Re: [Xen-devel] [PATCH v2] x86/hvm: fix interaction between internal and external emulation

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 05.12.17 at 15:11, <Paul.Durrant@xxxxxxxxxx> wrote:
>> From: Xen-devel [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxxx] On Behalf
>> Of Paul Durrant
>> Sent: 05 December 2017 14:00
>> To: 'Jan Beulich' <JBeulich@xxxxxxxx>
>> > From: Jan Beulich [mailto:JBeulich@xxxxxxxx]
>> > Sent: 05 December 2017 13:53
>> > >>> On 28.11.17 at 15:05, <paul.durrant@xxxxxxxxxx> wrote:
>> > > --- a/xen/arch/x86/hvm/io.c
>> > > +++ b/xen/arch/x86/hvm/io.c
>> > > @@ -88,7 +88,7 @@ bool
>> > hvm_emulate_one_insn(hvm_emulate_validate_t *validate, const char
>> > *descr)
>> > >
>> > >      rc = hvm_emulate_one(&ctxt);
>> > >
>> > > -    if ( hvm_vcpu_io_need_completion(vio) || vio->mmio_retry )
>> > > +    if ( hvm_vcpu_io_need_completion(vio) )
>> > >          vio->io_completion = HVMIO_mmio_completion;
>> > >      else
>> > >          vio->mmio_access = (struct npfec){};
>> >
>> > While I can't (yet) say why without this change things would have
>> > behaved better on that old AMD box which is causing the osstest
>> > failure, I think Andrew's suggestion that we might be trying to
>> > emulate from a stale instruction cache is spot on: Doesn't
>> 
>> Yes, I can't see how the above was ever correct.
> 
> I think I see why this worked before...
> 
> Setting up the io_completion value meant that when hvm_do_resume() called 
> handle_hvm_io_completion() there was apparently an mmio outstanding and thus 
> handle_mmio() was called. At some point handle_mmio() has become a static 
> inline that calls hvm_emulate_one_insn() and that took care of the remaining 
> reps.

That would have worked sometimes, but as we've recently clarified
handle_hvm_io_completion() won't be called every time before
exiting back to the guest. IOW I would have expected random
failures with the same pattern as we see now, but I don't recall
ever having seen such before. (Otoh I admit I don't always look
closely when a transient failure occurs and then disappears in the
next flight.)

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel