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
 126         /* Not ready to record updates yet. */
 127         if (ctxt == NULL || ctxt->loc_handle == NULL) {
 128                 llog_ctxt_put(ctxt);
 129                 return 0;
 130         }
 131
 132         rc = llog_declare_add(env, ctxt->loc_handle,
 133                               &record->lur_hdr, sub_th);
 134         if (rc < 0)
 135                 GOTO(out_put, rc);
 136
 137         while (left > ctxt->loc_chunk_size) {
 138                 rc = llog_declare_add(env, ctxt->loc_handle,
 139                                       &record->lur_hdr, sub_th);
 140                 if (rc < 0)
 141                         GOTO(out_put, rc);
 142
 143                 left -= ctxt->loc_chunk_size;
 144         }
 145
 146 out_put:
 147         llog_ctxt_put(ctxt);
 148
 149         return rc;
 150 }
 151
 152 /**
 153  * write update to sub device
 154  *
 155  * Write llog update record to the sub device during distribute
 156  * transaction. If it succeeds, llog cookie of the record will be
 157  * returned by @cookie.
 158  *
 159  * \param[in] env       execution environment
 160  * \param[in] record    update records being written
 161  * \param[in] sub_th    sub transaction handle
 162  * \param[out] cookie   llog cookie of the update record.
 163  *
 164  * \retval              1 if writing succeeds
 165  * \retval              negative errno if writing fails
 166  */
 167 static int sub_updates_write(const struct lu_env *env,
 168                              struct llog_update_record *record,
 169                              struct sub_thandle *sub_th)
 170 {
 171         struct dt_device        *dt = sub_th->st_dt;
 172         struct llog_ctxt        *ctxt;
 173         int                     rc;
 174         struct llog_update_record *lur = NULL;
 175         struct update_params    *params = NULL;
 176         __u32                   update_count = 0;
 177         __u32                   param_count = 0;
 178         __u32                   last_update_count = 0;
 179         __u32                   last_param_count = 0;
 180         void                    *src;
 181         void                    *start;
 182         void                    *next;
 183         struct sub_thandle_cookie *stc;
 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         if (likely(record->lur_hdr.lrh_len <= ctxt->loc_chunk_size)) {
 205                 OBD_ALLOC_PTR(stc);
 206                 if (stc == NULL)
 207                         GOTO(llog_put, rc = -ENOMEM);
 208                 INIT_LIST_HEAD(&stc->stc_list);
 209
 210                 rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
 211                               &stc->stc_cookie, sub_th->st_sub_th);
 212
 213                 CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u: rc = %d\n",
 214                        dt->dd_lu_dev.ld_obd->obd_name,
 215                        POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
 216                        stc->stc_cookie.lgc_index, rc);
 217
 218                 if (rc > 0) {
 219                         list_add(&stc->stc_list, &sub_th->st_cookie_list);
 220                         rc = 0;
 221                 } else {
 222                         OBD_FREE_PTR(stc);
 223                 }
 224
 225                 GOTO(llog_put, rc);
 226         }
 227
 228         /* Split the records into chunk_size update record */
 229         OBD_ALLOC_LARGE(lur, ctxt->loc_chunk_size);
 230         if (lur == NULL)
 231                 GOTO(llog_put, rc = -ENOMEM);
 232
 233         memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
 234         lur->lur_update_rec.ur_update_count = 0;
 235         lur->lur_update_rec.ur_param_count = 0;
 236         src = &record->lur_update_rec.ur_ops;
 237         start = next = src;
 238         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 239         params = update_records_get_params(&record->lur_update_rec);
 240         do {
 241                 size_t rec_len;
 242
 243                 if (update_count < record->lur_update_rec.ur_update_count) {
 244                         next = update_op_next_op((struct update_op *)src);
 245                 } else {
 246                         if (param_count == 0)
 247                                 next = update_records_get_params(
 248                                                 &record->lur_update_rec);
 249                         else
 250                                 next = (char *)src +
 251                                         object_update_param_size(
 252                                         (struct object_update_param *)src);
 253                 }
 254
 255                 rec_len = cfs_size_round((unsigned long)(next - src));
 256                 /* If its size > llog chunk_size, then write current chunk to
 257                  * the update llog. */
 258                 if (lur->lur_hdr.lrh_len + rec_len + LLOG_MIN_REC_SIZE >
 259                     ctxt->loc_chunk_size ||
 260                     param_count == record->lur_update_rec.ur_param_count) {
 261                         lur->lur_update_rec.ur_update_count =
 262                                 update_count > last_update_count ?
 263                                 update_count - last_update_count : 0;
 264                         lur->lur_update_rec.ur_param_count = param_count -
 265                                                              last_param_count;
 266
 267                         memcpy(&lur->lur_update_rec.ur_ops, start,
 268                                (unsigned long)(src - start));
 269                         if (last_update_count != 0)
 270                                 lur->lur_update_rec.ur_flags |=
 271                                                 UPDATE_RECORD_CONTINUE;
 272
 273                         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 274                         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 275                         LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
 276
 277                         OBD_ALLOC_PTR(stc);
 278                         if (stc == NULL)
 279                                 GOTO(llog_put, rc = -ENOMEM);
 280                         INIT_LIST_HEAD(&stc->stc_list);
 281
 282                         rc = llog_add(env, ctxt->loc_handle,
 283                                       &lur->lur_hdr,
 284                                       &stc->stc_cookie, sub_th->st_sub_th);
 285
 286                         CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u"
 287                                " rc = %d\n", dt->dd_lu_dev.ld_obd->obd_name,
 288                                POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
 289                                stc->stc_cookie.lgc_index, rc);
 290
 291                         if (rc > 0) {
 292                                 list_add(&stc->stc_list,
 293                                          &sub_th->st_cookie_list);
 294                                 rc = 0;
 295                         } else {
 296                                 OBD_FREE_PTR(stc);
 297                                 GOTO(llog_put, rc);
 298                         }
 299
 300                         last_update_count = update_count;
 301                         last_param_count = param_count;
 302                         start = src;
 303                         lur->lur_update_rec.ur_update_count = 0;
 304                         lur->lur_update_rec.ur_param_count = 0;
 305                         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 306                 }
 307
 308                 src = next;
 309                 lur->lur_hdr.lrh_len += cfs_size_round(rec_len);
 310                 if (update_count < record->lur_update_rec.ur_update_count)
 311                         update_count++;
 312                 else if (param_count < record->lur_update_rec.ur_param_count)
 313                         param_count++;
 314                 else
 315                         break;
 316         } while (1);
 317
 318 llog_put:
 319         if (lur != NULL)
 320                 OBD_FREE_LARGE(lur, ctxt->loc_chunk_size);
 321         llog_ctxt_put(ctxt);
 322
 323         RETURN(rc);
 324 }
 325
 326 /**
 327  * Prepare the update records.
 328  *
 329  * Merge params and ops into the update records, then initializing
 330  * the update buffer.
 331  *
 332  * During transaction execution phase, parameters and update ops
 333  * are collected in two different buffers (see lod_updates_pack()),
 334  * during transaction stop, it needs to be merged in one buffer,
 335  * so it will be written in the update log.
 336  *
 337  * \param[in] env       execution environment
 338  * \param[in] tmt       top_multiple_thandle for distribute txn
 339  *
 340  * \retval              0 if merging succeeds.
 341  * \retval              negaitive errno if merging fails.
