storage: fix a bug that auto repair may cause crash

The following sequence causes crash:

  1. connection1: open indexes
  2. connection2: open indexes
  3. connection1: start operations log check and detects crash!
  4. connection2: start operations log check and blocked.
  5. connection1: finish auto repair by recreating indexes.
  6. connection2: is unblocked and detects no crash but has opened indexes
     that are recreated connection1.
  7. connection2: use dropped index and is crashed.

TODO:

  * The same problem exists in wrapper mode.
  * Reduce lock scope. We can reduce lock scope by detecting "DISABLE
    KEYS" mode without opening indexes.

[groonga-dev,03608]

Reported by Hiroshi Kagami. Thanks!!!

ha_mroonga.cpp

@@ -4292,6 +4292,9 @@ int ha_mroonga::storage_open(const char *name, int mode, uint open_options)
   }
 
   if (!(open_options & HA_OPEN_FOR_REPAIR)) {
+    // TODO: Reduce lock scope
+    mrn::Lock lock(&mrn_operations_mutex);
+
     error = storage_open_indexes(name);
     if (error) {
       storage_close_columns();
@@ -4303,7 +4306,6 @@ int ha_mroonga::storage_open(const char *name, int mode, uint open_options)
     storage_set_keys_in_use();
 
     {
-      mrn::Lock lock(&mrn_operations_mutex);
       mrn::PathMapper mapper(name);
       const char *table_name = mapper.table_name();
       size_t table_name_size = strlen(table_name);