Bug #57002 Assert in upgrade_shared_lock_to_exclusive()

           for ALTER TABLE + MERGE tables

The patch for Bug#56292 changed how metadata locks are taken for MERGE
tables. After the patch, locking the MERGE table will also lock the
children tables with the same metadata lock type. This means that 
LOCK TABLES on a MERGE table also will implicitly do LOCK TABLES on
the children tables.

A consequence of this change, is that it is possible to do LOCK TABLES
on a child table both explicitly and implicitly with the same statement
and that these two locks can be of different strength. For example,
LOCK TABLES child READ, merge WRITE.

In LOCK TABLES mode, we are not allowed to take new locks and each
statement must therefore try to find an existing TABLE instance with
a suitable lock. The code that searched for a suitable TABLE instance,
only considered table level locks. If a child table was locked twice,
it was therefore possible for this code to find a TABLE instance with
suitable table level locks but without suitable metadata lock.

This problem caused the assert in upgrade_shared_lock_to_exclusive()
to be triggered as it tried to upgrade a MDL_SHARED lock to
EXCLUSIVE. The problem was a regression caused by the patch for
Bug#56292.

This patch fixes the problem by partially reverting the changes
done by Bug#56292. Now, the children tables will only use the
same metadata lock as the MERGE table for MDL_SHARED_NO_WRITE
when not in locked tables mode. This means that LOCK TABLE
on a MERGE table will not implicitly lock the children tables.
This still fixes the original problem in Bug#56292 without
causing a regression.

Test case added to merge.test.

1 files changed, 25 insertions, 2 deletions

diff --git a/storage/myisammrg/ha_myisammrg.cc b/storage/myisammrg/ha_myisammrg.cc
index 466c66ec769..26e54d61101 100644
--- a/storage/myisammrg/ha_myisammrg.cc
+++ b/storage/myisammrg/ha_myisammrg.cc
@@ -478,8 +478,31 @@ int ha_myisammrg::add_children_list(void)
     /* Set the expected table version, to not cause spurious re-prepare. */
     child_l->set_table_ref_id(mrg_child_def->get_child_table_ref_type(),
                               mrg_child_def->get_child_def_version());
-    /* Use the same metadata lock type for children. */
-    child_l->mdl_request.set_type(parent_l->mdl_request.type);
+    /*
+      For statements which acquire a SNW metadata lock on a parent table and
+      then later try to upgrade it to an X lock (e.g. ALTER TABLE), SNW
+      locks should be also taken on the children tables.
+
+      Otherwise we end up in a situation where the thread trying to upgrade SNW
+      to X lock on the parent also holds a SR metadata lock and a read
+      thr_lock.c lock on the child. As a result, another thread might be
+      blocked on the thr_lock.c lock for the child after successfully acquiring
+      a SR or SW metadata lock on it. If at the same time this second thread
+      has a shared metadata lock on the parent table or there is some other
+      thread which has a shared metadata lock on the parent and is waiting for
+      this second thread, we get a deadlock. This deadlock cannot be properly
+      detected by the MDL subsystem as part of the waiting happens within
+      thr_lock.c. By taking SNW locks on the child tables we ensure that any
+      thread which waits for a thread doing SNW -> X upgrade, does this within
+      the MDL subsystem and thus potential deadlocks are exposed to the deadlock
+      detector.
+
+      We don't do the same thing for SNRW locks as this would allow
+      DDL on implicitly locked underlying tables of a MERGE table.
+    */
+    if (! thd->locked_tables_mode &&
+        parent_l->mdl_request.type == MDL_SHARED_NO_WRITE)
+      child_l->mdl_request.set_type(MDL_SHARED_NO_WRITE);
     /* Link TABLE_LIST object into the children list. */
     if (this->children_last_l)
       child_l->prev_global= this->children_last_l;