lustre/target/update_trans.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2015, 2017, Intel Corporation.
  24  */
  25 /*
  26  * lustre/target/update_trans.c
  27  *
  28  * This file implements the update distribute transaction API.
  29  *
  30  * To manage the cross-MDT operation (distribute operation) transaction,
  31  * the transaction will also be separated two layers on MD stack, top
  32  * transaction and sub transaction.
  33  *
  34  * During the distribute operation, top transaction is created in the LOD
  35  * layer, and represent the operation. Sub transaction is created by
  36  * each OSD or OSP. Top transaction start/stop will trigger all of its sub
  37  * transaction start/stop. Top transaction (the whole operation) is committed
  38  * only all of its sub transaction are committed.
  39  *
  40  * there are three kinds of transactions
  41  * 1. local transaction: All updates are in a single local OSD.
  42  * 2. Remote transaction: All Updates are only in the remote OSD,
  43  *    i.e. locally all updates are in OSP.
  44  * 3. Mixed transaction: Updates are both in local OSD and remote
  45  *    OSD.
  46  *
  47  * Author: Di Wang <di.wang@intel.com>
  48  */
  49
  50 #define DEBUG_SUBSYSTEM S_CLASS
  51
  52 #include <linux/kthread.h>
  53 #include <lu_target.h>
  54 #include <lustre_log.h>
  55 #include <lustre_update.h>
  56 #include <obd.h>
  57 #include <obd_class.h>
  58 #include <tgt_internal.h>
  59
  60 #include <tgt_internal.h>
  61 /**
  62  * Dump top mulitple thandle
  63  *
  64  * Dump top multiple thandle and all of its sub thandle to the debug log.
  65  *
  66  * \param[in]mask       debug mask
  67  * \param[in]top_th     top_thandle to be dumped
  68  */
  69 static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
  70                                       __u32 mask)
  71 {
  72         struct sub_thandle      *st;
  73
  74         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
  75         CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d batchid %llu\n",
  76                tmt->tmt_master_sub_dt ?
  77                tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
  78                "NULL",
  79                tmt, atomic_read(&tmt->tmt_refcount), tmt->tmt_committed,
  80                tmt->tmt_result, tmt->tmt_batchid);
  81
  82         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
  83                 struct sub_thandle_cookie *stc;
  84
  85                 CDEBUG(mask, "st %p obd %s committed %d stopped %d sub_th %p\n",
  86                        st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
  87                        st->st_committed, st->st_stopped, st->st_sub_th);
  88
  89                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
  90                         CDEBUG(mask, " cookie "DFID".%u\n",
  91                                PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
  92                                stc->stc_cookie.lgc_index);
  93                 }
  94         }
  95 }
  96
  97 /**
  98  * Declare write update to sub device
  99  *
 100  * Declare Write updates llog records to the sub device during distribute
 101  * transaction.
 102  *
 103  * \param[in] env       execution environment
 104  * \param[in] record    update records being written
 105  * \param[in] sub_th    sub transaction handle
 106  * \param[in] record_size total update record size
 107  *
 108  * \retval              0 if writing succeeds
 109  * \retval              negative errno if writing fails
 110  */
 111 static int sub_declare_updates_write(const struct lu_env *env,
 112                                      struct llog_update_record *record,
 113                                      struct thandle *sub_th, size_t record_size)
 114 {
 115         struct llog_ctxt        *ctxt;
 116         struct dt_device        *dt = sub_th->th_dev;
 117         int                     left = record_size;
 118         int rc;
 119
 120         /* If ctxt is NULL, it means not need to write update,
 121          * for example if the the OSP is used to connect to OST */
 122         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 123                                 LLOG_UPDATELOG_ORIG_CTXT);
 124
 125         /* Not ready to record updates yet. */
 126         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 127                 llog_ctxt_put(ctxt);
 128                 return 0;
 129         }
 130
 131         rc = llog_declare_add(env, ctxt->loc_handle,
 132                               &record->lur_hdr, sub_th);
 133         if (rc < 0)
 134                 GOTO(out_put, rc);
 135
 136         while (left > ctxt->loc_chunk_size) {
 137                 rc = llog_declare_add(env, ctxt->loc_handle,
 138                                       &record->lur_hdr, sub_th);
 139                 if (rc < 0)
 140                         GOTO(out_put, rc);
 141
 142                 left -= ctxt->loc_chunk_size;
 143         }
 144
 145 out_put:
 146         llog_ctxt_put(ctxt);
 147
 148         return rc;
 149 }
 150
 151 /**
 152  * write update to sub device
 153  *
 154  * Write llog update record to the sub device during distribute
 155  * transaction. If it succeeds, llog cookie of the record will be
 156  * returned by @cookie.
 157  *
 158  * \param[in] env       execution environment
 159  * \param[in] record    update records being written
 160  * \param[in] sub_th    sub transaction handle
 161  * \param[out] cookie   llog cookie of the update record.
 162  *
 163  * \retval              1 if writing succeeds
 164  * \retval              negative errno if writing fails
 165  */
 166 static int sub_updates_write(const struct lu_env *env,
 167                              struct llog_update_record *record,
 168                              struct sub_thandle *sub_th)
 169 {
 170         struct dt_device *dt = sub_th->st_dt;
 171         struct llog_ctxt *ctxt;
 172         struct llog_update_record *lur = NULL;
 173         __u32 update_count = 0;
 174         __u32 param_count = 0;
 175         __u32 last_update_count = 0;
 176         __u32 last_param_count = 0;
 177         char *start;
 178         char *cur;
 179         char *next;
 180         struct sub_thandle_cookie *stc;
 181         size_t reclen;
 182         bool eof = false;
 183         int rc;
 184         ENTRY;
 185
 186         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 187                                 LLOG_UPDATELOG_ORIG_CTXT);
 188         /* If ctxt == NULL, then it means updates on OST (only happens
 189          * during migration), and we do not track those updates for now */
 190         /* If ctxt->loc_handle == NULL, then it does not need to record
 191          * update, usually happens in error handler path */
 192         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 193                 llog_ctxt_put(ctxt);
 194                 RETURN(0);
 195         }
 196
 197         /* Since the cross-MDT updates will includes both local
 198          * and remote updates, the update ops count must > 1 */
 199         LASSERT(record->lur_update_rec.ur_update_count > 1);
 200         LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
 201                  "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
 202                  llog_update_record_size(record));
 203
 204         /*
 205          * If its size > llog chunk_size, then write current chunk to the update
 206          * llog, NB the padding should >= LLOG_MIN_REC_SIZE.
 207          *
 208          * So check padding length is either >= LLOG_MIN_REC_SIZE or is 0
 209          * (record length just matches the chunk size).
