Re: __schedule #DF splat

[Gleb Natapov <gleb@kernel.org>]

On Sun, Jun 29, 2014 at 12:31:50PM +0200, Jan Kiszka wrote:
> On 2014-06-29 12:24, Gleb Natapov wrote:
> > On Sun, Jun 29, 2014 at 11:56:03AM +0200, Jan Kiszka wrote:
> >> On 2014-06-29 08:46, Gleb Natapov wrote:
> >>> On Sat, Jun 28, 2014 at 01:44:31PM +0200, Borislav Petkov wrote:
> >>>>  qemu-system-x86-20240 [006] ...1  9406.484134: kvm_page_fault: address 7fffb62ba318 error_code 2
> >>>>  qemu-system-x86-20240 [006] ...1  9406.484136: kvm_inj_exception: #PF (0x2)a
> >>>>
> >>>> kvm injects the #PF into the guest.
> >>>>
> >>>>  qemu-system-x86-20240 [006] d..2  9406.484136: kvm_entry: vcpu 1
> >>>>  qemu-system-x86-20240 [006] d..2  9406.484137: kvm_exit: reason PF excp rip 0xffffffff8161130f info 2 7fffb62ba318
> >>>>  qemu-system-x86-20240 [006] ...1  9406.484138: kvm_page_fault: address 7fffb62ba318 error_code 2
> >>>>  qemu-system-x86-20240 [006] ...1  9406.484141: kvm_inj_exception: #DF (0x0)
> >>>>
> >>>> Second #PF at the same address and kvm injects the #DF.
> >>>>
> >>>> BUT(!), why?
> >>>>
> >>>> I probably am missing something but WTH are we pagefaulting at a
> >>>> user address in context_switch() while doing a lockdep call, i.e.
> >>>> spin_release? We're not touching any userspace gunk there AFAICT.
> >>>>
> >>>> Is this an async pagefault or so which kvm is doing so that the guest
> >>>> rip is actually pointing at the wrong place?
> >>>>
> >>> There is nothing in the trace that point to async pagefault as far as I see.
> >>>
> >>>> Or something else I'm missing, most probably...
> >>>>
> >>> Strange indeed. Can you also enable kvmmmu tracing? You can also instrument
> >>> kvm_multiple_exception() to see which two exception are combined into #DF.
> >>>
> >>
> >> FWIW, I'm seeing the same issue here (likely) on an E-450 APU. It
> >> disappears with older KVM (didn't bisect yet, some 3.11 is fine) and
> >> when patch-disabling the vmport in QEMU.
> >>
> >> Let me know if I can help with the analysis.
> >>
> > Bisection would be great of course. Once thing that is special about
> > vmport that comes to mind is that it reads vcpu registers to userspace and
> > write them back. IIRC "info registers" does the same. Can you see if the
> > problem is reproducible with disabled vmport, but doing "info registers"
> > in qemu console? Although trace does not should any exists to userspace
> > near the failure...
> 
> Yes, info registers crashes the guest after a while as well (with
> different backtrace due to different context).
> 
Oh crap. Bisection would be most helpful. Just to be absolutely sure
that this is not QEMU problem: does exactly same QEMU version work with
older kernels?

--
			Gleb.