Re: 2.6.15-rc5: latency regression vs 2.6.14 in exit_mmap->free_pgtables

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

On Thu, 2006-01-19 at 07:29 +0000, Hugh Dickins wrote:
> On Wed, 18 Jan 2006, Lee Revell wrote:
> > On Wed, 2005-12-28 at 22:59 +0000, Hugh Dickins wrote:
> > > 
> > > On my list to work on; but the TLB always needs great care, and this
> > > goes down into architectural divergences, with truncation of a mapped
> > > file adding further awkward constraints.  I imagine 2.6.16-rc1 is only
> > > a couple of weeks away, so it's unlikely to be fixed in 2.6.16 either.
> > 
> > Is this believed to be fixed in 2.6.16-rc1?
> 
> Not at all, I'm afraid.  Do you think I ought to try to persuade Linus
> and Andrew to take that ugly free_pgtables #ifdef CONFIG_PREEMPT patch
> in the interim before we've the proper latency fix there?  (I doubt the
> mmu_gather rewrite in 2.6.17 too, but perhaps a reasonable compromise.)

For the time being, this is overshadowed by long latencies (~8ms or so)
in rt_garbage_collect() and rt_run_flush().  Previously I was under the
impression those could be worked around by sysctl/proc tuning, but this
is not the case.

Until those are fixed I would say it's low priority.

Lee