Re: [PATCH] proc: Add RLIMIT_RTTIME to /proc/<pid>/limits

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Fri, 2008-04-11 at 09:38 +0200, Michael Kerrisk wrote:
> On Thu, Feb 28, 2008 at 5:50 PM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> >
> > On Thu, 2008-02-28 at 16:44 +0100, Michael Kerrisk wrote:
> > > Peter,
> > >
> > > Thanks for the text.
> > >
> > > On Thu, Feb 28, 2008 at 4:21 PM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > > >
> > > >  On Thu, 2008-02-28 at 16:12 +0100, Michael Kerrisk wrote:
> > > >  > Peter,
> > > >  >
> > > >  > Could you please provide some text describing RLIMIT_RTTIMEfor the
> > > >  > getrlimit.2 man page.
> > > >
> > > >  The rlimit sets a timeout in [us] for SCHED_RR and SCHED_FIFO tasks.
> > > >  This time is measured between sleeps, so a schedule in RR or a
> > > >  preemption in either is not a sleep - the task needs to be dequeued and
> > > >  enqueued for the timer to reset.
> > >
> > > Just to clarify: sleep here means a call to some blocking syscall
> > > (e.g., nanosleep(), read(), select(), etc.), right?  Is there anything
> > > else that falls under the category of "sleep"?  What about a call to
> > > sched_yield() where the process explicitly lets go of the CPU?
> >
> > Yes, and yes, others would be blocking on futexes and the like.
> 
> Peter,
> 
> I've been testing this patch.  Above you seemed to be saying that
> doing a sched_yield() would be equivalent to a sleep, causing the rt
> counter to be reset to zero.  Howver, the results I'm seeing suggest
> that a sched_yield() does not cause the counter to be reset to zero
> (i.e., despite calling sched_yield() at frequent intervals, the
> process still encounters the RLIM_RTTIME soft limit and gets SIGXCPU).
>  Can you comment?

It appears you are right. I must have been staring at something else
than code when I said that :-(, yield() will indeed _not_ reset the
counter.

Now, I think it makes some sense to reset it, because we do try to play
nice by calling yield. OTOH we don't actually block and become
unrunnable - we'll still be contending for CPU time.