Re: [PATCH] crypto: gcm - fix cacheline sharing

[Ard Biesheuvel]

On Thu, 30 May 2019 at 16:27, Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On Thu, May 30, 2019 at 03:55:07PM +0200, Ard Biesheuvel wrote:
> >
> > > Would this work?
>
> I see.  You need to preserve the original IV.
>
> > > diff --git a/drivers/crypto/caam/caamalg.c b/drivers/crypto/caam/caamalg.c
> > > index c0ece44f303b..2ef2f76a3cb8 100644
> > > --- a/drivers/crypto/caam/caamalg.c
> > > +++ b/drivers/crypto/caam/caamalg.c
> > > @@ -1832,22 +1832,25 @@ static int skcipher_decrypt(struct
> > > skcipher_request *req)
> > >         struct caam_ctx *ctx = crypto_skcipher_ctx(skcipher);
> > >         int ivsize = crypto_skcipher_ivsize(skcipher);
> > >         struct device *jrdev = ctx->jrdev;
> > > +       u8 out_iv[AES_BLOCK_SIZE];
> > >         u32 *desc;
> > >         int ret = 0;
> > >
> > > -       /* allocate extended descriptor */
> > > -       edesc = skcipher_edesc_alloc(req, DESC_JOB_IO_LEN * CAAM_CMD_SZ);
> > > -       if (IS_ERR(edesc))
> > > -               return PTR_ERR(edesc);
> > > -
> > >         /*
> > >          * The crypto API expects us to set the IV (req->iv) to the last
> > >          * ciphertext block.
> > >          */
> > >         if (ivsize)
> > > -               scatterwalk_map_and_copy(req->iv, req->src, req->cryptlen -
> > > +               scatterwalk_map_and_copy(out_iv, req->src, req->cryptlen -
> > >                                          ivsize, ivsize, 0);
> > >
> > > +       /* allocate extended descriptor */
> > > +       edesc = skcipher_edesc_alloc(req, DESC_JOB_IO_LEN * CAAM_CMD_SZ);
> > > +       if (IS_ERR(edesc))
> > > +               return PTR_ERR(edesc);
> > > +
> > > +       memcpy(req->iv, out_iv, ivsize);
> > > +
> > >         /* Create and submit job descriptor*/
> > >         init_skcipher_job(req, edesc, false);
> > >         desc = edesc->hw_desc;
> >
> > Umm never mind
> >
> > /me hides in shame
>
> So why doesn't this work?
>

Because the memcpy() occurs while the buffer is mapped for DMA, so it
suffers from the exact same problem.