 342  */
 343 static int prepare_writing_updates(const struct lu_env *env,
 344                                    struct top_multiple_thandle *tmt)
 345 {
 346         struct thandle_update_records   *tur = tmt->tmt_update_records;
 347         struct llog_update_record       *lur;
 348         struct update_params *params;
 349         size_t params_size;
 350         size_t update_size;
 351
 352         if (tur == NULL || tur->tur_update_records == NULL ||
 353             tur->tur_update_params == NULL)
 354                 return 0;
 355
 356         lur = tur->tur_update_records;
 357         /* Extends the update records buffer if needed */
 358         params_size = update_params_size(tur->tur_update_params,
 359                                          tur->tur_update_param_count);
 360         LASSERT(lur->lur_update_rec.ur_param_count == 0);
 361         update_size = llog_update_record_size(lur);
 362         if (cfs_size_round(update_size + params_size) >
 363             tur->tur_update_records_buf_size) {
 364                 int rc;
 365
 366                 rc = tur_update_records_extend(tur,
 367                         cfs_size_round(update_size + params_size));
 368                 if (rc < 0)
 369                         return rc;
 370
 371                 lur = tur->tur_update_records;
 372         }
 373
 374         params = update_records_get_params(&lur->lur_update_rec);
 375         memcpy(params, tur->tur_update_params, params_size);
 376
 377         lur->lur_update_rec.ur_param_count = tur->tur_update_param_count;
 378         lur->lur_update_rec.ur_batchid = tmt->tmt_batchid;
 379         /* Init update record header */
 380         lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 381         lur->lur_hdr.lrh_type = UPDATE_REC;
 382
 383         /* Dump updates for debugging purpose */
 384         update_records_dump(&lur->lur_update_rec, D_INFO, true);
 385
 386         return 0;
 387 }
 388
 389 static inline int
 390 distribute_txn_commit_thread_running(struct lu_target *lut)
 391 {
 392         return lut->lut_tdtd_commit_thread.t_flags & SVC_RUNNING;
 393 }
 394
 395 static inline int
 396 distribute_txn_commit_thread_stopped(struct lu_target *lut)
 397 {
 398         return lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED;
 399 }
 400
 401 /**
 402  * Top thandle commit callback
 403  *
 404  * This callback will be called when all of sub transactions are committed.
 405  *
 406  * \param[in] th        top thandle to be committed.
 407  */
 408 static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
 409 {
 410         struct lu_target *lut;
 411         ENTRY;
 412
 413         LASSERT(atomic_read(&tmt->tmt_refcount) > 0);
 414
 415         top_multiple_thandle_dump(tmt, D_HA);
 416         tmt->tmt_committed = 1;
 417         lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
 418         if (distribute_txn_commit_thread_running(lut))
 419                 wake_up(&lut->lut_tdtd->tdtd_commit_thread_waitq);
 420         RETURN_EXIT;
 421 }
 422
 423 struct sub_thandle *lookup_sub_thandle(struct top_multiple_thandle *tmt,
 424                                        struct dt_device *dt_dev)
 425 {
 426         struct sub_thandle *st;
 427
 428         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 429                 if (st->st_dt == dt_dev)
 430                         return st;
 431         }
 432         return NULL;
 433 }
 434 EXPORT_SYMBOL(lookup_sub_thandle);
 435
 436 struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
 437                                        struct dt_device *dt_dev)
 438 {
 439         struct sub_thandle *st;
 440
 441         OBD_ALLOC_PTR(st);
 442         if (st == NULL)
 443                 RETURN(ERR_PTR(-ENOMEM));
 444
 445         INIT_LIST_HEAD(&st->st_sub_list);
 446         INIT_LIST_HEAD(&st->st_cookie_list);
 447         st->st_dt = dt_dev;
 448
 449         list_add(&st->st_sub_list, &tmt->tmt_sub_thandle_list);
 450         return st;
 451 }
 452
 453 /**
 454  * sub thandle commit callback
 455  *
 456  * Mark the sub thandle to be committed and if all sub thandle are committed
 457  * notify the top thandle.
 458  *
 459  * \param[in] env       execution environment
 460  * \param[in] sub_th    sub thandle being committed
 461  * \param[in] cb        commit callback
 462  * \param[in] err       trans result
 463  */
 464 static void sub_trans_commit_cb(struct lu_env *env,
 465                                 struct thandle *sub_th,
 466                                 struct dt_txn_commit_cb *cb, int err)
 467 {
 468         struct sub_thandle      *st;
 469         struct top_multiple_thandle *tmt = cb->dcb_data;
 470         bool                    all_committed = true;
 471         ENTRY;
 472
 473         /* Check if all sub thandles are committed */
 474         spin_lock(&tmt->tmt_sub_lock);
 475         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 476                 if (st->st_sub_th == sub_th) {
 477                         st->st_committed = 1;
 478                         st->st_result = err;
 479                 }
 480                 if (!st->st_committed)
 481                         all_committed = false;
 482         }
 483         spin_unlock(&tmt->tmt_sub_lock);
 484
 485         if (tmt->tmt_result == 0)
 486                 tmt->tmt_result = err;
 487
 488         if (all_committed)
 489                 top_trans_committed_cb(tmt);
 490
 491         top_multiple_thandle_dump(tmt, D_INFO);
 492         top_multiple_thandle_put(tmt);
 493         RETURN_EXIT;
 494 }
 495
 496 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
 497                                     struct top_multiple_thandle *tmt)
 498 {
 499         LASSERT(st->st_sub_th != NULL);
 500         top_multiple_thandle_get(tmt);
 501         st->st_commit_dcb.dcb_func = sub_trans_commit_cb;
 502         st->st_commit_dcb.dcb_data = tmt;
 503         INIT_LIST_HEAD(&st->st_commit_dcb.dcb_linkage);
 504         dt_trans_cb_add(st->st_sub_th, &st->st_commit_dcb);
 505 }
 506
 507 /**
 508  * Sub thandle stop call back
 509  *
 510  * After sub thandle is stopped, it will call this callback to notify
 511  * the top thandle.
 512  *
 513  * \param[in] th        sub thandle to be stopped
 514  * \param[in] rc        result of sub trans
 515  */
 516 static void sub_trans_stop_cb(struct lu_env *env,
 517                               struct thandle *sub_th,
 518                               struct dt_txn_commit_cb *cb, int err)
 519 {
 520         struct sub_thandle              *st;
 521         struct top_multiple_thandle     *tmt = cb->dcb_data;
 522         ENTRY;
 523
 524         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 525                 if (st->st_stopped)
 526                         continue;
 527
 528                 if (st->st_dt == sub_th->th_dev) {
 529                         st->st_stopped = 1;
 530                         st->st_result = err;
 531                         break;
 532                 }
 533         }
 534
 535         wake_up(&tmt->tmt_stop_waitq);
 536         RETURN_EXIT;
 537 }
 538
 539 static void sub_thandle_register_stop_cb(struct sub_thandle *st,
 540                                          struct top_multiple_thandle *tmt)
 541 {
 542         st->st_stop_dcb.dcb_func = sub_trans_stop_cb;
 543         st->st_stop_dcb.dcb_data = tmt;
 544         st->st_stop_dcb.dcb_flags = DCB_TRANS_STOP;
 545         INIT_LIST_HEAD(&st->st_stop_dcb.dcb_linkage);
 546         dt_trans_cb_add(st->st_sub_th, &st->st_stop_dcb);
 547 }
 548
 549 /**
 550  * Create sub thandle
 551  *
 552  * Create transaction handle for sub_thandle
 553  *
 554  * \param[in] env       execution environment
 555  * \param[in] th        top thandle
 556  * \param[in] st        sub_thandle
 557  *
 558  * \retval              0 if creation succeeds.
