From: Linus Walleij

Sent: 03 March 2021 15:19

On Tue, Feb 23, 2021 at 11:05 AM Will Deacon [off-list ref] wrote:

quoted

On Mon, Feb 22, 2021 at 01:50:06PM -0800, Jian Cai wrote:

quoted

I am not sure if there are any plans to protect assembly code and I
will leave it to the Arm folks since they know a whole lot better. But
even without that part, we should still have better protection,
especially when overhead does not look too bad: I did some preliminary
experiments on ChromeOS, code size of vmlinux increased 3%, and there
were no noticeable changes to run-time performance of the benchmarks I
used.

If the mitigation is required, I'm not sure I see a lot of point in only
doing a half-baked job of it. It feels a bit like a box-ticking exercise,
in which case any overhead is too much.

I wrote some suggestions on follow-ups in my reply, and I can
help out doing some of the patches, I think.

Since ARM32 RET is mov pc, <>
git grep 'mov.*pc,' | wc -l gives 93 sites in arch/arm.
I suppose these need to come out:

mov pc, lr
dsb(nsh);
isb();

Won't that go horribly wrong for conditional returns?

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)