Module signatures sometimes off by one or a few bytes

[Sean Mullan]

On 4/6/11 8:29 PM, Weijun Wang wrote:
> Do we still have time to change the file format?

Yes.

> say, put the signature in a
> separate file?

That would definitely solve this problem, but has other disadvantages because 
then you have 2 files to keep track of instead of 1.

You could break up the signatures into two, and have one at the beginning 
covering the header and module-info, and one at the end covering the rest of the 
content, but that seems overly complex.

Any type of streamable signature format when writing is going to conflict with 
the desire to have the signature near the front of the file so that the 
module-info and signer can be authenticated before downloading the entire file.

I guess I would like to see if there any objections to my proposed solution 
below first. This seems like a fairly simple solution with a small bit of 
performance impact.

--Sean

>
> Max
>
> On 04/07/2011 04:03 AM, Sean Mullan wrote:
>> Been stuck on a bug for a few days and finally figured it out.
>>
>> When jpkg is used to create signed modules, it first writes a dummy
>> signature to the module file. After the rest of the contents of the
>> module file is written and digested it then seeks back and replaces this
>> with the actual signature. This is because it cannot calculate the real
>> signature until all of the hashes are calculated. This works as long as
>> the size of the dummy signature is the same as the real signature.
>>
>> However, sometimes the length of the dummy signature is off by a couple
>> of bytes from the real signature. This causes the real signature to
>> overwrite into the first few bytes of the next section, and when the
>> signed module is installed, it fails because it cannot identify the
>> section following the signature. I don't know the exact reason the
>> length varies, but I don't think it is a bug in the underlying
>> cryptographic algorithms.
>>
>> I don't think we can assume the length of the signature will be the same
>> each time it is generated even over the same content. We could add some
>> padding bytes to try to accomodate the length changing, but that seems
>> messy. Also, with the addition of timestamps, creating the dummy
>> signature is expensive because it remotely contacts the TSA.
>>
>> I think we need a different way of addressing this. One simple solution
>> is to write the remaining contents of the file (and digest as we go) to
>> a temporary file, calculate the signature and then write the rest of the
>> module file from the contents of the temporary file.
>>
>> Comments?
>>
>> Thanks,
>> Sean