On Mon, Feb 14, 2022 at 10:25:48AM +0000, Daniel P. Berrangé wrote:
I understand the motivation, but please don't change this. Applications like OpenStack have configured ssh authorized_keys files with the specific command that libvirt invokes. So changes like this will break their SSH configs. We caused this pain when we first introduced the virt-ssh-helper, but at least that was giving them a functional improvement and they could use a URI parameter to force the old command string. This change is just prettiness for no functional improvement so is not worth breaking apps for.
Can you please provide pointers to the OpenStack implementation of this and the issue that resulted from introducing virt-ssh-helper?
I don't know where the code is. I just know that they were broken by our changes in this area.
AFAICT the only way to restrict what commands a user can run after successfully authenticating is to specify command=... before the corresponding key in authorized_keys and I don't see how this change, or indeed the one that happened when virt-ssh-helper was added, could interfere with that.
The command that was listed in the authorized_keys file no longer matched what libvirt was actually invoking, so it was rightly rejected.
I have found some of the issues filed at the time https://bugs.launchpad.net/tripleo/+bug/1918250 https://bugzilla.redhat.com/show_bug.cgi?id=1936804 and the corresponding fix https://github.com/rdo-packages/nova-distgit/commit/d5aba75f3b5589e156afeec9... The detection logic, as currently implemented, is a bit fragile, but updating it so that it keeps working even as we make minor adjustments to our ssh tunnel script wasn't particularly difficult https://github.com/andreabolognani/nova-distgit/commit/27cee8da127c1d447cfb6... I have posted a new version of this series which cleans up things further and actually addresses the original GitLab issue https://listman.redhat.com/archives/libvir-list/2022-February/msg00544.html but I will of course not push it until the nova-migration-wrapper changes mentioned above have been proposed and accepted, which I intend to try next. -- Andrea Bolognani / Red Hat / Virtualization