On 22/01/2025 09:51, Sabrina Dubroca wrote:
2025-01-22, 01:40:47 +0100, Antonio Quartulli wrote:
On 17/01/2025 12:48, Sabrina Dubroca wrote: [...]
With the delayed socket release (which is similar to what was in v11, but now with refcounting on the netdevice which should make rtnl_link_unregister in ovpn_cleanup wait [*]), we may return to userspace as if the peer was gone, but the socket hasn't been detached yet.
A userspace application that tries to remove the peer and immediately re-create it with the same socket could get EBUSY if the workqueue hasn't done its job yet. That would be quite confusing to the application.
So I would add a completion to wait here until the socket has been fully detached. Something like below.
[*] I don't think the current refcounting fully protects against that, I'll comment on 05/25
Sabrina, after the other changes I acknowledged, do you still have comments for 5/25?
The call_rcu vs _put was all I had for this.
Ok!
Note that you have to wait until ~Feb 4th before you can resubmit (since net-next is currently closed). I'll take another look at this revision next week, since I've only checked a few specific things (mainly related to peer and socket destruction) so far.
Alright! In the meantime I'll take these days to perform more tests (possibly extending the selftest suite a little bit).
If you want to compare this patchset with the very latest code (including your suggestions), you can check the main branch at:
https://github.com/openvpn/linux-kernel-ovpn
Thanks a lot!
Best Regards,