[MODERATED] Re: [RFC PATCH 4/6] kvm: handle host mode irqs 4

[Nicolai Stange <nstange@suse.de>]

speck for Peter Zijlstra  <speck@linutronix.de> writes:

> On Sun, Jul 22, 2018 at 11:35:38AM +0200, speck for Nicolai Stange wrote:
>
>> Implement "kernel mode irq tracking" to address this.
>> 
>> Introduce a per-cpu generation count, kernel_mode_irq_gen. Its semantics
>> are defined as follows: if its values as read at two points in time compare
>> equal, then an interrupt has not happened inbetween while in kernel mode
>> on that CPU.
>> 
>> Let the 32 and 64 bit interrupt specific entry code increment this
>> generation counter. Note that the increment is done unconditionally to
>> safe a branch. This doesn't matter though because nobody is watching
>> anyway when interrupts are taken from userspace.
>> 
>> NMIs are not tracked -- non-racy usage of kernel_mode_irq_gen requires
>> interrupts to be off.
>
> Should this not be part of irq_cpustat_t ? And have you looked at how
> much is already covered by kstat_incr_irqs_this_cpu() ? Also note that
> struct kernel_stat already includes a count vector for all softirqs.

kstat_incr_irqs_this_cpu() covers anything coming through do_IRQ(),
provided that irq_domain implementations call it from their associated
struct irq_desc ->handle_irq(). Those I checked do, but there's no
guarantee AFAICT.

Anyway, if we were to reuse the existing stats you suggested above, then
- either softirqs only or
- softirqs + do_IRQ()
could get made to trigger L1d flushes.

I'd still opt for also including the non-do_IRQ() irqs though, because
- The number of additional L1d flushes would be limited: Paolo
  mentioned somewhere that the expected time between vmexit -> vmenter
  where ->l1tf_flush_l1d would not have been set already is ~2500 clock
  cycles.
- The non-do_IRQ() irqs wouldn't have to get audited for potentially
  pulling in any sensitive data.

So if nobody objects, I'll go with what Paolo proposed at [1]: replace
the 'kernel_mode_irq_gen' counter with a latch flag and set that for any
irq (but NMIs for a start).

Thanks,

Nicolai


[1] 25bdef0a-d65a-a8f8-020c-f9a2056be76f@redhat.com

-- 
SUSE Linux GmbH, GF: Felix Imendörffer, Jane Smithard, Graham Norton,
HRB 21284 (AG Nürnberg)