Re: upcoming kerneloops.org item: get_page_from_freelist

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

On Wed, 24 Jun 2009 11:42:33 -0700 (PDT)
Linus Torvalds <torvalds@linux-foundation.org> wrote:

> So I'd suggest just doing this..
> 
> 			Linus
> ---
>  mm/page_alloc.c |    4 ++--
>  1 files changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index aecc9cd..5d714f8 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -1153,10 +1153,10 @@ again:
>  			 * properly detect and handle allocation failures.
>  			 *
>  			 * We most definitely don't want callers attempting to
> -			 * allocate greater than single-page units with
> +			 * allocate greater than order-1 page units with
>  			 * __GFP_NOFAIL.
>  			 */
> -			WARN_ON_ONCE(order > 0);
> +			WARN_ON_ONCE(order > 1);
>  		}
>  		spin_lock_irqsave(&zone->lock, flags);
>  		page = __rmqueue(zone, order, migratetype);

Well.  What is our overall objective here?

My original patch was motiviated by the horror at discovering that
we're using this thing (which was _never_ supposed to have new users)
for order>0 allocations.  We've gone backwards.

Ideally, we'd fix all callers to handle allocation faliures then remove
__GFP_NOFAIL.  But I don't know how to fix JBD.

So perhaps we should just revert that WARN_ON altogether, and I can go
on a little grep-empowered rampage, see if we can remove some of these
callsites.

It's not a huge problem, btw.  I don't think I've ever seen a report of
a machine getting stuck in a __GFP_NOFAIL allocation attempt.  But from
a design perspective it's Just Wrong.