Re: [BUG REPORT] [mm-hotplug, aio] aio ring_pages can't be offlined

[Mel Gorman <mgorman@suse.de>]

On Thu, Nov 29, 2012 at 11:55:02PM -0800, Andrew Morton wrote:
> On Fri, 30 Nov 2012 15:01:26 +0800 Lin Feng <linfeng@cn.fujitsu.com> wrote:
> 
> > 
> > 
> > On 11/30/2012 01:57 PM, Andrew Morton wrote:
> > > On Fri, 30 Nov 2012 11:42:05 +0800 Lin Feng <linfeng@cn.fujitsu.com> wrote:
> > > 
> > >> hi Andrew,
> > >>
> > >> On 11/30/2012 07:39 AM, Andrew Morton wrote:
> > >>> Tricky.
> > >>>
> > >>> I expect the same problem would occur with pages which are under
> > >>> O_DIRECT I/O.  Obviously O_DIRECT pages won't be pinned for such long
> > >>> periods, but the durations could still be lengthy (seconds).
> > >> the offline retry timeout duration is 2 minutes, so to O_DIRECT pages 
> > >> seem maybe not a problem for the moment.
> > >>>
> > >>> Worse is a futex page, which could easily remain pinned indefinitely.
> > >>>
> > >>> The best I can think of is to make changes in or around
> > >>> get_user_pages(), to steal the pages from userspace and replace them
> > >>> with non-movable ones before pinning them.  The performance cost of
> > >>> something like this would surely be unacceptable for direct-io, but
> > >>> maybe OK for the aio ring and futexes.
> > >> thanks for your advice.
> > >> I want to limit the impact as little as possible, as mentioned above,
> > >> direct-io seems not a problem, we needn't touch them. Maybe we can 
> > >> just change the use of get_user_pages()(in or around) such as aio 
> > >> ring pages. I will try to find a way to do this.
> > > 
> > > What about futexes?
> > hi Andrew,
> > 
> > Yes, better to find an approach to solve them all.
> >  
> > But I'm worried about that if we just confine get_user_pages() to use 
> > none-movable pages, it will drain the none-movable pages soon. Because
> > there are many places using get_user_pages() such as some drivers. 
> 
> Obviously we shouldn't change get_user_pages() for all callers.
> 
> > IMHO in most cases get_user_pages() callers should release the pages soon, 
> > so pages allocated from movable zone should be OK. But I'm not sure if
> > we get such rule upon get_user_pages(). 
> > And in other cases we specify get_user_pages() to allocate pages from
> > none-movable zone. 
> > 
> > So could we add a zone-alloc flags when we call get_user_pages()?
> 
> Well, that's a fairly low-level implementation detail.  A more typical
> approach would be to add a new get_user_pages_non_movable() or such. 
> That would probably have the same signature as get_user_pages(), with
> one additional argument.  Then get_user_pages() becomes a one-line
> wrapper which passes in a particular value of that argument.
> 

That is going in the direction that all pinned pages become MIGRATE_UNMOVABLE
allocations.  That will impact THP availability by increasing the number
of MIGRATE_UNMOVABLE blocks that exist and it would hit every user --
not just those that care about ZONE_MOVABLE.

I'm likely to NAK such a patch if it's only about node hot-remove because
it's much more of a corner case than wanting to use THP.

I would prefer if get_user_pages() checked if the page it was about to
pin was in ZONE_MOVABLE and if so, migrate it at that point before it's
pinned. It'll be expensive but will guarantee ZONE_MOVABLE availability
if that's what they want. The CMA people might also want to take
advantage of this if the page happened to be in the MIGRATE_CMA
pageblock.

-- 
Mel Gorman
SUSE Labs