[BUG] clockevents : clockevent_delta2ns usage

[BUG] clockevents : clockevent_delta2ns usage

[Marc Pignat]

Hello all!

We're currently working with David Brownell and Remy Bohmer on implementation
of clocksource/events on the at91rm9200 processor (arm arch).

While debugging our implementation, we came about a rounding problem, in the
setting of min_delta_ns field, and I think it can be in almost all clock_event_device.

Almost all clock_event_device initialize the field min_delta_ns this way:
	clkevt32k.mult = div_sc(CLOCK_TICK_RATE, NSEC_PER_SEC, clkevt32k.shift);
	clkevt32k.max_delta_ns = clockevent_delta2ns(0x000fffff, &clkevt32k);
	clkevt32k.min_delta_ns = clockevent_delta2ns(1, &clkevt32k);

The clockevent_delta2ns function rounds the result down, so the rounded min_delta_ns
can be smaller than the real min_delta_ns.

When clockevents_program_event is called, this rounding problem shows up:
...
	if (delta < dev->min_delta_ns)
		delta = dev->min_delta_ns;

	clc = delta * dev->mult;
	clc >>= dev->shift;

	return dev->set_next_event((unsigned long) clc, dev)

When delta is to small, min_delta_ns will be used, and with a wrong rounding,
will call set_next_event with a value smaller than the expected, and 
set_next_event will fail.

This will result in an infinite loop when clockevents_program_event is called
from tick_program_event with a too short delay.

In our driver, we fixed this problem using that:
-	clkevt32k.min_delta_ns = clockevent_delta2ns(1, &clkevt32k);
+	clkevt32k.min_delta_ns = clockevent_delta2ns(1, &clkevt32k)+1;

I think other clock_event_device can have the same problem.

Regards

Marc

Re: [BUG] clockevents : clockevent_delta2ns usage

[David Brownell]

On Monday 02 July 2007, Marc Pignat wrote:
> Hello all!
> 
> We're currently working with David Brownell and Remy Bohmer on implementation
> of clocksource/events on the at91rm9200 processor (arm arch).
> 
> While debugging our implementation, we came about a rounding problem, in the
> setting of min_delta_ns field, and I think it can be in almost all clock_event_device.
> 
> Almost all clock_event_device initialize the field min_delta_ns this way:
> 	clkevt32k.mult = div_sc(CLOCK_TICK_RATE, NSEC_PER_SEC, clkevt32k.shift);
> 	clkevt32k.max_delta_ns = clockevent_delta2ns(0x000fffff, &clkevt32k);
> 	clkevt32k.min_delta_ns = clockevent_delta2ns(1, &clkevt32k);
> 
> The clockevent_delta2ns function rounds the result down, so the rounded min_delta_ns
> can be smaller than the real min_delta_ns.

For concrete numbers:  CLOCK_TICK_RATE = 32768, making 30517.58 nsec
per tick  But min_delta_ns becomes 30517.  (And the shift is 32 bits;
in-kernel docs could stand to include guidance about how to choose
a good shift value.)


> When clockevents_program_event is called, this rounding problem shows up:
> ...
> 	if (delta < dev->min_delta_ns)
> 		delta = dev->min_delta_ns;
> 
> 	clc = delta * dev->mult;
> 	clc >>= dev->shift;
> 
> 	return dev->set_next_event((unsigned long) clc, dev)
> 
> When delta is to small, min_delta_ns will be used, and with a wrong rounding,
> will call set_next_event with a value smaller than the expected, and 
> set_next_event will fail.

Will be passed a value of zero, that is -- rather troublesome.
 

> This will result in an infinite loop when clockevents_program_event is called
> from tick_program_event with a too short delay.

Unless zero ticks gets special cased to return -ETIME and then the
backup strategies kick in, eventually asking for more ticks...

 
> In our driver, we fixed this problem using that:
> -	clkevt32k.min_delta_ns = clockevent_delta2ns(1, &clkevt32k);
> +	clkevt32k.min_delta_ns = clockevent_delta2ns(1, &clkevt32k)+1;
> 
> I think other clock_event_device can have the same problem.
> 
> Regards
> 
> Marc
>