Re: [PATCH 00/10] crypto: omap fixes towards 5.5

[Ard Biesheuvel <ard.biesheuvel@linaro.org>]

On Fri, 25 Oct 2019 at 14:18, Tero Kristo <t-kristo@ti.com> wrote:
>
> On 25/10/2019 15:05, Ard Biesheuvel wrote:
> > On Fri, 25 Oct 2019 at 13:56, Tero Kristo <t-kristo@ti.com> wrote:
> >>
> >> On 25/10/2019 14:55, Tero Kristo wrote:
> >>> On 25/10/2019 14:33, Ard Biesheuvel wrote:
> >>>> On Thu, 17 Oct 2019 at 14:26, Tero Kristo <t-kristo@ti.com> wrote:
> >>>>>
> >>>>> Hi,
> >>>>>
> >>>>> This series fixes a number of bugs with omap crypto implementation.
> >>>>> These have become evident with the changes to the cryptomanager, where
> >>>>> it adds some new test cases and modifies some existing, namely the split
> >>>>> update tests. Also, while fixing the cryptomanager induced bugs, some
> >>>>> other surfaced with tcrypt/IPSec tests, so fixed them aswell.
> >>>>>
> >>>>> Patch #9 is against crypto core modifying the crypto_wait_req
> >>>>> common API to have a timeout for it also, currently it waits forever
> >>>>> and it is kind of difficult to see what test fails with crypto manager.
> >>>>> This is not really needed for anything, but it is kind of nice to have
> >>>>> (makes debugging easier.)
> >>>>>
> >>>>> This series has been tested on top of 5.4-rc2, with following setups,
> >>>>> on AM57xx-beagle-x15 board:
> >>>>>
> >>>>> - crypto manager self tests
> >>>>> - tcrypt performance test
> >>>>> - ipsec test with strongswan
> >>>>>
> >>>>> This series depends on the skcipher API switch patch from Ard Biesheuvel
> >>>>> [1].
> >>>>>
> >>>>
> >>>> Hi Tero,
> >>>>
> >>>> On my BeagleBone White, I am hitting the following issues after
> >>>> applying these patches:
> >>>>
> >>>> [    7.493903] alg: skcipher: ecb-aes-omap encryption unexpectedly
> >>>> succeeded on test vector "random: len=531 klen=32";
> >>>> expected_error=-22, cfg="random: inplace may_sleep use_finup
> >>>> src_divs=[44.72%@+4028, <flush>14.70%@alignmask+3, 19.45%@+4070,
> >>>> 21.13%@+2728]"
> >>>> [    7.651103] alg: skcipher: cbc-aes-omap encryption unexpectedly
> >>>> succeeded on test vector "random: len=1118 klen=32";
> >>>> expected_error=-22, cfg="random: may_sleep use_final
> >>>> src_divs=[<reimport>41.87%@+31, <flush>58.13%@+2510]"
> >>>>
> >>>> These are simply a result of the ECB and CBC implementations not
> >>>> returning -EINVAL when the input is not a multiple of the block size.
> >>>>
> >>>> [    7.845527] alg: skcipher: blocksize for ctr-aes-omap (16) doesn't
> >>>> match generic impl (1)
> >>>>
> >>>> This means cra_blocksize is not set to 1 as it should. If your driver
> >>>> uses the skcipher walk API, it should set the walksize to
> >>>> AES_BLOCK_SIZE to ensure that the input is handled correctly. If you
> >>>> don't, then you can disregard that part.
> >>>>
> >>>> [    8.306491] alg: aead: gcm-aes-omap setauthsize unexpectedly
> >>>> succeeded on test vector "random: alen=3 plen=31 authsize=6 klen=9";
> >>>> expected_error=-22
> >>>>
> >>>> Another missing sanity check. GCM only permits certain authsizes.
> >>>>
> >>>> [    9.074703] omap_crypto_copy_sgs: Couldn't allocate pages for
> >>>> unaligned cases.
> >>>>
> >>>> This is not a bug, but I'm not sure if the below is related or not.
> >>>>
> >>>> I'll preserve the binaries, in case you need me to objdump anything.
> >>>
> >>> What are these tests you are executing? For me, the testmgr self test
> >>> suite is passing just fine. Any extra tests you have enabled somehow?
> >>>
> >
> > I enabled CONFIG_CRYPTO_MANAGER_EXTRA_TESTS, which enables a bunch of
> > fuzz tests of the offloaded algorithms against the generic
> > implementations.
>
> Ahha I see, let me give that a shot locally. I have so far only been
> testing with the standard suite.
>
> >
> >>> I am also running full test on different board though (am57xx), I
> >>> haven't been explicitly running anything on am335x.
> >>
> >> Oh, and btw, did you try without my series? I think the selftests are
> >> failing rather miserably without them...
> >>
> >
> > No, I just tried a branch with mine and your patches applied.
>
> Could you give it a shot without the CRYPTO_MANAGER_EXTRA_TESTS, that
> should pass with my series, and fail without?
>

The missing output IVs are fixed by this series, but it seems we need
some more work to get all the wrinkles ironed out. I sent some patches
on top that address a couple of them, but we still need a proper fix
for the situation where only assocdata is presented, and cryptlen == 0

Feel free to merge my patches into your series, or take bits and
pieces into your own patches where needed.