[dm-devel] Regression: wrong DIO alignment check with dm-crypt

[Keith Busch]

[Cc'ing Dmitrii, who also reported the same issue]

On Tue, Nov 01, 2022 at 08:11:15PM -0700, Eric Biggers wrote:
> Hi,
> 
> I happened to notice the following QEMU bug report:
> 
> https://gitlab.com/qemu-project/qemu/-/issues/1290
> 
> I believe it's a regression from the following kernel commit:
> 
>     commit b1a000d3b8ec582da64bb644be633e5a0beffcbf
>     Author: Keith Busch <kbusch at kernel.org>
>     Date:   Fri Jun 10 12:58:29 2022 -0700
> 
>         block: relax direct io memory alignment
> 
> The bug is that if a dm-crypt device is set up with a crypto sector size (and
> thus also a logical_block_size) of 4096, then the block layer now lets through
> direct I/O requests to dm-crypt when the user buffer has only 512-byte
> alignment, instead of the 4096-bytes expected by dm-crypt in that case.  This is
> because the dma_alignment of the device-mapper device is only 511 bytes.
> 
> This has two effects in this case:
> 
>     - The error code for DIO with a misaligned buffer is now EIO, instead of
>       EINVAL as expected and documented.  This is because the I/O reaches
>       dm-crypt instead of being rejected by the block layer.
> 
>     - STATX_DIOALIGN reports 512 bytes for stx_dio_mem_align, instead of the
>       correct value of 4096.  (Technically not a regression since STATX_DIOALIGN
>       is new in v6.1, but still a bug.)
> 
> Any thoughts on what the correct fix is here?  Maybe the device-mapper layer
> needs to set dma_alignment correctly?  Or maybe the block layer needs to set it
> to 'logical_block_size - 1' by default?

I think the quick fix is to have the device mapper override the default
queue stacking limits to align the dma mask to logical block size.

Does dm-crypt strictly require memory alignment to match the block size,
or is this just the way the current software works that we can change?
It may take me a moment to get to the bottem of that, but after a quick
glance, it looks like dm-crypt will work fine with the 512 offsets if we
happen to have a physically contiguous multi-page bvec, and will fail
otherwise due to a predetermined set of sg segments (looking at
crypt_convert_block_aead()).