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

[Adhemerval Zanella Netto]

On 07/03/23 08:11, Florian Weimer wrote:
> * Adhemerval Zanella:
> 
>> Different than gettimeofday and timespec_get, time uses
>> CLOCK_REALTIME_COARSE instead of CLOCK_REALTIME on Linux.  The
>> coarse time is used mostly as optimization, but it may show
>> divergence progression due the clock resolution.
>>
>> For x86_64 and powerpc64, it should add slight more latency since
>> it would call now clock_gettime internally.
> 
> It seems really significant on x86-64.
> 
> Before:
> 
> min:        14 ns
> 25%:        16 ns
> 50%:        17 ns
> 75%:        17 ns
> 95%:        18 ns
> 99%:        18 ns
> max:     18722 ns
> avg:   16.6606 ns
> 
> After:
> 
> min:        29 ns
> 25%:        31 ns
> 50%:        31 ns
> 75%:        32 ns
> 95%:        32 ns
> 99%:        33 ns
> max:     12161 ns
> avg:   31.2205 ns
> 
> And of those original 17 ns, quite a bit is overhead from the
> benchmarking loop.  I guess applications could work around it by having
> a background timer thread that increments a global variable and use that
> instead of the time function call, but that seems not a great approach.

Yes, this is expected since time call will be route through clock_gettime.
Another fix would be to convince kernels developers to use CLOCK_REALTIME
on vDSO as well.

> 
> Based on previous feedback, I expect we'd have to carry a downstream
> revert of this patch indefinitely, so I'm rather strongly against
> applying it upstrean.

To me it really seems like a over-optimization specially because 'time'
has only second resolution.