Re: [PATCH v4 1/2] y2038: linux: Provide __timerfd_gettime64 implementation

[Arnd Bergmann <arnd at arndb dot de>]

On Tue, Jan 7, 2020 at 9:16 PM Adhemerval Zanella
<adhemerval.zanella@linaro.org> wrote:
> On 07/01/2020 11:25, Lukasz Majewski wrote:
> >> On 07/01/2020 06:27, Lukasz Majewski wrote:
> >>
> >>>> As a side note, now that arch-syscall patch is upstream should we
> >>>> assume that for !__ASSUME_TIME64_SYSCALLS the
> >>>> __NR_timerfd_gettime64 should be defined (meaning that Linux
> >>>> supports time64 for all 32-bit architectures)?
> >>>
> >>> Only Linux version >= 5.1 supports 64 bit time on archs with
> >>> __WORDSIZE = 32. I do guess (but I may be wrong here) that the
> >>> arch-syscall is supposed to reflect the exact syscalls provided by
> >>> kernel headers used for building (to help with validation of Y2038
> >>> patches).
> >>
> >> The arch-syscall is now autogenerated from the latest kernel release
> >> defined in build-many-glibcs.py. So the question is whether Linux
> >> support and enforces time64 support on all and future 32-bit
> >> architectures or if there is still some missing ones (as it has
> >> happen on some syscall additions, where some architecture lag
> >> behind some releases).
> >
> > This question would be best answered by Arnd (CC'ed) IMHO. From what I
> > know all 32 bit architectures gained syscalls covered by
> > __ASSUME_TIME64_SYSCALLS from Linux 5.1+.
> >
> > The arch-syscall seems to me like a mean to test for example the time
> > related syscalls which use different versions (32bit time vs 64 bit) on
> > different archs. Notable example - clock_gettime(). Am I right?
>
> The arch-syscall is a way to decouple the build from the kernel header
> used on build, which might simplify the logic to use some kernel
> features.
>
> On the clock_gettime, for instance, as Arnd has indicated we can
> assume that __NR_clock_gettime64 will be always presented for
> !__ASSUME_TIME64_SYSCALLS.
>
> It would be interesting if kernel also could enforce that new
> generic syscalls would be wire-up, or at least the syscall number
> reserved; once a new generic syscall is introduced.  It would
> simplify the __ASSUME_* macro, not requiring the arch-specific
> overrides on some architectures.

We currently try to enforce it through review since the introduction
of the automatically generated asm/unistd.h header files, but his
has already failed one time for the recent clone3 system call that
accidentally was missing on one architecture.

I have some plans to enforce it using tooling, but it should
at least be a safe assumption that all new system calls have
the same numbers across all architectures and are added at
the same time.

I also have some vague plans to automatically generate not
only the asm/unistd.h header but also some trivial wrappers
around syscall() that provide an inline function with the correct
arguments, to allow calling any syscall without relying on libc
to provide a wrapper for it.

          Arnd