 559  * \retval              negative errno if creation fails.
 560  */
 561 int sub_thandle_trans_create(const struct lu_env *env,
 562                              struct top_thandle *top_th,
 563                              struct sub_thandle *st)
 564 {
 565         struct thandle *sub_th;
 566
 567         sub_th = dt_trans_create(env, st->st_dt);
 568         if (IS_ERR(sub_th))
 569                 return PTR_ERR(sub_th);
 570
 571         sub_th->th_top = &top_th->tt_super;
 572         st->st_sub_th = sub_th;
 573
 574         sub_th->th_wait_submit = 1;
 575         return 0;
 576 }
 577
 578 /**
 579  * Create the top transaction.
 580  *
 581  * Create the top transaction on the master device. It will create a top
 582  * thandle and a sub thandle on the master device.
 583  *
 584  * \param[in] env               execution environment
 585  * \param[in] master_dev        master_dev the top thandle will be created
 586  *
 587  * \retval                      pointer to the created thandle.
 588  * \retval                      ERR_PTR(errno) if creation failed.
 589  */
 590 struct thandle *
 591 top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
 592 {
 593         struct top_thandle      *top_th;
 594         struct thandle          *child_th;
 595
 596         OBD_ALLOC_GFP(top_th, sizeof(*top_th), __GFP_IO);
 597         if (top_th == NULL)
 598                 return ERR_PTR(-ENOMEM);
 599
 600         top_th->tt_super.th_top = &top_th->tt_super;
 601
 602         if (master_dev != NULL) {
 603                 child_th = dt_trans_create(env, master_dev);
 604                 if (IS_ERR(child_th)) {
 605                         OBD_FREE_PTR(top_th);
 606                         return child_th;
 607                 }
 608
 609                 child_th->th_top = &top_th->tt_super;
 610                 child_th->th_wait_submit = 1;
 611                 top_th->tt_master_sub_thandle = child_th;
 612
 613                 top_th->tt_super.th_tags |= child_th->th_tags;
 614         }
 615         return &top_th->tt_super;
 616 }
 617 EXPORT_SYMBOL(top_trans_create);
 618
 619 /**
 620  * Declare write update transaction
 621  *
 622  * Check if there are updates being recorded in this transaction,
 623  * it will write the record into the disk.
 624  *
 625  * \param[in] env       execution environment
 626  * \param[in] tmt       top multiple transaction handle
 627  *
 628  * \retval              0 if writing succeeds
 629  * \retval              negative errno if writing fails
 630  */
 631 static int declare_updates_write(const struct lu_env *env,
 632                                  struct top_multiple_thandle *tmt)
 633 {
 634         struct llog_update_record *record;
 635         struct sub_thandle *st;
 636         int rc = 0;
 637
 638         record = tmt->tmt_update_records->tur_update_records;
 639         /* Declare update write for all other target */
 640         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 641                 if (st->st_sub_th == NULL)
 642                         continue;
 643
 644                 rc = sub_declare_updates_write(env, record, st->st_sub_th,
 645                                                tmt->tmt_record_size);
 646                 if (rc < 0)
 647                         break;
 648         }
 649
 650         return rc;
 651 }
 652
 653 /**
 654  * Assign batchid to the distribute transaction.
 655  *
 656  * Assign batchid to the distribute transaction
 657  *
 658  * \param[in] tmt       distribute transaction
 659  */
 660 static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 661 {
 662         struct target_distribute_txn_data *tdtd;
 663         struct dt_device *dt = new->tmt_master_sub_dt;
 664
 665         LASSERT(dt != NULL);
 666         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 667         spin_lock(&tdtd->tdtd_batchid_lock);
 668         new->tmt_batchid = tdtd->tdtd_batchid++;
 669         list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
 670         spin_unlock(&tdtd->tdtd_batchid_lock);
 671         top_multiple_thandle_get(new);
 672         top_multiple_thandle_dump(new, D_INFO);
 673 }
 674
 675 /**
 676  * Insert distribute transaction to the distribute txn list.
 677  *
 678  * Insert distribute transaction to the distribute txn list.
 679  *
 680  * \param[in] new       the distribute txn to be inserted.
 681  */
 682 void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
 683 {
 684         struct dt_device *dt = new->tmt_master_sub_dt;
 685         struct top_multiple_thandle *tmt;
 686         struct target_distribute_txn_data *tdtd;
 687         bool    at_head = false;
 688
 689         LASSERT(dt != NULL);
 690         tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
 691
 692         spin_lock(&tdtd->tdtd_batchid_lock);
 693         list_for_each_entry_reverse(tmt, &tdtd->tdtd_list, tmt_commit_list) {
 694                 if (new->tmt_batchid > tmt->tmt_batchid) {
 695                         list_add(&new->tmt_commit_list, &tmt->tmt_commit_list);
 696                         break;
 697                 }
 698         }
 699         if (list_empty(&new->tmt_commit_list)) {
 700                 at_head = true;
 701                 list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
 702         }
 703         spin_unlock(&tdtd->tdtd_batchid_lock);
 704         top_multiple_thandle_get(new);
 705         top_multiple_thandle_dump(new, D_INFO);
 706         if (new->tmt_committed && at_head)
 707                 wake_up(&tdtd->tdtd_commit_thread_waitq);
 708 }
 709
 710 /**
 711  * Prepare cross-MDT operation.
 712  *
 713  * Create the update record buffer to record updates for cross-MDT operation,
 714  * add master sub transaction to tt_sub_trans_list, and declare the update
 715  * writes.
 716  *
 717  * During updates packing, all of parameters will be packed in
 718  * tur_update_params, and updates will be packed in tur_update_records.
 719  * Then in transaction stop, parameters and updates will be merged
 720  * into one updates buffer.
 721  *
 722  * And also master thandle will be added to the sub_th list, so it will be
 723  * easy to track the commit status.
 724  *
 725  * \param[in] env       execution environment
 726  * \param[in] th        top transaction handle
 727  *
 728  * \retval              0 if preparation succeeds.
 729  * \retval              negative errno if preparation fails.
 730  */
 731 static int prepare_multiple_node_trans(const struct lu_env *env,
 732                                        struct top_multiple_thandle *tmt)
 733 {
 734         struct thandle_update_records   *tur;
 735         int                             rc;
 736         ENTRY;
 737
 738         if (tmt->tmt_update_records == NULL) {
 739                 tur = &update_env_info(env)->uti_tur;
 740                 rc = check_and_prepare_update_record(env, tur);
 741                 if (rc < 0)
 742                         RETURN(rc);
 743
 744                 tmt->tmt_update_records = tur;
 745                 distribute_txn_assign_batchid(tmt);
 746         }
 747
 748         rc = declare_updates_write(env, tmt);
 749
 750         RETURN(rc);
 751 }
 752
 753 /**
 754  * start the top transaction.
 755  *
 756  * Start all of its sub transactions, then start master sub transaction.
 757  *
 758  * \param[in] env               execution environment
 759  * \param[in] master_dev        master_dev the top thandle will be start
 760  * \param[in] th                top thandle
 761  *
 762  * \retval                      0 if transaction start succeeds.
 763  * \retval                      negative errno if start fails.
