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) 2014, 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"
  76                "batchid "LPU64"\n",
  77                tmt->tmt_master_sub_dt ?
  78                tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
  79                "NULL",
  80                tmt, atomic_read(&tmt->tmt_refcount), tmt->tmt_committed,
  81                tmt->tmt_result, tmt->tmt_batchid);
  82
  83         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
  84                 CDEBUG(mask, "st %p obd %s committed %d sub_th %p "
  85                        " cookie "DOSTID": %u\n",
  86                        st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
  87                        st->st_committed, st->st_sub_th,
  88                        POSTID(&st->st_cookie.lgc_lgl.lgl_oi),
  89                        st->st_cookie.lgc_index);
  90         }
  91 }
  92
  93 /**
  94  * Declare write update to sub device
  95  *
  96  * Declare Write updates llog records to the sub device during distribute
  97  * transaction.
  98  *
  99  * \param[in] env       execution environment
 100  * \param[in] record    update records being written
 101  * \param[in] sub_th    sub transaction handle
 102  *
 103  * \retval              0 if writing succeeds
 104  * \retval              negative errno if writing fails
 105  */
 106 static int sub_declare_updates_write(const struct lu_env *env,
 107                                      struct llog_update_record *record,
 108                                      struct thandle *sub_th)
 109 {
 110         struct llog_ctxt        *ctxt;
 111         struct dt_device        *dt = sub_th->th_dev;
 112         int rc;
 113
 114         /* If ctxt is NULL, it means not need to write update,
 115          * for example if the the OSP is used to connect to OST */
 116         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 117                                 LLOG_UPDATELOG_ORIG_CTXT);
 118         LASSERT(ctxt != NULL);
 119
 120         /* Not ready to record updates yet. */
 121         if (ctxt->loc_handle == NULL) {
 122                 llog_ctxt_put(ctxt);
 123                 return 0;
 124         }
 125
 126         rc = llog_declare_add(env, ctxt->loc_handle, &record->lur_hdr,
 127                               sub_th);
 128
 129         llog_ctxt_put(ctxt);
 130
 131         return rc;
 132 }
 133
 134 /**
 135  * write update to sub device
 136  *
 137  * Write llog update record to the sub device during distribute
 138  * transaction. If it succeeds, llog cookie of the record will be
 139  * returned by @cookie.
 140  *
 141  * \param[in] env       execution environment
 142  * \param[in] record    update records being written
 143  * \param[in] sub_th    sub transaction handle
 144  * \param[out] cookie   llog cookie of the update record.
 145  *
 146  * \retval              1 if writing succeeds
 147  * \retval              negative errno if writing fails
 148  */
 149 static int sub_updates_write(const struct lu_env *env,
 150                              struct llog_update_record *record,
 151                              struct thandle *sub_th,
 152                              struct llog_cookie *cookie)
 153 {
 154         struct dt_device        *dt = sub_th->th_dev;
 155         struct llog_ctxt        *ctxt;
 156         int                     rc;
 157         ENTRY;
 158
 159         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 160                                 LLOG_UPDATELOG_ORIG_CTXT);
 161         LASSERT(ctxt != NULL);
 162
 163         /* Not ready to record updates yet, usually happens
 164          * in error handler path */
 165         if (ctxt->loc_handle == NULL) {
 166                 llog_ctxt_put(ctxt);
 167                 RETURN(0);
 168         }
 169
 170         /* Since the cross-MDT updates will includes both local
 171          * and remote updates, the update ops count must > 1 */
 172         LASSERT(record->lur_update_rec.ur_update_count > 1);
 173         LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
 174                  "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
 175                  llog_update_record_size(record));
 176
 177         rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
 178                       cookie, sub_th);
 179         llog_ctxt_put(ctxt);
 180
 181         CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u.\n",
 182                dt->dd_lu_dev.ld_obd->obd_name,
 183                POSTID(&cookie->lgc_lgl.lgl_oi), cookie->lgc_index);
 184
 185         if (rc > 0)
 186                 rc = 0;
 187
 188         RETURN(rc);
 189 }
 190
 191 /**
 192  * Prepare the update records.
 193  *
 194  * Merge params and ops into the update records, then initializing
 195  * the update buffer.
 196  *
 197  * During transaction execution phase, parameters and update ops
 198  * are collected in two different buffers (see lod_updates_pack()),
 199  * during transaction stop, it needs to be merged in one buffer,
 200  * so it will be written in the update log.
 201  *
 202  * \param[in] env       execution environment
 203  * \param[in] tmt       top_multiple_thandle for distribute txn
 204  *
 205  * \retval              0 if merging succeeds.
 206  * \retval              negaitive errno if merging fails.
 207  */
 208 static int prepare_writing_updates(const struct lu_env *env,
 209                                    struct top_multiple_thandle *tmt)
 210 {
 211         struct thandle_update_records   *tur = tmt->tmt_update_records;
 212         struct llog_update_record       *lur;
 213         struct update_params *params;
 214         size_t params_size;
 215         size_t update_size;
 216
 217         if (tur == NULL || tur->tur_update_records == NULL ||
 218             tur->tur_update_params == NULL)
 219                 return 0;
 220
 221         lur = tur->tur_update_records;
 222         /* Extends the update records buffer if needed */
 223         params_size = update_params_size(tur->tur_update_params,
 224                                          tur->tur_update_param_count);
 225         LASSERT(lur->lur_update_rec.ur_param_count == 0);
 226         update_size = llog_update_record_size(lur);
 227         if (cfs_size_round(update_size + params_size) >
 228             tur->tur_update_records_buf_size) {
 229                 int rc;
 230
 231                 rc = tur_update_records_extend(tur,
 232                         cfs_size_round(update_size + params_size));
 233                 if (rc < 0)
 234                         return rc;
 235
 236                 lur = tur->tur_update_records;
 237         }
 238
 239         params = update_records_get_params(&lur->lur_update_rec);
 240         memcpy(params, tur->tur_update_params, params_size);
 241
 242         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
 243         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
 244         /* Init update record header */
 245         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 246         lur->lur_hdr.lrh_type = UPDATE_REC;
 247
 248         /* Dump updates for debugging purpose */
 249         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 250
 251         return 0;
 252 }
 253
 254 static inline int
 255 distribute_txn_commit_thread_running(struct lu_target *lut)
 256 {
 257         return lut->lut_tdtd_commit_thread.t_flags & SVC_RUNNING;
 258 }
 259
 260 static inline int
 261 distribute_txn_commit_thread_stopped(struct lu_target *lut)
 262 {
 263         return lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED;
 264 }
 265
 266 /**
 267  * Top thandle commit callback
 268  *
 269  * This callback will be called when all of sub transactions are committed.
