Re: [PATCH 1/9] readahead: introduce PG_readahead

On Sat, 19 May 2007 20:30:31 +0800 Fengguang Wu [off-list ref] wrote:

On Fri, May 18, 2007 at 11:28:24PM -0700, Andrew Morton wrote:

quoted

On Thu, 17 May 2007 06:47:53 +0800 Fengguang Wu [off-list ref] wrote:

quoted

Introduce a new page flag: PG_readahead.

Is there any way in which we can avoid adding a new page flag?

We have the advantage that if the kernel very occasionally gets the wrong
result for PageReadahead(page), nothing particularly bad will happen, so we
can do racy things.

quoted

From a quick peek, it appears that PG_readahead is only ever set against

non-uptodate pages.  If true we could perhaps exploit that: say,
PageReadahead(page) == PG_referenced && !PG_uptodate?

PG_uptodate will flip to 1 before the reader touches the page :(

OK.

However, it may be possible to share the same bit with PG_reclaim or PG_booked.
Which one would be preferred?

I'd like PG_booked to go away too - I don't think we've put that under the
microscope yet.  If it remains then its scope will be "defined by the
filesystem", so readahead shouldn't use it.  PG_reclaim belongs to core VFS
so that's much better.

Let's not do anything ugly, slow or silly in there, but please do take a
look, see if there is an opportunity here.