Re: Poor granularity of usleep impl on Win32

[Daniel P. Berrange]

On Wed, Apr 23, 2014 at 08:38:07PM -0600, Eric Blake wrote:
> On 04/23/2014 04:26 AM, Daniel P. Berrange wrote:
> > The gnulib usleep replacement says
> > 
> >   /* This file is _intentionally_ light-weight.  Rather than using
> >      select or nanosleep, both of which drag in external libraries on
> >      some platforms, this merely rounds up to the nearest second if
> >      usleep() does not exist.  If sub-second resolution is important,
> >      then use a more powerful interface to begin with.  */
> 
> Modern mingw provides usleep - it's a version that can't sleep more than
> 1 second, but it at least has millisecond resolution.  The fallback to
> 1-second granularity only occurs for platforms that lack usleep()
> altogether.
> 
> > 
> > And the code confirms it
> > 
> >     int
> >     usleep (useconds_t micro)
> >     {
> >       unsigned int seconds = micro / 1000000;
> >       if (sizeof seconds < sizeof micro && micro / 1000000 != seconds)
> >         {
> >           errno = EINVAL;
> >           return -1;
> >         }
> >       if (!HAVE_USLEEP && micro % 1000000)
> >         seconds++;
> >       while ((seconds = sleep (seconds)) != 0);
> >     
> >     #undef usleep
> >     #if !HAVE_USLEEP
> 
> Have you confirmed the value of HAVE_USLEEP in config.h for the build of
> mingw that you are using?
> 
> >     # define usleep(x) 0
> >     #endif
> >       return usleep (micro % 1000000);
> >     }
> >     
> > 
> > The 'sleep' replacement on Win32 calls into the Win32-specific Sleep()
> > function which allows milli-second granularity. Why doesn't usleep()
> > call into Sleep() directly, so it gets milli-second granularity rather
> > than rounding up to the nearest second ?
> 
> As far as I can tell, mingw would be the only platform that would
> benefit by calling Sleep() - but if everyone these days is already using
> new enough mingw that provides usleep(), then the granularity problem is
> a red herring.  So at this point, I think there's nothing further worth
> patching in gnulib.

Hmm, I must admit I didn't check that. I was porting libgphoto2 (which
does NOT use gnulib) to mingw64 and hit build failures with lack of
usleep. I just assumed it was missing since Win32 doesn't provide it,
but I think in fact it must just have been wrong include files used.


Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|