Re: [RFC PATCH] x86/speculation: Add finer control for when to issue IBPB

[Anand K. Mistry]

> >
> > Signed-off-by: Anand K Mistry <amistry@xxxxxxxxxx>
> > Signed-off-by: Anand K Mistry <amistry@xxxxxxxxxxxx>
>
> Two SoBs by you, why?

Tooling issues probably. Not intentional.

>
> > ---
> > Background:
> > IBPB is slow on some CPUs.
> >
> > More detailed background:
> > On some CPUs, issuing an IBPB can cause the address space switch to be
> > 10x more expensive (yes, 10x, not 10%).
>
> Which CPUs are those?!

AMD A4-9120C. Probably the A6-9220C too, but I don't have one of those
machines to test with,

>
> > On a system that makes heavy use of processes, this can cause a very
> > significant performance hit.
>
> You're not really trying to convince reviewers for why you need to add
> more complexity to an already too complex and confusing code. "some
> CPUs" and "can cause" is not good enough.

On a simple ping-ping test between two processes (using a pair of
pipes), a process switch is ~7us with IBPB disabled. But with it
enabled, it increases to around 80us (tested with the powersave CPU
governor).

On Chrome's IPC system, a perftest running 50,000 ping-pong messages:
without IBPB    5579.49 ms
with IBPB        21396 ms
(~4x difference)

And, doing video playback in the browser (which is already very
optimised), the IBPB hit turns out to be ~2.5% of CPU cycles. Doing a
webrtc video call (tested using http://appr.tc), it's ~9% of CPU
cycles. I don't have exact numbers, but it's worse on some real VC
apps.

>
> > I understand this is likely to be very contentious. Obviously, this
> > isn't ready for code review, but I'm hoping to get some thoughts on the
> > problem and this approach.
>
> Yes, in the absence of hard performance data, I'm not convinced at all.

With this change, I can get a >80% reduction in CPU cycles consumed by
IBPB. A  video call on my test device goes from ~9% to ~0.80% cycles
used by IBPB. It doesn't sound like much, but it's a significant
difference on these devices.