroyed. */
1085         top_multiple_thandle_put(tmt);
1086         OBD_FREE_PTR(top_th);
1087         RETURN(rc);
1088 }
1089 EXPORT_SYMBOL(top_trans_stop);
1090
1091 /**
1092  * Create top_multiple_thandle for top_thandle
1093  *
1094  * Create top_mutilple_thandle to manage the mutiple node transaction
1095  * for top_thandle, and it also needs to add master sub thandle to the
1096  * sub trans list now.
1097  *
1098  * \param[in] env       execution environment
1099  * \param[in] top_th    the top thandle
1100  *
1101  * \retval      0 if creation succeeds
1102  * \retval      negative errno if creation fails
1103  */
1104 int top_trans_create_tmt(const struct lu_env *env,
1105                          struct top_thandle *top_th)
1106 {
1107         struct top_multiple_thandle *tmt;
1108
1109         OBD_ALLOC_PTR(tmt);
1110         if (tmt == NULL)
1111                 return -ENOMEM;
1112
1113         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1114         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1115         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1116         atomic_set(&tmt->tmt_refcount, 1);
1117         spin_lock_init(&tmt->tmt_sub_lock);
1118         init_waitqueue_head(&tmt->tmt_stop_waitq);
1119
1120         top_th->tt_multiple_thandle = tmt;
1121
1122         return 0;
1123 }
1124
1125 static struct sub_thandle *
1126 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1127                                 struct thandle *sub_th)
1128 {
1129         struct sub_thandle *st;
1130
1131         /* create and init sub th to the top trans list */
1132         st = create_sub_thandle(top_th->tt_multiple_thandle,
1133                                 sub_th->th_dev);
1134         if (IS_ERR(st))
1135                 return st;
1136
1137         st->st_sub_th = sub_th;
1138
1139         sub_th->th_top = &top_th->tt_super;
1140         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
1141         return st;
1142 }
1143
1144 /**
1145  * Get sub thandle.
1146  *
1147  * Get sub thandle from the top thandle according to the sub dt_device.
1148  *
1149  * \param[in] env       execution environment
1150  * \param[in] th        thandle on the top layer.
1151  * \param[in] sub_dt    sub dt_device used to get sub transaction
1152  *
1153  * \retval              thandle of sub transaction if succeed
1154  * \retval              PTR_ERR(errno) if failed
1155  */
1156 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1157                                       struct thandle *th,
1158                                       struct dt_device *sub_dt)
1159 {
1160         struct sub_thandle      *st = NULL;
1161         struct sub_thandle      *master_st = NULL;
1162         struct top_thandle      *top_th;
1163         struct thandle          *sub_th = NULL;
1164         int                     rc = 0;
1165         ENTRY;
1166
1167         top_th = container_of(th, struct top_thandle, tt_super);
1168
1169         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1170                 RETURN(top_th->tt_master_sub_thandle);
1171
1172         if (top_th->tt_multiple_thandle != NULL) {
1173                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1174                 if (st != NULL)
1175                         RETURN(st->st_sub_th);
1176         }
1177
1178         sub_th = dt_trans_create(env, sub_dt);
1179         if (IS_ERR(sub_th))
1180                 RETURN(sub_th);
1181
1182         /* Create top_multiple_thandle if necessary */
1183         if (top_th->tt_multiple_thandle == NULL) {
1184                 struct top_multiple_thandle *tmt;
1185
1186                 rc = top_trans_create_tmt(env, top_th);
1187                 if (rc < 0)
1188                         GOTO(stop_trans, rc);
1189
1190                 tmt = top_th->tt_multiple_thandle;
1191
1192                 /* Add master sub th to the top trans list */
1193                 tmt->tmt_master_sub_dt =
1194                         top_th->tt_master_sub_thandle->th_dev;
1195                 master_st = create_sub_thandle_with_thandle(top_th,
1196                                         top_th->tt_master_sub_thandle);
1197                 if (IS_ERR(master_st)) {
1198                         rc = PTR_ERR(master_st);
1199                         master_st = NULL;
1200                         GOTO(stop_trans, rc);
1201                 }
1202         }
1203
1204         /* create and init sub th to the top trans list */
1205         st = create_sub_thandle_with_thandle(top_th, sub_th);
1206         if (IS_ERR(st)) {
1207                 rc = PTR_ERR(st);
1208                 st = NULL;
1209                 GOTO(stop_trans, rc);
1210         }
1211         st->st_sub_th->th_wait_submit = 1;
1212 stop_trans:
1213         if (rc < 0) {
1214                 if (master_st != NULL) {
1215                         list_del(&master_st->st_sub_list);
1216                         OBD_FREE_PTR(master_st);
1217                 }
1218                 sub_th->th_result = rc;
1219                 dt_trans_stop(env, sub_dt, sub_th);
1220                 sub_th = ERR_PTR(rc);
1221         }
1222
1223         RETURN(sub_th);
1224 }
1225 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1226
1227 /**
1228  * Top multiple thandle destroy
1229  *
1230  * Destroy multiple thandle and all its sub thandle.
1231  *
1232  * \param[in] tmt       top_multiple_thandle to be destroyed.
1233  */
1234 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1235 {
1236         struct sub_thandle *st;
1237         struct sub_thandle *tmp;
1238
1239         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1240         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1241                                  st_sub_list) {
1242                 struct sub_thandle_cookie *stc;
1243                 struct sub_thandle_cookie *tmp;
1244
1245                 list_del(&st->st_sub_list);
1246                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1247                                          stc_list) {
1248                         list_del(&stc->stc_list);
1249                         OBD_FREE_PTR(stc);
1250                 }
1251                 OBD_FREE_PTR(st);
1252         }
1253         OBD_FREE_PTR(tmt);
1254 }
1255 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1256
1257 /**
1258  * Cancel the update log on MDTs
1259  *
1260  * Cancel the update log on MDTs then destroy the thandle.
1261  *
1262  * \param[in] env       execution environment
1263  * \param[in] tmt       the top multiple thandle whose updates records
1264  *                      will be cancelled.
1265  *
1266  * \retval              0 if cancellation succeeds.
1267  * \retval              negative errno if cancellation fails.
1268  */
1269 static int distribute_txn_cancel_records(const struct lu_env *env,
1270                                          struct top_multiple_thandle *tmt)
1271 {
1272         struct sub_thandle *st;
1273         ENTRY;
1274
1275         top_multiple_thandle_dump(tmt, D_INFO);
1276         /* Cancel update logs on other MDTs */
1277         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1278                 struct llog_ctxt        *ctxt;
1279                 struct obd_device       *obd;
1280                 struct llog_cookie      *cookie;
1281                 struct sub_thandle_cookie *stc;
1282                 int rc;
1283
1284                 obd = st->st_dt->dd_lu_dev.ld_obd;
1285                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1286                 if (ctxt == NULL)
1287                         continue;
1288                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1289                         cookie = &stc->stc_cookie;
1290                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1291                                 continue;
1292
1293                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1294                                                      cookie);
1295                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1296                                DFID".%u: rc = %d\n", obd->obd_name,
1297                                tmt->tmt_batchid,
1298                                PFID(&cookie->lgc_lgl.lgl_oi.oi_fid),
1299                                cookie->lgc_index, rc);
1300                 }
1301
1302                 llog_ctxt_put(ctxt);
1303         }
1304
1305         RETURN(0);
1306 }
1307
1308 /**
1309  * Check if there are committed transaction
1310  *
1311  * Check if there are committed transaction in the distribute transaction
1312  * list, then cancel the update records for those committed transaction.
1313  * Because the distribute transaction in the list are sorted by batchid,
1314  * and cancellation will be done by batchid order, so we only check the first
1315  * the transaction(with lowest batchid) in the list.