 270  *
 271  * \param[in] th        top thandle to be committed.
 272  */
 273 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
 274 {
 275         struct lu_target *lut;
 276         ENTRY;
 277
 278         LASSERT(atomic_read(&tmt->tmt_refcount) > 0);
 279
 280         top_multiple_thandle_dump(tmt, D_HA);
 281         tmt->tmt_committed = 1;
 282         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
 283         if (distribute_txn_commit_thread_running(lut))
 284                 wake_up(&lut->lut_tdtd->tdtd_commit_thread_waitq);
 285         RETURN_EXIT;
 286 }
 287
 288 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
 289                                        struct dt_device *dt_dev)
 290 {
 291         struct sub_thandle *st;
 292
 293         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 294                 if (st->st_dt == dt_dev)
 295                         return st;
 296         }
 297         return NULL;
 298 }
 299 EXPORT_SYMBOL(lookup_sub_thandle);
 300
 301 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
 302                                        struct dt_device *dt_dev)
 303 {
 304         struct sub_thandle *st;
 305
 306         OBD_ALLOC_PTR(st);
 307         if (st == NULL)
 308                 RETURN(ERR_PTR(-ENOMEM));
 309
 310         INIT_LIST_HEAD(&st->st_sub_list);
 311         st->st_dt = dt_dev;
 312
 313         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
 314         return st;
 315 }
 316
 317 /**
 318  * sub thandle commit callback
 319  *
 320  * Mark the sub thandle to be committed and if all sub thandle are committed
 321  * notify the top thandle.
 322  *
 323  * \param[in] sub_th    sub thandle being committed.
 324  */
 325 static void sub_trans_commit_cb(struct lu_env *env,
 326                                 struct thandle *sub_th,
 327                                 struct dt_txn_commit_cb *cb, int err)
 328 {
 329         struct sub_thandle      *st;
 330         struct top_multiple_thandle *tmt = cb->dcb_data;
 331         bool                    all_committed = true;
 332         ENTRY;
 333
 334         /* Check if all sub thandles are committed */
 335         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 336                 if (st->st_sub_th == sub_th) {
 337                         st->st_committed = 1;
 338                         st->st_result = err;
 339                 }
 340                 if (!st->st_committed)
 341                         all_committed = false;
 342         }
 343
 344         if (tmt->tmt_result == 0)
 345                 tmt->tmt_result = err;
 346
 347         if (all_committed)
 348                 top_trans_committed_cb(tmt);
 349
 350         top_multiple_thandle_dump(tmt, D_INFO);
 351         top_multiple_thandle_put(tmt);
 352         RETURN_EXIT;
 353 }
 354
 355 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
 356                                     struct top_multiple_thandle *tmt)
 357 {
 358         LASSERT(st->st_sub_th != NULL);
 359         top_multiple_thandle_get(tmt);
 360         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
 361         st->st_commit_dcb.dcb_data = tmt;
 362         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
 363         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
 364 }
 365
 366 /**
 367  * Create sub thandle
 368  *
 369  * Create transaction handle for sub_thandle
 370  *
 371  * \param[in] env       execution environment
 372  * \param[in] th        top thandle
 373  * \param[in] st        sub_thandle
 374  *
 375  * \retval              0 if creation succeeds.
 376  * \retval              negative errno if creation fails.
 377  */
 378 int sub_thandle_trans_create(const struct lu_env *env,
 379                              struct top_thandle *top_th,
 380                              struct sub_thandle *st)
 381 {
 382         struct thandle *sub_th;
 383
 384         sub_th = dt_trans_create(env, st->st_dt);
 385         if (IS_ERR(sub_th))
 386                 return PTR_ERR(sub_th);
 387
 388         sub_th->th_top = &top_th->tt_super;
 389         st->st_sub_th = sub_th;
 390
 391         sub_th->th_wait_submit = 1;
 392         return 0;
 393 }
 394
 395
 396 /**
 397  * Create the top transaction.
 398  *
 399  * Create the top transaction on the master device. It will create a top
 400  * thandle and a sub thandle on the master device.
 401  *
 402  * \param[in] env               execution environment
 403  * \param[in] master_dev        master_dev the top thandle will be created
 404  *
 405  * \retval                      pointer to the created thandle.
 406  * \retval                      ERR_PTR(errno) if creation failed.
 407  */
 408 struct thandle *
 409 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
 410 {
 411         struct top_thandle      *top_th;
 412         struct thandle          *child_th;
 413
 414         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
 415         if (top_th == NULL)
 416                 return ERR_PTR(-ENOMEM);
 417
 418         top_th->tt_super.th_top = &top_th->tt_super;
 419
 420         if (master_dev != NULL) {
 421                 child_th = dt_trans_create(env, master_dev);
 422                 if (IS_ERR(child_th)) {
 423                         OBD_FREE_PTR(top_th);
 424                         return child_th;
 425                 }
 426
 427                 child_th->th_top = &top_th->tt_super;
 428                 child_th->th_wait_submit = 1;
 429                 top_th->tt_master_sub_thandle = child_th;
 430
 431                 top_th->tt_super.th_tags |= child_th->th_tags;
 432         }
 433         return &top_th->tt_super;
 434 }
 435 EXPORT_SYMBOL(top_trans_create);
 436
 437 /**
 438  * Declare write update transaction
 439  *
 440  * Check if there are updates being recorded in this transaction,
 441  * it will write the record into the disk.