 764  */
 765 int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
 766                     struct thandle *th)
 767 {
 768         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 769                                                        tt_super);
 770         struct sub_thandle              *st;
 771         struct top_multiple_thandle     *tmt = top_th->tt_multiple_thandle;
 772         int                             rc = 0;
 773         ENTRY;
 774
 775         if (tmt == NULL) {
 776                 if (th->th_sync)
 777                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 778                 if (th->th_local)
 779                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 780                 top_th->tt_master_sub_thandle->th_tags = th->th_tags;
 781                 rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
 782                                     top_th->tt_master_sub_thandle);
 783                 RETURN(rc);
 784         }
 785
 786         tmt = top_th->tt_multiple_thandle;
 787         rc = prepare_multiple_node_trans(env, tmt);
 788         if (rc < 0)
 789                 RETURN(rc);
 790
 791         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 792                 if (st->st_sub_th == NULL)
 793                         continue;
 794                 if (th->th_sync)
 795                         st->st_sub_th->th_sync = th->th_sync;
 796                 if (th->th_local)
 797                         st->st_sub_th->th_local = th->th_local;
 798                 st->st_sub_th->th_tags = th->th_tags;
 799                 rc = dt_trans_start(env, st->st_sub_th->th_dev,
 800                                     st->st_sub_th);
 801                 if (rc != 0)
 802                         GOTO(out, rc);
 803
 804                 sub_thandle_register_stop_cb(st, tmt);
 805                 sub_thandle_register_commit_cb(st, tmt);
 806         }
 807 out:
 808         th->th_result = rc;
 809         RETURN(rc);
 810 }
 811 EXPORT_SYMBOL(top_trans_start);
 812
 813 /**
 814  * Check whether we need write updates record
 815  *
 816  * Check if the updates for the top_thandle needs to be writen
 817  * to all targets. Only if the transaction succeeds and the updates
 818  * number > 2, it will write the updates,
 819  *
 820  * \params [in] top_th  top thandle.
 821  *
 822  * \retval              true if it needs to write updates
 823  * \retval              false if it does not need to write updates
 824  **/
 825 static bool top_check_write_updates(struct top_thandle *top_th)
 826 {
 827         struct top_multiple_thandle     *tmt;
 828         struct thandle_update_records   *tur;
 829
 830         /* Do not write updates to records if the transaction fails */
 831         if (top_th->tt_super.th_result != 0)
 832                 return false;
 833
 834         tmt = top_th->tt_multiple_thandle;
 835         if (tmt == NULL)
 836                 return false;
 837
 838         tur = tmt->tmt_update_records;
 839         if (tur == NULL)
 840                 return false;
 841
 842         /* Hmm, false update records, since the cross-MDT operation
 843          * should includes both local and remote updates, so the
 844          * updates count should >= 2 */
 845         if (tur->tur_update_records == NULL ||
 846             tur->tur_update_records->lur_update_rec.ur_update_count <= 1)
 847                 return false;
 848
 849         return true;
 850 }
 851
 852 /**
 853  * Check if top transaction is stopped
 854  *
 855  * Check if top transaction is stopped, only if all sub transaction
 856  * is stopped, then the top transaction is stopped.
 857  *
 858  * \param [in] top_th   top thandle
 859  *
 860  * \retval              true if the top transaction is stopped.
 861  * \retval              false if the top transaction is not stopped.
 862  */
 863 static bool top_trans_is_stopped(struct top_thandle *top_th)
 864 {
 865         struct top_multiple_thandle     *tmt;
 866         struct sub_thandle              *st;
 867         bool                    all_stopped = true;
 868
 869         tmt = top_th->tt_multiple_thandle;
 870         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
 871                 if (!st->st_stopped && st->st_sub_th != NULL) {
 872                         all_stopped = false;
 873                         break;
 874                 }
 875
 876                 if (st->st_result != 0 &&
 877                     top_th->tt_super.th_result == 0)
 878                         top_th->tt_super.th_result = st->st_result;
 879         }
 880
 881         return all_stopped;
 882 }
 883
 884 /**
 885  * Wait result of top transaction
 886  *
 887  * Wait until all sub transaction get its result.
 888  *
 889  * \param [in] top_th   top thandle.
 890  *
 891  * \retval              the result of top thandle.
 892  */
 893 static int top_trans_wait_result(struct top_thandle *top_th)
 894 {
 895         struct l_wait_info      lwi = {0};
 896
 897         l_wait_event(top_th->tt_multiple_thandle->tmt_stop_waitq,
 898                      top_trans_is_stopped(top_th), &lwi);
 899
 900         RETURN(top_th->tt_super.th_result);
 901 }
 902
 903 /**
 904  * Stop the top transaction.
 905  *
 906  * Stop the transaction on the master device first, then stop transactions
 907  * on other sub devices.
 908  *
 909  * \param[in] env               execution environment
 910  * \param[in] master_dev        master_dev the top thandle will be created
 911  * \param[in] th                top thandle
 912  *
 913  * \retval                      0 if stop transaction succeeds.
 914  * \retval                      negative errno if stop transaction fails.
 915  */
 916 int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
 917                    struct thandle *th)
 918 {
 919         struct top_thandle      *top_th = container_of(th, struct top_thandle,
 920                                                        tt_super);
 921         struct sub_thandle              *st;
 922         struct sub_thandle              *master_st;
 923         struct top_multiple_thandle     *tmt;
 924         struct thandle_update_records   *tur;
 925         bool                            write_updates = false;
 926         int                     rc = 0;
 927         ENTRY;
 928
 929         if (likely(top_th->tt_multiple_thandle == NULL)) {
 930                 LASSERT(master_dev != NULL);
 931
 932                 if (th->th_sync)
 933                         top_th->tt_master_sub_thandle->th_sync = th->th_sync;
 934                 if (th->th_local)
 935                         top_th->tt_master_sub_thandle->th_local = th->th_local;
 936                 top_th->tt_master_sub_thandle->th_tags = th->th_tags;
 937                 rc = dt_trans_stop(env, master_dev,
 938                                    top_th->tt_master_sub_thandle);
 939                 OBD_FREE_PTR(top_th);
 940                 RETURN(rc);
 941         }
 942
 943         tmt = top_th->tt_multiple_thandle;
 944         tur = tmt->tmt_update_records;
 945
 946         /* Note: we need stop the master thandle first, then the stop
 947          * callback will fill the master transno in the update logs,
 948          * then these update logs will be sent to other MDTs */
 949         /* get the master sub thandle */
 950         master_st = lookup_sub_thandle(tmt, tmt->tmt_master_sub_dt);
 951         write_updates = top_check_write_updates(top_th);
 952
 953         /* Step 1: write the updates log on Master MDT */
 954         if (master_st != NULL && master_st->st_sub_th != NULL &&
 955             write_updates) {
 956                 struct llog_update_record *lur;
 957
 958                 /* Merge the parameters and updates into one buffer */
 959                 rc = prepare_writing_updates(env, tmt);
 960                 if (rc < 0) {
 961                         CERROR("%s: cannot prepare updates: rc = %d\n",
 962                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 963                         th->th_result = rc;
 964                         GOTO(stop_master_trans, rc);
 965                 }
 966
 967                 lur = tur->tur_update_records;
 968                 /* Write updates to the master MDT */
 969                 rc = sub_updates_write(env, lur, master_st);
 970
 971                 /* Cleanup the common parameters in the update records,
 972                  * master transno callback might add more parameters.
