On Fri, 21 Jun 2019 at 09:06, Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote: > > On Fri, 21 Jun 2019 at 09:01, Milan Broz <gmazyland@xxxxxxxxx> wrote: > > > > On 20/06/2019 15:52, Ard Biesheuvel wrote: > > >>>> Does this include configurations that combine authenc with essiv? > > >>> > > >>> Hm, seems that we are missing these in luks2-integrity-test. I'll add them there. > > >>> > > >>> I also used this older test > > >>> https://gitlab.com/omos/dm-crypt-test-scripts/blob/master/root/test_dmintegrity.sh > > >>> > > >>> (just aes-gcm-random need to be commented out, we never supported this format, it was > > >>> written for some devel version) > > >>> > > >>> But seems ESSIV is there tested only without AEAD composition... > > >>> > > >>> So yes, this AEAD part need more testing. > > >> > > >> And unfortunately it does not work - it returns EIO on sectors where it should not be data corruption. > > >> > > >> I added few lines with length-preserving mode with ESSIV + AEAD, please could you run luks2-integrity-test > > >> in cryptsetup upstream? > > >> > > >> This patch adds the tests: > > >> https://gitlab.com/cryptsetup/cryptsetup/commit/4c74ff5e5ae328cb61b44bf99f98d08ffee3366a > > >> > > >> It is ok on mainline kernel, fails with the patchset: > > >> > > >> # ./luks2-integrity-test > > >> [aes-cbc-essiv:sha256:hmac-sha256:128:512][FORMAT][ACTIVATE]sha256sum: /dev/mapper/dmi_test: Input/output error > > >> [FAIL] > > >> Expecting ee501705a084cd0ab6f4a28014bcf62b8bfa3434de00b82743c50b3abf06232c got . > > >> > > >> FAILED backtrace: > > >> 77 ./luks2-integrity-test > > >> 112 intformat ./luks2-integrity-test > > >> 127 main ./luks2-integrity-test > > >> > > > > > > OK, I will investigate. > > > > > > I did my testing in a VM using a volume that was created using a > > > distro kernel, and mounted and used it using a kernel with these > > > changes applied. > > > > > > Likewise, if I take a working key.img and mode-test.img, i can mount > > > it and use it on the system running these patches. > > > > > > I noticed that this test uses algif_skcipher not algif_aead when it > > > formats the volume, and so I wonder if the way userland creates the > > > image is affected by this? > > > > Not sure if I understand the question, but I do not think userspace even touch data area here > > (except direct-io wiping after the format, but it does not read it back). > > > > It only encrypts keyslots - and here we cannot use AEAD (in fact it is already > > authenticated by a LUKS digest). > > > > So if the data area uses AEAD (or composition of length-preserving mode and > > some authentication tag like HMAC), we fallback to non-AEAD for keyslot encryption. > > > > In short, to test it, you need to activate device (that works ok with your patches) > > and *access* the data, testing LUKS format and just keyslot access will never use AEAD. > > > > So init the data by direct-io writes, and try to read them back (with dd). > > > > For testing data on dm-integrity (or dm-crypt with AEAD encryption stacked oved dm-integrity) > > I used small utility, maybe it could be useful https://github.com/mbroz/dm_int_tools > > > > Thanks. > > It appears that my code generates the wrong authentication tags on > encryption, but on decryption it works fine. > I'll keep digging ... OK, mystery solved. The skcipher inside authenc() was corrupting the IV before the hmac got a chance to read it. I'll send out an updated version of the series.