Re: [patch 07/11] [PATCH v2 07/10] Linux Patch #7

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, 23 Apr 2018, speck for Jon Masters wrote:

> On 04/23/2018 02:05 PM, speck for Linus Torvalds wrote:
> > 
> > 
> > On Mon, 23 Apr 2018, speck for Konrad Rzeszutek Wilk wrote:
> >>
> >> I believe (and Linus, please correct me here) that the question of
> >> toggling on/off SPEC_CTRL MSR on user-space entrance is a no-go.
> > 
> > Absolutely. That would be entirely crazy.
> 
> Yep, totally nuts, and nobody is asking for that :)
> 
> > Is there a known situation where that would actually make sense?
> 
> Intel point out that globally disabling MD by default has typically a
> few percent hit, but sometimes up to 30%. Therefore, they want it to be
> per-process controllable. As I said, e.g. with a prctrl that happens to
> control an MSR underneath, but it's not userspace setting an MSR.

As we discussed before:

 1) Yet another patch site in the entry/exit path which if patched to
    functional contains yet another conditonal.

    The amount of NOOPs in entry/exit is horrendous already if this whole
    nonsense is patched out.

 2) The prctl is a handwavy idea. The semantics are blury at best. Is it
    opt-in or opt-out? Which processes should set it? What's the chance
    that the applications get actually patched? This is the ideal target
    for bitrot.

Anyway, if Intel thinks that this is desired, then they can send patches on
top of the big hammer (MD on/off globally) and provide the answers to #2
and proper numbers.

It's fully orthogonal and does not change the plan of having the global MD
on/off switch there ASAP and first.

Thanks,

	tglx