Re: [PATCH 2/2] x86/speculation: Support "Enhanced IBRS" on future CPUs

[David Woodhouse <dwmw2@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 2796 bytes --]

On Fri, 2018-02-16 at 12:04 +0100, Paolo Bonzini wrote:
> On 16/02/2018 11:21, David Woodhouse wrote:
> > 
> > Why? With IBRS_ALL the guest *never* gets to affect the actual hardware
> > MSR, which is always on. The MSR is purely an emulated no-op. Why does
> > that affect migration?
> Because even if the host has IBRS_ALL, as long as you want to migrate to
> a system without IBRS_ALL the guest will likely not have it.  You can
> fake IBRS_ALL on the older system after migration, and forcing the guest
> to always run with IBRS=1 even when in user mode; that is slow.  Or...

No you can't; it's also a barrier. You can't just leave it on.

> > Even if the guest doesn't have/support IBRS_ALL, and is frobbing the
> > (now emulated) MSR on every kernel entry/exit, that's *still* going to
> > be a metric shitload faster than what it *thought* it was doing.
>
> ... you are making every kernel entry/exit 3 times slower by adding two
> KVM exits (both hypervisor traps and syscalls are in the 1000-1500 clock
> cycles ballpark).  That cannot be fast at all.

I'm not convinced I care. It's still on a par with the performance of
*actually* frobbing IBRS every time. If the guest cares about
performance, they'll be using retpoline instead and not doing that.

We really don't want to penalise the *host*, and other guests, for one
guest which does stupid things.

And if we have a conditional 'set IBRS back on because we're using
IBRS_ALL and the guest had access to it' in the vmexit path, only the
*first* clause (IBRS_ALL) of that condition can be done with
alternatives. The other part is necessarily a runtime thing, and thus
needs the 'else lfence' to make sure it really happens, penalising
*all* guests. (Unless there's something else guaranteed to be
serialising in that path but after Arjan mentioned it last time I took
a quick look and didn't see anything unconditional).

An alternative would be to put the SPEC_CTRL MSR into the list of MSRs 
to be automatically saved/reset at vmexit, but we've already carefully
*changed* from doing that for the non-IBRS_ALL case because doing it
manually in the host is faster. I don't know that we want that
additional complexity — that might be the last straw that makes us tell
Intel "no, in that case we don't want IBRS_ALL as it is. Define a new
bit which is like IBRS_ALL but also promises that it's always like that
and the IBRS bit in the MSR is a no-op". That new bit would be set on
all future hardware except a small handful of current chips, I believe.

I think we can live with trapping and emulating SPEC_CTRL for the
guests which use it, for now.

If we *really* want to explore optimising it somehow, there's nothing
to prevent us doing that in a subsequent patch.

[-- Attachment #2: smime.p7s --]
[-- Type: application/x-pkcs7-signature, Size: 5213 bytes --]