Re: [Qemu-devel] [PULL 15/54] build: convert pci.mak to Kconfig

[Michael S. Tsirkin]

On Thu, Mar 14, 2019 at 01:53:48PM +0100, Andrea Bolognani wrote:
> On Mon, 2019-03-04 at 19:19 +0100, Paolo Bonzini wrote:
> > Instead of including the same list of devices for each target,
> > set CONFIG_PCI to true, and make the devices default to present
> > whenever PCI is available.  However, s390x does not want all the
> > PCI devices, so there is a separate symbol to enable them.
> [...]
> > +++ b/default-configs/riscv32-softmmu.mak
> > @@ -1,8 +1,8 @@
> >  # Default configuration for riscv-softmmu
> >  
> > -include pci.mak
> >  include usb.mak
> > -
> > +CONFIG_PCI=y
> > +CONFIG_PCI_DEVICES=y
> >  CONFIG_SERIAL=y
> >  CONFIG_VIRTIO_MMIO=y
> 
> I *think* this might have caused some unwanted changes for RISC-V.
> 
> pcie-root-port is built into qemu-system-riscv64 by default as of
> dbbc277510aa (along with ioh3420!), but if you actually try to use it
> you'll get:
> 
>   $ ./riscv64-softmmu/qemu-system-riscv64 \
>     -M virt \
>     -device pcie-root-port
>   qemu-system-riscv64: -device pcie-root-port: MSI-X is not
>                        supported by interrupt controller
> 
> This is a limitation we have been aware of, and the plan was to
> enable the device in QEMU once it had been addressed: from the
> libvirt side, the availability of the device would have meant that it
> was safe to use it, but if the device is enabled in QEMU before it
> can actually be used, then that makes detection on the libvirt side
> problematic.
> 
> I haven't spent time digging further - and I'm not familiar enough
> with the QEMU build system anyway O:-) - but I wouldn't be surprised
> if the same happened for other architectures, too.

I'd say just disable it at build time by default.
How about a dependency on BROKEN? Developers can
enable it temporarily. Paolo?

> -- 
> Andrea Bolognani / Red Hat / Virtualization