maria-developers team mailing list archive

[Marko Mäkelä <marko.makela@xxxxxxxxxxx>]

Hi Otto, all,

For what it is worth, the amd64-ubuntu-2204-msan builder is only being
set up. Unlike the older clang-13 based builder, it revealed a few
surprises, as you can read in
https://jira.mariadb.org/browse/MDEV-30936. I helped with fixing
trouble related to that, because I heard that there is a plan to make
it a mandatory builder for branch protection.

I took a deeper look at one of the failures that you pointed to:
https://buildbot.mariadb.org/#/builders/561/builds/119
(amd64-ubuntu-2204-msan failure for 10.8). The client-side log ends
like this:

stress.wrapper                           w4 [ skipped ]  Only to be
run with mtr --stress
stress.deadlock_drop_table               w9 [ pass ]     64
sys_vars.alter_algorithm_basic           w4 [ pass ]     18
stress.misc                              w10 [ pass ]   3252
innodb.innodb_bug52663 'innodb'          w5 [ pass ]  50025
innodb_gis.rtree_search 'innodb'         w3 [ pass ]   7166
innodb_fts.sync_ddl 'innodb,orig'        w2 [ pass ]  15247
innodb.log_corruption 'innodb'           w1 [ pass ]  44124
stress.ddl_memory                        w6 [ pass ]  159732
stress.ddl_myisam                        w7 [ pass ]  162212
command timed out: 600 seconds without output running [b'bash',
b'-xc', b'cd mysql-test && LD_LIBRARY_PATH=/msan-libs
MSAN_OPTIONS=abort_on_error=1:poison_in_dtor=0 ./mtr --mem --big-test
--force --retry=0 --skip-test=.*compression.* --max-test-fail=100
--parallel=$(expr 5 \\* 2)'], attempting to kill
process killed by signal 9
program finished with exit code -1
elapsedTime=2771.840262

After some analysis, I found the log of the worker that is apparently hung:
https://ci.mariadb.org/33865/logs/amd64-ubuntu-2204-msan/mysqld.1.err.8

The last test in that log corresponds to the following in the client-side log:
innodb.innodb_page_compressed 'innodb,lzma' w8 [ fail ]
This test fails because the LZMA library is available in the test
environment, but not in MSAN-instrumented form. In MemorySanitizer,
any data returned by uninstrumented code will be assumed to be
uninitialized. I hope that the CI maintainers will address this
failure soon, if it was not done already.

After that test failure, that server was restarted for running the
test versioning.alter, but there is no mention of that test in the
client log, and no sign of an attempted server shutdown either. For
this type of test driver timeout, the ticket
https://jira.mariadb.org/browse/MDEV-30836 has been filed, because the
claimed fix in https://jira.mariadb.org/browse/MDEV-29023 only appears
to actually work in the 10.3 branch.

> https://github.com/MariaDB/server/tree/10.3 / 4c4939b
> NOT OK:
> - amd64-debian-10-debug-embedded fails on sql_sequence.kill
> - amd64-fedora-36 fails on main.func_math

The undefined behaviour in main.func_math was only fixed starting with
the 10.5 branch: https://jira.mariadb.org/browse/MDEV-26645
Maybe we should just disable building 10.3 and 10.4 on Fedora 36 or later?

> https://github.com/MariaDB/server/tree/10.5 / ac5a534
> NOT OK:
> - amd64-ubuntu-2004-msan fails on main.selectivity_innodb
> - amd64-windows fails to compile

Apparently the Windows build has been rerun. The failed run is
https://buildbot.mariadb.org/#/builders/234/builds/17303 and it failed
due to the following:

   195>CUSTOMBUILD : CMake error : Cannot restore timestamp
"D:/Buildbot/amd64-windows/build/storage/perfschema/unittest/CMakeFiles/generate.stamp":
Access is denied.
[D:\Buildbot\amd64-windows\build\storage\perfschema\unittest\pfs_instr-t.vcxproj]

This is a mystery error that occurs from time to time.

> https://github.com/MariaDB/server/tree/10.7 / bc656c4

The 10.7 branch has reached its end of life. The commit that you
mention is for the final mariadb-10.7.8 release, February 1.

Best regards,

Marko