On Wed, Feb 23, 2022 at 05:47:25PM -0800, Eric Biggers wrote:
> On Wed, Feb 23, 2022 at 10:04:00AM +0200, Gilad Ben-Yossef wrote:
> > the drbg code was binding the same buffer to two different
> > scatter gather lists and submitting those as source and
> > destination to a crypto api operation, thus potentially
> > causing HW crypto drivers to perform overlapping DMA
> > mappings which are not aware it is the same buffer.
> >
> > This can have serious consequences of data corruption of
> > internal DRBG buffers and wrong RNG output.
> >
> > Fix this by reusing the same scatter gatther list for both
> > src and dst.
> >
> > Signed-off-by: Gilad Ben-Yossef <gilad@xxxxxxxxxxxxx>
> > Reported-by: Corentin Labbe <clabbe.montjoie@xxxxxxxxx>
> > Tested-by: Corentin Labbe <clabbe.montjoie@xxxxxxxxx>
> > Tested-on: r8a7795-salvator-x
> > Tested-on: xilinx-zc706
> > Fixes: 43490e8046b5d ("crypto: drbg - in-place cipher operation for CTR")
> > Cc: stable@xxxxxxxxxxxxxxx
>
> Where is it documented and tested that the API doesn't allow this?
> I wasn't aware of this case; it sounds perfectly allowed to me.
> There might be a lot of other users who do this, not just drbg.c.
>
Just quickly looking through the code I maintain, there is another place that
uses scatterlists like this: in fscrypt_crypt_block() in fs/crypto/crypto.c, the
source and destination can be the same. That's just the code I maintain; I'm
sure if you looked through the whole kernel you'd find a lot more.
This sounds more like a driver bug, and a case we need to add self-tests for.
- Eric