 210          */
 211
 212         reclen = record->lur_hdr.lrh_len;
 213         if (reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 214             reclen == ctxt->loc_chunk_size) {
 215                 OBD_ALLOC_PTR(stc);
 216                 if (stc == NULL)
 217                         GOTO(llog_put, rc = -ENOMEM);
 218                 INIT_LIST_HEAD(&stc->stc_list);
 219
 220                 rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
 221                               &stc->stc_cookie, sub_th->st_sub_th);
 222
 223                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 224                        dt->dd_lu_dev.ld_obd->obd_name,
 225                        PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
 226                        stc->stc_cookie.lgc_index, rc);
 227
 228                 if (rc > 0) {
 229                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 230                         rc = 0;
 231                 } else {
 232                         OBD_FREE_PTR(stc);
 233                 }
 234
 235                 GOTO(llog_put, rc);
 236         }
 237
 238         /* Split the records into chunk_size update record */
 239         OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
 240         if (lur == NULL)
 241                 GOTO(llog_put, rc = -ENOMEM);
 242
 243         memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
 244         lur->lur_update_rec.ur_update_count = 0;
 245         lur->lur_update_rec.ur_param_count = 0;
 246         start = (char *)&record->lur_update_rec.ur_ops;
 247         cur = next = start;
 248         do {
 249                 if (update_count < record->lur_update_rec.ur_update_count)
 250                         next = (char *)update_op_next_op(
 251                                                 (struct update_op *)cur);
 252                 else if (param_count < record->lur_update_rec.ur_param_count)
 253                         next = (char *)update_param_next_param(
 254                                                 (struct update_param *)cur);
 255                 else
 256                         eof = true;
 257
 258                 reclen = __llog_update_record_size(
 259                                 __update_records_size(next - start));
 260                 if ((reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
 261                      reclen == ctxt->loc_chunk_size) &&
 262                     !eof) {
 263                         cur = next;
 264
 265                         if (update_count <
 266                             record->lur_update_rec.ur_update_count)
 267                                 update_count++;
 268                         else if (param_count <
 269                                  record->lur_update_rec.ur_param_count)
 270                                 param_count++;
 271                         continue;
 272                 }
 273
 274                 lur->lur_update_rec.ur_update_count = update_count -
 275                                                       last_update_count;
 276                 lur->lur_update_rec.ur_param_count = param_count -
 277                                                      last_param_count;
 278                 memcpy(&lur->lur_update_rec.ur_ops, start, cur - start);
 279                 lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 280
 281                 LASSERT(lur->lur_hdr.lrh_len ==
 282                          __llog_update_record_size(
 283                                 __update_records_size(cur - start)));
 284                 LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
 285
 286                 update_records_dump(&lur->lur_update_rec, D_INFO, true);
 287
 288                 OBD_ALLOC_PTR(stc);
 289                 if (stc == NULL)
 290                         GOTO(llog_put, rc = -ENOMEM);
 291                 INIT_LIST_HEAD(&stc->stc_list);
 292
 293                 rc = llog_add(env, ctxt->loc_handle, &lur->lur_hdr,
 294                               &stc->stc_cookie, sub_th->st_sub_th);
 295
 296                 CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
 297                         dt->dd_lu_dev.ld_obd->obd_name,
 298                         PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
 299                         stc->stc_cookie.lgc_index, rc);
 300
 301                 if (rc > 0) {
 302                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 303                         rc = 0;
 304                 } else {
 305                         OBD_FREE_PTR(stc);
 306                         GOTO(llog_put, rc);
 307                 }
 308
 309                 last_update_count = update_count;
 310                 last_param_count = param_count;
 311                 start = cur;
 312                 lur->lur_update_rec.ur_update_count = 0;
 313                 lur->lur_update_rec.ur_param_count = 0;
 314                 lur->lur_update_rec.ur_flags |= UPDATE_RECORD_CONTINUE;
 315         } while (!eof);
 316
 317 llog_put:
 318         if (lur != NULL)
 319                 OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
 320         llog_ctxt_put(ctxt);
 321
 322         RETURN(rc);
 323 }
 324
 325 /**
 326  * Prepare the update records.
 327  *
 328  * Merge params and ops into the update records, then initializing
 329  * the update buffer.
 330  *
 331  * During transaction execution phase, parameters and update ops
 332  * are collected in two different buffers (see lod_updates_pack()),
 333  * during transaction stop, it needs to be merged in one buffer,
 334  * so it will be written in the update log.
 335  *
 336  * \param[in] env       execution environment
 337  * \param[in] tmt       top_multiple_thandle for distribute txn
 338  *
 339  * \retval              0 if merging succeeds.
 340  * \retval              negaitive errno if merging fails.
 341  */
 342 static int prepare_writing_updates(const struct lu_env *env,
 343                                    struct top_multiple_thandle *tmt)
 344 {
 345         struct thandle_update_records   *tur = tmt->tmt_update_records;
 346         struct llog_update_record       *lur;
 347         struct update_params *params;
 348         size_t params_size;
 349         size_t update_size;
 350
 351         if (tur == NULL || tur->tur_update_records == NULL ||
 352             tur->tur_update_params == NULL)
 353                 return 0;
 354
 355         lur = tur->tur_update_records;
 356         /* Extends the update records buffer if needed */
 357         params_size = update_params_size(tur->tur_update_params,
 358                                          tur->tur_update_param_count);
 359         LASSERT(lur->lur_update_rec.ur_param_count == 0);
 360         update_size = llog_update_record_size(lur);
 361         if (cfs_size_round(update_size + params_size) >
 362             tur->tur_update_records_buf_size) {
 363                 int rc;
 364
 365                 rc = tur_update_records_extend(tur,
 366                         cfs_size_round(update_size + params_size));
 367                 if (rc < 0)
 368                         return rc;
 369
 370                 lur = tur->tur_update_records;
 371         }
 372
 373         params = update_records_get_params(&lur->lur_update_rec);
 374         memcpy(params, tur->tur_update_params, params_size);
 375
 376         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
 377         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
 378         /* Init update record header */
 379         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 380         lur->lur_hdr.lrh_type = UPDATE_REC;
 381
 382         /* Dump updates for debugging purpose */
 383         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 384
 385         return 0;
 386 }
 387
 388 static inline int
 389 distribute_txn_commit_thread_running(struct lu_target *lut)
 390 {
 391         return lut->lut_tdtd_commit_thread.t_flags & SVC_RUNNING;
 392 }
 393
 394 static inline int
 395 distribute_txn_commit_thread_stopped(struct lu_target *lut)
 396 {
 397         return lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED;
 398 }
 399
 400 /**
 401  * Top thandle commit callback
 402  *
 403  * This callback will be called when all of sub transactions are committed.
 404  *
 405  * \param[in] th        top thandle to be committed.
 406  */
 407 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
 408 {
 409         struct lu_target *lut;
 410         ENTRY;
 411
 412         LASSERT(atomic_read(&tmt->tmt_refcount) > 0);
 413
 414         top_multiple_thandle_dump(tmt, D_HA);
 415         tmt->tmt_committed = 1;
 416         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
 417         if (distribute_txn_commit_thread_running(lut))
 418                 wake_up(&lut->lut_tdtd->tdtd_commit_thread_waitq);
 419         RETURN_EXIT;
 420 }
 421
 422 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
 423                                        struct dt_device *dt_dev)
 424 {
 425         struct sub_thandle *st;
 426
 427         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 428                 if (st->st_dt == dt_dev)
 429                         return st;
 430         }
 431         return NULL;
 432 }
 433 EXPORT_SYMBOL(lookup_sub_thandle);
 434
 435 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
 436                                        struct dt_device *dt_dev)
 437 {
 438         struct sub_thandle *st;
 439
 440         OBD_ALLOC_PTR(st);
 441         if (st == NULL)
 442                 RETURN(ERR_PTR(-ENOMEM));
 443
 444         INIT_LIST_HEAD(&st->st_sub_list);
 445         INIT_LIST_HEAD(&st->st_cookie_list);
 446         st->st_dt = dt_dev;
 447
 448         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
 449         return st;
 450 }
 451
 452 static void sub_trans_commit_cb_internal(struct top_multiple_thandle *tmt,
 453                                          struct thandle *sub_th, int err)
 454 {
 455         struct sub_thandle      *st;
 456         bool                    all_committed = true;
 457
 458         /* Check if all sub thandles are committed */
 459         spin_lock(&tmt->tmt_sub_lock);
 460         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 461                 if (st->st_sub_th == sub_th) {
 462                         st->st_committed = 1;
 463                         st->st_result = err;
 464                 }
 465                 if (!st->st_committed)
 466                         all_committed = false;
 467         }
 468         spin_unlock(&tmt->tmt_sub_lock);
 469
 470         if (tmt->tmt_result == 0)
 471                 tmt->tmt_result = err;
 472
 473         if (all_committed)
 474                 top_trans_committed_cb(tmt);
 475
 476         top_multiple_thandle_dump(tmt, D_INFO);
 477         top_multiple_thandle_put(tmt);
 478         RETURN_EXIT;
 479 }
 480
 481 /**
 482  * sub thandle commit callback
 483  *
 484  * Mark the sub thandle to be committed and if all sub thandle are committed
 485  * notify the top thandle.
