psw/ae: blank out ld-linux.so interpretor path from AEs

[berrange]

The enclaves are getting built as ELF executables, and thus the linker will embed the current ld-linux.so path for the host OS environment in the binary:

$ readelf -a libsgx_pce.signed.so | grep interpreter
[Requesting program interpreter: /nix/store/xmprbk52mlcdsljz66m8yf7cf0xf36n1-glibc-2.38-44/lib/ld-linux-x86-64.so.2]

The SGX enclaves are never loaded using ld-linux.so, as SGX has custom code for loading enclaves in the required manner.

This embedded ld-linux.so path thus serves no functional purpose, while also making it harder to do a reproducible build of the enclaves outside of the NixOS environment.

This patch blanks out the NixOS interpretor path, by setting it to the empty string.

Fixes: #1040

Two alternatives I considered:

• Extending the strip command to add --remove-section=.interp. This has the undesirable side effect that 'readelf' will complain that the AE ELF file is malformed, because ELF executables are expected to have an .interp section
• Changing LDTFLAGS to add -Wl,--shared. This tells the linker to create an ELF file that is a shared library rather than an executable, and thus avoids creating a .interp section in the first place. This has a bunch of changes to the overall ELF file, however, and I was not confident that the result would be functionally correct for the SGX enclave loader.

I still believe that using -Wl,--shared is probably the more semantically logical approach, but using -Wl,--dynamic-loader, is the approach that has fewer chances of unexpected behaviour changes, hence I picked the latter here.

[berrange]

The DCAP counterpart of this is intel/confidential-computing.tee.dcap#435

[haitaohuang]

I realize we probably do not want to re-sign AEs for this change only if this affects the sigstruct (likely not but did not check)

[berrange]

I realize we probably do not want to re-sign AEs for this change only if this affects the sigstruct (likely not but did not check)

In Fedora context, I've got a local workaround to do reproducible builds by matching the current nixos path. So I'm fine if this only applies to future AEs, no need to re-sign anything that already exists IMHO.