lkml.org 
[lkml]   [2014]   [Nov]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 2/3] kernel: irq: use a kmem_cache for allocating struct irq_desc
On 11/20/2014 02:52 AM, David Rientjes wrote:
> On Mon, 10 Nov 2014, Andrey Ryabinin wrote:
>
>> After enabling alignment checks in UBSan I've noticed a lot of
>> reports like this:
>>
>> UBSan: Undefined behaviour in ../kernel/irq/chip.c:195:14
>> member access within misaligned address ffff88003e80d6f8
>> for type 'struct irq_desc' which requires 16 byte alignment
>>
>> struct irq_desc declared with ____cacheline_internodealigned_in_smp
>> attribute. However in some cases it allocated dynamically via kmalloc().
>> In general case kmalloc() guaranties only sizeof(void *) alignment.
>> We should use a separate slab cache to make struct irq_desc
>> properly aligned on SMP configuration.
>>
>> This also could slightly reduce memory usage on some configurations.
>> E.g. in my setup sizeof(struct irq_desc) == 320. Which means that
>> kmalloc-512 will be used for allocating irg_desc via kmalloc().
>> In that case using separate slab cache will save us 192 bytes per
>> each irq_desc.
>>
>> Note: UBSan reports says that 'struct irq_desc' requires 16 byte alignment.
>> It's wrong, in my setup it should be 64 bytes. This looks like a gcc bug,
>> but it doesn't change the fact that irq_desc is misaligned.
>>
>> Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com>
>
> I think this is just fine, I would just prefer that you do the memset()

I'd rather do kmem_cache_alloc_node(irq_desc_cachep, gfp | __GFP_ZERO, node)
instead of memset.

> explicitly rather than introduce the new slab function for such a
> specialized purpose (unless there's other examples in the kernel where
> this would be useful).
>

I've counted 7 places where kmem_cache_alloc_node(..., gfp | __GFP_ZERO, ...); called.



\
 
 \ /
  Last update: 2014-11-20 10:21    [W:0.778 / U:0.176 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site