 486  *
 487  * \param[in] env       execution environment
 488  * \param[in] sub_th    sub thandle being committed
 489  * \param[in] cb        commit callback
 490  * \param[in] err       trans result
 491  */
 492 static void sub_trans_commit_cb(struct lu_env *env,
 493                                 struct thandle *sub_th,
 494                                 struct dt_txn_commit_cb *cb, int err)
 495 {
 496         struct top_multiple_thandle *tmt = cb->dcb_data;
 497
 498         sub_trans_commit_cb_internal(tmt, sub_th, err);
 499 }
 500
 501 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
 502                                     struct top_multiple_thandle *tmt)
 503 {
 504         LASSERT(st->st_sub_th != NULL);
 505         top_multiple_thandle_get(tmt);
 506         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
 507         st->st_commit_dcb.dcb_data = tmt;
 508         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
 509         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
 510 }
 511
 512 /**
 513  * Sub thandle stop call back
 514  *
 515  * After sub thandle is stopped, it will call this callback to notify
 516  * the top thandle.
 517  *
 518  * \param[in] th        sub thandle to be stopped
 519  * \param[in] rc        result of sub trans
 520  */
 521 static void sub_trans_stop_cb(struct lu_env *env,
 522                               struct thandle *sub_th,
 523                               struct dt_txn_commit_cb *cb, int err)
 524 {
 525         struct sub_thandle              *st;
 526         struct top_multiple_thandle     *tmt = cb->dcb_data;
 527         ENTRY;
 528
 529         spin_lock(&tmt->tmt_sub_lock);
 530         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 531                 if (st->st_stopped)
 532                         continue;
 533
 534                 if (st->st_dt == sub_th->th_dev) {
 535                         st->st_stopped = 1;
 536                         st->st_result = err;
 537                         break;
 538                 }
 539         }
 540         spin_unlock(&tmt->tmt_sub_lock);
 541
 542         wake_up(&tmt->tmt_stop_waitq);
 543         RETURN_EXIT;
 544 }
 545
 546 static void sub_thandle_register_stop_cb(struct sub_thandle *st,
 547                                          struct top_multiple_thandle *tmt)
 548 {
 549         st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
 550         st->st_stop_dcb.dcb_data = tmt;
 551         st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
 552         INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
 553         dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
 554 }
 555
 556 /**
 557  * Create sub thandle
 558  *
 559  * Create transaction handle for sub_thandle
 560  *
 561  * \param[in] env       execution environment
 562  * \param[in] th        top thandle
 563  * \param[in] st        sub_thandle
 564  *
 565  * \retval              0 if creation succeeds.
 566  * \retval              negative errno if creation fails.
 567  */
 568 int sub_thandle_trans_create(const struct lu_env *env,
 569                              struct top_thandle *top_th,
 570                              struct sub_thandle *st)
 571 {
 572         struct thandle *sub_th;
 573
 574         sub_th = dt_trans_create(env, st->st_dt);
 575         if (IS_ERR(sub_th))
 576                 return PTR_ERR(sub_th);
 577
 578         sub_th->th_top = &top_th->tt_super;
 579         st->st_sub_th = sub_th;
 580
 581         sub_th->th_wait_submit = 1;
 582         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
 583         return 0;
 584 }
 585
 586 /**
 587  * Create the top transaction.
 588  *
 589  * Create the top transaction on the master device. It will create a top
 590  * thandle and a sub thandle on the master device.
 591  *
 592  * \param[in] env               execution environment
 593  * \param[in] master_dev        master_dev the top thandle will be created
 594  *
 595  * \retval                      pointer to the created thandle.
 596  * \retval                      ERR_PTR(errno) if creation failed.
 597  */
 598 struct thandle *
 599 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
 600 {
 601         struct top_thandle      *top_th;
 602         struct thandle          *child_th;
 603
 604         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
 605         if (top_th == NULL)
 606                 return ERR_PTR(-ENOMEM);
 607
 608         top_th->tt_super.th_top = &top_th->tt_super;
 609
 610         if (master_dev != NULL) {
 611                 child_th = dt_trans_create(env, master_dev);
 612                 if (IS_ERR(child_th)) {
 613                         OBD_FREE_PTR(top_th);
 614                         return child_th;
 615                 }
 616
 617                 child_th->th_top = &top_th->tt_super;
 618                 child_th->th_wait_submit = 1;
 619                 top_th->tt_master_sub_thandle = child_th;
 620         }
 621         return &top_th->tt_super;
 622 }
 623 EXPORT_SYMBOL(top_trans_create);
 624
 625 /**
 626  * Declare write update transaction
 627  *
 628  * Check if there are updates being recorded in this transaction,
 629  * it will write the record into the disk.
 630  *
 631  * \param[in] env       execution environment
 632  * \param[in] tmt       top multiple transaction handle
 633  *
 634  * \retval              0 if writing succeeds
 635  * \retval              negative errno if writing fails
 636  */
 637 static int declare_updates_write(const struct lu_env *env,
 638                                  struct top_multiple_thandle *tmt)
 639 {
 640         struct llog_update_record *record;
 641         struct sub_thandle *st;
 642         int rc = 0;
 643
 644         record = tmt->tmt_update_records->tur_update_records;
 645         /* Declare update write for all other target */
 646         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 647                 if (st->st_sub_th == NULL)
 648                         continue;
 649
 650                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
 651                                                tmt->tmt_record_size);
 652                 if (rc < 0)
 653                         break;
 654         }
 655
 656         return rc;
 657 }
 658
 659 /**
 660  * Assign batchid to the distribute transaction.
 661  *
 662  * Assign batchid to the distribute transaction
 663  *
 664  * \param[in] tmt       distribute transaction
 665  */
 666 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 667 {
 668         struct target_distribute_txn_data *tdtd;
 669         struct dt_device *dt = new->tmt_master_sub_dt;
 670         struct sub_thandle *st;
 671
 672         LASSERT(dt != NULL);
 673         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 674         spin_lock(&tdtd->tdtd_batchid_lock);
 675         new->tmt_batchid = tdtd->tdtd_batchid++;
 676         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
 677         spin_unlock(&tdtd->tdtd_batchid_lock);
 678         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 679                 if (st->st_sub_th != NULL)
 680                         sub_thandle_register_commit_cb(st, new);
 681         }
 682         top_multiple_thandle_get(new);
 683         top_multiple_thandle_dump(new, D_INFO);
 684 }
 685
 686 /**
 687  * Insert distribute transaction to the distribute txn list.
