There's something very wrong going on when using btrfs for the test
images, namely:
- there's a significant performance hit, i.e. the Arch Linux run is
~20% slower, in the coverage run the situation is even worse
- intermittent boot failures
- intermittent "No space left on device" errors (even though there's
enough free space)
Since debugging this might take a while, let's temporarily revert back
to ext4 to make the CI stable again.
This reverts commit 7eb7e3ec4f.
Othewise test images are missing the tmpfiles snippets used to create the very
basic files at boot, which can be useful when a test wants to reuse the OS tree
(is already running in) for spawning a new container in pristine state.
Arch finally made dbus-broker the default dbus daemon [0], but unlike
Fedora they don't use Alias=dbus.service to make the dbus.symlink under
/etc, instead they create the symlink manually under /usr/lib, so let's
account for that.
[0] b24d15795a
If a binary built with ASan crashes for a reason unrelated to ASan
stuff, we're left with pretty much nothing, as there is neither an ASan
trace nor a coredump. Let's make this slightly more debug-able by
allowing such binaries to dump a core, but without the huge shadow map
(we should be actually fine by just setting disable_coredump=0, since
use_madv_dontdump defaults to true, but let's play it safe and not
potentially dump a 16+ TB core file).
We can't use the systemd-journal-upload user here, since it's created
dynamically by DynamicUser=yes. However, we can use the group specified
in SupplementaryGroups=, so do exactly that.
Turns out that redirecting a lot of output to the console can have some
funny effects, like random kernel soft lockups. I spotted this in
various CIs, but it remained almost entirely hidden thanks to
`softlockup_panic=1`, until 1a36d2672f which introduced a couple of
tests that log quite a lot in a short amount of time. This, in
combination with newer kernel version, which, for some reason, seem to
be more susceptible to such soft lockups, made the Arch Linux jobs soft
lockup quite a lot, see [0].
While debugging this I also noticed that runs which don't redirect
stdout/stderr to the console are noticeably faster, e.g.:
# TEST-71 nspawn + QEMU (KVM), StandardOutput=journal+console
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:24.64
# TEST-71 nspawn + QEMU (KVM), StandardOutput=journal
Elapsed (wall clock) time (h:mm:ss or m:ss): 0:17.95
# TEST-71 nspawn + QEMU, StandardOutput=journal+console
Elapsed (wall clock) time (h:mm:ss or m:ss): 2:04.70
# TEST-71 nspawn + QEMU, StandardOutput=journal
Elapsed (wall clock) time (h:mm:ss or m:ss): 1:44.48
# TEST-04 QEMU, StandardOutput=journal+console
Elapsed (wall clock) time (h:mm:ss or m:ss): 4:22.70
# TEST-04 QEMU, StandardOutput=console
Elapsed (wall clock) time (h:mm:ss or m:ss): 5:04.67
Given all this, let's effectively revert ba7abf79a5, and dump the
testsuite-related journal messages only after the test finishes, so they
don't go through the slow console.
Resolves: systemd/systemd-centos-ci#660
[0] https://github.com/systemd/systemd-centos-ci/issues/660
Let's save all journals from the test machine instead of calling export
on each journal file separately, which makes the code less complicated
(and probably faster).
I'm pretty sure this is not the only case, but it's the one I recently
noticed. Even though we call ddebug() from a function, that function is
called before ddebug() is defined, resulting in the same issue as if we
called just ddebug() in its place, i.e.:
..//test-functions: line 276: ddebug: command not found
Currently the test works only with policy shipped by Fedora, which makes
it pretty much useless in most of our CIs. Let's drop the custom module
and make the test more generic, so it works with the refpolicy as well,
which should allow us to run it on Arch and probably even in Ubuntu CI.
Let's just rely on the word splitting done by bash instead of messing
with that ourselves, as it's just adding extra complexity to appease one
ShellCheck check. Also, this apparently never worked for the nspawn
stuff anyway, since I forgot to set $IFS to an appropriate value, so it
always put all arguments from $KERNEL_APPEND into a single array item
with an extra newline, which then made systemd sad:
~# readarray arr <<< "foo bar baz"; for i in "${arr[@]}"; do echo "'$i'"; done
'foo bar baz
'
~# make -C test/TEST-45-TIMEDATE/ clean setup run BUILD_DIR=$PWD/build TEST_NO_QEMU=1 KERNEL_APPEND="systemd.log_level=console"
...
~# journalctl -o short-monotonic --no-hostname --file /var/tmp/systemd-tests/systemd-test.XaDX67/system.journal --grep "Failed to parse" -p info --no-pager
[551138.986882] systemd-tmpfiles[21]: Failed to parse log level 'console
[551138.987179] systemd-remount-fs[20]: Failed to parse log level 'console
[551138.993125] systemd-sysusers[23]: Failed to parse log level 'console
[551138.998685] journalctl[29]: Failed to parse log level 'console
Resolves: #29945
Let's put this back in, as it could help with occasional machine lock ups
on overloaded systems (and it didn't help with the original issue
anyway).
This reverts commit 3a89904e45.
On slower/overloaded systems it may take a bit for the swtpm socket
to show up:
I: Started swtpm as PID 189419 with state dir /tmp/tmp.pWqUutuGUj
I: Configured emulated TPM2 device tpm-spapr
+ tee /var/tmp/systemd-test-TEST-70-TPM2_1/console.log
+ timeout --foreground 1200 /bin/qemu-system-ppc64le -smp 4 ...
qemu-system-ppc64le: -chardev socket,id=chrtpm,path=/tmp/tmp.pWqUutuGUj/sock: Failed to connect to '/tmp/tmp.pWqUutuGUj/sock': No such file or directory
E: qemu failed with exit code 1
Spotted regularly in the ppc64le cron job and in some Ubuntu CI/CentOS CI
pr runs [0].
[0] https://github.com/systemd/systemd/pull/29183#issuecomment-1721727927
We can't do anything about them anyway, and most importantly this seems
to alleviate systemd/systemd-centos-ci#660, which should make the CIs
a bit less angry (at least until the issue is addressed properly).
The pipe stuff introduced in 701e0c2660 causes nspawn to switch the
console from 'interactive' into 'read-only' which is a bit useless when
debugging. Let's set --console=interactive explicitly in such case.
Follow-up to 701e0c2660.
This metadata (EXTENSION_RELOAD_MANAGER) can be set to "1" to reload the manager
when merging/refreshing/unmerging a system extension image. This can be useful in case the sysext
image provides systemd units that need to be loaded.
With `--no-reload`, one can deactivate the EXTENSION_RELOAD_MANAGER metadata interpretation.
Signed-off-by: Mathieu Tortuyaux <mtortuyaux@microsoft.com>
