Re: [PATCH bpf-next v2] bpf: bpf_{g,s}etsockopt for struct bpf_sock

[sdf@google.com]

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@gmail.com>
> > Acked-by: Martin KaFai Lau <kafai@fb.com>
> > Signed-off-by: Stanislav Fomichev <sdf@google.com>
> [...]
> > +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?