 688  *
 689  * Insert distribute transaction to the distribute txn list.
 690  *
 691  * \param[in] new       the distribute txn to be inserted.
 692  */
 693 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
 694 {
 695         struct dt_device *dt = new->tmt_master_sub_dt;
 696         struct top_multiple_thandle *tmt;
 697         struct target_distribute_txn_data *tdtd;
 698         struct sub_thandle *st;
 699         bool    at_head = false;
 700
 701         LASSERT(dt != NULL);
 702         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 703
 704         spin_lock(&tdtd->tdtd_batchid_lock);
 705         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
 706                 if (new->tmt_batchid > tmt->tmt_batchid) {
 707                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
 708                         break;
 709                 }
 710         }
 711         if (list_empty(&new->tmt_commit_list)) {
 712                 at_head = true;
 713                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
 714         }
 715         spin_unlock(&tdtd->tdtd_batchid_lock);
 716
 717         list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
 718                 if (st->st_sub_th != NULL)
 719                         sub_thandle_register_commit_cb(st, new);
 720         }
 721
 722         top_multiple_thandle_get(new);
 723         top_multiple_thandle_dump(new, D_INFO);
 724         if (new->tmt_committed && at_head)
 725                 wake_up(&tdtd->tdtd_commit_thread_waitq);
 726 }
 727
 728 /**
 729  * Prepare cross-MDT operation.
 730  *
 731  * Create the update record buffer to record updates for cross-MDT operation,
 732  * add master sub transaction to tt_sub_trans_list, and declare the update
 733  * writes.
 734  *
 735  * During updates packing, all of parameters will be packed in
 736  * tur_update_params, and updates will be packed in tur_update_records.
 737  * Then in transaction stop, parameters and updates will be merged
 738  * into one updates buffer.
 739  *
 740  * And also master thandle will be added to the sub_th list, so it will be
 741  * easy to track the commit status.
 742  *
 743  * \param[in] env       execution environment
 744  * \param[in] th        top transaction handle
 745  *
 746  * \retval              0 if preparation succeeds.
 747  * \retval              negative errno if preparation fails.
 748  */
 749 static int prepare_multiple_node_trans(const struct lu_env *env,
 750                                        struct top_multiple_thandle *tmt)
 751 {
 752         struct thandle_update_records   *tur;
 753         int                             rc;
 754         ENTRY;
 755
 756         if (tmt->tmt_update_records == NULL) {
 757                 tur = &update_env_info(env)->uti_tur;
 758                 rc = check_and_prepare_update_record(env, tur);
 759                 if (rc < 0)
 760                         RETURN(rc);
 761
 762                 tmt->tmt_update_records = tur;
 763                 distribute_txn_assign_batchid(tmt);
 764         }
 765
 766         rc = declare_updates_write(env, tmt);
 767
 768         RETURN(rc);
 769 }
 770
 771 /**
 772  * start the top transaction.
 773  *
 774  * Start all of its sub transactions, then start master sub transaction.
 775  *
 776  * \param[in] env               execution environment
 777  * \param[in] master_dev        master_dev the top thandle will be start
 778  * \param[in] th                top thandle
 779  *
 780  * \retval                      0 if transaction start succeeds.
 781  * \retval                      negative errno if start fails.
 782  */
 783 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
 784                     struct thandle *th)
 785 {
 786         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 787                                                        tt_super);
 788         struct sub_thandle              *st;
 789         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
 790         int                             rc = 0;
 791         ENTRY;
 792
 793         if (tmt == NULL) {
 794                 if (th->th_sync)
 795                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 796                 if (th->th_local)
 797                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 798                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
 799                                     top_th->tt_master_sub_thandle);
 800                 RETURN(rc);
 801         }
 802
 803         tmt = top_th->tt_multiple_thandle;
 804         rc = prepare_multiple_node_trans(env, tmt);
 805         if (rc < 0)
 806                 RETURN(rc);
 807
 808         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 809                 if (st->st_sub_th == NULL)
 810                         continue;
 811                 if (th->th_sync)
 812                         st->st_sub_th->th_sync = th->th_sync;
 813                 if (th->th_local)
 814                         st->st_sub_th->th_local = th->th_local;
 815                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
 816                                     st->st_sub_th);
 817                 if (rc != 0)
 818                         GOTO(out, rc);
 819
 820                 LASSERT(st->st_started == 0);
 821                 st->st_started = 1;
 822         }
 823 out:
 824         th->th_result = rc;
 825         RETURN(rc);
 826 }
 827 EXPORT_SYMBOL(top_trans_start);
 828
 829 /**
 830  * Check whether we need write updates record
 831  *
 832  * Check if the updates for the top_thandle needs to be writen
 833  * to all targets. Only if the transaction succeeds and the updates
 834  * number > 2, it will write the updates,
 835  *
 836  * \params [in] top_th  top thandle.
 837  *
 838  * \retval              true if it needs to write updates
 839  * \retval              false if it does not need to write updates
 840  **/
 841 static bool top_check_write_updates(struct top_thandle *top_th)
 842 {
 843         struct top_multiple_thandle     *tmt;
 844         struct thandle_update_records   *tur;
 845
 846         /* Do not write updates to records if the transaction fails */
 847         if (top_th->tt_super.th_result != 0)
 848                 return false;
 849
 850         tmt = top_th->tt_multiple_thandle;
 851         if (tmt == NULL)
 852                 return false;
 853
 854         tur = tmt->tmt_update_records;
 855         if (tur == NULL)
 856                 return false;
 857
 858         /* Hmm, false update records, since the cross-MDT operation
 859          * should includes both local and remote updates, so the
 860          * updates count should >= 2 */
 861         if (tur->tur_update_records == NULL ||
 862             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
 863                 return false;
 864
 865         return true;
 866 }
 867
 868 /**
 869  * Check if top transaction is stopped
 870  *
 871  * Check if top transaction is stopped, only if all sub transaction
 872  * is stopped, then the top transaction is stopped.
 873  *
 874  * \param [in] top_th   top thandle
 875  *
 876  * \retval              true if the top transaction is stopped.
 877  * \retval              false if the top transaction is not stopped.
 878  */
 879 static bool top_trans_is_stopped(struct top_thandle *top_th)
 880 {
 881         struct top_multiple_thandle     *tmt;
 882         struct sub_thandle              *st;
 883         bool                    all_stopped = true;
 884
 885         tmt = top_th->tt_multiple_thandle;
 886         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 887                 if (!st->st_stopped && st->st_sub_th != NULL) {
 888                         all_stopped = false;
 889                         break;
 890                 }
 891
 892                 if (st->st_result != 0 &&
 893                     top_th->tt_super.th_result == 0)
 894                         top_th->tt_super.th_result = st->st_result;
 895         }
 896
 897         return all_stopped;
 898 }
 899
 900 /**
 901  * Wait result of top transaction
 902  *
 903  * Wait until all sub transaction get its result.
 904  *
 905  * \param [in] top_th   top thandle.
 906  *
 907  * \retval              the result of top thandle.
 908  */
 909 static int top_trans_wait_result(struct top_thandle *top_th)
 910 {
 911         struct l_wait_info      lwi = {0};
 912
 913         l_wait_event(top_th->tt_multiple_thandle->tmt_stop_waitq,
 914                      top_trans_is_stopped(top_th), &lwi);
 915
 916         RETURN(top_th->tt_super.th_result);
 917 }
 918
 919 /**
 920  * Stop the top transaction.
