Re: [PATCH 1/2] KEYS: fix signature length passed to KEYCTL_PKEY_VERIFY

From: Rodolfo Giometti

Date: Tue Nov 17 2026 - 08:21:48 EST


On 11/11/2026 16:56, Florian Bauckholt wrote:
> Commit cb481e59ea6c ("KEYS: fix overflow in keyctl_pkey_params_get_2()")
> made keyctl_pkey_params_get_2() set params->out_len to the maximum
> output size of the operation instead of the size provided by the caller.
> It does so for KEYCTL_PKEY_VERIFY as well, but verify has no output
> buffer: out_len shares a union with in2_len, the length of the
> signature.
>
> As a result keyctl_pkey_verify() ignores the signature length passed by
> the caller. It copies max_sig_size bytes from the signature pointer and
> passes max_sig_size as the signature size to the verification.
>
> This makes no difference for signatures that are always max_sig_size
> bytes long, such as RSA. A shorter signature, e.g. a DER-encoded ECDSA
> signature, is followed by whatever is behind it in the caller's memory,
> or the call fails with -EFAULT if that memory is not readable. The
> x962 ECDSA template currently ignores data following the signature,
> which hides the problem.

The test suite of craf (https://github.com/giometti/crypto-afalg,
test3008) hits this with RSA too: it cuts a genuine 256-byte RSA-2048
PKCS#1 signature to 255 bytes and expects the verify to fail. On
Ubuntu's 6.8.0-147 kernel, which carries the stable backport of
cb481e59ea6c, the cut signature verifies about once in 256 runs. Over
2000 runs per build:

plain: ok=1991 ko=9
asan: ok=1997 ko=3

and a failing run prints (line wrapped here):

libkcrypto [pkey-cert]: kc-check pkey-verify cert:test3008.cert
enc=pkcs1 hash=sha256 test3008.d256 test3008.bad: exit 0, not 1:
KO!

To reproduce it (the count printed is the number of failures; here it
gave 3 out of 1000 runs):

git clone https://github.com/giometti/crypto-afalg
cd crypto-afalg/tests
export TESTS_TMPDIR=$(mktemp -d)
for i in $(seq 1000); do ./do_test.sh 3008; done | grep -c 'KO!'

I also replayed the inputs of one failing run on a UML kernel of
current mainline (666539eecdf4, after v7.4-rc3): the 255-byte signature
verifies 20 times out of 20, and with your patch applied it is refused
20 times out of 20.

Maybe it is worth saying so in the commit message, since this is going
to stable?

I'll send my Tested-by once test3008 has run 2000 times on a kernel
with your patch applied.

>
> Use the length provided by the caller, which has already been checked
> against max_sig_size.
>
> Fixes: cb481e59ea6c ("KEYS: fix overflow in keyctl_pkey_params_get_2()")
> Cc: stable@xxxxxxxxxxxxxxx
> Link: https://lore.kernel.org/all/20260919221742.3709658-1-benquike@xxxxxxxxx/
> Signed-off-by: Florian Bauckholt <florian.bauckholt@xxxxxxxx>
> ---
> security/keys/keyctl_pkey.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/security/keys/keyctl_pkey.c b/security/keys/keyctl_pkey.c
> index 15b6bf6399b5..794fa3d01911 100644
> --- a/security/keys/keyctl_pkey.c
> +++ b/security/keys/keyctl_pkey.c
> @@ -160,7 +160,7 @@ static int keyctl_pkey_params_get_2(const struct keyctl_pkey_params __user *_par
> uparams.in2_len > info.max_sig_size)
> return -EINVAL;
>
> - params->out_len = info.max_sig_size;
> + params->in2_len = uparams.in2_len;
> break;
> default:
> return -EOPNOTSUPP;
>

Ciao,

Rodolfo

--
GNU/Linux Solutions e-mail: giometti@xxxxxxxxxxxx
Linux Device Driver giometti@xxxxxxxx
Embedded Systems phone: +39 349 2432127
UNIX programming