Re: [PATCH] rtc: adapt allowed RTC update error

[Jason Gunthorpe]

On Tue, Dec 01, 2020 at 03:38:35PM +0100, Miroslav Lichvar wrote:
> When the system clock is marked as synchronized via adjtimex(), the
> kernel is expected to copy the system time to the RTC every 11 minutes.
> 
> There are reports that it doesn't always work reliably. It seems the
> current requirement for the RTC update to happen within 5 ticks of the
> target time in some cases can consistently fail for hours or even days.
> 
> It is better to set the RTC with a larger error than let it drift for
> too long.
> 
> Add a static variable to rtc_tv_nsec_ok() to count the checks. With each
> failed check, relax the requirement by one jiffie, and reset the counter
> when it finally succeeds. This should allow the RTC update to happen in
> a minute at most.
> 
> Signed-off-by: Miroslav Lichvar <mlichvar@xxxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: John Stultz <john.stultz@xxxxxxxxxx>
> Cc: Prarit Bhargava <prarit@xxxxxxxxxx>
> Cc: Jason Gunthorpe <jgg@xxxxxxxx>
>  include/linux/rtc.h | 17 +++++++++++++----
>  1 file changed, 13 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/rtc.h b/include/linux/rtc.h
> index 22d1575e4991..8d105f10ef6a 100644
> +++ b/include/linux/rtc.h
> @@ -218,21 +218,30 @@ static inline bool rtc_tv_nsec_ok(s64 set_offset_nsec,
>  				  struct timespec64 *to_set,
>  				  const struct timespec64 *now)
>  {
> -	/* Allowed error in tv_nsec, arbitarily set to 5 jiffies in ns. */
> -	const unsigned long TIME_SET_NSEC_FUZZ = TICK_NSEC * 5;
>  	struct timespec64 delay = {.tv_sec = 0,
>  				   .tv_nsec = set_offset_nsec};
> +	unsigned long time_set_nsec_fuzz;
> +	static unsigned int attempt;

Adding a static value instide a static inline should not be done

I'm not sure using a static like this is the best idea anyhow, if you
want something like this it should be per-rtc, not global

Did you look at why time has become so in-accurate in your system? 5
jiffies is usually a pretty long time?

Jason