Re: [PATCH v36 21/24] x86/vdso: Implement a vDSO for Intel SGX enclave call

[Jarkko Sakkinen <jarkko.sakkinen@linux.intel.com>]

On Tue, Aug 18, 2020 at 05:52:41PM +0300, Jarkko Sakkinen wrote:
> On Mon, Aug 10, 2020 at 03:23:17PM -0700, Sean Christopherson wrote:
> > > This can be done implicitly by wrapping the struct
> > > sgx_enclave_exception in another structure and then using techniques
> > > like container_of() to find another field. However, this is made more
> > > difficult by the fact that the sgx_enclave_exit_handler_t is not
> > > really using the x86_64 sysv calling convention. Therefore, the
> > > sgx_enclave_exit_handler_t MUST be written in assembly.
> > 
> > What bits of the x86-64 ABI require writing the handler in assembly?  There
> > are certainly restrictions on what the handler can do without needing an
> > assembly trampoline, but I was under the impression that vanilla C code is
> > compatible with the exit handler patch.  Is Rust more picky about calling
> > convention?
> > 
> > Side topic, the documentation for vdso_sgx_enter_enclave_t is wrong, it
> > states the EFLAGS.DF is not cleared before invoking the handler, but that's
> > a lie.
> 
> If handler requires the use of setjmp/longjmp API for sudden exits, that
> is considered bad even with C++, as it is not compatible with stack
> unwinding. The handler has a lot of constraints for its environment, and
> is somewhat unappealing to use.
> 
> That's why I started today thinking a possibility of using a bpf program
> as a middle-man. BPF programs can be used to execute code by the kernel
> in behalf of user in a domain defined sandbox. The execution context is
> just a buffer passed in R1 to the BPF interpreter. It can be defined by
> application.

Something like

1. An exception is triggered.
2. Kernel executes an eBPF program behalf of the caller, if one was
   given.
3. vDSO calls a fixed exit handler that based on the outcome calls
   ERESUME/EENTER.

Possibly an ioctl could be used to attach an eBPF program to an
enclave and vDSO would only get a context struct.

/Jarkko