Re: [PATCH -next] crypto: asymmetric_keys: verify_pefile: fix integer underflow in pefile_strip_sig_wrapper

[Ignat Korchagin]

On Mon, Nov 2, 2026 at 7:32 AM Gu Bowen <gubowen5@xxxxxxxxxx> wrote:

nit: the subject line looks malformed: weird "-next" suffix pointing
nowhere. Otherwise:

> The pefile_strip_sig_wrapper() function validates the win_certificate
> wrapper that precedes the PKCS#7 signature in a PE binary.  It checks
> that the wrapper length does not exceed the certificate table size
> (``wrapper.length > ctx->sig_len``) but fails to verify that the
> wrapper length is at least ``sizeof(struct win_certificate)`` (8 bytes).
>
> When a crafted PE image supplies a wrapper with ``length`` in the range
> [1, 7], the subsequent arithmetic::
>
>     ctx->sig_len = wrapper.length;      /* e.g. 1 */
>     ctx->sig_offset += sizeof(wrapper);
>     ctx->sig_len -= sizeof(wrapper);    /* 1 - 8 = 0xFFFFFFF9 (unsigned) */
>
> underflows the unsigned ``ctx->sig_len`` to a value near UINT_MAX.  The
> following ``sig_len < 4`` sanity check is bypassed because 0xFFFFFFF9 is
> greater than 4.
>
> The underflowed ``sig_len`` is then passed to verify_pkcs7_signature()
> and onwards to asn1_ber_decoder().  The ASN.1 decoder's own
> ``datalen > 65535`` entry guard can be satisfied by constructing a
> PKCS#7 length field of exactly 65535 (e.g. ``30 82 ff fb``), allowing
> the decoder to proceed and attempt to read OID octets beyond the
> allocated vmalloc buffer, triggering a page fault in look_up_OID()::
>
>   BUG: unable to handle page fault for address: ffffd1e986cb2000
>   #PF: supervisor read access in kernel mode
>   RIP: 0010:look_up_OID+0x14/0x100
>   Call Trace:
>    pkcs7_note_OID+0x33/0xa0
>    asn1_ber_decoder+0x363/0xbd0
>    pkcs7_parse_message+0xd9/0x1b0
>    verify_pkcs7_signature+0x24/0x80
>    verify_pefile_signature+0x7b/0xf0
>    kexec_kernel_verify_pe_sig+0x1b/0x40
>    kimage_file_alloc_init+0x1d8/0x3f0
>    __se_sys_kexec_file_load+0xc6/0x270
>
> This is reachable through the kexec_file_load syscall on architectures
> that verify PE signatures (x86 bzImage with
> CONFIG_KEXEC_BZIMAGE_VERIFY_SIG, arm64 Image with
> CONFIG_KEXEC_IMAGE_VERIFY_SIG), requiring CAP_SYS_BOOT.
>
> The oops occurs while the caller holds kexec_lock; if the faulting task
> is killed, the lock is never released, leaving all subsequent kexec and
> kdump operations permanently returning -EBUSY until reboot.
>
> Add a lower-bound check on wrapper.length before the subtraction to
> prevent the underflow.
>
> Fixes: 09dacbbda935 ("pefile: Strip the wrapper off of the cert data block")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Gu Bowen <gubowen5@xxxxxxxxxx>

Reviewed-by: Ignat Korchagin <ignat@xxxxxxxxx>

> ---
>  crypto/asymmetric_keys/verify_pefile.c | 5 +++++
>  1 file changed, 5 insertions(+)
>
> diff --git a/crypto/asymmetric_keys/verify_pefile.c b/crypto/asymmetric_keys/verify_pefile.c
> index cec99db14129..cfa85ccb1cf7 100644
> --- a/crypto/asymmetric_keys/verify_pefile.c
> +++ b/crypto/asymmetric_keys/verify_pefile.c
> @@ -146,6 +146,11 @@ static int pefile_strip_sig_wrapper(const void *pebuf,
>                         ctx->sig_len, wrapper.length);
>                 return -ELIBBAD;
>         }
> +       if (wrapper.length < sizeof(wrapper)) {
> +               pr_warn("Signature wrapper too small (%x < %zx)\n",
> +                       wrapper.length, sizeof(wrapper));
> +               return -ELIBBAD;
> +       }
>         if (wrapper.revision != WIN_CERT_REVISION_2_0) {
>                 pr_warn("Signature is not revision 2.0\n");
>                 return -ENOTSUPP;
> --
> 2.43.0
>
>