Re: [PATCH] x86,kvm: move qemu/guest FPU switching out to vcpu_run

[Rik van Riel <riel@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2398 bytes --]

On Tue, 2017-11-14 at 20:40 +0100, David Hildenbrand wrote:
> On 14.11.2017 19:07, Rik van Riel wrote:
> > On Tue, 2017-11-14 at 17:57 +0100, David Hildenbrand wrote:
> > > 
> > > > diff --git a/arch/x86/include/asm/kvm_host.h
> > > > b/arch/x86/include/asm/kvm_host.h
> > > > index c73e493adf07..92e66685249e 100644
> > > > --- a/arch/x86/include/asm/kvm_host.h
> > > > +++ b/arch/x86/include/asm/kvm_host.h
> > > 
> > > We should also get rid of guest_fpu_loaded now, right?
> > 
> > Indeed, we no longer need that member. I'll get rid of it.
> > 
> > > emulator_get_fpu() does a kvm_load_guest_fpu(). Doesn't that mean
> > > that
> > > this is now not needed anymore? (at least when emulator code is
> > > called
> > > from inside the loop?)
> > 
> > Now that is a very good question!
> > 
> > When called from inside the loop, it is indeed not
> > needed.
> > 
> > My question is, can the in-kernel emulator code ever
> > be called from OUTSIDE the KVM_RUN ioctl loop?
> > 
> > If so, we need to restore the user FPU context before
> > returning from the emulator code. Given that the current
> > emulator code does not do that, I suspect this is not
> > the case. I also see no path from the kvm ioctl into
> > the emulator code, other than via KVM_RUN.
> > 
> > The FPU and XSAVE ioctls all work on the saved
> > vcpu->arch.guest_fpu data, and never directly on the
> > registers.
> > 
> > Looks like we can completely get rid of .get_fpu and
> > .put_fpu...
> > 
> > Unless Paolo has any objection, I'll go do that :)
> 
> 
> I think we should check all get/put_fpu callers if they need
> preempt_disable().
> 
> E.g. em_fxrstor() needs disabled preemption as we temporarily
> save + restore some host register (via fxsave + fxrstor) under some
> circumstances that are not saved/restored when switching to/back from
> another process. We should double check.
> 
> @Paolo what about complete_userspace_io? It can end up calling
> emulate_instruction(). So maybe we have to move load/put fpu further
> out
> or add special handling.

It looks like all complete_userspace_io causes is for
the vcpu_run loop to exit, and return to userspace
from the KVM_RUN ioctl code.

In other words, the userspace qemu FPU context should
be restored before we return to userspace, even with
my patch (v2 on the way).

-- 
All rights reversed

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 473 bytes --]