Re: [RFC] wait for VX855 RTC to become ready during resume

[Andrew Morton <akpm@google.com>]

On Mon, 19 Sep 2011 11:36:37 +0100 (BST)
Daniel Drake <dsd@laptop.org> wrote:

> Hi,
> 
> As diagnosed at http://dev.laptop.org/ticket/10757 we are finding that
> the RTC on the VIA VX855 frequently is not 'ready' during system resume.
> 
> Reading from it produces zero. This causes the resume code to fail to
> increment the system clock for the amount of time that was spent sleeping,
> so the time is then wrong.
> 
> We have found that if we wait for the RTC_REF_CLCK_32KHZ signal, the RTC
> does soon start returning 'good' values.
> 
> The patch below demonstrates the solution we have found, but will probably
> make x86 maintainers cry. What would be a good approach to get this
> appropriately quirked and worked around? Detect VX855 southbridge on PCI bus
> and apply the quirk in this case? Do it from OLPC platform code?
> 
> Thanks,
> Daniel
> 
> 
> ---
>  arch/x86/kernel/rtc.c |   15 +++++++++++++++
>  1 files changed, 15 insertions(+), 0 deletions(-)
> 
> diff --git a/arch/x86/kernel/rtc.c b/arch/x86/kernel/rtc.c
> index 88ee70d..3ad2e47 100644
> --- a/arch/x86/kernel/rtc.c
> +++ b/arch/x86/kernel/rtc.c
> @@ -100,6 +100,21 @@ unsigned long mach_get_cmos_time(void)
>  	unsigned int status, year, mon, day, hour, min, sec, century = 0;
>  
>  	/*
> +	 * dev.laptop.org #10757, sometimes the RTC returns zero data
> +	 * immediately after a resume, so this code will give it some
> +	 * more time by waiting for the 32KHz clock bit to be set, in the
> +	 * hope that this means it is ready to be used.
> +	 */
> +	if (!(CMOS_READ(RTC_FREQ_SELECT) & RTC_REF_CLCK_32KHZ)) {
> +		int n = 0;
> +		printk(KERN_INFO "rtc is insane, waiting for it\n");
> +		while (!(CMOS_READ(RTC_FREQ_SELECT) & RTC_REF_CLCK_32KHZ)) {
> +			cpu_relax();
> +			n++;
> +		}
> +		printk(KERN_INFO "rtc was insane for %d counts\n", n);
> +	}

If the rtc is permanently insane, the kernel locks up.  I'd suggest
adding a timeout.

Perhaps also replace the cpu_relax() with a udelay(1), so the timeout
can be properly implemented and so that the information in that printk
has some useful meaning: "102 microseconds" is better than "11,434 counts".