Re: [PATCH RESEND] x86/speculation: Fix user-mode spectre-v2 protection with KERNEL_IBRS

On Mon, Feb 20, 2023 at 10:56:38AM -0800, KP Singh wrote:

Sure, it looks like an omission to me, we wrote a POC on Skylake that was able to do cross-thread training with the current set of mitigations.

Right.

STIBP with IBRS is still correct if spectre_v2=ibrs had really meant IBRS everywhere,

Yeah, IBRS everywhere got shot down as a no-no very early in the game, for apparent reasons.

but just means KERNEL_IBRS, which means only kernel is protected, userspace is still unprotected.

Yes, that was always the intent with IBRS: enable on kernel entry and disable on exit.

Thx.