On 11/12/26 21:58, Daniel P. Berrangé wrote:
On Thu, Nov 12, 2026 at 09:45:09PM +0300, Michael Tokarev wrote:
From: Matthew White <[email protected]>
Upstream commit c4ce04cfb7460d46c0262d437a7f35bb5c5fc449 changed the
configure file to set container_command regardless of use_containers.
As a result, it runs `docker.py probe` to detect a Docker-compatible
application, such as Podman, even when --disable-containers is used.
On Gentoo, this reopens https://bugs.gentoo.org/732972. As reported
in https://bugs.gentoo.org/984747, the `docker.py probe` command can
trigger SANDBOX ACCESS VIOLATION errors under QEMU v11.1.2:
* ACCESS DENIED: open_wr_creat: /var/lib/rpm/rpmdb.sqlite-wal
* ACCESS DENIED: fchown: /var/lib/rpm/rpmdb.sqlite-wal
...
I'm puzzled why docker.py probe would touch the RPM database.
Yeah, that was puzzling for me too, but that wasn't the reason
I took this patch (which turned out not to be trivial in the end).
As long as the configure script does not break though, having
the system log sandbox warnings is not a bug from QEMU's POV
IMHO.
My immediate reaction is - why probe at all once --disable-containers
is given.
There might be different reasons for the error (or lack thereof), but
once it's been told to disable containers, there's no reason to probe
container environment, end of story....
The probe can be disabled by using --container-command=podman
or probably even --container-command=false would work.
This patch is to prevent `docker.py probe` from running and to clear
container_command when --disable-containers is used.
This reintroduces the problem that c4ce04cfb7460d46c0262d437a7f35bb5c5fc449
was fixing, because tests/docker/Makefile will now re-run the docker
probe on every command.
The --disable-containers arg was only intended to disable the use of
cross compilation in containers for the purposes of TCG testsuite.
...And this is the actual problem. It isn't --disable-containers-for-tests.
It is --disable-container. It's more, the help currently says:
--disable-containers don't use containers for cross-building
so it isn't obvious at all this is only used for tests.
And as long as it is actually doing something else, not what the name suggests,
people will try to use it for something it isn't intended for.
As long as these two options (coming together) are interpreted in such
an awkward way, there will be more attempts to do the same, like this one.
(Why this change seemed trivial to me: because it restores the status-quo,
everything seems to match now, disable-containers means empty container
command, end of question).
I re-sent the pull request without this change. But this should be
fixed one way or another, I think :)
And I should be more careful next time.
Thanks,
/mjt