On Wed, Apr 30, 2014 at 05:56:53PM +0100, Sebastian Andrzej Siewior wrote:

* Will Deacon | 2014-04-30 14:26:28 [+0100]:

quoted

I don't think that's the problem I was referring to. What I mean is that a
clocksource might overflow at any number of bits, so the delay calculation
needs to take this into account when it does:

while ((get_cycles() - start) < cycles)

because a premature overflow from get_cycles() will cause us to return
early. The solution is to mask the result of the subtraction before the
comparison to match the width of the clock.

So I got this:

[...]

Is this what you had in mind? If so, there is one user of
register_current_timer_delay() which I didn't convert. That is
arch_timer_delay_timer_register(). It returns arch_counter_get_cntvct()
which seems to return an u64 (which is truncated to 32bit). However
arch_counter_register() registers the clocksource with 56bits. So this
does not look too good, right?

That should be fine, I think there's only an issue if you can overflow
twice during a single delay operation, so the thing would need to be
ticking at quite a frequency for that to happen!

The other thing I noticed is
|arch/arm/include/asm/timex.h:typedef unsigned long cycles_t;

This works for clocksource because timekeeping is using
|include/linux/clocksource.h:typedef u64 cycle_t;

instead.
Do I assume correct, that the arch_timer is really providing a number
wider than 32bit? Shouldn't I then promote cycles_t to 64bit if that
timer is active? Unless you have better suggestions of course :)

The architected timer is guaranteed to be at least 56 bits wide, but I
think we can safely truncate delay sources to 32-bit.

So actually, we only have a problem if people want to register delay clocks
smaller than 32-bit. Maybe it's simpler to enforce at least 32-bit precision
and don't bother with the registration if the clock is smaller than that?
You could use sizeof(cycles_t) to parameterise that.

Will