Whamcloud - gitweb
LU-15420 build: fixes to support building on Ubuntu 22.04 LTS
[fs/lustre-release.git] / lustre / target / update_trans.c
index db92713..aa13f94 100644 (file)
@@ -20,7 +20,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2014, Intel Corporation.
+ * Copyright (c) 2015, 2017, Intel Corporation.
  */
 /*
  * lustre/target/update_trans.c
@@ -72,8 +72,7 @@ static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
        struct sub_thandle      *st;
 
        LASSERT(tmt->tmt_magic == TOP_THANDLE_MAGIC);
-       CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d"
-              "batchid "LPU64"\n",
+       CDEBUG(mask, "%s tmt %p refcount %d committed %d result %d batchid %llu\n",
               tmt->tmt_master_sub_dt ?
               tmt->tmt_master_sub_dt->dd_lu_dev.ld_obd->obd_name :
               "NULL",
@@ -83,13 +82,15 @@ static void top_multiple_thandle_dump(struct top_multiple_thandle *tmt,
        list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
                struct sub_thandle_cookie *stc;
 
-               CDEBUG(mask, "st %p obd %s committed %d sub_th %p\n",
+               CDEBUG(mask, "st %p obd %s committed %d started %d stopped %d "
+                      "result %d sub_th %p\n",
                       st, st->st_dt->dd_lu_dev.ld_obd->obd_name,
-                      st->st_committed, st->st_sub_th);
+                      st->st_committed, st->st_started, st->st_stopped,
+                      st->st_result, st->st_sub_th);
 
                list_for_each_entry(stc, &st->st_cookie_list, stc_list) {
-                       CDEBUG(mask, " cookie "DOSTID": %u\n",
-                              POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
+                       CDEBUG(mask, " cookie "DFID".%u\n",
+                              PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
                               stc->stc_cookie.lgc_index);
                }
        }
@@ -168,19 +169,20 @@ static int sub_updates_write(const struct lu_env *env,
                             struct llog_update_record *record,
                             struct sub_thandle *sub_th)
 {
-       struct dt_device        *dt = sub_th->st_dt;
-       struct llog_ctxt        *ctxt;
-       int                     rc;
+       struct dt_device *dt = sub_th->st_dt;
+       struct llog_ctxt *ctxt;
        struct llog_update_record *lur = NULL;
-       struct update_params    *params = NULL;
-       __u32                   update_count = 0;
-       __u32                   param_count = 0;
-       __u32                   last_update_count = 0;
-       __u32                   last_param_count = 0;
-       void                    *src;
-       void                    *start;
-       void                    *next;
+       __u32 update_count = 0;
+       __u32 param_count = 0;
+       __u32 last_update_count = 0;
+       __u32 last_param_count = 0;
+       char *start;
+       char *cur;
+       char *next;
        struct sub_thandle_cookie *stc;
+       size_t reclen;
+       bool eof = false;
+       int rc;
        ENTRY;
 
        ctxt = llog_get_context(dt->dd_lu_dev.ld_obd,
@@ -201,7 +203,17 @@ static int sub_updates_write(const struct lu_env *env,
                 "lrh_len %u record_size %zu\n", record->lur_hdr.lrh_len,
                 llog_update_record_size(record));
 
-       if (likely(record->lur_hdr.lrh_len <= ctxt->loc_chunk_size)) {
+       /*
+        * If its size > llog chunk_size, then write current chunk to the update
+        * llog, NB the padding should >= LLOG_MIN_REC_SIZE.
+        *
+        * So check padding length is either >= LLOG_MIN_REC_SIZE or is 0
+        * (record length just matches the chunk size).
+        */
+
+       reclen = record->lur_hdr.lrh_len;
+       if (reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
+           reclen == ctxt->loc_chunk_size) {
                OBD_ALLOC_PTR(stc);
                if (stc == NULL)
                        GOTO(llog_put, rc = -ENOMEM);
@@ -210,9 +222,9 @@ static int sub_updates_write(const struct lu_env *env,
                rc = llog_add(env, ctxt->loc_handle, &record->lur_hdr,
                              &stc->stc_cookie, sub_th->st_sub_th);
 
-               CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u: rc = %d\n",
+               CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
                       dt->dd_lu_dev.ld_obd->obd_name,
-                      POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
+                      PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
                       stc->stc_cookie.lgc_index, rc);
 
