Re: [PATCH 3/3] mm,migration: Remove straggling migration PTEswhen page tables are being moved after the VMA has already moved

[Mel Gorman]

Thanks to you both for looking into this. I far prefer this general approach
than cleaning up the migration PTEs as the page tables get copied. While it
might "work", it's sloppy in the same way as having migration_entry_wait()
do the cleanup was sloppy. It's far preferable to make the VMA move and
page table copy atomic with anon_vma->lock.

On Wed, Apr 28, 2010 at 04:28:38PM +0900, KAMEZAWA Hiroyuki wrote:
> On Wed, 28 Apr 2010 11:49:44 +0900
> KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> 
> > On Wed, 28 Apr 2010 04:42:27 +0200
> > Andrea Arcangeli <aarcange@xxxxxxxxxx> wrote:
>  
> > > migrate.c requires rmap to be able to find all ptes mapping a page at
> > > all times, otherwise the migration entry can be instantiated, but it
> > > can't be removed if the second rmap_walk fails to find the page.
> > > 
> > > So shift_arg_pages must run atomically with respect of rmap_walk, and
> > > it's enough to run it under the anon_vma lock to make it atomic.
> > > 
> > > And split_huge_page() will have the same requirements as migrate.c
> > > already has.
> > > 
> > > Signed-off-by: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> > 
> > Seems good.
> > Reviewed-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> > 
> > I'll test this and report if I see trouble again.
> > 
> > Unfortunately, I'll have a week of holidays (in Japan) in 4/29-5/05,
> > my office is nearly closed. So, please consider no-mail-from-me is
> > good information.
> > 
> Here is bad news. When move_page_tables() fails, "some ptes" are moved
> but others are not and....there is no rollback routine.
> 

The biggest problem is that the reverse mapping is temporarily out of
sync until do_exit gets rid of the mess, but how serious is that really?

If there is a migration entry in there, the mapcount should already be zero and
migration holds a reference to the page to prevent it going away. rmap_walk()
may then miss the migration_pte so it gets left behind. Ordinarily this
would be bad but in exec(), we cannot be faulting this page so we won't
trigger the bug in swapops. Instead, do_exit ultimately will skip over the
migration PTE doing nothing with the page but as the mapcount is still zero,
the page won't leak.

> I bet the best way to fix this mess up is 
>  - disable overlap moving of arg pages
>  - use do_mremap().
> 
> But maybe you guys want to fix this directly.
> Here is a temporal fix from me. But don't trust me..

I see the point of your patch but I'm not yet seeing why it is
necessary to back out if move_page_tables fails.

That said, both patches have a greater problem. Both of them hold a spinlock
(anon_vma->lock) while calling into the page allocator with GFP_KERNEL (to
allocate the page tables). We don't want to change that to GFP_ATOMIC so
either we need to allocate the pages in advance or special case rmap_walk()
to not walk processes that are in exec.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/