[pve-devel] [PATCH pve-http-server v2 1/1] fix unexpected EOF for client when closing TLS session

[Fabian Grünbichler]

> Fiona Ebner <f.ebner at proxmox.com> hat am 08.04.2025 09:27 CEST geschrieben:
> 
>  
> Am 05.03.25 um 22:45 schrieb Rob Rozestraten via pve-devel:
> > When pve-http-server initiates the closure of a TLS session, it does not
> > send a TLS close notify, resulting in an unexpected EOF error on systems
> > with recent crypto policies. This can break functionality with other
> > applications, such as Foreman[0].
> > 
> > This behavior can be observed in the following cases:
> > 
> >  * client uses HTTP/1.0 (no keepalive; server closes connection)
> >  * client sends no data for 5 sec (timeout; server closes connection)
> >  * server responds with 400 (no keepalive; server closes connection)
> > 
> > This patch sends the TLS close notify prior to socket teardown,
> > resulting in clean closure of TLS connections and no client error.
> > 
> > It also moves shutdown() to after the clearing of handlers. The reason
> > for this is stoptls() must come before shutdown(), but it also triggers
> > on_drain(), which calls client_do_disconnect() again. The extra call to
> > client_do_disconnect() is avoided inside accept_connections() by commit
> > f737984, but perhaps clearing the handlers prior to shutdown() will
> > avoid it in all cases.
> > 
> > [0]: https://github.com/theforeman/foreman_fog_proxmox/issues/325
> > 
> 
> I feel like the questions regarding blocking/missing client ack from
> Fabian from v1 are not answered yet:
> 
> > If I read the docs right, this could block (would that be an issue here?) and could potentially destroy the handle (so that might need to be rechecked afterwards to prevent spurious warnings?)
> > 
> > what happens if we initiate the teardown, and the client never acks it?

there was some more input in a separate mail:

https://lore.proxmox.com/pve-devel/mailman.799.1741211155.293.pve-devel@lists.proxmox.com/