lustre/target/update_trans.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2014, Intel Corporation.
  24  */
  25 /*
  26  * lustre/target/update_trans.c
  27  *
  28  * This file implements the update distribute transaction API.
  29  *
  30  * To manage the cross-MDT operation (distribute operation) transaction,
  31  * the transaction will also be separated two layers on MD stack, top
  32  * transaction and sub transaction.
  33  *
  34  * During the distribute operation, top transaction is created in the LOD
  35  * layer, and represent the operation. Sub transaction is created by
  36  * each OSD or OSP. Top transaction start/stop will trigger all of its sub
  37  * transaction start/stop. Top transaction (the whole operation) is committed
  38  * only all of its sub transaction are committed.
  39  *
  40  * there are three kinds of transactions
  41  * 1. local transaction: All updates are in a single local OSD.
  42  * 2. Remote transaction: All Updates are only in the remote OSD,
  43  *    i.e. locally all updates are in OSP.
  44  * 3. Mixed transaction: Updates are both in local OSD and remote
  45  *    OSD.
  46  *
  47  * Author: Di Wang <di.wang@intel.com>
  48  */
  49
  50 #define DEBUG_SUBSYSTEM S_CLASS
  51
  52 #include <linux/kthread.h>
  53 #include <lu_target.h>
  54 #include <lustre_log.h>
  55 #include <lustre_update.h>
  56 #include <obd.h>
  57 #include <obd_class.h>
  58 #include <tgt_internal.h>
  59
  60 #include <tgt_internal.h>
  61 /**
  62  * Dump top mulitple thandle
  63  *
  64  * Dump top multiple thandle and all of its sub thandle to the debug log.
  65  *
  66  * \param[in]mask       debug mask
  67  * \param[in]top_th     top_thandle to be dumped
  68  */
  69 static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
  70                                       __u32 mask)
  71 {
  72         struct sub_thandle      *st;
  73
  74         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
  75         CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d"
  76                "batchid "LPU64"\n",
  77                tmt->tmt_master_sub_dt ?
  78                tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
  79                "NULL",
  80                tmt, atomic_read(&tmt->tmt_refcount), tmt->tmt_committed,
  81                tmt->tmt_result, tmt->tmt_batchid);
  82
  83         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
  84                 struct sub_thandle_cookie *stc;
  85
  86                 CDEBUG(mask, "st %p obd %s committed %d sub_th %p\n",
  87                        st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
  88                        st->st_committed, st->st_sub_th);
  89
  90                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
  91                         CDEBUG(mask, " cookie "DOSTID": %u\n",
  92                                POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
  93                                stc->stc_cookie.lgc_index);
  94                 }
  95         }
  96 }
  97
  98 /**
  99  * Declare write update to sub device
 100  *
 101  * Declare Write updates llog records to the sub device during distribute
 102  * transaction.
 103  *
 104  * \param[in] env       execution environment
 105  * \param[in] record    update records being written
 106  * \param[in] sub_th    sub transaction handle
 107  * \param[in] record_size total update record size
 108  *
 109  * \retval              0 if writing succeeds
 110  * \retval              negative errno if writing fails
 111  */
 112 static int sub_declare_updates_write(const struct lu_env *env,
 113                                      struct llog_update_record *record,
 114                                      struct thandle *sub_th, size_t record_size)
 115 {
 116         struct llog_ctxt        *ctxt;
 117         struct dt_device        *dt = sub_th->th_dev;
 118         int                     left = record_size;
 119         int rc;
 120
 121         /* If ctxt is NULL, it means not need to write update,
 122          * for example if the the OSP is used to connect to OST */
 123         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 124                                 LLOG_UPDATELOG_ORIG_CTXT);
 125         LASSERT(ctxt != NULL);
 126
 127         /* Not ready to record updates yet. */
 128         if (ctxt->loc_handle == NULL)
 129                 GOTO(out_put, rc = 0);
 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         int                     rc;
 173         struct llog_update_record *lur = NULL;
 174         struct update_params    *params = NULL;
 175         __u32                   update_count = 0;
 176         __u32                   param_count = 0;
 177         __u32                   last_update_count = 0;
 178         __u32                   last_param_count = 0;
 179         void                    *src;
 180         void                    *start;
 181         void                    *next;
 182         struct sub_thandle_cookie *stc;
 183         ENTRY;
 184
 185         ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
 186                                 LLOG_UPDATELOG_ORIG_CTXT);
 187         LASSERT(ctxt != NULL);
 188
 189         /* Not ready to record updates yet, usually happens
 190          * in error handler path */
 191         if (ctxt->loc_handle == NULL)
 192                 GOTO(llog_put, rc = 0);
 193
 194         /* Since the cross-MDT updates will includes both local
 195          * and remote updates, the update ops count must > 1 */
 196         LASSERT(record->lur_update_rec.ur_update_count > 1);
 197         LASSERTF(record->lur_hdr.lrh_len == llog_update_record_size(record),
 198                  "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
 199                  llog_update_record_size(record));
 200
 201         if (likely(record->lur_hdr.lrh_len <= ctxt->loc_chunk_size)) {
 202                 OBD_ALLOC_PTR(stc);
 203                 if (stc == NULL)
 204                         GOTO(llog_put, rc = -ENOMEM);
 205                 INIT_LIST_HEAD(&stc->stc_list);
 206
 207                 rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
 208                               &stc->stc_cookie, sub_th->st_sub_th);
 209
 210                 CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u: rc = %d\n",
 211                        dt->dd_lu_dev.ld_obd->obd_name,
 212                        POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
 213                        stc->stc_cookie.lgc_index, rc);
 214
 215                 if (rc > 0) {
 216                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 217                         rc = 0;
 218                 } else {
 219                         OBD_FREE_PTR(stc);
 220                 }
 221
 222                 GOTO(llog_put, rc);
 223         }
 224
 225         /* Split the records into chunk_size update record */
 226         OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
 227         if (lur == NULL)
 228                 GOTO(llog_put, rc = -ENOMEM);
 229
 230         memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
 231         lur->lur_update_rec.ur_update_count = 0;
 232         lur->lur_update_rec.ur_param_count = 0;
 233         src = &record->lur_update_rec.ur_ops;
 234         start = next = src;
 235         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 236         params = update_records_get_params(&record->lur_update_rec);
 237         do {
 238                 size_t rec_len;
 239
 240                 if (update_count < record->lur_update_rec.ur_update_count) {
 241                         next = update_op_next_op((struct update_op *)src);
 242                 } else {
 243                         if (param_count == 0)
 244                                 next = update_records_get_params(
 245                                                 &record->lur_update_rec);
 246                         else
 247                                 next = (char *)src +
 248                                         object_update_param_size(
 249                                         (struct object_update_param *)src);
 250                 }
 251
 252                 rec_len = cfs_size_round((unsigned long)(next - src));
 253                 /* If its size > llog chunk_size, then write current chunk to
 254                  * the update llog. */
 255                 if (lur->lur_hdr.lrh_len + rec_len + LLOG_MIN_REC_SIZE >
 256                     ctxt->loc_chunk_size ||
 257                     param_count == record->lur_update_rec.ur_param_count) {
 258                         lur->lur_update_rec.ur_update_count =
 259                                 update_count > last_update_count ?
