Re: [GIT PULL] SLAB changes for v3.10

[Glauber Costa <glommer@parallels.com>]

On 05/08/2013 04:26 PM, Pekka Enberg wrote:
> On Wed, May 8, 2013 at 2:58 PM, Glauber Costa <glommer@parallels.com> wrote:
>> My first guess is that it hit a NULL cache. Being a NULL pointer
>> dereference, the thing among all that has the biggest chances of being
>> NULL and accessed unconditionally is the cache pointer itself.
>>
>> Due to the size being too big. But if that were the case, he would have
>> hit the WARN_ON recently introduced:
>>
>>                 if (WARN_ON_ONCE(size > KMALLOC_MAX_SIZE))
>>                         return NULL;
>>
>>
>> Is this WARN hit ?
> 
> I doubt it:
> 
> [    0.000000] r7 : 00000000  r6 : 600001d3  r5 : 00000000  r4 : 00008000
> [    0.000000] r3 : 00000050  r2 : c06ec000  r1 : c06f77c8  r0 : c00eda9c
> 
> [    0.000000] [<c00edab4>] (kmem_cache_alloc_trace+0x50/0x178) from
> [<c0086958>] (alloc_desc+0x24/0xb4)
> 
> It's the kzalloc_node() in kernel/irq/irqdesc.c::alloc_desc() and
> AFAICT based on r4 it's a 32 KB allocation. It's more likely that
> KMALLOC_SHIFT_HIGH is less than 25 and because kmalloc_index() doesn't
> respect it, we return a pointer to an uninitialized kmalloc cache.
> 

Exactly, but then the index is calculated from the size. If we are
allocating with a size that would lead to an invalid index, we should
WARN. If this is not happening, that WARN is really really badly placed.