 442  *
 443  * \param[in] env       execution environment
 444  * \param[in] tmt       top multiple transaction handle
 445  *
 446  * \retval              0 if writing succeeds
 447  * \retval              negative errno if writing fails
 448  */
 449 static int declare_updates_write(const struct lu_env *env,
 450                                  struct top_multiple_thandle *tmt)
 451 {
 452         struct llog_update_record *record;
 453         struct sub_thandle *st;
 454         int rc = 0;
 455
 456         record = tmt->tmt_update_records->tur_update_records;
 457         /* Declare update write for all other target */
 458         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 459                 if (st->st_sub_th == NULL)
 460                         continue;
 461
 462                 rc = sub_declare_updates_write(env, record, st->st_sub_th);
 463                 if (rc < 0)
 464                         break;
 465         }
 466
 467         return rc;
 468 }
 469
 470 /**
 471  * Assign batchid to the distribute transaction.
 472  *
 473  * Assign batchid to the distribute transaction
 474  *
 475  * \param[in] tmt       distribute transaction
 476  */
 477 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 478 {
 479         struct target_distribute_txn_data *tdtd;
 480         struct dt_device *dt = new->tmt_master_sub_dt;
 481
 482         LASSERT(dt != NULL);
 483         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 484         spin_lock(&tdtd->tdtd_batchid_lock);
 485         new->tmt_batchid = tdtd->tdtd_batchid++;
 486         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
 487         spin_unlock(&tdtd->tdtd_batchid_lock);
 488         top_multiple_thandle_get(new);
 489         top_multiple_thandle_dump(new, D_INFO);
 490 }
 491
 492 /**
 493  * Insert distribute transaction to the distribute txn list.
 494  *
 495  * Insert distribute transaction to the distribute txn list.
 496  *
 497  * \param[in] new       the distribute txn to be inserted.
 498  */
 499 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
 500 {
 501         struct dt_device *dt = new->tmt_master_sub_dt;
 502         struct top_multiple_thandle *tmt;
 503         struct target_distribute_txn_data *tdtd;
 504         bool    at_head = false;
 505
 506         LASSERT(dt != NULL);
 507         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 508
 509         spin_lock(&tdtd->tdtd_batchid_lock);
 510         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
 511                 if (new->tmt_batchid > tmt->tmt_batchid) {
 512                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
 513                         break;
 514                 }
 515         }
 516         if (list_empty(&new->tmt_commit_list)) {
 517                 at_head = true;
 518                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
 519         }
 520         spin_unlock(&tdtd->tdtd_batchid_lock);
 521         top_multiple_thandle_get(new);
 522         top_multiple_thandle_dump(new, D_INFO);
 523         if (new->tmt_committed && at_head)
 524                 wake_up(&tdtd->tdtd_commit_thread_waitq);
 525 }
 526
 527 /**
 528  * Prepare cross-MDT operation.
 529  *
 530  * Create the update record buffer to record updates for cross-MDT operation,
 531  * add master sub transaction to tt_sub_trans_list, and declare the update
 532  * writes.
 533  *
 534  * During updates packing, all of parameters will be packed in
 535  * tur_update_params, and updates will be packed in tur_update_records.
 536  * Then in transaction stop, parameters and updates will be merged
 537  * into one updates buffer.
 538  *
 539  * And also master thandle will be added to the sub_th list, so it will be
 540  * easy to track the commit status.
 541  *
 542  * \param[in] env       execution environment
 543  * \param[in] th        top transaction handle
 544  *
 545  * \retval              0 if preparation succeeds.
 546  * \retval              negative errno if preparation fails.
 547  */
 548 static int prepare_multiple_node_trans(const struct lu_env *env,
 549                                        struct top_multiple_thandle *tmt)
 550 {
 551         struct thandle_update_records   *tur;
 552         int                             rc;
 553         ENTRY;
 554
 555         if (tmt->tmt_update_records == NULL) {
 556                 tur = &update_env_info(env)->uti_tur;
 557                 rc = check_and_prepare_update_record(env, tur);
 558                 if (rc < 0)
 559                         RETURN(rc);
 560
 561                 tmt->tmt_update_records = tur;
 562                 distribute_txn_assign_batchid(tmt);
 563         }
 564
 565         rc = declare_updates_write(env, tmt);
 566
 567         RETURN(rc);
 568 }
 569
 570 /**
 571  * start the top transaction.
 572  *
 573  * Start all of its sub transactions, then start master sub transaction.
 574  *
 575  * \param[in] env               execution environment
 576  * \param[in] master_dev        master_dev the top thandle will be start
 577  * \param[in] th                top thandle
 578  *
 579  * \retval                      0 if transaction start succeeds.
 580  * \retval                      negative errno if start fails.
 581  */
 582 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
 583                     struct thandle *th)
 584 {
 585         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 586                                                        tt_super);
 587         struct sub_thandle              *st;
 588         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
 589         int                             rc = 0;
 590         ENTRY;
 591
 592         if (tmt == NULL) {
 593                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
 594                                     top_th->tt_master_sub_thandle);
 595                 RETURN(rc);
 596         }
 597
 598         tmt = top_th->tt_multiple_thandle;
 599         rc = prepare_multiple_node_trans(env, tmt);
 600         if (rc < 0)
 601                 RETURN(rc);
 602
 603         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 604                 if (st->st_sub_th == NULL)
 605                         continue;
 606                 st->st_sub_th->th_sync = th->th_sync;
 607                 st->st_sub_th->th_local = th->th_local;
 608                 st->st_sub_th->th_tags = th->th_tags;
 609                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
 610                                     st->st_sub_th);
 611                 if (rc != 0)
 612                         GOTO(out, rc);
 613
 614                 sub_thandle_register_commit_cb(st, tmt);
 615         }
 616 out:
 617         th->th_result = rc;
 618         RETURN(rc);
 619 }
 620 EXPORT_SYMBOL(top_trans_start);
 621
 622 /**
 623  * Check whether we need write updates record
 624  *
 625  * Check if the updates for the top_thandle needs to be writen
 626  * to all targets. Only if the transaction succeeds and the updates
 627  * number > 2, it will write the updates,
 628  *
 629  * \params [in] top_th  top thandle.
