On Sat, 5 Oct 2019 at 18:15, Gert Robben <t2@xxxxxxx> wrote:
>
> Op 05-10-2019 om 11:11 schreef Ard Biesheuvel:
> > Commit 79c65d179a40e145 ("crypto: cbc - Convert to skcipher") updated
> > the generic CBC template wrapper from a blkcipher to a skcipher algo,
> > to get away from the deprecated blkcipher interface. However, as a side
> > effect, drivers that instantiate CBC transforms using the blkcipher as
> > a fallback no longer work, since skciphers can wrap blkciphers but not
> > the other way around. This broke the geode-aes driver.
> >
> > So let's fix it by moving to the sync skcipher interface when allocating
> > the fallback. At the same time, align with the generic API for ECB and
> > CBC by rejecting inputs that are not a multiple of the AES block size.
> >
> > Fixes: 79c65d179a40e145 ("crypto: cbc - Convert to skcipher")
> > Cc: <stable@xxxxxxxxxxxxxxx> # v4.20+ ONLY
> > Signed-off-by: Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx>
> > ---
> > v2: pass dst and src scatterlist in the right order
> > reject inputs that are not a multiple of the block size
>
> Yes, with this patch, the CRYPTO_MANAGER_EXTRA_TESTS output nothing
> (apart from "extra crypto tests enabled").
> All items in /proc/crypto have "selftest: passed" mentioned.
> "openssl speed -evp aes-128-cbc -elapsed -engine afalg" reaches the
> proper speed.
> And nginx (correctly) transfers files about 40% faster than without
> geode-aes.
>
> I didn't think about testing ecb before, because I don't use it.
> Now that I did, I tried the same openssl benchmark for ecb.
> But that only reaches software AES speed, and "time" also shows the work
> is being done in "user" instead of "sys" (see below).
> Yet I see no errors.
> (Maybe this is normal/expected, so I didn't look much further into it).
>
> Thank you,
> Gert
>
> # time openssl speed -evp aes-128-cbc -elapsed -engine afalg
> - - - 8< - - -
> type 16 bytes 64 bytes 256 bytes 1024 bytes 8192
> bytes 16384 bytes
> aes-128-cbc 135.82k 539.29k 2087.90k 7491.16k
> 29221.69k 34943.67k
>
> real 0m18.081s
> user 0m0.516s
> sys 0m17.541s
>
> # time openssl speed -evp aes-128-ecb -elapsed -engine afalg
> - - - 8< - - -
> type 16 bytes 64 bytes 256 bytes 1024 bytes 8192
> bytes 16384 bytes
> aes-128-ecb 4480.65k 5137.66k 5336.94k 5410.19k
> 5409.91k 5409.91k
>
> real 0m18.084s
> user 0m18.046s
> sys 0m0.012s
>
It seems likely that the ECB code in OpenSSL is not invoking the afalg
code at all. Since ECB is just the bare block cipher applied to each
block in the input, I wonder if it even uses a skcipher like interface
internally.
