Re: [PATCH] [GIT PULL v2] x86: Workaround for NMI iret woes

[Ingo Molnar <mingo@elte.hu>]

* Steven Rostedt <rostedt@goodmis.org> wrote:

> > > +	pushq_cfi $repeat_nmi
> > > +
> > > +	/* Put stack back */
> > > +	addq $(11*8), %rsp
> 
> This is where we put the stack back to the original position. 
> Is CFI notation really necessary here?

i'd add it if it's not hard or ugly - in theory we could get a
#MC exception in that window.

> > Note that the IRQ return checks are needed because NMI path 
> > can set the irq-work TIF. Might be worth putting into the 
> > comment - NMIs are not *entirely* passive entities.
> 
> The NMI path can set the TIF flags? Then where should they be 
> processed. There was an assumption that NMIs shouldn't do 
> that. I could have been wrong with that. What work needs to be 
> done and when? This is the change that Linus made. If that's 
> the case, we need to work something else out.

Hm, you are right, we at most access them (for 32-bit compat 
checks for example) but don't modify them - we have switched to 
using the special irq work self-IPI.

So the change is fine.

> > Something like nmi_postprocess_retry_preprocess()?
> 
> Not sure what would be good, as i386 does the retry, x86_64 
> just switches the idt. The two archs do two different things. 
> The above name would be confusing as it doesn't match what 
> x86_64 does.

Yeah, that assymetry is bothering me too. I guess we can keep it 
as-is, no strong feelings. The whole thing *feels* fragile.

Thanks,

	Ingo