"marked as crashed", "zerofilling" and "wrong data in bitmap" when recovering Aria tables

[Philip Stoev]

After running the engine_stress RQG scenario and recovery, mysqld reported the following:

# 2010-09-10T13:26:31 100910 13:26:31 [ERROR] mysqld: Table './test/table100_maria_key_pk_parts_2_int_autoinc#P#p0' is marked as crashed and should be repaired
# 2010-09-10T13:26:31 100910 13:26:31 [Note] Zerofilling table: './test/table100_maria_key_pk_parts_2_int_autoinc'
# 2010-09-10T13:26:31 100910 13:26:31 [Note] Zerofilling table: './test/table100_maria_key_pk_parts_2_int_autoinc'

then CHECK TABLE `test`.`table100_maria_key_pk_parts_2_int_autoinc` EXTENDED reported:

            'Page 1: Wrong data in bitmap. Page_type: 1 full: 1 empty_space: 6418 Bitmap-bits: 3'

This happened on a mysqld-based recovery using existing tables. Recovery using maria_read_log and empty datadir worked as expected.

bzr version-info:

date: 2010-09-10 02:42:12 +0300
build-date: 2010-09-10 16:52:51 +0300
revno: 2928
branch-nick: 5.1

RQG command line:

/home/buildbot/randgen/runall.pl \
--queries=1M \
--engine=Maria \
--mysqld=--safe-mode \
--mysqld=--sync-sys=0 \
--mysqld=--log-output=file \
--mysqld=--maria_log_purge_type=at_flush \
--reporters=ErrorLog,Backtrace,Recovery,Shutdown \
--duration=120 \
--threads=5 \
--rows=100 \
--mask-level=0 \
--mysqld=--maria-repair-threads=2 \
--mysqld=--loose-maria-group-commit=soft \
--mysqld=--loose-maria_group_commit_interval=100 \
--mysqld=--maria-checkpoint-interval=1 \
--mysqld=--maria-block-size=16K \
--mask=59718 \
--queries=100000000 \
--mask=59718 \
--seed=time \
--basedir=. \
--vardir=/home/buildbot/rqg-buildbot-slave/rqg-weekend/vardir-maria_recovery_weekend/current \
--grammar=conf/engines/engine_stress.yy \
--gendata=conf/engines/engine_stress.zz

[Michael Widenius (monty)]

Found bug, fixed, tested and pushed in 5.1
- Tables marked as opened was not properly unmarked on recovery if there was not changes since checkpoint
- zerofill of tables put wrong data in bitmap if directory for page was full
- Tables was thought as 'moved' during recovery if they had a create_lsn bigger than the lsn in the control file.