On Sat, Jul 18, 2026 at 10:28:17PM +0200, Fabrice Derepas wrote:
> Two defects in trusted_dcp_unseal() combine to allow a heap
> out-of-bounds write in kernel context.
>
> The primary defect is a missing upper-bound check:
> trusted_dcp_unseal() reads p->key_len from an attacker-supplied blob
> with no validation that it is at most MAX_KEY_SIZE (128). It then
> passes p->key_len + DCP_BLOB_AUTHLEN to do_aead_crypto() as the
> operation length. Because p->key is only MAX_KEY_SIZE + 1 = 129
> bytes, any payload_len above 128 in the blob overruns p->key. On
> implementations where AES-128-GCM decryption writes plaintext before
> verifying the authentication tag, up to 330 bytes of out-of-bounds
> heap overwrite can occur before -EBADMSG is returned.
>
> The secondary defect is an integer overflow in calc_blob_len(), which
> computes its sum in size_t but returns unsigned int, truncating the
> result on 64-bit platforms. This allows a crafted blob with
> payload_len near UINT_MAX to bypass the sanity check
> (blen != p->blob_len). Due to a second wrap in the expression
> p->key_len + DCP_BLOB_AUTHLEN (computed in unsigned int at the call
> site), the len value that reaches do_aead_crypto() via this path is
> at most 15 bytes and does not directly amplify the OOB write, but
> the integrity check bypass must be fixed.
>
> Fix the primary OOB by validating p->key_len against MIN_KEY_SIZE
> and MAX_KEY_SIZE immediately after reading it from the blob, matching
> the validation already performed in trusted_core.c on the Opt_new
> path. Fix the overflow by changing the return type of calc_blob_len()
> to size_t; update the two blen declarations from int to size_t to
> avoid narrowing-conversion warnings, and update the pr_err format
> specifier for blen accordingly.
>
> The seal path is not affected: p->key_len is validated by
> trusted_core.c before trusted_dcp_seal() is called, and the
> blen > MAX_BLOB_SIZE guard provides defence in depth there.
>
> Exploitation requires high privileges. The unseal path is reached
> via add_key("trusted", ..., KEY_SPEC_*) with the Opt_load command,
> which requires write permission to a keyring that accepts trusted
> keys -- in practice CAP_SYS_ADMIN. Additionally,
> CONFIG_TRUSTED_KEYS_DCP must be enabled. The attacker must also supply
> a crafted sealed blob, which in the typical dm-crypt use case means
> either physical access to modify on-disk key material or a prior
> compromise of a privileged process that writes the blob.
This is hypothesis of an exploit at most.
The change look while but this paragraph is way too abstract
to qualify for anything.
>
> How this was found:
> Found by applying the Squeeze Loop strategy ("The Squeeze Loop
> Strategy: Catching Coherent-and-Wrong Artifacts with an
> Author-Independent Executable Oracle," Zenodo
> DOI 10.5281/zenodo.20787816, 2026) on its C terrain using
> deductive verification. The formal proof of calc_blob_len (12/12
> goals discharged) required a no_wrap precondition on payload_len;
> removing that precondition caused 3 of 6 proof goals to fail.
> Tracing the caller path confirmed the bounds bypass and the missing
> key_len validation.
>
> Signed-off-by: Fabrice Derepas <[email protected]>
> ---
> security/keys/trusted-keys/trusted_dcp.c | 13 ++++++++++---
> 1 file changed, 9 insertions(+), 4 deletions(-)
>
> diff --git a/security/keys/trusted-keys/trusted_dcp.c
> b/security/keys/trusted-keys/trusted_dcp.c
> index XXXXXXXXXXXXXXXX..YYYYYYYYYYYYYYYY 100644
> --- a/security/keys/trusted-keys/trusted_dcp.c
> +++ b/security/keys/trusted-keys/trusted_dcp.c
> @@ -69,7 +69,7 @@ module_param_named(dcp_skip_zk_test, skip_zk_test, bool, 0);
> MODULE_PARM_DESC(dcp_skip_zk_test, "Don't test whether device keys are
> zero'ed");
>
> -static unsigned int calc_blob_len(unsigned int payload_len)
> +static size_t calc_blob_len(unsigned int payload_len)
> {
> return sizeof(struct dcp_blob_fmt) + payload_len + DCP_BLOB_AUTHLEN;
> }
> @@ -200,7 +200,8 @@ static int trusted_dcp_seal(struct trusted_key_payload
> *p, char *datablob)
> {
> struct dcp_blob_fmt *b = (struct dcp_blob_fmt *)p->blob;
> - int blen, ret;
> + size_t blen;
> + int ret;
> u8 *plain_blob_key;
>
> blen = calc_blob_len(p->key_len);
> @@ -242,21 +243,25 @@ static int trusted_dcp_unseal(struct
> trusted_key_payload *p, char *datablob)
> {
> struct dcp_blob_fmt *b = (struct dcp_blob_fmt *)p->blob;
> - int blen, ret;
> + size_t blen;
> + int ret;
> u8 *plain_blob_key = NULL;
>
> if (b->fmt_version != DCP_BLOB_VERSION) {
> pr_err("DCP blob has bad version: %i, expected %i\n",
> b->fmt_version, DCP_BLOB_VERSION);
> ret = -EINVAL;
> goto out;
> }
>
> p->key_len = le32_to_cpu(b->payload_len);
> + if (p->key_len < MIN_KEY_SIZE || p->key_len > MAX_KEY_SIZE)
> + return -EINVAL;
> +
> blen = calc_blob_len(p->key_len);
> if (blen != p->blob_len) {
> - pr_err("DCP blob has bad length: %i != %i\n", blen,
> + pr_err("DCP blob has bad length: %zu != %u\n", blen,
> p->blob_len);
> ret = -EINVAL;
> goto out;
> --
> 2.43.0
As said, code changes themselves look reasonable but the commit message
is painting a picture of impact that this bug really does not have.
BR, Jarkko
