数据库内核月报

数据库内核月报 - 2018 / 11

MySQL · 引擎特性 · InnoDB MVCC 相关实现

Author: mianren

InnoDB支持MVCC来提高系统读写并发性能。InnoDB MVCC的实现基于Undo log,通过回滚段来构建需要的版本记录。通过ReadView来判断哪些版本的数据可见。同时Purge线程是通过ReadView来清理旧版本数据。MVCC的相关知识在过去的月报中已有涉及,这里笔者从部分相关实现的角度做一个学习与分享。代码基于MySQL8.0。

之前月报涉及相关知识的有:

MySQL · 引擎特性 · InnoDB undo log 漫游

MySQL · 引擎特性 · InnoDB 事务系统

MySQL · 引擎特性 · InnoDB 事务子系统介绍

Undo log

Undo log可以用来做事务的回滚操作,保证事务的原子性。同时可以用来构建数据修改之前的版本,支持多版本读。

InnoDB表数据组织方式是主键聚簇索引。二级索引通过索引键值加主键值组合来唯一确定一条记录。聚簇索引和二级索引都包含了DELETED BIT标记位来标识记录是否被删除,真正的删除在事务commit之后且没有读会引用该版本数据的时候。在聚簇索引上还有一些额外信息会存储,6字节的DB_TRX_ID字段,表示最近一次插入或者更新该记录的事务ID。7字节的DB_ROLL_PTR字段,指向该记录的rollback segment的undo log记录。6字节的DB_ROW_ID,当有新数据插入的时候会自动递增。当表上没有用户主键的时候,InnoDB会自动产生聚集索引,包含DB_ROW_ID字段。

对于聚簇索引,更新是在原记录位置更新,通过记录指向undo log的隐藏列来重构早期版本的数据。但对于二级索引,是没有聚簇索引上的这些隐藏列的,因此无法在原记录位置更新。当二级索引更新的时候,需要将原记录标记为删除,再插入新的数据记录。当快照读通过二级索引读取数据发现deleted标识或者更新的时候,如果二级索引页上无法判断可见性,InnoDB会查看聚簇索引上的记录行,通过行上的DB_TRX_ID判断可见性,找到正确的可见版本数据。

当用mvcc读取的时候(row_search_mvcc),对于聚簇索引,当拿到一条记录后,会先通过函数lock_clust_rec_cons_read_sees判断可见性,如果不可见会再构建老版本数据row_vers_build_for_consistent_read。

dberr_t row_search_mvcc(byte *buf, page_cur_mode_t mode,
                        row_prebuilt_t *prebuilt, ulint match_mode,
                        ulint direction) {
  ...
  /* This is a non-locking consistent read: if necessary, fetch
  a previous version of the record */

  if (trx->isolation_level == TRX_ISO_READ_UNCOMMITTED) {
      /* Do nothing: we let a non-locking SELECT read the
      latest version of the record */

  } else if (index == clust_index) {
    /* Fetch a previous version of the row if the current
    one is not visible in the snapshot; if we have a very
    high force recovery level set, we try to avoid crashes
    by skipping this lookup */

    if (srv_force_recovery < 5 &&
        !lock_clust_rec_cons_read_sees(rec, index, offsets,
        trx_get_read_view(trx))) {
      rec_t *old_vers;
      /* The following call returns 'offsets' associated with 'old_vers' */
      err = row_sel_build_prev_vers_for_mysql(
          trx->read_view, clust_index, prebuilt, rec, &offsets, &heap,
          &old_vers, need_vrow ? &vrow : NULL, &mtr,
          prebuilt->get_lob_undo());
}

对于二级索引,拿到记录会先调用lock_sec_rec_cons_read_sees判断page上记录的最近一次修改trx id是否小于m_up_limit_id,如果小于即该page上数据可见,否则即调用row_search_idx_cond_check检查可见性,对于ICP,索引条件下推的,可以先判断索引条件是否满足条件,这样避免不满足条件的行回表;对于满足条件的行则回表查看可见性。

dberr_t row_search_mvcc(byte *buf, page_cur_mode_t mode,
                        row_prebuilt_t *prebuilt, ulint match_mode,
                        ulint direction) {
  ...
  /* This is a non-locking consistent read: if necessary, fetch
  a previous version of the record */

  if (trx->isolation_level == TRX_ISO_READ_UNCOMMITTED) {
      /* Do nothing: we let a non-locking SELECT read the
      latest version of the record */

  } else if (index == clust_index) {
    ...
  } else {
    /* We are looking into a non-clustered index,
    and to get the right version of the record we
    have to look also into the clustered index: this
    is necessary, because we can only get the undo
    information via the clustered index record. */

    ut_ad(!index->is_clustered());

    if (!srv_read_only_mode &&
        !lock_sec_rec_cons_read_sees(rec, index, trx->read_view)) {
      /* We should look at the clustered index.
      However, as this is a non-locking read,
      we can skip the clustered index lookup if
      the condition does not match the secondary
      index entry. */
      switch (row_search_idx_cond_check(buf, prebuilt, rec, offsets)) {
        case ICP_NO_MATCH:
          goto next_rec;
        case ICP_OUT_OF_RANGE:
          err = DB_RECORD_NOT_FOUND;
          goto idx_cond_failed;
        case ICP_MATCH:
          goto requires_clust_rec;
      }
      ...
}

bool lock_sec_rec_cons_read_sees(
    const rec_t *rec,          /*!< in: user record which
                               should be read or passed over
                               by a read cursor */
    const dict_index_t *index, /*!< in: index */
    const ReadView *view)      /*!< in: consistent read view */
{
  ...
  trx_id_t max_trx_id = page_get_max_trx_id(page_align(rec));

  ut_ad(max_trx_id > 0);

  return (view->sees(max_trx_id));
}

在Undo log中会记录TRX_UNDO_TRX_ID事务ID和TRX_UNDO_TRX_NO事务Commit时的number值。其他的信息可以参考MySQL · 引擎特性 · InnoDB undo log 漫游

当事务为读写事务的时候,事务会获取trx_id。

/** Allocates a new transaction id.
 @return new, allocated trx id */
UNIV_INLINE
trx_id_t trx_sys_get_new_trx_id() {
  ut_ad(trx_sys_mutex_own());

  /* VERY important: after the database is started, max_trx_id value is
  divisible by TRX_SYS_TRX_ID_WRITE_MARGIN, and the following if
  will evaluate to TRUE when this function is first time called,
  and the value for trx id will be written to disk-based header!
  Thus trx id values will not overlap when the database is
  repeatedly started! */

  if (!(trx_sys->max_trx_id % TRX_SYS_TRX_ID_WRITE_MARGIN)) {
    trx_sys_flush_max_trx_id();
  }

  return (trx_sys->max_trx_id++);
}

当事务commit时会获取新的系统trx id作为trx_no。

trx_commit_low->trx_write_serialisation_history->trx_serialisation_number_get

/** Set the transaction serialisation number.
 @return true if the transaction number was added to the serialisation_list. */
static bool trx_serialisation_number_get(
    trx_t *trx,                         /*!< in/out: transaction */
    trx_undo_ptr_t *redo_rseg_undo_ptr, /*!< in/out: Set trx
                                        serialisation number in
                                        referred undo rseg. */
    trx_undo_ptr_t *temp_rseg_undo_ptr) /*!< in/out: Set trx
                                        serialisation number in
                                        referred undo rseg. */
{
    ...
    trx->no = trx_sys_get_new_trx_id();
    ...                                    
}

由于Undo log会保留直到事务提交同时没有其他快照读引用后才会purge。所以需要尽量避免长语句或长事务的执行,避免因此导致的undo堆积或者undo链太长使读取变慢。

Read View

ReadView主要结构

ReadView可见性判断:

bool changes_visible(trx_id_t id, const table_name_t &name) const
    MY_ATTRIBUTE((warn_unused_result)) {
  ut_ad(id > 0);

  if (id < m_up_limit_id || id == m_creator_trx_id) {
    return (true);
  }

  check_trx_id_sanity(id, name);

  if (id >= m_low_limit_id) {
    return (false);

  } else if (m_ids.empty()) {
    return (true);
  }

  const ids_t::value_type *p = m_ids.data();

  return (!std::binary_search(p, p + m_ids.size(), id));
}

Class MVCC封装了ReadView相关的访问。内部成员变量有 m_free存放释放的read view用来reuse避免重新构造。m_views存放active和closed状态的read view。该类提供的主要函数有

Semi consistent read

对于RC隔离级别或者设置innodb_locks_unsafe_for_binlog的情况下,当发生表扫描的UPDATE语句,如果数据行上有锁,UPDATE会先查看最近一次提交的数据是否满足条件,利用undo构建最近一次提交的数据。当满足条件再去读最新修改的行,这一次再等锁加锁,避免锁的等待。

row_search_mvcc()
{
  case DB_LOCK_WAIT:
    /* Lock wait for R-tree should already
    be handled in sel_set_rtr_rec_lock() */
    ut_ad(!dict_index_is_spatial(index));
    /* Never unlock rows that were part of a conflict. */
    std::fill_n(prebuilt->new_rec_lock, row_prebuilt_t::LOCK_COUNT, false);

    if (UNIV_LIKELY(prebuilt->row_read_type !=
                    ROW_READ_TRY_SEMI_CONSISTENT) ||
        unique_search || index != clust_index) {
      goto lock_wait_or_error;
    }

    /* The following call returns 'offsets' associated with 'old_vers' */
    row_sel_build_committed_vers_for_mysql(clust_index, prebuilt, rec,
                                           &offsets, &heap, &old_vers,
                                           need_vrow ? &vrow : NULL, &mtr);
}

这里当查询为unique_search并没有走semi consistent read,即对于’update t set … where pk = xx’的语句不会走semi consistent read。这里原因是bug#52663,在部分代码实现约束仅table scan的执行才可以。

同时Semi consistent read由于采用了非冲突串行化的处理方式,因此只能用在RC隔离级别或者设置innodb_locks_unsafe_for_binlog的情况下使用。

总结

InnoDB的多版本并不是直接存储多个版本的数据,而是所有更改操作利用行锁做并发控制,这样对某一行的更新操作是串行化的,然后用Undo log记录串行化的结果。当快照读的时候,利用Undo log重建需要读取版本的数据,从而实现读写并发。