Rewrite the SGX backend as an elf loader #478
Conversation
Signed-off-by: Nathaniel McCallum <nathaniel@congru.us>
This makes it easier to reason about. Signed-off-by: Nathaniel McCallum <nathaniel@congru.us>
This makes it clearer where the actual entrypoint into the program is. Signed-off-by: Nathaniel McCallum <nathaniel@congru.us>
|
|
||
| /// This segment contains unmeasured pages. | ||
| #[cfg(feature = "backend-sgx")] | ||
| pub const PF_ENARX_SGX_UNMEASURED: u32 = 1 << 21; |
There was a problem hiding this comment.
Where is this used in the linker script?
There was a problem hiding this comment.
It isn't (yet?). But I want it well defined in case we choose to use this strategy in the shim in the future. That way it won't require a new server side. Only client side changes will be required.
| code.region().into(), | ||
| ); | ||
| // Find the offset for loading the code. | ||
| let slot = Span::from(shim.find_header(PT_ENARX_CODE).unwrap().vm_range()); |
There was a problem hiding this comment.
Maybe add an anyhow context to the error and return with ? instead of unwrap(). This makes debugging a little bit easier.
Same for the rest.
There was a problem hiding this comment.
I can do that now. However, my next change is to pull out the common behavior between the backends and make the Backend API more intelligent. This was going to include better error handling processes. Do you want to wait for that? Or do you want me to improve error handling now?
This is defensive in case the host manages to resume the enclave during an exit() syscall. If so, this syscall will be repeated endlessly. Signed-off-by: Nathaniel McCallum <nathaniel@congru.us>
This matches the behavior under kvm. Signed-off-by: Nathaniel McCallum <nathaniel@congru.us>
Signed-off-by: Nathaniel McCallum <nathaniel@congru.us>
| tcs0 PT_LOAD FLAGS(1 << 20); /* PF_ENARX_SGX_TCS */ | ||
| ssa0 PT_LOAD; | ||
|
|
||
| exec 0x634A0003 FLAGS(0); /* PT_ENARX_EXEC */ |
There was a problem hiding this comment.
PT_ENARX_EXEC or PT_ENARX_CODE or PT_ENARX_INIT ??
There was a problem hiding this comment.
I propose we rename the "code" layer to the executable or "exec" layer.
Merge these first: