Re: [PATCH] Radix-tree pagecache for 2.5

[Arjan van de Ven]

Rik van Riel wrote:
> 
> On Fri, 1 Feb 2002, Alan Cox wrote:
> 
> > > the prefetch engine will have to restart every 4kB, so we would want to
> > > use 16MB pages if possible.
> > >
> > > How would we allocate large pages? Would there be a boot option to
> > > reserve an area of RAM for large pages only?
> >
> > If you have an rmap all you have to do is to avoid smearing kernel objects
> > around lots of 16Mb page sets. If need be you can then get a 16Mb page
> > back just by shuffling user pages.
> >
> > It does make the performance analysis much more interesting though.
> 
> Actually, I suspect that for most workloads the amount of
> large pages vs. the amount of small pages should be fairly
> static.
> 
> In that case we can just reclaim an old large page from
> the inactive_clean list whenever we want to allocate a new
> one.
> 
> As for not putting kernel objects everywhere, this comes
> naturally with HIGHMEM ;)

well except when you start doing pagetables high, as Andrea is doing
(and it makes tons of sense to do that)