On Mon, 12 Aug 2019 at 16:51, Milan Broz <gmazyland@xxxxxxxxx> wrote: > > On 12/08/2019 09:50, Ard Biesheuvel wrote: > > On Mon, 12 Aug 2019 at 10:44, Milan Broz <gmazyland@xxxxxxxxx> wrote: > >> > >> On 12/08/2019 08:54, Ard Biesheuvel wrote: > >>> On Mon, 12 Aug 2019 at 09:33, Milan Broz <gmazyland@xxxxxxxxx> wrote: > >>>> Try for example > >>>> # cryptsetup luksFormat /dev/sdc -c aes-cbc-essiv:sha256 --integrity hmac-sha256 -q -i1 > >>>> > >>>> It should produce Crypto API string > >>>> authenc(hmac(sha256),essiv(cbc(aes),sha256)) > >>>> while it produces > >>>> essiv(authenc(hmac(sha256),cbc(aes)),sha256) > >>>> (and fails). > >>>> > >>> > >>> No. I don't know why it fails, but the latter is actually the correct > >>> string. The essiv template is instantiated either as a skcipher or as > >>> an aead, and it encapsulates the entire transformation. (This is > >>> necessary considering that the IV is passed via the AAD and so the > >>> ESSIV handling needs to touch that as well) > >> > >> Hm. Constructing these strings seems to be more confusing than dmcrypt mode combinations :-) > >> > >> But you are right, I actually tried the former string (authenc(hmac(sha256),essiv(cbc(aes),sha256))) > >> and it worked, but I guess the authenticated IV (AAD) was actually the input to IV (plain sector number) > >> not the output of ESSIV? Do I understand it correctly now? > >> > > > > Indeed. The former string instantiates the skcipher version of the > > ESSIV template, and so the AAD handling is omitted, and we end up > > using the plain IV in the authentication rather than the encrypted IV. > > > > So when using the latter string, does it produce any error messages > > when it fails? > > The error is > table: 253:1: crypt: Error decoding and setting key > > and it is failing in crypt_setkey() int this crypto_aead_setkey(); > > And it is because it now wrongly calculates MAC key length. > (We have two keys here - one for length-preserving CBC-ESSIV encryption > and one for HMAC.) > > This super-ugly hotfix helps here... I guess it can be done better :-) > Weird. It did work fine before, but now that I have dropped the 'md: dm-crypt: infer ESSIV block cipher from cipher string directly' patch, we are probably taking a different code path and hitting this error. I'll try to fix this cleanly. Thanks for doing the diagnosis. > diff --git a/drivers/md/dm-crypt.c b/drivers/md/dm-crypt.c > index e9a0093c88ee..7b06d975a2e1 100644 > --- a/drivers/md/dm-crypt.c > +++ b/drivers/md/dm-crypt.c > @@ -2342,6 +2342,9 @@ static int crypt_ctr_auth_cipher(struct crypt_config *cc, char *cipher_api) > char *start, *end, *mac_alg = NULL; > struct crypto_ahash *mac; > > + if (strstarts(cipher_api, "essiv(authenc(")) > + cipher_api += strlen("essiv("); > + > if (!strstarts(cipher_api, "authenc(")) > return 0; > > Milan