lkml.org 
[lkml]   [2009]   [Jun]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/3] Properly account for the number of page cache pages zone_reclaim() can reclaim
On Tue, Jun 09, 2009 at 04:27:29PM +0800, Mel Gorman wrote:
> On Tue, Jun 09, 2009 at 10:25:49AM +0800, Wu Fengguang wrote:
> > On Mon, Jun 08, 2009 at 09:01:29PM +0800, Mel Gorman wrote:
> > > On NUMA machines, the administrator can configure zone_relcaim_mode that
> > > is a more targetted form of direct reclaim. On machines with large NUMA
> > > distances for example, a zone_reclaim_mode defaults to 1 meaning that clean
> > > unmapped pages will be reclaimed if the zone watermarks are not being met.
> > >
> > > There is a heuristic that determines if the scan is worthwhile but the
> > > problem is that the heuristic is not being properly applied and is basically
> > > assuming zone_reclaim_mode is 1 if it is enabled.
> > >
> > > This patch makes zone_reclaim() makes a better attempt at working out how
> > > many pages it might be able to reclaim given the current reclaim_mode. If it
> > > cannot clean pages, then NR_FILE_DIRTY number of pages are not candidates. If
> > > it cannot swap, then NR_FILE_MAPPED are not. This indirectly addresses tmpfs
> > > as those pages tend to be dirty as they are not cleaned by pdflush or sync.
> >
> > No, tmpfs pages are not accounted in NR_FILE_DIRTY because of the
> > BDI_CAP_NO_ACCT_AND_WRITEBACK bits.
> >
>
> Ok, that explains why the dirty page count was not as high as I was
> expecting. Thanks.
>
> > > The ideal would be that the number of tmpfs pages would also be known
> > > and account for like NR_FILE_MAPPED as swap is required to discard them.
> > > A means of working this out quickly was not obvious but a comment is added
> > > noting the problem.
> >
> > I'd rather prefer it be accounted separately than to muck up NR_FILE_MAPPED :)
> >
>
> Maybe I used a poor choice of words. What I meant was that the ideal would
> be we had a separate count for tmpfs pages. As tmpfs pages and mapped pages
> both have to be unmapped and potentially, they are "like" each other with
> respect to the zone_reclaim_mode and how it behaves. We would end up
> with something like
>
> pagecache_reclaimable -= zone_page_state(zone, NR_FILE_MAPPED);
> pagecache_reclaimable -= zone_page_state(zone, NR_FILE_TMPFS);

OK. But tmpfs pages may be mapped, so there will be double counting.
We must at least make sure pagecache_reclaimable won't get underflowed.
(Or make another LRU list for tmpfs pages?)

> > > + int pagecache_reclaimable;
> > > +
> > > + /*
> > > + * Work out how many page cache pages we can reclaim in this mode.
> > > + *
> > > + * NOTE: Ideally, tmpfs pages would be accounted as if they were
> > > + * NR_FILE_MAPPED as swap is required to discard those
> > > + * pages even when they are clean. However, there is no
> > > + * way of quickly identifying the number of tmpfs pages
> > > + */
> >
> > So can you remove the note on NR_FILE_MAPPED?
> >
>
> Why would I remove the note? I can alter the wording but the intention is
> to show we cannot count the number of tmpfs pages quickly and it would be
> nice if we could. Maybe this is clearer?
>
> Note: Ideally tmpfs pages would be accounted for as NR_FILE_TMPFS or
> similar and treated similar to NR_FILE_MAPPED as both require
> unmapping from page tables and potentially swap to reclaim.
> However, no such counter exists.

That's better. Thanks.

> > > + pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
> > > + if (!(zone_reclaim_mode & RECLAIM_WRITE))
> > > + pagecache_reclaimable -= zone_page_state(zone, NR_FILE_DIRTY);
> >
> > > + if (!(zone_reclaim_mode & RECLAIM_SWAP))
> > > + pagecache_reclaimable -= zone_page_state(zone, NR_FILE_MAPPED);
> >
> > So the "if" can be removed because NR_FILE_MAPPED is not related to swapping?
> >
>
> It's partially related with respect to what zone_reclaim() is doing.
> Once something is mapped, we need RECLAIM_SWAP set on the
> zone_reclaim_mode to do anything useful with them.

You are referring to mapped anonymous/tmpfs pages? But I mean
NR_FILE_MAPPED pages won't goto swap when unmapped.

Thanks,
Fengguang

> > > /*
> > > * Zone reclaim reclaims unmapped file backed pages and
> > > @@ -2391,8 +2406,7 @@ int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
> > > * if less than a specified percentage of the zone is used by
> > > * unmapped file backed pages.
> > > */
> > > - if (zone_page_state(zone, NR_FILE_PAGES) -
> > > - zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
> > > + if (pagecache_reclaimable <= zone->min_unmapped_pages
> > > && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
> > > <= zone->min_slab_pages)
> > > return 0;
> > > --
> > > 1.5.6.5
> >
>
> --
> Mel Gorman
> Part-time Phd Student Linux Technology Center
> University of Limerick IBM Dublin Software Lab


\
 
 \ /
  Last update: 2009-06-09 10:49    [W:0.052 / U:2.000 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site