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 started %d stopped %d "
  86                        "result %d sub_th %p\n",
  87                        st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
  88                        st->st_committed, st->st_started, st->st_stopped,
  89                        st->st_result, st->st_sub_th);
  90
  91                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
  92                         CDEBUG(mask, " cookie "DFID".%u\n",
  93                                PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
  94                                stc->stc_cookie.lgc_index);
  95                 }
  96         }
  97 }
  98
  99 /**
 100  * Declare write update to sub device
 101  *
 102  * Declare Write updates llog records to the sub device during distribute
 103  * transaction.
 104  *
 105  * \param[in] env       execution environment
 106  * \param[in] record    update records being written
 107  * \param[in] sub_th    sub transaction handle
 108  * \param[in] record_size total update record size
 109  *
 110  * \retval              0 if writing succeeds
 111  * \retval              negative errno if writing fails
 112  */
 113 static int sub_declare_updates_write(const struct lu_env *env,
 114                                      struct llog_update_record *record,
 115                                      struct thandle *sub_th, size_t record_size)
 116 {
 117         struct llog_ctxt        *ctxt;
 118         struct dt_device        *dt = sub_th->th_dev;
 119         int                     left = record_size;
 120         int rc;
 121
 122         /* If ctxt is NULL, it means not need to write update,
 123          * for example if the the OSP is used to connect to OST */
 124         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 125                                 LLOG_UPDATELOG_ORIG_CTXT);
 126
 127         /* Not ready to record updates yet. */
 128         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 129                 llog_ctxt_put(ctxt);
 130                 return 0;
 131         }
 132
 133         rc = llog_declare_add(env, ctxt->loc_handle,
 134                               &record->lur_hdr, sub_th);
 135         if (rc < 0)
 136                 GOTO(out_put, rc);
 137
 138         while (left > ctxt->loc_chunk_size) {
 139                 rc = llog_declare_add(env, ctxt->loc_handle,
 140                                       &record->lur_hdr, sub_th);
 141                 if (rc < 0)
 142                         GOTO(out_put, rc);
 143
 144                 left -= ctxt->loc_chunk_size;
 145         }
 146
 147 out_put:
 148         llog_ctxt_put(ctxt);
 149
 150         return rc;
 151 }
 152
 153 /**
 154  * write update to sub device
 155  *
 156  * Write llog update record to the sub device during distribute
 157  * transaction. If it succeeds, llog cookie of the record will be
 158  * returned by @cookie.
 159  *
 160  * \param[in] env       execution environment
 161  * \param[in] record    update records being written
 162  * \param[in] sub_th    sub transaction handle
 163  * \param[out] cookie   llog cookie of the update record.
 164  *
 165  * \retval              1 if writing succeeds
 166  * \retval              negative errno if writing fails
 167  */
 168 static int sub_updates_write(const struct lu_env *env,
 169                              struct llog_update_record *record,
 170                              struct sub_thandle *sub_th)
 171 {
 172         struct dt_device *dt = sub_th->st_dt;
 173         struct llog_ctxt *ctxt;
 174         struct llog_update_record *lur = NULL;
 175         __u32 update_count = 0;
 176         __u32 param_count = 0;
 177         __u32 last_update_count = 0;
 178         __u32 last_param_count = 0;
 179         char *start;
 180         char *cur;
 181         char *next;
 182         struct sub_thandle_cookie *stc;
 183         size_t reclen;
 184         bool eof = false;
 185         int rc;
 186         ENTRY;
 187
 188         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 189                                 LLOG_UPDATELOG_ORIG_CTXT);
 190         /* If ctxt == NULL, then it means updates on OST (only happens
 191          * during migration), and we do not track those updates for now */
 192         /* If ctxt->loc_handle == NULL, then it does not need to record
 193          * update, usually happens in error handler path */
 194         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 195                 llog_ctxt_put(ctxt);
 196                 RETURN(0);
 197         }
 198
 199         /* Since the cross-MDT updates will includes both local
 200          * and remote updates, the update ops count must > 1 */
 201         LASSERT(record->lur_update_rec.ur_update_count > 1);
 202         LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
 203                  "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
 204                  llog_update_record_size(record));
 205
 206         /*
 207          * If its size > llog chunk_size, then write current chunk to the update
 208          * llog, NB the padding should >= LLOG_MIN_REC_SIZE.
 209          *
 210          * So check padding length is either >= LLOG_MIN_REC_SIZE or is 0
 211          * (record length just matches the chunk size).
 212          */
 213
 214         reclen = record->lur_hdr.lrh_len;
 215         if (reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 216             reclen == ctxt->loc_chunk_size) {
 217                 OBD_ALLOC_PTR(stc);
 218                 if (stc == NULL)
 219                         GOTO(llog_put, rc = -ENOMEM);
 220                 INIT_LIST_HEAD(&stc->stc_list);
 221
 222                 rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
 223                               &stc->stc_cookie, sub_th->st_sub_th);
 224
 225                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 226                        dt->dd_lu_dev.ld_obd->obd_name,
 227                        PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
 228                        stc->stc_cookie.lgc_index, rc);
 229
 230                 if (rc > 0) {
 231                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 232                         rc = 0;
 233                 } else {
 234                         OBD_FREE_PTR(stc);
 235                 }
 236
 237                 GOTO(llog_put, rc);
 238         }
 239
 240         /* Split the records into chunk_size update record */
 241         OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
 242         if (lur == NULL)
 243                 GOTO(llog_put, rc = -ENOMEM);
 244
 245         memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
 246         lur->lur_update_rec.ur_update_count = 0;
 247         lur->lur_update_rec.ur_param_count = 0;
 248         start = (char *)&record->lur_update_rec.ur_ops;
 249         cur = next = start;
 250         do {
 251                 if (update_count < record->lur_update_rec.ur_update_count)
 252                         next = (char *)update_op_next_op(
 253                                                 (struct update_op *)cur);
 254                 else if (param_count < record->lur_update_rec.ur_param_count)
 255                         next = (char *)update_param_next_param(
 256                                                 (struct update_param *)cur);
 257                 else
 258                         eof = true;
 259
 260                 reclen = __llog_update_record_size(
 261                                 __update_records_size(next - start));
 262                 if ((reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 263                      reclen == ctxt->loc_chunk_size) &&
 264                     !eof) {
 265                         cur = next;
 266
 267                         if (update_count <
 268                             record->lur_update_rec.ur_update_count)
 269                                 update_count++;
 270                         else if (param_count <
 271                                  record->lur_update_rec.ur_param_count)
 272                                 param_count++;
 273                         continue;
 274                 }
 275
 276                 lur->lur_update_rec.ur_update_count = update_count -
 277                                                       last_update_count;
 278                 lur->lur_update_rec.ur_param_count = param_count -
 279                                                      last_param_count;
 280                 memcpy(&lur->lur_update_rec.ur_ops, start, cur - start);
 281                 lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 282
 283                 LASSERT(lur->lur_hdr.lrh_len ==
 284                          __llog_update_record_size(
 285                                 __update_records_size(cur - start)));
 286                 LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
 287
 288                 update_records_dump(&lur->lur_update_rec, D_INFO, true);
 289
 290                 OBD_ALLOC_PTR(stc);
 291                 if (stc == NULL)
 292                         GOTO(llog_put, rc = -ENOMEM);
 293                 INIT_LIST_HEAD(&stc->stc_list);
 294
 295                 rc = llog_add(env, ctxt->loc_handle, &lur->lur_hdr,
 296                               &stc->stc_cookie, sub_th->st_sub_th);
 297
 298                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 299                         dt->dd_lu_dev.ld_obd->obd_name,
 300                         PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
 301                         stc->stc_cookie.lgc_index, rc);
 302
 303                 if (rc > 0) {
 304                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 305                         rc = 0;
 306                 } else {
 307                         OBD_FREE_PTR(stc);
 308                         GOTO(llog_put, rc);
 309                 }
 310
 311                 last_update_count = update_count;
 312                 last_param_count = param_count;
 313                 start = cur;
 314                 lur->lur_update_rec.ur_update_count = 0;
 315                 lur->lur_update_rec.ur_param_count = 0;
 316                 lur->lur_update_rec.ur_flags |= UPDATE_RECORD_CONTINUE;
 317         } while (!eof);
 318
 319 llog_put:
 320         if (lur != NULL)
 321                 OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
 322         llog_ctxt_put(ctxt);
 323
 324         RETURN(rc);
 325 }
 326
 327 /**
 328  * Prepare the update records.
