On Tue, Apr 25, 2023 at 10:59 AM Kui-Feng Lee <sinquersw@xxxxxxxxx> wrote:
>
>
>
> On 4/18/23 09:47, Stanislav Fomichev wrote:
> > On 04/17, Martin KaFai Lau wrote:
> >> On 4/14/23 6:55 PM, Stanislav Fomichev wrote:
> >>> On 04/13, Stanislav Fomichev wrote:
> >>>> On Thu, Apr 13, 2023 at 7:38 AM Aleksandr Mikhalitsyn
> >>>> <aleksandr.mikhalitsyn@xxxxxxxxxxxxx> wrote:
> >>>>>
> >>>>> On Thu, Apr 13, 2023 at 4:22 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> >>>>>>
> >>>>>> On Thu, Apr 13, 2023 at 3:35 PM Alexander Mikhalitsyn
> >>>>>> <aleksandr.mikhalitsyn@xxxxxxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>> During work on SO_PEERPIDFD, it was discovered (thanks to Christian),
> >>>>>>> that bpf cgroup hook can cause FD leaks when used with sockopts which
> >>>>>>> install FDs into the process fdtable.
> >>>>>>>
> >>>>>>> After some offlist discussion it was proposed to add a blacklist of
> >>>>>>
> >>>>>> We try to replace this word by either denylist or blocklist, even in changelogs.
> >>>>>
> >>>>> Hi Eric,
> >>>>>
> >>>>> Oh, I'm sorry about that. :( Sure.
> >>>>>
> >>>>>>
> >>>>>>> socket options those can cause troubles when BPF cgroup hook is enabled.
> >>>>>>>
> >>>>>>
> >>>>>> Can we find the appropriate Fixes: tag to help stable teams ?
> >>>>>
> >>>>> Sure, I will add next time.
> >>>>>
> >>>>> Fixes: 0d01da6afc54 ("bpf: implement getsockopt and setsockopt hooks")
> >>>>>
> >>>>> I think it's better to add Stanislav Fomichev to CC.
> >>>>
> >>>> Can we use 'struct proto' bpf_bypass_getsockopt instead? We already
> >>>> use it for tcp zerocopy, I'm assuming it should work in this case as
> >>>> well?
> >>>
> >>> Jakub reminded me of the other things I wanted to ask here bug forgot:
> >>>
> >>> - setsockopt is probably not needed, right? setsockopt hook triggers
> >>> before the kernel and shouldn't leak anything
> >>> - for getsockopt, instead of bypassing bpf completely, should we instead
> >>> ignore the error from the bpf program? that would still preserve
> >>> the observability aspect
> >>
> >> stealing this thread to discuss the optlen issue which may make sense to
> >> bypass also.
> >>
> >> There has been issue with optlen. Other than this older post related to
> >> optlen > PAGE_SIZE:
> >> https://lore.kernel.org/bpf/5c8b7d59-1f28-2284-f7b9-49d946f2e982@xxxxxxxxx/,
> >> the recent one related to optlen that we have seen is
> >> NETLINK_LIST_MEMBERSHIPS. The userspace passed in optlen == 0 and the kernel
> >> put the expected optlen (> 0) and 'return 0;' to userspace. The userspace
> >> intention is to learn the expected optlen. This makes 'ctx.optlen >
> >> max_optlen' and __cgroup_bpf_run_filter_getsockopt() ends up returning
> >> -EFAULT to the userspace even the bpf prog has not changed anything.
> >
> > (ignoring -EFAULT issue) this seems like it needs to be
> >
> > if (optval && (ctx.optlen > max_optlen || ctx.optlen < 0)) {
> > /* error */
> > }
> >
> > ?
> >
> >> Does it make sense to also bypass the bpf prog when 'ctx.optlen >
> >> max_optlen' for now (and this can use a separate patch which as usual
> >> requires a bpf selftests)?
> >
> > Yeah, makes sense. Replacing this -EFAULT with WARN_ON_ONCE or something
> > seems like the way to go. It caused too much trouble already :-(
> >
> > Should I prepare a patch or do you want to take a stab at it?
> >
> >> In the future, does it make sense to have a specific cgroup-bpf-prog (a
> >> specific attach type?) that only uses bpf_dynptr kfunc to access the optval
> >> such that it can enforce read-only for some optname and potentially also
> >> track if bpf-prog has written a new optval? The bpf-prog can only return 1
> >> (OK) and only allows using bpf_set_retval() instead. Likely there is still
> >> holes but could be a seed of thought to continue polishing the idea.
> >
> > Ack, let's think about it.
> >
> > Maybe we should re-evaluate 'getsockopt-happens-after-the-kernel' idea
> > as well? If we can have a sleepable hook that can copy_from_user/copy_to_user,
> > and we have a mostly working bpf_getsockopt (after your refactoring),
> > I don't see why we need to continue the current scheme of triggering
> > after the kernel?
>
> Since a sleepable hook would cause some restrictions, perhaps, we could
> introduce something like the promise pattern. In our case here, BPF
> program call an async version of copy_from_user()/copy_to_user() to
> return a promise.
Having a promise might work. This is essentially what we already do
with sockets/etc with acquire/release pattern.
What are the sleepable restrictions you're hinting about? I feel like
with the sleepable bpf, we can also remove all the temporary buffer
management / extra copies which sounds like a win to me. (we have this
ugly heuristics with BPF_SOCKOPT_KERN_BUF_SIZE) The program can
allocate temporary buffers if needed..
> >>> - or maybe we can even have a per-proto bpf_getsockopt_cleanup call that
> >>> gets called whenever bpf returns an error to make sure protocols have
> >>> a chance to handle that condition (and free the fd)
> >>>
> >>
> >>
