[Qemu-devel] Re: qemu-kvm vs. qemu: Terminate cpu loop on reset?

[Gleb Natapov <gleb@redhat.com>]

On Fri, Jan 07, 2011 at 07:24:00PM +0100, Jan Kiszka wrote:
> Am 07.01.2011 18:53, Gleb Natapov wrote:
> > On Fri, Jan 07, 2011 at 06:30:57PM +0100, Jan Kiszka wrote:
> >> Am 07.01.2011 18:16, Gleb Natapov wrote:
> >>> On Fri, Jan 07, 2011 at 05:59:34PM +0100, Jan Kiszka wrote:
> >>>> Am 07.01.2011 17:53, Gleb Natapov wrote:
> >>>>> On Fri, Jan 07, 2011 at 04:57:31PM +0100, Jan Kiszka wrote:
> >>>>>> Hi,
> >>>>>>
> >>>>>> does anyone immediately know if this hunk from vl.c
> >>>>>>
> >>>>>> @@ -1278,6 +1197,10 @@ void qemu_system_reset_request(void)
> >>>>>>      } else {
> >>>>>>          reset_requested = 1;
> >>>>>>      }
> >>>>>> +    if (cpu_single_env) {
> >>>>>> +        cpu_single_env->stopped = 1;
> >>>>>> +        cpu_exit(cpu_single_env);
> >>>>>> +    }
> >>>>>>      qemu_notify_event();
> >>>>>>  }
> >>>>>>
> >>>>>> is (semantically) relevant for upstream as well? IIUC, it ensures that
> >>>>>> the kvm cpu loop is not continued if an IO access called into
> >>>>>> qemu_system_reset_request.
> >>>>>>
> >>>>> I don't know TCG enough to tell. If TCG can continue vcpu execution
> >>>>> after io without checking reset_requested then it is relevant for
> >>>>> upstream too.
> >>>>
> >>>> I was first of all thinking about kvm upstream, but their handling
> >>>> differ much less upstream than in current qemu-kvm. Anyway, need to dig
> >>>> into the details.
> >>>>
> >>>>>
> >>>>>> If yes, then it would be a good time to push a patch: these bits will
> >>>>>> fall to dust on next merge from upstream (vl.c no longer has access to
> >>>>>> the cpu state).
> >>>>>>
> >>>>> On a next merge cpu state will have to be exposed to vl.c then. This
> >>>>> code cannot be dropped in qemu-kvm.
> >>>>
> >>>> I think a cleaner approach, even if it's only temporarily required, is
> >>>> to move that code to cpus.c. That's likely also the way when we need it
> >>>> upstream. 
> >>> It doesn't matter where the code resides as long as it is called on
> >>> reset.
> >>
> >> It technically matters for the build process (vl.c is built once these
> >> days, cpus.c is built per target).
> >>
> > Yes, I understand the build requirement. Runtime behaviour should not
> > change.
> 
> Yep, for sure.
> 
> BTW, the self-IPI on pending exit request is there for a reason I but.
> In order to complete half-done string-io or something like that? Would
> be the next patch for upstream then.
> 
The (documented) rule of KVM is that if exit to userspace happens
during instruction emulation KVM_RUN has to be called again to complete
instruction emulation.

> > 
> >> In any case, we apparently need to fix upstream, I'm playing with some
> >> approach.
> >>
Note to self: need to write unit test to check that vcpu is not executed
after it issues reset by doing pio.

> >>>
> >>>>            If upstream does not need it, we have to understand why and
> >>>> maybe adopt its pattern (the ultimate goal is unification anyway).
> >>>>
> >>> I don't consider kvm upstream as working product. The goal should be
> >>> moving to qemu-kvm code in upstream preserving all the knowledge we
> >>> acquired while making it production grade code.
> >>
> >> We had this discussion before. My goal remains to filter the remaining
> >> upstream fixes out of the noise, adjust both versions so that they are
> >> apparently identical, and then switch to a single version.
> >>
> > I thought there was an agreement to accept qemu-kvm implementation as is
> > into upstream (without some parts like device assignment). If you look
> > at qemu-kvm you'll see that upstream implementation is marked as
> > OBSOLETE_KVM_IMPL.
> 
> You can't merge both trees without introducing regressions, either in
> the kvm part or some other section that qemu-kvm did not stress. IMO,
> there is no way around understanding all the nice little "fixes" that
> piled up over the years and translate them into proper, documented patches.
OBSOLETE_KVM_IMPL should be just dropped, not merged.

> 
> > 
> >> We are on a good track now. I predict that we will be left with only one
> >> or two major additional features in qemu-kvm in a few months from now,
> >> no more duplications with subtle differences, and production-grade kvm
> >> upstream stability.
> >>
> > You are optimistic. My prediction is that it will take at least one major RHEL
> > release until such merged code base will become production-grade. That
> > is when most bugs that were introduced by eliminating subtle differences
> > between working and non-working version will be found :)
> 
> The more upstream code qemu-kvm stresses, the faster this convergence
> will become. And there is really not that much left. E.g, I've a
> qemu-kvm-x86.c here that is <400 LOC.
> 
That's what I don't get. Why working qemu-kvm should stress non working
upstream code? Just remove upstream code and replace it with qemu-kvm
version.

> > 
> > BTW Do you have a plan how to move upstream to thread per vcpu?
> 
> Upstream has this already, but it's - once again - a different
> implementation. Understanding those differences is one of the next steps.
> 
I see only two threads on upstream no matter how much vcpus I configure.

> In fact, as posted recently, unifying the execution model
> implementations is the only big problem I see. In-kernel irqchips and
> device assignment are things that can live in qemu-kvm without much
> conflicts until they are finally mergable.
> 
Upstream kvm is kinda useless without in-kernel irqchips.

--
			Gleb.