1316  *
1317  * \param[in] lod       lod device where cancel thread is
1318  *
1319  * \retval              true if it is ready
1320  * \retval              false if it is not ready
1321  */
1322 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1323 {
1324         struct top_multiple_thandle     *tmt = NULL;
1325         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1326         bool    ready = false;
1327
1328         spin_lock(&tdtd->tdtd_batchid_lock);
1329         if (!list_empty(&tdtd->tdtd_list)) {
1330                 tmt = list_entry(tdtd->tdtd_list.next,
1331                                  struct top_multiple_thandle, tmt_commit_list);
1332                 if (tmt->tmt_committed &&
1333                     (!obd->obd_recovering || (obd->obd_recovering &&
1334                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1335                         ready = true;
1336         }
1337         spin_unlock(&tdtd->tdtd_batchid_lock);
1338
1339         return ready;
1340 }
1341
1342 struct distribute_txn_bid_data {
1343         struct dt_txn_commit_cb  dtbd_cb;
1344         struct target_distribute_txn_data      *dtbd_tdtd;
1345         __u64                    dtbd_batchid;
1346 };
1347
1348 /**
1349  * callback of updating commit batchid
1350  *
1351  * Updating commit batchid then wake up the commit thread to cancel the
1352  * records.
1353  *
1354  * \param[in]env        execution environment
1355  * \param[in]th         thandle to updating commit batchid
1356  * \param[in]cb         commit callback
1357  * \param[in]err        result of thandle
1358  */
1359 static void distribute_txn_batchid_cb(struct lu_env *env,
1360                                       struct thandle *th,
1361                                       struct dt_txn_commit_cb *cb,
1362                                       int err)
1363 {
1364         struct distribute_txn_bid_data          *dtbd = NULL;
1365         struct target_distribute_txn_data       *tdtd;
1366
1367         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1368         tdtd = dtbd->dtbd_tdtd;
1369
1370         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1371               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1372         spin_lock(&tdtd->tdtd_batchid_lock);
1373         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1374             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1375                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1376         spin_unlock(&tdtd->tdtd_batchid_lock);
1377         atomic_dec(&tdtd->tdtd_refcount);
1378         wake_up(&tdtd->tdtd_commit_thread_waitq);
1379
1380         OBD_FREE_PTR(dtbd);
1381 }
1382
1383 /**
1384  * Update the commit batchid in disk
1385  *
1386  * Update commit batchid in the disk, after this is committed, it can start
1387  * to cancel the update records.
1388  *
1389  * \param[in] env       execution environment
1390  * \param[in] tdtd      distribute transaction structure
1391  * \param[in] batchid   commit batchid to be updated
1392  *
1393  * \retval              0 if update succeeds.
1394  * \retval              negative errno if update fails.
1395  */
1396 static int
1397 distribute_txn_commit_batchid_update(const struct lu_env *env,
1398                               struct target_distribute_txn_data *tdtd,
1399                               __u64 batchid)
1400 {
1401         struct distribute_txn_bid_data  *dtbd = NULL;
1402         struct thandle          *th;
1403         struct lu_buf            buf;
1404         __u64                    tmp;
1405         __u64                    off;
1406         int                      rc;
1407         ENTRY;
1408
1409         OBD_ALLOC_PTR(dtbd);
1410         if (dtbd == NULL)
1411                 RETURN(-ENOMEM);
1412         dtbd->dtbd_batchid = batchid;
1413         dtbd->dtbd_tdtd = tdtd;
1414         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1415         atomic_inc(&tdtd->tdtd_refcount);
1416
1417         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1418         if (IS_ERR(th)) {
1419                 atomic_dec(&tdtd->tdtd_refcount);
1420                 OBD_FREE_PTR(dtbd);
1421                 RETURN(PTR_ERR(th));
1422         }
1423
1424         tmp = cpu_to_le64(batchid);
1425         buf.lb_buf = &tmp;
1426         buf.lb_len = sizeof(tmp);
1427         off = 0;
1428
1429         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1430                                      th);
1431         if (rc < 0)
1432                 GOTO(stop, rc);
1433
1434         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1435         if (rc < 0)
1436                 GOTO(stop, rc);
1437
1438         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1439         if (rc < 0)
1440                 GOTO(stop, rc);
1441
1442         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1443                              &off, th);
1444
1445         CDEBUG(D_INFO, "%s: update batchid %llu: rc = %d\n",
1446                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1447
1448 stop:
1449         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1450         if (rc < 0) {
1451                 atomic_dec(&tdtd->tdtd_refcount);
1452                 OBD_FREE_PTR(dtbd);
1453         }
1454         RETURN(rc);
1455 }
1456
1457 /**
1458  * Init commit batchid for distribute transaction.