 921  *
 922  * Stop the transaction on the master device first, then stop transactions
 923  * on other sub devices.
 924  *
 925  * \param[in] env               execution environment
 926  * \param[in] master_dev        master_dev the top thandle will be created
 927  * \param[in] th                top thandle
 928  *
 929  * \retval                      0 if stop transaction succeeds.
 930  * \retval                      negative errno if stop transaction fails.
 931  */
 932 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
 933                    struct thandle *th)
 934 {
 935         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 936                                                        tt_super);
 937         struct sub_thandle              *st;
 938         struct sub_thandle              *master_st;
 939         struct top_multiple_thandle     *tmt;
 940         struct thandle_update_records   *tur;
 941         bool                            write_updates = false;
 942         int                     rc = 0;
 943         ENTRY;
 944
 945         if (likely(top_th->tt_multiple_thandle == NULL)) {
 946                 LASSERT(master_dev != NULL);
 947
 948                 if (th->th_sync)
 949                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 950                 if (th->th_local)
 951                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 952                 rc = dt_trans_stop(env, master_dev,
 953                                    top_th->tt_master_sub_thandle);
 954                 OBD_FREE_PTR(top_th);
 955                 RETURN(rc);
 956         }
 957
 958         tmt = top_th->tt_multiple_thandle;
 959         tur = tmt->tmt_update_records;
 960
 961         /* Note: we need stop the master thandle first, then the stop
 962          * callback will fill the master transno in the update logs,
 963          * then these update logs will be sent to other MDTs */
 964         /* get the master sub thandle */
 965         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
 966         write_updates = top_check_write_updates(top_th);
 967
 968         /* Step 1: write the updates log on Master MDT */
 969         if (master_st != NULL && master_st->st_sub_th != NULL &&
 970             write_updates) {
 971                 struct llog_update_record *lur;
 972
 973                 /* Merge the parameters and updates into one buffer */
 974                 rc = prepare_writing_updates(env, tmt);
 975                 if (rc < 0) {
 976                         CERROR("%s: cannot prepare updates: rc = %d\n",
 977                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 978                         th->th_result = rc;
 979                         write_updates = false;
 980                         GOTO(stop_master_trans, rc);
 981                 }
 982
 983                 lur = tur->tur_update_records;
 984                 /* Write updates to the master MDT */
 985                 rc = sub_updates_write(env, lur, master_st);
 986
 987                 /* Cleanup the common parameters in the update records,
 988                  * master transno callback might add more parameters.
 989                  * and we need merge the update records again in the
 990                  * following */
 991                 if (tur->tur_update_params != NULL)
 992                         lur->lur_update_rec.ur_param_count = 0;
 993
 994                 if (rc < 0) {
 995                         CERROR("%s: write updates failed: rc = %d\n",
 996                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 997                         th->th_result = rc;
 998                         write_updates = false;
 999                         GOTO(stop_master_trans, rc);
1000                 }
1001         }
1002
1003 stop_master_trans:
1004         /* Step 2: Stop the transaction on the master MDT, and fill the
1005          * master transno in the update logs to other MDT. */
1006         if (master_st != NULL && master_st->st_sub_th != NULL) {
1007                 if (th->th_local)
1008                         master_st->st_sub_th->th_local = th->th_local;
1009                 if (th->th_sync)
1010                         master_st->st_sub_th->th_sync = th->th_sync;
1011                 master_st->st_sub_th->th_result = th->th_result;
1012                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
1013                 /* If it does not write_updates, then we call submit callback
1014                  * here, otherwise callback is done through
1015                  * osd(osp)_trans_commit_cb() */
1016                 if (!master_st->st_started &&
1017                     !list_empty(&tmt->tmt_commit_list))
1018                         sub_trans_commit_cb_internal(tmt,
1019                                                 master_st->st_sub_th, rc);
1020                 if (rc < 0) {
1021                         th->th_result = rc;
1022                         GOTO(stop_other_trans, rc);
1023                 } else if (tur != NULL && tur->tur_update_records != NULL) {
1024                         struct llog_update_record *lur;
1025
1026                         lur = tur->tur_update_records;
1027                         if (lur->lur_update_rec.ur_master_transno == 0)
1028                                 /* Update master transno after master stop
1029                                  * callback */
1030                                 lur->lur_update_rec.ur_master_transno =
1031                                                 tgt_th_info(env)->tti_transno;
1032                 }
1033         }
1034
1035         /* Step 3: write updates to other MDTs */
1036         if (write_updates) {
1037                 struct llog_update_record *lur;
1038
1039                 /* Stop callback of master will add more updates and also update
1040                  * master transno, so merge the parameters and updates into one
1041                  * buffer again */
1042                 rc = prepare_writing_updates(env, tmt);
1043                 if (rc < 0) {
1044                         CERROR("%s: prepare updates failed: rc = %d\n",
1045                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1046                         th->th_result = rc;
1047                         GOTO(stop_other_trans, rc);
1048                 }
1049                 lur = tur->tur_update_records;
1050                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1051                                     st_sub_list) {
1052                         if (st->st_sub_th == NULL || st == master_st ||
1053                             st->st_sub_th->th_result < 0)
1054                                 continue;
1055
1056                         rc = sub_updates_write(env, lur, st);
1057                         if (rc < 0) {
1058                                 th->th_result = rc;
1059                                 break;
1060                         }
1061                 }
1062         }
1063
1064 stop_other_trans:
1065         /* Step 4: Stop the transaction on other MDTs */
1066         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1067                 if (st == master_st || st->st_sub_th == NULL)
1068                         continue;
1069
1070                 if (th->th_sync)
1071                         st->st_sub_th->th_sync = th->th_sync;
1072                 if (th->th_local)
1073                         st->st_sub_th->th_local = th->th_local;
1074                 st->st_sub_th->th_result = th->th_result;
1075                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1076                                    st->st_sub_th);
1077                 if (unlikely(rc < 0 && th->th_result == 0))
1078                         th->th_result = rc;
1079         }
1080
1081         rc = top_trans_wait_result(top_th);
1082
1083         tmt->tmt_result = rc;
1084
1085         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1086          * multiple node transaction, the top transaction will be destroyed. */
1087         top_multiple_thandle_put(tmt);
1088         OBD_FREE_PTR(top_th);
1089         RETURN(rc);
1090 }
1091 EXPORT_SYMBOL(top_trans_stop);
1092
1093 /**
1094  * Create top_multiple_thandle for top_thandle
1095  *
1096  * Create top_mutilple_thandle to manage the mutiple node transaction
1097  * for top_thandle, and it also needs to add master sub thandle to the
1098  * sub trans list now.