 260                                 update_count - last_update_count : 0;
 261                         lur->lur_update_rec.ur_param_count = param_count -
 262                                                              last_param_count;
 263
 264                         memcpy(&lur->lur_update_rec.ur_ops, start,
 265                                (unsigned long)(src - start));
 266                         if (last_update_count != 0)
 267                                 lur->lur_update_rec.ur_flags |=
 268                                                 UPDATE_RECORD_CONTINUE;
 269
 270                         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 271                         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 272                         LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
 273
 274                         OBD_ALLOC_PTR(stc);
 275                         if (stc == NULL)
 276                                 GOTO(llog_put, rc = -ENOMEM);
 277                         INIT_LIST_HEAD(&stc->stc_list);
 278
 279                         rc = llog_add(env, ctxt->loc_handle,
 280                                       &lur->lur_hdr,
 281                                       &stc->stc_cookie, sub_th->st_sub_th);
 282
 283                         CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u"
 284                                " rc = %d\n", dt->dd_lu_dev.ld_obd->obd_name,
 285                                POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
 286                                stc->stc_cookie.lgc_index, rc);
 287
 288                         if (rc > 0) {
 289                                 list_add(&stc->stc_list,
 290                                          &sub_th->st_cookie_list);
 291                                 rc = 0;
 292                         } else {
 293                                 OBD_FREE_PTR(stc);
 294                                 GOTO(llog_put, rc);
 295                         }
 296
 297                         last_update_count = update_count;
 298                         last_param_count = param_count;
 299                         start = src;
 300                         lur->lur_update_rec.ur_update_count = 0;
 301                         lur->lur_update_rec.ur_param_count = 0;
 302                         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 303                 }
 304
 305                 src = next;
 306                 lur->lur_hdr.lrh_len += cfs_size_round(rec_len);
 307                 if (update_count < record->lur_update_rec.ur_update_count)
 308                         update_count++;
 309                 else if (param_count < record->lur_update_rec.ur_param_count)
 310                         param_count++;
 311                 else
 312                         break;
 313         } while (1);
 314
 315 llog_put:
 316         if (lur != NULL)
 317                 OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
 318         llog_ctxt_put(ctxt);
 319
 320         RETURN(rc);
 321 }
 322
 323 /**
 324  * Prepare the update records.
 325  *
 326  * Merge params and ops into the update records, then initializing
 327  * the update buffer.
 328  *
 329  * During transaction execution phase, parameters and update ops
 330  * are collected in two different buffers (see lod_updates_pack()),
 331  * during transaction stop, it needs to be merged in one buffer,
 332  * so it will be written in the update log.
 333  *
 334  * \param[in] env       execution environment
 335  * \param[in] tmt       top_multiple_thandle for distribute txn
 336  *
 337  * \retval              0 if merging succeeds.
 338  * \retval              negaitive errno if merging fails.
 339  */
 340 static int prepare_writing_updates(const struct lu_env *env,
 341                                    struct top_multiple_thandle *tmt)
 342 {
 343         struct thandle_update_records   *tur = tmt->tmt_update_records;
 344         struct llog_update_record       *lur;
 345         struct update_params *params;
 346         size_t params_size;
 347         size_t update_size;
 348
 349         if (tur == NULL || tur->tur_update_records == NULL ||
 350             tur->tur_update_params == NULL)
 351                 return 0;
 352
 353         lur = tur->tur_update_records;
 354         /* Extends the update records buffer if needed */
 355         params_size = update_params_size(tur->tur_update_params,
 356                                          tur->tur_update_param_count);
 357         LASSERT(lur->lur_update_rec.ur_param_count == 0);
 358         update_size = llog_update_record_size(lur);
 359         if (cfs_size_round(update_size + params_size) >
 360             tur->tur_update_records_buf_size) {
 361                 int rc;
 362
 363                 rc = tur_update_records_extend(tur,
 364                         cfs_size_round(update_size + params_size));
 365                 if (rc < 0)
 366                         return rc;
 367
 368                 lur = tur->tur_update_records;
 369         }
 370
 371         params = update_records_get_params(&lur->lur_update_rec);
 372         memcpy(params, tur->tur_update_params, params_size);
 373
 374         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
 375         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
 376         /* Init update record header */
 377         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 378         lur->lur_hdr.lrh_type = UPDATE_REC;
 379
 380         /* Dump updates for debugging purpose */
 381         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 382
 383         return 0;
 384 }
 385
 386 static inline int
 387 distribute_txn_commit_thread_running(struct lu_target *lut)
 388 {
 389         return lut->lut_tdtd_commit_thread.t_flags & SVC_RUNNING;
 390 }
 391
 392 static inline int
 393 distribute_txn_commit_thread_stopped(struct lu_target *lut)
 394 {
 395         return lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED;
 396 }
 397
 398 /**
 399  * Top thandle commit callback
 400  *
 401  * This callback will be called when all of sub transactions are committed.
 402  *
 403  * \param[in] th        top thandle to be committed.
 404  */
 405 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
 406 {
 407         struct lu_target *lut;
 408         ENTRY;
 409
 410         LASSERT(atomic_read(&tmt->tmt_refcount) > 0);
 411
 412         top_multiple_thandle_dump(tmt, D_HA);
 413         tmt->tmt_committed = 1;
 414         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
 415         if (distribute_txn_commit_thread_running(lut))
 416                 wake_up(&lut->lut_tdtd->tdtd_commit_thread_waitq);
 417         RETURN_EXIT;
 418 }
 419
 420 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
 421                                        struct dt_device *dt_dev)
 422 {
 423         struct sub_thandle *st;
 424
 425         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 426                 if (st->st_dt == dt_dev)
 427                         return st;
 428         }
 429         return NULL;
 430 }
 431 EXPORT_SYMBOL(lookup_sub_thandle);
 432
 433 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
 434                                        struct dt_device *dt_dev)
 435 {
 436         struct sub_thandle *st;
 437
 438         OBD_ALLOC_PTR(st);
 439         if (st == NULL)
 440                 RETURN(ERR_PTR(-ENOMEM));
 441
 442         INIT_LIST_HEAD(&st->st_sub_list);
 443         INIT_LIST_HEAD(&st->st_cookie_list);
 444         st->st_dt = dt_dev;
 445
 446         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
 447         return st;
 448 }
 449
 450 /**
 451  * sub thandle commit callback
 452  *
 453  * Mark the sub thandle to be committed and if all sub thandle are committed
 454  * notify the top thandle.