1459  *
1460  * Initialize the batchid object and get commit batchid from the object.
1461  *
1462  * \param[in] env       execution environment
1463  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1464  *
1465  * \retval              0 if initialization succeeds.
1466  * \retval              negative errno if initialization fails.
1467  **/
1468 static int
1469 distribute_txn_commit_batchid_init(const struct lu_env *env,
1470                                    struct target_distribute_txn_data *tdtd)
1471 {
1472         struct tgt_thread_info  *tti = tgt_th_info(env);
1473         struct lu_target        *lut = tdtd->tdtd_lut;
1474         struct lu_attr          *attr = &tti->tti_attr;
1475         struct lu_fid           *fid = &tti->tti_fid1;
1476         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1477         struct dt_object        *dt_obj = NULL;
1478         struct lu_buf           buf;
1479         __u64                   tmp;
1480         __u64                   off;
1481         int                     rc;
1482         ENTRY;
1483
1484         memset(attr, 0, sizeof(*attr));
1485         attr->la_valid = LA_MODE;
1486         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1487         dof->dof_type = dt_mode_to_dft(S_IFREG);
1488
1489         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1490
1491         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1492                                    attr);
1493         if (IS_ERR(dt_obj)) {
1494                 rc = PTR_ERR(dt_obj);
1495                 dt_obj = NULL;
1496                 GOTO(out_put, rc);
1497         }
1498
1499         tdtd->tdtd_batchid_obj = dt_obj;
1500
1501         buf.lb_buf = &tmp;
1502         buf.lb_len = sizeof(tmp);
1503         off = 0;
1504         rc = dt_read(env, dt_obj, &buf, &off);
1505         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1506                 CERROR("%s can't read last committed batchid: rc = %d\n",
1507                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1508                 if (rc > 0)
1509                         rc = -EINVAL;
1510                 GOTO(out_put, rc);
1511         } else if (rc == buf.lb_len) {
1512                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1513                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1514                        tdtd->tdtd_lut->lut_obd->obd_name,
1515                        tdtd->tdtd_committed_batchid);
1516                 rc = 0;
1517         }
1518
1519 out_put:
1520         if (rc < 0 && dt_obj != NULL) {
1521                 dt_object_put(env, dt_obj);
1522                 tdtd->tdtd_batchid_obj = NULL;
1523         }
1524         return rc;
1525 }
1526
1527 /**
1528  * manage the distribute transaction thread
1529  *
1530  * Distribute transaction are linked to the list, and once the distribute
1531  * transaction is committed, it will update the last committed batchid first,
1532  * after it is committed, it will cancel the records.
1533  *
1534  * \param[in] _arg      argument for commit thread
1535  *
1536  * \retval              0 if thread is running successfully
1537  * \retval              negative errno if the thread can not be run.
