Re: [PATCH] x86/speculation: Use Indirect Branch Prediction Barrier in context switch

From: Josh Poimboeuf
Date: Tue Jan 30 2018 - 22:59:42 EST


On Tue, Jan 30, 2018 at 01:23:17PM -0800, Tim Chen wrote:
> On 01/30/2018 09:48 AM, Josh Poimboeuf wrote:
> > On Mon, Jan 29, 2018 at 10:04:47PM +0000, David Woodhouse wrote:
> >> From: Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx>
> >>
> >> Flush indirect branches when switching into a process that marked itself
> >> non dumpable. This protects high value processes like gpg better,
> >> without having too high performance overhead.
> >
> > I wonder what the point of this patch is. An audit of my laptop shows
> > only a single user of PR_SET_DUMPABLE: systemd-coredump.
>
> This is an opt in approach. For processes who need extra
> security, it set itself as non-dumpable. Then it can
> ensure that it doesn't see any poisoned BTB.

I don't want other users reading my applications' memory.

I don't want other containers reading my containers' memory.

I don't want *any* user tasks reading root daemons' memory.

Those are not unreasonable expectations.

So now I have to go and modify all my containers and applications to set
PR_SET_DUMPABLE? That seems highly impractical and unlikely.

Plus, I happen to *like* core dumps.

The other option is to rebuild the entire userland with retpolines, but
again, that would make this patch completely pointless.

> > [ And yes, I have gpg-agent running. Also, a grep of the gnupg source
> > doesn't show any evidence of it being used there. So the gpg thing
> > seems to be a myth. ]
>
> I'm less familiar with gpg-agent. Dave was the one who
> put in comments about gpg-agent in this patch so perhaps
> he can comment.
>
> >
> > But also, I much preferred the original version of the patch which only
> > skipped IBPB when 'prev' could ptrace 'next'.
>
> For the A->kernel thread->B scenario, you will need context of A
> to decide if you need IBPB when switching to B. You need to
> worry about whether the context of A has been released ... etc if
> you want to use ptrace.

Is that why the ptrace approach was abandoned? Surely that's a solvable
problem? We have some smart people on lkml. And anyway I didn't see it
discussed anywhere. In the worst case we could just always do IBPB when
switching between kernel and user tasks.

--
Josh