 455  *
 456  * \param[in] env       execution environment
 457  * \param[in] sub_th    sub thandle being committed
 458  * \param[in] cb        commit callback
 459  * \param[in] err       trans result
 460  */
 461 static void sub_trans_commit_cb(struct lu_env *env,
 462                                 struct thandle *sub_th,
 463                                 struct dt_txn_commit_cb *cb, int err)
 464 {
 465         struct sub_thandle      *st;
 466         struct top_multiple_thandle *tmt = cb->dcb_data;
 467         bool                    all_committed = true;
 468         ENTRY;
 469
 470         /* Check if all sub thandles are committed */
 471         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 472                 if (st->st_sub_th == sub_th) {
 473                         st->st_committed = 1;
 474                         st->st_result = err;
 475                 }
 476                 if (!st->st_committed)
 477                         all_committed = false;
 478         }
 479
 480         if (tmt->tmt_result == 0)
 481                 tmt->tmt_result = err;
 482
 483         if (all_committed)
 484                 top_trans_committed_cb(tmt);
 485
 486         top_multiple_thandle_dump(tmt, D_INFO);
 487         top_multiple_thandle_put(tmt);
 488         RETURN_EXIT;
 489 }
 490
 491 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
 492                                     struct top_multiple_thandle *tmt)
 493 {
 494         LASSERT(st->st_sub_th != NULL);
 495         top_multiple_thandle_get(tmt);
 496         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
 497         st->st_commit_dcb.dcb_data = tmt;
 498         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
 499         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
 500 }
 501
 502 /**
 503  * Sub thandle stop call back
 504  *
 505  * After sub thandle is stopped, it will call this callback to notify
 506  * the top thandle.
 507  *
 508  * \param[in] th        sub thandle to be stopped
 509  * \param[in] rc        result of sub trans
 510  */
 511 static void sub_trans_stop_cb(struct lu_env *env,
 512                               struct thandle *sub_th,
 513                               struct dt_txn_commit_cb *cb, int err)
 514 {
 515         struct sub_thandle              *st;
 516         struct top_multiple_thandle     *tmt = cb->dcb_data;
 517         ENTRY;
 518
 519         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 520                 if (st->st_stopped)
 521                         continue;
 522
 523                 if (st->st_dt == sub_th->th_dev) {
 524                         st->st_stopped = 1;
 525                         st->st_result = err;
 526                         break;
 527                 }
 528         }
 529
 530         wake_up(&tmt->tmt_stop_waitq);
 531         RETURN_EXIT;
 532 }
 533
 534 static void sub_thandle_register_stop_cb(struct sub_thandle *st,
 535                                          struct top_multiple_thandle *tmt)
 536 {
 537         st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
 538         st->st_stop_dcb.dcb_data = tmt;
 539         st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
 540         INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
 541         dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
 542 }
 543
 544 /**
 545  * Create sub thandle
 546  *
 547  * Create transaction handle for sub_thandle
 548  *
 549  * \param[in] env       execution environment
 550  * \param[in] th        top thandle
 551  * \param[in] st        sub_thandle
 552  *
 553  * \retval              0 if creation succeeds.
 554  * \retval              negative errno if creation fails.
 555  */
 556 int sub_thandle_trans_create(const struct lu_env *env,
 557                              struct top_thandle *top_th,
 558                              struct sub_thandle *st)
 559 {
 560         struct thandle *sub_th;
 561
 562         sub_th = dt_trans_create(env, st->st_dt);
 563         if (IS_ERR(sub_th))
 564                 return PTR_ERR(sub_th);
 565
 566         sub_th->th_top = &top_th->tt_super;
 567         st->st_sub_th = sub_th;
 568
 569         sub_th->th_wait_submit = 1;
 570         return 0;
 571 }
 572
 573 /**
 574  * Create the top transaction.
 575  *
 576  * Create the top transaction on the master device. It will create a top
 577  * thandle and a sub thandle on the master device.
 578  *
 579  * \param[in] env               execution environment
 580  * \param[in] master_dev        master_dev the top thandle will be created
 581  *
 582  * \retval                      pointer to the created thandle.
 583  * \retval                      ERR_PTR(errno) if creation failed.
 584  */
 585 struct thandle *
 586 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
 587 {
 588         struct top_thandle      *top_th;
 589         struct thandle          *child_th;
 590
 591         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
 592         if (top_th == NULL)
 593                 return ERR_PTR(-ENOMEM);
 594
 595         top_th->tt_super.th_top = &top_th->tt_super;
 596
 597         if (master_dev != NULL) {
 598                 child_th = dt_trans_create(env, master_dev);
 599                 if (IS_ERR(child_th)) {
 600                         OBD_FREE_PTR(top_th);
 601                         return child_th;
 602                 }
 603
 604                 child_th->th_top = &top_th->tt_super;
 605                 child_th->th_wait_submit = 1;
 606                 top_th->tt_master_sub_thandle = child_th;
 607
 608                 top_th->tt_super.th_tags |= child_th->th_tags;
 609         }
 610         return &top_th->tt_super;
 611 }
 612 EXPORT_SYMBOL(top_trans_create);
 613
 614 /**
 615  * Declare write update transaction
 616  *
 617  * Check if there are updates being recorded in this transaction,
 618  * it will write the record into the disk.
 619  *
 620  * \param[in] env       execution environment
 621  * \param[in] tmt       top multiple transaction handle
 622  *
 623  * \retval              0 if writing succeeds
 624  * \retval              negative errno if writing fails
 625  */
 626 static int declare_updates_write(const struct lu_env *env,
 627                                  struct top_multiple_thandle *tmt)
 628 {
 629         struct llog_update_record *record;
 630         struct sub_thandle *st;
 631         int rc = 0;
 632
 633         record = tmt->tmt_update_records->tur_update_records;
 634         /* Declare update write for all other target */
 635         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 636                 if (st->st_sub_th == NULL)
 637                         continue;
 638
 639                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
 640                                                tmt->tmt_record_size);
 641                 if (rc < 0)
 642                         break;
 643         }
 644
 645         return rc;
 646 }
 647
 648 /**
 649  * Assign batchid to the distribute transaction.
 650  *
 651  * Assign batchid to the distribute transaction
 652  *
 653  * \param[in] tmt       distribute transaction
 654  */
 655 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 656 {
 657         struct target_distribute_txn_data *tdtd;
 658         struct dt_device *dt = new->tmt_master_sub_dt;
 659
 660         LASSERT(dt != NULL);
 661         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 662         spin_lock(&tdtd->tdtd_batchid_lock);
 663         new->tmt_batchid = tdtd->tdtd_batchid++;
 664         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
 665         spin_unlock(&tdtd->tdtd_batchid_lock);
 666         top_multiple_thandle_get(new);
 667         top_multiple_thandle_dump(new, D_INFO);
 668 }
 669
 670 /**
 671  * Insert distribute transaction to the distribute txn list.
 672  *
 673  * Insert distribute transaction to the distribute txn list.
 674  *
 675  * \param[in] new       the distribute txn to be inserted.
