Re: [PATCH v06 18/36] uapi linux/errqueue.h: include linux/time.h in user space

[Mikko Rapeli]

On Sun, Aug 06, 2017 at 05:42:13PM -0400, Willem de Bruijn wrote:
> On Sun, Aug 6, 2017 at 5:33 PM, Mikko Rapeli <mikko.rapeli@xxxxxx> wrote:
> > On Sun, Aug 06, 2017 at 05:24:20PM -0400, Willem de Bruijn wrote:
> >> >> > +#ifdef __KERNEL__
> >> >> > +#include <linux/time.h>
> >> >> > +#else
> >> >> > +#include <time.h>
> >> >> > +#endif /* __KERNEL__ */
> >> >>
> >> >> This will break applications that include <linux/time.h> manually.
> >> >> I previously sent a patch to use libc-compat to make compilation succeed
> >> >> when both are included in the case where <linux/time.h> is included after
> >> >> <time.h>.
> >> >>
> >> >>   https://lkml.org/lkml/2016/9/12/872
> >> >>
> >> >> The inverse will require changes to the libc header to avoid redefining
> >> >> symbols already defined by <linux/time.h>
> >> >>
> >> >> The second patch in that 2-patch set included <linux/time.h>
> >> >> unconditionally after the fix. This broke builds that also included
> >> >> <time.h> in the wrong order. I did not resubmit the first patch as a
> >> >> stand-alone, as it is not sufficient to avoid breakage.
> >> >
> >> > I wasn't aware of your change, but I was about to send this to fix the
> >> > case when glibc <time.h> is included before <linux/time.h>:
> >> >
> >> > https://github.com/mcfrisk/linux/commit/f3952a27b8a21c6478d26e6246055383483f6a66
> >>
> >> There are a few differences between the two. Including <time.h> does not
> >> unconditionally define all the symbols. Some are conditional on additional
> >> state, such as __timespec_defined.
> >
> > Yep, your patch seems better for libc-compat.h. Could you send it again?
> 
> Okay. Or feel free to include it in the patchset if that helps resolve
> dependencies.

If you don't have the time, I will send tomorrow a new version of this
patch which fixes the commit topic and before that your libc-compat.h change
so both could be applied together.

Feel free to be faster :)

> >> > I don't like leaving a few dozen non-compiling header files into uapi.
> >>
> >> I agree, but I do not see a simple solution.
> >>
> >> Unless libc has the analogous change, including either <time.h> or
> >> <linux/time.h> in userspace can unfortunately cause breakage.
> >>
> >> The added include if __KERNEL__ is defined should be safe, though.
> >
> > Yes, for the kernel side, but your libc-compat change would nice for
> > userspace, where something will break for sure, but providing source
> > API compatibility is sometimes impossible.
> >
> > To summarize, this change from me, and your libc-compat.c for time.h, or?
> 
> I'm still afraid that this patch as is will break builds that include
> <linux/time.h> first.

I agree, but I also want uapi headers to cleanly compile. I know this might
break stuff on userspace side which rely on these broken header file
dependencies, but if the fix to just re-order include
statements I'm fine with it, also when the complaints hit my inbox.

If I had the CPU time, memory and disk space, I'd do a full yocto distro
build to see how badly userspace could break but I don't at home.

-Mikko