[PATCH] lib/crypto/poly1305: Fix arm64's poly1305_blocks_arch()

[Eric Biggers]

On Mon, Jun 16, 2025 at 11:27:09AM +0800, Herbert Xu wrote:
> Eric Biggers <ebiggers at kernel.org> wrote:
> > From: Eric Biggers <ebiggers at google.com>
> > 
> > For some reason arm64's Poly1305 code got changed to ignore the padbit
> > argument.  As a result, the output is incorrect when the message length
> > is not a multiple of 16 (which is not reached with the standard
> > ChaCha20Poly1305, but bcachefs could reach this).  Fix this.
> 
> Sorry, it was a cut-n-paste error since I copy the code from
> the update function where the padbit is always 1.
> 
> > diff --git a/arch/arm64/lib/crypto/poly1305-glue.c b/arch/arm64/lib/crypto/poly1305-glue.c
> > index 6a661cf048213..c9a74766785bd 100644
> > --- a/arch/arm64/lib/crypto/poly1305-glue.c
> > +++ b/arch/arm64/lib/crypto/poly1305-glue.c
> > @@ -36,18 +36,18 @@ void poly1305_blocks_arch(struct poly1305_block_state *state, const u8 *src,
> >        if (static_branch_likely(&have_neon)) {
> >                do {
> >                        unsigned int todo = min_t(unsigned int, len, SZ_4K);
> > 
> >                        kernel_neon_begin();
> > -                       poly1305_blocks_neon(state, src, todo, 1);
> > +                       poly1305_blocks_neon(state, src, todo, padbit);
> 
> This would do the wrong thing if someone ever tried to pad a
> message more than 4K and called the block function with padbit == 0.
> Fortunately it can't happen today as there is no digest interface
> to poly1305.

The final partial block is (and needs to be) processed with its own call to
poly1305_blocks().

> Looking around it seems that this pattern is replicated across
> all of our poly1305 implementations so it isn't a big deal.
> 
> I presume you will be picking this up via the lib/crypto tree?

Yes.

- Eric