Re: [PATCH] Implement getloadavg()

[Jon Turney <jon dot turney at dronecode dot org dot uk>]

On 20/03/2017 15:40, Corinna Vinschen wrote:
> On Mar 20 15:04, Jon Turney wrote:
> > On 20/03/2017 10:37, Corinna Vinschen wrote:
> > > On Mar 17 17:50, Jon Turney wrote:
> > > The load average is global, non-critical data.  So what about storing it
> > > in shared_info instead?  This way, only the first call of the first
> > > Cygwin process returns all zero.
> >
> > Ok.
> >
> > > > +static bool load_init (void)
> > > > +{
> > > > +  static bool tried = false;
> > > > +  static bool initialized = false;
> > > > +
> > > > +  if (!tried) {
> > > > +    tried = true;
> > > > +
> > > > +    if ((PdhOpenQueryA (NULL, 0, &query) == ERROR_SUCCESS) &&
> > > > +	(PdhAddEnglishCounterA (query, "\\Processor(_Total)\\% Processor Time",
> > > > +				0, &counter1) == ERROR_SUCCESS) &&
> > > > +	(PdhAddEnglishCounterA (query, "\\System\\Processor Queue Length",
> > > > +				0, &counter2) == ERROR_SUCCESS)) {
> > > > +      initialized = true;
> > > > +    } else {
> > > > +      debug_printf("loadavg PDH initialization failed\n");
> > > > +    }
> > > > +  }
> > > > +
> > > > +  return initialized;
> > > > +}
> > >
> > > How slow is that initialization?  Would it {make sense|hurt} to call it
> > > once in the initalization of Cygwin's shared mem in shared_info::initialize?
> >
> > I don't think that's particularly heavyweight, and I didn't see anything to
> > suggest that PDH query handles can be shared between processes, but I'll
> > look into it.
>
> Oh, right, that might pose a problem.

I can't find anything which documents these handles as shareable.

In practise, they seem to randomly stop working after a while after the 
process which created the handle exits, or something like that. :S

> But even then:
>
> The first process creating shared_info could call this and prime the values
> with a first call to getloadavg.  Each other process would have to init its

We cannot determine an initial value of load when shared_info is 
created, as the %CPU is not measured instantaneously, but over an 
interval.  This means that we can't have a load estimate until the 2nd 
time a process calls PdhCollectQueryData()

I've tweaked things slightly in v2 so the loadavg is initialized to the 
current load, rather than converging on it from 0.0.