 630  *
 631  * \retval              true if it needs to write updates
 632  * \retval              false if it does not need to write updates
 633  **/
 634 static bool top_check_write_updates(struct top_thandle *top_th)
 635 {
 636         struct top_multiple_thandle     *tmt;
 637         struct thandle_update_records   *tur;
 638
 639         /* Do not write updates to records if the transaction fails */
 640         if (top_th->tt_super.th_result != 0)
 641                 return false;
 642
 643         tmt = top_th->tt_multiple_thandle;
 644         if (tmt == NULL)
 645                 return false;
 646
 647         tur = tmt->tmt_update_records;
 648         if (tur == NULL)
 649                 return false;
 650
 651         /* Hmm, false update records, since the cross-MDT operation
 652          * should includes both local and remote updates, so the
 653          * updates count should >= 2 */
 654         if (tur->tur_update_records == NULL ||
 655             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
 656                 return false;
 657
 658         return true;
 659 }
 660
 661 /**
 662  * Stop the top transaction.
 663  *
 664  * Stop the transaction on the master device first, then stop transactions
 665  * on other sub devices.
 666  *
 667  * \param[in] env               execution environment
 668  * \param[in] master_dev        master_dev the top thandle will be created
 669  * \param[in] th                top thandle
 670  *
 671  * \retval                      0 if stop transaction succeeds.
 672  * \retval                      negative errno if stop transaction fails.
 673  */
 674 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
 675                    struct thandle *th)
 676 {
 677         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 678                                                        tt_super);
 679         struct sub_thandle              *st;
 680         struct sub_thandle              *master_st;
 681         struct top_multiple_thandle     *tmt;
 682         struct thandle_update_records   *tur;
 683         bool                            write_updates = false;
 684         int                     rc = 0;
 685         ENTRY;
 686
 687         if (likely(top_th->tt_multiple_thandle == NULL)) {
 688                 LASSERT(master_dev != NULL);
 689                 rc = dt_trans_stop(env, master_dev,
 690                                    top_th->tt_master_sub_thandle);
 691                 OBD_FREE_PTR(top_th);
 692                 RETURN(rc);
 693         }
 694
 695         tmt = top_th->tt_multiple_thandle;
 696         tur = tmt->tmt_update_records;
 697
 698         /* Note: we need stop the master thandle first, then the stop
 699          * callback will fill the master transno in the update logs,
 700          * then these update logs will be sent to other MDTs */
 701         /* get the master sub thandle */
 702         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
 703         write_updates = top_check_write_updates(top_th);
 704
 705         /* Step 1: write the updates log on Master MDT */
 706         if (master_st != NULL && master_st->st_sub_th != NULL &&
 707             write_updates) {
 708                 struct llog_update_record *lur;
 709
 710                 /* Merge the parameters and updates into one buffer */
 711                 rc = prepare_writing_updates(env, tmt);
 712                 if (rc < 0) {
 713                         CERROR("%s: cannot prepare updates: rc = %d\n",
 714                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 715                         th->th_result = rc;
 716                         GOTO(stop_master_trans, rc);
 717                 }
 718
 719                 lur = tur->tur_update_records;
 720                 /* Write updates to the master MDT */
 721                 rc = sub_updates_write(env, lur, master_st->st_sub_th,
 722                                        &master_st->st_cookie);
 723
 724                 /* Cleanup the common parameters in the update records,
 725                  * master transno callback might add more parameters.
 726                  * and we need merge the update records again in the
 727                  * following */
 728                 if (tur->tur_update_params != NULL)
 729                         lur->lur_update_rec.ur_param_count = 0;
 730
 731                 if (rc < 0) {
 732                         CERROR("%s: write updates failed: rc = %d\n",
 733                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 734                         th->th_result = rc;
 735                         GOTO(stop_master_trans, rc);
 736                 }
 737         }
 738
 739 stop_master_trans:
 740         /* Step 2: Stop the transaction on the master MDT, and fill the
 741          * master transno in the update logs to other MDT. */
 742         if (master_st != NULL && master_st->st_sub_th != NULL) {
 743                 master_st->st_sub_th->th_local = th->th_local;
 744                 master_st->st_sub_th->th_sync = th->th_sync;
 745                 master_st->st_sub_th->th_tags = th->th_tags;
 746                 master_st->st_sub_th->th_result = th->th_result;
 747                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
 748                 if (rc < 0) {
 749                         th->th_result = rc;
 750                         GOTO(stop_other_trans, rc);
 751                 } else if (tur != NULL && tur->tur_update_records != NULL) {
 752                         struct llog_update_record *lur;
 753
 754                         lur = tur->tur_update_records;
 755                         if (lur->lur_update_rec.ur_master_transno == 0)
 756                                 /* Update master transno after master stop
 757                                  * callback */
 758                                 lur->lur_update_rec.ur_master_transno =
 759                                                 tgt_th_info(env)->tti_transno;
 760                 }
 761         }
 762
 763         /* Step 3: write updates to other MDTs */
 764         if (write_updates) {
 765                 struct llog_update_record *lur;
 766
 767                 /* Stop callback of master will add more updates and also update
 768                  * master transno, so merge the parameters and updates into one
 769                  * buffer again */
 770                 rc = prepare_writing_updates(env, tmt);
 771                 if (rc < 0) {
 772                         CERROR("%s: prepare updates failed: rc = %d\n",
 773                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 774                         th->th_result = rc;
 775                         GOTO(stop_other_trans, rc);
 776                 }
 777                 lur = tur->tur_update_records;
 778                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
 779                                     st_sub_list) {
 780                         if (st->st_sub_th == NULL || st == master_st ||
 781                             st->st_sub_th->th_result < 0)
 782                                 continue;
 783
 784                         rc = sub_updates_write(env, lur, st->st_sub_th,
 785                                                &st->st_cookie);
 786                         if (rc < 0) {
 787                                 th->th_result = rc;
 788                                 break;
 789                         }
 790                 }
 791         }
 792
 793 stop_other_trans:
 794         /* Step 4: Stop the transaction on other MDTs */
 795         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 796                 if (st == master_st || st->st_sub_th == NULL)
 797                         continue;
 798
 799                 st->st_sub_th->th_sync = th->th_sync;
 800                 st->st_sub_th->th_local = th->th_local;
 801                 st->st_sub_th->th_tags = th->th_tags;
 802                 st->st_sub_th->th_result = th->th_result;
 803                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
 804                                    st->st_sub_th);
 805                 if (unlikely(rc < 0 && th->th_result == 0))
 806                         th->th_result = rc;
 807         }
 808
 809         tmt->tmt_result = rc;
 810         /* Balance for the refcount in top_trans_create, Note: if it is NOT
 811          * multiple node transaction, the top transaction will be destroyed. */
 812         top_multiple_thandle_put(tmt);
 813         OBD_FREE_PTR(top_th);
 814         RETURN(rc);
 815 }
 816 EXPORT_SYMBOL(top_trans_stop);
 817
 818 /**
 819  * Create top_multiple_thandle for top_thandle
 820  *
 821  * Create top_mutilple_thandle to manage the mutiple node transaction
 822  * for top_thandle, and it also needs to add master sub thandle to the
 823  * sub trans list now.
