libbsd select timeout issue(critial!!)

[Gedare Bloom]

On Tue, Nov 10, 2020 at 7:51 PM RUI Zhengxin <ruizx at 126.com> wrote:
>
> This issue is very critial, it can make the select abnormally return timeout after long run time.
> If the timeout is set to 5ms,  when the system run after 70000s(19.4h),
> the select function will wait 10ms timeout, which is two times of setting value.
>
>
> This issue is created by calculating timeout watchdog expire tick using the absolute time since libbsd5 version.
>
> 在 2020-11-03 10:26:39，"RUI Zhengxin" <ruizx at 126.com> 写道：
>
> Hello everyone,
>
>
> We meet an issue of select function timeout drift in rtems-libbsd5.1.
>
> The select timeout can drift 1ms every 14000 seconds.
>
>
> Kernel put the select thread to sleep queue in sleepq_set_timeout_sbt,
>
> the expire value is calculated by this
>
> expire = (sbt - SBT_1S + sbt_per_tick - 1) / sbt_per_tick;
>
>
> sbt is the absolute time,
>
> Precision is a constant value, no time drift.
>
> but the sbt_per_tick is not an accurate value,
>
> It can make the expire vaule bigger and bigger while sbt is increasing.
>
>
> in rtems_bsd_initialize function,
>
> rtems_bsd_sbt_per_watchdog_tick = SBT_1S / tps;
>
> in our system the tick rate is set to 1000Hz,
>
> The real value of rtems_bsd_sbt_per_watchdog_tick is 1<<32/1000=4294967.296
>
> but in integer type the value is 4294967.
>
> The error is 0.296/4294967 = 0.069ppm,
>
> and it can increase the expire value 1ms every 14000 seconds.
>

I don't think this bug can be fixed simply by arithmetic. Mixing
power-of-2 timeouts with power-of-10 ticks means the arithmetic will
always be subject to roundoff error.

The choices are to use a power-of-2 hertz, or to determine the drift
amount and reset your clock when you hit a round number. In your case,
the error is 0.296 per tick, so after 1000 ticks you should adjust by
296.

Or, if you can use HZ=1024 (if that is supported by your clock source)
then you will get 1<<32/1024 = 1<<22 precisely.

> Zhengxin
> Best regards.
>
>
>
>
> _______________________________________________
> users mailing list
> users at rtems.org
> http://lists.rtems.org/mailman/listinfo/users