Re: [PATCH next/mmotm] slub: partly fix freeze in __slab_free

[Pekka Enberg <penberg@kernel.org>]

On 7/11/11 10:39 PM, Eric Dumazet wrote:
> Le lundi 11 juillet 2011 à 11:58 -0700, Hugh Dickins a écrit :
>> My load tests on PowerPC freeze within minutes in __slab_free().
>> I happened to try PPC first, didn't try without this fix on x86.
>>
>> It looks as if the author was interrupted while devising the new
>> cmpxchg_double_slab() version of __slab_free(): its decision to
>> spin_lock_irqsave() depends on several uninitialized fields,
>> and fixing that (by copying page to new) mostly fixes it.
>>
>> But I didn't think about it very much, and this may well not be what
>> the author intends; and I have seen a couple of much rarer freezes
>> in __slab_free() on PPC (not yet on x86) even after applying this.
>>
>> Signed-off-by: Hugh Dickins<hughd@google.com>
>> ---
>>   mm/slub.c |    1 +
>>   1 file changed, 1 insertion(+)
>>
>> --- mmotm/mm/slub.c	2011-07-08 18:59:44.135443127 -0700
>> +++ linux/mm/slub.c	2011-07-10 05:07:08.000000000 -0700
>> @@ -2217,6 +2217,7 @@ static void __slab_free(struct kmem_cach
>>   		return;
>>
>>   	do {
>> +		new = *page;
>>   		prior = page->freelist;
>>   		counters = page->counters;
>>   		set_freepointer(s, object, prior);
>> --
> I suspect you hit the bug on 32bit arch ?
>
> What about following patch instead ?
>
> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> index 3d76a43..1351d28 100644
> --- a/include/linux/mm_types.h
> +++ b/include/linux/mm_types.h
> @@ -62,7 +62,7 @@ struct page {
>   		struct {			/* SLUB cmpxchg_double area */
>   			void *freelist;
>   			union {
> -				unsigned long counters;
> +				u64	counters;
>   				struct {
>   					unsigned inuse:16;
>   					unsigned objects:15;
>

Christoph?