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

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

On Fri, 2008-04-11 at 11:27 +0200, Michael Kerrisk wrote:
> On Fri, Apr 11, 2008 at 11:21 AM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> >
> > On Fri, 2008-04-11 at 11:16 +0200, Michael Kerrisk wrote:
> > > On Fri, Apr 11, 2008 at 11:01 AM, Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > > > On Fri, 2008-04-11 at 10:56 +0200, Michael Kerrisk wrote:
> > > > > On Thu, Feb 28, 2008 at 5:21 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.
> > > > > >
> > > > > > Upon reaching the cur limit we start giving SIGXCPU every second, upon
> > > > > > reaching the hard limit we give SIGKILL - matching RLIMIT_CPU.
> > > > > >
> > > > > > Time is measured in tick granularity (for now).
> > > > >
> > > > > So I have another question: why is the granularity of this rlimit
> > > > > microseconds?  On the one hand, specifying limits down at the
> > > > > microsecond level seems (to my naive eye) unlikely to be useful.  (But
> > > > > perhaps I have missed a thread where this was explained.)  On the
> > > > > other hand, it means that on 32-bit the largest time limit we can set
> > > > > is ~4000 seconds, and I wonder if there are scenarios where it might
> > > > > be useful to have larger limits than that.
> > > > >
> > > > > Why not, for example, have a granularity of milliseconds?
> > > >
> > > > The us scale seemed the best fit in that it allows sub-ms granularity
> > > > while still allowing for quite long periods too. I'd preferred ns scale
> > > > as that is what we use throughout the scheduler where possible - but
> > > > that seemed too restrictive at the high end.
> > > >
> > > > No real hard arguments either way.
> > >
> > > I'm curious: what scenarios require sub-millisecond timeouts?
> >
> > I'm not sure, nor will they actually work atm since its tick based.
> 
> Just to make sure me and the man page are clear: by tick-based, you
> mean the granularity is in jiffies, right?

Yes, they are currently jiffy based - but I could do hrtimer if someone
shows need.

> > But
> > I'm not wanting to exclude too many things, and 4k second upper limit is
> > plenty large.
> 
> Okay.
> 
> And following on from my other conversation in this thread...  What
> should/will be the specified behavior w.r.t. resetting or not
> resetting the timer on a sched_yield()?

I think I'll keep it as is; so sched_yield() will _not_ reset the
counter. The rationale is that the process didn't actually stop running
- it just got scheduled away.