Re: poll-related "scheduling while atomic", 2.5.44-mm6

[Andrew Morton <akpm@digeo.com>]

Paolo Ciarrocchi wrote:
> 
> >> So my guess is somewhere between -mm5 and -mm6 we
> >> screwed up the atomicity count.
> >Mine too.  I'll check it out, thanks.
> 
> The same here as well
> 

This'll fix it up.  Whoever invented cut-n-paste has a lot to
answer for.


--- 25/mm/swap.c~preempt-count-fix	Tue Oct 29 22:19:54 2002
+++ 25-akpm/mm/swap.c	Tue Oct 29 22:20:16 2002
@@ -90,11 +90,12 @@ void lru_cache_add_active(struct page *p
 
 void lru_add_drain(void)
 {
-	struct pagevec *pvec = &per_cpu(lru_add_pvecs, get_cpu());
+	int cpu = get_cpu();
+	struct pagevec *pvec = &per_cpu(lru_add_pvecs, cpu);
 
 	if (pagevec_count(pvec))
 		__pagevec_lru_add(pvec);
-	pvec = &per_cpu(lru_add_active_pvecs, get_cpu());
+	pvec = &per_cpu(lru_add_active_pvecs, cpu);
 	if (pagevec_count(pvec))
 		__pagevec_lru_add_active(pvec);
 	put_cpu();

.

I had a crash while testing SMP+preempt btw.  Nasty one - took a
pagefault from userspace but do_page_fault() decided that the
fault was in-kernel or something.  It fell all the way through
to die() and, well, died.  I saw the same happen some months ago.