 824  *
 825  * \param[in] env       execution environment
 826  * \param[in] top_th    the top thandle
 827  *
 828  * \retval      0 if creation succeeds
 829  * \retval      negative errno if creation fails
 830  */
 831 int top_trans_create_tmt(const struct lu_env *env,
 832                          struct top_thandle *top_th)
 833 {
 834         struct top_multiple_thandle *tmt;
 835
 836         OBD_ALLOC_PTR(tmt);
 837         if (tmt == NULL)
 838                 return -ENOMEM;
 839
 840         tmt->tmt_magic = TOP_THANDLE_MAGIC;
 841         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
 842         INIT_LIST_HEAD(&tmt->tmt_commit_list);
 843         atomic_set(&tmt->tmt_refcount, 1);
 844
 845         top_th->tt_multiple_thandle = tmt;
 846
 847         return 0;
 848 }
 849
 850 static struct sub_thandle *
 851 create_sub_thandle_with_thandle(struct top_thandle *top_th,
 852                                 struct thandle *sub_th)
 853 {
 854         struct sub_thandle *st;
 855
 856         /* create and init sub th to the top trans list */
 857         st = create_sub_thandle(top_th->tt_multiple_thandle,
 858                                 sub_th->th_dev);
 859         if (IS_ERR(st))
 860                 return st;
 861
 862         st->st_sub_th = sub_th;
 863         sub_th->th_top = &top_th->tt_super;
 864         return st;
 865 }
 866
 867 /**
 868  * Get sub thandle.
 869  *
 870  * Get sub thandle from the top thandle according to the sub dt_device.
 871  *
 872  * \param[in] env       execution environment
 873  * \param[in] th        thandle on the top layer.
 874  * \param[in] sub_dt    sub dt_device used to get sub transaction
 875  *
 876  * \retval              thandle of sub transaction if succeed
 877  * \retval              PTR_ERR(errno) if failed
 878  */
 879 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
 880                                       struct thandle *th,
 881                                       struct dt_device *sub_dt)
 882 {
 883         struct sub_thandle      *st = NULL;
 884         struct top_thandle      *top_th;
 885         struct thandle          *sub_th = NULL;
 886         int                     rc = 0;
 887         ENTRY;
 888
 889         top_th = container_of(th, struct top_thandle, tt_super);
 890
 891         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
 892                 RETURN(top_th->tt_master_sub_thandle);
 893
 894         if (top_th->tt_multiple_thandle != NULL) {
 895                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
 896                 if (st != NULL)
 897                         RETURN(st->st_sub_th);
 898         }
 899
 900         sub_th = dt_trans_create(env, sub_dt);
 901         if (IS_ERR(sub_th))
 902                 RETURN(sub_th);
 903
 904         /* Create top_multiple_thandle if necessary */
 905         if (top_th->tt_multiple_thandle == NULL) {
 906                 struct top_multiple_thandle *tmt;
 907
 908                 rc = top_trans_create_tmt(env, top_th);
 909                 if (rc < 0)
 910                         GOTO(stop_trans, rc);
 911
 912                 tmt = top_th->tt_multiple_thandle;
 913
 914                 /* Add master sub th to the top trans list */
 915                 tmt->tmt_master_sub_dt =
 916                         top_th->tt_master_sub_thandle->th_dev;
 917                 st = create_sub_thandle_with_thandle(top_th,
 918                                 top_th->tt_master_sub_thandle);
 919                 if (IS_ERR(st))
 920                         GOTO(stop_trans, rc = PTR_ERR(st));
 921                 top_th->tt_master_sub_thandle->th_sync = 1;
 922                 top_th->tt_super.th_sync = 1;
 923         }
 924
 925         /* create and init sub th to the top trans list */
 926         st = create_sub_thandle_with_thandle(top_th, sub_th);
 927         st->st_sub_th->th_wait_submit = 1;
 928 stop_trans:
 929         if (rc < 0) {
 930                 if (st != NULL)
 931                         OBD_FREE_PTR(st);
 932                 sub_th->th_result = rc;
 933                 dt_trans_stop(env, sub_dt, sub_th);
 934                 sub_th = ERR_PTR(rc);
 935         }
 936
 937         RETURN(sub_th);
 938 }
 939 EXPORT_SYMBOL(thandle_get_sub_by_dt);
 940
 941 /**
 942  * Top multiple thandle destroy
 943  *
 944  * Destroy multiple thandle and all its sub thandle.
 945  *
 946  * \param[in] tmt       top_multiple_thandle to be destroyed.