 676  */
 677 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
 678 {
 679         struct dt_device *dt = new->tmt_master_sub_dt;
 680         struct top_multiple_thandle *tmt;
 681         struct target_distribute_txn_data *tdtd;
 682         bool    at_head = false;
 683
 684         LASSERT(dt != NULL);
 685         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 686
 687         spin_lock(&tdtd->tdtd_batchid_lock);
 688         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
 689                 if (new->tmt_batchid > tmt->tmt_batchid) {
 690                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
 691                         break;
 692                 }
 693         }
 694         if (list_empty(&new->tmt_commit_list)) {
 695                 at_head = true;
 696                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
 697         }
 698         spin_unlock(&tdtd->tdtd_batchid_lock);
 699         top_multiple_thandle_get(new);
 700         top_multiple_thandle_dump(new, D_INFO);
 701         if (new->tmt_committed && at_head)
 702                 wake_up(&tdtd->tdtd_commit_thread_waitq);
 703 }
 704
 705 /**
 706  * Prepare cross-MDT operation.
 707  *
 708  * Create the update record buffer to record updates for cross-MDT operation,
 709  * add master sub transaction to tt_sub_trans_list, and declare the update
 710  * writes.
 711  *
 712  * During updates packing, all of parameters will be packed in
 713  * tur_update_params, and updates will be packed in tur_update_records.
 714  * Then in transaction stop, parameters and updates will be merged
 715  * into one updates buffer.
 716  *
 717  * And also master thandle will be added to the sub_th list, so it will be
 718  * easy to track the commit status.
 719  *
 720  * \param[in] env       execution environment
 721  * \param[in] th        top transaction handle
 722  *
 723  * \retval              0 if preparation succeeds.
 724  * \retval              negative errno if preparation fails.
 725  */
 726 static int prepare_multiple_node_trans(const struct lu_env *env,
 727                                        struct top_multiple_thandle *tmt)
 728 {
 729         struct thandle_update_records   *tur;
 730         int                             rc;
 731         ENTRY;
 732
 733         if (tmt->tmt_update_records == NULL) {
 734                 tur = &update_env_info(env)->uti_tur;
 735                 rc = check_and_prepare_update_record(env, tur);
 736                 if (rc < 0)
 737                         RETURN(rc);
 738
 739                 tmt->tmt_update_records = tur;
 740                 distribute_txn_assign_batchid(tmt);
 741         }
 742
 743         rc = declare_updates_write(env, tmt);
 744
 745         RETURN(rc);
 746 }
 747
 748 /**
 749  * start the top transaction.
 750  *
 751  * Start all of its sub transactions, then start master sub transaction.
 752  *
 753  * \param[in] env               execution environment
 754  * \param[in] master_dev        master_dev the top thandle will be start
 755  * \param[in] th                top thandle
 756  *
 757  * \retval                      0 if transaction start succeeds.
 758  * \retval                      negative errno if start fails.
 759  */
 760 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
 761                     struct thandle *th)
 762 {
 763         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 764                                                        tt_super);
 765         struct sub_thandle              *st;
 766         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
 767         int                             rc = 0;
 768         ENTRY;
 769
 770         if (tmt == NULL) {
 771                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
 772                                     top_th->tt_master_sub_thandle);
 773                 RETURN(rc);
 774         }
 775
 776         tmt = top_th->tt_multiple_thandle;
 777         rc = prepare_multiple_node_trans(env, tmt);
 778         if (rc < 0)
 779                 RETURN(rc);
 780
 781         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 782                 if (st->st_sub_th == NULL)
 783                         continue;
 784                 if (th->th_sync)
 785                         st->st_sub_th->th_sync = th->th_sync;
 786                 st->st_sub_th->th_local = th->th_local;
 787                 st->st_sub_th->th_tags = th->th_tags;
 788                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
 789                                     st->st_sub_th);
 790                 if (rc != 0)
 791                         GOTO(out, rc);
 792
 793                 sub_thandle_register_stop_cb(st, tmt);
 794                 sub_thandle_register_commit_cb(st, tmt);
 795         }
 796 out:
 797         th->th_result = rc;
 798         RETURN(rc);
 799 }
 800 EXPORT_SYMBOL(top_trans_start);
 801
 802 /**
 803  * Check whether we need write updates record
 804  *
 805  * Check if the updates for the top_thandle needs to be writen
 806  * to all targets. Only if the transaction succeeds and the updates
 807  * number > 2, it will write the updates,
 808  *
 809  * \params [in] top_th  top thandle.
 810  *
 811  * \retval              true if it needs to write updates
 812  * \retval              false if it does not need to write updates
 813  **/
 814 static bool top_check_write_updates(struct top_thandle *top_th)
 815 {
 816         struct top_multiple_thandle     *tmt;
 817         struct thandle_update_records   *tur;
 818
 819         /* Do not write updates to records if the transaction fails */
 820         if (top_th->tt_super.th_result != 0)
 821                 return false;
 822
 823         tmt = top_th->tt_multiple_thandle;
 824         if (tmt == NULL)
 825                 return false;
 826
 827         tur = tmt->tmt_update_records;
 828         if (tur == NULL)
 829                 return false;
 830
 831         /* Hmm, false update records, since the cross-MDT operation
 832          * should includes both local and remote updates, so the
 833          * updates count should >= 2 */
 834         if (tur->tur_update_records == NULL ||
 835             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
 836                 return false;
 837
 838         return true;
 839 }
 840
 841 /**
 842  * Check if top transaction is stopped
 843  *
 844  * Check if top transaction is stopped, only if all sub transaction
 845  * is stopped, then the top transaction is stopped.
 846  *
 847  * \param [in] top_th   top thandle
 848  *
 849  * \retval              true if the top transaction is stopped.
 850  * \retval              false if the top transaction is not stopped.
 851  */
 852 static bool top_trans_is_stopped(struct top_thandle *top_th)
 853 {
 854         struct top_multiple_thandle     *tmt;
 855         struct sub_thandle              *st;
 856         bool                    all_stopped = true;
 857
 858         tmt = top_th->tt_multiple_thandle;
 859         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 860                 if (!st->st_stopped && st->st_sub_th != NULL) {
 861                         all_stopped = false;
 862                         break;
 863                 }
 864
 865                 if (st->st_result != 0 &&
 866                     top_th->tt_super.th_result == 0)
 867                         top_th->tt_super.th_result = st->st_result;
 868         }
 869
 870         return all_stopped;
 871 }
 872
 873 /**
 874  * Wait result of top transaction
 875  *
 876  * Wait until all sub transaction get its result.
 877  *
 878  * \param [in] top_th   top thandle.
 879  *
 880  * \retval              the result of top thandle.
 881  */
 882 static int top_trans_wait_result(struct top_thandle *top_th)
 883 {
 884         struct l_wait_info      lwi = {0};
 885
 886         l_wait_event(top_th->tt_multiple_thandle->tmt_stop_waitq,
 887                      top_trans_is_stopped(top_th), &lwi);
 888
 889         RETURN(top_th->tt_super.th_result);
 890 }
 891
 892 /**
 893  * Stop the top transaction.
 894  *
 895  * Stop the transaction on the master device first, then stop transactions
 896  * on other sub devices.
 897  *
 898  * \param[in] env               execution environment
 899  * \param[in] master_dev        master_dev the top thandle will be created
 900  * \param[in] th                top thandle
 901  *
 902  * \retval                      0 if stop transaction succeeds.
 903  * \retval                      negative errno if stop transaction fails.
