On Tue, Mar 17, 2020 at 12:28:58PM -0400, Nathaniel McCallum wrote: > On Mon, Mar 16, 2020 at 6:53 PM Sean Christopherson > <sean.j.christopherson@xxxxxxxxx> wrote: > > > > On Mon, Mar 16, 2020 at 11:38:24PM +0200, Jarkko Sakkinen wrote: > > > On Mon, 2020-03-16 at 10:01 -0400, Nathaniel McCallum wrote: > > > > On Mon, Mar 16, 2020 at 9:56 AM Jarkko Sakkinen > > > > <jarkko.sakkinen@xxxxxxxxxxxxxxx> wrote: > > > > > On Sun, 2020-03-15 at 13:53 -0400, Nathaniel McCallum wrote: > > > > > > On Sat, Mar 14, 2020 at 9:25 PM Jarkko Sakkinen > > > > > > <jarkko.sakkinen@xxxxxxxxxxxxxxx> wrote: > > > > > > > On Wed, Mar 11, 2020 at 01:30:07PM -0400, Nathaniel McCallum wrote: > > > > > > > > Currently, the selftest has a wrapper around > > > > > > > > __vdso_sgx_enter_enclave() which preserves all x86-64 ABI callee-saved > > > > > > > > registers (CSRs), though it uses none of them. Then it calls this > > > > > > > > function which uses %rbx but preserves none of the CSRs. Then it jumps > > > > > > > > into an enclave which zeroes all these registers before returning. > > > > > > > > Thus: > > > > > > > > > > > > > > > > 1. wrapper saves all CSRs > > > > > > > > 2. wrapper repositions stack arguments > > > > > > > > 3. __vdso_sgx_enter_enclave() modifies, but does not save %rbx > > > > > > > > 4. selftest zeros all CSRs > > > > > > > > 5. wrapper loads all CSRs > > > > > > > > > > > > > > > > I'd like to propose instead that the enclave be responsible for saving > > > > > > > > and restoring CSRs. So instead of the above we have: > > > > > > > > 1. __vdso_sgx_enter_enclave() saves %rbx > > > > > > > > 2. enclave saves CSRs > > > > > > > > 3. enclave loads CSRs > > > > > > > > 4. __vdso_sgx_enter_enclave() loads %rbx > > > > > > > > > > > > > > > > I know that lots of other stuff happens during enclave transitions, > > > > > > > > but at the very least we could reduce the number of instructions > > > > > > > > through this critical path. > > > > > > > > > > > > > > What Jethro said and also that it is a good general principle to cut > > > > > > > down the semantics of any vdso as minimal as possible. > > > > > > > > > > > > > > I.e. even if saving RBX would make somehow sense it *can* be left > > > > > > > out without loss in terms of what can be done with the vDSO. > > > > > > > > > > > > Please read the rest of the thread. Sean and I have hammered out some > > > > > > sensible and effective changes. > > > > > > > > > > Have skimmed through that discussion but it comes down how much you get > > > > > by obviously degrading some of the robustness. Complexity of the calling > > > > > pattern is not something that should be emphasized as that is something > > > > > that is anyway hidden inside the runtime. > > > > > > > > My suggestions explicitly maintained robustness, and in fact increased > > > > it. If you think we've lost capability, please speak with specificity > > > > rather than in vague generalities. Under my suggestions we can: > > > > 1. call the vDSO from C > > > > 2. pass context to the handler > > > > 3. have additional stack manipulation options in the handler > > > > > > > > The cost for this is a net 2 additional instructions. No existing > > > > capability is lost. > > > > > > My vague generality in this case is just that the whole design > > > approach so far has been to minimize the amount of wrapping to > > > EENTER. > > > > Yes and no. If we wanted to minimize the amount of wrapping around the > > vDSO's ENCLU then we wouldn't have the exit handler shenanigans in the > > first place. The whole process has been about balancing the wants of each > > use case against the overall quality of the API and code. > > > > > And since this has been kind of agreed by most of the > > > stakeholders doing something against the chosen strategy is > > > something I do hold some resistance. > > > > Up until Nathaniel joined the party, the only stakeholder in terms of the > > exit handler was the Intel SDK. > > I would hope that having additional stakeholders would ease the path > to adoption. > > > There was a general consensus to pass > > registers as-is when there isn't a strong reason to do otherwise. Note > > that Nathaniel has also expressed approval of that approach. > > I still approve that approach. > > > So I think the question that needs to be answered is whether the benefits > > of using %rcx instead of %rax to pass @leaf justify the "pass registers > > as-is" guideline. We've effectively already given this waiver for %rbx, > > as the whole reason why the TCS is passed in on the stack instead of via > > %rbx is so that it can be passed to the exit handler. E.g. the vDSO > > could take the TCS in %rbx and save it on the stack, but we're throwing > > the baby out with the bathwater at that point. > > > > The major benefits being that the vDSO would be callable from C and that > > the kernel could define a legitimate prototype instead of a frankenstein > > prototype that's half assembly and half C. For me, those are significant > > benefits and well worth the extra MOV, PUSH and POP. For some use cases > > it would eliminate the need for an assembly wrapper. For runtimes that > > need an assembly wrapper for whatever reason, it's probably still a win as > > a well designed runtime can avoid register shuffling in the wrapper. And > > if there is a runtime that isn't covered by the above, it's at worst an > > extra MOV. > > > Guys, maybe it is just enough discussing. I see things go in circles at least. Just send a patch against current tree and we'll look into it then? I'm a strong believer of "good enough" well, in everything in life. With a legit patch it is easier to evaluate if what we get is just a different version of good enough, or perhaps we might get some useful value out of it. If you think that you get together something C callable, please *prove* that by also updating the self test. Fair enough? /Jarkko
