Re: [PATCH net-next 7/9] net-timestamp: open gate for bpf_setsockopt

[Jason Xing <kerneljasonxing@gmail.com>]

On Wed, Oct 9, 2024 at 9:23 PM Willem de Bruijn
<willemdebruijn.kernel@gmail.com> wrote:
>
> Jason Xing wrote:
> > On Wed, Oct 9, 2024 at 3:19 PM Martin KaFai Lau <martin.lau@linux.dev> wrote:
> > >
> > > On 10/8/24 2:51 AM, Jason Xing wrote:
> > > > From: Jason Xing <kernelxing@tencent.com>
> > > >
> > > > Now we allow users to set tsflags through bpf_setsockopt. What I
> > > > want to do is passing SOF_TIMESTAMPING_RX_SOFTWARE flag, so that
> > > > we can generate rx timestamps the moment the skb traverses through
> > > > driver.
> > > >
> > > > Here is an example:
> > > >
> > > > case BPF_SOCK_OPS_ACTIVE_ESTABLISHED_CB:
> > > > case BPF_SOCK_OPS_PASSIVE_ESTABLISHED_CB:
> > > >       sock_opt = SOF_TIMESTAMPING_RX_SOFTWARE;
> > > >       bpf_setsockopt(skops, SOL_SOCKET, SO_TIMESTAMPING,
> > > >                      &sock_opt, sizeof(sock_opt));
> > > >       break;
> > > >
> > > > In this way, we can use bpf program that help us generate and report
> > > > rx timestamp.
> > > >
> > > > Signed-off-by: Jason Xing <kernelxing@tencent.com>
> > > > ---
> > > >   net/core/filter.c | 3 +++
> > > >   1 file changed, 3 insertions(+)
> > > >
> > > > diff --git a/net/core/filter.c b/net/core/filter.c
> > > > index bd0d08bf76bb..9ce99d320571 100644
> > > > --- a/net/core/filter.c
> > > > +++ b/net/core/filter.c
> > > > @@ -5225,6 +5225,9 @@ static int sol_socket_sockopt(struct sock *sk, int optname,
> > > >               break;
> > > >       case SO_BINDTODEVICE:
> > > >               break;
> > > > +     case SO_TIMESTAMPING_NEW:
> > > > +     case SO_TIMESTAMPING_OLD:
> > >
> > > I believe this change was proposed before. It will change the user expectation
> > > on the sk_error_queue. It needs some bits/fields/knobs for bpf. I think this
> > > point is similar to other's earlier comments in this thread.
> >
> > Thanks for your reply.
> >
> > After seeing what you mentioned, I searched through the mailing list
> > and found one [1] which was designed to fetch hardware timestamps.
> >
> > [1]：https://lore.kernel.org/bpf/51fd5249-140a-4f1b-b20e-703f159e88a3@linux.dev/T/
> >
> > >
> > > I only have a chance to briefly look at it. I think it is useful. This
> > > bpf/timestamp feature request has come up before.
> >
> > At the very beginning, I had no intention to use bpf_setsockopt() to
> > retrieve the rx timestamp because it will override sk_tsflags, but I
> > cannot implement a good way like what I did to tx path: only setting
> > skb's field. I'm not sure if this override behaviour is acceptable, so
> > I post it to know what the bpf experts' suggestions are.
> >
> > >
> > > A high level comment. The current timestamp should work for non tcp sock? The
> > > bpf/timestamp solution should be able to also.
> >
> > For now, it only supports TCP proto. I would like to quickly implement
> > a framework which is also suitable for other protos. TCP is just a
> > start point.
> >
> > >
> > > sockops is tcp centric. From looking at patch 9 that needs to initialize 4 args,
> > > this interface feels old and not sure we want to extend to other sock types.
> > > This needs some thoughts.
> >
> > For me, I have interests to extend to other sock types. But I'm
> > supposed to ask Willem's opinion first.
> >
> > +Willem de Bruijn Do you want this bpf extension feature to extend to
> > other protos?
>
> There would likely be users for other protocols too, just like
> SO_TIMESTAMPING. Though TCP is probably the most widely used case by
> far.

Agreed !