Re: [ck] Re: 2.6.20-ck1

[Andrew Morton <akpm@linux-foundation.org>]

On Sun, 18 Feb 2007 08:00:06 +1100 Con Kolivas <kernel@kolivas.org> wrote:

> On Sunday 18 February 2007 05:45, Chuck Ebbert wrote:
> ...
> > But the one I like, mm-filesize_dependant_lru_cache_add.patch,
> > has an on-off switch.
> >
>
> ...
>
> Do you still want this patch for mainline?... 

Don't think so.  The problems I see are:

- It's a system-wide knob.  In many situations this will do the wrong
  thing.  Controlling pagecache should be per-process.

- Its heuristics for working out when to invalidate the pagecache will be
  too much for some situations and too little for others.

- Whatever we do, there will be some applications in some situations
  which are hurt badly by changes like this: they'll do heaps of extra IO.


Generally, the penalties for getting this stuff wrong are very very high:
orders of magnitude slowdowns in the right situations.  Which I suspect
will make any system-wide knob ultimately unsuccessful.

The ideal way of getting this *right* is to change every application in the
world to get smart about using sync_page_range() and/or posix_fadvise(),
then to add a set of command-line options to each application in the world
so the user can control its pagecache handling.

Obviously that isn't practical.  But what _could_ be done is to put these
pagecache smarts into glibc's read() and write() code.  So the user can do:

	MAX_PAGECACHE=4M MAX_DIRTY_PAGECACHE=2M rsync foo bar

This will provide pagecache control for pretty much every application.  It
has limitations (fork+exec behaviour??) but will be useful.


A kernel-based solution might use new rlimits, but would not be as flexible
or successful as a libc-based one, I suspect.