Re: block: DMA alignment of IO buffer allocated from slab

[Andrey Ryabinin]

On 09/24/2018 06:08 PM, Bart Van Assche wrote:
> On Mon, 2018-09-24 at 17:43 +0300, Andrey Ryabinin wrote:
>>
>> On 09/24/2018 05:19 PM, Bart Van Assche wrote:
>>> On 9/24/18 2:46 AM, Andrey Ryabinin wrote:
>>>> On 09/24/2018 01:42 AM, Ming Lei wrote:
>>>>> On Fri, Sep 21, 2018 at 03:04:18PM +0200, Vitaly Kuznetsov wrote:
>>>>>> Christoph Hellwig <hch@xxxxxx> writes:
>>>>>>
>>>>>>> On Wed, Sep 19, 2018 at 05:15:43PM +0800, Ming Lei wrote:
>>>>>>>> 1) does kmalloc-N slab guarantee to return N-byte aligned buffer?  If
>>>>>>>> yes, is it a stable rule?
>>>>>>>
>>>>>>> This is the assumption in a lot of the kernel, so I think if somethings
>>>>>>> breaks this we are in a lot of pain.
>>>>
>>>> This assumption is not correct. And it's not correct at least from the beginning of the
>>>> git era, which is even before SLUB allocator appeared. With CONFIG_DEBUG_SLAB=y
>>>> the same as with CONFIG_SLUB_DEBUG_ON=y kmalloc return 'unaligned' objects.
>>>> The guaranteed arch-and-config-independent alignment of kmalloc() result is "sizeof(void*)".
>>
>> Correction sizeof(unsigned long long), so 8-byte alignment guarantee.
>>
>>>>
>>>> If objects has higher alignment requirement, the could be allocated via specifically created kmem_cache.
>>>
>>> Hello Andrey,
>>>
>>> The above confuses me. Can you explain to me why the following comment is present in include/linux/slab.h?
>>>
>>> /*
>>>  * kmalloc and friends return ARCH_KMALLOC_MINALIGN aligned
>>>  * pointers. kmem_cache_alloc and friends return ARCH_SLAB_MINALIGN
>>>  * aligned pointers.
>>>  */
>>>
>>
>> ARCH_KMALLOC_MINALIGN - guaranteed alignment of the kmalloc() result.
>> ARCH_SLAB_MINALIGN - guaranteed alignment of kmem_cache_alloc() result.
>>
>> If the 'align' argument passed into kmem_cache_create() is bigger than ARCH_SLAB_MINALIGN
>> than kmem_cache_alloc() from that cache should return 'align'-aligned pointers.
> 
> Hello Andrey,
> 
> Do you realize that that comment from <linux/slab.h> contradicts what you
> wrote about kmalloc() if ARCH_KMALLOC_MINALIGN > sizeof(unsigned long long)?
> 

No, I don't see the contradiction. I said that arch-and-config-independent alignment is 8-bytes (at first I said that sizeof(void*), but corrected later)
If some arch defines "ARCH_KMALLOC_MINALIGN > sizeof(unsigned long long)" than on that arch kmalloc() guarantee to return > 8 bytes
aligned pointer, but that become arch-dependent alignment.

I just realized that my phrase "kmalloc return 'unaligned' objects" is very confusing.
By 'unaligned' objects, I meant that kmalloc-N doesn't return N-bytes aligned object.
ARCH_KMALLOC_MINALIGN alignment is always guaranteed.

> Additionally, shouldn't CONFIG_DEBUG_SLAB=y and CONFIG_SLUB_DEBUG_ON=y
> provide the same guarantees as with debugging disabled, namely that kmalloc()
> buffers are aligned on ARCH_KMALLOC_MINALIGN boundaries? Since buffers
> allocated with kmalloc() are often used for DMA, how otherwise is DMA assumed
> to work?
> 

Yes, with CONFIG_DEBUG_SLAB=y, CONFIG_SLUB_DEBUG_ON=y kmalloc() guarantees that result is aligned on ARCH_KMALLOC_MINALIGN boundary.


> Thanks,
> 
> Bart.
>