Re: UNTRAIN_RET in native_irq_return_ldt

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Jul 12, 2022 at 08:20:44PM +0200, Alexandre Chartre wrote:
> 
> Hi,
> 
> I think there is an issue in native_irq_return_ldt: UNTRAIN_RET is used and can
> clobber %rax which is expected to be the user rax.

and cx and dx..

> A simple fix would be to preserve %rax:
> 
> diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
> index a4ba162e52c3..f1fe05289d84 100644
> --- a/arch/x86/entry/entry_64.S
> +++ b/arch/x86/entry/entry_64.S
> @@ -728,7 +728,11 @@ native_irq_return_ldt:
>         pushq   %rdi                            /* Stash user RDI */
>         swapgs                                  /* to kernel GS */
>         SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi   /* to kernel CR3 */
> +
> +       /* UNTRAIN_RET can clobber %rax, so preserve it */
> +       movq    %rax, %rdi
>         UNTRAIN_RET
> +       movq    %rdi, %rax
>         movq    PER_CPU_VAR(espfix_waddr), %rdi
>         movq    %rax, (0*8)(%rdi)               /* user RAX */
> 

Which still leaves cx and dx scrambled.

> But I wonder if we really need to use UNTRAIN_RET in native_irq_return_ldt because
> I think we reach this point from the kernel after untrain has already be done,
> and it looks like we don't do ret afterward (the code just fixup the stack and
> then iret).

Yes, I think removing it is fine, the objtool unret validation also
doesn't complain about it not being there (I really should have written
that validation before doing these patches, not after, oh well).