Re: 2.6.22-stable causes oomkiller to be invoked

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

On Sat, 15 Dec 2007 09:22:00 +0530 Dhaval Giani <dhaval@linux.vnet.ibm.com> wrote:

> > Is it really the case that the bug only turns up when you run tests like
> > 
> > 	while echo; do cat /sys/kernel/kexec_crash_loaded; done
> > and
> > 	while echo; do cat /sys/kernel/uevent_seqnum ; done;
> > 
> > or will any fork-intensive workload also do it?  Say,
> > 
> > 	while echo ; do true ; done
> > 
> 
> This does not leak, but having a simple text file and reading it in a
> loop causes it.

hm.

> > ?
> > 
> > Another interesting factoid here is that after the oomkilling you slabinfo has
> > 
> > mm_struct             38     98    584    7    1 : tunables   32   16    8 : slabdata     14     14      0 : globalstat    2781    196    49   31 				   0    1    0    0    0 : cpustat 368800  11864 368920  11721
> > 
> > so we aren't leaking mm_structs.  In fact we aren't leaking anything from
> > slab.   But we are leaking pgds.
> > 
> > iirc the most recent change we've made in the pgd_t area is the quicklist
> > management which went into 2.6.22-rc1.  You say the bug was present in
> > 2.6.22.  Can you test 2.6.21?  
> 
> Nope, leak is not present in 2.6.21.7

Could you try this debug patch please?

It might need some fiddling to get useful output.  Basic idea is to see if
we are failing to empty the quicklists.

--- a/include/linux/quicklist.h~a
+++ a/include/linux/quicklist.h
@@ -69,6 +69,8 @@ static inline void __quicklist_free(int 
 	*(void **)p = q->page;
 	q->page = p;
 	q->nr_pages++;
+	if (q->nr_pages && !(q->nr_pages % 1000))
+		printk("eek: %d\n", q->nr_pages);
 	put_cpu_var(quicklist);
 }
 
_

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>