Re: [patch 00/11] x86/vdso: Cleanups, simmplifications and CLOCK_TAI support

[Arnd Bergmann <arnd@arndb.de>]

On Mon, Sep 17, 2018 at 3:00 PM Thomas Gleixner <tglx@linutronix.de> wrote:
>
> On Fri, 14 Sep 2018, Arnd Bergmann wrote:
> > On Fri, Sep 14, 2018 at 2:52 PM Thomas Gleixner <tglx@linutronix.de> wrote:
> > A couple of architectures (s390, ia64, riscv, powerpc, arm64)
> > implement the vdso as assembler code at the moment, so they
> > won't be as easy to consolidate (other than outright replacing all
> > the code).
> >
> > The other five:
> > arch/x86/entry/vdso/vclock_gettime.c
> > arch/sparc/vdso/vclock_gettime.c
> > arch/nds32/kernel/vdso/gettimeofday.c
> > arch/mips/vdso/gettimeofday.c
> > arch/arm/vdso/vgettimeofday.c
> >
> > are basically all minor variations of the same code base and could be
> > consolidated to some degree.
> > Any suggestions here? Should we plan to do that consolitdation based on
> > your new version, or just add clock_gettime64 in arm32 and x86-32, and then
> > be done with it? The other ones will obviously still be fast for 32-bit time_t
> > and will have a working non-vdso sys_clock_getttime64().
>
> In principle consolidating all those implementations should be possible to
> some extent and probably worthwhile. What's arch specific are the actual
> accessors to the hardware clocks.

Ok.

> > I also wonder about clock_getres(): half the architectures seem to implement
> > it in vdso, but notably arm32 and x86 don't, and I had not expected it to be
> > performance critical given that the result is easily cached in user space.
>
> getres() is not really performance critical, but adding it does not create
> a huge problem either.

Right, I'd just not add a getres_time64() to the vdso then.

        Arnd