Re: Fix nanosleep returning negative rem

[Corinna Vinschen <corinna-cygwin@cygwin.com>]

On Jul 21 16:02, David Allsopp wrote:
> Corinna Vinschen wrote:
> > Sent: 21 July 2021 10:33
> > To: cygwin-patches@cygwin.com
> > Subject: Re: Fix nanosleep returning negative rem
> > 
> > On Jul 21 11:30, Corinna Vinschen wrote:
> > > I wrote a quick STC using the NT API calls and I can't reproduce the
> > > problem with this code either.  The output is either
> > >
> > >   SignalState: 1 TimeRemaining: -5354077459183
> > >
> > > or
> > >
> > >   SignalState: 0 TimeRemaining: 653
> > >
> > > I never get a small negative value in the latter case.  Can you
> > > reproduce your problem with this testcase or tweak it to reproduce it?
> > 
> > Now I actually attached the code :}
> 
> :) Yes, I can reproduce - I didn't even need a loop! Third time:
> 
>   dra@Thor /cygdrive/c/Scratch/nanosleep
>   $ ./timer
>   SignalState: 0 TimeRemaining: -1151
> 
> That said, I can get it easily get this on my desktop (AMD Ryzen
> Threadripper 3990X) but not at all on my laptop (Intel Core i7-8650U).
> On the laptop, ignoring the couple of signalled cases, 747 runs of
> timer.c give values between 131597-149947 with a very narrow SD
> (~4000) whereas on the AMD chip, 738 runs gives a range of -2722 to
> 149896 with a relatively wider SD of ~23000.
> 
> The CI system where this was first seen is an virtualised Intel system
> so it doesn't appear to be as simple as CPU manufacturer or even core
> count.

Weird.  I pushed your patch.

> That said, I'm not at all familiar with the details of how this
> works, but I expect the timer for these things is part of the chipset,
> not the CPU?!

As timers go, the timer behind the NtCreateTimer scenes is a simple
interrupt based timer with a default resolution of 15 ms.


Thanks,
Corinna