 973                  * and we need merge the update records again in the
 974                  * following */
 975                 if (tur->tur_update_params != NULL)
 976                         lur->lur_update_rec.ur_param_count = 0;
 977
 978                 if (rc < 0) {
 979                         CERROR("%s: write updates failed: rc = %d\n",
 980                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
 981                         th->th_result = rc;
 982                         GOTO(stop_master_trans, rc);
 983                 }
 984         }
 985
 986 stop_master_trans:
 987         /* Step 2: Stop the transaction on the master MDT, and fill the
 988          * master transno in the update logs to other MDT. */
 989         if (master_st != NULL && master_st->st_sub_th != NULL) {
 990                 if (th->th_local)
 991                         master_st->st_sub_th->th_local = th->th_local;
 992                 if (th->th_sync)
 993                         master_st->st_sub_th->th_sync = th->th_sync;
 994                 master_st->st_sub_th->th_tags = th->th_tags;
 995                 master_st->st_sub_th->th_result = th->th_result;
 996                 rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
 997                 if (rc < 0) {
 998                         th->th_result = rc;
 999                         GOTO(stop_other_trans, rc);
1000                 } else if (tur != NULL && tur->tur_update_records != NULL) {
1001                         struct llog_update_record *lur;
1002
1003                         lur = tur->tur_update_records;
1004                         if (lur->lur_update_rec.ur_master_transno == 0)
1005                                 /* Update master transno after master stop
1006                                  * callback */
1007                                 lur->lur_update_rec.ur_master_transno =
1008                                                 tgt_th_info(env)->tti_transno;
1009                 }
1010         }
1011
1012         /* Step 3: write updates to other MDTs */
1013         if (write_updates) {
1014                 struct llog_update_record *lur;
1015
1016                 /* Stop callback of master will add more updates and also update
1017                  * master transno, so merge the parameters and updates into one
1018                  * buffer again */
1019                 rc = prepare_writing_updates(env, tmt);
1020                 if (rc < 0) {
1021                         CERROR("%s: prepare updates failed: rc = %d\n",
1022                                master_dev->dd_lu_dev.ld_obd->obd_name, rc);
1023                         th->th_result = rc;
1024                         GOTO(stop_other_trans, rc);
1025                 }
1026                 lur = tur->tur_update_records;
1027                 list_for_each_entry(st, &tmt->tmt_sub_thandle_list,
1028                                     st_sub_list) {
1029                         if (st->st_sub_th == NULL || st == master_st ||
1030                             st->st_sub_th->th_result < 0)
1031                                 continue;
1032
1033                         rc = sub_updates_write(env, lur, st);
1034                         if (rc < 0) {
1035                                 th->th_result = rc;
1036                                 break;
1037                         }
1038                 }
1039         }
1040
1041 stop_other_trans:
1042         /* Step 4: Stop the transaction on other MDTs */
1043         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1044                 if (st == master_st || st->st_sub_th == NULL)
1045                         continue;
1046
1047                 if (th->th_sync)
1048                         st->st_sub_th->th_sync = th->th_sync;
1049                 if (th->th_local)
1050                         st->st_sub_th->th_local = th->th_local;
1051                 st->st_sub_th->th_tags = th->th_tags;
1052                 st->st_sub_th->th_result = th->th_result;
1053                 rc = dt_trans_stop(env, st->st_sub_th->th_dev,
1054                                    st->st_sub_th);
1055                 if (unlikely(rc < 0 && th->th_result == 0))
1056                         th->th_result = rc;
1057         }
1058
1059         rc = top_trans_wait_result(top_th);
1060
1061         tmt->tmt_result = rc;
1062
1063         /* Balance for the refcount in top_trans_create, Note: if it is NOT
1064          * multiple node transaction, the top transaction will be destroyed. */
1065         top_multiple_thandle_put(tmt);
1066         OBD_FREE_PTR(top_th);
1067         RETURN(rc);
1068 }
1069 EXPORT_SYMBOL(top_trans_stop);
1070
1071 /**
1072  * Create top_multiple_thandle for top_thandle
1073  *
1074  * Create top_mutilple_thandle to manage the mutiple node transaction
1075  * for top_thandle, and it also needs to add master sub thandle to the
1076  * sub trans list now.
1077  *
1078  * \param[in] env       execution environment
1079  * \param[in] top_th    the top thandle
1080  *
1081  * \retval      0 if creation succeeds
1082  * \retval      negative errno if creation fails
1083  */
1084 int top_trans_create_tmt(const struct lu_env *env,
1085                          struct top_thandle *top_th)
1086 {
1087         struct top_multiple_thandle *tmt;
1088
1089         OBD_ALLOC_PTR(tmt);
1090         if (tmt == NULL)
1091                 return -ENOMEM;
1092
1093         tmt->tmt_magic = TOP_THANDLE_MAGIC;
1094         INIT_LIST_HEAD(&tmt->tmt_sub_thandle_list);
1095         INIT_LIST_HEAD(&tmt->tmt_commit_list);
1096         atomic_set(&tmt->tmt_refcount, 1);
1097         spin_lock_init(&tmt->tmt_sub_lock);
1098         init_waitqueue_head(&tmt->tmt_stop_waitq);
1099
1100         top_th->tt_multiple_thandle = tmt;
1101
1102         return 0;
1103 }
1104
1105 static struct sub_thandle *
1106 create_sub_thandle_with_thandle(struct top_thandle *top_th,
1107                                 struct thandle *sub_th)
1108 {
1109         struct sub_thandle *st;
1110
1111         /* create and init sub th to the top trans list */
1112         st = create_sub_thandle(top_th->tt_multiple_thandle,
1113                                 sub_th->th_dev);
1114         if (IS_ERR(st))
1115                 return st;
1116
1117         st->st_sub_th = sub_th;
1118
1119         sub_th->th_top = &top_th->tt_super;
1120         return st;
1121 }
1122
1123 /**
1124  * Get sub thandle.
1125  *
1126  * Get sub thandle from the top thandle according to the sub dt_device.
1127  *
1128  * \param[in] env       execution environment
1129  * \param[in] th        thandle on the top layer.
