On Thu, Mar 11, 2021 at 10:47:50AM +0800, Jia Zhang wrote: > > > On 2021/3/11 上午5:39, Jarkko Sakkinen wrote: > > On Wed, Mar 10, 2021 at 08:44:44PM +0800, Jia Zhang wrote: > >> > >> > >> On 2021/3/2 下午9:47, Jarkko Sakkinen wrote: > >>> On Mon, Mar 01, 2021 at 09:54:37PM -0800, Andy Lutomirski wrote: > >>>> On Mon, Mar 1, 2021 at 9:06 PM Tianjia Zhang > >>>> <tianjia.zhang@xxxxxxxxxxxxxxxxx> wrote: > >>>>> > >>>>> > >>>>> > >>>>> On 3/1/21 5:54 PM, Jarkko Sakkinen wrote: > >>>>>> On Mon, Mar 01, 2021 at 01:18:36PM +0800, Tianjia Zhang wrote: > >>>>>>> q2 is not always 384-byte length. Sometimes it only has 383-byte. > >>>>>> > >>>>>> What does determine this? > >>>>>> > >>>>>>> In this case, the valid portion of q2 is reordered reversely for > >>>>>>> little endian order, and the remaining portion is filled with zero. > >>>>>> > >>>>>> I'm presuming that you want to say "In this case, q2 needs to be reversed because...". > >>>>>> > >>>>>> I'm lacking these details: > >>>>>> > >>>>>> 1. Why the length of Q2 can vary? > >>>>>> 2. Why reversing the bytes is the correct measure to counter-measure > >>>>>> this variation? > >>>>>> > >>>>>> /Jarkko > >>>>>> > >>>>> > >>>>> When use openssl to generate a key instead of using the built-in > >>>>> sign_key.pem, there is a probability that will encounter this problem. > >>>>> > >>>>> Here is a problematic key I encountered. The calculated q1 and q2 of > >>>>> this key are both 383 bytes, If the length is not processed, the > >>>>> hardware signature will fail. > >>>> > >>>> Presumably the issue is that some keys have parameters that have > >>>> enough leading 0 bits to be effectively shorter. The openssl API > >>>> (and, sadly, a bunch of the ASN.1 stuff) treats these parameters as > >>>> variable-size integers. > >>> > >>> But the test uses a static key. It used to generate a key on fly but > >> > >> IMO even though the test code, it comes from the linux kernel, meaning > >> that its quality has a certain guarantee and it is a good reference, so > >> the test code still needs to ensure its correctness. > > > > Hmm... what is working incorrectly then? > > In current implementation, it is working well, after all the static key > can derive the full 384-byte length of q1 and q2. As mentioned above, if > someone refers to the design of signing tool from selftest code, it is > quite possible that the actual implementation will use dynamical or > external signing key deriving shorter q1 and/or q2 in length. A self-test needs is not meant to be generic to be directly used in 3rd party code. With the current key there is not issue => there is no issue. > > Going back the technical background, I'm not a crypto expert, so I > choose to check the code, doc and make experiment. > > Accorindg to intel sdm vol3 37.14: > > ``` > SIGSTRUCT includes four 3072-bit integers (MODULUS, SIGNATURE, Q1, Q2). > Each such integer is represented as a byte strings of length 384, with > the most significant byte at the position “offset + 383”, and the least > significant byte at position “offset”. > > ... > > The 3072-bit integers Q1 and Q2 are defined by: > q1 = floor(Signature^2 / Modulus); > q2 = floor((Signature^3 - q1 * Signature * Modulus) / Modulus); > ``` > > and the implementation of singing tool from Intel SGX SDK > (https://github.com/intel/linux-sgx/blob/master/sdk/sign_tool/SignTool/sign_tool.cpp#L381 > ), the most significant byte shuld be at the position “offset + > q1/q2_actuall_len”, and the padding portion should be filled with zero, > and don't involve the order reverse, but the selftest code always does. > This is the root cause of SGX_INVALID_SIGNATURE. > > Just simplily enforce size_q1 and size_q2 to SE_KEY_SIZE, and generate > randome siging key with `openssl genrsa -3 -out signing_key.pem 3072`, > the SGX_INVALID_SIGNATURE error will appear up quickly. > > Jia > > > > > /Jarkko > > > /Jarkko