 329  *
 330  * Merge params and ops into the update records, then initializing
 331  * the update buffer.
 332  *
 333  * During transaction execution phase, parameters and update ops
 334  * are collected in two different buffers (see lod_updates_pack()),
 335  * during transaction stop, it needs to be merged in one buffer,
 336  * so it will be written in the update log.
 337  *
 338  * \param[in] env       execution environment
 339  * \param[in] tmt       top_multiple_thandle for distribute txn
 340  *
 341  * \retval              0 if merging succeeds.
 342  * \retval              negaitive errno if merging fails.
 343  */
 344 static int prepare_writing_updates(const struct lu_env *env,
 345                                    struct top_multiple_thandle *tmt)
 346 {
 347         struct thandle_update_records   *tur = tmt->tmt_update_records;
 348         struct llog_update_record       *lur;
 349         struct update_params *params;
 350         size_t params_size;
 351         size_t update_size;
 352
 353         if (tur == NULL || tur->tur_update_records == NULL ||
 354             tur->tur_update_params == NULL)
 355                 return 0;
 356
 357         lur = tur->tur_update_records;
 358         /* Extends the update records buffer if needed */
 359         params_size = update_params_size(tur->tur_update_params,
 360                                          tur->tur_update_param_count);
 361         LASSERT(lur->lur_update_rec.ur_param_count == 0);
 362         update_size = llog_update_record_size(lur);
 363         if (cfs_size_round(update_size + params_size) >
 364             tur->tur_update_records_buf_size) {
 365                 int rc;
 366
 367                 rc = tur_update_records_extend(tur,
 368                         cfs_size_round(update_size + params_size));
 369                 if (rc < 0)
 370                         return rc;
 371
 372                 lur = tur->tur_update_records;
 373         }
 374
 375         params = update_records_get_params(&lur->lur_update_rec);
 376         memcpy(params, tur->tur_update_params, params_size);
 377
 378         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
 379         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
 380         /* Init update record header */
 381         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 382         lur->lur_hdr.lrh_type = UPDATE_REC;
 383
 384         /* Dump updates for debugging purpose */
 385         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 386
 387         return 0;
 388 }
 389
 390 /**
 391  * Top thandle commit callback
 392  *
 393  * This callback will be called when all of sub transactions are committed.
 394  *
 395  * \param[in] th        top thandle to be committed.
 396  */
 397 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
 398 {
 399         struct lu_target *lut;
 400         ENTRY;
 401
 402         LASSERT(atomic_read(&tmt->tmt_refcount) > 0);
 403
 404         top_multiple_thandle_dump(tmt, D_HA);
 405         tmt->tmt_committed = 1;
 406         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
 407         if (lut->lut_tdtd && lut->lut_tdtd->tdtd_commit_task)
 408                 wake_up_process(lut->lut_tdtd->tdtd_commit_task);
 409
 410         RETURN_EXIT;
 411 }
 412
 413 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
 414                                        struct dt_device *dt_dev)
 415 {
 416         struct sub_thandle *st;
 417
 418         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 419                 if (st->st_dt == dt_dev)
 420                         return st;
 421         }
 422         return NULL;
 423 }
 424 EXPORT_SYMBOL(lookup_sub_thandle);
 425
 426 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
 427                                        struct dt_device *dt_dev)
 428 {
 429         struct sub_thandle *st;
 430
 431         OBD_ALLOC_PTR(st);
 432         if (st == NULL)
 433                 RETURN(ERR_PTR(-ENOMEM));
 434
 435         INIT_LIST_HEAD(&st->st_sub_list);
 436         INIT_LIST_HEAD(&st->st_cookie_list);
 437         st->st_dt = dt_dev;
 438
 439         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
 440         return st;
 441 }
 442
 443 static void sub_trans_commit_cb_internal(struct top_multiple_thandle *tmt,
 444                                          struct thandle *sub_th, int err)
 445 {
 446         struct sub_thandle      *st;
 447         bool                    all_committed = true;
 448
 449         /* Check if all sub thandles are committed */
 450         spin_lock(&tmt->tmt_sub_lock);
 451         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 452                 if (st->st_sub_th == sub_th) {
 453                         st->st_committed = 1;
 454                         st->st_result = err;
 455                 }
 456                 if (!st->st_committed)
 457                         all_committed = false;
 458         }
 459         spin_unlock(&tmt->tmt_sub_lock);
 460
 461         if (tmt->tmt_result == 0)
 462                 tmt->tmt_result = err;
 463
 464         if (all_committed)
 465                 top_trans_committed_cb(tmt);
 466
 467         top_multiple_thandle_dump(tmt, D_INFO);
 468         top_multiple_thandle_put(tmt);
 469         RETURN_EXIT;
 470 }
 471
 472 /**
 473  * sub thandle commit callback
 474  *
 475  * Mark the sub thandle to be committed and if all sub thandle are committed
 476  * notify the top thandle.
 477  *
 478  * \param[in] env       execution environment
 479  * \param[in] sub_th    sub thandle being committed
 480  * \param[in] cb        commit callback
 481  * \param[in] err       trans result
 482  */
 483 static void sub_trans_commit_cb(struct lu_env *env,
 484                                 struct thandle *sub_th,
 485                                 struct dt_txn_commit_cb *cb, int err)
 486 {
 487         struct top_multiple_thandle *tmt = cb->dcb_data;
 488
 489         sub_trans_commit_cb_internal(tmt, sub_th, err);
 490 }
 491
 492 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
 493                                     struct top_multiple_thandle *tmt)
 494 {
 495         LASSERT(st->st_sub_th != NULL);
 496         top_multiple_thandle_get(tmt);
 497         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
 498         st->st_commit_dcb.dcb_data = tmt;
 499         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
 500         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
 501 }
 502
 503 /**
 504  * Sub thandle stop call back
 505  *
 506  * After sub thandle is stopped, it will call this callback to notify
 507  * the top thandle.
