Re: [PATCH v3] net/tls: Fix return values to avoid ENOTSUPP

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

On Thu, Dec 5, 2019 at 3:44 PM Valentin Vidić
<vvidic@valentin-vidic.from.hr> wrote:
>
> On Thu, Dec 05, 2019 at 03:06:55PM -0500, Willem de Bruijn wrote:
> > On Thu, Dec 5, 2019 at 2:34 PM Jakub Kicinski
> > <jakub.kicinski@netronome.com> wrote:
> > >
> > > On Thu,  5 Dec 2019 07:41:18 +0100, Valentin Vidic wrote:
> > > > ENOTSUPP is not available in userspace, for example:
> > > >
> > > >   setsockopt failed, 524, Unknown error 524
> > > >
> > > > Signed-off-by: Valentin Vidic <vvidic@valentin-vidic.from.hr>
> > >
> > > > diff --git a/net/tls/tls_device.c b/net/tls/tls_device.c
> > > > index 0683788bbef0..cd91ad812291 100644
> > > > --- a/net/tls/tls_device.c
> > > > +++ b/net/tls/tls_device.c
> > > > @@ -429,7 +429,7 @@ static int tls_push_data(struct sock *sk,
> > > >
> > > >       if (flags &
> > > >           ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
> > > > -             return -ENOTSUPP;
> > > > +             return -EOPNOTSUPP;
> > > >
> > > >       if (unlikely(sk->sk_err))
> > > >               return -sk->sk_err;
> > > > @@ -571,7 +571,7 @@ int tls_device_sendpage(struct sock *sk, struct page *page,
> > > >       lock_sock(sk);
> > > >
> > > >       if (flags & MSG_OOB) {
> > > > -             rc = -ENOTSUPP;
> > > > +             rc = -EOPNOTSUPP;
> > >
> > > Perhaps the flag checks should return EINVAL? Willem any opinions?
> >
> > No strong opinion. Judging from do_tcp_sendpages MSG_OOB is a
> > supported flag in general for sendpage, so signaling that the TLS
> > variant cannot support that otherwise valid request sounds fine to me.
>
> I based these on the description from the sendmsg manpage, but you decide:
>
> EOPNOTSUPP
>     Some bit in the flags argument is inappropriate for the socket type.

Interesting. That is a narrower interpretation than asm-generic/errno.h

  #define EOPNOTSUPP      95      /* Operation not supported on
transport endpoint */

which is also the string that strerror() generates.

>
> > > > diff --git a/net/tls/tls_main.c b/net/tls/tls_main.c
> > > > index bdca31ffe6da..5830b8e02a36 100644
> > > > --- a/net/tls/tls_main.c
> > > > +++ b/net/tls/tls_main.c
> > > > @@ -496,7 +496,7 @@ static int do_tls_setsockopt_conf(struct sock *sk, char __user *optval,
> > > >       /* check version */
> > > >       if (crypto_info->version != TLS_1_2_VERSION &&
> > > >           crypto_info->version != TLS_1_3_VERSION) {
> > > > -             rc = -ENOTSUPP;
> > > > +             rc = -EINVAL;
> > >
> > > This one I think Willem asked to be EOPNOTSUPP OTOH.
> >
> > Indeed (assuming no one disagrees). Based on the same rationale: the
> > request may be valid, it just cannot be accommodated (yet).
>
> In this case other checks in the same function like crypto_info->cipher_type
> return EINVAL, so I used the same here.

That makes sense.

I think there is a fundamental difference between, say, passing an
argument of incorrect length (optlen < sizeof(..)) and asking for a
possibly unsupported cipher mode. But consistency trumps that.

I don't mean to drag this out by bike-shedding.

Happy to defer to maintainers on whether the errno on published code
can and should be changed, which is the more fundamental issue than
the exact errno.

FWIW, I also did not see existing openssl and gnutls callers test the
specific errno. The calls just fail on any setsockopt return value -1.