 904  */
 905 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
 906                    struct thandle *th)
 907 {
 908         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 909                                                        tt_super);
 910         struct sub_thandle              *st;
 911         struct sub_thandle              *master_st;
 912         struct top_multiple_thandle     *tmt;
 913         struct thandle_update_records   *tur;
 914         bool                            write_updates = false;
 915         int                     rc = 0;
 916         ENTRY;
 917
 918         if (likely(top_th->tt_multiple_thandle == NULL)) {
 919                 LASSERT(master_dev != NULL);
 920                 rc = dt_trans_stop(env, master_dev,
 921                                    top_th->tt_master_sub_thandle);
 922                 OBD_FREE_PTR(top_th);
 923                 RETURN(rc);
 924         }
 925
 926         tmt = top_th->tt_multiple_thandle;
 927         tur = tmt->tmt_update_records;
 928
 929         /* Note: we need stop the master thandle first, then the stop
 930          * callback will fill the master transno in the update logs,
 931          * then these update logs will be sent to other MDTs */
 932         /* get the master sub thandle */
 933         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
 934         write_updates = top_check_write_updates(top_th);
 935
 936         /* Step 1: write the updates log on Master MDT */
 937         if (master_st != NULL && master_st->st_sub_th != NULL &&
 938             write_updates) {
 939                 struct llog_update_record *lur;
 940
 941                 /* Merge the parameters and updates into one buffer */
 942                 rc = prepare_writing_updates(env, tmt);
 943                 if (rc < 0) {
 944                         CERROR("%s: cannot prepare updates: rc = %d\n",
 945                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 946                         th->th_result = rc;
 947                         GOTO(stop_master_trans, rc);
 948                 }
 949
 950                 lur = tur->tur_update_records;
 951                 /* Write updates to the master MDT */
 952                 rc = sub_updates_write(env, lur, master_st);
 953
 954                 /* Cleanup the common parameters in the update records,
 955                  * master transno callback might add more parameters.
 956                  * and we need merge the update records again in the
 957                  * following */
 958                 if (tur->tur_update_params != NULL)
 959                         lur->lur_update_rec.ur_param_count = 0;
 960
 961                 if (rc < 0) {
 962                         CERROR("%s: write updates failed: rc = %d\n",
 963                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 964                         th->th_result = rc;
 965                         GOTO(stop_master_trans, rc);
 966                 }
 967         }
 968
 969 stop_master_trans:
 970         /* Step 2: Stop the transaction on the master MDT, and fill the
 971          * master transno in the update logs to other MDT. */
 972         if (master_st != NULL && master_st->st_sub_th != NULL) {
 973                 master_st->st_sub_th->th_local = th->th_local;
 974                 if (th->th_sync)
 975                         master_st->st_sub_th->th_sync = th->th_sync;
 976                 master_st->st_sub_th->th_tags = th->th_tags;
 977                 master_st->st_sub_th->th_result = th->th_result;
 978                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
 979                 if (rc < 0) {
 980                         th->th_result = rc;
 981                         GOTO(stop_other_trans, rc);
 982                 } else if (tur != NULL && tur->tur_update_records != NULL) {
 983                         struct llog_update_record *lur;
 984
 985                         lur = tur->tur_update_records;
 986                         if (lur->lur_update_rec.ur_master_transno == 0)
 987                                 /* Update master transno after master stop
 988                                  * callback */
 989                                 lur->lur_update_rec.ur_master_transno =
 990                                                 tgt_th_info(env)->tti_transno;
 991                 }
 992         }
 993
 994         /* Step 3: write updates to other MDTs */
 995         if (write_updates) {
 996                 struct llog_update_record *lur;
 997
 998                 /* Stop callback of master will add more updates and also update
 999                  * master transno, so merge the parameters and updates into one
1000                  * buffer again */
1001                 rc = prepare_writing_updates(env, tmt);
1002                 if (rc < 0) {
1003                         CERROR("%s: prepare updates failed: rc = %d\n",
1004                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1005                         th->th_result = rc;
1006                         GOTO(stop_other_trans, rc);
1007                 }
1008                 lur = tur->tur_update_records;
1009                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1010                                     st_sub_list) {
1011                         if (st->st_sub_th == NULL || st == master_st ||
1012                             st->st_sub_th->th_result < 0)
1013                                 continue;
1014
1015                         rc = sub_updates_write(env, lur, st);
1016                         if (rc < 0) {
1017                                 th->th_result = rc;
1018                                 break;
1019                         }
1020                 }
1021         }
1022
1023 stop_other_trans:
1024         /* Step 4: Stop the transaction on other MDTs */
1025         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1026                 if (st == master_st || st->st_sub_th == NULL)
1027                         continue;
1028
1029                 if (th->th_sync)
1030                         st->st_sub_th->th_sync = th->th_sync;
1031                 st->st_sub_th->th_local = th->th_local;
1032                 st->st_sub_th->th_tags = th->th_tags;
1033                 st->st_sub_th->th_result = th->th_result;
1034                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1035                                    st->st_sub_th);
1036                 if (unlikely(rc < 0 && th->th_result == 0))
1037                         th->th_result = rc;
1038         }
1039
1040         rc = top_trans_wait_result(top_th);
1041
1042         tmt->tmt_result = rc;
1043
1044         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1045          * multiple node transaction, the top transaction will be destroyed. */
1046         top_multiple_thandle_put(tmt);
1047         OBD_FREE_PTR(top_th);
1048         RETURN(rc);
1049 }
1050 EXPORT_SYMBOL(top_trans_stop);
1051
1052 /**
1053  * Create top_multiple_thandle for top_thandle
1054  *
1055  * Create top_mutilple_thandle to manage the mutiple node transaction
1056  * for top_thandle, and it also needs to add master sub thandle to the
1057  * sub trans list now.
1058  *
1059  * \param[in] env       execution environment
1060  * \param[in] top_th    the top thandle
1061  *
1062  * \retval      0 if creation succeeds
1063  * \retval      negative errno if creation fails
1064  */
1065 int top_trans_create_tmt(const struct lu_env *env,
1066                          struct top_thandle *top_th)
1067 {
1068         struct top_multiple_thandle *tmt;
1069
1070         OBD_ALLOC_PTR(tmt);
1071         if (tmt == NULL)
1072                 return -ENOMEM;
1073
1074         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1075         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1076         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1077         atomic_set(&tmt->tmt_refcount, 1);
1078
1079         init_waitqueue_head(&tmt->tmt_stop_waitq);
1080         top_th->tt_multiple_thandle = tmt;
1081
1082         return 0;
1083 }
1084
1085 static struct sub_thandle *
1086 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1087                                 struct thandle *sub_th)
1088 {
1089         struct sub_thandle *st;
1090
1091         /* create and init sub th to the top trans list */
1092         st = create_sub_thandle(top_th->tt_multiple_thandle,
1093                                 sub_th->th_dev);
1094         if (IS_ERR(st))
1095                 return st;
1096
1097         st->st_sub_th = sub_th;
1098
1099         sub_th->th_top = &top_th->tt_super;
1100         return st;
1101 }
1102
1103 /**
1104  * Get sub thandle.
