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

[Bart Van Assche <bvanassche@acm.org>]

On 9/19/18 2:15 AM, Ming Lei wrote:
> Hi Guys,
> 
> Some storage controllers have DMA alignment limit, which is often set via
> blk_queue_dma_alignment(), such as 512-byte alignment for IO buffer.
> 
> Block layer now only checks if this limit is respected for buffer of
> pass-through request,
> see blk_rq_map_user_iov(), bio_map_user_iov().
> 
> The userspace buffer for direct IO is checked in dio path, see
> do_blockdev_direct_IO().
> IO buffer from page cache should be fine wrt. this limit too.
> 
> However, some file systems, such as XFS, may allocate single sector IO buffer
> via slab. Usually I guess kmalloc-512 should be fine to return
> 512-aligned buffer.
> But once KASAN or other slab debug options are enabled, looks this
> isn't true any
> more, kmalloc-512 may not return 512-aligned buffer. Then data corruption
> can be observed because the IO buffer from fs layer doesn't respect the DMA
> alignment limit any more.
> 
> Follows several related questions:
> 
> 1) does kmalloc-N slab guarantee to return N-byte aligned buffer?  If
> yes, is it a stable rule?
> 
> 2) If it is a rule for kmalloc-N slab to return N-byte aligned buffer,
> seems KASAN violates this
> rule?
> 
> 3) If slab can't guarantee to return 512-aligned buffer, how to fix
> this data corruption issue?

I don't think that (1) is correct, especially if N is not a power of 
two. In the skd driver I addressed this problem by using 
kmem_cache_create() and kmem_cache_alloc() instead of kmalloc(). 
kmem_cache_create() allows to specify the alignment explicitly.

Bart.