Re: [PATCH 6/6] xen/hybrid: Enable grant table and xenbus

[Sheng Yang <sheng@linux.intel.com>]

On Wednesday 03 February 2010 02:09:03 Ian Campbell wrote:
> On Tue, 2010-02-02 at 16:46 +0000, Sheng Yang wrote:
> > On Wednesday 03 February 2010 00:24:04 Ian Campbell wrote:
> > > On Tue, 2010-02-02 at 13:24 +0000, Sheng Yang wrote:
> > > > I am not sure if I understand you right, but I think the issue is,
> > > > there is no PVonHVM drivers in Linux upstream. The drivers are
> > > > currently maintained by OSVs, and the one in Xen upstream code only
> > > > support 2.6.18. So I didn't take them into consideration at the time.
> > >
> > > True, but this is something which should be taken care of by the core
> > > Xen-aware code not something which should be pushed down into each
> > > driver.
> > >
> > > Someone who wants to add PVonHVM functionality shouldn't have to go and
> > > remove a bunch of conditionals from each driver (or worse add
> > > alternative clauses to each check!).
> > >
> > > > I think the "xen_evtchn_enable()" looks much better. Would replace
> > > > these ugly lines in the next version.
> > >
> > > I think it would be cleaner to encapsulate this in the evtchn code
> > > rather than leaking platform knowledge into each driver. IOW the evtchn
> > > functions should return failure if event channels are not enabled and
> > > the driver should cope with this gracefully.
> >
> > Agree. That what I suppose to do. What the drivers should only know is,
> > if event channel is enabled.
> >
> > > Or perhaps at the xenbus driver level we should be deciding whether or
> > > not we have enough paravirtualisation to be worth probing the drivers
> > > at all?
> >
> > I think current scheme is direct enough for now. We can improve it later.
> 
> Taking a step back I don't think any of these checks are necessary at
> all -- in order to get as far as actually probing the devices xenbus
> needs to be up and running, which implies event channels, as well as
> everything else required for PV drivers to function, are working.
> 
> Drivers for PCI devices don't all start with "if (pci_bus_available())",
> they just register the driver and let the kernel's driver core take care
> of things. If someone is worried about the overhead of an extra driver
> being registered, well that is what modular kernels are for.
> 
> I think that even the existing "if (xen_domain())" check is unnecessary,
> at least in the frontend drivers.

Um, very reasonable. I would provide another patch for this.

-- 
regards
Yang, Sheng