1105  *
1106  * Get sub thandle from the top thandle according to the sub dt_device.
1107  *
1108  * \param[in] env       execution environment
1109  * \param[in] th        thandle on the top layer.
1110  * \param[in] sub_dt    sub dt_device used to get sub transaction
1111  *
1112  * \retval              thandle of sub transaction if succeed
1113  * \retval              PTR_ERR(errno) if failed
1114  */
1115 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1116                                       struct thandle *th,
1117                                       struct dt_device *sub_dt)
1118 {
1119         struct sub_thandle      *st = NULL;
1120         struct top_thandle      *top_th;
1121         struct thandle          *sub_th = NULL;
1122         int                     rc = 0;
1123         ENTRY;
1124
1125         top_th = container_of(th, struct top_thandle, tt_super);
1126
1127         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1128                 RETURN(top_th->tt_master_sub_thandle);
1129
1130         if (top_th->tt_multiple_thandle != NULL) {
1131                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1132                 if (st != NULL)
1133                         RETURN(st->st_sub_th);
1134         }
1135
1136         sub_th = dt_trans_create(env, sub_dt);
1137         if (IS_ERR(sub_th))
1138                 RETURN(sub_th);
1139
1140         /* Create top_multiple_thandle if necessary */
1141         if (top_th->tt_multiple_thandle == NULL) {
1142                 struct top_multiple_thandle *tmt;
1143
1144                 rc = top_trans_create_tmt(env, top_th);
1145                 if (rc < 0)
1146                         GOTO(stop_trans, rc);
1147
1148                 tmt = top_th->tt_multiple_thandle;
1149
1150                 /* Add master sub th to the top trans list */
1151                 tmt->tmt_master_sub_dt =
1152                         top_th->tt_master_sub_thandle->th_dev;
1153                 st = create_sub_thandle_with_thandle(top_th,
1154                                 top_th->tt_master_sub_thandle);
1155                 if (IS_ERR(st))
1156                         GOTO(stop_trans, rc = PTR_ERR(st));
1157         }
1158
1159         /* create and init sub th to the top trans list */
1160         st = create_sub_thandle_with_thandle(top_th, sub_th);
1161         st->st_sub_th->th_wait_submit = 1;
1162 stop_trans:
1163         if (rc < 0) {
1164                 if (st != NULL)
1165                         OBD_FREE_PTR(st);
1166                 sub_th->th_result = rc;
1167                 dt_trans_stop(env, sub_dt, sub_th);
1168                 sub_th = ERR_PTR(rc);
1169         }
1170
1171         RETURN(sub_th);
1172 }
1173 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1174
1175 /**
1176  * Top multiple thandle destroy
1177  *
1178  * Destroy multiple thandle and all its sub thandle.
1179  *
1180  * \param[in] tmt       top_multiple_thandle to be destroyed.
1181  */
1182 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1183 {
1184         struct sub_thandle *st;
1185         struct sub_thandle *tmp;
1186
1187         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1188         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1189                                  st_sub_list) {
1190                 struct sub_thandle_cookie *stc;
1191                 struct sub_thandle_cookie *tmp;
1192
1193                 list_del(&st->st_sub_list);
1194                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1195                                          stc_list) {
1196                         list_del(&stc->stc_list);
1197                         OBD_FREE_PTR(stc);
1198                 }
1199                 OBD_FREE_PTR(st);
1200         }
1201         OBD_FREE_PTR(tmt);
1202 }
1203 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1204
1205 /**
1206  * Cancel the update log on MDTs
1207  *
1208  * Cancel the update log on MDTs then destroy the thandle.
1209  *
1210  * \param[in] env       execution environment
1211  * \param[in] tmt       the top multiple thandle whose updates records
1212  *                      will be cancelled.
1213  *
1214  * \retval              0 if cancellation succeeds.
1215  * \retval              negative errno if cancellation fails.
1216  */
1217 static int distribute_txn_cancel_records(const struct lu_env *env,
1218                                          struct top_multiple_thandle *tmt)
1219 {
1220         struct sub_thandle *st;
1221         ENTRY;
1222
1223         top_multiple_thandle_dump(tmt, D_INFO);
1224         /* Cancel update logs on other MDTs */
1225         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1226                 struct llog_ctxt        *ctxt;
1227                 struct obd_device       *obd;
1228                 struct llog_cookie      *cookie;
1229                 struct sub_thandle_cookie *stc;
1230                 int rc;
1231
1232                 obd = st->st_dt->dd_lu_dev.ld_obd;
1233                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1234                 LASSERT(ctxt);
1235                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1236                         cookie = &stc->stc_cookie;
1237                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1238                                 continue;
1239
1240                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1241                                                      cookie);
1242                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1243                                DOSTID ".%u : rc = %d\n", obd->obd_name,
1244                                tmt->tmt_batchid,
1245                                POSTID(&cookie->lgc_lgl.lgl_oi),
1246                                cookie->lgc_index, rc);
1247                 }
1248
1249                 llog_ctxt_put(ctxt);
1250         }
1251
1252         RETURN(0);
1253 }
1254
1255 /**
1256  * Check if there are committed transaction
1257  *
1258  * Check if there are committed transaction in the distribute transaction
1259  * list, then cancel the update records for those committed transaction.
1260  * Because the distribute transaction in the list are sorted by batchid,
1261  * and cancellation will be done by batchid order, so we only check the first
1262  * the transaction(with lowest batchid) in the list.
1263  *
1264  * \param[in] lod       lod device where cancel thread is
1265  *
1266  * \retval              true if it is ready
1267  * \retval              false if it is not ready
1268  */
1269 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1270 {
1271         struct top_multiple_thandle     *tmt = NULL;
1272         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1273         bool    ready = false;
1274
1275         spin_lock(&tdtd->tdtd_batchid_lock);
1276         if (!list_empty(&tdtd->tdtd_list)) {
1277                 tmt = list_entry(tdtd->tdtd_list.next,
1278                                  struct top_multiple_thandle, tmt_commit_list);
1279                 if (tmt->tmt_committed &&
1280                     (!obd->obd_recovering || (obd->obd_recovering &&
1281                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1282                         ready = true;
1283         }
1284         spin_unlock(&tdtd->tdtd_batchid_lock);
1285
1286         return ready;
1287 }
1288
1289 struct distribute_txn_bid_data {
1290         struct dt_txn_commit_cb  dtbd_cb;
1291         struct target_distribute_txn_data      *dtbd_tdtd;
1292         __u64                    dtbd_batchid;
1293 };
1294
1295 /**
1296  * callback of updating commit batchid
1297  *
1298  * Updating commit batchid then wake up the commit thread to cancel the
1299  * records.
