MySQL 派生表查询导致 Crash 的根源分析与解决方案

一、问题发现

在之前的 MySQL 8.0.32 使用中,发现使用以下带有派生表的 SQL 会导致 MySQL Crash,以下的sequence_table(2)替换为任何非常量表都行:

仅 MySQL 8.0.32 版本有影响。
EXPLAIN FORMAT=TREE 
select  
    trim(ref_15.c_ogj),
    0<>0 as c_lrcm63eani
  from 
    (select 
    0<>0 as c_ogj
  from 
   sequence_table(2) t1
  where 0<>0
  order by c_ogj asc) as ref_15;

Crash 的堆栈如下:

Thread 55 "mysqld" received signal SIGSEGV, Segmentation fault.
Item_view_ref::used_tables (this=0x7fff2418f410)
    at sql/item.h:6670
6670            table_map inner_map = ref_item()->used_tables(); ==> ref_item()为空指针,因此crash了
(gdb) bt
#0  Item_view_ref::used_tables (this=0x7fff2418f410)
    at sql/item.h:6670
#1  0x0000555558e978d1 in Item::const_item (this=0x7fff2418f410)
    at sql/item.h:2342
#2  0x0000555558ecc765 in Item_ref::print (this=0x7fff2418f410, thd=0x7fff24001050, 
    str=0x7fffc83ee7e0, query_type=(QT_TO_SYSTEM_CHARSET | QT_SHOW_SELECT_NUMBER))
    at sql/item.cc:9993
#3  0x000055555903b839 in Item_func_trim::print (this=0x7fff24120d20, thd=0x7fff24001050, 
    str=0x7fffc83ee7e0, query_type=(QT_TO_SYSTEM_CHARSET | QT_SHOW_SELECT_NUMBER))
    at sql/item_strfunc.cc:3244
#4  0x0000555558ea7fc5 in Item::print_item_w_name (this=0x7fff24120d20, thd=0x7fff24001050, 
    str=0x7fffc83ee7e0, query_type=(QT_TO_SYSTEM_CHARSET | QT_SHOW_SELECT_NUMBER))
    at sql/item.cc:727
#5  0x00005555593f18c0 in Query_block::print_item_list (this=0x7fff24120768, thd=0x7fff24001050, 
    str=0x7fffc83ee7e0, query_type=(QT_TO_SYSTEM_CHARSET | QT_SHOW_SELECT_NUMBER))
    at sql/sql_lex.cc:4041
#6  0x00005555593efb50 in Query_block::print_query_block (this=0x7fff24120768, 
    thd=0x7fff24001050, str=0x7fffc83ee7e0, 
    query_type=(QT_TO_SYSTEM_CHARSET | QT_SHOW_SELECT_NUMBER))
    at sql/sql_lex.cc:3614
#7  0x00005555593efa3d in Query_block::print (this=0x7fff24120768, thd=0x7fff24001050, 
    str=0x7fffc83ee7e0, query_type=(QT_TO_SYSTEM_CHARSET | QT_SHOW_SELECT_NUMBER))
    at sql/sql_lex.cc:3598
#8  0x00005555593ee556 in Query_expression::print (this=0x7fff24120670, thd=0x7fff24001050, 
    str=0x7fffc83ee7e0, query_type=(QT_TO_SYSTEM_CHARSET | QT_SHOW_SELECT_NUMBER))
    at sql/sql_lex.cc:3232
#9  0x0000555559a89c2c in print_query_for_explain (query_thd=0x7fff24001050, 
    unit=0x7fff24120670, str=0x7fffc83ee7e0)
    at sql/opt_explain.cc:2288
#10 0x0000555559a10b11 in PrintQueryPlan[abi:cxx11](THD*, THD const*, Query_expression*) (
    ethd=0x7fff24001050, query_thd=0x7fff24001050, unit=0x7fff24120670)
    at sql/join_optimizer/explain_access_path.cc:1894
#11 0x0000555559a8985a in ExplainIterator (ethd=0x7fff24001050, query_thd=0x7fff24001050, 
    unit=0x7fff24120670) at sql/opt_explain.cc:2205
#12 0x0000555559a89e91 in explain_query (explain_thd=0x7fff24001050, query_thd=0x7fff24001050, 
    unit=0x7fff24120670) at sql/opt_explain.cc:2359
#13 0x000055555955cd46 in Sql_cmd_dml::execute_inner (this=0x7fff24165630, thd=0x7fff24001050)

二、问题调查过程

调查执行 SQL 的 optimize 的过程,分析发现该 SQL 的 SQL 变换情况如下:

以下的 trim(ref_15.c_ogj) 执行完 find_order_in_list 后,Item_func_trimargs[0]->m_ref_item[0] 等于0<>0 as c_lrcm63eani,而不是0<>0 as c_ogj,这是因为c_lrcm63eani和c_ogj的名字都一样,都是0<>0,在find_order_in_list函数里面由于名字一样因此内层字段被外层替代了。而后在Item::clean_up_after_removal执行的时候,Item_func_ne即c_lrcm63eani因为出现了2次,因此执行了2次decrement_ref_count(),然而在Query_block::delete_unused_merged_columns函数却把0<>0 as c_lrcm63eani的Item置为空了,因为这个时候c_lrcm63eani的item->decrement_ref_count()以后ref_count()为0因此继续执行Item::clean_up_after_removal了。

EXPLAIN FORMAT=TREE 
select  
    trim(ref_15.c_ogj),
    0<>0 as c_lrcm63eani
  from 
    (select 
    0<>0 as c_ogj
  from 
   sequence_table(2) t1
  where 0<>0
  order by c_ogj asc) as ref_15;

查看函数调用过程发现 Query_block 在 prepare 的时候执行了 delete_unused_merged_columns,

-- 函数调用过程: Query_block::prepare -> Query_block::apply_local_transforms -> Query_block::delete_unused_merged_columns

bool find_order_in_list() {
  if (select_item != not_found_item) {
      if ((*order->item)->real_item() != (*select_item)->real_item()) {
        Item::Cleanup_after_removal_context ctx(
            thd->lex->current_query_block());

        (*order->item)
            ->walk(&Item::clean_up_after_removal, walk_options,  ==>Item_func_ne执行了2次,也执行了2次decrement_ref_count()
                   pointer_cast<uchar *>(&ctx));
      }
  }
}

bool Query_block::apply_local_transforms(THD *thd, bool prune) {
  DBUG_TRACE;

  assert(first_execution);
  -- 这个函数把((Item_func *)&fields[0][0])->args[0]->m_ref_item[0]给删了
  if (derived_table_count) delete_unused_merged_columns(&m_table_nest);
}

void Query_block::delete_unused_merged_columns(
    mem_root_deque<Table_ref *> *tables) {
  DBUG_TRACE;

  for (Table_ref *tl : *tables) {
    if (tl->nested_join == nullptr) continue;
    if (tl->is_merged()) {
      for (Field_translator *transl = tl->field_translation;
           transl < tl->field_translation_end; transl++) {
        Item *const item = transl->item;
        // Decrement the ref count as its no more used in
        // select list.
        if (item->decrement_ref_count()) continue; -- 因为执行完decrement_ref_count()以后返回的m_ref_count=0因此不会跳出这个循环

        // Cleanup the item since its not referenced from
        // anywhere.
        assert(item->fixed);
        Item::Cleanup_after_removal_context ctx(this);
        item->walk(&Item::clean_up_after_removal, walk_options,
                   pointer_cast<uchar *>(&ctx));
        transl->item = nullptr; -- 这个地方把Item_view_ref引用的Item_func_ne对象置为空了,即把trim函数参数的c_lrcm63eani列删除了
      }
    }
    delete_unused_merged_columns(&tl->nested_join->m_tables);
  }
}

三、解决方案

通过上面的分析,我们可以发现问题在于多执行了一次Item::clean_up_after_removal,随后在 MySQL 最新代码尝试执行以上 SQL 发现该 BUG 已经被修复,找到相关修复代码,可以发现以下修复代码。

相关commit ID号为: 2171a1260e2cdbbd379646be8ff6413a92fd48f4

-- 相关修复代码如下:
@@ -7575,7 +7865,6 @@ bool Item::clean_up_after_removal(uchar *arg) {
 
   if (reference_count() > 1) {
     (void)decrement_ref_count();
+    ctx->stop_at(this);
   }
   return false;
 }

修改完查看一下这个函数的堆栈信息:

#0  Item::clean_up_after_removal (this=0x2, 
    arg=0x41 <error: Cannot access memory at address 0x41>)
    at sql/item.cc:9236
#1  0x0000555558fea5a8 in Item::walk (this=0x7fff2c338db8, processor=&virtual table offset 864, 
    walk=7, arg=0x7fffc83ee4b0 "") at sql/item.h:2543
#2  0x00005555596cc6f2 in find_order_in_list (thd=0x7fff2c001070, ref_item_array=..., 
    tables=0x7fff2c330b90, order=0x7fff2c32eae8, fields=0x7fff2c32fb20, is_group_field=false, 
    is_window_order=false) at sql/sql_resolver.cc:4625
#3  0x00005555596cd0ae in setup_order (thd=0x7fff2c001070, ref_item_array=..., 
    tables=0x7fff2c330b90, fields=0x7fff2c32fb20, order=0x7fff2c32eae8)
    at sql/sql_resolver.cc:4811
#4  0x00005555596bf528 in Query_block::prepare (this=0x7fff2c32fae0, thd=0x7fff2c001070, 
    insert_field_list=0x0) at sql/sql_resolver.cc:400
#5  0x00005555597d035d in Query_expression::prepare (this=0x7fff2c32f9e8, thd=0x7fff2c001070, 
    sel_result=0x7fff2c33b2a8, insert_field_list=0x0, added_options=0, removed_options=0)
    at sql/sql_union.cc:758
#6  0x0000555559590772 in Table_ref::resolve_derived (this=0x7fff2c339790, thd=0x7fff2c001070, 
    apply_semijoin=true) at sql/sql_derived.cc:451
#7  0x00005555596c2a80 in Query_block::resolve_placeholder_tables (this=0x7fff2c333f08, 
    thd=0x7fff2c001070, apply_semijoin=true)
    at sql/sql_resolver.cc:1408
#8  0x00005555596bea62 in Query_block::prepare (this=0x7fff2c333f08, thd=0x7fff2c001070, 
    insert_field_list=0x0) at sql/sql_resolver.cc:265

对于0<>0 as c_lrcm63eani这个Item_func_ne对象,执行到Item::clean_up_after_removal的时候,因为reference_count() > 1因此会执行新添加的ctx->stop_at(this),等到下一次再执行到这个Item_func_neclean_up_after_removal()函数的时候,就会因为ctx->is_stopped(this)而直接返回,不再执行一次decrement_ref_count(),从而避免了执行后面的transl->item = nullptr

bool find_order_in_list() {
  if (select_item != not_found_item) {
      if ((*order->item)->real_item() != (*select_item)->real_item()) {
        Item::Cleanup_after_removal_context ctx(
            thd->lex->current_query_block());

        (*order->item)
            ->walk(&Item::clean_up_after_removal, walk_options,  -- Item_func_ne执行了2次,而只执行了一次decrement_ref_count()
                   pointer_cast<uchar *>(&ctx));
      }
  }
}

void Query_block::delete_unused_merged_columns(
    mem_root_deque<Table_ref *> *tables) {
  DBUG_TRACE;

  for (Table_ref *tl : *tables) {
    if (tl->nested_join == nullptr) continue;
    if (tl->is_merged()) {
      for (Field_translator *transl = tl->field_translation;
           transl < tl->field_translation_end; transl++) {
        Item *const item = transl->item;
        // Decrement the ref count as its no more used in
        // select list.
        if (item->decrement_ref_count()) continue; 因为执行完decrement_ref_count()以后返回的m_ref_count=1因此不会继续执行后面的置空设置

        // Cleanup the item since its not referenced from
        // anywhere.
        assert(item->fixed);
        Item::Cleanup_after_removal_context ctx(this);
        item->walk(&Item::clean_up_after_removal, walk_options,
                   pointer_cast<uchar *>(&ctx));
        transl->item = nullptr; ==>这个地方不会运行到
      }
    }
    delete_unused_merged_columns(&tl->nested_join->m_tables);
  }
}

四、问题总结

通过以上分析我们可以发现,对于复杂的 SQL 会执行复杂的 Item 变换和删除不需要的 Item,但是正是由于这样才更容易导致 Crash 的出现。分析类似这样的 Crash 问题的时候,因为涉及代码量大,代码逻辑复杂往往很难找到相关修复代码,因此需要对代码运行流程比较熟悉,同时要有相关复杂问题解决的经验才能更好的应对这类问题。


Enjoy GreatSQL :)

## 关于 GreatSQL

GreatSQL是适用于金融级应用的国内自主开源数据库,具备高性能、高可靠、高易用性、高安全等多个核心特性,可以作为MySQL或Percona Server的可选替换,用于线上生产环境,且完全免费并兼容MySQL或Percona Server。

相关链接: GreatSQL社区 Gitee GitHub Bilibili

GreatSQL社区:

社区博客有奖征稿详情:https://greatsql.cn/thread-100-1-1.html

image-20230105161905827

技术交流群:

微信:扫码添加GreatSQL社区助手微信好友,发送验证信息加群

image-20221030163217640


GreatSQL社区
8 声望6 粉丝

GreatSQL是由万里数据库维护的MySQL分支,专注于提升MGR可靠性及性能,支持InnoDB并行查询特性,是适用于金融级应用的MySQL分支版本。