1538  */
1539 static int distribute_txn_commit_thread(void *_arg)
1540 {
1541         struct target_distribute_txn_data *tdtd = _arg;
1542         struct lu_target        *lut = tdtd->tdtd_lut;
1543         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1544         struct l_wait_info       lwi = { 0 };
1545         struct lu_env            env;
1546         struct list_head         list;
1547         int                      rc;
1548         struct top_multiple_thandle *tmt;
1549         struct top_multiple_thandle *tmp;
1550         __u64                    batchid = 0, committed;
1551
1552         ENTRY;
1553
1554         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1555         if (rc != 0)
1556                 RETURN(rc);
1557
1558         spin_lock(&tdtd->tdtd_batchid_lock);
1559         thread->t_flags = SVC_RUNNING;
1560         spin_unlock(&tdtd->tdtd_batchid_lock);
1561         wake_up(&thread->t_ctl_waitq);
1562         INIT_LIST_HEAD(&list);
1563
1564         CDEBUG(D_HA, "%s: start commit thread committed batchid %llu\n",
1565                tdtd->tdtd_lut->lut_obd->obd_name,
1566                tdtd->tdtd_committed_batchid);
1567
1568         while (distribute_txn_commit_thread_running(lut)) {
1569                 spin_lock(&tdtd->tdtd_batchid_lock);
1570                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1571                                          tmt_commit_list) {
1572                         if (tmt->tmt_committed == 0)
1573                                 break;
1574
1575                         /* Note: right now, replay is based on master MDT
1576                          * transno, but cancellation is based on batchid.
1577                          * so we do not try to cancel the update log until
1578                          * the recoverying is done, unless the update records
1579                          * batchid < committed_batchid. */
1580                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1581                                 list_move_tail(&tmt->tmt_commit_list, &list);
1582                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1583                                 LASSERTF(tmt->tmt_batchid >= batchid,
1584                                          "tmt %p tmt_batchid: %llu, batchid "
1585                                           "%llu\n", tmt, tmt->tmt_batchid,
1586                                          batchid);
1587                                 /* There are three types of distribution
1588                                  * transaction result
1589                                  *
1590                                  * 1. If tmt_result < 0, it means the
1591                                  * distribution transaction fails, which should
1592                                  * be rare, because once declare phase succeeds,
1593                                  * the operation should succeeds anyway. Note in
1594                                  * this case, we will still update batchid so
1595                                  * cancellation would be stopped.
1596                                  *
1597                                  * 2. If tmt_result == 0, it means the
1598                                  * distribution transaction succeeds, and we
1599                                  * will update batchid.
1600                                  *
1601                                  * 3. If tmt_result > 0, it means distribute
1602                                  * transaction is not yet committed on every
1603                                  * node, but we need release this tmt before
1604                                  * that, which usuually happens during umount.
1605                                  */
1606                                 if (tmt->tmt_result <= 0)
1607                                         batchid = tmt->tmt_batchid;
1608                                 list_move_tail(&tmt->tmt_commit_list, &list);
1609                         }
1610                 }
1611                 spin_unlock(&tdtd->tdtd_batchid_lock);
1612
1613                 CDEBUG(D_HA, "%s: batchid: %llu committed batchid "
1614                        "%llu\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1615                        tdtd->tdtd_committed_batchid);
1616                 /* update globally committed on a storage */
1617                 if (batchid > tdtd->tdtd_committed_batchid) {
1618                         rc = distribute_txn_commit_batchid_update(&env, tdtd,
1619                                                              batchid);
1620                         if (rc == 0)
1621                                 batchid = 0;
1622                 }
1623                 /* cancel the records for committed batchid's */
1624                 /* XXX: should we postpone cancel's till the end of recovery? */
1625                 committed = tdtd->tdtd_committed_batchid;
1626                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1627                         if (tmt->tmt_batchid > committed)
1628                                 break;
1629                         list_del_init(&tmt->tmt_commit_list);
1630                         if (tmt->tmt_result <= 0)
1631                                 distribute_txn_cancel_records(&env, tmt);
1632                         top_multiple_thandle_put(tmt);
1633                 }
1634
1635                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1636                              !distribute_txn_commit_thread_running(lut) ||
1637                              committed < tdtd->tdtd_committed_batchid ||
1638                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1639         };
1640
1641         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1642                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1643
1644         spin_lock(&tdtd->tdtd_batchid_lock);
1645         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1646                                  tmt_commit_list)
1647                 list_move_tail(&tmt->tmt_commit_list, &list);
1648         spin_unlock(&tdtd->tdtd_batchid_lock);
1649
1650         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1651                tdtd->tdtd_lut->lut_obd->obd_name);
1652         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1653                 list_del_init(&tmt->tmt_commit_list);
1654                 top_multiple_thandle_dump(tmt, D_HA);
1655                 top_multiple_thandle_put(tmt);
1656         }
1657
1658         thread->t_flags = SVC_STOPPED;
1659         lu_env_fini(&env);
1660         wake_up(&thread->t_ctl_waitq);
1661
1662         RETURN(0);
1663 }
1664
1665 /**
1666  * Start llog cancel thread
1667  *
1668  * Start llog cancel(master/slave) thread on LOD
1669  *
1670  * \param[in]lclt       cancel log thread to be started.
