Re: [RFC PATCH] arm64: Enable data independent timing (DIT) in the kernel

[Catalin Marinas]

On Fri, Nov 04, 2022 at 10:29:10AM +0100, Ard Biesheuvel wrote:
> On Fri, 4 Nov 2022 at 09:09, Eric Biggers <ebiggers@xxxxxxxxxx> wrote:
> > On Thu, Oct 27, 2022 at 01:27:41PM +0200, Ard Biesheuvel wrote:
> > > Given that running privileged code with DIT disabled on a CPU that
> > > implements support for it may result in a side channel that exposes
> > > privileged data to unprivileged user space processes, let's enable DIT
> > > while running in the kernel if supported by all CPUs.
> >
> > This patch looks good to me, though I'm not an expert in low-level arm64 stuff.
> > It's a bit unfortunate that we have to manually create the .inst to enable DIT
> > instead of just using the assembler.  But it looks like there's a reason for it
> > (it's done for PAN et al. too), and based on the manual it looks correct.
> 
> Yes. The reason is that the assembler requires -march=armv8.2-a to be
> passed when using the DIT register (and similar requirements apply to
> the other registers). However, doing so may result in object code that
> can no longer run on pre-v8.2 cores, whereas the DIT accesses
> themselves are only emitted in a carefully controlled manner anyway,
> so keeping the arch baseline to v8.0 and using .inst is the cleanest
> way around this.

We worked around this already by defining asm-arch in
arch/arm64/Makefile to the latest that the assembler supports while
keeping the C compiler on armv8.0. Unlike the C compiler, the assembler
shouldn't generate new instructions unless specifically asked through
inline asm or .S files. We use this trick for MTE already (and LSE
atomics), though we needed another __MTE_PREAMBLE as armv8.5-a wasn't
sufficient for these instructions.

I think we ended up with .inst initially as binutils did not support
some of those instructions. We could try to clean them up but it's a bit
of a hassle to check which versions your binutils supports.

-- 
Catalin