1300  *
1301  * \param[in]env        execution environment
1302  * \param[in]th         thandle to updating commit batchid
1303  * \param[in]cb         commit callback
1304  * \param[in]err        result of thandle
1305  */
1306 static void distribute_txn_batchid_cb(struct lu_env *env,
1307                                       struct thandle *th,
1308                                       struct dt_txn_commit_cb *cb,
1309                                       int err)
1310 {
1311         struct distribute_txn_bid_data          *dtbd = NULL;
1312         struct target_distribute_txn_data       *tdtd;
1313
1314         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1315         tdtd = dtbd->dtbd_tdtd;
1316
1317         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1318               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1319         spin_lock(&tdtd->tdtd_batchid_lock);
1320         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1321             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1322                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1323         spin_unlock(&tdtd->tdtd_batchid_lock);
1324         atomic_dec(&tdtd->tdtd_refcount);
1325         wake_up(&tdtd->tdtd_commit_thread_waitq);
1326
1327         OBD_FREE_PTR(dtbd);
1328 }
1329
1330 /**
1331  * Update the commit batchid in disk
1332  *
1333  * Update commit batchid in the disk, after this is committed, it can start
1334  * to cancel the update records.
1335  *
1336  * \param[in] env       execution environment
1337  * \param[in] tdtd      distribute transaction structure
1338  * \param[in] batchid   commit batchid to be updated
1339  *
1340  * \retval              0 if update succeeds.
1341  * \retval              negative errno if update fails.
1342  */
1343 static int
1344 distribute_txn_commit_batchid_update(const struct lu_env *env,
1345                               struct target_distribute_txn_data *tdtd,
1346                               __u64 batchid)
1347 {
1348         struct distribute_txn_bid_data  *dtbd = NULL;
1349         struct thandle          *th;
1350         struct lu_buf            buf;
1351         __u64                    tmp;
1352         __u64                    off;
1353         int                      rc;
1354         ENTRY;
1355
1356         OBD_ALLOC_PTR(dtbd);
1357         if (dtbd == NULL)
1358                 RETURN(-ENOMEM);
1359         dtbd->dtbd_batchid = batchid;
1360         dtbd->dtbd_tdtd = tdtd;
1361         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1362         atomic_inc(&tdtd->tdtd_refcount);
1363
1364         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1365         if (IS_ERR(th)) {
1366                 OBD_FREE_PTR(dtbd);
1367                 RETURN(PTR_ERR(th));
1368         }
1369
1370         tmp = cpu_to_le64(batchid);
1371         buf.lb_buf = &tmp;
1372         buf.lb_len = sizeof(tmp);
1373         off = 0;
1374
1375         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1376                                      th);
1377         if (rc < 0)
1378                 GOTO(stop, rc);
1379
1380         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1381         if (rc < 0)
1382                 GOTO(stop, rc);
1383
1384         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1385         if (rc < 0)
1386                 GOTO(stop, rc);
1387
1388         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1389                              &off, th);
1390
1391         CDEBUG(D_INFO, "%s: update batchid "LPU64": rc = %d\n",
1392                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1393
1394 stop:
1395         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1396         if (rc < 0)
1397                 OBD_FREE_PTR(dtbd);
1398         RETURN(rc);
1399 }
1400
1401 /**
1402  * Init commit batchid for distribute transaction.
1403  *
1404  * Initialize the batchid object and get commit batchid from the object.
1405  *
1406  * \param[in] env       execution environment
1407  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1408  *
1409  * \retval              0 if initialization succeeds.
1410  * \retval              negative errno if initialization fails.
1411  **/
1412 static int
1413 distribute_txn_commit_batchid_init(const struct lu_env *env,
1414                                    struct target_distribute_txn_data *tdtd)
1415 {
1416         struct tgt_thread_info  *tti = tgt_th_info(env);
1417         struct lu_target        *lut = tdtd->tdtd_lut;
1418         struct lu_attr          *attr = &tti->tti_attr;
1419         struct lu_fid           *fid = &tti->tti_fid1;
1420         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1421         struct dt_object        *dt_obj = NULL;
1422         struct lu_buf           buf;
1423         __u64                   tmp;
1424         __u64                   off;
1425         int                     rc;
1426         ENTRY;
1427
1428         memset(attr, 0, sizeof(*attr));
1429         attr->la_valid = LA_MODE;
1430         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1431         dof->dof_type = dt_mode_to_dft(S_IFREG);
1432
1433         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1434
1435         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1436                                    attr);
1437         if (IS_ERR(dt_obj)) {
1438                 rc = PTR_ERR(dt_obj);
1439                 dt_obj = NULL;
1440                 GOTO(out_put, rc);
1441         }
1442
1443         tdtd->tdtd_batchid_obj = dt_obj;
1444
1445         buf.lb_buf = &tmp;
1446         buf.lb_len = sizeof(tmp);
1447         off = 0;
1448         rc = dt_read(env, dt_obj, &buf, &off);
1449         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1450                 CERROR("%s can't read last committed batchid: rc = %d\n",
1451                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1452                 if (rc > 0)
1453                         rc = -EINVAL;
1454                 GOTO(out_put, rc);
1455         } else if (rc == buf.lb_len) {
1456                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1457                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1458                        tdtd->tdtd_lut->lut_obd->obd_name,
1459                        tdtd->tdtd_committed_batchid);
1460                 rc = 0;
1461         }
1462
1463 out_put:
1464         if (rc < 0 && dt_obj != NULL) {
1465                 lu_object_put(env, &dt_obj->do_lu);
1466                 tdtd->tdtd_batchid_obj = NULL;
1467         }
1468         return rc;
1469 }
1470
1471 /**
1472  * manage the distribute transaction thread
1473  *
1474  * Distribute transaction are linked to the list, and once the distribute
1475  * transaction is committed, it will update the last committed batchid first,
1476  * after it is committed, it will cancel the records.
1477  *
1478  * \param[in] _arg      argument for commit thread
1479  *
1480  * \retval              0 if thread is running successfully
1481  * \retval              negative errno if the thread can not be run.
1482  */
1483 static int distribute_txn_commit_thread(void *_arg)
1484 {
1485         struct target_distribute_txn_data *tdtd = _arg;
1486         struct lu_target        *lut = tdtd->tdtd_lut;
1487         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1488         struct l_wait_info       lwi = { 0 };
1489         struct lu_env            env;
1490         struct list_head         list;
1491         int                      rc;
1492         struct top_multiple_thandle *tmt;
1493         struct top_multiple_thandle *tmp;
1494         __u64                    batchid = 0, committed;
1495
1496         ENTRY;
1497
1498         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1499         if (rc != 0)
1500                 RETURN(rc);
1501
1502         spin_lock(&tdtd->tdtd_batchid_lock);
1503         thread->t_flags = SVC_RUNNING;
1504         spin_unlock(&tdtd->tdtd_batchid_lock);
1505         wake_up(&thread->t_ctl_waitq);
1506         INIT_LIST_HEAD(&list);
1507
1508         CDEBUG(D_HA, "%s: start commit thread committed batchid "LPU64"\n",
1509                tdtd->tdtd_lut->lut_obd->obd_name,
1510                tdtd->tdtd_committed_batchid);
1511
1512         while (distribute_txn_commit_thread_running(lut)) {
1513                 spin_lock(&tdtd->tdtd_batchid_lock);
1514                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1515                                          tmt_commit_list) {
1516                         if (tmt->tmt_committed == 0)
1517                                 break;
1518
1519                         /* Note: right now, replay is based on master MDT
1520                          * transno, but cancellation is based on batchid.
