Re: [PATCH v3 net-next 1/7] net: skb_drop_reason: add document for drop reasons

[Menglong Dong]

On Thu, Feb 10, 2022 at 1:12 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
>
> On Thu, 10 Feb 2022 11:19:49 +0800 Menglong Dong wrote:
> > I'm doing the job of using kfree_skb_reason() for the TCP layer,
> > and I have some puzzles.
> >
> > When collecting drop reason for tcp_v4_inbound_md5_hash() in
> > tcp_v4_rcv(), I come up with 2 ways:
> >
> > First way: pass the address of reason to tcp_v4_inbound_md5_hash()
> > like this:
> >
> >  static bool tcp_v4_inbound_md5_hash(const struct sock *sk,
> >                       const struct sk_buff *skb,
> > -                    int dif, int sdif)
> > +                    int dif, int sdif,
> > +                    enum skb_drop_reason *reason)
> >
> > This can work, but many functions like tcp_v4_inbound_md5_hash()
> > need to do such a change.
> >
> > Second way: introduce a 'drop_reason' field to 'struct sk_buff'. Therefore,
> > drop reason can be set by 'skb->drop_reason = SKB_DROP_REASON_XXX'
> > anywhere.
> >
> > For TCP, there are many cases where you can't get a drop reason in
> > the place where skb is freed, so I think there needs to be a way to
> > deeply collect drop reasons. The second can resolve this problem
> > easily, but extra fields may have performance problems.
> >
> > Do you have some better ideas?
>
> On a quick look tcp_v4_inbound_md5_hash() returns a drop / no drop
> decision, so you could just change the return type to enum
> skb_drop_reason. SKB_DROP_REASON_NOT_SPECIFIED is 0 is false,
> so if (reason) goto drop; logic will hold up.

Yeah, that's an idea. But some functions are more complex, such as
tcp_rcv_state_process() and tcp_rcv_state_process()->tcp_v4_conn_request().
The return value of tcp_rcv_state_process() can't be reused, and it's hard
to add a function param of type 'enum skb_drop_reason *' to
tcp_v4_conn_request().

There are some nice drop reasons in tcp_v4_conn_request(), it's a pity to
give up them.

How about introducing a field to 'struct sock' for drop reasons? As sk is
locked during the packet process in tcp_v4_do_rcv(), this seems to work.

Thanks!
Menglong Dong