1130  * \param[in] sub_dt    sub dt_device used to get sub transaction
1131  *
1132  * \retval              thandle of sub transaction if succeed
1133  * \retval              PTR_ERR(errno) if failed
1134  */
1135 struct thandle *thandle_get_sub_by_dt(const struct lu_env *env,
1136                                       struct thandle *th,
1137                                       struct dt_device *sub_dt)
1138 {
1139         struct sub_thandle      *st = NULL;
1140         struct sub_thandle      *master_st = NULL;
1141         struct top_thandle      *top_th;
1142         struct thandle          *sub_th = NULL;
1143         int                     rc = 0;
1144         ENTRY;
1145
1146         top_th = container_of(th, struct top_thandle, tt_super);
1147
1148         if (likely(sub_dt == top_th->tt_master_sub_thandle->th_dev))
1149                 RETURN(top_th->tt_master_sub_thandle);
1150
1151         if (top_th->tt_multiple_thandle != NULL) {
1152                 st = lookup_sub_thandle(top_th->tt_multiple_thandle, sub_dt);
1153                 if (st != NULL)
1154                         RETURN(st->st_sub_th);
1155         }
1156
1157         sub_th = dt_trans_create(env, sub_dt);
1158         if (IS_ERR(sub_th))
1159                 RETURN(sub_th);
1160
1161         /* Create top_multiple_thandle if necessary */
1162         if (top_th->tt_multiple_thandle == NULL) {
1163                 struct top_multiple_thandle *tmt;
1164
1165                 rc = top_trans_create_tmt(env, top_th);
1166                 if (rc < 0)
1167                         GOTO(stop_trans, rc);
1168
1169                 tmt = top_th->tt_multiple_thandle;
1170
1171                 /* Add master sub th to the top trans list */
1172                 tmt->tmt_master_sub_dt =
1173                         top_th->tt_master_sub_thandle->th_dev;
1174                 master_st = create_sub_thandle_with_thandle(top_th,
1175                                         top_th->tt_master_sub_thandle);
1176                 if (IS_ERR(master_st)) {
1177                         rc = PTR_ERR(master_st);
1178                         master_st = NULL;
1179                         GOTO(stop_trans, rc);
1180                 }
1181         }
1182
1183         /* create and init sub th to the top trans list */
1184         st = create_sub_thandle_with_thandle(top_th, sub_th);
1185         if (IS_ERR(st)) {
1186                 rc = PTR_ERR(st);
1187                 st = NULL;
1188                 GOTO(stop_trans, rc);
1189         }
1190         st->st_sub_th->th_wait_submit = 1;
1191 stop_trans:
1192         if (rc < 0) {
1193                 if (master_st != NULL) {
1194                         list_del(&master_st->st_sub_list);
1195                         OBD_FREE_PTR(master_st);
1196                 }
1197                 sub_th->th_result = rc;
1198                 dt_trans_stop(env, sub_dt, sub_th);
1199                 sub_th = ERR_PTR(rc);
1200         }
1201
1202         RETURN(sub_th);
1203 }
1204 EXPORT_SYMBOL(thandle_get_sub_by_dt);
1205
1206 /**
1207  * Top multiple thandle destroy
1208  *
1209  * Destroy multiple thandle and all its sub thandle.
1210  *
1211  * \param[in] tmt       top_multiple_thandle to be destroyed.
1212  */
1213 void top_multiple_thandle_destroy(struct top_multiple_thandle *tmt)
1214 {
1215         struct sub_thandle *st;
1216         struct sub_thandle *tmp;
1217
1218         LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
1219         list_for_each_entry_safe(st, tmp, &tmt->tmt_sub_thandle_list,
1220                                  st_sub_list) {
1221                 struct sub_thandle_cookie *stc;
1222                 struct sub_thandle_cookie *tmp;
1223
1224                 list_del(&st->st_sub_list);
1225                 list_for_each_entry_safe(stc, tmp, &st->st_cookie_list,
1226                                          stc_list) {
1227                         list_del(&stc->stc_list);
1228                         OBD_FREE_PTR(stc);
1229                 }
1230                 OBD_FREE_PTR(st);
1231         }
1232         OBD_FREE_PTR(tmt);
1233 }
1234 EXPORT_SYMBOL(top_multiple_thandle_destroy);
1235
1236 /**
1237  * Cancel the update log on MDTs
1238  *
1239  * Cancel the update log on MDTs then destroy the thandle.
1240  *
1241  * \param[in] env       execution environment
1242  * \param[in] tmt       the top multiple thandle whose updates records
1243  *                      will be cancelled.
1244  *
1245  * \retval              0 if cancellation succeeds.
1246  * \retval              negative errno if cancellation fails.
1247  */
1248 static int distribute_txn_cancel_records(const struct lu_env *env,
1249                                          struct top_multiple_thandle *tmt)
1250 {
1251         struct sub_thandle *st;
1252         ENTRY;
1253
1254         top_multiple_thandle_dump(tmt, D_INFO);
1255         /* Cancel update logs on other MDTs */
1256         list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
1257                 struct llog_ctxt        *ctxt;
1258                 struct obd_device       *obd;
1259                 struct llog_cookie      *cookie;
1260                 struct sub_thandle_cookie *stc;
1261                 int rc;
1262
1263                 obd = st->st_dt->dd_lu_dev.ld_obd;
1264                 ctxt = llog_get_context(obd, LLOG_UPDATELOG_ORIG_CTXT);
1265                 if (ctxt == NULL)
1266                         continue;
1267                 list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
1268                         cookie = &stc->stc_cookie;
1269                         if (fid_is_zero(&cookie->lgc_lgl.lgl_oi.oi_fid))
1270                                 continue;
1271
1272                         rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
1273                                                      cookie);
1274                         CDEBUG(D_HA, "%s: batchid %llu cancel update log "
1275                                DOSTID ".%u : rc = %d\n", obd->obd_name,
1276                                tmt->tmt_batchid,
1277                                POSTID(&cookie->lgc_lgl.lgl_oi),
1278                                cookie->lgc_index, rc);
1279                 }
1280
1281                 llog_ctxt_put(ctxt);
1282         }
1283
1284         RETURN(0);
1285 }
1286
1287 /**
1288  * Check if there are committed transaction
1289  *
1290  * Check if there are committed transaction in the distribute transaction
1291  * list, then cancel the update records for those committed transaction.
1292  * Because the distribute transaction in the list are sorted by batchid,
1293  * and cancellation will be done by batchid order, so we only check the first
1294  * the transaction(with lowest batchid) in the list.
1295  *
1296  * \param[in] lod       lod device where cancel thread is
1297  *
1298  * \retval              true if it is ready
1299  * \retval              false if it is not ready
1300  */
1301 static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
1302 {
1303         struct top_multiple_thandle     *tmt = NULL;
1304         struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
1305         bool    ready = false;
1306
1307         spin_lock(&tdtd->tdtd_batchid_lock);
1308         if (!list_empty(&tdtd->tdtd_list)) {
1309                 tmt = list_entry(tdtd->tdtd_list.next,
1310                                  struct top_multiple_thandle, tmt_commit_list);
1311                 if (tmt->tmt_committed &&
1312                     (!obd->obd_recovering || (obd->obd_recovering &&
1313                     tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
1314                         ready = true;
1315         }
1316         spin_unlock(&tdtd->tdtd_batchid_lock);
1317
1318         return ready;
1319 }
1320
1321 struct distribute_txn_bid_data {
1322         struct dt_txn_commit_cb  dtbd_cb;
1323         struct target_distribute_txn_data      *dtbd_tdtd;
1324         __u64                    dtbd_batchid;
1325 };
1326
1327 /**
1328  * callback of updating commit batchid
1329  *
1330  * Updating commit batchid then wake up the commit thread to cancel the
1331  * records.
1332  *
1333  * \param[in]env        execution environment
1334  * \param[in]th         thandle to updating commit batchid
1335  * \param[in]cb         commit callback
1336  * \param[in]err        result of thandle
1337  */
1338 static void distribute_txn_batchid_cb(struct lu_env *env,
1339                                       struct thandle *th,
1340                                       struct dt_txn_commit_cb *cb,
1341                                       int err)
1342 {
1343         struct distribute_txn_bid_data          *dtbd = NULL;
1344         struct target_distribute_txn_data       *tdtd;
1345
1346         dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
1347         tdtd = dtbd->dtbd_tdtd;
1348
1349         CDEBUG(D_HA, "%s: %llu batchid updated\n",
1350               tdtd->tdtd_lut->lut_obd->obd_name, dtbd->dtbd_batchid);
1351         spin_lock(&tdtd->tdtd_batchid_lock);
1352         if (dtbd->dtbd_batchid > tdtd->tdtd_committed_batchid &&
1353             !tdtd->tdtd_lut->lut_obd->obd_no_transno)
1354                 tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
1355         spin_unlock(&tdtd->tdtd_batchid_lock);
1356         atomic_dec(&tdtd->tdtd_refcount);
1357         wake_up(&tdtd->tdtd_commit_thread_waitq);
1358
1359         OBD_FREE_PTR(dtbd);
1360 }
1361
1362 /**
1363  * Update the commit batchid in disk
1364  *
1365  * Update commit batchid in the disk, after this is committed, it can start
1366  * to cancel the update records.
