Re: [PATCH v2 0/3] arm_arch_timer: VDSO preparation, code consolidation

[Catalin Marinas]

On Mon, Sep 22, 2014 at 11:30:23PM +0100, Russell King - ARM Linux wrote:
> On Mon, Sep 22, 2014 at 04:39:19PM +0100, Will Deacon wrote:
> > On Thu, Sep 18, 2014 at 03:59:32PM +0100, Nathan Lynch wrote:
> > > This series contains the necessary changes to allow architected timer
> > > access from user-space on 32-bit ARM.  This allows the VDSO to support
> > > high resolution timestamps for clock_gettime and gettimeofday.  This
> > > also merges substantially similar code from arm and arm64 into the
> > > core arm_arch_timer driver.
> > > 
> > > The functional changes are:
> > > - When available, CNTVCT is made readable by user space on arm, as it
> > >   is on arm64.
> > > - The clocksource name becomes "arch_mem_counter" if CP15 access to
> > >   the counter is not available.
> > > 
> > > These changes have been carried as part of the ARM VDSO patch set over
> > > the last several months, but I am splitting them out here as I assume
> > > they should go through the clocksource maintainers.
> > 
> > For the series:
> > 
> >   Acked-by: Will Deacon <will.deacon@xxxxxxx>
> > 
> > I'm not sure which tree the arch-timer stuff usually goes through, but
> > the arm/arm64 bits look fine so I'm happy for them to merged together.
> 
> I raised a while back with Will whether there's much point to having
> this on ARM.  While it's useful for virtualisation, the majority of
> 32-bit ARM doesn't run virtualised.

This has nothing to do with virtualisation. The main reason we use
CNTVCT is to not require kernel binary differences when running the OS
as host or guest. But it does _not_ mean that it is only used when
running as a guest.

> So there's little point in having the VDSO on the majority of
> platforms - it will just add additional unnecessary cycles slowing
> down the system calls that the VDSO is designed to try to speed up.

A good reason for VDSO is to avoid a system call for gettimeofday when
you can read the clocks source from user space. That's a significant
improvement on CPUs like A7, A15.

Now, I understand there will be a few cycles lost with the syscall
indirection via VDSO when the architected timers are not available. If
you are concerned about this, maybe the VDSO (or the corresponding
gettimeofday) function should be disabled at boot time. IIRC, the VDSO
gettimeofday function would override the glibc one at dynamic loading
time but if the VDSO isn't present, glibc would still work as usual.

-- 
Catalin
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/