"./mtr --mysqld=--default-storage-engine=maria mysqldump".
First problem was use of INSERT DELAYED and MERGE tables without
specifying that the tables to create should always be MyISAM.
After fixing this, no rows were returned by the final SELECT of the
"BUG 19025" portion of the test. Simplified problem was:
LOCK TABLES `t1` WRITE;
/*!40000 ALTER TABLE `t1` DISABLE KEYS */;
INSERT INTO `t1` VALUES ('bla',1000),('bla',1001),('bla',1002);
/*!40000 ALTER TABLE `t1` ENABLE KEYS */;
UNLOCK TABLES;
select * from t1;
The SELECT would find no rows. Reason: ENABLE KEYS does a maria_repair();
but data pages are still in the page cache and not on disk (because
they were not flushed because maria_lock_database(F_UNLCK) was
not called at the end of INSERT because under LOCK TABLES).
At start of maria_repair(), sort_info.filelength is set to the
physical size of the data file (=> too small because pages are in
cache and not on disk).
Then in sort_get_next_record(), when seeing end-of-file, this is done:
sort_param->max_pos= sort_info->filelength;
Further in maria_repair(), this is done:
info->state->data_file_length= sort_param.max_pos;
and so data_file_length is smaller (0) than reality (16384).
This makes SELECT think EOF is where it is not, and thus find
no rows.
This is fixed by flushing all data pages at the start of maria_repair()
(no performance problem is introduced as in common cases where
ALTER TABLE is not under LOCK TABLES, the previous statement did
this flush anyway).
Another reason to do this flush is that, if not doing it, old cached
pages might go down onto the repaired data file at a later point
and thus corrupt it (assume a REPAIR non-QUICK).
A similar bug is fixed:
LOCK TABLES WRITE; INSERT; CHECK TABLE;
reports "Size of datafile is: 0         Should be: 16384"
again because the physical size was read without a preliminary
page cache flush.


mysql-test/r/maria.result:
  result update
mysql-test/r/mysqldump.result:
  result update
mysql-test/t/maria.test:
  adding test for fixed bug in LOCK TABLES + CHECK TABLE + block format.
  Disabling portion which hits "incorrect key file" but still
  letting it make the test fail (Monty to fix).
mysql-test/t/mysqldump.test:
  in places where test expects engine to support INSERT DELAYED and
  be includable in a MERGE table, i.e. be MyISAM, we explicitely
  ask for MyISAM.
storage/maria/ma_check.c:
  Before reading the data file's physical size with my_seek(MY_SEEK_END)
  during maria_chk_size() and maria_repair(), we must flush this
  data file, otherwise physical size is misleading and leads to
  - CHECK TABLE finding the table corrupted ("size of datafile should be"
  error)
  - ALTER TABLE ENABLE KEYS losing rows (maria_repair()
  setting data_file_length to a too small value => later SELECT does
  not find rows though they are in the data file).
  This fixes the "mysqldump.test" failure.
  sort_info.filelength contains the physical size, re-using it.