Re: [PATCH net-next 2/3] bpf: add TCP_BPF_RTO_MAX for bpf_setsockopt

[Stanislav Fomichev <stfomichev@gmail.com>]

On 02/14, Jason Xing wrote:
> On Fri, Feb 14, 2025 at 7:41 AM Stanislav Fomichev <stfomichev@gmail.com> wrote:
> >
> > On 02/13, Jason Xing wrote:
> > > Support bpf_setsockopt() to set the maximum value of RTO for
> > > BPF program.
> > >
> > > Signed-off-by: Jason Xing <kerneljasonxing@gmail.com>
> > > ---
> > >  Documentation/networking/ip-sysctl.rst | 3 ++-
> > >  include/uapi/linux/bpf.h               | 2 ++
> > >  net/core/filter.c                      | 6 ++++++
> > >  tools/include/uapi/linux/bpf.h         | 2 ++
> > >  4 files changed, 12 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/Documentation/networking/ip-sysctl.rst b/Documentation/networking/ip-sysctl.rst
> > > index 054561f8dcae..78eb0959438a 100644
> > > --- a/Documentation/networking/ip-sysctl.rst
> > > +++ b/Documentation/networking/ip-sysctl.rst
> > > @@ -1241,7 +1241,8 @@ tcp_rto_min_us - INTEGER
> > >
> > >  tcp_rto_max_ms - INTEGER
> > >       Maximal TCP retransmission timeout (in ms).
> > > -     Note that TCP_RTO_MAX_MS socket option has higher precedence.
> > > +     Note that TCP_BPF_RTO_MAX and TCP_RTO_MAX_MS socket option have the
> > > +     higher precedence for configuring this setting.
> >
> > The cover letter needs more explanation about the motivation. And
> > the precedence as well.
> 
> I am targeting the net-next tree because of recent changes[1] made by
> Eric. It probably hasn't merged into the bpf-next tree.
> 
> [1]: https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?id=ae9b3c0e79bc
> 
> >
> > WRT precedence, can you install setsockopt cgroup program and filter out
> > calls to TCP_RTO_MAX_MS?
> 
> Yesterday, as suggested by Kuniyuki, I decided to re-use the same
> logic of TCP_RTO_MAX_MS for bpf_setsockopt():
> diff --git a/net/core/filter.c b/net/core/filter.c
> index 2ec162dd83c4..ffec7b4357f9 100644
> --- a/net/core/filter.c
> +++ b/net/core/filter.c
> @@ -5382,6 +5382,7 @@ static int sol_tcp_sockopt(struct sock *sk, int optname,
>         case TCP_USER_TIMEOUT:
>         case TCP_NOTSENT_LOWAT:
>         case TCP_SAVE_SYN:
> +       case TCP_RTO_MAX_MS:
>                 if (*optlen != sizeof(int))
>                         return -EINVAL;
>                 break;
> 
> Are you referring to using the previous way (by introducing a new flag
> for BPF) because we need to know the explicit precedence between
> setsockopt() and bpf_setsockopt() or other reasons? If so, I think
> there are more places than setsockopt() to modify.
> 
> And, sorry that I don't follow what you meant by saying "install
> setsockopt cgroup program" here. Please provide more hints.

Ah, sorry, I misread it as bpf options taking precedence over tcp ones;
ignore the suggestion about setsockopt cgroup prog.

And yes, reusing the logic of TCP_RTO_MAX_MS looks better!