On Tue, Oct 21, 2025 at 02:14:02PM +0200, Laurent Vivier wrote:
> On 20/10/2025 03:34, David Gibson wrote:
> > > @@ -570,7 +572,8 @@ static void tcp_timer_ctl(const struct ctx *c, struct tcp_tap_conn *conn)
> > >   		}
> > >   		conn->timer = fd;
> > > -		if (epoll_ctl(c->epollfd, EPOLL_CTL_ADD, conn->timer, &ev)) {
> > > +		if (epoll_ctl(flow_epollfd(&conn->f), EPOLL_CTL_ADD,
> > > +			      conn->timer, &ev)) {
> > Possibly a question for an earlier patch, but is there a reason we
> > can't use epoll_add() here?
> > 
> 
> Yes, the fd we use in epoll_ctl() and the fd stored in ref are not the same.
> 
> The fd in ref is conn->sock (epoll_add() takes fd from ref) but in
> epoll_ctl() we add conn->timer (the fd from timerfd_create()).

Huh.  That seems like a pre-existing bug.  I'm pretty sure the epoll
ref for the timer ought to have the timer's fd in it.  I suspect we
just never previously used the ref.fd field for this case.

-- 
David Gibson (he or they)	| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you, not the other way
				| around.
http://www.ozlabs.org/~dgibson