Re: [PATCH] SELinux: Dontaudit access to dri devices

[Paul Holzinger <pholzing@redhat.com>]

Hi,

On 31/03/2026 21:47, Stefano Brivio wrote:
> On Tue, 31 Mar 2026 09:00:47 +0200
> Johannes Segitz <jsegitz@suse.de> wrote:
>
>> Hi,
>>
>> On Mon, Mar 30, 2026 at 05:15:42PM +0200, Stefano Brivio wrote:
>>> On Mon, 30 Mar 2026 13:05:57 +0200 Johannes Segitz <jsegitz@suse.de> wrote:
>>>> Currently podman can pass a FD to a DRI device to pasta, leading to AVCs
>>>> like this:
>>>> avc:  denied  { read write }
>>>> comm="pasta" path="/dev/dri/renderD128"
>>>> scontext=unconfined_u:unconfined_r:pasta_t:s0-s0:c0.c1023
>>>> tcontext=system_u:object_r:dri_device_t:s0
>>>> tclass=chr_file
>>>> These are harmless, so dontaudit them
>>>>
>>>> Signed-off-by: Johannes Segitz <jsegitz@suse.de>
>>> Thanks for the patch.
>>>
>>> I'm wondering how can this still happen though, as commit 09603cab28f9
>>> ("passt, util: Close any open file that the parent might have leaked")
>>> should take care of those. Do you happen to know?
>> No, I just read the code and it seems like this should prevent this. I
>> unfortunately can't debug this in depth, because it doesn't happen on my
>> system. The reporter is helpful with debugging, but going into gdb sessions
>> with remote hands doesn't sound feasible ;)
> Yeah, I see. :) ...I finally had a quick look at it.
>
> I think it's actually good that SELinux caught this, because *maybe* we
> just call close_range() too late and things would be fine otherwise,
> but if not, we would leave pasta running with access to an unrelated
> device, which isn't great, even though we don't run as root so it's
> unlikely we could really do something with it.
>
> By the way I wonder if it's similar to this report:
>
>    https://bugzilla.redhat.com/show_bug.cgi?id=2374197
>
> which I never really tried to figure out.

I described here I think: 
https://bugzilla.redhat.com/show_bug.cgi?id=2374291#c10

There is never time to close fds earlier, it validates sometime during 
execve(). My guess because that is the point where it transitions into 
the pasta_t context so it checks all files against the new policy?

So any leak warnings are impossible to avoid on the pasta side and must 
be fixed where we leak them.

Johannes it would be good if you can find out where that files gets 
opened? Is it Podman? The go runtime uses O_CLOEXEC but we do have some 
manual open calls where we might forget it and as such leaks, it would 
be good to find and fix them. It would be good if you have a reproducer.

>>> Perhaps the access happens before we call isolate_initial()... but then
>>> I guess we should try to close leaked files before that point, to be on
>>> the safe side?
>> Would be worth a try. If you have a patch for that I can provide an updated
>> package to the reporter and ask him to test it
> Assuming that the kernel version is >= 5.9 (otherwise we don't have
> close_range() at all), you could try something like this:
>
> ---
> diff --git a/passt.c b/passt.c
> index f84419c..d5dad4c 100644
> --- a/passt.c
> +++ b/passt.c
> @@ -340,6 +340,8 @@ int main(int argc, char **argv)
>   	struct timespec now;
>   	struct sigaction sa;
>   
> +	close_range(STDERR_FILENO + 1, ~0U, CLOSE_RANGE_UNSHARE);
> +
>   	if (clock_gettime(CLOCK_MONOTONIC, &log_start))
>   		die_perror("Failed to get CLOCK_MONOTONIC time");
>   
> ---
>
> and, if it doesn't work, try to close standard error as well (no idea
> why a DRI device would be passed as standard error but I'm not sure why
> we would have it as any other open file descriptor either):
>
> 	close_range(STDERR_FILENO, ~0U, CLOSE_RANGE_UNSHARE);
>
> ...all the way to:
>
> 	close_range(0, ~0U, CLOSE_RANGE_UNSHARE);
>
> ...if that doesn't work, the next thing I would try is to add a delay
> in the Podman thread starting pasta (or even there at the beginning
> of main() in pasta itself), use that delay to attach with strace (as
> root, pasta won't let you do that otherwise), correlate timestamps
> with SELinux logs, and from there everything should be clear.
>
-- 
Paul Holzinger