 508  *
 509  * \param[in] th        sub thandle to be stopped
 510  * \param[in] rc        result of sub trans
 511  */
 512 static void sub_trans_stop_cb(struct lu_env *env,
 513                               struct thandle *sub_th,
 514                               struct dt_txn_commit_cb *cb, int err)
 515 {
 516         struct sub_thandle              *st;
 517         struct top_multiple_thandle     *tmt = cb->dcb_data;
 518         ENTRY;
 519
 520         spin_lock(&tmt->tmt_sub_lock);
 521         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 522                 if (st->st_stopped)
 523                         continue;
 524
 525                 if (st->st_dt == sub_th->th_dev) {
 526                         st->st_stopped = 1;
 527                         st->st_result = err;
 528                         break;
 529                 }
 530         }
 531         spin_unlock(&tmt->tmt_sub_lock);
 532
 533         wake_up(&tmt->tmt_stop_waitq);
 534         RETURN_EXIT;
 535 }
 536
 537 static void sub_thandle_register_stop_cb(struct sub_thandle *st,
 538                                          struct top_multiple_thandle *tmt)
 539 {
 540         st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
 541         st->st_stop_dcb.dcb_data = tmt;
 542         st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
 543         INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
 544         dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
 545 }
 546
 547 /**
 548  * Create sub thandle
 549  *
 550  * Create transaction handle for sub_thandle
 551  *
 552  * \param[in] env       execution environment
 553  * \param[in] th        top thandle
 554  * \param[in] st        sub_thandle
 555  *
 556  * \retval              0 if creation succeeds.
 557  * \retval              negative errno if creation fails.
 558  */
 559 int sub_thandle_trans_create(const struct lu_env *env,
 560                              struct top_thandle *top_th,
 561                              struct sub_thandle *st)
 562 {
 563         struct thandle *sub_th;
 564
 565         sub_th = dt_trans_create(env, st->st_dt);
 566         if (IS_ERR(sub_th))
 567                 return PTR_ERR(sub_th);
 568
 569         sub_th->th_top = &top_th->tt_super;
 570         st->st_sub_th = sub_th;
 571
 572         sub_th->th_wait_submit = 1;
 573         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
 574         return 0;
 575 }
 576
 577 /**
 578  * Create the top transaction.
 579  *
 580  * Create the top transaction on the master device. It will create a top
 581  * thandle and a sub thandle on the master device.
 582  *
 583  * \param[in] env               execution environment
 584  * \param[in] master_dev        master_dev the top thandle will be created
 585  *
 586  * \retval                      pointer to the created thandle.
 587  * \retval                      ERR_PTR(errno) if creation failed.
 588  */
 589 struct thandle *
 590 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
 591 {
 592         struct top_thandle      *top_th;
 593         struct thandle          *child_th;
 594
 595         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
 596         if (top_th == NULL)
 597                 return ERR_PTR(-ENOMEM);
 598
 599         top_th->tt_super.th_top = &top_th->tt_super;
 600
 601         if (master_dev != NULL) {
 602                 child_th = dt_trans_create(env, master_dev);
 603                 if (IS_ERR(child_th)) {
 604                         OBD_FREE_PTR(top_th);
 605                         return child_th;
 606                 }
 607
 608                 child_th->th_top = &top_th->tt_super;
 609                 child_th->th_wait_submit = 1;
 610                 top_th->tt_master_sub_thandle = child_th;
 611         }
 612         return &top_th->tt_super;
 613 }
 614 EXPORT_SYMBOL(top_trans_create);
 615
 616 /**
 617  * Declare write update transaction
 618  *
 619  * Check if there are updates being recorded in this transaction,
 620  * it will write the record into the disk.
 621  *
 622  * \param[in] env       execution environment
 623  * \param[in] tmt       top multiple transaction handle
 624  *
 625  * \retval              0 if writing succeeds
 626  * \retval              negative errno if writing fails
 627  */
 628 static int declare_updates_write(const struct lu_env *env,
 629                                  struct top_multiple_thandle *tmt)
 630 {
 631         struct llog_update_record *record;
 632         struct sub_thandle *st;
 633         int rc = 0;
 634
 635         record = tmt->tmt_update_records->tur_update_records;
 636         /* Declare update write for all other target */
 637         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 638                 if (st->st_sub_th == NULL)
 639                         continue;
 640
 641                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
 642                                                tmt->tmt_record_size);
 643                 if (rc < 0)
 644                         break;
 645         }
 646
 647         return rc;
 648 }
 649
 650 /**
 651  * Assign batchid to the distribute transaction.
 652  *
 653  * Assign batchid to the distribute transaction
 654  *
 655  * \param[in] tmt       distribute transaction
 656  */
 657 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 658 {
 659         struct target_distribute_txn_data *tdtd;
 660         struct dt_device *dt = new->tmt_master_sub_dt;
 661         struct sub_thandle *st;
 662
 663         LASSERT(dt != NULL);
 664         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 665         spin_lock(&tdtd->tdtd_batchid_lock);
 666         new->tmt_batchid = tdtd->tdtd_batchid++;
 667         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
 668         spin_unlock(&tdtd->tdtd_batchid_lock);
 669         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 670                 if (st->st_sub_th != NULL)
 671                         sub_thandle_register_commit_cb(st, new);
 672         }
 673         top_multiple_thandle_get(new);
 674         top_multiple_thandle_dump(new, D_INFO);
 675 }
 676
 677 /**
 678  * Insert distribute transaction to the distribute txn list.
 679  *
 680  * Insert distribute transaction to the distribute txn list.
 681  *
 682  * \param[in] new       the distribute txn to be inserted.
 683  */
 684 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
 685 {
 686         struct dt_device *dt = new->tmt_master_sub_dt;
 687         struct top_multiple_thandle *tmt;
 688         struct target_distribute_txn_data *tdtd;
 689         struct sub_thandle *st;
 690         bool    at_head = false;
 691
 692         LASSERT(dt != NULL);
 693         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 694
 695         spin_lock(&tdtd->tdtd_batchid_lock);
 696         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
 697                 if (new->tmt_batchid > tmt->tmt_batchid) {
 698                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
 699                         break;
 700                 }
 701         }
 702         if (list_empty(&new->tmt_commit_list)) {
 703                 at_head = true;
 704                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
 705         }
 706         spin_unlock(&tdtd->tdtd_batchid_lock);
 707
 708         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 709                 if (st->st_sub_th != NULL)
 710                         sub_thandle_register_commit_cb(st, new);
 711         }
 712
 713         top_multiple_thandle_get(new);
 714         top_multiple_thandle_dump(new, D_INFO);
 715         if (new->tmt_committed && at_head && tdtd->tdtd_commit_task)
 716                 wake_up_process(tdtd->tdtd_commit_task);
 717 }
 718
 719 /**
 720  * Prepare cross-MDT operation.
 721  *
 722  * Create the update record buffer to record updates for cross-MDT operation,
 723  * add master sub transaction to tt_sub_trans_list, and declare the update
 724  * writes.
 725  *
 726  * During updates packing, all of parameters will be packed in
 727  * tur_update_params, and updates will be packed in tur_update_records.
