On Fri, Dec 14, 2018 at 10:20:39AM -0800, Josh Triplett wrote: > On Fri, Dec 14, 2018 at 09:03:11AM -0800, Sean Christopherson wrote: > > On Fri, Dec 14, 2018 at 07:38:30AM -0800, Sean Christopherson wrote: > > > On Fri, Dec 14, 2018 at 07:12:04AM -0800, Sean Christopherson wrote: > > > > On Fri, Dec 14, 2018 at 09:55:49AM +0000, Jethro Beekman wrote: > > > > > On 2018-12-14 03:01, Sean Christopherson wrote: > > > > > >+2: pop %rbx > > > > > >+ pop %r12 > > > > > >+ pop %r13 > > > > > >+ pop %r14 > > > > > >+ pop %r15 > > > > > >+ pop %rbp > > > > > >+ ret > > > > > > > > > > x86-64 ABI requires that you call CLD here (enclave may set it). > > > > > > > > Ugh. Technically MXCSR and the x87 CW also need to be preserved. > > > > > > > > What if rather than treating the enclave as hostile we require it to be > > > > compliant with the x86-64 ABI like any other function? That would solve > > > > the EFLAGS.DF, MXCSR and x87 issues without adding unnecessary overhead. > > > > And we wouldn't have to save/restore R12-R15. It'd mean we couldn't use > > > > the stack's red zone to hold @regs and @e, but that's poor form anyways. > > > > > > Grr, except the processor crushes R12-R15, FCW and MXCSR on asynchronous > > > exits. But not EFLAGS.DF, that's real helpful. > > > > I can think of three options that are at least somewhat reasonable: > > > > 1) Save/restore MXCSR and FCW > > > > + 100% compliant with the x86-64 ABI > > + Callable from any code > > + Minimal documentation required > > - Restoring MXCSR/FCW is likely unnecessary 99% of the time > > - Slow > > > > 2) Clear EFLAGS.DF but not save/restore MXCSR and FCW > > > > + Mostly compliant with the x86-64 ABI > > + Callable from any code that doesn't use SIMD registers > > - Need to document deviations from x86-64 ABI > > > > 3) Require the caller to save/restore everything. > > > > + Fast > > + Userspace can pass all GPRs to the enclave (minus EAX, RBX and RCX) > > - Completely custom ABI > > - For all intents and purposes must be called from an assembly wrapper > > > > > > Option (3) actually isn't all that awful. RCX can be used to pass an > > optional pointer to a 'struct sgx_enclave_exception' and we can still > > return standard error codes, e.g. -EFAULT. > > Entering and exiting a syscall requires an assembly wrapper, and that > doesn't seem completely unreasonable. It's an easy bit of inline > assembly. The code I posted had a few typos (stupid AT&T syntax), but with those fixed the idea checks out. My initial reaction to a barebones ABI was that it would be a "documentation nightmare", but it's not too bad if it returns actual error codes and fills in a struct on exceptions instead of stuffing registers. And with the MXCSR/FCW issues it might actually be less documentation in the long run since we can simply say that all state is the caller's responsibility. I *really* like that we basically eliminate bikeshedding on which GPRs to pass to/from the enclave.
