Re: [Qemu-devel] [PATCH 6/8] pci: Simpler implementation of primary PCI bus

[David Gibson]

On Wed, May 29, 2013 at 03:22:29PM +0300, Michael S. Tsirkin wrote:
> On Wed, May 29, 2013 at 09:04:00PM +1000, David Gibson wrote:
> > On Wed, May 29, 2013 at 01:17:13PM +0300, Michael S. Tsirkin wrote:
> > > On Wed, May 29, 2013 at 08:06:42PM +1000, David Gibson wrote:
> > > > On Wed, May 29, 2013 at 12:55:53PM +0300, Michael S. Tsirkin wrote:
> > > > > On Wed, May 29, 2013 at 07:43:41PM +1000, David Gibson wrote:
> > > > > > On Thu, May 23, 2013 at 10:16:27PM +1000, David Gibson wrote:
> > > > > > > On Thu, May 23, 2013 at 02:22:30PM +0300, Michael S. Tsirkin 
> > > > > > > wrote:
> > > > > > > > On Thu, May 09, 2013 at 10:31:10AM +1000, David Gibson wrote:
> > > > > > > > > Currently pci_get_primary_bus() searches the list of root 
> > > > > > > > > buses for one
> > > > > > > > > with domain 0.  But since host buses are always registered 
> > > > > > > > > with domain 0,
> > > > > > > > > this just amounts to finding the only PCI host bus.
> > > > > > > > > 
> > > > > > > > > This simplifies the implementation by defining the primary 
> > > > > > > > > PCI bus to
> > > > > > > > > be the first one registered, using a global variable to track 
> > > > > > > > > it.
> > > > > > > > > 
> > > > > > > > > Signed-off-by: David Gibson <address@hidden>
> > > > > > > > 
> > > > > > > > Or better: can we just fail if there is more than
> > > > > > > > one root?
> > > > > > > 
> > > > > > > That might work, I'll look into doing that.
> > > > > > 
> > > > > > So, the difficulty with this is that then any machine with multiple
> > > > > > PCI bridges could not use pci_nic_init(), since it calls
> > > > > > pci_get_bus_devfn() which calls pci_find_primary_bus() which would
> > > > > > always fail.  And using pci_nic_init() is more or less mandatory in
> > > > > > the machine_init function to support old-style nic configuration.
> > > > > > 
> > > > > > Suggestions?
> > > > > 
> > > > > You mean multiple PCI roots?
> > > > > Well, there are no legacy machines with multiple roots to support, are
> > > > > there?  So why do we need to support legacy flags for these new
> > > > > configurations?
> > > > 
> > > > Because people expect them.
> > > 
> > > People can learn, somehow they will learn to add a new root, so they can
> > > learn to use -device too.
> > 
> > Hrm.  I'd kind of like a second (third?) opinion on that.  Anthony?
> > 
> > > So let's make it fail on multiple roots, and output a message along the
> > > lines of "please use -device virtio-net-pci instead".
> > 
> > How to produce a meaningful error like that isn't totally obvious,
> > since the test for multiple roots is down in find_primary_pci_bus()
> > (or whatever), and once we get back up to pci_nic_init() we just know
> > that pci_get_bus_devfn() failed for some reason.
> 
> What other possible reason for it to fail?

Unparseable address (it can be user specified) or internal failure to
initialize the device are the first two that spring to mind..

> > > > Plus on spapr we already support the
> > > > legacy nic options; it would be very strange for them to suddenly
> > > > break when we add a second host bridge.
> > > 
> > > Not sure who "we" is here. IMHO user should ask for a new
> > > machine type with two roots explicitly.
> > 
> > You seem to be thinking of the number of host bridges as a fixed
> > property of the platform, which it isn't on spapr.  PCI host bridges
> > are just another device.  Large scale real hardware can easily have
> > dozens of them.
> 
> Absolutely. I'm not thinking of it as fixed.
> I'm thinking of the *default* number of pci host bridges
> as fixed. If a user is smart enough to use -device to create
> a host bridge, said user can learn about -device for creating
> a nic.

Hm, I guess.  I'm still uncomfortable with breaking a documented
option, even if its not the preferred method these days.

> > In qemu we create one always as a convenience, but
> > users can (and will have to, for vfio) create additional ones
> > trivially with -device.
> 
> So they know about -device then.
> 
> > [Which raises another complication as a tangent.  People (and libvirt)
> > don't generally expect -nodefaults to remove the PCI bridge, but
> > arguably it should on spapr, since a PAPR guest with no PCI is
> > perfectly viable but there's currently no way to specify such a
> > thing.]
> 
> I guess the problem is not what they expect generally,
> but specifically that some users might rely on spapr with
> -nodefaults having PCI?

I'm pretty sure libvirt will rely on that, if nothing else.

> I don't have any ideas besides introducing a new machine type
> that is same as spapr but without the default PCI host bridge.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: Digital signature