 728  * Then in transaction stop, parameters and updates will be merged
 729  * into one updates buffer.
 730  *
 731  * And also master thandle will be added to the sub_th list, so it will be
 732  * easy to track the commit status.
 733  *
 734  * \param[in] env       execution environment
 735  * \param[in] th        top transaction handle
 736  *
 737  * \retval              0 if preparation succeeds.
 738  * \retval              negative errno if preparation fails.
 739  */
 740 static int prepare_multiple_node_trans(const struct lu_env *env,
 741                                        struct top_multiple_thandle *tmt)
 742 {
 743         struct thandle_update_records   *tur;
 744         int                             rc;
 745         ENTRY;
 746
 747         if (tmt->tmt_update_records == NULL) {
 748                 tur = &update_env_info(env)->uti_tur;
 749                 rc = check_and_prepare_update_record(env, tur);
 750                 if (rc < 0)
 751                         RETURN(rc);
 752
 753                 tmt->tmt_update_records = tur;
 754                 distribute_txn_assign_batchid(tmt);
 755         }
 756
 757         rc = declare_updates_write(env, tmt);
 758
 759         RETURN(rc);
 760 }
 761
 762 /**
 763  * start the top transaction.
 764  *
 765  * Start all of its sub transactions, then start master sub transaction.
 766  *
 767  * \param[in] env               execution environment
 768  * \param[in] master_dev        master_dev the top thandle will be start
 769  * \param[in] th                top thandle
 770  *
 771  * \retval                      0 if transaction start succeeds.
 772  * \retval                      negative errno if start fails.
 773  */
 774 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
 775                     struct thandle *th)
 776 {
 777         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 778                                                        tt_super);
 779         struct sub_thandle              *st;
 780         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
 781         int                             rc = 0;
 782         ENTRY;
 783
 784         if (tmt == NULL) {
 785                 if (th->th_sync)
 786                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 787                 if (th->th_local)
 788                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 789                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
 790                                     top_th->tt_master_sub_thandle);
 791                 RETURN(rc);
 792         }
 793
 794         tmt = top_th->tt_multiple_thandle;
 795         rc = prepare_multiple_node_trans(env, tmt);
 796         if (rc < 0)
 797                 RETURN(rc);
 798
 799         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 800                 if (st->st_sub_th == NULL)
 801                         continue;
 802                 if (th->th_sync)
 803                         st->st_sub_th->th_sync = th->th_sync;
 804                 if (th->th_local)
 805                         st->st_sub_th->th_local = th->th_local;
 806                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
 807                                     st->st_sub_th);
 808                 if (rc != 0)
 809                         GOTO(out, rc);
 810
 811                 LASSERT(st->st_started == 0);
 812                 st->st_started = 1;
 813         }
 814 out:
 815         th->th_result = rc;
 816         RETURN(rc);
 817 }
 818 EXPORT_SYMBOL(top_trans_start);
 819
 820 /**
 821  * Check whether we need write updates record
 822  *
 823  * Check if the updates for the top_thandle needs to be writen
 824  * to all targets. Only if the transaction succeeds and the updates
 825  * number > 2, it will write the updates,
 826  *
 827  * \params [in] top_th  top thandle.
 828  *
 829  * \retval              true if it needs to write updates
 830  * \retval              false if it does not need to write updates
 831  **/
 832 static bool top_check_write_updates(struct top_thandle *top_th)
 833 {
 834         struct top_multiple_thandle     *tmt;
 835         struct thandle_update_records   *tur;
 836
 837         /* Do not write updates to records if the transaction fails */
 838         if (top_th->tt_super.th_result != 0)
 839                 return false;
 840
 841         tmt = top_th->tt_multiple_thandle;
 842         if (tmt == NULL)
 843                 return false;
 844
 845         tur = tmt->tmt_update_records;
 846         if (tur == NULL)
 847                 return false;
 848
 849         /* Hmm, false update records, since the cross-MDT operation
 850          * should includes both local and remote updates, so the
 851          * updates count should >= 2 */
 852         if (tur->tur_update_records == NULL ||
 853             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
 854                 return false;
 855
 856         return true;
 857 }
 858
 859 /**
 860  * Check if top transaction is stopped
 861  *
 862  * Check if top transaction is stopped, only if all sub transaction
 863  * is stopped, then the top transaction is stopped.
 864  *
 865  * \param [in] top_th   top thandle
 866  *
 867  * \retval              true if the top transaction is stopped.
 868  * \retval              false if the top transaction is not stopped.
 869  */
 870 static bool top_trans_is_stopped(struct top_thandle *top_th)
 871 {
 872         struct top_multiple_thandle     *tmt;
 873         struct sub_thandle              *st;
 874         bool                    all_stopped = true;
 875
 876         tmt = top_th->tt_multiple_thandle;
 877         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 878                 if (!st->st_stopped && st->st_sub_th != NULL) {
 879                         all_stopped = false;
 880                         break;
 881                 }
 882
 883                 if (st->st_result != 0 &&
 884                     top_th->tt_super.th_result == 0)
 885                         top_th->tt_super.th_result = st->st_result;
 886         }
 887
 888         return all_stopped;
 889 }
 890
 891 /**
 892  * Wait result of top transaction
 893  *
 894  * Wait until all sub transaction get its result.
 895  *
 896  * \param [in] top_th   top thandle.
 897  *
 898  * \retval              the result of top thandle.
 899  */
 900 static int top_trans_wait_result(struct top_thandle *top_th)
 901 {
 902         wait_event_idle(top_th->tt_multiple_thandle->tmt_stop_waitq,
 903                         top_trans_is_stopped(top_th));
 904
 905         RETURN(top_th->tt_super.th_result);
 906 }
 907
 908 /**
 909  * Stop the top transaction.
 910  *
 911  * Stop the transaction on the master device first, then stop transactions
 912  * on other sub devices.
 913  *
 914  * \param[in] env               execution environment
 915  * \param[in] master_dev        master_dev the top thandle will be created
 916  * \param[in] th                top thandle
 917  *
 918  * \retval                      0 if stop transaction succeeds.
 919  * \retval                      negative errno if stop transaction fails.
 920  */
 921 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
 922                    struct thandle *th)
 923 {
 924         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 925                                                        tt_super);
 926         struct sub_thandle              *st;
 927         struct sub_thandle              *master_st;
 928         struct top_multiple_thandle     *tmt;
 929         struct thandle_update_records   *tur;
 930         bool                            write_updates = false;
 931         int                     rc = 0;
 932         ENTRY;
 933
 934         if (likely(top_th->tt_multiple_thandle == NULL)) {
 935                 LASSERT(master_dev != NULL);
 936
 937                 if (th->th_sync)
 938                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 939                 if (th->th_local)
 940                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 941                 rc = dt_trans_stop(env, master_dev,
 942                                    top_th->tt_master_sub_thandle);
 943                 OBD_FREE_PTR(top_th);
 944                 RETURN(rc);
 945         }
 946
 947         tmt = top_th->tt_multiple_thandle;
 948         tur = tmt->tmt_update_records;
 949
 950         /* Note: we need stop the master thandle first, then the stop
 951          * callback will fill the master transno in the update logs,
 952          * then these update logs will be sent to other MDTs */
 953         /* get the master sub thandle */
 954         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
 955         write_updates = top_check_write_updates(top_th);
 956
 957         /* Step 1: write the updates log on Master MDT */
 958         if (master_st != NULL && master_st->st_sub_th != NULL &&
 959             write_updates) {
 960                 struct llog_update_record *lur;
 961
 962                 /* Merge the parameters and updates into one buffer */
 963                 rc = prepare_writing_updates(env, tmt);
 964                 if (rc < 0) {
 965                         CERROR("%s: cannot prepare updates: rc = %d\n",
 966                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 967                         th->th_result = rc;
 968                         write_updates = false;
 969                         GOTO(stop_master_trans, rc);
 970                 }
 971
 972                 lur = tur->tur_update_records;
 973                 /* Write updates to the master MDT */
 974                 rc = sub_updates_write(env, lur, master_st);
 975
 976                 /* Cleanup the common parameters in the update records,
 977                  * master transno callback might add more parameters.
