radeon breaks with clocksource_jiffies

radeon breaks with clocksource_jiffies

[David Miller]

While probeing PLL information via radeon_get_pllinfo(), it does a
"gettimeofday(); do_something(); gettimeofday();" type sequence
explicitly with interrupts disabled, so ends up with a zero
measurement which then results in a bunch of divisions by zero.

We should decide whether gettimeofday() can be expected to advance with
interrupts disabled, or that clocksource_jiffies is simply invalid because
of this behavior.

Re: radeon breaks with clocksource_jiffies

[David Miller]

From: David Miller <davem@davemloft.net>
Date: Fri, 23 Feb 2007 13:45:05 -0800 (PST)

> While probeing PLL information via radeon_get_pllinfo(), it does a
> "gettimeofday(); do_something(); gettimeofday();" type sequence
> explicitly with interrupts disabled, so ends up with a zero
> measurement which then results in a bunch of divisions by zero.
> 
> We should decide whether gettimeofday() can be expected to advance with
> interrupts disabled, or that clocksource_jiffies is simply invalid because
> of this behavior.

Actually, with clocksource based gettimeofday(), radeon built-in cannot
work at all.

The reason is that the clocksource code will not jump over to a
clock source other than clocksource_jiffies until the late initcalls
are run, because of this code:

/* clocksource_done_booting - Called near the end of bootup
 *
 * Hack to avoid lots of clocksource churn at boot time
 */
static int __init clocksource_done_booting(void)
{
	finished_booting = 1;
	return 0;
}
late_initcall(clocksource_done_booting);
 ...
struct clocksource *clocksource_get_next(void)
{
	unsigned long flags;

	spin_lock_irqsave(&clocksource_lock, flags);
	if (next_clocksource && finished_booting) {
		curr_clocksource = next_clocksource;
		next_clocksource = NULL;
	}
	spin_unlock_irqrestore(&clocksource_lock, flags);

	return curr_clocksource;
}

So even if other clock sources are provided, they'll never show up until
after the radeon driver tries to initialize.

I understand the intention of the clocksource_get_next() code, it doesn't
want to hop onto several different clocksource implementations as the
drivers for those startup, but going through all the module_init() calls
for various drivers without a sane clocksource is just as bad of a problem
and is in fact fatal in this radeon case.

Re: radeon breaks with clocksource_jiffies

[john stultz-lkml]

On 2/23/07, David Miller <davem@davemloft.net> wrote:
> > While probeing PLL information via radeon_get_pllinfo(), it does a
> > "gettimeofday(); do_something(); gettimeofday();" type sequence
> > explicitly with interrupts disabled, so ends up with a zero
> > measurement which then results in a bunch of divisions by zero.
> >
> > We should decide whether gettimeofday() can be expected to advance with
> > interrupts disabled, or that clocksource_jiffies is simply invalid because
> > of this behavior.
>
> Actually, with clocksource based gettimeofday(), radeon built-in cannot
> work at all.
>
> The reason is that the clocksource code will not jump over to a
> clock source other than clocksource_jiffies until the late initcalls
> are run, because of this code:

Check out the patch I sent out yesterday. It should resolve this problem.

thanks
-john

Re: radeon breaks with clocksource_jiffies

[David Miller]

From: "john stultz-lkml" <johnstul.lkml@gmail.com>
Date: Fri, 23 Feb 2007 14:10:07 -0800

> Check out the patch I sent out yesterday. It should resolve this problem.

Yep, changing that finished_booting setting into an fs_initcall()
fixes this issue.

Re: radeon breaks with clocksource_jiffies

[john stultz]

Crud, my poor gmail skills dropped lkml on the CC list for this one.

On 2/23/07, john stultz <johnstul@us.ibm.com> wrote:
> On 2/23/07, David Miller <davem@davemloft.net> wrote:
> > While probeing PLL information via radeon_get_pllinfo(), it does a
> > "gettimeofday(); do_something(); gettimeofday();" type sequence
> > explicitly with interrupts disabled, so ends up with a zero
> > measurement which then results in a bunch of divisions by zero.
>
> This is at module init time?  If so I just sent out a patch yesterday
> to akpm that might help this issue (assuming other clocksources are
> available on the hardware).
>
>
> > We should decide whether gettimeofday() can be expected to advance with
> > interrupts disabled, or that clocksource_jiffies is simply invalid because
> > of this behavior.
>
> Some arches have no fine-grained timekeeping, so I don't think it can
> always be assumed, but where hardware is available it should function.
>
> It should be noted that with older kernels, if interrupts were
> disabled right before a tick, it would be possible that
> gettimeofday()'s limiting code would cause time to stop advancing
> until interrupts were re-enabled. So theoretically its not really a
> new issue.
>
> The timekeeping_is_continuous() function could be used to flag this
> sort of "software driven vs continuous" clocksource cases.
>
> thanks
> -john
>