Hi Tyler,

On 08/05/17 20:59, Baicar, Tyler wrote:

On 5/8/2017 11:28 AM, James Morse wrote:

quoted

I was tidying up the masking/unmasking in entry.S, something I wasn't aware of
that leads to a bug:
entry.S will unmask interrupts for instruction/data aborts that came from a
context with interrupts enabled. This makes sense for get_user() and friends...
For do_sea() we pull nmi_enter() as this can interrupt interrupts-masked code,
such as APEI, but if we end up in here with interrupts unmasked we can take an
IRQ from this 'NMI' context, which will inherit the in_nmi() and could lead to
the deadlock we were originally trying to avoid.

Teaching entry.S to spot external aborts is messy. I think the two choices are
to either mask interrupts when calling nmi_enter() (as these things should be
mutually exclusive), or to conditionally call nmi_enter() based on
interrupts_enabled(regs). I prefer the second one as it matches the notify_sea()
while interruptible that happens when KVM takes one of these.

Thinking about this some more: the KVM case is different as we know it was a
guest that triggered the external abort. Nothing the host kernel does is likely
to trigger either the same error or a related one.

But I can't think of a way this would trip twice on the host... yes your
suggestion looks fine.

(When we add SError/SEI support too we will need to change it as SEA may
interrupt SEI, and nmi_enter() has a BUG_ON(in_nmi()), so this nesting will need
explicitly checking.)


Thanks,

James