1521                          * so we do not try to cancel the update log until
1522                          * the recoverying is done, unless the update records
1523                          * batchid < committed_batchid. */
1524                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1525                                 list_move_tail(&tmt->tmt_commit_list, &list);
1526                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1527                                 LASSERTF(tmt->tmt_batchid >= batchid,
1528                                          "tmt %p tmt_batchid: "LPU64", batchid "
1529                                           LPU64"\n", tmt, tmt->tmt_batchid,
1530                                          batchid);
1531                                 /* There are three types of distribution
1532                                  * transaction result
1533                                  *
1534                                  * 1. If tmt_result < 0, it means the
1535                                  * distribution transaction fails, which should
1536                                  * be rare, because once declare phase succeeds,
1537                                  * the operation should succeeds anyway. Note in
1538                                  * this case, we will still update batchid so
1539                                  * cancellation would be stopped.
1540                                  *
1541                                  * 2. If tmt_result == 0, it means the
1542                                  * distribution transaction succeeds, and we
1543                                  * will update batchid.
1544                                  *
1545                                  * 3. If tmt_result > 0, it means distribute
1546                                  * transaction is not yet committed on every
1547                                  * node, but we need release this tmt before
1548                                  * that, which usuually happens during umount.
1549                                  */
1550                                 if (tmt->tmt_result <= 0)
1551                                         batchid = tmt->tmt_batchid;
1552                                 list_move_tail(&tmt->tmt_commit_list, &list);
1553                         }
1554                 }
1555                 spin_unlock(&tdtd->tdtd_batchid_lock);
1556
1557                 CDEBUG(D_HA, "%s: batchid: "LPU64" committed batchid "
1558                        LPU64"\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1559                        tdtd->tdtd_committed_batchid);
1560                 /* update globally committed on a storage */
1561                 if (batchid > tdtd->tdtd_committed_batchid) {
1562                         distribute_txn_commit_batchid_update(&env, tdtd,
1563                                                              batchid);
1564                         spin_lock(&tdtd->tdtd_batchid_lock);
1565                         if (batchid > tdtd->tdtd_batchid) {
1566                                 /* This might happen during recovery,
1567                                  * batchid is initialized as last transno,
1568                                  * and the batchid in the update records
1569                                  * on other MDTs might be bigger than
1570                                  * the batchid, so we need update it to
1571                                  * avoid duplicate batchid. */
1572                                 CDEBUG(D_HA, "%s update batchid from "LPU64
1573                                        " to "LPU64"\n",
1574                                        tdtd->tdtd_lut->lut_obd->obd_name,
1575                                        tdtd->tdtd_batchid, batchid);
1576                                 tdtd->tdtd_batchid = batchid;
1577                         }
1578                         spin_unlock(&tdtd->tdtd_batchid_lock);
1579                 }
1580                 /* cancel the records for committed batchid's */
1581                 /* XXX: should we postpone cancel's till the end of recovery? */
1582                 committed = tdtd->tdtd_committed_batchid;
1583                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1584                         if (tmt->tmt_batchid > committed)
1585                                 break;
1586                         list_del_init(&tmt->tmt_commit_list);
1587                         if (tmt->tmt_result <= 0)
1588                                 distribute_txn_cancel_records(&env, tmt);
1589                         top_multiple_thandle_put(tmt);
1590                 }
1591
1592                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1593                              !distribute_txn_commit_thread_running(lut) ||
1594                              committed < tdtd->tdtd_committed_batchid ||
1595                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1596         };
1597
1598         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1599                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1600
1601         spin_lock(&tdtd->tdtd_batchid_lock);
1602         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1603                                  tmt_commit_list)
1604                 list_move_tail(&tmt->tmt_commit_list, &list);
1605         spin_unlock(&tdtd->tdtd_batchid_lock);
1606
1607         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1608                tdtd->tdtd_lut->lut_obd->obd_name);
1609         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1610                 list_del_init(&tmt->tmt_commit_list);
1611                 top_multiple_thandle_dump(tmt, D_HA);
1612                 top_multiple_thandle_put(tmt);
1613         }
1614
1615         thread->t_flags = SVC_STOPPED;
1616         lu_env_fini(&env);
1617         wake_up(&thread->t_ctl_waitq);
1618
1619         RETURN(0);
1620 }
1621
1622 /**
1623  * Start llog cancel thread
1624  *
1625  * Start llog cancel(master/slave) thread on LOD
1626  *
1627  * \param[in]lclt       cancel log thread to be started.
1628  *
1629  * \retval              0 if the thread is started successfully.
1630  * \retval              negative errno if the thread is not being
1631  *                      started.
1632  */
1633 int distribute_txn_init(const struct lu_env *env,
1634                         struct lu_target *lut,
1635                         struct target_distribute_txn_data *tdtd,
1636                         __u32 index)
1637 {
1638         struct task_struct      *task;
1639         struct l_wait_info       lwi = { 0 };
1640         int                     rc;
1641         ENTRY;
1642
1643         spin_lock_init(&tdtd->tdtd_batchid_lock);
1644         INIT_LIST_HEAD(&tdtd->tdtd_list);
1645
1646         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1647
1648         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1649         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1650         atomic_set(&tdtd->tdtd_refcount, 0);
1651
1652         tdtd->tdtd_lut = lut;
1653         rc = distribute_txn_commit_batchid_init(env, tdtd);
1654         if (rc != 0)
1655                 RETURN(rc);
1656
1657         task = kthread_run(distribute_txn_commit_thread, tdtd, "tdtd-%u",
1658                            index);
1659         if (IS_ERR(task))
1660                 RETURN(PTR_ERR(task));
1661
1662         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1663                      distribute_txn_commit_thread_running(lut) ||
1664                      distribute_txn_commit_thread_stopped(lut), &lwi);
1665         RETURN(0);
1666 }
1667 EXPORT_SYMBOL(distribute_txn_init);
1668
1669 /**
1670  * Stop llog cancel thread
1671  *
1672  * Stop llog cancel(master/slave) thread on LOD and also destory
1673  * all of transaction in the list.
1674  *
1675  * \param[in]lclt       cancel log thread to be stopped.
1676  */
1677 void distribute_txn_fini(const struct lu_env *env,
1678                          struct target_distribute_txn_data *tdtd)
1679 {
1680         struct lu_target *lut = tdtd->tdtd_lut;
1681
1682         /* Stop cancel thread */
1683         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1684                 return;
1685
1686         spin_lock(&tdtd->tdtd_batchid_lock);
1687         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1688         spin_unlock(&tdtd->tdtd_batchid_lock);
1689         wake_up(&tdtd->tdtd_commit_thread_waitq);
1690         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1691                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1692
1693         dtrq_list_destroy(tdtd);
1694         if (tdtd->tdtd_batchid_obj != NULL) {
1695                 lu_object_put(env, &tdtd->tdtd_batchid_obj->do_lu);
1696                 tdtd->tdtd_batchid_obj = NULL;
1697         }
1698 }
1699 EXPORT_SYMBOL(distribute_txn_fini);