Re: [PATCH] mm/filemap.c: unconditionally call mark_page_accessed

[Andrea Arcangeli <andrea@suse.de>]

On Thu, Mar 15, 2007 at 06:15:45PM -0500, Dave Kleikamp wrote:
> On Thu, 2007-03-15 at 23:59 +0100, Andrea Arcangeli wrote:
> > On Thu, Mar 15, 2007 at 05:44:01PM +0000, Hugh Dickins wrote:
> > > who removed the !offset condition, he should be consulted on its
> > > reintroduction.
> > 
> > the !offset check looks a pretty broken heuristic indeed, it would
> > break random I/O.
> 
> I wouldn't call it broken.  At worst, I'd say it's imperfect.  But
> that's the nature of a heuristic.  It most likely works in a huge
> majority of cases.

well, IMHO in the huge majority of cases the prev_page check isn't
necessary in the first place (and IMHO it hurts a lot more than it can
help, as demonstrated by specweb, since we'll bite on the good guys to
help the bad guys).

The only case where I can imagine the prev_page to make sense is to
handle contiguous I/O made with a small buffer, so clearly an
inefficient code in the first place. But if this guy is reading with
<PAGE_SIZE buffer there's no guarantee that he's reading f_pos aligned
either, hence the need of taking last_offset into account too so at
least it's a "perfect" heuristic that will reliably detect contiguous
I/O no matter in what shape or form you execute it, as long as it is
contiguous I/O.

Any other variation of behavior besides the autodetection of
contiguous I/O run in whatever buffer/aligned form, should be mandated
by userland through fadvise/madvise IMHO or we run into the toes of
the good guys.