Dropping "depends on SMP" for HAVE_ARM_TWD -- take 2

Russell King - ARM Linux linux at arm.linux.org.uk
Mon Oct 5 08:25:22 PDT 2015

On Mon, Oct 05, 2015 at 08:13:03AM -0700, Sören Brinkmann wrote:
> On Mon, 2015-10-05 at 05:19PM +0530, Afzal Mohammed wrote:
> > Hi,
> > 
> > On Sun, Oct 04, 2015 at 10:46:52PM -0700, Sören Brinkmann wrote:
> > > On Sat, 2015-10-03 at 11:12AM +0100, Marc Zyngier wrote:
> > 
> > > > Indeed, I cannot see any code that does that in the GT driver. But if
> > > > you have an A9 MP, you probably want to stick to TWD, which gives you a
> > > > per-cpu timer instead of a global timer that will require IPIs to other
> > > > CPUs.
> > > 
> > > I think the TWD only provides a clock_event device. Clocksource and
> > > schedclock would have to be provided by something else.
> > 
> > If no clocksource, sched clock is provided, default jiffies based ones
> > would be sufficient for single core, no ?, though not a preferred one.
> Yes that is correct. My point was more, that twd+cpufreq is a rather
> easy case since it doesn't provide a clocksource/schedclock. A driver
> that provides those needs quite some more work if it depends on the CPU
> frequency and needs to be usable with cpufreq.

_And_ you need to take the decision that you don't care about time
keeping accuracy - because each and every time you change the
clocksource's frequency, you will never be able to know exactly when
the frequency change took effect in comparison with the counter value.

You could do a very tight "read counter before, write clock rate, read
counter after" and use that to calculate the number of timer ticks at
the old rate, and re-set the timing parameters with the new initial
counter value after the change, but even that's not perfect as you don't
actually know when the write hits the hardware compared to the reads.

FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

More information about the linux-arm-kernel mailing list