1099  *
1100  * \param[in] env       execution environment
1101  * \param[in] top_th    the top thandle
1102  *
1103  * \retval      0 if creation succeeds
1104  * \retval      negative errno if creation fails
1105  */
1106 int top_trans_create_tmt(const struct lu_env *env,
1107                          struct top_thandle *top_th)
1108 {
1109         struct top_multiple_thandle *tmt;
1110
1111         OBD_ALLOC_PTR(tmt);
1112         if (tmt == NULL)
1113                 return -ENOMEM;
1114
1115         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1116         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1117         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1118         atomic_set(&tmt->tmt_refcount, 1);
1119         spin_lock_init(&tmt->tmt_sub_lock);
1120         init_waitqueue_head(&tmt->tmt_stop_waitq);
1121
1122         top_th->tt_multiple_thandle = tmt;
1123
1124         return 0;
1125 }
1126
1127 static struct sub_thandle *
1128 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1129                                 struct thandle *sub_th)
1130 {
1131         struct sub_thandle *st;
1132
1133         /* create and init sub th to the top trans list */
1134         st = create_sub_thandle(top_th->tt_multiple_thandle,
1135                                 sub_th->th_dev);
1136         if (IS_ERR(st))
1137                 return st;
1138
1139         st->st_sub_th = sub_th;
1140
1141         sub_th->th_top = &top_th->tt_super;
1142         sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
1143         return st;
1144 }
1145
1146 /**
1147  * Get sub thandle.
1148  *
1149  * Get sub thandle from the top thandle according to the sub dt_device.
1150  *
1151  * \param[in] env       execution environment
1152  * \param[in] th        thandle on the top layer.
1153  * \param[in] sub_dt    sub dt_device used to get sub transaction
1154  *
1155  * \retval              thandle of sub transaction if succeed
1156  * \retval              PTR_ERR(errno) if failed
1157  */
1158 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1159                                       struct thandle *th,
1160                                       struct dt_device *sub_dt)
1161 {
1162         struct sub_thandle      *st = NULL;
1163         struct sub_thandle      *master_st = NULL;
1164         struct top_thandle      *top_th;
1165         struct thandle          *sub_th = NULL;
1166         int                     rc = 0;
1167         ENTRY;
1168
1169         top_th = container_of(th, struct top_thandle, tt_super);
1170
1171         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1172                 RETURN(top_th->tt_master_sub_thandle);
1173
1174         if (top_th->tt_multiple_thandle != NULL) {
1175                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1176                 if (st != NULL)
1177                         RETURN(st->st_sub_th);
1178         }
1179
1180         sub_th = dt_trans_create(env, sub_dt);
1181         if (IS_ERR(sub_th))
1182                 RETURN(sub_th);
1183
1184         /* Create top_multiple_thandle if necessary */
1185         if (top_th->tt_multiple_thandle == NULL) {
1186                 struct top_multiple_thandle *tmt;
1187
1188                 rc = top_trans_create_tmt(env, top_th);
1189                 if (rc < 0)
1190                         GOTO(stop_trans, rc);
1191
1192                 tmt = top_th->tt_multiple_thandle;
1193
1194                 /* Add master sub th to the top trans list */
1195                 tmt->tmt_master_sub_dt =
1196                         top_th->tt_master_sub_thandle->th_dev;
1197                 master_st = create_sub_thandle_with_thandle(top_th,
1198                                         top_th->tt_master_sub_thandle);
1199                 if (IS_ERR(master_st)) {
1200                         rc = PTR_ERR(master_st);
1201                         master_st = NULL;
1202                         GOTO(stop_trans, rc);
1203                 }
1204         }
1205
1206         /* create and init sub th to the top trans list */
1207         st = create_sub_thandle_with_thandle(top_th, sub_th);
1208         if (IS_ERR(st)) {
1209                 rc = PTR_ERR(st);
1210                 st = NULL;
1211                 GOTO(stop_trans, rc);
1212         }
1213         st->st_sub_th->th_wait_submit = 1;
1214 stop_trans:
1215         if (rc < 0) {
1216                 if (master_st != NULL) {
1217                         list_del(&master_st->st_sub_list);
1218                         OBD_FREE_PTR(master_st);
1219                 }
1220                 sub_th->th_result = rc;
1221                 dt_trans_stop(env, sub_dt, sub_th);
1222                 sub_th = ERR_PTR(rc);
1223         }
1224
1225         RETURN(sub_th);
1226 }
1227 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1228
1229 /**
1230  * Top multiple thandle destroy
1231  *
1232  * Destroy multiple thandle and all its sub thandle.
1233  *
1234  * \param[in] tmt       top_multiple_thandle to be destroyed.
1235  */
1236 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1237 {
1238         struct sub_thandle *st;
1239         struct sub_thandle *tmp;
1240
1241         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1242         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1243                                  st_sub_list) {
1244                 struct sub_thandle_cookie *stc;
1245                 struct sub_thandle_cookie *tmp;
1246
1247                 list_del(&st->st_sub_list);
1248                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1249                                          stc_list) {
1250                         list_del(&stc->stc_list);
1251                         OBD_FREE_PTR(stc);
1252                 }
1253                 OBD_FREE_PTR(st);
1254         }
1255         OBD_FREE_PTR(tmt);
1256 }
1257 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1258
1259 /**
1260  * Cancel the update log on MDTs
1261  *
1262  * Cancel the update log on MDTs then destroy the thandle.
1263  *
1264  * \param[in] env       execution environment
1265  * \param[in] tmt       the top multiple thandle whose updates records
1266  *                      will be cancelled.
1267  *
1268  * \retval              0 if cancellation succeeds.
1269  * \retval              negative errno if cancellation fails.
1270  */
1271 static int distribute_txn_cancel_records(const struct lu_env *env,
1272                                          struct top_multiple_thandle *tmt)
1273 {
1274         struct sub_thandle *st;
1275         ENTRY;
1276
1277         top_multiple_thandle_dump(tmt, D_INFO);
1278         /* Cancel update logs on other MDTs */
1279         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1280                 struct llog_ctxt        *ctxt;
1281                 struct obd_device       *obd;
1282                 struct llog_cookie      *cookie;
1283                 struct sub_thandle_cookie *stc;
1284                 int rc;
1285
1286                 obd = st->st_dt->dd_lu_dev.ld_obd;
1287                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1288                 if (ctxt == NULL)
1289                         continue;
1290                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1291                         cookie = &stc->stc_cookie;
1292                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1293                                 continue;
1294
1295                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1296                                                      cookie);
1297                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1298                                DFID".%u: rc = %d\n", obd->obd_name,
1299                                tmt->tmt_batchid,
1300                                PFID(&cookie->lgc_lgl.lgl_oi.oi_fid),
1301                                cookie->lgc_index, rc);
1302                 }
1303
1304                 llog_ctxt_put(ctxt);
1305         }
1306
1307         RETURN(0);
1308 }
1309
1310 /**
1311  * Check if there are committed transaction
1312  *
1313  * Check if there are committed transaction in the distribute transaction
1314  * list, then cancel the update records for those committed transaction.
1315  * Because the distribute transaction in the list are sorted by batchid,
1316  * and cancellation will be done by batchid order, so we only check the first
1317  * the transaction(with lowest batchid) in the list.
1318  *
1319  * \param[in] lod       lod device where cancel thread is
1320  *
1321  * \retval              true if it is ready
1322  * \retval              false if it is not ready
1323  */
1324 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1325 {
1326         struct top_multiple_thandle     *tmt = NULL;
1327         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1328         bool    ready = false;
1329
1330         spin_lock(&tdtd->tdtd_batchid_lock);
1331         if (!list_empty(&tdtd->tdtd_list)) {
1332                 tmt = list_entry(tdtd->tdtd_list.next,
1333                                  struct top_multiple_thandle, tmt_commit_list);
1334                 if (tmt->tmt_committed &&
1335                     (!obd->obd_recovering || (obd->obd_recovering &&
1336                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1337                         ready = true;
1338         }
1339         spin_unlock(&tdtd->tdtd_batchid_lock);
1340
1341         return ready;
1342 }
1343
1344 struct distribute_txn_bid_data {
1345         struct dt_txn_commit_cb  dtbd_cb;
1346         struct target_distribute_txn_data      *dtbd_tdtd;
1347         __u64                    dtbd_batchid;
1348 };
1349
1350 /**
1351  * callback of updating commit batchid
1352  *
1353  * Updating commit batchid then wake up the commit thread to cancel the
1354  * records.
