Re: [PATCH v14 09/19] x86/mm: x86/sgx: Signal SEGV_SGXERR for #PFs w/ PF_SGX

[Sean Christopherson <sean.j.christopherson@intel.com>]

On Wed, 2018-09-26 at 14:15 -0700, Andy Lutomirski wrote:
> On Wed, Sep 26, 2018 at 1:55 PM Dave Hansen <dave.hansen@intel.com> wrote:
> > 
> > 
> > On 09/26/2018 01:44 PM, Sean Christopherson wrote:
> > > 
> > > On Wed, Sep 26, 2018 at 01:16:59PM -0700, Dave Hansen wrote:
> > > > 
> > > > We also need to clarify how this can happen.  Is it through something
> > > > than an app does, or is it solely when the hardware does something under
> > > > the covers, like suspend/resume.
> > > Are you looking for something in the changelog, the comment, or just
> > > a response?  If it's the latter...
> > Comments, please.
> > 
> > > 
> > > On bare metal with a bug-free kernel, the only scenario I'm aware of
> > > where we'll encounter these faults is when hardware pulls the rug out
> > > from under us.  In a virtualized environment all bets are off because
> > > the architecture allows VMMs to silently "destroy" the EPC at will,
> > > e.g. KVM, and I believe Hyper-V, will take advantage of this behavior
> > > to support live migration.  Post migration, the destination system
> > > will generate PF_SGX because the EPC{M} can't be migrated between
> > > system, i.e. the destination EPCM sees all EPC pages as invalid.
> > OK, cool.
> > 
> > That's good background fodder for the changelog.
> > 
> > But, for the comment, I'm happy with something like this:
> > 
> >         /*
> >          * The fault resulted from violation of SGX-specific access-
> >          * controls.  This is expected to be the result of some lower
> >          * layer action (CPU suspend/resume, VM migration) and is
> >          * not related to anything the OS did.  Treat it as an access
> >          * error to ensure it is passed up to the app via a signal where
> >          * it can be handled.
> >          */
> > 
> > I really don't think we need to delve too deeply into the relationship
> > between EPCM and PTEs or anything.  Let's just say, "it's not the
> > kernel's fault, it's not the app's fault, so throw up our hands".
> There is a non-nitpicky consideration here.  Logically, user code is
> going to do this (totally made-up pseudocode):
> 
> enclave_t enclave = load_and_init_enclave(...);
> int ret = sgx_run(enclave, some pointers to non-enclave-memory buffers, ...);
> 
> and, with the code in this patch, a correct implementation of
> sgx_run() requires installing a signal handler.  This is nasty, since
> signal handlers, expecially for something like SIGSEGV or SIGBUS, are
> not fantastic to say the least in libraries.
>
> Could we perhaps have a little vDSO entry (or syscall, I suppose) that
> runs an enclave an returns an error code, and rig up the #PF handler
> to check if the error happened in the vDSO entry and fix it up rather
> than sending a signal?


If we want to avoid having to install a signal handler then I'm pretty
sure we'd need to fixup all #GPs and "bad access" #PFs that occur on
EENTER or in the enclave, not just PF_SGX faults.  SGX1 hardware takes
a #GP instead of a #PF on EPCM faults, and SGX2 hardware allows enclaves
to allocate/free/adjust EPC pages at runtime, e.g. an enclave runtime
might want to intercept #PFs from within the enclave so that the enclave
can dynamically grow its stack.

> On Windows, this is much less of a concern, because Windows has real
> scoped fault handling. But Linux doesn't, at least not yet.
> 
> 
> --
> Andy Lutomirski
> AMA Capital Management, LLC