                if (rc > 0) {
@@ -233,87 +245,76 @@ static int sub_updates_write(const struct lu_env *env,
        memcpy(lur, &record->lur_hdr, sizeof(record->lur_hdr));
        lur->lur_update_rec.ur_update_count = 0;
        lur->lur_update_rec.ur_param_count = 0;
-       src = &record->lur_update_rec.ur_ops;
-       start = next = src;
-       lur->lur_hdr.lrh_len = llog_update_record_size(lur);
-       params = update_records_get_params(&record->lur_update_rec);
+       start = (char *)&record->lur_update_rec.ur_ops;
+       cur = next = start;
        do {
-               size_t rec_len;
-
-               if (update_count < record->lur_update_rec.ur_update_count) {
-                       next = update_op_next_op((struct update_op *)src);
-               } else {
-                       if (param_count == 0)
-                               next = update_records_get_params(
-                                               &record->lur_update_rec);
-                       else
-                               next = (char *)src +
-                                       object_update_param_size(
-                                       (struct object_update_param *)src);
+               if (update_count < record->lur_update_rec.ur_update_count)
+                       next = (char *)update_op_next_op(
+                                               (struct update_op *)cur);
+               else if (param_count < record->lur_update_rec.ur_param_count)
+                       next = (char *)update_param_next_param(
+                                               (struct update_param *)cur);
+               else
+                       eof = true;
+
+               reclen = __llog_update_record_size(
+                               __update_records_size(next - start));
+               if ((reclen + LLOG_MIN_REC_SIZE <= ctxt->loc_chunk_size ||
+                    reclen == ctxt->loc_chunk_size) &&
+                   !eof) {
+                       cur = next;
+
+                       if (update_count <
+                           record->lur_update_rec.ur_update_count)
+                               update_count++;
+                       else if (param_count <
+                                record->lur_update_rec.ur_param_count)
+                               param_count++;
+                       continue;
                }
 
-               rec_len = cfs_size_round((unsigned long)(next - src));
-               /* If its size > llog chunk_size, then write current chunk to
-                * the update llog. */
-               if (lur->lur_hdr.lrh_len + rec_len + LLOG_MIN_REC_SIZE >
-                   ctxt->loc_chunk_size ||
-                   param_count == record->lur_update_rec.ur_param_count) {
-                       lur->lur_update_rec.ur_update_count =
-                               update_count > last_update_count ?
-                               update_count - last_update_count : 0;
-                       lur->lur_update_rec.ur_param_count = param_count -
-                                                            last_param_count;
-
-                       memcpy(&lur->lur_update_rec.ur_ops, start,
-                              (unsigned long)(src - start));
-                       if (last_update_count != 0)
-                               lur->lur_update_rec.ur_flags |=
-                                               UPDATE_RECORD_CONTINUE;
-
-                       update_records_dump(&lur->lur_update_rec, D_INFO, true);
-                       lur->lur_hdr.lrh_len = llog_update_record_size(lur);
-                       LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
-
-                       OBD_ALLOC_PTR(stc);
-                       if (stc == NULL)
-                               GOTO(llog_put, rc = -ENOMEM);
-                       INIT_LIST_HEAD(&stc->stc_list);
-
-                       rc = llog_add(env, ctxt->loc_handle,
-                                     &lur->lur_hdr,
-                                     &stc->stc_cookie, sub_th->st_sub_th);
-
-                       CDEBUG(D_INFO, "%s: Add update log "DOSTID":%u"
-                              " rc = %d\n", dt->dd_lu_dev.ld_obd->obd_name,
-                              POSTID(&stc->stc_cookie.lgc_lgl.lgl_oi),
-                              stc->stc_cookie.lgc_index, rc);
-
-                       if (rc > 0) {
-                               list_add(&stc->stc_list,
-                                        &sub_th->st_cookie_list);
-                               rc = 0;
-                       } else {
-                               OBD_FREE_PTR(stc);
-                               GOTO(llog_put, rc);
-                       }
+               lur->lur_update_rec.ur_update_count = update_count -
+                                                     last_update_count;
+               lur->lur_update_rec.ur_param_count = param_count -
+                                                    last_param_count;
+               memcpy(&lur->lur_update_rec.ur_ops, start, cur - start);
+               lur->lur_hdr.lrh_len = llog_update_record_size(lur);
 
-                       last_update_count = update_count;
-                       last_param_count = param_count;
-                       start = src;
-                       lur->lur_update_rec.ur_update_count = 0;
-                       lur->lur_update_rec.ur_param_count = 0;
-                       lur->lur_hdr.lrh_len = llog_update_record_size(lur);
+               LASSERT(lur->lur_hdr.lrh_len ==
+                        __llog_update_record_size(
+                               __update_records_size(cur - start)));
+               LASSERT(lur->lur_hdr.lrh_len <= ctxt->loc_chunk_size);
+
+               update_records_dump(&lur->lur_update_rec, D_INFO, true);
+
+               OBD_ALLOC_PTR(stc);
+               if (stc == NULL)
+                       GOTO(llog_put, rc = -ENOMEM);
+               INIT_LIST_HEAD(&stc->stc_list);
+
+               rc = llog_add(env, ctxt->loc_handle, &lur->lur_hdr,
+                             &stc->stc_cookie, sub_th->st_sub_th);
+
+               CDEBUG(D_INFO, "%s: Add update log "DFID".%u: rc = %d\n",
+                       dt->dd_lu_dev.ld_obd->obd_name,
+                       PFID(&stc->stc_cookie.lgc_lgl.lgl_oi.oi_fid),
+                       stc->stc_cookie.lgc_index, rc);
+
+               if (rc > 0) {
+                       list_add(&stc->stc_list, &sub_th->st_cookie_list);
+                       rc = 0;
+               } else {
+                       OBD_FREE_PTR(stc);
+                       GOTO(llog_put, rc);
                }
 