1367  *
1368  * \param[in] env       execution environment
1369  * \param[in] tdtd      distribute transaction structure
1370  * \param[in] batchid   commit batchid to be updated
1371  *
1372  * \retval              0 if update succeeds.
1373  * \retval              negative errno if update fails.
1374  */
1375 static int
1376 distribute_txn_commit_batchid_update(const struct lu_env *env,
1377                               struct target_distribute_txn_data *tdtd,
1378                               __u64 batchid)
1379 {
1380         struct distribute_txn_bid_data  *dtbd = NULL;
1381         struct thandle          *th;
1382         struct lu_buf            buf;
1383         __u64                    tmp;
1384         __u64                    off;
1385         int                      rc;
1386         ENTRY;
1387
1388         OBD_ALLOC_PTR(dtbd);
1389         if (dtbd == NULL)
1390                 RETURN(-ENOMEM);
1391         dtbd->dtbd_batchid = batchid;
1392         dtbd->dtbd_tdtd = tdtd;
1393         dtbd->dtbd_cb.dcb_func = distribute_txn_batchid_cb;
1394         atomic_inc(&tdtd->tdtd_refcount);
1395
1396         th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
1397         if (IS_ERR(th)) {
1398                 OBD_FREE_PTR(dtbd);
1399                 RETURN(PTR_ERR(th));
1400         }
1401
1402         tmp = cpu_to_le64(batchid);
1403         buf.lb_buf = &tmp;
1404         buf.lb_len = sizeof(tmp);
1405         off = 0;
1406
1407         rc = dt_declare_record_write(env, tdtd->tdtd_batchid_obj, &buf, off,
1408                                      th);
1409         if (rc < 0)
1410                 GOTO(stop, rc);
1411
1412         rc = dt_trans_start_local(env, tdtd->tdtd_lut->lut_bottom, th);
1413         if (rc < 0)
1414                 GOTO(stop, rc);
1415
1416         rc = dt_trans_cb_add(th, &dtbd->dtbd_cb);
1417         if (rc < 0)
1418                 GOTO(stop, rc);
1419
1420         rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
1421                              &off, th);
1422
1423         CDEBUG(D_INFO, "%s: update batchid "LPU64": rc = %d\n",
1424                tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
1425
1426 stop:
1427         dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
1428         if (rc < 0)
1429                 OBD_FREE_PTR(dtbd);
1430         RETURN(rc);
1431 }
1432
1433 /**
1434  * Init commit batchid for distribute transaction.
1435  *
1436  * Initialize the batchid object and get commit batchid from the object.
1437  *
1438  * \param[in] env       execution environment
1439  * \param[in] tdtd      distribute transaction whose batchid is initialized.
1440  *
1441  * \retval              0 if initialization succeeds.
1442  * \retval              negative errno if initialization fails.
1443  **/
1444 static int
1445 distribute_txn_commit_batchid_init(const struct lu_env *env,
1446                                    struct target_distribute_txn_data *tdtd)
1447 {
1448         struct tgt_thread_info  *tti = tgt_th_info(env);
1449         struct lu_target        *lut = tdtd->tdtd_lut;
1450         struct lu_attr          *attr = &tti->tti_attr;
1451         struct lu_fid           *fid = &tti->tti_fid1;
1452         struct dt_object_format *dof = &tti->tti_u.update.tti_update_dof;
1453         struct dt_object        *dt_obj = NULL;
1454         struct lu_buf           buf;
1455         __u64                   tmp;
1456         __u64                   off;
1457         int                     rc;
1458         ENTRY;
1459
1460         memset(attr, 0, sizeof(*attr));
1461         attr->la_valid = LA_MODE;
1462         attr->la_mode = S_IFREG | S_IRUGO | S_IWUSR;
1463         dof->dof_type = dt_mode_to_dft(S_IFREG);
1464
1465         lu_local_obj_fid(fid, BATCHID_COMMITTED_OID);
1466
1467         dt_obj = dt_find_or_create(env, lut->lut_bottom, fid, dof,
1468                                    attr);
1469         if (IS_ERR(dt_obj)) {
1470                 rc = PTR_ERR(dt_obj);
1471                 dt_obj = NULL;
1472                 GOTO(out_put, rc);
1473         }
1474
1475         tdtd->tdtd_batchid_obj = dt_obj;
1476
1477         buf.lb_buf = &tmp;
1478         buf.lb_len = sizeof(tmp);
1479         off = 0;
1480         rc = dt_read(env, dt_obj, &buf, &off);
1481         if (rc < 0 || (rc < buf.lb_len && rc > 0)) {
1482                 CERROR("%s can't read last committed batchid: rc = %d\n",
1483                        tdtd->tdtd_lut->lut_obd->obd_name, rc);
1484                 if (rc > 0)
1485                         rc = -EINVAL;
1486                 GOTO(out_put, rc);
1487         } else if (rc == buf.lb_len) {
1488                 tdtd->tdtd_committed_batchid = le64_to_cpu(tmp);
1489                 CDEBUG(D_HA, "%s: committed batchid %llu\n",
1490                        tdtd->tdtd_lut->lut_obd->obd_name,
1491                        tdtd->tdtd_committed_batchid);
1492                 rc = 0;
1493         }
1494
1495 out_put:
1496         if (rc < 0 && dt_obj != NULL) {
1497                 lu_object_put(env, &dt_obj->do_lu);
1498                 tdtd->tdtd_batchid_obj = NULL;
1499         }
1500         return rc;
1501 }
1502
1503 /**
1504  * manage the distribute transaction thread
1505  *
1506  * Distribute transaction are linked to the list, and once the distribute
1507  * transaction is committed, it will update the last committed batchid first,
1508  * after it is committed, it will cancel the records.
1509  *
1510  * \param[in] _arg      argument for commit thread
1511  *
1512  * \retval              0 if thread is running successfully
1513  * \retval              negative errno if the thread can not be run.
1514  */
1515 static int distribute_txn_commit_thread(void *_arg)
1516 {
1517         struct target_distribute_txn_data *tdtd = _arg;
1518         struct lu_target        *lut = tdtd->tdtd_lut;
1519         struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
1520         struct l_wait_info       lwi = { 0 };
1521         struct lu_env            env;
1522         struct list_head         list;
1523         int                      rc;
1524         struct top_multiple_thandle *tmt;
1525         struct top_multiple_thandle *tmp;
1526         __u64                    batchid = 0, committed;
1527
1528         ENTRY;
1529
1530         rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
1531         if (rc != 0)
1532                 RETURN(rc);
1533
1534         spin_lock(&tdtd->tdtd_batchid_lock);
1535         thread->t_flags = SVC_RUNNING;
1536         spin_unlock(&tdtd->tdtd_batchid_lock);
1537         wake_up(&thread->t_ctl_waitq);
1538         INIT_LIST_HEAD(&list);
1539
1540         CDEBUG(D_HA, "%s: start commit thread committed batchid "LPU64"\n",
1541                tdtd->tdtd_lut->lut_obd->obd_name,
1542                tdtd->tdtd_committed_batchid);
1543
1544         while (distribute_txn_commit_thread_running(lut)) {
1545                 spin_lock(&tdtd->tdtd_batchid_lock);
1546                 list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1547                                          tmt_commit_list) {
1548                         if (tmt->tmt_committed == 0)
1549                                 break;
1550
1551                         /* Note: right now, replay is based on master MDT
1552                          * transno, but cancellation is based on batchid.
