lkml.org 
[lkml]   [2020]   [Feb]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v26 10/22] x86/sgx: Linux Enclave Driver
On Thu, Feb 20, 2020 at 10:33:36AM -0800, Jordan Hand wrote:
> On 2/20/20 10:13 AM, Sean Christopherson wrote:
> > There are essentially two paths we can take:
> >
> > 1) Exempt EPC pages from RIE during mmap()/mprotect(), i.e. don't add
> > PROT_EXEC for enclaves.
> >
> > 2) Punt the issue to userspace.
> >
> > Option (1) is desirable in some ways:
> >
> > - Enclaves will get an executable stack if and only if the loader/creator
> > intentionally configures it to have an executable stack.
> >
> > - Separates enclaves from the personality of the loader.
> >
> > - Userspace doesn't have to do anything for the common case of not
> > wanting an executable stack for its enclaves.
> >
> > The big down side to (1) is that it'd require an ugly hook in architecture
> > agnostic code. And arguably, it reduces the overall security of the
> > platform (more below).
> >
> > For (2), userspace has a few options:
> >
> > a) Tell the linker the enclave loader doesn't need RIE, either via a .note
> > in assembly files or via the global "-z noexecstack" flag.
> >
> > b) Spawn a separate process to run/map the enclave if the enclave loader
> > needs RIE.
> >
> > c) Require enclaves to allow PROT_EXEC on all pages. Note, this is an
> > absolutely terrible idea and only included for completeness.
> >
> > As shown by the lack of a mmap()/mprotect() hook in this series to squash
> > RIE, we chose option (2). Given that enclave loaders are not legacy code
> > and hopefully following decent coding practices, option (2a) should suffice
> > for all loaders. The security benefit mentioned above is that forcing
> > enclave loaders to squash RIE eliminates an exectuable stack as an attack
> > vector on the loader.
>
> I see your point and I do agree that there are security benefits to (2a)
> and I think we could do that for our loader. That said, it does concern
> me that this breaks perfectly valid userspace behavior. If a userspace
> process decides to use RIE, I don't know that the SGX driver should
> disobey that decision.
>
> So option (3) would be to just honor RIE for enclave pages and when page
> permissions are set to PROT_READ in sgx_encl_page_alloc and RIE is set,
> also add PROT_EXEC.

Ah, right, IIRC that idea also came up in our internal discussions. Note,
SGX would need to implement this option by checking for RIE in
sgx_encl_may_map(), as the process that built the enclave may not be the
same process that is running the enclave.

> I understand your concerns that this using RIE is bad security practice
> and I'm not convinced that (3) is the way to go, but from a philosophy
> perspective I don't know that the kernel should be in the business of
> stopping userspace from doing valid things.
>
> If option (3) can't/shouldn't be done for some reason, option (1) at
> least keeps from breaking expected userspace behavior. But I do agree
> that (1) is ugly to implement.

My biggest concern for allowing PROT_EXEC if RIE is that it would result
in #PF(SGX) (#GP on Skylake) due to an EPCM violation if the enclave
actually tried to execute from such a page. This isn't a problem for the
kernel as the fault will be reported cleanly through the vDSO (or get
delivered as a SIGSEGV if the enclave isn't entered through the vDSO), but
it's a bit weird for userspace as userspace will see the #PF(SGX) and
likely assume the EPC was lost, e.g. silently restart the enclave instead
of logging an error that the enclave is broken.

\
 
 \ /
  Last update: 2020-02-20 19:49    [W:0.111 / U:2.364 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site