Re: [RFC][PATCH 2/2] readahead: avoid page-by-page reads on POSIX_FADV_RANDOM

[Wu Fengguang <fengguang.wu@intel.com>]

On Thu, Dec 31, 2009 at 02:02:38AM +0800, Andi Kleen wrote:
> Wu Fengguang <fengguang.wu@intel.com> writes:
> >   * the ra fields can be accessed concurrently in a racy way.
> > --- linux.orig/mm/fadvise.c	2009-12-30 13:02:03.000000000 +0800
> > +++ linux/mm/fadvise.c	2009-12-30 13:23:05.000000000 +0800
> > @@ -77,12 +77,14 @@ SYSCALL_DEFINE(fadvise64_64)(int fd, lof
> >  	switch (advice) {
> >  	case POSIX_FADV_NORMAL:
> >  		file->f_ra.ra_pages = bdi->ra_pages;
> > +		file->f_ra.flags &= ~RA_FLAG_RANDOM;
> >  		break;
> >  	case POSIX_FADV_RANDOM:
> > -		file->f_ra.ra_pages = 0;
> > +		file->f_ra.flags |= RA_FLAG_RANDOM;
> 
> What prevents this from racing with a parallel readahead
> state modification, losing the bits?

Oh I pretended that the problem don't exist..

To be serious, the race only exist inside a mutithread application,
where one single fd is shared between two threads, one is doing
fadvise, another doing readahead.

A sane application won't do fadvise(POSIX_FADV_RANDOM) while active
reads are going one concurrently: this leads to indeterminate behavior
by itself -- from which request the random hint takes effect?

fadvise() shall always be in the same streamline with all reads.

In real workloads, 1% applications may do POSIX_FADV_RANDOM, among
which 1% applications may be broken. And if the race does happen, the
impact is very small. So I choose to just ignore the race and use
non-atomic operations..

Thanks,
Fengguang