On Fri 2018-01-12 09:34:03, Linus Torvalds wrote:
> On Fri, Jan 12, 2018 at 5:23 AM, Arnd Bergmann <arnd@arndb.de> wrote:
> >
> > However, there is an important difference between KPTI and X86_4G:
> > The former unmaps the kernel pages from the user space page tables,
> > but keeps both the linear mapping and the user pages visible in
> > kernel mode, while the latter must have also unmapped user space
> > pages from kernel mode, requiring a more expensive get_user/put_user
> > implementation.
> 
> Indeed. And I think that the 4G:4G patches do things wrong.

Yeah. But if there's copy around for something recent, I'd still like
to see it.

> People are already complaining about the PTI costs. Separating user
> space entirely is much much worse, and makes all user accesses from
> kernel space too painful for words.
> 
> Honestly, I didn't merge the old 4G:4G patches originally, and I'm not
> going to merge them this time around either.

I'll try to do the right thing. OTOH... I don't like the fact that
kernel memory on my machine is currently readable, probably even from
javascript.

I tried disabling CPU caches. Just like that, off, boom. My system
will not survive that, and it looks like 100x slowdown. So 2x slowdown
would be an improvement (and 4G:4G can probably do better than that).

									Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html