 947  */
 948 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
 949 {
 950         struct sub_thandle *st;
 951         struct sub_thandle *tmp;
 952
 953         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
 954         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
 955                                  st_sub_list) {
 956                 list_del(&st->st_sub_list);
 957                 OBD_FREE_PTR(st);
 958         }
 959         OBD_FREE_PTR(tmt);
 960 }
 961 EXPORT_SYMBOL(top_multiple_thandle_destroy);
 962
 963 /**
 964  * Cancel the update log on MDTs
 965  *
 966  * Cancel the update log on MDTs then destroy the thandle.
 967  *
 968  * \param[in] env       execution environment
 969  * \param[in] tmt       the top multiple thandle whose updates records
 970  *                      will be cancelled.
 971  *
 972  * \retval              0 if cancellation succeeds.
 973  * \retval              negative errno if cancellation fails.
 974  */
 975 static int distribute_txn_cancel_records(const struct lu_env *env,
 976                                          struct top_multiple_thandle *tmt)
 977 {
 978         struct sub_thandle *st;
 979         ENTRY;
 980
 981         top_multiple_thandle_dump(tmt, D_INFO);
 982         /* Cancel update logs on other MDTs */
 983         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 984                 struct llog_ctxt        *ctxt;
 985                 struct obd_device       *obd;
 986                 struct llog_cookie      *cookie;
 987                 int rc;
 988
 989                 cookie = &st->st_cookie;
 990                 if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
 991                         continue;
 992
 993                 obd = st->st_dt->dd_lu_dev.ld_obd;
 994                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
 995                 LASSERT(ctxt);
 996
 997                 rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
 998                                              cookie);
 999
1000                 llog_ctxt_put(ctxt);
1001                 CDEBUG(D_HA, "%s: batchid %llu cancel update log "DOSTID
1002                        ".%u : rc = %d\n", obd->obd_name, tmt->tmt_batchid,
1003                        POSTID(&cookie->lgc_lgl.lgl_oi), cookie->lgc_index, rc);
1004         }
1005
1006         RETURN(0);
1007 }
1008
1009 /**
1010  * Check if there are committed transaction
1011  *
1012  * Check if there are committed transaction in the distribute transaction
1013  * list, then cancel the update records for those committed transaction.
1014  * Because the distribute transaction in the list are sorted by batchid,
1015  * and cancellation will be done by batchid order, so we only check the first
1016  * the transaction(with lowest batchid) in the list.
1017  *
1018  * \param[in] lod       lod device where cancel thread is
1019  *
1020  * \retval              true if it is ready
1021  * \retval              false if it is not ready
1022  */
1023 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1024 {
1025         struct top_multiple_thandle     *tmt = NULL;
1026         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1027         bool    ready = false;
1028
1029         spin_lock(&tdtd->tdtd_batchid_lock);
1030         if (!list_empty(&tdtd->tdtd_list)) {
1031                 tmt = list_entry(tdtd->tdtd_list.next,
1032                                  struct top_multiple_thandle, tmt_commit_list);
1033                 if (tmt->tmt_committed &&
1034                     (!obd->obd_recovering || (obd->obd_recovering &&
1035                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1036                         ready = true;
1037         }
1038         spin_unlock(&tdtd->tdtd_batchid_lock);
1039
1040         return ready;
1041 }
1042
1043 struct distribute_txn_bid_data {
1044         struct dt_txn_commit_cb  dtbd_cb;
1045         struct target_distribute_txn_data      *dtbd_tdtd;
1046         __u64                    dtbd_batchid;
1047 };
1048
1049 /**
1050  * callback of updating commit batchid
1051  *
1052  * Updating commit batchid then wake up the commit thread to cancel the
1053  * records.
1054  *
1055  * \param[in]env        execution environment
1056  * \param[in]th         thandle to updating commit batchid
1057  * \param[in]cb         commit callback
1058  * \param[in]err        result of thandle
1059  */
1060 static void distribute_txn_batchid_cb(struct lu_env *env,
1061                                       struct thandle *th,
1062                                       struct dt_txn_commit_cb *cb,
1063                                       int err)
1064 {
1065         struct distribute_txn_bid_data          *dtbd = NULL;
1066         struct target_distribute_txn_data       *tdtd;
1067
1068         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1069         tdtd = dtbd->dtbd_tdtd;
1070
1071         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1072               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1073         spin_lock(&tdtd->tdtd_batchid_lock);
1074         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1075             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1076                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1077         spin_unlock(&tdtd->tdtd_batchid_lock);
1078         atomic_dec(&tdtd->tdtd_refcount);
1079         wake_up(&tdtd->tdtd_commit_thread_waitq);
1080
1081         OBD_FREE_PTR(dtbd);
1082 }
1083
1084 /**
1085  * Update the commit batchid in disk
1086  *
1087  * Update commit batchid in the disk, after this is committed, it can start
1088  * to cancel the update records.
1089  *
1090  * \param[in] env       execution environment
1091  * \param[in] tdtd      distribute transaction structure
1092  * \param[in] batchid   commit batchid to be updated
1093  *
1094  * \retval              0 if update succeeds.
1095  * \retval              negative errno if update fails.
1096  */
1097 static int
1098 distribute_txn_commit_batchid_update(const struct lu_env *env,
1099                               struct target_distribute_txn_data *tdtd,
1100                               __u64 batchid)
1101 {
1102         struct distribute_txn_bid_data  *dtbd = NULL;
1103         struct thandle          *th;
1104         struct lu_buf            buf;
1105         __u64                    tmp;
1106         __u64                    off;
1107         int                      rc;
1108         ENTRY;
1109
1110         OBD_ALLOC_PTR(dtbd);
1111         if (dtbd == NULL)
1112                 RETURN(-ENOMEM);
1113         dtbd->dtbd_batchid = batchid;
1114         dtbd->dtbd_tdtd = tdtd;
1115         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1116         atomic_inc(&tdtd->tdtd_refcount);
1117
1118         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1119         if (IS_ERR(th)) {
1120                 OBD_FREE_PTR(dtbd);
1121                 RETURN(PTR_ERR(th));
1122         }
1123
1124         tmp = cpu_to_le64(batchid);
1125         buf.lb_buf = &tmp;
1126         buf.lb_len = sizeof(tmp);
1127         off = 0;
1128
1129         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1130                                      th);
1131         if (rc < 0)
1132                 GOTO(stop, rc);
1133
1134         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1135         if (rc < 0)
1136                 GOTO(stop, rc);
1137
1138         dt_trans_cb_add(th, &dtbd->dtbd_cb);
1139         if (rc < 0)
1140                 GOTO(stop, rc);
1141
1142         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1143                              &off, th);
1144
1145         CDEBUG(D_INFO, "%s: update batchid "LPU64": rc = %d\n",
1146                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1147
1148 stop:
1149         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1150         if (rc < 0)
1151                 OBD_FREE_PTR(dtbd);
1152         RETURN(rc);
1153 }
1154
1155 /**
1156  * Init commit batchid for distribute transaction.