 978                  * and we need merge the update records again in the
 979                  * following */
 980                 if (tur->tur_update_params != NULL)
 981                         lur->lur_update_rec.ur_param_count = 0;
 982
 983                 if (rc < 0) {
 984                         CERROR("%s: write updates failed: rc = %d\n",
 985                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 986                         th->th_result = rc;
 987                         write_updates = false;
 988                         GOTO(stop_master_trans, rc);
 989                 }
 990         }
 991
 992 stop_master_trans:
 993         /* Step 2: Stop the transaction on the master MDT, and fill the
 994          * master transno in the update logs to other MDT. */
 995         if (master_st != NULL && master_st->st_sub_th != NULL) {
 996                 if (th->th_local)
 997                         master_st->st_sub_th->th_local = th->th_local;
 998                 if (th->th_sync)
 999                         master_st->st_sub_th->th_sync = th->th_sync;
1000                 master_st->st_sub_th->th_result = th->th_result;
1001                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
1002                 /* If it does not write_updates, then we call submit callback
1003                  * here, otherwise callback is done through
1004                  * osd(osp)_trans_commit_cb() */
1005                 if (!master_st->st_started &&
1006                     !list_empty(&tmt->tmt_commit_list))
1007                         sub_trans_commit_cb_internal(tmt,
1008                                                 master_st->st_sub_th, rc);
1009                 if (rc < 0) {
1010                         CERROR("%s: stop trans failed: rc = %d\n",
1011                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1012                         th->th_result = rc;
1013                         GOTO(stop_other_trans, rc);
1014                 } else if (tur != NULL && tur->tur_update_records != NULL) {
1015                         struct llog_update_record *lur;
1016
1017                         lur = tur->tur_update_records;
1018                         if (lur->lur_update_rec.ur_master_transno == 0)
1019                                 /* Update master transno after master stop
1020                                  * callback */
1021                                 lur->lur_update_rec.ur_master_transno =
1022                                                 tgt_th_info(env)->tti_transno;
1023                 }
1024         }
1025
1026         /* Step 3: write updates to other MDTs */
1027         if (write_updates) {
1028                 struct llog_update_record *lur;
1029                 if (OBD_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
1030                         if (cfs_fail_val == 1) {
1031                                 long timeout = cfs_time_seconds(1) / 10;
1032
1033                                 OBD_RACE(OBD_FAIL_OUT_OBJECT_MISS);
1034                                 set_current_state(TASK_UNINTERRUPTIBLE);
1035                                 schedule_timeout(schedule_timeout(timeout));
1036                                 cfs_fail_loc = 0;
1037                         }
1038                         cfs_fail_val++;
1039                 }
1040
1041                 /* Stop callback of master will add more updates and also update
1042                  * master transno, so merge the parameters and updates into one
1043                  * buffer again */
1044                 rc = prepare_writing_updates(env, tmt);
1045                 if (rc < 0) {
1046                         CERROR("%s: prepare updates failed: rc = %d\n",
1047                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1048                         th->th_result = rc;
1049                         GOTO(stop_other_trans, rc);
1050                 }
1051                 lur = tur->tur_update_records;
1052                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1053                                     st_sub_list) {
1054                         if (st->st_sub_th == NULL || st == master_st ||
1055                             st->st_sub_th->th_result < 0)
1056                                 continue;
1057
1058                         rc = sub_updates_write(env, lur, st);
1059                         if (rc < 0) {
1060                                 CERROR("%s: write updates failed: rc = %d\n",
1061                                        st->st_dt->dd_lu_dev.ld_obd->obd_name,
1062                                        rc);
1063                                 th->th_result = rc;
1064                                 break;
1065                         }
1066                 }
1067         }
1068
1069 stop_other_trans:
1070         /* Step 4: Stop the transaction on other MDTs */
1071         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1072                 if (st == master_st || st->st_sub_th == NULL)
1073                         continue;
1074
1075                 if (th->th_sync)
1076                         st->st_sub_th->th_sync = th->th_sync;
1077                 if (th->th_local)
1078                         st->st_sub_th->th_local = th->th_local;
1079                 st->st_sub_th->th_result = th->th_result;
1080                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1081                                    st->st_sub_th);
1082                 if (rc < 0) {
1083                         CERROR("%s: stop trans failed: rc = %d\n",
1084                                st->st_dt->dd_lu_dev.ld_obd->obd_name, rc);
1085                         if (th->th_result == 0)
1086                                 th->th_result = rc;
1087                 }
1088         }
1089
1090         rc = top_trans_wait_result(top_th);
1091
1092         tmt->tmt_result = rc;
1093
1094         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1095          * multiple node transaction, the top transaction will be destroyed. */
1096         top_multiple_thandle_put(tmt);
1097         OBD_FREE_PTR(top_th);
1098         RETURN(rc);
1099 }
1100 EXPORT_SYMBOL(top_trans_stop);
1101
1102 /**
1103  * Create top_multiple_thandle for top_thandle
1104  *
1105  * Create top_mutilple_thandle to manage the mutiple node transaction
1106  * for top_thandle, and it also needs to add master sub thandle to the
1107  * sub trans list now.
1108  *
1109  * \param[in] env       execution environment
1110  * \param[in] top_th    the top thandle
1111  *
1112  * \retval      0 if creation succeeds
1113  * \retval      negative errno if creation fails
1114  */
1115 int top_trans_create_tmt(const struct lu_env *env,
1116                          struct top_thandle *top_th)
1117 {
1118         struct top_multiple_thandle *tmt;
1119
1120         OBD_ALLOC_PTR(tmt);
1121         if (tmt == NULL)
1122                 return -ENOMEM;
1123
1124         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1125         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1126         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1127         atomic_set(&tmt->tmt_refcount, 1);
1128         spin_lock_init(&tmt->tmt_sub_lock);
1129         init_waitqueue_head(&tmt->tmt_stop_waitq);
1130
1131         top_th->tt_multiple_thandle = tmt;
1132
1133         return 0;
1134 }
1135
1136 static struct sub_thandle *
1137 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1138                                 struct thandle *sub_th)
1139 {
1140         struct sub_thandle *st;
1141
1142         /* create and init sub th to the top trans list */
1143         st = create_sub_thandle(top_th->tt_multiple_thandle,
1144                                 sub_th->th_dev);
1145         if (IS_ERR(st))
1146                 return st;
1147
1148         st->st_sub_th = sub_th;
1149
1150         sub_th->th_top = &top_th->tt_super;
1151         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
1152         return st;
1153 }
1154
1155 /**
1156  * Get sub thandle.
