Re: 2.6.14-rt22 (and mainline) excessive latency

[Lee Revell <rlrevell@joe-job.com>]

On Wed, 2005-12-21 at 05:36 -0800, Paul E. McKenney wrote:
> On Tue, Dec 20, 2005 at 10:32:48PM -0500, Lee Revell wrote:
> > On Tue, 2005-12-20 at 17:47 -0800, Paul E. McKenney wrote:
> > > On Tue, Dec 20, 2005 at 05:24:42AM +0100, Ingo Molnar wrote:
> > > > 
> > > > * Lee Revell <rlrevell@joe-job.com> wrote:
> > > > 
> > > > > I captured this 3+ ms latency trace when killing a process with a few 
> > > > > thousand threads.  Can a cond_resched be added to this code path?
> > > > 
> > > > >     bash-17992 0.n.1   29us : eligible_child (do_wait)
> > > > > 
> > > > >     [ 3000+ of these deleted ]
> > > > > 
> > > > >     bash-17992 0.n.1 3296us : eligible_child (do_wait)
> > > > 
> > > > Atomicity of signal delivery is pretty much a must, so i'm not sure this 
> > > > particular latency can be fixed, short of running PREEMPT_RT. Paul E.  
> > > > McKenney is doing some excellent stuff by RCU-ifying the task lookup and 
> > > > signal code, but i'm not sure whether it could cover do_wait().
> > > 
> > > Took a quick break from repeatedly shooting myself in the foot with
> > > RCU read-side priority boosting (still have a few toes left) to take
> > > a quick look at this.  The TASK_TRACED and TASK_STOPPED cases seem
> > > non-trivial, and I am concerned about races with exit.
> > > 
> > > Any thoughts on whether the latency is due to contention on the
> > > tasklist lock vs. the "goto repeat" in do_wait()?
> > 
> > It's a UP system so I'd be surprised if there were any contention.
> 
> Couldn't there be contention due to preemption of someone holding
> the tasklist lock?

But I'm running with PREEMPT_DESKTOP (specifically I configured a system
to have the exact same preemption model as mainline - PREEMPT_DESKTOP
with no soft/hardirq preemption) so holding a spinlock will disable
preemption.

Lee