1157  *
1158  * Initialize the batchid object and get commit batchid from the object.
1159  *
1160  * \param[in] env       execution environment
1161  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1162  *
1163  * \retval              0 if initialization succeeds.
1164  * \retval              negative errno if initialization fails.
1165  **/
1166 static int
1167 distribute_txn_commit_batchid_init(const struct lu_env *env,
1168                                    struct target_distribute_txn_data *tdtd)
1169 {
1170         struct tgt_thread_info  *tti = tgt_th_info(env);
1171         struct lu_target        *lut = tdtd->tdtd_lut;
1172         struct lu_attr          *attr = &tti->tti_attr;
1173         struct lu_fid           *fid = &tti->tti_fid1;
1174         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1175         struct dt_object        *dt_obj = NULL;
1176         struct lu_buf           buf;
1177         __u64                   tmp;
1178         __u64                   off;
1179         int                     rc;
1180         ENTRY;
1181
1182         memset(attr, 0, sizeof(*attr));
1183         attr->la_valid = LA_MODE;
1184         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1185         dof->dof_type = dt_mode_to_dft(S_IFREG);
1186
1187         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1188
1189         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1190                                    attr);
1191         if (IS_ERR(dt_obj))
1192                 GOTO(out_put, rc = PTR_ERR(dt_obj));
1193
1194         tdtd->tdtd_batchid_obj = dt_obj;
1195
1196         buf.lb_buf = &tmp;
1197         buf.lb_len = sizeof(tmp);
1198         off = 0;
1199         rc = dt_read(env, dt_obj, &buf, &off);
1200         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1201                 CERROR("%s can't read last committed batchid: rc = %d\n",
1202                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1203                 if (rc > 0)
1204                         rc = -EINVAL;
1205                 GOTO(out_put, rc);
1206         } else if (rc == buf.lb_len) {
1207                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1208                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1209                        tdtd->tdtd_lut->lut_obd->obd_name,
1210                        tdtd->tdtd_committed_batchid);
1211                 rc = 0;
1212         }
1213
1214 out_put:
1215         if (rc < 0 && dt_obj != NULL) {
1216                 lu_object_put(env, &dt_obj->do_lu);
1217                 tdtd->tdtd_batchid_obj = NULL;
1218         }
1219         return rc;
1220 }
1221
1222 /**
1223  * manage the distribute transaction thread
1224  *
1225  * Distribute transaction are linked to the list, and once the distribute
1226  * transaction is committed, it will update the last committed batchid first,
1227  * after it is committed, it will cancel the records.
1228  *
1229  * \param[in] _arg      argument for commit thread
1230  *
1231  * \retval              0 if thread is running successfully
1232  * \retval              negative errno if the thread can not be run.
1233  */
1234 static int distribute_txn_commit_thread(void *_arg)
1235 {
1236         struct target_distribute_txn_data *tdtd = _arg;
1237         struct lu_target        *lut = tdtd->tdtd_lut;
1238         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1239         struct l_wait_info       lwi = { 0 };
1240         struct lu_env            env;
1241         struct list_head         list;
1242         int                      rc;
1243         struct top_multiple_thandle *tmt;
1244         struct top_multiple_thandle *tmp;
1245         __u64                    batchid = 0, committed;
1246
1247         ENTRY;
1248
1249         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1250         if (rc != 0)
1251                 RETURN(rc);
1252
1253         spin_lock(&tdtd->tdtd_batchid_lock);
1254         thread->t_flags = SVC_RUNNING;
1255         spin_unlock(&tdtd->tdtd_batchid_lock);
1256         wake_up(&thread->t_ctl_waitq);
1257         INIT_LIST_HEAD(&list);
1258
1259         CDEBUG(D_HA, "%s: start commit thread committed batchid "LPU64"\n",
1260                tdtd->tdtd_lut->lut_obd->obd_name,
1261                tdtd->tdtd_committed_batchid);
1262
1263         while (distribute_txn_commit_thread_running(lut)) {
1264                 spin_lock(&tdtd->tdtd_batchid_lock);
1265                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1266                                          tmt_commit_list) {
1267                         if (tmt->tmt_committed == 0)
1268                                 break;
1269
1270                         /* Note: right now, replay is based on master MDT
1271                          * transno, but cancellation is based on batchid.
1272                          * so we do not try to cancel the update log until
1273                          * the recoverying is done, unless the update records
1274                          * batchid < committed_batchid. */
1275                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1276                                 list_move_tail(&tmt->tmt_commit_list, &list);
1277                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1278                                 LASSERTF(tmt->tmt_batchid >= batchid,
1279                                          "tmt %p tmt_batchid: "LPU64", batchid "
1280                                           LPU64"\n", tmt, tmt->tmt_batchid,
1281                                          batchid);
1282                                 /* There are three types of distribution
1283                                  * transaction result
1284                                  *
1285                                  * 1. If tmt_result < 0, it means the
1286                                  * distribution transaction fails, which should
1287                                  * be rare, because once declare phase succeeds,
1288                                  * the operation should succeeds anyway. Note in
1289                                  * this case, we will still update batchid so
1290                                  * cancellation would be stopped.
1291                                  *
1292                                  * 2. If tmt_result == 0, it means the
1293                                  * distribution transaction succeeds, and we
1294                                  * will update batchid.
