Re: [rfc] lockless get_user_pages for dio (and more)

[Dave Kleikamp <shaggy@linux.vnet.ibm.com>]

On Wed, 2007-12-12 at 15:57 +1100, Nick Piggin wrote:
> On Tuesday 11 December 2007 08:30, Dave Kleikamp wrote:

> > Nick,
> > I've played with the fast_gup patch a bit.  I was able to find a problem
> > in follow_hugetlb_page() that Adam Litke fixed.  I'm haven't been brave
> > enough to implement it on any other architectures, but I did add  a
> > default that takes mmap_sem and calls the normal get_user_pages() if the
> > architecture doesn't define fast_gup().  I put it in linux/mm.h, for
> > lack of a better place, but it's a little kludgy since I didn't want
> > mm.h to have to include sched.h.  This patch is against 2.6.24-rc4.
> > It's not ready for inclusion yet, of course.
> 
> Hi Dave,
> 
> Thanks so much. This makes it much more a complete patch (although
> still missing the "normal page" detection).
> 
> I think I missed -- or forgot -- what was the follow_hugetlb_page
> problem?

Badari found a problem running some tests and handed it off to me to
look at.  I didn't share it publicly.  Anyway, we were finding that
fastgup was taking the slow path almost all the time with huge pages.
The problem was that follow_hugetlb_page was failing to fault on a
non-writable page when it needed a writable one.  So we'd keep seeing a
non-writable page over and over.  This is fixed in 2.6.24-rc5.

> Anyway, I am hoping that someone will one day and test if this and
> find it helps their workload, but on the other hand, if it doesn't
> help anyone then we don't have to worry about adding it to the
> kernel ;) I don't have any real setups that hammers DIO with threads.
> I'm guessing DB2 and/or Oracle does?

I'll try to get someone to run a DB2 benchmark and see what it looks
like.
-- 
David Kleikamp
IBM Linux Technology Center

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/