maria-developers team mailing list archive

[Sergey Vojtovich <svoj@xxxxxxxxxxx>]

Sergei,

On Thu, Jun 22, 2017 at 07:57:21PM +0200, Sergei Golubchik wrote:
> Hi, Sergey!
> 
> On Jun 22, Sergey Vojtovich wrote:
> > revision-id: 922f9700a7fb084e3f7d6cff56e27e24fc603ed7 (mariadb-10.2.6-58-g922f970)
> > parent(s): 62021f391a42c5577190aa43cb8ad91e56235b46
> > committer: Sergey Vojtovich
> > timestamp: 2017-06-22 17:35:36 +0400
> > message:
> > 
> > MDEV-12882 - Assertion failed in MDL_context::upgrade_shared_lock
> > 
> > Relaxed assertion (in MySQL it was removed).
> > For "LOCK TABLES t1 WRITE CONCURRENT, t1 READ" upgrade lock to weakest
> > existing suitable lock, which is MDL_SHARED_NO_READ_WRITE.
> > 
> > diff --git a/mysql-test/t/mdl.test b/mysql-test/t/mdl.test
> > new file mode 100644
> > index 0000000..5a74ae5
> > --- /dev/null
> > +++ b/mysql-test/t/mdl.test
> > @@ -0,0 +1,15 @@
> 
> Hm. I'd totally expect main.mdl test to exist for years :)
Me too. I guess in most cases tests for MDL need debug_sync, thus most tests
went into mdl_sync.test.

> 
> > diff --git a/sql/sql_parse.cc b/sql/sql_parse.cc
> > index 412cd1d..7c910e8 100644
> > --- a/sql/sql_parse.cc
> > +++ b/sql/sql_parse.cc
> > @@ -2787,6 +2787,7 @@ static bool lock_tables_open_and_lock_tables(THD *thd, TABLE_LIST *tables)
> >                 ! table->prelocking_placeholder &&
> >                 table->table->file->lock_count() == 0)
> >        {
> > +        enum enum_mdl_type lock_type;
> >          /*
> >            In case when LOCK TABLE ... READ LOCAL was issued for table with
> >            storage engine which doesn't support READ LOCAL option and doesn't
> > @@ -2799,9 +2800,12 @@ static bool lock_tables_open_and_lock_tables(THD *thd, TABLE_LIST *tables)
> >          deadlock_handler.init();
> >          thd->push_internal_handler(&deadlock_handler);
> >  
> > +        lock_type= table->table->mdl_ticket->get_type() == MDL_SHARED_WRITE ?
> > +                   MDL_SHARED_NO_READ_WRITE : MDL_SHARED_READ_ONLY;
> > +
> >          bool result= thd->mdl_context.upgrade_shared_lock(
> >                                          table->table->mdl_ticket,
> > -                                        MDL_SHARED_READ_ONLY,
> > +                                        lock_type,
> >                                          thd->variables.lock_wait_timeout);
> 
> This only works if you lock WRITE, then READ. If you do it the other way
> around, you'll get an error. Isn't it inconsistent?
Well, it works the other way around too. But in this case connection holds two
locks: READ_ONLY and WRITE. Which is even better than NO_READ_WRITE, but I
thought connection never acquires more than one lock for a table.

Apparently number of locks acquired may depends on the oreder they were
requested. I added "reverse" test to my patch.

Thanks,
Sergey