On Wed, Jan 31, 2018 at 2:10 PM, David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote:
>
> Given that IBRS_ALL is supposed to be a sanely-performing option, I'd
> rather convince Intel to just make it unconditional. If they've added
> the appropriate tagging to the BTB, why even *have* this deliberately
> insecure mode when IBRS==0?
>
> I understand that until/unless they get a *proper* fix, software is
> still going to have to use IBPB as appropriate. But there's no need for
> the IBRS bit to do *anything*.
Amen, brother!
Please please please can Amazon and friends push this? The current
situation with IBRS_ALL is complete nasty horrible garbage. It's
pointless on current CPU's, and it's not well-defined enough on future
CPU's.
The whole "you can enable this, but performance may or may not be
acceptable, and we won't tell you" thing is some bad mumbo-jumbo.
Before IBRS is good, we'll do retpoline and BTB stuffing and have
those (hopefully very rare) IBPB's. So the whole "badly performing
IBRS_ALL" is completely pointless, and actively wrong.
Linus