1157  *
1158  * Get sub thandle from the top thandle according to the sub dt_device.
1159  *
1160  * \param[in] env       execution environment
1161  * \param[in] th        thandle on the top layer.
1162  * \param[in] sub_dt    sub dt_device used to get sub transaction
1163  *
1164  * \retval              thandle of sub transaction if succeed
1165  * \retval              PTR_ERR(errno) if failed
1166  */
1167 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1168                                       struct thandle *th,
1169                                       struct dt_device *sub_dt)
1170 {
1171         struct sub_thandle      *st = NULL;
1172         struct sub_thandle      *master_st = NULL;
1173         struct top_thandle      *top_th;
1174         struct thandle          *sub_th = NULL;
1175         int                     rc = 0;
1176         ENTRY;
1177
1178         top_th = container_of(th, struct top_thandle, tt_super);
1179
1180         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1181                 RETURN(top_th->tt_master_sub_thandle);
1182
1183         if (top_th->tt_multiple_thandle != NULL) {
1184                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1185                 if (st != NULL)
1186                         RETURN(st->st_sub_th);
1187         }
1188
1189         sub_th = dt_trans_create(env, sub_dt);
1190         if (IS_ERR(sub_th))
1191                 RETURN(sub_th);
1192
1193         /* Create top_multiple_thandle if necessary */
1194         if (top_th->tt_multiple_thandle == NULL) {
1195                 struct top_multiple_thandle *tmt;
1196
1197                 rc = top_trans_create_tmt(env, top_th);
1198                 if (rc < 0)
1199                         GOTO(stop_trans, rc);
1200
1201                 tmt = top_th->tt_multiple_thandle;
1202
1203                 /* Add master sub th to the top trans list */
1204                 tmt->tmt_master_sub_dt =
1205                         top_th->tt_master_sub_thandle->th_dev;
1206                 master_st = create_sub_thandle_with_thandle(top_th,
1207                                         top_th->tt_master_sub_thandle);
1208                 if (IS_ERR(master_st)) {
1209                         rc = PTR_ERR(master_st);
1210                         master_st = NULL;
1211                         GOTO(stop_trans, rc);
1212                 }
1213         }
1214
1215         /* create and init sub th to the top trans list */
1216         st = create_sub_thandle_with_thandle(top_th, sub_th);
1217         if (IS_ERR(st)) {
1218                 rc = PTR_ERR(st);
1219                 st = NULL;
1220                 GOTO(stop_trans, rc);
1221         }
1222         st->st_sub_th->th_wait_submit = 1;
1223 stop_trans:
1224         if (rc < 0) {
1225                 if (master_st != NULL) {
1226                         list_del(&master_st->st_sub_list);
1227                         OBD_FREE_PTR(master_st);
1228                 }
1229                 sub_th->th_result = rc;
1230                 dt_trans_stop(env, sub_dt, sub_th);
1231                 sub_th = ERR_PTR(rc);
1232         }
1233
1234         RETURN(sub_th);
1235 }
1236 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1237
1238 /**
1239  * Top multiple thandle destroy
1240  *
1241  * Destroy multiple thandle and all its sub thandle.
1242  *
1243  * \param[in] tmt       top_multiple_thandle to be destroyed.
1244  */
1245 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1246 {
1247         struct sub_thandle *st;
1248         struct sub_thandle *tmp;
1249
1250         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1251         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1252                                  st_sub_list) {
1253                 struct sub_thandle_cookie *stc;
1254                 struct sub_thandle_cookie *tmp;
1255
1256                 list_del(&st->st_sub_list);
1257                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1258                                          stc_list) {
1259                         list_del(&stc->stc_list);
1260                         OBD_FREE_PTR(stc);
1261                 }
1262                 OBD_FREE_PTR(st);
1263         }
1264         OBD_FREE_PTR(tmt);
1265 }
1266 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1267
1268 /**
1269  * Cancel the update log on MDTs
1270  *
1271  * Cancel the update log on MDTs then destroy the thandle.
1272  *
1273  * \param[in] env       execution environment
1274  * \param[in] tmt       the top multiple thandle whose updates records
1275  *                      will be cancelled.
1276  *
1277  * \retval              0 if cancellation succeeds.
1278  * \retval              negative errno if cancellation fails.
1279  */
1280 static int distribute_txn_cancel_records(const struct lu_env *env,
1281                                          struct top_multiple_thandle *tmt)
1282 {
1283         struct sub_thandle *st;
1284         ENTRY;
1285
1286         top_multiple_thandle_dump(tmt, D_INFO);
1287         /* Cancel update logs on other MDTs */
1288         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1289                 struct llog_ctxt        *ctxt;
1290                 struct obd_device       *obd;
1291                 struct llog_cookie      *cookie;
1292                 struct sub_thandle_cookie *stc;
1293                 int rc;
1294
1295                 obd = st->st_dt->dd_lu_dev.ld_obd;
1296                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1297                 if (ctxt == NULL)
1298                         continue;
1299                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1300                         cookie = &stc->stc_cookie;
1301                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1302                                 continue;
1303
1304                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1305                                                      cookie);
1306                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1307                                DFID".%u: rc = %d\n", obd->obd_name,
1308                                tmt->tmt_batchid,
1309                                PFID(&cookie->lgc_lgl.lgl_oi.oi_fid),
1310                                cookie->lgc_index, rc);
1311                 }
1312
1313                 llog_ctxt_put(ctxt);
1314         }
1315
1316         RETURN(0);
1317 }
1318
1319 struct distribute_txn_bid_data {
1320         struct dt_txn_commit_cb  dtbd_cb;
1321         struct target_distribute_txn_data      *dtbd_tdtd;
1322         __u64                    dtbd_batchid;
1323 };
1324
1325 /**
1326  * callback of updating commit batchid
1327  *
1328  * Updating commit batchid then wake up the commit thread to cancel the
1329  * records.
1330  *
1331  * \param[in]env        execution environment
1332  * \param[in]th         thandle to updating commit batchid
1333  * \param[in]cb         commit callback
1334  * \param[in]err        result of thandle
1335  */
1336 static void distribute_txn_batchid_cb(struct lu_env *env,
1337                                       struct thandle *th,
1338                                       struct dt_txn_commit_cb *cb,
1339                                       int err)
1340 {
1341         struct distribute_txn_bid_data          *dtbd = NULL;
1342         struct target_distribute_txn_data       *tdtd;
1343
1344         dtbd = container_of(cb, struct distribute_txn_bid_data, dtbd_cb);
1345         tdtd = dtbd->dtbd_tdtd;
1346
1347         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1348               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1349         spin_lock(&tdtd->tdtd_batchid_lock);
1350         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1351             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1352                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1353         spin_unlock(&tdtd->tdtd_batchid_lock);
1354         if (atomic_dec_and_test(&tdtd->tdtd_refcount))
1355                 wake_up_process(tdtd->tdtd_commit_task);
1356
1357         OBD_FREE_PTR(dtbd);
1358 }
1359
1360 /**
1361  * Update the commit batchid in disk
1362  *
1363  * Update commit batchid in the disk, after this is committed, it can start
1364  * to cancel the update records.
