On Wed, Apr 29, 2020 at 04:33:12PM -0700, sdf@xxxxxxxxxx wrote:
> On 04/30, Daniel Borkmann wrote:
> > On 4/29/20 7:05 PM, Stanislav Fomichev wrote:
> > > Currently, bpf_getsocktop and bpf_setsockopt helpers operate on the
> > > 'struct bpf_sock_ops' context in BPF_PROG_TYPE_SOCK_OPS program.
> > > Let's generalize them and make the first argument be 'struct bpf_sock'.
> > > That way, in the future, we can allow those helpers in more places.
> > >
> > > BPF_PROG_TYPE_SOCK_OPS still has the existing helpers that operate
> > > on 'struct bpf_sock_ops', but we add new bpf_{g,s}etsockopt that work
> > > on 'struct bpf_sock'. [Alternatively, for BPF_PROG_TYPE_SOCK_OPS,
> > > we can enable them both and teach verifier to pick the right one
> > > based on the context (bpf_sock_ops vs bpf_sock).]
> > >
> > > As an example, let's allow those 'struct bpf_sock' based helpers to
> > > be called from the BPF_CGROUP_INET{4,6}_CONNECT hooks. That way
> > > we can override CC before the connection is made.
> > >
> > > v2:
> > > * s/BPF_PROG_TYPE_CGROUP_SOCKOPT/BPF_PROG_TYPE_SOCK_OPS/
> > >
> > > Acked-by: John Fastabend <john.fastabend@xxxxxxxxx>
> > > Acked-by: Martin KaFai Lau <kafai@xxxxxx>
> > > Signed-off-by: Stanislav Fomichev <sdf@xxxxxxxxxx>
> > [...]
> > > +BPF_CALL_5(bpf_setsockopt, struct sock *, sk,
> > > + int, level, int, optname, char *, optval, int, optlen)
> > > +{
> > > + u32 flags = 0;
> > > + return _bpf_setsockopt(sk, level, optname, optval, optlen, flags);
> > > +}
> > > +
> > > +static const struct bpf_func_proto bpf_setsockopt_proto = {
> > > + .func = bpf_setsockopt,
> > > + .gpl_only = false,
> > > + .ret_type = RET_INTEGER,
> > > + .arg1_type = ARG_PTR_TO_SOCKET,
> > > + .arg2_type = ARG_ANYTHING,
> > > + .arg3_type = ARG_ANYTHING,
> > > + .arg4_type = ARG_PTR_TO_MEM,
> > > + .arg5_type = ARG_CONST_SIZE,
> > > +};
> > > +
> > > +BPF_CALL_5(bpf_getsockopt, struct sock *, sk,
> > > + int, level, int, optname, char *, optval, int, optlen)
> > > +{
> > > + return _bpf_getsockopt(sk, level, optname, optval, optlen);
> > > +}
> > > +
> > > static const struct bpf_func_proto bpf_getsockopt_proto = {
> > > .func = bpf_getsockopt,
> > > .gpl_only = false,
> > > .ret_type = RET_INTEGER,
> > > + .arg1_type = ARG_PTR_TO_SOCKET,
> > > + .arg2_type = ARG_ANYTHING,
> > > + .arg3_type = ARG_ANYTHING,
> > > + .arg4_type = ARG_PTR_TO_UNINIT_MEM,
> > > + .arg5_type = ARG_CONST_SIZE,
> > > +};
> > > +
> > [...]
> > > @@ -6043,6 +6098,22 @@ sock_addr_func_proto(enum bpf_func_id func_id,
> > const struct bpf_prog *prog)
> > > return &bpf_sk_storage_get_proto;
> > > case BPF_FUNC_sk_storage_delete:
> > > return &bpf_sk_storage_delete_proto;
> > > + case BPF_FUNC_setsockopt:
> > > + switch (prog->expected_attach_type) {
> > > + case BPF_CGROUP_INET4_CONNECT:
> > > + case BPF_CGROUP_INET6_CONNECT:
> > > + return &bpf_setsockopt_proto;
>
> > Hm, I'm not sure this is safe. In the sock_addr_func_proto() we also have
> > other helpers callable from connect hooks like sk_lookup_{tcp,udp} which
> > return a PTR_TO_SOCKET_OR_NULL, and now we can pass those sockets also
> > into
> > bpf_{get,set}sockopt() helper after lookup to change various sk related
> > stuff
> > but w/o being under lock. Doesn't the sock_owned_by_me() yell here at
> > minimum
> > (I'd expect so)?
> Ugh, good point, I missed the fact that sk_lookup_{tcp,udp} are there
> for sock_addr :-( I can try to do a simple test case to verify
> that sock_owned_by_me triggers, but I'm pretty certain it should
> (I've been calling bpf_{s,g}etsockopt for context socket so it's quiet).
>
> I don't think there is any helper similar to sock_owned_by_me() that
> I can call to verify that the socket is held by current thread
> (without the lockdep splat) and bail out?
>
> In this case, is something like adding new PTR_TO_LOCKED_SOCKET_OR_NULL
> is the way to go? Any other ideas?
Looks like networking will benefit from sleepable progs too.
We could have just did lock_sock() inside bpf_setsockopt
before setting cong control.
In the mean time how about introducing try_lock_sock()
that will bail out if it cannot grab the lock?
For most practical cases that would work and eventually we
can convert it to full lock_sock ?