1671  *
1672  * \retval              0 if the thread is started successfully.
1673  * \retval              negative errno if the thread is not being
1674  *                      started.
1675  */
1676 int distribute_txn_init(const struct lu_env *env,
1677                         struct lu_target *lut,
1678                         struct target_distribute_txn_data *tdtd,
1679                         __u32 index)
1680 {
1681         struct task_struct      *task;
1682         struct l_wait_info       lwi = { 0 };
1683         int                     rc;
1684         ENTRY;
1685
1686         INIT_LIST_HEAD(&tdtd->tdtd_list);
1687         INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
1688         INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
1689         spin_lock_init(&tdtd->tdtd_batchid_lock);
1690         spin_lock_init(&tdtd->tdtd_replay_list_lock);
1691         tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
1692         tdtd->tdtd_replay_ready = 0;
1693
1694         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1695
1696         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1697         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1698         init_waitqueue_head(&tdtd->tdtd_recovery_threads_waitq);
1699         atomic_set(&tdtd->tdtd_refcount, 0);
1700         atomic_set(&tdtd->tdtd_recovery_threads_count, 0);
1701
1702         tdtd->tdtd_lut = lut;
1703         if (lut->lut_bottom->dd_rdonly)
1704                 RETURN(0);
1705
1706         rc = distribute_txn_commit_batchid_init(env, tdtd);
1707         if (rc != 0)
1708                 RETURN(rc);
1709
1710         task = kthread_run(distribute_txn_commit_thread, tdtd, "dist_txn-%u",
1711                            index);
1712         if (IS_ERR(task))
1713                 RETURN(PTR_ERR(task));
1714
1715         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1716                      distribute_txn_commit_thread_running(lut) ||
1717                      distribute_txn_commit_thread_stopped(lut), &lwi);
1718         RETURN(0);
1719 }
1720 EXPORT_SYMBOL(distribute_txn_init);
1721
1722 /**
1723  * Stop llog cancel thread
1724  *
1725  * Stop llog cancel(master/slave) thread on LOD and also destory
1726  * all of transaction in the list.
1727  *
1728  * \param[in]lclt       cancel log thread to be stopped.
1729  */
1730 void distribute_txn_fini(const struct lu_env *env,
1731                          struct target_distribute_txn_data *tdtd)
1732 {
1733         struct lu_target *lut = tdtd->tdtd_lut;
1734
1735         /* Stop cancel thread */
1736         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1737                 return;
1738
1739         spin_lock(&tdtd->tdtd_batchid_lock);
1740         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1741         spin_unlock(&tdtd->tdtd_batchid_lock);
1742         wake_up(&tdtd->tdtd_commit_thread_waitq);
1743         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1744                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1745
1746         dtrq_list_destroy(tdtd);
1747         if (tdtd->tdtd_batchid_obj != NULL) {
1748                 dt_object_put(env, tdtd->tdtd_batchid_obj);
1749                 tdtd->tdtd_batchid_obj = NULL;
1750         }
1751 }
1752 EXPORT_SYMBOL(distribute_txn_fini);