1365  *
1366  * \param[in] env       execution environment
1367  * \param[in] tdtd      distribute transaction structure
1368  * \param[in] batchid   commit batchid to be updated
1369  *
1370  * \retval              0 if update succeeds.
1371  * \retval              negative errno if update fails.
1372  */
1373 static int
1374 distribute_txn_commit_batchid_update(const struct lu_env *env,
1375                               struct target_distribute_txn_data *tdtd,
1376                               __u64 batchid)
1377 {
1378         struct distribute_txn_bid_data  *dtbd = NULL;
1379         struct thandle          *th;
1380         struct lu_buf            buf;
1381         __u64                    tmp;
1382         __u64                    off;
1383         int                      rc;
1384         ENTRY;
1385
1386         OBD_ALLOC_PTR(dtbd);
1387         if (dtbd == NULL)
1388                 RETURN(-ENOMEM);
1389         dtbd->dtbd_batchid = batchid;
1390         dtbd->dtbd_tdtd = tdtd;
1391         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1392         atomic_inc(&tdtd->tdtd_refcount);
1393
1394         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1395         if (IS_ERR(th)) {
1396                 atomic_dec(&tdtd->tdtd_refcount);
1397                 OBD_FREE_PTR(dtbd);
1398                 RETURN(PTR_ERR(th));
1399         }
1400
1401         tmp = cpu_to_le64(batchid);
1402         buf.lb_buf = &tmp;
1403         buf.lb_len = sizeof(tmp);
1404         off = 0;
1405
1406         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1407                                      th);
1408         if (rc < 0)
1409                 GOTO(stop, rc);
1410
1411         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1412         if (rc < 0)
1413                 GOTO(stop, rc);
1414
1415         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1416         if (rc < 0)
1417                 GOTO(stop, rc);
1418
1419         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1420                              &off, th);
1421
1422         CDEBUG(D_INFO, "%s: update batchid %llu: rc = %d\n",
1423                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1424
1425 stop:
1426         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1427         if (rc < 0) {
1428                 atomic_dec(&tdtd->tdtd_refcount);
1429                 OBD_FREE_PTR(dtbd);
1430         }
1431         RETURN(rc);
1432 }
1433
1434 /**
1435  * Init commit batchid for distribute transaction.
1436  *
1437  * Initialize the batchid object and get commit batchid from the object.
1438  *
1439  * \param[in] env       execution environment
1440  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1441  *
1442  * \retval              0 if initialization succeeds.
1443  * \retval              negative errno if initialization fails.
1444  **/
1445 static int
1446 distribute_txn_commit_batchid_init(const struct lu_env *env,
1447                                    struct target_distribute_txn_data *tdtd)
1448 {
1449         struct tgt_thread_info  *tti = tgt_th_info(env);
1450         struct lu_target        *lut = tdtd->tdtd_lut;
1451         struct lu_attr          *attr = &tti->tti_attr;
1452         struct lu_fid           *fid = &tti->tti_fid1;
1453         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1454         struct dt_object        *dt_obj = NULL;
1455         struct lu_buf           buf;
1456         __u64                   tmp;
1457         __u64                   off;
1458         int                     rc;
1459         ENTRY;
1460
1461         memset(attr, 0, sizeof(*attr));
1462         attr->la_valid = LA_MODE;
1463         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1464         dof->dof_type = dt_mode_to_dft(S_IFREG);
1465
1466         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1467
1468         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1469                                    attr);
1470         if (IS_ERR(dt_obj)) {
1471                 rc = PTR_ERR(dt_obj);
1472                 dt_obj = NULL;
1473                 GOTO(out_put, rc);
1474         }
1475
1476         tdtd->tdtd_batchid_obj = dt_obj;
1477
1478         buf.lb_buf = &tmp;
1479         buf.lb_len = sizeof(tmp);
1480         off = 0;
1481         rc = dt_read(env, dt_obj, &buf, &off);
1482         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1483                 CERROR("%s can't read last committed batchid: rc = %d\n",
1484                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1485                 if (rc > 0)
1486                         rc = -EINVAL;
1487                 GOTO(out_put, rc);
1488         } else if (rc == buf.lb_len) {
1489                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1490                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1491                        tdtd->tdtd_lut->lut_obd->obd_name,
1492                        tdtd->tdtd_committed_batchid);
1493                 rc = 0;
1494         }
1495
1496 out_put:
1497         if (rc < 0 && dt_obj != NULL) {
1498                 dt_object_put(env, dt_obj);
1499                 tdtd->tdtd_batchid_obj = NULL;
1500         }
1501         return rc;
1502 }
1503
1504 #ifndef TASK_IDLE
1505 #define TASK_IDLE TASK_INTERRUPTIBLE
1506 #endif
1507
1508 /**
1509  * manage the distribute transaction thread
1510  *
1511  * Distribute transaction are linked to the list, and once the distribute
1512  * transaction is committed, it will update the last committed batchid first,
1513  * after it is committed, it will cancel the records.
1514  *
1515  * \param[in] _arg      argument for commit thread
1516  *
1517  * \retval              0 if thread is running successfully
1518  * \retval              negative errno if the thread can not be run.
1519  */
1520 static int distribute_txn_commit_thread(void *_arg)
1521 {
1522         struct target_distribute_txn_data *tdtd = _arg;
1523         struct lu_env           *env = &tdtd->tdtd_env;
1524         LIST_HEAD(list);
1525         int                      rc;
1526         struct top_multiple_thandle *tmt;
1527         struct top_multiple_thandle *tmp;
1528         __u64                    batchid = 0, committed;
1529
1530         ENTRY;
1531
1532
1533         CDEBUG(D_HA, "%s: start commit thread committed batchid %llu\n",
1534                tdtd->tdtd_lut->lut_obd->obd_name,
1535                tdtd->tdtd_committed_batchid);
1536
1537         while (({set_current_state(TASK_IDLE);
1538                  !kthread_should_stop(); })) {
1539                 spin_lock(&tdtd->tdtd_batchid_lock);
1540                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1541                                          tmt_commit_list) {
1542                         if (tmt->tmt_committed == 0)
1543                                 break;
1544
1545                         /* Note: right now, replay is based on master MDT
1546                          * transno, but cancellation is based on batchid.
1547                          * so we do not try to cancel the update log until
1548                          * the recoverying is done, unless the update records
1549                          * batchid < committed_batchid. */
1550                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1551                                 __set_current_state(TASK_RUNNING);
1552                                 list_move_tail(&tmt->tmt_commit_list, &list);
1553                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1554                                 __set_current_state(TASK_RUNNING);
1555                                 LASSERTF(tmt->tmt_batchid >= batchid,
1556                                          "tmt %p tmt_batchid: %llu, batchid "
1557                                           "%llu\n", tmt, tmt->tmt_batchid,
1558                                          batchid);
1559                                 /* There are three types of distribution
1560                                  * transaction result
1561                                  *
1562                                  * 1. If tmt_result < 0, it means the
1563                                  * distribution transaction fails, which should
1564                                  * be rare, because once declare phase succeeds,
1565                                  * the operation should succeeds anyway. Note in
1566                                  * this case, we will still update batchid so
1567                                  * cancellation would be stopped.