1295                                  *
1296                                  * 3. If tmt_result > 0, it means distribute
1297                                  * transaction is not yet committed on every
1298                                  * node, but we need release this tmt before
1299                                  * that, which usuually happens during umount.
1300                                  */
1301                                 if (tmt->tmt_result <= 0)
1302                                         batchid = tmt->tmt_batchid;
1303                                 list_move_tail(&tmt->tmt_commit_list, &list);
1304                         }
1305                 }
1306                 spin_unlock(&tdtd->tdtd_batchid_lock);
1307
1308                 CDEBUG(D_HA, "%s: batchid: "LPU64" committed batchid "
1309                        LPU64"\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1310                        tdtd->tdtd_committed_batchid);
1311                 /* update globally committed on a storage */
1312                 if (batchid > tdtd->tdtd_committed_batchid) {
1313                         distribute_txn_commit_batchid_update(&env, tdtd,
1314                                                              batchid);
1315                         spin_lock(&tdtd->tdtd_batchid_lock);
1316                         if (batchid > tdtd->tdtd_batchid) {
1317                                 /* This might happen during recovery,
1318                                  * batchid is initialized as last transno,
1319                                  * and the batchid in the update records
1320                                  * on other MDTs might be bigger than
1321                                  * the batchid, so we need update it to
1322                                  * avoid duplicate batchid. */
1323                                 CDEBUG(D_HA, "%s update batchid from "LPU64
1324                                        " to "LPU64"\n",
1325                                        tdtd->tdtd_lut->lut_obd->obd_name,
1326                                        tdtd->tdtd_batchid, batchid);
1327                                 tdtd->tdtd_batchid = batchid;
1328                         }
1329                         spin_unlock(&tdtd->tdtd_batchid_lock);
1330                 }
1331                 /* cancel the records for committed batchid's */
1332                 /* XXX: should we postpone cancel's till the end of recovery? */
1333                 committed = tdtd->tdtd_committed_batchid;
1334                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1335                         if (tmt->tmt_batchid > committed)
1336                                 break;
1337                         list_del_init(&tmt->tmt_commit_list);
1338                         if (tmt->tmt_result <= 0)
1339                                 distribute_txn_cancel_records(&env, tmt);
1340                         top_multiple_thandle_put(tmt);
1341                 }
1342
1343                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1344                              !distribute_txn_commit_thread_running(lut) ||
1345                              committed < tdtd->tdtd_committed_batchid ||
1346                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1347         };
1348
1349         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1350                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1351
1352         spin_lock(&tdtd->tdtd_batchid_lock);
1353         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1354                                  tmt_commit_list)
1355                 list_move_tail(&tmt->tmt_commit_list, &list);
1356         spin_unlock(&tdtd->tdtd_batchid_lock);
1357
1358         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1359                tdtd->tdtd_lut->lut_obd->obd_name);
1360         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1361                 list_del_init(&tmt->tmt_commit_list);
1362                 top_multiple_thandle_dump(tmt, D_HA);
1363                 top_multiple_thandle_put(tmt);
1364         }
1365
1366         thread->t_flags = SVC_STOPPED;
1367         lu_env_fini(&env);
1368         wake_up(&thread->t_ctl_waitq);
1369
1370         RETURN(0);
1371 }
1372
1373 /**
1374  * Start llog cancel thread
1375  *
1376  * Start llog cancel(master/slave) thread on LOD
1377  *
1378  * \param[in]lclt       cancel log thread to be started.
1379  *
1380  * \retval              0 if the thread is started successfully.
1381  * \retval              negative errno if the thread is not being
1382  *                      started.
1383  */
1384 int distribute_txn_init(const struct lu_env *env,
1385                         struct lu_target *lut,
1386                         struct target_distribute_txn_data *tdtd,
1387                         __u32 index)
1388 {
1389         struct task_struct      *task;
1390         struct l_wait_info       lwi = { 0 };
1391         int                     rc;
1392         ENTRY;
1393
1394         spin_lock_init(&tdtd->tdtd_batchid_lock);
1395         INIT_LIST_HEAD(&tdtd->tdtd_list);
1396
1397         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1398
1399         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1400         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1401         atomic_set(&tdtd->tdtd_refcount, 0);
1402
1403         tdtd->tdtd_lut = lut;
1404         rc = distribute_txn_commit_batchid_init(env, tdtd);
1405         if (rc != 0)
1406                 RETURN(rc);
1407
1408         task = kthread_run(distribute_txn_commit_thread, tdtd, "tdtd-%u",
1409                            index);
1410         if (IS_ERR(task))
1411                 RETURN(PTR_ERR(task));
1412
1413         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1414                      distribute_txn_commit_thread_running(lut) ||
1415                      distribute_txn_commit_thread_stopped(lut), &lwi);
1416         RETURN(0);
1417 }
1418 EXPORT_SYMBOL(distribute_txn_init);
1419
1420 /**
1421  * Stop llog cancel thread
1422  *
1423  * Stop llog cancel(master/slave) thread on LOD and also destory
1424  * all of transaction in the list.
1425  *
1426  * \param[in]lclt       cancel log thread to be stopped.
1427  */
1428 void distribute_txn_fini(const struct lu_env *env,
1429                          struct target_distribute_txn_data *tdtd)
1430 {
1431         struct lu_target *lut = tdtd->tdtd_lut;
1432
1433         /* Stop cancel thread */
1434         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1435                 return;
1436
1437         spin_lock(&tdtd->tdtd_batchid_lock);
1438         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1439         spin_unlock(&tdtd->tdtd_batchid_lock);
1440         wake_up(&tdtd->tdtd_commit_thread_waitq);
1441         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1442                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1443
1444         dtrq_list_destroy(tdtd);
1445         if (tdtd->tdtd_batchid_obj != NULL) {
1446                 lu_object_put(env, &tdtd->tdtd_batchid_obj->do_lu);
1447                 tdtd->tdtd_batchid_obj = NULL;
1448         }
1449 }
1450 EXPORT_SYMBOL(distribute_txn_fini);