AMD Adds Basic RPC Mechanism To LLVM libc For GPUs

[phoronix]

Phoronix: AMD Adds Basic RPC Mechanism To LLVM libc For GPUs

AMD has upstreamed a basic RPC (remote procedure call) mechanism for GPU use to LLVM's libc and wired it up for AMDGPU use...

https://www.phoronix.com/news/AMDGPU-LLVM-libc-RPC

[cj.wijtmans]

Is that not a security risk?

[coder]

Is that not a security risk?

If there's one of these buffers for each host process using the GPU, then I think it should still be okay.

The host userspace process should be what's implementing the backend support for the libc functions, so its GPU code will effectively have the same privileges it does.

GPUs have MMUs on them and supposedly all of the security mechanisms needed to protect different processes' GPU code from each other, as well as arbitrary host memory. That should avoid one process using the GPU to stuff commands into the mailbox of another.

[cj.wijtmans]

If there's one of these buffers for each host process using the GPU, then I think it should still be okay.

The host userspace process should be what's implementing the backend support for the libc functions, so its GPU code will effectively have the same privileges it does.

GPUs have MMUs on them and supposedly all of the security mechanisms needed to protect different processes' GPU code from each other, as well as arbitrary host memory. That should avoid one process using the GPU to stuff commands into the mailbox of another.

I just think we should be very caareful because you never know when they force a vendor to implement a new backdoor when one gets patched.

[coder]

I just think we should be very caareful because you never know when they force a vendor to implement a new backdoor when one gets patched.

You should have a look at the code, then. My statements are 100% speculation.