-               src = next;
-               lur->lur_hdr.lrh_len += cfs_size_round(rec_len);
-               if (update_count < record->lur_update_rec.ur_update_count)
-                       update_count++;
-               else if (param_count < record->lur_update_rec.ur_param_count)
-                       param_count++;
-               else
-                       break;
-       } while (1);
+               last_update_count = update_count;
+               last_param_count = param_count;
+               start = cur;
+               lur->lur_update_rec.ur_update_count = 0;
+               lur->lur_update_rec.ur_param_count = 0;
+               lur->lur_update_rec.ur_flags |= UPDATE_RECORD_CONTINUE;
+       } while (!eof);
 
 llog_put:
        if (lur != NULL)
@@ -386,18 +387,6 @@ static int prepare_writing_updates(const struct lu_env *env,
        return 0;
 }
 
-static inline int
-distribute_txn_commit_thread_running(struct lu_target *lut)
-{
-       return lut->lut_tdtd_commit_thread.t_flags & SVC_RUNNING;
-}
-
-static inline int
-distribute_txn_commit_thread_stopped(struct lu_target *lut)
-{
-       return lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED;
-}
-
 /**
  * Top thandle commit callback
  *
@@ -415,8 +404,9 @@ static void top_trans_committed_cb(struct top_multiple_thandle *tmt)
        top_multiple_thandle_dump(tmt, D_HA);
        tmt->tmt_committed = 1;
        lut = dt2lu_dev(tmt->tmt_master_sub_dt)->ld_site->ls_tgt;
-       if (distribute_txn_commit_thread_running(lut))
-               wake_up(&lut->lut_tdtd->tdtd_commit_thread_waitq);
+       if (lut->lut_tdtd && lut->lut_tdtd->tdtd_commit_task)
+               wake_up_process(lut->lut_tdtd->tdtd_commit_task);
+
        RETURN_EXIT;
 }
 
@@ -450,25 +440,11 @@ struct sub_thandle *create_sub_thandle(struct top_multiple_thandle *tmt,
        return st;
 }
 
-/**
- * sub thandle commit callback
- *
- * Mark the sub thandle to be committed and if all sub thandle are committed
- * notify the top thandle.
- *
- * \param[in] env      execution environment
- * \param[in] sub_th   sub thandle being committed
- * \param[in] cb       commit callback
- * \param[in] err      trans result
- */
-static void sub_trans_commit_cb(struct lu_env *env,
-                               struct thandle *sub_th,
-                               struct dt_txn_commit_cb *cb, int err)
+static void sub_trans_commit_cb_internal(struct top_multiple_thandle *tmt,
+                                        struct thandle *sub_th, int err)
 {
        struct sub_thandle      *st;
-       struct top_multiple_thandle *tmt = cb->dcb_data;
        bool                    all_committed = true;
-       ENTRY;
 
        /* Check if all sub thandles are committed */
        spin_lock(&tmt->tmt_sub_lock);
@@ -493,6 +469,26 @@ static void sub_trans_commit_cb(struct lu_env *env,
        RETURN_EXIT;
 }
 
+/**
+ * sub thandle commit callback
+ *
+ * Mark the sub thandle to be committed and if all sub thandle are committed
+ * notify the top thandle.
+ *
+ * \param[in] env      execution environment
+ * \param[in] sub_th   sub thandle being committed
+ * \param[in] cb       commit callback
+ * \param[in] err      trans result
+ */
+static void sub_trans_commit_cb(struct lu_env *env,
+                               struct thandle *sub_th,
+                               struct dt_txn_commit_cb *cb, int err)
+{
+       struct top_multiple_thandle *tmt = cb->dcb_data;
+
+       sub_trans_commit_cb_internal(tmt, sub_th, err);
+}
+
 static void sub_thandle_register_commit_cb(struct sub_thandle *st,
                                    struct top_multiple_thandle *tmt)
 {
@@ -521,6 +517,7 @@ static void sub_trans_stop_cb(struct lu_env *env,
        struct top_multiple_thandle     *tmt = cb->dcb_data;
        ENTRY;
 
+       spin_lock(&tmt->tmt_sub_lock);
        list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
                if (st->st_stopped)
                        continue;
@@ -531,6 +528,7 @@ static void sub_trans_stop_cb(struct lu_env *env,
                        break;
                }
        }
+       spin_unlock(&tmt->tmt_sub_lock);
 
        wake_up(&tmt->tmt_stop_waitq);
        RETURN_EXIT;
@@ -572,6 +570,7 @@ int sub_thandle_trans_create(const struct lu_env *env,
        st->st_sub_th = sub_th;
 
        sub_th->th_wait_submit = 1;
+       sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
        return 0;
 }
 
@@ -609,8 +608,6 @@ top_trans_create(const struct lu_env *env, struct dt_device *master_dev)
                child_th->th_top = &top_th->tt_super;
                child_th->th_wait_submit = 1;
                top_th->tt_master_sub_thandle = child_th;
-
-               top_th->tt_super.th_tags |= child_th->th_tags;
        }
        return &top_th->tt_super;
 }
@@ -661,6 +658,7 @@ static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
 {
        struct target_distribute_txn_data *tdtd;
        struct dt_device *dt = new->tmt_master_sub_dt;
+       struct sub_thandle *st;
 
        LASSERT(dt != NULL);
        tdtd = dt2lu_dev(dt)->ld_site->ls_tgt->lut_tdtd;
@@ -668,6 +666,10 @@ static void distribute_txn_assign_batchid(struct top_multiple_thandle *new)
        new->tmt_batchid = tdtd->tdtd_batchid++;
        list_add_tail(&new->tmt_commit_list, &tdtd->tdtd_list);
        spin_unlock(&tdtd->tdtd_batchid_lock);
+       list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
+               if (st->st_sub_th != NULL)
+                       sub_thandle_register_commit_cb(st, new);
+       }
        top_multiple_thandle_get(new);
        top_multiple_thandle_dump(new, D_INFO);
 }
@@ -684,6 +686,7 @@ void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
        struct dt_device *dt = new->tmt_master_sub_dt;
        struct top_multiple_thandle *tmt;
        struct target_distribute_txn_data *tdtd;
+       struct sub_thandle *st;
        bool    at_head = false;
 
        LASSERT(dt != NULL);
@@ -701,10 +704,16 @@ void distribute_txn_insert_by_batchid(struct top_multiple_thandle *new)
                list_add(&new->tmt_commit_list, &tdtd->tdtd_list);
        }
        spin_unlock(&tdtd->tdtd_batchid_lock);
+
+       list_for_each_entry(st, &new->tmt_sub_thandle_list, st_sub_list) {
+               if (st->st_sub_th != NULL)
+                       sub_thandle_register_commit_cb(st, new);
+       }
+
        top_multiple_thandle_get(new);
        top_multiple_thandle_dump(new, D_INFO);
-       if (new->tmt_committed && at_head)
-               wake_up(&tdtd->tdtd_commit_thread_waitq);
+       if (new->tmt_committed && at_head && tdtd->tdtd_commit_task)
+               wake_up_process(tdtd->tdtd_commit_task);
 }
 
 /**
@@ -777,7 +786,6 @@ int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
                        top_th->tt_master_sub_thandle->th_sync = th->th_sync;
                if (th->th_local)
                        top_th->tt_master_sub_thandle->th_local = th->th_local;
-               top_th->tt_master_sub_thandle->th_tags = th->th_tags;
                rc = dt_trans_start(env, top_th->tt_master_sub_thandle->th_dev,
                                    top_th->tt_master_sub_thandle);
                RETURN(rc);
@@ -795,14 +803,13 @@ int top_trans_start(const struct lu_env *env, struct dt_device *master_dev,
                        st->st_sub_th->th_sync = th->th_sync;
                if (th->th_local)
                        st->st_sub_th->th_local = th->th_local;
-               st->st_sub_th->th_tags = th->th_tags;
                rc = dt_trans_start(env, st->st_sub_th->th_dev,
                                    st->st_sub_th);
                if (rc != 0)
                        GOTO(out, rc);
 
-               sub_thandle_register_stop_cb(st, tmt);
-               sub_thandle_register_commit_cb(st, tmt);
+               LASSERT(st->st_started == 0);
+               st->st_started = 1;
        }
 out:
        th->th_result = rc;
@@ -892,10 +899,8 @@ static bool top_trans_is_stopped(struct top_thandle *top_th)
  */
 static int top_trans_wait_result(struct top_thandle *top_th)
 {
-       struct l_wait_info      lwi = {0};
-
-       l_wait_event(top_th->tt_multiple_thandle->tmt_stop_waitq,
-                    top_trans_is_stopped(top_th), &lwi);
+       wait_event_idle(top_th->tt_multiple_thandle->tmt_stop_waitq,
+                       top_trans_is_stopped(top_th));
 
        RETURN(top_th->tt_super.th_result);
 }
@@ -933,7 +938,6 @@ int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
                        top_th->tt_master_sub_thandle->th_sync = th->th_sync;
                if (th->th_local)
                        top_th->tt_master_sub_thandle->th_local = th->th_local;
-               top_th->tt_master_sub_thandle->th_tags = th->th_tags;
                rc = dt_trans_stop(env, master_dev,
                                   top_th->tt_master_sub_thandle);
                OBD_FREE_PTR(top_th);
@@ -961,6 +965,7 @@ int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
                        CERROR("%s: cannot prepare updates: rc = %d\n",
                               master_dev->dd_lu_dev.ld_obd->obd_name, rc);
                        th->th_result = rc;
+                       write_updates = false;
                        GOTO(stop_master_trans, rc);
                }
 
@@ -979,6 +984,7 @@ int top_trans_stop(const struct lu_env *env, struct dt_device *master_dev,
                        CERROR("%s: write updates failed: rc = %d\n",
                               master_dev->dd_lu_dev.ld_obd->obd_name, rc);
                        th->th_result = rc;
+                       write_updates = false;
                        GOTO(stop_master_trans, rc);
                }
        }
@@ -991,10 +997,18 @@ stop_master_trans:
                        master_st->st_sub_th->th_local = th->th_local;
                if (th->th_sync)
                        master_st->st_sub_th->th_sync = th->th_sync;
-               master_st->st_sub_th->th_tags = th->th_tags;
                master_st->st_sub_th->th_result = th->th_result;
                rc = dt_trans_stop(env, master_st->st_dt, master_st->st_sub_th);
+               /* If it does not write_updates, then we call submit callback
+                * here, otherwise callback is done through
+                * osd(osp)_trans_commit_cb() */
+               if (!master_st->st_started &&
+                   !list_empty(&tmt->tmt_commit_list))
+                       sub_trans_commit_cb_internal(tmt,
+                                               master_st->st_sub_th, rc);
                if (rc < 0) {
+                       CERROR("%s: stop trans failed: rc = %d\n",
+                              master_dev->dd_lu_dev.ld_obd->obd_name, rc);
                        th->th_result = rc;
                        GOTO(stop_other_trans, rc);
                } else if (tur != NULL && tur->tur_update_records != NULL) {
@@ -1012,6 +1026,17 @@ stop_master_trans:
        /* Step 3: write updates to other MDTs */
        if (write_updates) {
                struct llog_update_record *lur;
+               if (OBD_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
+                       if (cfs_fail_val == 1) {
+                               long timeout = cfs_time_seconds(1) / 10;
+
+                               OBD_RACE(OBD_FAIL_OUT_OBJECT_MISS);
+                               set_current_state(TASK_UNINTERRUPTIBLE);
+                               schedule_timeout(schedule_timeout(timeout));
+                               cfs_fail_loc = 0;
+                       }
+                       cfs_fail_val++;
+               }
 
                /* Stop callback of master will add more updates and also update
                 * master transno, so merge the parameters and updates into one
@@ -1032,6 +1057,9 @@ stop_master_trans:
 
                        rc = sub_updates_write(env, lur, st);
                        if (rc < 0) {
+                               CERROR("%s: write updates failed: rc = %d\n",
+                                      st->st_dt->dd_lu_dev.ld_obd->obd_name,
+                                      rc);
                                th->th_result = rc;
                                break;
                        }
@@ -1048,12 +1076,15 @@ stop_other_trans:
                        st->st_sub_th->th_sync = th->th_sync;
                if (th->th_local)
                        st->st_sub_th->th_local = th->th_local;
-               st->st_sub_th->th_tags = th->th_tags;
                st->st_sub_th->th_result = th->th_result;
                rc = dt_trans_stop(env, st->st_sub_th->th_dev,
                                   st->st_sub_th);
-               if (unlikely(rc < 0 && th->th_result == 0))
-                       th->th_result = rc;
+               if (rc < 0) {
+                       CERROR("%s: stop trans failed: rc = %d\n",
+                              st->st_dt->dd_lu_dev.ld_obd->obd_name, rc);
+                       if (th->th_result == 0)
+                               th->th_result = rc;
+               }
        }
 
        rc = top_trans_wait_result(top_th);
@@ -1117,6 +1148,7 @@ create_sub_thandle_with_thandle(struct top_thandle *top_th,
        st->st_sub_th = sub_th;
 
        sub_th->th_top = &top_th->tt_super;
+       sub_thandle_register_stop_cb(st, top_th->tt_multiple_thandle);
        return st;
 }
 
@@ -1251,6 +1283,9 @@ static int distribute_txn_cancel_records(const struct lu_env *env,
        struct sub_thandle *st;
        ENTRY;
 
+       if (OBD_FAIL_CHECK(OBD_FAIL_TGT_TXN_NO_CANCEL))
+               RETURN(0);
+
        top_multiple_thandle_dump(tmt, D_INFO);
        /* Cancel update logs on other MDTs */
        list_for_each_entry(st, &tmt->tmt_sub_thandle_list, st_sub_list) {
@@ -1272,9 +1307,9 @@ static int distribute_txn_cancel_records(const struct lu_env *env,
                        rc = llog_cat_cancel_records(env, ctxt->loc_handle, 1,
                                                     cookie);
                        CDEBUG(D_HA, "%s: batchid %llu cancel update log "
-                              DOSTID ".%u : rc = %d\n", obd->obd_name,
+                              DFID".%u: rc = %d\n", obd->obd_name,
                               tmt->tmt_batchid,
-                              POSTID(&cookie->lgc_lgl.lgl_oi),
+                              PFID(&cookie->lgc_lgl.lgl_oi.oi_fid),
                               cookie->lgc_index, rc);
                }
 
@@ -1284,40 +1319,6 @@ static int distribute_txn_cancel_records(const struct lu_env *env,
        RETURN(0);
 }
 
-/**
- * Check if there are committed transaction
- *
- * Check if there are committed transaction in the distribute transaction
- * list, then cancel the update records for those committed transaction.
- * Because the distribute transaction in the list are sorted by batchid,
- * and cancellation will be done by batchid order, so we only check the first
- * the transaction(with lowest batchid) in the list.
- *
- * \param[in] lod      lod device where cancel thread is
- *
- * \retval             true if it is ready
- * \retval             false if it is not ready
- */
-static bool tdtd_ready_for_cancel_log(struct target_distribute_txn_data *tdtd)
-{
-       struct top_multiple_thandle     *tmt = NULL;
-       struct obd_device               *obd = tdtd->tdtd_lut->lut_obd;
-       bool    ready = false;
-
-       spin_lock(&tdtd->tdtd_batchid_lock);
-       if (!list_empty(&tdtd->tdtd_list)) {
-               tmt = list_entry(tdtd->tdtd_list.next,
-                                struct top_multiple_thandle, tmt_commit_list);
-               if (tmt->tmt_committed &&
-                   (!obd->obd_recovering || (obd->obd_recovering &&
-                   tmt->tmt_batchid <= tdtd->tdtd_committed_batchid)))
-                       ready = true;
-       }
-       spin_unlock(&tdtd->tdtd_batchid_lock);
-
-       return ready;
-}
-
 struct distribute_txn_bid_data {
        struct dt_txn_commit_cb  dtbd_cb;
        struct target_distribute_txn_data      *dtbd_tdtd;
@@ -1343,7 +1344,7 @@ static void distribute_txn_batchid_cb(struct lu_env *env,
        struct distribute_txn_bid_data          *dtbd = NULL;
        struct target_distribute_txn_data       *tdtd;
 
-       dtbd = container_of0(cb, struct distribute_txn_bid_data, dtbd_cb);
+       dtbd = container_of(cb, struct distribute_txn_bid_data, dtbd_cb);
        tdtd = dtbd->dtbd_tdtd;
 
        CDEBUG(D_HA, "%s: %llu batchid updated\n",
@@ -1353,8 +1354,8 @@ static void distribute_txn_batchid_cb(struct lu_env *env,
            !tdtd->tdtd_lut->lut_obd->obd_no_transno)
                tdtd->tdtd_committed_batchid = dtbd->dtbd_batchid;
        spin_unlock(&tdtd->tdtd_batchid_lock);
-       atomic_dec(&tdtd->tdtd_refcount);
-       wake_up(&tdtd->tdtd_commit_thread_waitq);
+       if (atomic_dec_and_test(&tdtd->tdtd_refcount))
+               wake_up_process(tdtd->tdtd_commit_task);
 
        OBD_FREE_PTR(dtbd);
 }
@@ -1395,6 +1396,7 @@ distribute_txn_commit_batchid_update(const struct lu_env *env,
 
        th = dt_trans_create(env, tdtd->tdtd_lut->lut_bottom);
        if (IS_ERR(th)) {
+               atomic_dec(&tdtd->tdtd_refcount);
                OBD_FREE_PTR(dtbd);
                RETURN(PTR_ERR(th));
        }
@@ -1420,13 +1422,15 @@ distribute_txn_commit_batchid_update(const struct lu_env *env,
        rc = dt_record_write(env, tdtd->tdtd_batchid_obj, &buf,
                             &off, th);
 
-       CDEBUG(D_INFO, "%s: update batchid "LPU64": rc = %d\n",
+       CDEBUG(D_INFO, "%s: update batchid %llu: rc = %d\n",
               tdtd->tdtd_lut->lut_obd->obd_name, batchid, rc);
 
 stop:
        dt_trans_stop(env, tdtd->tdtd_lut->lut_bottom, th);
-       if (rc < 0)
+       if (rc < 0) {
+               atomic_dec(&tdtd->tdtd_refcount);
                OBD_FREE_PTR(dtbd);
+       }
        RETURN(rc);
 }
 
@@ -1494,12 +1498,16 @@ distribute_txn_commit_batchid_init(const struct lu_env *env,
 
 out_put:
        if (rc < 0 && dt_obj != NULL) {
-               lu_object_put(env, &dt_obj->do_lu);
+               dt_object_put(env, dt_obj);
                tdtd->tdtd_batchid_obj = NULL;
        }
        return rc;
 }
 
+#ifndef TASK_IDLE
+#define TASK_IDLE TASK_INTERRUPTIBLE
+#endif
+
 /**
  * manage the distribute transaction thread
  *
@@ -1515,11 +1523,8 @@ out_put:
 static int distribute_txn_commit_thread(void *_arg)
 {
        struct target_distribute_txn_data *tdtd = _arg;
-       struct lu_target        *lut = tdtd->tdtd_lut;
-       struct ptlrpc_thread    *thread = &lut->lut_tdtd_commit_thread;
-       struct l_wait_info       lwi = { 0 };
-       struct lu_env            env;
-       struct list_head         list;
+       struct lu_env           *env = &tdtd->tdtd_env;
+       LIST_HEAD(list);
        int                      rc;
        struct top_multiple_thandle *tmt;
        struct top_multiple_thandle *tmp;
@@ -1527,21 +1532,13 @@ static int distribute_txn_commit_thread(void *_arg)
 
        ENTRY;
 
-       rc = lu_env_init(&env, LCT_LOCAL | LCT_MD_THREAD);
-       if (rc != 0)
-               RETURN(rc);
-
-       spin_lock(&tdtd->tdtd_batchid_lock);
-       thread->t_flags = SVC_RUNNING;
-       spin_unlock(&tdtd->tdtd_batchid_lock);
-       wake_up(&thread->t_ctl_waitq);
-       INIT_LIST_HEAD(&list);
 
-       CDEBUG(D_HA, "%s: start commit thread committed batchid "LPU64"\n",
+       CDEBUG(D_HA, "%s: start commit thread committed batchid %llu\n",
               tdtd->tdtd_lut->lut_obd->obd_name,
               tdtd->tdtd_committed_batchid);
 
-       while (distribute_txn_commit_thread_running(lut)) {
+       while (({set_current_state(TASK_IDLE);
+                !kthread_should_stop(); })) {
                spin_lock(&tdtd->tdtd_batchid_lock);
                list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
                                         tmt_commit_list) {
@@ -1554,11 +1551,13 @@ static int distribute_txn_commit_thread(void *_arg)
                         * the recoverying is done, unless the update records
                         * batchid < committed_batchid. */
                        if (tmt->tmt_batchid <= tdtd->tdtd_committed_batchid) {
+                               __set_current_state(TASK_RUNNING);
                                list_move_tail(&tmt->tmt_commit_list, &list);
                        } else if (!tdtd->tdtd_lut->lut_obd->obd_recovering) {
+                               __set_current_state(TASK_RUNNING);
                                LASSERTF(tmt->tmt_batchid >= batchid,
-                                        "tmt %p tmt_batchid: "LPU64", batchid "
-                                         LPU64"\n", tmt, tmt->tmt_batchid,
+                                        "tmt %p tmt_batchid: %llu, batchid "
+                                         "%llu\n", tmt, tmt->tmt_batchid,
                                         batchid);
                                /* There are three types of distribution
                                 * transaction result
@@ -1586,12 +1585,12 @@ static int distribute_txn_commit_thread(void *_arg)
                }
                spin_unlock(&tdtd->tdtd_batchid_lock);
 
-               CDEBUG(D_HA, "%s: batchid: "LPU64" committed batchid "
-                      LPU64"\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
+               CDEBUG(D_HA, "%s: batchid: %llu committed batchid "
+                      "%llu\n", tdtd->tdtd_lut->lut_obd->obd_name, batchid,
                       tdtd->tdtd_committed_batchid);
                /* update globally committed on a storage */
                if (batchid > tdtd->tdtd_committed_batchid) {
-                       rc = distribute_txn_commit_batchid_update(&env, tdtd,
+                       rc = distribute_txn_commit_batchid_update(env, tdtd,
                                                             batchid);
                        if (rc == 0)
                                batchid = 0;
@@ -1602,20 +1601,26 @@ static int distribute_txn_commit_thread(void *_arg)
                list_for_each_entry_safe(tmt, tmp, &list, tmt_commit_list) {
                        if (tmt->tmt_batchid > committed)
                                break;
+                       __set_current_state(TASK_RUNNING);
                        list_del_init(&tmt->tmt_commit_list);
                        if (tmt->tmt_result <= 0)
-                               distribute_txn_cancel_records(&env, tmt);
+                               distribute_txn_cancel_records(env, tmt);
                        top_multiple_thandle_put(tmt);
                }
 
-               l_wait_event(tdtd->tdtd_commit_thread_waitq,
-                            !distribute_txn_commit_thread_running(lut) ||
-                            committed < tdtd->tdtd_committed_batchid ||
-                            tdtd_ready_for_cancel_log(tdtd), &lwi);
-       };
+               if (!task_is_running(current))
+                       schedule();
 
-       l_wait_event(tdtd->tdtd_commit_thread_waitq,
-                    atomic_read(&tdtd->tdtd_refcount) == 0, &lwi);
+               if (OBD_FAIL_PRECHECK(OBD_FAIL_OUT_OBJECT_MISS)) {
+                       set_current_state(TASK_UNINTERRUPTIBLE);
+                       schedule_timeout(cfs_time_seconds(5));
+               }
+       }
+
+       while (({set_current_state(TASK_IDLE);
+                atomic_read(&tdtd->tdtd_refcount) != 0; }))
+               schedule();
+       __set_current_state(TASK_RUNNING);
 
        spin_lock(&tdtd->tdtd_batchid_lock);
        list_for_each_entry_safe(tmt, tmp, &tdtd->tdtd_list,
@@ -1630,11 +1635,6 @@ static int distribute_txn_commit_thread(void *_arg)
                top_multiple_thandle_dump(tmt, D_HA);
                top_multiple_thandle_put(tmt);
        }
-
-       thread->t_flags = SVC_STOPPED;
-       lu_env_fini(&env);
-       wake_up(&thread->t_ctl_waitq);
-
        RETURN(0);
 }
 
@@ -1655,7 +1655,6 @@ int distribute_txn_init(const struct lu_env *env,
                        __u32 index)
 {
        struct task_struct      *task;
-       struct l_wait_info       lwi = { 0 };
        int                     rc;
        ENTRY;
 
@@ -1669,23 +1668,31 @@ int distribute_txn_init(const struct lu_env *env,
 
        tdtd->tdtd_batchid = lut->lut_last_transno + 1;
 
-       init_waitqueue_head(&lut->lut_tdtd_commit_thread.t_ctl_waitq);
-       init_waitqueue_head(&tdtd->tdtd_commit_thread_waitq);
+       init_waitqueue_head(&tdtd->tdtd_recovery_threads_waitq);
        atomic_set(&tdtd->tdtd_refcount, 0);
+       atomic_set(&tdtd->tdtd_recovery_threads_count, 0);
 
        tdtd->tdtd_lut = lut;
+       if (lut->lut_bottom->dd_rdonly)
+               RETURN(0);
+
        rc = distribute_txn_commit_batchid_init(env, tdtd);
        if (rc != 0)
                RETURN(rc);
 
-       task = kthread_run(distribute_txn_commit_thread, tdtd, "tdtd-%u",
-                          index);
-       if (IS_ERR(task))
+       rc = lu_env_init(&tdtd->tdtd_env, LCT_LOCAL | LCT_MD_THREAD);
+       if (rc)
+               RETURN(rc);
+
+       task = kthread_create(distribute_txn_commit_thread, tdtd, "dist_txn-%u",
+                             index);
+       if (IS_ERR(task)) {
+               lu_env_fini(&tdtd->tdtd_env);
                RETURN(PTR_ERR(task));
+       }
+       tdtd->tdtd_commit_task = task;
+       wake_up_process(task);
 
-       l_wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
-                    distribute_txn_commit_thread_running(lut) ||
-                    distribute_txn_commit_thread_stopped(lut), &lwi);
        RETURN(0);
 }
 EXPORT_SYMBOL(distribute_txn_init);
@@ -1701,22 +1708,35 @@ EXPORT_SYMBOL(distribute_txn_init);
 void distribute_txn_fini(const struct lu_env *env,
                         struct target_distribute_txn_data *tdtd)
 {
-       struct lu_target *lut = tdtd->tdtd_lut;
+       struct top_multiple_thandle *tmt;
+       LIST_HEAD(list);
 
        /* Stop cancel thread */
-       if (lut == NULL || !distribute_txn_commit_thread_running(lut))
+       if (!tdtd->tdtd_commit_task)
                return;
 
+       kthread_stop(tdtd->tdtd_commit_task);
+       tdtd->tdtd_commit_task = NULL;
+
        spin_lock(&tdtd->tdtd_batchid_lock);
-       lut->lut_tdtd_commit_thread.t_flags = SVC_STOPPING;
+       list_splice_init(&tdtd->tdtd_list, &list);
        spin_unlock(&tdtd->tdtd_batchid_lock);
-       wake_up(&tdtd->tdtd_commit_thread_waitq);
-       wait_event(lut->lut_tdtd_commit_thread.t_ctl_waitq,
-                  lut->lut_tdtd_commit_thread.t_flags & SVC_STOPPED);
+
+       CDEBUG(D_INFO, "%s stopping distribute txn commit thread.\n",
+              tdtd->tdtd_lut->lut_obd->obd_name);
+       while ((tmt = list_first_entry_or_null(&list,
+                                              struct top_multiple_thandle,
+                                              tmt_commit_list)) != NULL) {
+               list_del_init(&tmt->tmt_commit_list);
+               top_multiple_thandle_dump(tmt, D_HA);
+               top_multiple_thandle_put(tmt);
+       }
+
+       lu_env_fini(&tdtd->tdtd_env);
 
        dtrq_list_destroy(tdtd);
        if (tdtd->tdtd_batchid_obj != NULL) {
-               lu_object_put(env, &tdtd->tdtd_batchid_obj->do_lu);
+               dt_object_put(env, tdtd->tdtd_batchid_obj);
                tdtd->tdtd_batchid_obj = NULL;
        }
 }