Re: [PATCH net-next v10 08/15] quic: add path management

On Tue, Mar 3, 2026 at 3:23=E2=80=AFAM Paolo Abeni <[email protected]> wrot=
e:
>
> On 2/25/26 3:34 AM, Xin Long wrote:
> > +/* Binds a QUIC path to a local port and sets up a UDP socket. */
> > +int quic_path_bind(struct sock *sk, struct quic_path_group *paths, u8 =
path)
> > +{
> > +     union quic_addr *a =3D quic_path_saddr(paths, path);
> > +     int rover, low, high, remaining;
> > +     struct net *net =3D sock_net(sk);
> > +     struct quic_uhash_head *head;
> > +     struct quic_udp_sock *us;
> > +     u16 port;
> > +
> > +     port =3D ntohs(a->v4.sin_port);
> > +     if (port) {
> > +             head =3D quic_udp_sock_head(net, port);
> > +             mutex_lock(&head->lock);
> > +             us =3D quic_udp_sock_lookup(sk, a, port);
> > +             if (us) {
>
> When the quick socket is already bound to a local port, reusing an
> existing udp tunnel sock is allowed, but when the quick socket is not
> bound, UDP tunnel sock reused is prevented. This looks confusing and not
> documented, please clarify the behavior and/or make it consistent.
>
Yes,

/* Reuse of an existing UDP tunnel socket is allowed,
* but if it is currently being freed asynchronously by the workqueue,
* it cannot be used now =E2=80=94 retry later.
*/

Let me know if it's still not clear.
>
> > +                     if (!quic_udp_sock_get(us)) { /* Releasing in wor=
kqueue; retry later. */
> > +                             mutex_unlock(&head->lock);
> > +                             return -EAGAIN;
>
> Why not -EADDRINUSE here?
Because in this case, the us is being released in workqueue, a retry
will likely succeed.

>
> > +                     }
> > +             } else {
> > +                     us =3D quic_udp_sock_create(sk, a);
> > +                     if (!us) {
> > +                             mutex_unlock(&head->lock);
> > +                             return -EINVAL;
>
> It's probably better to propagate an error code (PTR_ERR) from
> quic_udp_sock_create(), or use -ENOMEM
>
Changing to PTR_ERR looks better.

> [...]
> > @@ -332,6 +333,12 @@ static __init int quic_init(void)
> >       if (err)
> >               goto err_hash;
> >
> > +     quic_wq =3D create_workqueue("quic_workqueue");
> > +     if (!quic_wq) {
> > +             err =3D -ENOMEM;
> > +             goto err_wq;
> > +     }
>
> AI review noted that:
>
> This isn't a bug, but create_workqueue() is a legacy API marked with
> __WQ_LEGACY in include/linux/workqueue.h. Should new subsystem code use
> alloc_workqueue() with explicit flags instead?
>
> Looking at include/linux/workqueue.h, create_workqueue() implicitly sets
> WQ_PERCPU, creating per-CPU worker threads. Since quic_wq only handles
> infrequent UDP socket cleanup via quic_udp_sock_put_work() in path.c, is
> per-CPU allocation necessary here? Would alloc_workqueue("quic_workqueue"=
,
> WQ_MEM_RECLAIM, 0) be more appropriate, or could this simply use system_w=
q
> if memory reclaim safety is not required?
>
This workqueue is also used for processing some backlog packets, which requ=
ires
process context.

I will move the infrequent UDP socket cleanup to system_wq as the AI sugges=
ts,
and leave this workqueue for the backlog packets processing only.
Then the allocation becomes:

quic_wq =3D alloc_workqueue("quic_workqueue", WQ_PERCPU, 0);

Thanks.