Re: [PATCH 0/8] Avoiding fragmentation with subzone groupings v25

[Mel Gorman]

On Fri, 8 Sep 2006, Peter Zijlstra wrote:

> On Fri, 2006-09-08 at 09:36 +0100, Mel Gorman wrote:
> > On Thu, 7 Sep 2006, Andrew Morton wrote:
> >
> > > On Thu,  7 Sep 2006 20:03:42 +0100 (IST)
> > > Mel Gorman <mel@xxxxxxxxx> wrote:
> > >
> > > > When a page is allocated, the page-flags
> > > > are updated with a value indicating it's type of reclaimability so that it
> > > > is placed on the correct list on free.
> > >
> > > We're getting awful tight on page-flags.
> >
> > Yeah, I know :(
> >
> > > Would it be possible to avoid adding the flag?  Say, have a per-zone bitmap
> > > of size (zone->present_pages/(1<<MAX_ORDER)) bits, then do a lookup in
> > > there to work out whether a particular page is within a MAX_ORDER clump of
> > > easy-reclaimable pages?
> >
> > An early version of the patches created such a bitmap and it was heavily
> > resisted for two reasons. It put more pressure on the cache and it needed
> > to be resized during hot-add and hot-remove. It was the latter issue
> > people had more problems with. However, I can reimplement it if people
> > want to take a look. As I see it currently, there are five choices that
> > could be taken to avoid using an additional pageflag
> >
> > 1. Re-use existing page flags. This is what I currently do in a later
> >     patch for the software suspend flags
> >     pros: Straight-forward implementation, appears to use no additional flags
> >     cons: When swsusp stops using the flags, anti-frag takes them right back
> >           Makes anti-frag mutually exclusive with swsusp
> >
> > 2. Create a per-zone bitmap for every MAX_ORDER block
> >     pros: Straight-forward implementation initially
> >     cons: Needs resizing during hotadd which could get complicated
> >           Bit more cache pressure
> >
> > 3. Use the low two bits of page->lru
> >     pros: Uses existing struct page field
> >     cons: It's a bit funky looking
> >
> > 4. Use the page->flags of the struct page backing the pages used
> >     for the memmap.
> >     pros: Similar to the bitmap idea except with less hotadd problems
> >     cons: Bit more cache pressure
> >
> > 5. Add an additional field page->hintsflags used for non-critical flags.
> >     There are patches out there like guest page hinting that want to
> >     consume flags but not for any vital purpose and usually for machines
> >     that have ample amounts of memory. For these features, add an
> >     additional page->hintsflags
> >     pros: Straight-forward to implement
> >     cons: Increses struct page size for some kernel features.
> >
> > I am leaning towards option 3 because it uses no additional memory but I'm
> > not sure how people feel about using pointer magic like this.
> >
> > Any opinions?
>
> If, as you stated in a previous mail, you'd like to have flags per
> MAX_ORDER block, you'd already have to suffer the extra cache pressure.
> In that case I vote for 4.

Originally, I wanted flags per MAX_ORDER block but I no longer have data 
on whether this is a good idea or not. It could turn out that we steal 
back and forth a lot when pageblock flags are used.

> Otherwise 3 sounds doable, we already hide PAGE_MAPPING_ANON in a
> pointer, so hiding flags is not new to struct page. It's just a question
> of how good the implementation will look, I hope you'll not have to
> visit all the list ops.

One way to find out for sure! I reckon I'll go off and implement options 3 
and 4 as add-on patches that avoid the use of page->flags and see what 
they look like. As you said, pointer magic in struct page is not new.

--
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/