Re: minimum compiler for Linux UAPI (was Re: [PATCH v3] ethtool: Replace 0-length array with flexible array)

[Kees Cook <keescook@chromium.org>]

On Fri, Jan 06, 2023 at 11:25:14PM +0900, Vincent MAILHOL wrote:
> On Fri. 6 Jan 2023 at 22:19, Jann Horn <jannh@google.com> wrote:
> > On Fri, Jan 6, 2023 at 5:28 AM Kees Cook <keescook@chromium.org> wrote:
> > > Zero-length arrays are deprecated[1]. Replace struct ethtool_rxnfc's
> > > "rule_locs" 0-length array with a flexible array. Detected with GCC 13,
> > > using -fstrict-flex-arrays=3:
> [...]
> > > diff --git a/include/uapi/linux/ethtool.h b/include/uapi/linux/ethtool.h
> > > index 58e587ba0450..3135fa0ba9a4 100644
> > > --- a/include/uapi/linux/ethtool.h
> > > +++ b/include/uapi/linux/ethtool.h
> > > @@ -1183,7 +1183,7 @@ struct ethtool_rxnfc {
> > >                 __u32                   rule_cnt;
> > >                 __u32                   rss_context;
> > >         };
> > > -       __u32                           rule_locs[0];
> > > +       __u32                           rule_locs[];
> >
> > Stupid question: Is this syntax allowed in UAPI headers despite not
> > being part of standard C90 or C++? Are we relying on all C/C++
> > compilers for pre-C99 having gcc/clang extensions?
> 
> The [0] isn't part of the C90 standard either. So having to choose
> between [0] and [], the latter is the most portable nowadays.
> 
> If I do a bit of speleology, I can see that C99 flexible array members
> were used as early as v2.6.19 (released in November 2006):
> 
>   https://elixir.bootlin.com/linux/v2.6.19/source/include/linux/usb/audio.h#L36
> 
> This is prior to the include/linux and include/uapi/linux split, but
> believe me, this usb/audio.h file is indeed part of the uapi.
> So, yes, using C99 flexible array members in the UAPI is de facto
> allowed because it was used for the last 16 years.
> 
> An interesting sub question would be:
> 
>   What are the minimum compiler requirements to build a program using
> the Linux UAPI?

You're right -- we haven't explicitly documented this. C99 seems like
the defacto minimum, though.

> And, after research, I could not find the answer. The requirements to
> build the kernel are well documented:
> 
>   https://docs.kernel.org/process/changes.html#changes
> 
> But no clue for the uapi. I guess that at one point in 2006, people
> decided that it was time to set the minimum requirement to C99. Maybe
> this matches the end of life of the latest pre-C99 GCC version? The
> detailed answer must be hidden somewhere on lkml.

I would make the argument that the requirements for building Linux UAPI
should match that of building the kernel...

-- 
Kees Cook