On 2020-01-24, Sargun Dhillon [off-list ref] wrote:

On Fri, Jan 24, 2020 at 10:03 AM Tycho Andersen [off-list ref] wrote:

quoted

On Fri, Jan 24, 2020 at 01:17:42AM -0800, Sargun Dhillon wrote:

quoted

Currently, this just opens the group leader of the thread that triggere
the event, as pidfds (currently) are limited to group leaders.

I don't love the semantics of this; when they're not limited to thread
group leaders any more, we won't be able to change this. Is that work
far off?

Tycho

We would be able to change this in the future if we introduced a flag like
SECCOMP_USER_NOTIF_FLAG_PIDFD_THREAD which would send a
pidfd that's for the thread, and not just the group leader. The flag could
either be XOR with SECCOMP_USER_NOTIF_FLAG_PIDFD, or
could require both. Alternatively, we can rename
SECCOMP_USER_NOTIF_FLAG_PIDFD to
SECCOMP_USER_NOTIF_FLAG_GROUP_LEADER_PIDFD.

Possibly unpopular proposal -- would it make sense to just store the
pidfd_open(2) flags rather than coming up with our own set for
SECCOMP_USER_NOTIF? If/when pidfds are expanded to include non-leaders
there will be a corresponding flag for pidfd_open(2). Something like:

	struct seccomp_notif {
		__u64 id;
		__u32 pid;
		__u32 flags;
		struct seccomp_data data;
		__u64 pidfd_flags; // or __u32 -- not sure what Christian plans
		__u32 pidfd;
		__u32 __padding;
	};

This does mean there'll be an additional flags field, but I think it's a
slightly more consistent way to indicate "SECCOMP_USER_NOTIF_FLAG_PIDFD
implies a pidfd_open(2) on the traced task".

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>