Re: [PATCH] x86: Enable IBT in Rust if enabled in C

[Matthew Maurer <mmaurer@google.com>]

On Tue, Oct 10, 2023 at 7:24 AM Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Tue, Oct 10, 2023 at 07:06:32AM -0700, Matthew Maurer wrote:
>
> > > > +KBUILD_RUSTFLAGS += -Zcf-protection=branch -Zno-jump-tables
> > >
> > > One question, -Zcf-protection=branch, will that ever emit NOTRACK
> > > prefix? The kernel very explicitly does not support (enable) NOTRACK.
>
> > rustc does this via LLVM, so its code generation works very similarly to clang.
> > It does not create its own explicit NOTRACKs, but LLVM will by default
> > with just -Zcf-protection-branch.
> > I've linked a godbolt showing that at least for the basic case, your
> > no-jump-tables approach from clang ports over.
> > https://godbolt.org/z/bc4n6sq5q
> > Whether rust generates NOTRACK should end up being roughly equivalent
> > to whether clang generates it, and if LLVM gains a code generation
> > flag for NOTRACK being disallowed some day, we can pass that through
> > as well.
>
> IIRC C++ will also emit NOTRACK for things like catch/throw and other
> stack/scope unwinds. Obviously C doesn't have that, but does Rust? (as
> might be obvious, I *really* don't know the language).
>
That's fine - Rust does have stack/scope unwinds with the
`panic=unwind` strategy. In the kernel, we use `panic=abort` and are
unlikely to ever change this approach. There are a host of other
complications that come from unwinding without NOTRACK getting
involved :)

In case you find `catch_unwind` - this function only has an effect
with `panic=unwind`. When `panic=abort`, there's nothing analogous to
catch/throw anymore, and `catch_unwind` becomes a no-op.

Are there other features you expect might trigger NOTRACK?
> ISTR HJL had a GCC patch to force-disable NOTRACK, but I've no idea what
> happened to that.
>