Re: RFC: get_user_pages_locked|unlocked to leverage VM_FAULT_RETRY

[Andrea Arcangeli]

On Thu, Oct 02, 2014 at 02:56:38PM +0200, Peter Zijlstra wrote:
> On Thu, Oct 02, 2014 at 02:50:52PM +0200, Peter Zijlstra wrote:
> > On Thu, Oct 02, 2014 at 02:31:17PM +0200, Andrea Arcangeli wrote:
> > > On Wed, Oct 01, 2014 at 05:36:11PM +0200, Peter Zijlstra wrote:
> > > > For all these and the other _fast() users, is there an actual limit to
> > > > the nr_pages passed in? Because we used to have the 64 pages limit from
> > > > DIO, but without that we get rather long IRQ-off latencies.
> > > 
> > > Ok, I would tend to think this is an issue to solve in gup_fast
> > > implementation, I wouldn't blame or modify the callers for it.
> > > 
> > > I don't think there's anything that prevents gup_fast to enable irqs
> > > after certain number of pages have been taken, nop; and disable the
> > > irqs again.
> > > 
> > 
> > Agreed, I once upon a time had a patch set converting the 2 (x86 and
> > powerpc) gup_fast implementations at the time, but somehow that never
> > got anywhere.
> > 
> > Just saying we should probably do that before we add callers with
> > unlimited nr_pages.
> 
> https://lkml.org/lkml/2009/6/24/457
> 
> Clearly there's more work these days. Many more archs grew a gup.c

What about this? The alternative is that I do s/gup_fast/gup_unlocked/
to still retain the mmap_sem scalability benefit. It'd be still better
than the current plain gup() (and it would be equivalent for
userfaultfd point of view).

Or if the below is ok, should I modify all other archs too or are the
respective maintainers going to fix it themself? For example the arm*
gup_fast is a moving target in development on linux-mm right now and I
should only patch the gup_rcu version that didn't hit upstream yet. In
fact after that gup_rcu merge, supposedly the powerpc and sparc
gup_fast can be dropped from arch/* entirely and they can use the
generic version (otherwise having the arm gup_fast in mm/ instead of
arch/ would be a mistake). Right now, I wouldn't touch at least
arm/sparc/powerpc until the gup_rcu hit upstream as those are all
about to disappear.

Thanks,
Andrea