Re: 2.6.24-rc6-mm1

[Peter Zijlstra <peterz@infradead.org>]

On Wed, 2008-01-02 at 21:31 +1100, Nick Piggin wrote:
> On Monday 31 December 2007 00:10, Ingo Molnar wrote:
> > * Herbert Xu <herbert@gondor.apana.org.au> wrote:
> > > > Ingo, it's not good that we have cond_resched() definitions
> > > > conditionally duplicated in kernel.h - that's increasing the risk of
> > > > bugs like this one.
> > >
> > > Actually, why do we even have cond_resched when real preemption is on?
> > > It seems to be a waste of space and time.
> >
> > due to the BKL. cond_resched() in BKL code breaks up BKL latencies.
> >
> > i dont mind not doing that though - we should increase the pain for BKL
> > users, so that subsystems finally get rid of it altogether.
> > lock_kernel() use within the kernel is still rampant - there are still
> > more than 400 callsites to lock_kernel().
> 
> It would be silly to potentially increase latency in some areas
> for CONFIG_PREEMPT kernels, though.
> 
> Better may be to detect when there is CONFIG_PREEMPT and
> CONFIG_PREEMPT_BKL, and ifdef away the cond_resched in that case
> (or -- why do we even make CONFIG_PREEMPT_BKL an option? Are there
> really workloads left where it causes throughput regressions?)

I've seen 1s+ desktop latencies due to PREEMPT_BKL when I was still
using reiserfs.

Both reiserfs and tty were fighting for the bkl and massive prio
inversion ensued. Turning PREEMPT_BKL off made the system usable again.