Re: [PATCH 2/3] binfmt_elf: Verify signature of signed elf binary

["H. Peter Anvin" <hpa@zytor.com>]

You then get into issues like: do we have to ban prelink as a result?

Mimi Zohar <zohar@linux.vnet.ibm.com> wrote:

>On Thu, 2013-01-17 at 10:51 -0500, Vivek Goyal wrote:
>> On Thu, Jan 17, 2013 at 10:37:01AM -0500, Mimi Zohar wrote:
>> > On Tue, 2013-01-15 at 16:34 -0500, Vivek Goyal wrote:
>> > > If a binary is signed, verify its signature. If signature is not
>valid, do
>> > > not allow execution. If binary is not signed, execution is
>allowed
>> > > unconditionally.
>> > > 
>> > > CONFIG_BINFMT_ELF_SIGNATURE controls whether elf binary signature
>support
>> > > is compiled in or not.
>> > > 
>> > > Signature are expected to be present in elf section ".section".
>This code
>> > > is written along the lines of module signature verification code.
>Just
>> > > that I have removed the magic string. It is not needed as
>signature is
>> > > expected to be present in a specific section.
>> > 
>> > Right, it's written along the lines of the original module
>signature
>> > verification code, where the signature was in the ELF header, not
>the
>> > version that was upstreamed, where the signature was appended.
>> > 
>> > > I put the signature into a section, instead of appending it so
>that
>> > > strip operation works fine.
>> > 
>> > Wouldn't the original reasons for not embedding the signature in
>the ELF
>> > header for modules apply here too?  
>> 
>> I think rusty wanted to append signatures. He thought it is much
>easier.
>> Adding a .signature section makes life easier for user space tools.
>One
>> can strip files even after signing.
>> 
>> Not that I am married to the idea of putting signature in a section.
>Just
>> that it sounded reasonable enough to do for an RFC. So if appending 
>> signature proves to be better, it is easy to implement that.
>> > 
>> > > One signs and verifies  all the areas mapped by PT_LOAD segments
>of elf
>> > > binary. Typically Elf header is mapped in first PT_LOAD segment.
>As adding
>> > > .signature section can change three elf header fields (e_shoff,
>e_shnum
>> > > and e_shstrndx), these fields are excluded from digest
>calculation
>> > > 
>> > > Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
>> > 
>> > At this point, why would you want yet another method for signing
>files?
>> 
>> Are you saying that append signature instead of putting them in a
>section
>> or are you saying that just use IMA.
>> 
>> - For the first, I am fine with appending too if that works better.
>So
>>   what's wrong with current implementation. Just because we append
>the
>>   signatures in case of modules, we should follow the same thing for
>>   executables too?
>
>No, I was saying that if this patch set were to be upstreamed, then the
>signature verification, at least for ELF modules and ELF executables,
>should be the same.  The patch would then be a lot smaller.
>
>> - If above comment is w.r.t use of IMA, then I have no issues in
>using
>>   IMA as long as it can meet all the requirements. Looks like there
>is
>>   a long TODO list before we get there. In fact for some things its
>not
>>   even clear whether they fit in IMA or somehwere else.
>> 
>>   - Make sure IMA/EVM stuff chains into secureboot chain of trust.
>
>For sure.
>
>>   - Sort out all the memory locking related issues I mentioned in
>>     other mail. You seemed to be of opinion that it is out of scope
>>     for IMA, but I think it probably is nice extenstion.
>
>Yes, it would be, but I'm not sure that your method of mmaping a file
>with MAP_LOCKED is scalable, when all executables are signed.  If it
>is,
>then why not do it in general?  Otherwise there needs to be a more
>scalable solution.
>
>>     Or somehow a way needs to be found to make sure nobody can modify
>>     process address space without processe's knowledge.
>
>I'm not sure I understand.  Does your patch already address this?
>
>>   - Once all this works, then one needs to figure out all the RPM
>stuff
>>     and plugins to make sure files can be singed on build server and
>>     installed properly on target system.
>
>Yes, we need the distros involvment in this to sign all files.
>Immutable files should be signed with digital signatures.  Mutable
>files
>should have hashes.
>
>thanks,
>
>Mimi

-- 
Sent from my mobile phone. Please excuse brevity and lack of formatting.