On Mon, Jun 25, 2018 at 12:28:34PM +0200, Paolo Bonzini wrote: > On 25/06/2018 08:50, Markus Armbruster wrote: > > Paolo Bonzini <pbonz...@redhat.com> writes: > > > >> On 22/06/2018 21:35, Eduardo Habkost wrote: > >>>>> Why is this better than using KVM by default if it's available? > >>>> The answer is (as almost always): Compatibility with migration. Nobody > >>>> dares to sacrifice that chicken :-( > >>> We can now kill it if we announce the feature as deprecated a > >>> couple of releases in advance. > >>> > >>> If we declare that compatibility when the accelerator is omitted > >>> is deprecated in 3.0, in QEMU 3.3 we will be free to choose a > >>> different default accelerator. > >> > >> We can, we don't necessarily want it. > >> > >> The status quo is that people using KVM are invoking qemu as qemu-kvm, > >> people using TCG are invoking qemu as qemu-system-*. All distros are > >> shipping a qemu-kvm or more rarely kvm binary, which is invariably a > >> wrapper script except for RHEL because RHEL doesn't have a qemu-system-* > >> binary at all. > >> > >> By the way, changing qemu-system-*'s default to e.g. RHEL's "kvm or tcg" > >> would not help distros that have "-accel kvm" in their /usr/bin/qemu-kvm > >> script. > > > > It wouldn't hurt them, either. > > Right; to sum up, it does make things a little less consistent for their > users in two ways: > > - qemu-system-<native> behaves differently from qemu-system-<others>. > For example, for ARM the default CPU model might not work for KVM, so > you would have to add a "-cpu xxx" option. > > - qemu-system-<native> would still need an accelerator option on OS X or > Windows, where there is not quite parity between TCG and the native > accelerator, in terms of either features or stability. Because of this > we wouldn't be able to change the default to "whatever virtualizing > accelerators are available followed by TCG". > > > Attentive distros could even replace the wrapper script by a link. > > If they are okay with replacing the "KVM only" semantics with "KVM or > TCG", which I think is generally worse.
If we can't get agreement on what's the right default for each QEMU binary, I think that's yet another reason to document that upstream QEMU won't guarantee ABI compatibility if -accel is omitted. If downstream distributions want to keep promising ABI compatibility, it will be up to them. -- Eduardo
