[Qemu-devel] Re: [PATCH v3 08/13] pcie root port: implement pcie root port.

[Michael S. Tsirkin]

On Tue, Sep 28, 2010 at 08:01:15AM +0900, Isaku Yamahata wrote:
> On Mon, Sep 27, 2010 at 12:40:12PM +0200, Michael S. Tsirkin wrote:
> > On Mon, Sep 27, 2010 at 03:22:43PM +0900, Isaku Yamahata wrote:
> > > On Sun, Sep 26, 2010 at 02:50:42PM +0200, Michael S. Tsirkin wrote:
> > > > On Fri, Sep 24, 2010 at 02:38:09PM +0900, Isaku Yamahata wrote:
> > > > > On Wed, Sep 22, 2010 at 01:25:59PM +0200, Michael S. Tsirkin wrote:
> > > > > 
> > > > > > > +PCIESlot *pcie_root_init(PCIBus *bus, int devfn, bool 
> > > > > > > multifunction,
> > > > > > > +                         const char *bus_name, pci_map_irq_fn 
> > > > > > > map_irq,
> > > > > > > +                         uint8_t port, uint8_t chassis, uint16_t 
> > > > > > > slot);
> > > > > > > +
> > > > > > 
> > > > > > I am a bit unhappy about all these _init functions.
> > > > > > Can devices be created with qdev? If they were
> > > > > > it would be possible to configure the system completely
> > > > > > from qemu command line.
> > > > > 
> > > > > That's very reasonable question.
> > > > > Once machine configuration file is supported, those initialization
> > > > > functions will go away.
> > > > > I.e. when the initialization code like pc_init1() in pc_piix.c 
> > > > > disappears,
> > > > > those functions will also go away.
> > > > > 
> > > > > Until that, those initialization glues will stay like pci_create 
> > > > > family
> > > > > or other many initialization glues unfortunately.
> > > > > This is the result of qdev missing a feature, not the cause.
> > > > > It would be a long-term issue to add machine configuration file 
> > > > > support.
> > > > 
> > > > Just to clarify, if I wanted to have a flag to make virtio-net
> > > > a pci express device, how would I do this?
> > > 
> > > the following preparation is needed.
> > > - register PCIDeviceInfo with name like "virtio-net-pci-xen"
> > >   with PCIDeviceInfo::is_express = true.
> > > - in initialization function, initialize express capability.
> > > - in write config function, call related function.
> > > 
> > > And then,
> > > if (express)
> > >    create "virtio-net-pci-xen"
> > > else
> > >    create "virtio-net-pci"
> > 
> > Sounds pretty bad: we can't double the number of devices
> > with each capability we add.
> > Can we make it so setting is_express on command line
> > will convert the device to PCI express?
> 
> I don't see your point. Capability isn't something that should be
> genericly customizable.
> If any, it would not be a generic option, but one specific to device.
> There is no generic way to automatically convert pci device into
> express device. It means designing a new express device.
> Maybe what you want is
> - set is_express = true
>   This result in always allocating 4k-sized configuration space.
>   (Possibly we need more fine-grained parameter in PCIDeviceInfo
>    than is_express.)
> - introduce device specific property that a user can turn on/off
> - In initialization function/config write function and where necessary,
>   check the property.

I think what I want is (At some point) automatically convert all virtio
users to express devices, but have a fallback option for old machine
types.

> -- 
> yamahata