Re: [PATCH net-next v5 03/11] tcp: use drop reasons in cookie check for ipv4

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From: Jason Xing <kerneljasonxing@gmail.com>
Date: Fri, 16 Feb 2024 09:28:26 +0800
> On Fri, Feb 16, 2024 at 5:09 AM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > From: Jason Xing <kerneljasonxing@gmail.com>
> > Date: Thu, 15 Feb 2024 09:20:19 +0800
> > > From: Jason Xing <kernelxing@tencent.com>
> > >
> > > Now it's time to use the prepared definitions to refine this part.
> > > Four reasons used might enough for now, I think.
> > >
> > > Signed-off-by: Jason Xing <kernelxing@tencent.com>
> > > --
> > > v5:
> > > Link: https://lore.kernel.org/netdev/CANn89i+iELpsoea6+C-08m6+=JkneEEM=nAj-28eNtcOCkwQjw@mail.gmail.com/
> > > Link: https://lore.kernel.org/netdev/632c6fd4-e060-4b8e-a80e-5d545a6c6b6c@kernel.org/
> > > 1. Use SKB_DROP_REASON_IP_OUTNOROUTES instead of introducing a new one (Eric, David)
> > > 2. Reuse SKB_DROP_REASON_NOMEM to handle failure of request socket allocation (Eric)
> > > 3. Reuse NO_SOCKET instead of introducing COOKIE_NOCHILD
> > > ---
> > >  net/ipv4/syncookies.c | 18 +++++++++++++-----
> > >  1 file changed, 13 insertions(+), 5 deletions(-)
> > >
> > > diff --git a/net/ipv4/syncookies.c b/net/ipv4/syncookies.c
> > > index 38f331da6677..aeb61c880fbd 100644
> > > --- a/net/ipv4/syncookies.c
> > > +++ b/net/ipv4/syncookies.c
> > > @@ -421,8 +421,10 @@ struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb)
> > >               if (IS_ERR(req))
> > >                       goto out;

I noticed in this case (ret == sk) we can set drop reason in
tcp_v4_do_rcv() as INVALID_COOKIE or something else.


> > >       }
> > > -     if (!req)
> > > +     if (!req) {
> > > +             SKB_DR_SET(reason, NOMEM);
> >
> > NOMEM is not appropriate when mptcp_subflow_init_cookie_req() fails.
> 
> Thanks for your careful check. It's true. I didn't check the MPTCP
> path about how to handle it.
> 
> It also means that what I did to the cookie_v6_check() is also wrong.

Yes, same for the v6 change.


> 
> [...]
> > >       /* Try to redo what tcp_v4_send_synack did. */
> > >       req->rsk_window_clamp = tp->window_clamp ? :dst_metric(&rt->dst, RTAX_WINDOW);
> > > @@ -476,10 +482,12 @@ struct sock *cookie_v4_check(struct sock *sk, struct sk_buff *skb)
> > >       /* ip_queue_xmit() depends on our flow being setup
> > >        * Normal sockets get it right from inet_csk_route_child_sock()
> > >        */
> > > -     if (ret)
> > > +     if (ret) {
> > >               inet_sk(ret)->cork.fl.u.ip4 = fl4;
> > > -     else
> > > +     } else {
> > > +             SKB_DR_SET(reason, NO_SOCKET);
> >
> > This also seems wrong to me.
> >
> > e.g. syn_recv_sock() could fail with sk_acceptq_is_full(sk),
> > then the listener is actually found.
> 
> Initially I thought using a not-that-clear name could be helpfull,
> though. NO_SOCKET here means no child socket can be used if I add a
> new description to SKB_DROP_REASON_NO_SOCKET.

Currently, NO_SOCKET is used only when sk lookup fails.  Mixing
different reasons sounds like pushing it back to NOT_SPECIFIED.
We could distinguish them by the caller IP though.


> 
> If the idea is proper, how about using NO_SOCKET for the first point
> you said to explain that there is no request socket that can be used?
> 
> If not, for both of the points you mentioned, it seems I have to add
> back those two new reasons (perhaps with a better name updated)?
> 1. Using SKB_DROP_REASON_REQSK_ALLOC for the first point (request
> socket allocation in cookie_v4/6_check())
> 2. Using SKB_DROP_REASON_GET_SOCK for the second point (child socket
> fetching in cookie_v4/6_check())
> 
> Now I'm struggling with the name and whether I should introduce some
> new reasons like what I did in the old version of the series :S

Why naming is hard would be because there are multiple reasons of
failure.  One way to be more specific is moving kfree_skb_reason()
into callee as you did in patch 2.


> If someone comes up with a good name or a good way to explain them,
> please tell me, thanks!

For 1. no idea :p

For 2. Maybe VALID_COOKIE ?  we drop the valid cookie in the same
function, but due to LSM or L3 layer, so the reason could be said
as L4-specific ?


> 
> also cc Eric, David
> 
> Thanks,
> Jason
>