Re: [PATCH v5 05/27] arm64: Use daifflag_restore after bp_hardening

[Julien Thierry]

On 12/09/18 13:28, James Morse wrote:
> On 12/09/18 12:11, Julien Thierry wrote:
> > On 12/09/18 11:32, James Morse wrote:
> > > On 28/08/18 16:51, Julien Thierry wrote:
> > > > For EL0 entries requiring bp_hardening, daif status is kept at
> > > > DAIF_PROCCTX_NOIRQ until after hardening has been done. Then interrupts
> > > > are enabled through local_irq_enable().
> > > >
> > > > Before using local_irq_* functions, daifflags should be properly restored
> > > > to a state where IRQs are enabled.
> > >
> > > > Enable IRQs by restoring DAIF_PROCCTX state after bp hardening.
> > >
> > > Is this just for symmetry, or are you going on to add something to the daifflags
> > > state that local_irq_* functions won't change? (if so, could you allude to that
> > > in the commit message)
>
> > What happens is that once we use ICC_PMR_EL1, local_irq_enable will not touch
> > PSR.I. And we are coming back from an entry where PSR.I was kept to 1 so
> > local_irq_enable was not actually enabling the interrupts. On the otherhand,
> > restore will affect both.
>
> Got it. Thanks!
>
> Does this mean stop_machine()s local_save_flags()/local_irq_restore() will not
> be symmetric around __apply_alternatives_multi_stop()?
> I see you add alternatives in these in patch 15, but I haven't got that far yet)

It's a good point but it should be fine.
The changes to the irqflags make save/restore takes everything into 
consideration (PMR + PSR.I) because of situtations like this, 
enable/disable only toggle the PMR (so the goal is to not have PSR.I set 
before reaching path calling enable/disable).
Maybe I should add a comment for this in asm/irqflags.f when I add the 
alternatives, so that at least arch code can be wary of this.

> > Another option is to have the asm macro "enable_da_f" also switch to PMR usage
> > (i.e. "just keep normal interrupts disabled"). Overall it would probably be
> > easier to reason with, but I'm just unsure whether it is acceptable to receive a
> > Pseudo NMI before having applied the bp_hardening.
>
> Wouldn't this give the interrupt controller a headache? I assume IRQs really are
> masked when handle_arch_irq is called. (I know nothing about the gic)

Yes, you're right, I missed that da_f gets unmasked right before the 
irq_handler... So unless I do some special case for irqs, enable_da_f is 
not the way to go.

Thanks,

--
Julien Thierry