lkml.org 
[lkml]   [2009]   [Feb]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] mm: disable preemption in apply_to_pte_range
From
Date
On Sat, 2009-02-14 at 01:30 +1100, Nick Piggin wrote:
> On Saturday 14 February 2009 01:16:51 Peter Zijlstra wrote:
> > On Sat, 2009-02-14 at 00:30 +1100, Nick Piggin wrote:
> > > On Friday 13 February 2009 22:48:30 Peter Zijlstra wrote:
> > > > On Thu, 2009-02-12 at 17:39 -0800, Jeremy Fitzhardinge wrote:
> > > > > In general the model for lazy updates is that you're batching the
> > > > > updates in some queue somewhere, which is almost certainly a piece of
> > > > > percpu state being maintained by someone. Its therefore broken
> > > > > and/or meaningless to have the code making the updates wandering
> > > > > between cpus for the duration of the lazy updates.
> > > > >
> > > > > > If so, should we do the preempt_disable/enable within those
> > > > > > functions? Probably not worth the cost, I guess.
> > > > >
> > > > > The specific rules are that
> > > > > arch_enter_lazy_mmu_mode()/arch_leave_lazy_mmu_mode() require you to
> > > > > be holding the appropriate pte locks for the ptes you're updating, so
> > > > > preemption is naturally disabled in that case.
> > > >
> > > > Right, except on -rt where the pte lock is a mutex.
> > > >
> > > > > This all goes a bit strange with init_mm's non-requirement for taking
> > > > > pte locks. The caller has to arrange for some kind of serialization
> > > > > on updating the range in question, and that could be a mutex.
> > > > > Explicitly disabling preemption in enter_lazy_mmu_mode would make
> > > > > sense for this case, but it would be redundant for the common case of
> > > > > batched updates to usermode ptes.
> > > >
> > > > I really utterly hate how you just plonk preempt_disable() in there
> > > > unconditionally and without very clear comments on how and why.
> > >
> > > And even on mainline kernels, builds without the lazy mmu mode stuff
> > > don't need preemption disabled here either, so it is technically a
> > > regression in those cases too.
> >
> > Well, normally we'd be holding the pte lock, which on regular kernels
> > already disable preemption, as Jeremy noted. So in that respect it
> > doesn't change things too much.
>
> But not (necessarily) in the init_mm case.

Right.

> > Its just that slapping preempt_disable()s around like there's not
> > tomorrow is horridly annoying, its like using the BKL -- there's no data
> > affinity what so ever, so trying to unravel the dependencies a year
> > later when you notice its a latency concern is a massive pain in the
> > backside.
>
> Or like using memory barriers. Any of them are OK if they're properly
> commented though, I guess.

Yes, given sufficient comments and a good reason most things can be
gotten away with ;-)

> > > > I'd rather we'd fix up the init_mm to also have a pte lock.
> > >
> > > Well that wouldn't fix -rt; there would need to be a preempt_disable
> > > within arch_enter_lazy_mmu_mode(), which I think is the cleanest
> > > solution.
> >
> > Hmm, so you're saying we need to be cpu-affine for the lazy mmu stuff?
> > Otherwise a -rt would just convert the init_mm pte lock to a mutex along
> > with all other pte locks and there'd be no issue.
>
> Well I don't see any other reason why it should have to use preempt_disable.
> Not necessarily just cpu-affine, but perhaps it is using per-cpu data in
> non-trivial way so cannot get switched out either.

If the lazy mmu code relies on per-cpu data, then it should be the lazy
mmu's responsibility to ensure stuff is properly serialized. Eg. it
should do get_cpu_var() and put_cpu_var().

Those constructs can usually be converted to preemptable variants quite
easily, as it clearly shows what data needs to be protected.


\
 
 \ /
  Last update: 2009-02-13 15:39    [W:0.039 / U:47.024 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site