1355  *
1356  * \param[in]env        execution environment
1357  * \param[in]th         thandle to updating commit batchid
1358  * \param[in]cb         commit callback
1359  * \param[in]err        result of thandle
1360  */
1361 static void distribute_txn_batchid_cb(struct lu_env *env,
1362                                       struct thandle *th,
1363                                       struct dt_txn_commit_cb *cb,
1364                                       int err)
1365 {
1366         struct distribute_txn_bid_data          *dtbd = NULL;
1367         struct target_distribute_txn_data       *tdtd;
1368
1369         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1370         tdtd = dtbd->dtbd_tdtd;
1371
1372         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1373               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1374         spin_lock(&tdtd->tdtd_batchid_lock);
1375         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1376             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1377                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1378         spin_unlock(&tdtd->tdtd_batchid_lock);
1379         atomic_dec(&tdtd->tdtd_refcount);
1380         wake_up(&tdtd->tdtd_commit_thread_waitq);
1381
1382         OBD_FREE_PTR(dtbd);
1383 }
1384
1385 /**
1386  * Update the commit batchid in disk
1387  *
1388  * Update commit batchid in the disk, after this is committed, it can start
1389  * to cancel the update records.
1390  *
1391  * \param[in] env       execution environment
1392  * \param[in] tdtd      distribute transaction structure
1393  * \param[in] batchid   commit batchid to be updated
1394  *
1395  * \retval              0 if update succeeds.
1396  * \retval              negative errno if update fails.
1397  */
1398 static int
1399 distribute_txn_commit_batchid_update(const struct lu_env *env,
1400                               struct target_distribute_txn_data *tdtd,
1401                               __u64 batchid)
1402 {
1403         struct distribute_txn_bid_data  *dtbd = NULL;
1404         struct thandle          *th;
1405         struct lu_buf            buf;
1406         __u64                    tmp;
1407         __u64                    off;
1408         int                      rc;
1409         ENTRY;
1410
1411         OBD_ALLOC_PTR(dtbd);
1412         if (dtbd == NULL)
1413                 RETURN(-ENOMEM);
1414         dtbd->dtbd_batchid = batchid;
1415         dtbd->dtbd_tdtd = tdtd;
1416         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1417         atomic_inc(&tdtd->tdtd_refcount);
1418
1419         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1420         if (IS_ERR(th)) {
1421                 atomic_dec(&tdtd->tdtd_refcount);
1422                 OBD_FREE_PTR(dtbd);
1423                 RETURN(PTR_ERR(th));
1424         }
1425
1426         tmp = cpu_to_le64(batchid);
1427         buf.lb_buf = &tmp;
1428         buf.lb_len = sizeof(tmp);
1429         off = 0;
1430
1431         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1432                                      th);
1433         if (rc < 0)
1434                 GOTO(stop, rc);
1435
1436         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1437         if (rc < 0)
1438                 GOTO(stop, rc);
1439
1440         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1441         if (rc < 0)
1442                 GOTO(stop, rc);
1443
1444         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1445                              &off, th);
1446
1447         CDEBUG(D_INFO, "%s: update batchid %llu: rc = %d\n",
1448                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1449
1450 stop:
1451         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1452         if (rc < 0) {
1453                 atomic_dec(&tdtd->tdtd_refcount);
1454                 OBD_FREE_PTR(dtbd);
1455         }
1456         RETURN(rc);
1457 }
1458
1459 /**
1460  * Init commit batchid for distribute transaction.
1461  *
1462  * Initialize the batchid object and get commit batchid from the object.
1463  *
1464  * \param[in] env       execution environment
1465  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1466  *
1467  * \retval              0 if initialization succeeds.
1468  * \retval              negative errno if initialization fails.
1469  **/
1470 static int
1471 distribute_txn_commit_batchid_init(const struct lu_env *env,
1472                                    struct target_distribute_txn_data *tdtd)
1473 {
1474         struct tgt_thread_info  *tti = tgt_th_info(env);
1475         struct lu_target        *lut = tdtd->tdtd_lut;
1476         struct lu_attr          *attr = &tti->tti_attr;
1477         struct lu_fid           *fid = &tti->tti_fid1;
1478         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1479         struct dt_object        *dt_obj = NULL;
1480         struct lu_buf           buf;
1481         __u64                   tmp;
1482         __u64                   off;
1483         int                     rc;
1484         ENTRY;
1485
1486         memset(attr, 0, sizeof(*attr));
1487         attr->la_valid = LA_MODE;
1488         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1489         dof->dof_type = dt_mode_to_dft(S_IFREG);
1490
1491         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1492
1493         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1494                                    attr);
1495         if (IS_ERR(dt_obj)) {
1496                 rc = PTR_ERR(dt_obj);
1497                 dt_obj = NULL;
1498                 GOTO(out_put, rc);
1499         }
1500
1501         tdtd->tdtd_batchid_obj = dt_obj;
1502
1503         buf.lb_buf = &tmp;
1504         buf.lb_len = sizeof(tmp);
1505         off = 0;
1506         rc = dt_read(env, dt_obj, &buf, &off);
1507         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1508                 CERROR("%s can't read last committed batchid: rc = %d\n",
1509                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1510                 if (rc > 0)
1511                         rc = -EINVAL;
1512                 GOTO(out_put, rc);
1513         } else if (rc == buf.lb_len) {
1514                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1515                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1516                        tdtd->tdtd_lut->lut_obd->obd_name,
1517                        tdtd->tdtd_committed_batchid);
1518                 rc = 0;
1519         }
1520
1521 out_put:
1522         if (rc < 0 && dt_obj != NULL) {
1523                 dt_object_put(env, dt_obj);
1524                 tdtd->tdtd_batchid_obj = NULL;
1525         }
1526         return rc;
1527 }
1528
1529 /**
1530  * manage the distribute transaction thread
1531  *
1532  * Distribute transaction are linked to the list, and once the distribute
1533  * transaction is committed, it will update the last committed batchid first,
1534  * after it is committed, it will cancel the records.
1535  *
1536  * \param[in] _arg      argument for commit thread
1537  *
1538  * \retval              0 if thread is running successfully
1539  * \retval              negative errno if the thread can not be run.
1540  */
1541 static int distribute_txn_commit_thread(void *_arg)
1542 {
1543         struct target_distribute_txn_data *tdtd = _arg;
1544         struct lu_target        *lut = tdtd->tdtd_lut;
1545         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1546         struct l_wait_info       lwi = { 0 };
1547         struct lu_env            env;
1548         struct list_head         list;
1549         int                      rc;
1550         struct top_multiple_thandle *tmt;
1551         struct top_multiple_thandle *tmp;
1552         __u64                    batchid = 0, committed;
1553
1554         ENTRY;
1555
1556         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1557         if (rc != 0)
1558                 RETURN(rc);
1559
1560         spin_lock(&tdtd->tdtd_batchid_lock);
1561         thread->t_flags = SVC_RUNNING;
1562         spin_unlock(&tdtd->tdtd_batchid_lock);
1563         wake_up(&thread->t_ctl_waitq);
1564         INIT_LIST_HEAD(&list);
1565
1566         CDEBUG(D_HA, "%s: start commit thread committed batchid %llu\n",
1567                tdtd->tdtd_lut->lut_obd->obd_name,
1568                tdtd->tdtd_committed_batchid);
1569
1570         while (distribute_txn_commit_thread_running(lut)) {
1571                 spin_lock(&tdtd->tdtd_batchid_lock);
1572                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1573                                          tmt_commit_list) {
1574                         if (tmt->tmt_committed == 0)
1575                                 break;
1576
1577                         /* Note: right now, replay is based on master MDT
1578                          * transno, but cancellation is based on batchid.
