Re: Dealing with the NMI mess

[Willy Tarreau <w@1wt.eu>]

On Thu, Jul 23, 2015 at 02:13:16PM -0700, Linus Torvalds wrote:
> On Thu, Jul 23, 2015 at 1:52 PM, Willy Tarreau <w@1wt.eu> wrote:
> >
> > What's the worst case that can happen with RF cleared when returing
> > to user space ?
> 
> Not a good idea. We are fine breaking breakpoints on the kernel ("use
> the tracing infrastructure instead"). Breaking it in user space is not
> really an option.

But that wouldn't disable the breakpoint, just make it strike again,
so the user would not be hurt.

> And we really don't need to. We'd only use 'ret' when returning to
> kernel code. And not even for the usual case, only for the "interrupts
> are off" case.  If somebody tries to put a breakpoint on something
> that is used in an irq-off situation, they are doing something very
> specialized, and we cna tell them: "sorry, we had to break your use
> case because it's crazy any other way".
> 
> Those kind of people are by definition not "users". They are mucking
> with kernel internals. Breaking them is not a regression.
> 
> Btw, we should still ask Intel for that "fast iret that doesn't
> re-enable NMI". So for possible future CPU's we might let people do
> crazy things again.

I'm just thinking that there should be an option for this : task switching.
You can store the EFLAGS in the TSS, so by preparing a dummy task with
everything needed to emulate iret, we might be able to do it without the
iret instruction. Or is this a stupid idea ? At least now I've well
understood that ugliness is not an excuse for not proposing something :-)

Willy