On Fri, Oct 16, 2015 at 10:52:00AM +0200, Laurent Vivier wrote:
>
>
> On 13/10/2015 15:55, Peter Maydell wrote:
> > On 13 October 2015 at 14:25, Peter Maydell <peter.mayd...@linaro.org> wrote:
> >> On 13 October 2015 at 14:19, Michael S. Tsirkin <m...@redhat.com> wrote:
> >>> On Tue, Oct 13, 2015 at 02:12:34PM +0100, Peter Maydell wrote:
> >>>> Well, if we're going to go around and fix the machines which
> >>>> don't get things right, I guess. It's a shame the default for
> >>>> the global is "this machine is broken", because now every
> >>>> new machine will default unnecessarily to broken, and there's
> >>>> no way to grep the source tree for machines which need fixing.
> >>
> >>> It'd be easy enough to revert the logic if someone's willing to start on
> >>> this. I'm reluctant to make this patchset depend on changing all
> >>> machines, but if you think I'm wrong, pls let me know.
> >>
> >> Most machines don't have a PCI controller, luckily.
> >> I'll have a look at how many files it would touch...
> >
> > So, first up I'm happy that the gpex and versatile
> > pci controllers don't have this problem (the way they
> > set up the mmio and io windows means there won't be
> > overlaps). That leaves the following pci controllers,
> > which I've listed with the source files which define machines
> > that use them:
> >
> > hw/pci-host/pbm
> > hw/sparc64/sun4u.c
> > hw/pci-host/bonito
> > hw/mips/mips_fulong2e.c
> > hw/pci-host/grackle
> > hw/pc/mac_oldworld.c
> > hw/pci-host/piix
> > [the x86 systems]
> > hw/pci-host/ppce500
> > hw/ppc/e500.c
> > hw/pci-host/prep
> > hw/ppc/prep.c
> > hw/pci-host/q35
> > [x86 systems]
> > hw/pci-host/uninorth
> > hw/ppc/mac_newworld.c
> > hw/alpha/typhoon
> > used in hw/alpha/dp264.c
> > hw/mips/gt64xxx_pci
> > hw/mips/mips_malta.c
> > hw/ppc/ppc4xx_pci
> > hw/ppc/ppc440_bamboo.c
> > hw/ppc/spapr_pci
> > hw/ppc/spapr.c [already marked as '0 addrs ok']
> > hw/s390/s390-pci-bus
> > hw/s390/s390-virtio-ccw.c
> > hw/sh4/sh_pci
> > hw/sh4/rd2.c
> >
> > So we'd need to touch perhaps fifteen files in total.
> >
> > I don't insist we do that rather than applying this particular
> > patch, but I don't think it's a huge effort.
>
> I'm going to remove the zero address checking and try to start a qemu
> for each of them to see which ones are broken.
Can't hurt, but I'm not sure that will be enough,
as overlaps might depend on what does the guest do.
It's a question of looking at how are mmio and
io windows for pci set up.
> I remember from a previous discussion on my initial patch that x86
> should be ok.
Yes because on x86 PCI has priority -1 - lower than all other
system devices.
That was handled by this patch:
commit 83d08f2673504a299194dcac1657a13754b5932a
Author: Michael S. Tsirkin <m...@redhat.com>
Date: Tue Oct 29 13:57:34 2013 +0100
pc: map PCI address space as catchall region for not mapped addresses
> > thanks
> > -- PMM
> >
> Laurent
