9.10. Advisory TFV-10 (CVE-2022-47630)
Title |
Incorrect validation of X.509 certificate extensions can result in an out-of-bounds read. |
---|---|
CVE ID |
|
Date |
Reported on 12 Dec 2022 |
Versions Affected |
v1.2 to v2.8 |
Configurations Affected |
BL1 and BL2 with Trusted Boot enabled with custom,
downstream usages of |
Impact |
Out-of-bounds read. |
Fix Version |
Note that 72460f50e2437a85 is not fixing any vulnerability per se but it is required for f5c51855d36e399e to apply cleanly. |
Credit |
Demi Marie Obenour, Invisible Things Lab |
This security advisory describes a vulnerability in the X.509 parser used to parse boot certificates in TF-A trusted boot: it is possible for a crafted certificate to cause an out-of-bounds memory read.
Note that upstream platforms are not affected by this. Only downstream platforms may be, if (and only if) the interfaces described below are used in a different context than seen in upstream code. Details of such context is described in the rest of this document.
To fully understand this security advisory, it is recommended to refer to the following standards documents:
RFC 5280, Internet X.509 Public Key Infrastructure Certificate and Certificate Revocation List (CRL) Profile.
ITU-T X.690, ASN.1 encoding rules: Specification of Basic Encoding Rules (BER), Canonical Encoding Rules (CER) and Distinguished Encoding Rules (DER).
9.10.1. Bug 1: Insufficient certificate validation
The vulnerability lies in the following source file:
drivers/auth/mbedtls/mbedtls_x509_parser.c
. By design, get_ext()
does
not check the return value of the various mbedtls_*()
functions, as
cert_parse()
is assumed to have guaranteed that they will always succeed.
However, it passes the end of an extension as the end pointer to these
functions, whereas cert_parse()
passes the end of the TBSCertificate
.
Furthermore, cert_parse()
does not check that the contents of the extension
have the same length as the extension itself. It also does not check that the
extension block extends to the end of the TBSCertificate
.
This is a problem, as mbedtls_asn1_get_tag()
leaves *p
and *len
undefined on failure. In practice, this results in get_ext()
continuing to
parse at different offsets than were used (and validated) by cert_parse()
,
which means that the in-bounds guarantee provided by cert_parse()
no longer
holds. The result is that it is possible for get_ext()
to read memory past
the end of the certificate. This could potentially access memory with dangerous
read side effects, or leak microarchitectural state that could theoretically be
retrieved through some side-channel attacks as part of a more complex attack.
9.10.2. Bug 2: Missing bounds check in auth_nvctr()
auth_nvctr()
does not check that the buffer provided is
long enough to hold an ASN.1 INTEGER
. Since auth_nvctr()
will only ever
read 6 bytes, it is possible to read up to 6 bytes past the end of the buffer.
9.10.3. Exploitability Analysis
9.10.3.1. Upstream TF-A Code
In upstream TF-A code, the only caller of auth_nvctr()
takes its input from
get_ext()
, which means that the second bug is exploitable, so is the first.
Therefore, only the first bug need be considered.
All standard chains of trust provided in TF-A source tree (that is, under
drivers/auth/
) require that the certificate’s signature has already been
validated prior to calling get_ext()
, or any function that calls get_ext()
.
Platforms taking their chain of trust from a dynamic configuration file (such as
fdts/tbbr_cot_descriptors.dtsi
) are also safe, as signature verification will
always be done prior to any calls to get_ext()
or auth_nvctr()
in this
case, no matter the order of the properties in the file. Therefore, it is not
possible to exploit this vulnerability pre-authentication in upstream TF-A.
Furthermore, the data read through get_ext()
only
ever gets used by the authentication framework (drivers/auth/auth_mod.c
),
which greatly reduces the range of inputs it will ever receive and thus the
impact this has. Specifically, the authentication framework uses get_ext()
in three cases:
Retrieving a hash from an X.509 certificate to check the integrity of a child certificate (see
auth_hash()
).Retrieving the signature details from an X.509 certificate to check its authenticity and integrity (see
auth_signature()
).Retrieving the security counter value from an X.509 certificate to protect it from unauthorized rollback to a previous version (see
auth_nvctr()
).
None of these uses authentication framework write to the out-of-bounds memory, so no memory corruption is possible.
In summary, there are 2 separate issues - one in get_ext()
and another one
in auth_nvctr()
- but neither of these can be exploited in the context of
TF-A upstream code.
Only in the following 2 cases do we expect this vulnerability to be triggerable prior to authentication:
The platform uses a custom chain of trust which uses the non-volatile counter authentication method (
AUTH_METHOD_NV_CTR
) before the cryptographic authentication method (AUTH_METHOD_SIG
).The chain of trust uses a custom authentication method that calls
get_ext()
before cryptographic authentication.
9.10.3.2. Custom Image Parsers
If the platform uses a custom image parser instead of the certificate parser,
the bug in the certificate parser is obviously not relevant. The bug in
auth_nvctr()
may be relevant, but only if the returned data is:
Taken from an untrusted source (meaning that it is read prior to authentication).
Not already checked to be a primitively-encoded ASN.1 tag.
In particular, if the custom image parser implementation wraps a 32-bit integer
in an ASN.1 INTEGER
, it is not affected.