Re: net/core/sock.c lacks some SO_TIMESTAMPING_NEW support

[Willem de Bruijn <willemdebruijn.kernel@gmail.com>]

Jörn-Thorben Hinz wrote:
> On Wed, 2023-12-20 at 09:53 -0500, Willem de Bruijn wrote:
> > Jörn-Thorben Hinz wrote:
> > > Hi Arnd,
> > > 
> > > thanks for indirectly pinging me here about the unfinished patches.
> > > I
> > > kinda forgot about them over other things happening.
> > > 
> > > Happy to look back into them, it looks like it would be helpful to
> > > apply them. Is it fine to just answer the remarks from earlier this
> > > year, after a few months, in the same mail thread? Or preferable to
> > > resubmit the series[1] first?
> > 
> > Please resubmit instead of reviving the old thread. Thanks for
> > reviving
> > that.
> Thanks for the hint, will do so! (Maybe after Christmas.)
> 
> > 
> > IIRC the only open item was to limit the new BPF user to the new API?
> > That only applies to patch 2/2.
> Another point was to not change the behavior of
> getsockopt(SO_TIMESTAMPING_OLD), that’s just a minor change.
> 
> About limiting BPF to the SO_TIMESTAMPING_NEW, I am unsure if this is
> feasible, necessary, or even makes a difference (for a BPF program). In
> many places, BPF just passes-through calls like to get-/setsockopt(),
> only testing whether this call is explicitly allowed from BPF space.
> 
> Also, due to its nature, BPF code often has to re-provide defines, see
> for example tools/testing/selftests/bpf/progs/bpf_tracing_net.h This is
> also the case for SO_TIMESTAMPING_*. A limitation of BPF to
> SO_TIMESTAMPING_NEW could only be done in the allowed get-/setsockopt()
> calls, not through any BPF-provided defines.
> 
> I will take another look at this aspect and add my comments/findings to
> a resubmission.
> 
> > 
> > The missing sk_getsockopt SO_TIMESTAMPING_NEW might be breaking
> > users,
> > so is best sent stand-alone to net, rather than net-next.
> Hmm, I initially sent both patches together and to bpf-next since the
> second, BPF-related patch depends (for the included selftest) on the
> first one already being applied.
> 
> I’m unsure how to split them because of the dependency. Would one add a
> comment that commit X needs to be pulled in from net for commit Y to be
> applied in bpf-next? (That sounds bound to break something.)
>
> Also, getsockopt(SO_TIMESTAMPING_NEW) has been missing since 2019,
> since SO_TIMESTAMPING_NEW was added. Do you think it is still "urgent"
> enough to provide it through net instead of net-next/bpf-next?

net gets pulled into net-next at least once a week. If you submit this
patch now, it will likely be in bpf-next by the time we get to the
second more involved patch.

This report was a reminder that the current omission can actually
break users, so having it as a fix that goes to stable is warranted.
The Fixes tag will be

Fixes: 9718475e6908 ("socket: Add SO_TIMESTAMPING_NEW")


> > 
> > > Thorben
> > > 
> > > [1]
> > > https://lore.kernel.org/lkml/20230703175048.151683-1-jthinz@mailbox.tu-berlin.de/
> > > 
> > > On Wed, 2023-12-20 at 09:43 +0000, Arnd Bergmann wrote:
> > > > On Wed, Dec 20, 2023, at 04:00, Willem de Bruijn wrote:
> > > > > Thomas Lange wrote:
> > > > > > diff --git a/net/core/sock.c b/net/core/sock.c
> > > > > > index 16584e2dd648..a56ec1d492c9 100644
> > > > > > --- a/net/core/sock.c
> > > > > > +++ b/net/core/sock.c
> > > > > > @@ -2821,6 +2821,7 @@ int __sock_cmsg_send(struct sock *sk,
> > > > > > struct cmsghdr *cmsg,
> > > > > >                  sockc->mark = *(u32 *)CMSG_DATA(cmsg);
> > > > > >                  break;
> > > > > >          case SO_TIMESTAMPING_OLD:
> > > > > > +       case SO_TIMESTAMPING_NEW:
> > > > > >                  if (cmsg->cmsg_len != CMSG_LEN(sizeof(u32)))
> > > > > >                          return -EINVAL;
> > > > > > 
> > > > > > However, looking through the module, it seems that
> > > > > > sk_getsockopt() has no
> > > > > > support for SO_TIMESTAMPING_NEW either, but sk_setsockopt()
> > > > > > has.
> > > > > 
> > > > > Good point. Adding the author to see if this was a simple
> > > > > oversight
> > > > > or
> > > > > there was a rationale at the time for leaving it out.
> > > > 
> > > > I'm fairly sure this was just a mistake on our side. For the cmsg
> > > > case,
> > > > I think we just missed it because there is no corresponding
> > > > SO_TIMESTAMP{,NS}
> > > > version of this, so it fell through the cracks.
> > > > 
> > > > In the patch above, I'm not entirely sure about what needs to
> > > > happen
> > > > with the old/new format, i.e. the
> > > > 
> > > >    sock_valbool_flag(sk, SOCK_TSTAMP_NEW, optname ==
> > > > SO_TIMESTAMPING_NEW)
> > > > 
> > > > from setsockopt(). Is __sock_cmsg_send() allowed to turn on
> > > > timestamping
> > > > without it being first enabled using setsockopt()? If so, I think
> > > > we need to set the flag here the same way that setsockopt does.
> > > > If
> > > > not, then I think we instead should check that the old/new format
> > > > in the option sent via cmsg is the same that was set earlier with
> > > > setsockopt.
> > 
> > __sock_cmsg_send can only modify a subset of the bits in the
> > timestamping feature bitmap, so a call to setsockopt is still needed
> > 
> > But there is no ordering requirement, so the __sock_cmsg_send call
> > can
> > come before the setsockopt call. It would be odd, but the API allows
> > it.
> > > > 
> > > > For the missing getsockopt, there was even a patch earlier this
> > > > year
> > > > by Jörn-Thorben Hinz [1], but I failed to realize that we need
> > > > patch
> > > > 1/2 from his series regardless of patch 2/2.
> > > > 
> > > >      Arnd
> > > > 
> > > > [1]
> > > > https://lore.kernel.org/lkml/20230703175048.151683-2-jthinz@mailbox.tu-berlin.de/
> > > 
> > 
> > 
>