Re: [RFC PATCH v2 6/6] x86/entry/pti: don't switch PGD on when pti_disable is set

[Andy Lutomirski]

On Wed, Jan 10, 2018 at 11:50 AM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Wed, Jan 10, 2018 at 11:21 AM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>
>> I really dislike state that isn't cleared on execve().  I'm assuming
>> that this is so you can run time pwn_me_without_pti whatever?  Surely
>> LD_PRELOAD can do this, too?
>
> Andy, what the hell is wrong with you?
>
> You are actively trying to screw this whole interface up, aren't you?
>
> LD_PRELOAD cannot work for a wrapper, for the simple reason that it
> runs in the same context as the process. So if you want to say "I want
> to run this process without PTI", but you don't want to run the
> process with elevated privileges, LD_PRELOAD doesn't work.

Oh, right, duh.  Brain was off.

> The "let's do it per thread" made no sense either, since that's
> fundamentally not how page tables work, and it's complete broken shit.

I still disagree with you here.  The whole concept of per-thread or
per-mm or per-whatever PTI disablement is if the admin for some reason
trusts some piece of code not to try to exploit Meltdown.  But just
imagine a program like a web browser.  The browser will do some
performance critical stuff (networking) and some
absolutely-no-fucking-way-would-I-turn-off-PTI stuff (running
scripts).  So per-thread seems totally sensible to me.  No one sane
would ever do this for a web browser, but I can easily imagine it for
something like a web *server* or even a database server.

Just logically, too, per-thread is the obvious semantics.  Whether we
rewrite CR3 when we go to usermode is a thing affecting that thread.
The only reason the mm has anything to do with it is the NX trick.