1579                          * so we do not try to cancel the update log until
1580                          * the recoverying is done, unless the update records
1581                          * batchid < committed_batchid. */
1582                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1583                                 list_move_tail(&tmt->tmt_commit_list, &list);
1584                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1585                                 LASSERTF(tmt->tmt_batchid >= batchid,
1586                                          "tmt %p tmt_batchid: %llu, batchid "
1587                                           "%llu\n", tmt, tmt->tmt_batchid,
1588                                          batchid);
1589                                 /* There are three types of distribution
1590                                  * transaction result
1591                                  *
1592                                  * 1. If tmt_result < 0, it means the
1593                                  * distribution transaction fails, which should
1594                                  * be rare, because once declare phase succeeds,
1595                                  * the operation should succeeds anyway. Note in
1596                                  * this case, we will still update batchid so
1597                                  * cancellation would be stopped.
1598                                  *
1599                                  * 2. If tmt_result == 0, it means the
1600                                  * distribution transaction succeeds, and we
1601                                  * will update batchid.
1602                                  *
1603                                  * 3. If tmt_result > 0, it means distribute
1604                                  * transaction is not yet committed on every
1605                                  * node, but we need release this tmt before
1606                                  * that, which usuually happens during umount.
1607                                  */
1608                                 if (tmt->tmt_result <= 0)
1609                                         batchid = tmt->tmt_batchid;
1610                                 list_move_tail(&tmt->tmt_commit_list, &list);
1611                         }
1612                 }
1613                 spin_unlock(&tdtd->tdtd_batchid_lock);
1614
1615                 CDEBUG(D_HA, "%s: batchid: %llu committed batchid "
1616                        "%llu\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1617                        tdtd->tdtd_committed_batchid);
1618                 /* update globally committed on a storage */
1619                 if (batchid > tdtd->tdtd_committed_batchid) {
1620                         rc = distribute_txn_commit_batchid_update(&env, tdtd,
1621                                                              batchid);
1622                         if (rc == 0)
1623                                 batchid = 0;
1624                 }
1625                 /* cancel the records for committed batchid's */
1626                 /* XXX: should we postpone cancel's till the end of recovery? */
1627                 committed = tdtd->tdtd_committed_batchid;
1628                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1629                         if (tmt->tmt_batchid > committed)
1630                                 break;
1631                         list_del_init(&tmt->tmt_commit_list);
1632                         if (tmt->tmt_result <= 0)
1633                                 distribute_txn_cancel_records(&env, tmt);
1634                         top_multiple_thandle_put(tmt);
1635                 }
1636
1637                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1638                              !distribute_txn_commit_thread_running(lut) ||
1639                              committed < tdtd->tdtd_committed_batchid ||
1640                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1641         };
1642
1643         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1644                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1645
1646         spin_lock(&tdtd->tdtd_batchid_lock);
1647         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1648                                  tmt_commit_list)
1649                 list_move_tail(&tmt->tmt_commit_list, &list);
1650         spin_unlock(&tdtd->tdtd_batchid_lock);
1651
1652         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1653                tdtd->tdtd_lut->lut_obd->obd_name);
1654         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1655                 list_del_init(&tmt->tmt_commit_list);
1656                 top_multiple_thandle_dump(tmt, D_HA);
1657                 top_multiple_thandle_put(tmt);
1658         }
1659
1660         thread->t_flags = SVC_STOPPED;
1661         lu_env_fini(&env);
1662         wake_up(&thread->t_ctl_waitq);
1663
1664         RETURN(0);
1665 }
1666
1667 /**
1668  * Start llog cancel thread
1669  *
1670  * Start llog cancel(master/slave) thread on LOD
1671  *
1672  * \param[in]lclt       cancel log thread to be started.
1673  *
1674  * \retval              0 if the thread is started successfully.
1675  * \retval              negative errno if the thread is not being
1676  *                      started.
1677  */
1678 int distribute_txn_init(const struct lu_env *env,
1679                         struct lu_target *lut,
1680                         struct target_distribute_txn_data *tdtd,
1681                         __u32 index)
1682 {
1683         struct task_struct      *task;
1684         struct l_wait_info       lwi = { 0 };
1685         int                     rc;
1686         ENTRY;
1687
1688         INIT_LIST_HEAD(&tdtd->tdtd_list);
1689         INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
1690         INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
1691         spin_lock_init(&tdtd->tdtd_batchid_lock);
1692         spin_lock_init(&tdtd->tdtd_replay_list_lock);
1693         tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
1694         tdtd->tdtd_replay_ready = 0;
1695
1696         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1697
1698         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1699         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1700         init_waitqueue_head(&tdtd->tdtd_recovery_threads_waitq);
1701         atomic_set(&tdtd->tdtd_refcount, 0);
1702         atomic_set(&tdtd->tdtd_recovery_threads_count, 0);
1703
1704         tdtd->tdtd_lut = lut;
1705         if (lut->lut_bottom->dd_rdonly)
1706                 RETURN(0);
1707
1708         rc = distribute_txn_commit_batchid_init(env, tdtd);
1709         if (rc != 0)
1710                 RETURN(rc);
1711
1712         task = kthread_run(distribute_txn_commit_thread, tdtd, "dist_txn-%u",
1713                            index);
1714         if (IS_ERR(task))
1715                 RETURN(PTR_ERR(task));
1716
1717         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1718                      distribute_txn_commit_thread_running(lut) ||
1719                      distribute_txn_commit_thread_stopped(lut), &lwi);
1720         RETURN(0);
1721 }
1722 EXPORT_SYMBOL(distribute_txn_init);
1723
1724 /**
1725  * Stop llog cancel thread
1726  *
1727  * Stop llog cancel(master/slave) thread on LOD and also destory
1728  * all of transaction in the list.
1729  *
1730  * \param[in]lclt       cancel log thread to be stopped.
1731  */
1732 void distribute_txn_fini(const struct lu_env *env,
1733                          struct target_distribute_txn_data *tdtd)
1734 {
1735         struct lu_target *lut = tdtd->tdtd_lut;
1736
1737         /* Stop cancel thread */
1738         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1739                 return;
1740
1741         spin_lock(&tdtd->tdtd_batchid_lock);
1742         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1743         spin_unlock(&tdtd->tdtd_batchid_lock);
1744         wake_up(&tdtd->tdtd_commit_thread_waitq);
1745         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1746                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1747
1748         dtrq_list_destroy(tdtd);
1749         if (tdtd->tdtd_batchid_obj != NULL) {
1750                 dt_object_put(env, tdtd->tdtd_batchid_obj);
1751                 tdtd->tdtd_batchid_obj = NULL;
1752         }
1753 }
1754 EXPORT_SYMBOL(distribute_txn_fini);