On Sat, Mar 21, 2020 at 02:55:28AM +0200, Jarkko Sakkinen wrote: > On Fri, Mar 20, 2020 at 04:25:12PM -0700, Sean Christopherson wrote: > > On Fri, Mar 20, 2020 at 02:57:24AM +0200, Jarkko Sakkinen wrote: > > > On Wed, Mar 18, 2020 at 06:11:22PM -0700, Sean Christopherson wrote: > > > > Sean Christopherson (8): > > > > x86/sgx: vdso: Remove an incorrect statement the enter enclave comment > > > > x86/sgx: vdso: Make the %rsp fixup on return from handler relative > > > > x86/sgx: vdso: Make __vdso_sgx_enter_enclave() callable from C code > > > > x86/sgx: vdso: Define a typedef for __vdso_sgx_enter_enclave > > > > selftests/x86: sgx: Zero out @result before invoking vDSO sub-test > > > > selftests/x86: sgx: Pass EENTER to vDSO wrapper instead of hardcoding > > > > selftests/x86: sgx: Stop clobbering non-volatile registers > > > > selftests/x86: Add selftest to invoke __vsgx_enter_enclave() from C > > > > > > > > arch/x86/entry/vdso/vsgx_enter_enclave.S | 72 ++----------------- > > > > arch/x86/include/uapi/asm/sgx.h | 61 ++++++++++++++++ > > > > .../selftests/x86/sgx/encl_bootstrap.S | 6 +- > > > > tools/testing/selftests/x86/sgx/main.c | 17 ++++- > > > > tools/testing/selftests/x86/sgx/sgx_call.S | 1 - > > > > tools/testing/selftests/x86/sgx/sgx_call.h | 2 +- > > > > 6 files changed, 85 insertions(+), 74 deletions(-) > > > > > > > > -- > > > > 2.24.1 > > > > > > > > > > Might be a grazy idea but better to go through this anyway. > > > > > > Given the premise that we need the carry the callback anyway in all > > > cases, why won't just have the callback. > > > > > > Why we absolutely need the code path that fills exception info given > > > that we no matter what need to have a callback route? > > > > > > Would simplify considerably to have only clear flow. > > > > Invoking the callback uses a retpoline, which is non-trivial overhead. > > For runtimes that need an assembly wrapper for other reasons, and aren't > > using the untrusted stack, forcing them to implement a handler would be > > painful. > > The non-callback route only exists because we did not know that we have > to do the callback route. It does not make sense to add something for > any other reason than absolute necessity. The non-callback route exists because it's simpler for the runtime to use if it already has an asm wrapper and doesn't do stack shenanigans, e.g. the runtime doesn't need to play games to get context information when handling an exit. And using a callback in conjunction with calling the vDSO from C is by no means necessary, e.g. if the runtime is happy to die on exceptions, or if it's doing clever clobbering, function annotation, longjump(), etc... to avoid consuming corrupted state on an exception. > Things would simplify in the vDSO implementation considerably. Now it is > overwhelmingly complex for no good reason. No it wouldn't, it literally saves zero instructions (unless the vDSO wanted to crash the caller on a NULL pointer exception). It'd do nothing more than move this code: /* Invoke userspace's exit handler if one was provided. */ .Lhandle_exit: cmp $0, 0x20(%rbp) jne .Linvoke_userspace_handler up to the beginning of the code as /* Return -EINVAL if there's no userspace exit handler. */ cmp $0, 0x20(%rbp) je .Linvalid_input and the exit handler invocation would get inlined. Unless I'm missing some clever massaging of code, that's it. At best the above could be dropped, but that's a whopping two instructions and a single uop.