1568                                  *
1569                                  * 2. If tmt_result == 0, it means the
1570                                  * distribution transaction succeeds, and we
1571                                  * will update batchid.
1572                                  *
1573                                  * 3. If tmt_result > 0, it means distribute
1574                                  * transaction is not yet committed on every
1575                                  * node, but we need release this tmt before
1576                                  * that, which usuually happens during umount.
1577                                  */
1578                                 if (tmt->tmt_result <= 0)
1579                                         batchid = tmt->tmt_batchid;
1580                                 list_move_tail(&tmt->tmt_commit_list, &list);
1581                         }
1582                 }
1583                 spin_unlock(&tdtd->tdtd_batchid_lock);
1584
1585                 CDEBUG(D_HA, "%s: batchid: %llu committed batchid "
1586                        "%llu\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1587                        tdtd->tdtd_committed_batchid);
1588                 /* update globally committed on a storage */
1589                 if (batchid > tdtd->tdtd_committed_batchid) {
1590                         rc = distribute_txn_commit_batchid_update(env, tdtd,
1591                                                              batchid);
1592                         if (rc == 0)
1593                                 batchid = 0;
1594                 }
1595                 /* cancel the records for committed batchid's */
1596                 /* XXX: should we postpone cancel's till the end of recovery? */
1597                 committed = tdtd->tdtd_committed_batchid;
1598                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1599                         if (tmt->tmt_batchid > committed)
1600                                 break;
1601                         __set_current_state(TASK_RUNNING);
1602                         list_del_init(&tmt->tmt_commit_list);
1603                         if (tmt->tmt_result <= 0)
1604                                 distribute_txn_cancel_records(env, tmt);
1605                         top_multiple_thandle_put(tmt);
1606                 }
1607
1608                 if (current->state)
1609                         schedule();
1610
1611                 if (OBD_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
1612                         set_current_state(TASK_UNINTERRUPTIBLE);
1613                         schedule_timeout(cfs_time_seconds(5));
1614                 }
1615         }
1616
1617         while (({set_current_state(TASK_IDLE);
1618                  atomic_read(&tdtd->tdtd_refcount) != 0; }))
1619                 schedule();
1620         __set_current_state(TASK_RUNNING);
1621
1622         spin_lock(&tdtd->tdtd_batchid_lock);
1623         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1624                                  tmt_commit_list)
1625                 list_move_tail(&tmt->tmt_commit_list, &list);
1626         spin_unlock(&tdtd->tdtd_batchid_lock);
1627
1628         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1629                tdtd->tdtd_lut->lut_obd->obd_name);
1630         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1631                 list_del_init(&tmt->tmt_commit_list);
1632                 top_multiple_thandle_dump(tmt, D_HA);
1633                 top_multiple_thandle_put(tmt);
1634         }
1635         RETURN(0);
1636 }
1637
1638 /**
1639  * Start llog cancel thread
1640  *
1641  * Start llog cancel(master/slave) thread on LOD
1642  *
1643  * \param[in]lclt       cancel log thread to be started.
1644  *
1645  * \retval              0 if the thread is started successfully.
1646  * \retval              negative errno if the thread is not being
1647  *                      started.
1648  */
1649 int distribute_txn_init(const struct lu_env *env,
1650                         struct lu_target *lut,
1651                         struct target_distribute_txn_data *tdtd,
1652                         __u32 index)
1653 {
1654         struct task_struct      *task;
1655         int                     rc;
1656         ENTRY;
1657
1658         INIT_LIST_HEAD(&tdtd->tdtd_list);
1659         INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
1660         INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
1661         spin_lock_init(&tdtd->tdtd_batchid_lock);
1662         spin_lock_init(&tdtd->tdtd_replay_list_lock);
1663         tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
1664         tdtd->tdtd_replay_ready = 0;
1665
1666         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1667
1668         init_waitqueue_head(&tdtd->tdtd_recovery_threads_waitq);
1669         atomic_set(&tdtd->tdtd_refcount, 0);
1670         atomic_set(&tdtd->tdtd_recovery_threads_count, 0);
1671
1672         tdtd->tdtd_lut = lut;
1673         if (lut->lut_bottom->dd_rdonly)
1674                 RETURN(0);
1675
1676         rc = distribute_txn_commit_batchid_init(env, tdtd);
1677         if (rc != 0)
1678                 RETURN(rc);
1679
1680         rc = lu_env_init(&tdtd->tdtd_env, LCT_LOCAL | LCT_MD_THREAD);
1681         if (rc)
1682                 RETURN(rc);
1683
1684         task = kthread_create(distribute_txn_commit_thread, tdtd, "dist_txn-%u",
1685                               index);
1686         if (IS_ERR(task)) {
1687                 lu_env_fini(&tdtd->tdtd_env);
1688                 RETURN(PTR_ERR(task));
1689         }
1690         tdtd->tdtd_commit_task = task;
1691         wake_up_process(task);
1692
1693         RETURN(0);
1694 }
1695 EXPORT_SYMBOL(distribute_txn_init);
1696
1697 /**
1698  * Stop llog cancel thread
1699  *
1700  * Stop llog cancel(master/slave) thread on LOD and also destory
1701  * all of transaction in the list.
1702  *
1703  * \param[in]lclt       cancel log thread to be stopped.
1704  */
1705 void distribute_txn_fini(const struct lu_env *env,
1706                          struct target_distribute_txn_data *tdtd)
1707 {
1708         struct top_multiple_thandle *tmt;
1709         LIST_HEAD(list);
1710
1711         /* Stop cancel thread */
1712         if (!tdtd->tdtd_commit_task)
1713                 return;
1714
1715         kthread_stop(tdtd->tdtd_commit_task);
1716         tdtd->tdtd_commit_task = NULL;
1717
1718         spin_lock(&tdtd->tdtd_batchid_lock);
1719         list_splice_init(&tdtd->tdtd_list, &list);
1720         spin_unlock(&tdtd->tdtd_batchid_lock);
1721
1722         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1723                tdtd->tdtd_lut->lut_obd->obd_name);
1724         while ((tmt = list_first_entry_or_null(&list,
1725                                                struct top_multiple_thandle,
1726                                                tmt_commit_list)) != NULL) {
1727                 list_del_init(&tmt->tmt_commit_list);
1728                 top_multiple_thandle_dump(tmt, D_HA);
1729                 top_multiple_thandle_put(tmt);
1730         }
1731
1732         lu_env_fini(&tdtd->tdtd_env);
1733
1734         dtrq_list_destroy(tdtd);
1735         if (tdtd->tdtd_batchid_obj != NULL) {
1736                 dt_object_put(env, tdtd->tdtd_batchid_obj);
1737                 tdtd->tdtd_batchid_obj = NULL;
1738         }
1739 }
1740 EXPORT_SYMBOL(distribute_txn_fini);