FIPS 140-3 requires us to indicate whether an operation was using
approved services or not. The FIPS 140-3 implementation guidelines
provide two basic approaches to doing this: implicit indicators, and
explicit indicators.
Implicit indicators are basically the concept of "if the operation
passes, it was approved". We were originally aiming for implicit
indicators in our copy of OpenSSL. However, this proved to be a problem,
because we wanted to certify a signature service, and FIPS 140-3
requires that a signature service computes the digest to be signed
within the boundaries of the FIPS module. Since we were planning to
certify fips.so only, this means that EVP_PKEY_sign/EVP_PKEY_verify
would have to be blocked. Unfortunately, EVP_SignFinal uses
EVP_PKEY_sign internally, but outside of fips.so and thus outside of the
FIPS module boundary. This means that using implicit indicators in
combination with certifying only fips.so would require us to block both
EVP_PKEY_sign and EVP_SignFinal, which are the two APIs currently used
by most users of OpenSSL for signatures.
EVP_DigestSign would be acceptable, but has only been added in 3.0 and
is thus not yet widely used.
As a consequence, we've decided to introduce explicit indicators so that
EVP_PKEY_sign and EVP_SignFinal can continue to work for now, but
FIPS-aware applications can query the explicit indicator to check
whether the operation was approved.
To avoid affecting the ABI and public API too much, this is implemented
as an exported symbol in fips.so and a private header, so applications
that wish to use this will have to dlopen(3) fips.so, locate the
function using dlsym(3), and then call it. These applications will have
to build against the private header in order to use the returned
pointer.
Modify util/mkdef.pl to support exposing a symbol only for a specific
provider identified by its name and path.
Signed-off-by: Clemens Lang <cllang@redhat.com>
Resolves: rhbz#2087147
Disabling EVP_PKEY_sign and EVP_PKEY_verify also breaks EVP_SignFinal,
which is used by many applications, among them OpenSSH. This change thus
broke sshd in FIPS mode. Revert it for now until we found a better
solution.
Related: rhbz#2087147
Signed-off-by: Clemens Lang <cllang@redhat.com>
1. Deny SHA-1 signature verification in FIPS provider
For RHEL, we already disable SHA-1 signatures by default in the default
provider, so it is unexpected that the FIPS provider would have a more
lenient configuration in this regard. Additionally, we do not think
continuing to accept SHA-1 signatures is a good idea due to the
published chosen-prefix collision attacks.
As a consequence, disable verification of SHA-1 signatures in the FIPS
provider.
This requires adjusting a few tests that would otherwise fail:
- 30-test_acvp: Remove the test vectors that use SHA-1.
- 30-test_evp: Mark tests in evppkey_rsa_common.txt and
evppkey_ecdsa.txt that use SHA-1 digests as "Availablein = default",
which will not run them when the FIPS provider is enabled.
- 80-test_cms: Re-generate all certificates in test/smime-certificates
using the mksmime-certs.sh script, because most of them were signed
with SHA-1 and thus fail verification in the FIPS provider. Keep
smec3.pem, which was used to sign static test data in
test/recipes/80-test_cms_data/ciphertext_from_1_1_1.cms, which would
otherwise no longer verify. Note that smec3.pem was signed with
a smroot.pem, which was now re-generated. This does not affect the
test.
Fix some other tests by explicitly running them in the default
provider, where SHA-1 is available.
- 80-test_ssl_old: Skip tests that rely on SSLv3 and SHA-1 when run with
the FIPS provider.
2. Disable EVP_PKEY_{sign,verify} in FIPS provider
The APIs to compute both digest and signature in one step,
EVP_DigestSign*/EVP_DigestVerify* and EVP_Sign*/EVP_Verify*, should be
used instead. This ensures that the digest is computed inside of the
FIPS module, and that only approved digests are used.
Update documentation for EVP_PKEY_{sign,verify} to reflect this.
Since the KATs use EVP_PKEY_sign/EVP_PKEY_verify, modify the tests to
set the OSSL_SIGNATURE_PARAM_KAT parameter and use EVP_PKEY_sign_init_ex
and EVP_PKEY_verify_init_ex where these parameters can be passed on
creation and allow EVP_PKEY_sign/EVP_PKEY_verify when this parameter is
set and evaluates as true.
Move tests that use the EVP_PKEY API to only run in the default
provider, since they would fail in the FIPS provider. This also affects
a number of CMS tests where error handling is insufficient and failure
to sign would only show up when verifying the CMS structure due to
a parse error.
Resolves: rhbz#2087147
Signed-off-by: Clemens Lang <cllang@redhat.com>
Include a hash of specfile, patches, and sources in the FIPS module
version. This should allow us to uniquely identify a build that we do,
so that we can be sure which specific binary is being submitted for
validation and was certified.
The previous solution used $(date +%Y%m%d), which had some risks related
to build server timezone and build date differences on different
architectures.
Resolves: rhbz#2070550
Signed-off-by: Clemens Lang <cllang@redhat.com>
Invocations of EVP_PKEY_CTX_set_rsa_padding(RSA_PKCS1_PSS_PADDING)
before setting an allowed digest with EVP_PKEY_CTX_set_signature_md()
would fail with SHA-1 use in signatures disabled, because OpenSSL's
internal default for the digest was SHA-1.
This isn't documented in any of the manpages, hence we expect users to
always call both EVP_PKEY_CTX_set_rsa_padding() and
EVP_PKEY_CTX_set_signature_md(). We do not want set_rsa_padding() to
fail if users set a non-SHA-1 signature algorithm after setting the
padding mode, though, so change the internal default to SHA-256 if SHA-1
is disabled.
Resolves: rhbz#2062640
We want legacy policy to be able to talk to older RHEL that only
supports SHA1 signature algorithms, so allow SHA1 signatures even in
seclevel 2 if rh-allow-sha1-signatures is set to yes.
Resolves: rhbz#2060510
Signed-off-by: Clemens Lang <cllang@redhat.com>
providers/implementations/signature/{ec,}dsa_sig.c accept a NID_undef
digest, so to prevent SHA1 from working with ECDSA and DSA, we must
return a negative value in securitycheck.c.
Resolves: rhbz#2031742
The EVP_DigestSign API is used in TLS to compute a SHA1 HMAC, which is
OK from our point of view, but was blocked so far. Modify
0049-Selectively-disallow-SHA1-signatures.patch to check the EVP_PKEY
type for HMAC (and TLS1-PRF and HKDF), and allow SHA1 for these cases.
Note that TLS1.1 signs a MD5-SHA1 hash with a private key, which does
not work with rh-allow-sha1-signatures = no, so the minimum TLS version
will be TLS 1.2.
Resolves: rhbz#2031742
Signed-off-by: Clemens Lang <cllang@redhat.com>
