What's with people sending out replies without subjects?
On Tue, 15 Nov 2016 14:29:16 -0600 (CST)
Christoph Lameter <[email protected]> wrote:
> > > There is a deadlock, Peter!!!
> >
> > Describe please? Also, have you tried disabling RT_RUNTIME_SHARE ?
> >
>
>
> The description was given earlier in the the thread and the drawbacks of
> using RT_RUNTIME_SHARE as well.
Peter is saying to disable that, because of the drawbacks. I think the
question is, do you still have a deadlock after disabling it? If not,
then that's your solution.
-- Steve
On Tue, Nov 15, 2016 at 02:29:16PM -0600, Christoph Lameter wrote:
>
> > > There is a deadlock, Peter!!!
> >
> > Describe please? Also, have you tried disabling RT_RUNTIME_SHARE ?
> >
>
>
> The description was given earlier in the the thread and the drawbacks of
> using RT_RUNTIME_SHARE as well.
I've not seen a deadlock described. It either was an unbounded priority
inversion or a starvation issue, both of which are 'design' features of
the !rt kernel.
Neither things are new, so its not a regression either.
And, as stated, I'm not really happy to muck with this known troublesome
code and add features for which we must then maintain feature parity
when replacing it either.
On top of which, the implementation had issues; now I know you're the
blinder kind of person that disregards everything not in his immediate
interest, but if you'd looked at the patch you'd have seen he'd added
code the idle entry path, which will slow down every single to-idle
transition.
On Wed, 16 Nov 2016 11:40:14 +0100
Peter Zijlstra <[email protected]> wrote:
> On top of which, the implementation had issues; now I know you're the
> blinder kind of person that disregards everything not in his immediate
> interest, but if you'd looked at the patch you'd have seen he'd added
> code the idle entry path, which will slow down every single to-idle
> transition.
Isn't to-idle a bit bloated anyway? Or has that been fixed. I know
there was some issues with idle_balance() which can add latency to
wakeups. idle_balance() is also in the to-idle path.
Note, that this is a sched feature which would be a nop (jump_label)
when disabled. And I'm sure it could also be optimized to be a static
inline as well when it is enabled.
I'm not saying we need to go this approach, but I'm just saying that
the to-idle issue is a bit of a red herring.
-- Steve
On Wed, Nov 16, 2016 at 09:25:43AM -0500, Steven Rostedt wrote:
> On Wed, 16 Nov 2016 11:40:14 +0100
> Peter Zijlstra <[email protected]> wrote:
>
>
> > On top of which, the implementation had issues; now I know you're the
> > blinder kind of person that disregards everything not in his immediate
> > interest, but if you'd looked at the patch you'd have seen he'd added
> > code the idle entry path, which will slow down every single to-idle
> > transition.
>
> Isn't to-idle a bit bloated anyway? Or has that been fixed. I know
> there was some issues with idle_balance() which can add latency to
> wakeups. idle_balance() is also in the to-idle path.
>
Yes it is too heavy as is, but just stacking more crap in just because
its already expensive seems to wrong way around.