Re: [PATCH] time: Use CLOCK_REALTIME for time (BZ #30200)

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 08/03/23 14:44, Bruno Haible wrote:
> Adhemerval Zanella Netto wrote:
>> ... all the trouble gnulib is pushing?
> 
> The trouble does not originate in gnulib; it originate through the
> use of inconsistent time APIs.

Yes I understand it and I agree with that this inconsistency is not
desirable, that exactly why I have sent this patch.  The 'trouble'
I meant all this boilerplate wrapper code you are adding on gnulib
to 'fix' this glibc optimization.

> 
> On a typical system, there are
>   - hundreds of programs that call gettimeofday() or timespec_get(),
>   - hundreds of programs that call time().
> Find attached the lists of programs in /usr/bin on my machine:
>   - 425 programs that call gettimeofday() or timespec_get(),
>   - 730 programs that call time().
> 
> Each time a program calls gettimeofday() or timespec_get(), and then,
> within 1 to 3 milliseconds, a program (the same or a different one)
> calls time(), there is the potential for trouble because the time
> appears to go backwards. This trouble can occur with probability
> between 0.1% and 0.3%, according to the data in
> https://sourceware.org/bugzilla/show_bug.cgi?id=30200 .
> 
> Gnulib attempts to reduce the trouble for specific programs, by
> moving them from the "call time()" camp to the "call gettimeofday() or
> timespec_get()" camp.
> 
> Bruno
>