1553                          * so we do not try to cancel the update log until
1554                          * the recoverying is done, unless the update records
1555                          * batchid < committed_batchid. */
1556                         if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
1557                                 list_move_tail(&tmt->tmt_commit_list, &list);
1558                         } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
1559                                 LASSERTF(tmt->tmt_batchid >= batchid,
1560                                          "tmt %p tmt_batchid: "LPU64", batchid "
1561                                           LPU64"\n", tmt, tmt->tmt_batchid,
1562                                          batchid);
1563                                 /* There are three types of distribution
1564                                  * transaction result
1565                                  *
1566                                  * 1. If tmt_result < 0, it means the
1567                                  * distribution transaction fails, which should
1568                                  * be rare, because once declare phase succeeds,
1569                                  * the operation should succeeds anyway. Note in
1570                                  * this case, we will still update batchid so
1571                                  * cancellation would be stopped.
1572                                  *
1573                                  * 2. If tmt_result == 0, it means the
1574                                  * distribution transaction succeeds, and we
1575                                  * will update batchid.
1576                                  *
1577                                  * 3. If tmt_result > 0, it means distribute
1578                                  * transaction is not yet committed on every
1579                                  * node, but we need release this tmt before
1580                                  * that, which usuually happens during umount.
1581                                  */
1582                                 if (tmt->tmt_result <= 0)
1583                                         batchid = tmt->tmt_batchid;
1584                                 list_move_tail(&tmt->tmt_commit_list, &list);
1585                         }
1586                 }
1587                 spin_unlock(&tdtd->tdtd_batchid_lock);
1588
1589                 CDEBUG(D_HA, "%s: batchid: "LPU64" committed batchid "
1590                        LPU64"\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
1591                        tdtd->tdtd_committed_batchid);
1592                 /* update globally committed on a storage */
1593                 if (batchid > tdtd->tdtd_committed_batchid) {
1594                         rc = distribute_txn_commit_batchid_update(&env, tdtd,
1595                                                              batchid);
1596                         if (rc == 0)
1597                                 batchid = 0;
1598                 }
1599                 /* cancel the records for committed batchid's */
1600                 /* XXX: should we postpone cancel's till the end of recovery? */
1601                 committed = tdtd->tdtd_committed_batchid;
1602                 list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1603                         if (tmt->tmt_batchid > committed)
1604                                 break;
1605                         list_del_init(&tmt->tmt_commit_list);
1606                         if (tmt->tmt_result <= 0)
1607                                 distribute_txn_cancel_records(&env, tmt);
1608                         top_multiple_thandle_put(tmt);
1609                 }
1610
1611                 l_wait_event(tdtd->tdtd_commit_thread_waitq,
1612                              !distribute_txn_commit_thread_running(lut) ||
1613                              committed < tdtd->tdtd_committed_batchid ||
1614                              tdtd_ready_for_cancel_log(tdtd), &lwi);
1615         };
1616
1617         l_wait_event(tdtd->tdtd_commit_thread_waitq,
1618                      atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
1619
1620         spin_lock(&tdtd->tdtd_batchid_lock);
1621         list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
1622                                  tmt_commit_list)
1623                 list_move_tail(&tmt->tmt_commit_list, &list);
1624         spin_unlock(&tdtd->tdtd_batchid_lock);
1625
1626         CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
1627                tdtd->tdtd_lut->lut_obd->obd_name);
1628         list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
1629                 list_del_init(&tmt->tmt_commit_list);
1630                 top_multiple_thandle_dump(tmt, D_HA);
1631                 top_multiple_thandle_put(tmt);
1632         }
1633
1634         thread->t_flags = SVC_STOPPED;
1635         lu_env_fini(&env);
1636         wake_up(&thread->t_ctl_waitq);
1637
1638         RETURN(0);
1639 }
1640
1641 /**
1642  * Start llog cancel thread
1643  *
1644  * Start llog cancel(master/slave) thread on LOD
1645  *
1646  * \param[in]lclt       cancel log thread to be started.
1647  *
1648  * \retval              0 if the thread is started successfully.
1649  * \retval              negative errno if the thread is not being
1650  *                      started.
1651  */
1652 int distribute_txn_init(const struct lu_env *env,
1653                         struct lu_target *lut,
1654                         struct target_distribute_txn_data *tdtd,
1655                         __u32 index)
1656 {
1657         struct task_struct      *task;
1658         struct l_wait_info       lwi = { 0 };
1659         int                     rc;
1660         ENTRY;
1661
1662         INIT_LIST_HEAD(&tdtd->tdtd_list);
1663         INIT_LIST_HEAD(&tdtd->tdtd_replay_finish_list);
1664         INIT_LIST_HEAD(&tdtd->tdtd_replay_list);
1665         spin_lock_init(&tdtd->tdtd_batchid_lock);
1666         spin_lock_init(&tdtd->tdtd_replay_list_lock);
1667         tdtd->tdtd_replay_handler = distribute_txn_replay_handle;
1668         tdtd->tdtd_replay_ready = 0;
1669
1670         tdtd->tdtd_batchid = lut->lut_last_transno + 1;
1671
1672         init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
1673         init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
1674         atomic_set(&tdtd->tdtd_refcount, 0);
1675
1676         tdtd->tdtd_lut = lut;
1677         rc = distribute_txn_commit_batchid_init(env, tdtd);
1678         if (rc != 0)
1679                 RETURN(rc);
1680
1681         task = kthread_run(distribute_txn_commit_thread, tdtd, "tdtd-%u",
1682                            index);
1683         if (IS_ERR(task))
1684                 RETURN(PTR_ERR(task));
1685
1686         l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1687                      distribute_txn_commit_thread_running(lut) ||
1688                      distribute_txn_commit_thread_stopped(lut), &lwi);
1689         RETURN(0);
1690 }
1691 EXPORT_SYMBOL(distribute_txn_init);
1692
1693 /**
1694  * Stop llog cancel thread
1695  *
1696  * Stop llog cancel(master/slave) thread on LOD and also destory
1697  * all of transaction in the list.
1698  *
1699  * \param[in]lclt       cancel log thread to be stopped.
1700  */
1701 void distribute_txn_fini(const struct lu_env *env,
1702                          struct target_distribute_txn_data *tdtd)
1703 {
1704         struct lu_target *lut = tdtd->tdtd_lut;
1705
1706         /* Stop cancel thread */
1707         if (lut == NULL || !distribute_txn_commit_thread_running(lut))
1708                 return;
1709
1710         spin_lock(&tdtd->tdtd_batchid_lock);
1711         lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
1712         spin_unlock(&tdtd->tdtd_batchid_lock);
1713         wake_up(&tdtd->tdtd_commit_thread_waitq);
1714         wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
1715                    lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
1716
1717         dtrq_list_destroy(tdtd);
1718         if (tdtd->tdtd_batchid_obj != NULL) {
1719                 lu_object_put(env, &tdtd->tdtd_batchid_obj->do_lu);
1720                 tdtd->tdtd_batchid_obj = NULL;
1721         }
1722 }
1723 EXPORT_SYMBOL(distribute_txn_fini);