Re: [Qemu-devel] [v5][PATCH 2/5] xen, gfx passthrough: create pseudo intel isa bridge

[Chen, Tiejun]

On 2014/7/1 13:47, Michael S. Tsirkin wrote:
> On Mon, Jun 30, 2014 at 08:34:46PM +0100, Stefano Stabellini wrote:
> > On Fri, 27 Jun 2014, Chen, Tiejun wrote:
> > > On 2014/6/25 17:58, Chen, Tiejun wrote:
> > > > On 2014/6/25 17:44, Michael S. Tsirkin wrote:
> > > > > On Wed, Jun 25, 2014 at 05:28:48PM +0800, Chen, Tiejun wrote:
> > > > > > On 2014/6/25 17:21, Michael S. Tsirkin wrote:
> > > > > > > On Wed, Jun 25, 2014 at 05:14:30PM +0800, Chen, Tiejun wrote:
> > > > > > > > On 2014/6/25 17:04, Michael S. Tsirkin wrote:
> > > > > > > > > On Wed, Jun 25, 2014 at 04:48:02PM +0800, Chen, Tiejun wrote:
> > > > > > > > > > On 2014/6/25 16:43, Michael S. Tsirkin wrote:
> > > > > > > > > > > On Wed, Jun 25, 2014 at 04:39:07PM +0800, Chen, Tiejun wrote:
> > > > > > > > > > > > > In fact it's exactly what passthrough does.
> > > > > > > > > > > > > I wonder if more bits from ./hw/i386/kvm/pci-assign.c
> > > > > > > > > > > > > can be reused. How do you poke at the host device? sysfs?
> > > > > > > > > > > >
> > > > > > > > > > > > Yes, sysfs.
> > > > > > > > > > > >
> > > > > > > > > > > > Thanks
> > > > > > > > > > > > Tiejun
> > > > > > > > > > >
> > > > > > > > > > > Then you should be able to re-use large chunks of
> > > > > > > > > > > ./hw/i386/kvm/pci-assign.c: basically everything
> > > > > > > > > > > that deals with emulation.
> > > > > > > > > >
> > > > > > > > > > Do you mean those hooks to get info from the real device? Xen
> > > > > > > > > > have its own
> > > > > > > > > > wrapper, xen_host_pci_get_block(), so we always go there in xen
> > > > > > > > > > scenario.
> > > > > > > > > >
> > > > > > > > > > Thanks
> > > > > > > > > > Tiejun
> > > > > > > > >
> > > > > > > > > Yes and that's not good.  We have two pieces of code doing mostly
> > > > > > > > > identical things slightly differently.
> > > > > > > > > hw/i386/kvm/pci-assign.c is a bit younger so it's cleaner,
> > > > > > > > > but these really need to be unified.
> > > > > > > >
> > > > > > > > Sorry, take a look at this again,
> > > > > > > >
> > > > > > > > xen_host_pci_get_block(XenHostPCIDevice *d, int pos, uint8_t *buf,
> > > > > > > > int len)
> > > > > > > >      |
> > > > > > > >      + xen_host_pci_config_read(d, pos, buf, len)
> > > > > > > >          |
> > > > > > > >          + pread(d->config_fd, buf, len, pos)
> > > > > > > >
> > > > > > > > I thinks this should be same as kvm.
> > > > > > > >
> > > > > > > > Thanks
> > > > > > > > Tiejun
> > > > > > >
> > > > > > > get_block is trivial.
> > > > > > >
> > > > > > > I really mean the whole PT infrastructure for
> > > > > > > - discovering host devices through sysfs
> > > > > > > - virtualizing devices
> > > > > > >
> > > > > > > rom, bars, msi ...
> > > > > > > the list goes on.
> > > > > > >
> > > > > > > logic is mostly the same.
> > > > > >
> > > > > > Looks you mean we can unify the entire PT infrastructure between kvm
> > > > > > and xen
> > > > > > inside qemu. But I'm afraid its not easy to do in a short time, so
> > > > > > maybe we
> > > > > > can queue this as next phase.
> > > > > >
> > > > > > Thanks
> > > > > > Tiejun
> > > > >
> > > > > I'm afraid once we merge your code, you'll lose interest :)
> > > >
> > > > Currently we have to push this feature into upstream as our first
> > > > priority, so unless something is really needed to address. Of course I
> > > > hope this point what we're talking is not such a thing :)
> > > >
> > > > But I can promise here I'd like to do this optimization with your guide
> > > > next :)
> > > >
> > > > > At least, don't add duplicate code for ROM.
> > > >
> > > > Let me try this.
> > >
> > > Its not easy as expected.
> > >
> > > kvm always work with this structure, AssignedDevice, and especially this is
> > > just activated in kvm_enabled(). And then set all properties to this
> > > structure.
> > >
> > > In xen case, the similar structure, XenHostPCIDevice, is not easy transferred
> > > into the structure, AssignedDevice. So this mean we have to split
> > > assigned_dev_load_option_rom() as line by line for xen and kvm, respectively.
> > >
> > > I really agree we definitely need to unify PT infrastructure between kvm and
> > > xen after this try since I can't understand why we originally introduce same
> > > way to do same thing :(
> > >
> > > Do you have better idea? If not, I prefer we open this completely as next
> > > action to follow-up. But this time I'm afraid I can't get in this.
> >
> > The Xen PT code came first. At the time Anthony Liguori and I argued for
> > sharing the PT code with KVM going forward but Avi wanted to retain the
> > KVM PT code as is in order not to introduce any regressions on KVM.
> >
> > Ah.. the good old times :-)
> >
> > http://lists.xen.org/archives/html/xen-devel/2011-10/msg00195.html
>
> Right, I remember that there was an idea to first make the code
> reusable before merging either Xen or KVM PT, but in the end
> the plan to merge both and then abstract it out, won.
> We never got to abstracting it out but we should!
>
> Yet, I do not think we should require abstracting all of Xen PT as
> a pre-requisite for this work, but I think we can
> avoid adding to the duplication without a lot of effort.
> To start sharing with ROM code, I think we should rename
> AssignedDevice->KvmAssignedDevice, and move fields that can be shared
> with Xen from there to AssignedDevice.

Its not simple to share something between AssignedDevice and the similar 
structure, XenHostPCIDevice.

As I said previously AssignedDevice is just activated in kvm_enabled(), 
then set all properties to this structure.

But kvm_enabled() is zero since we always disable this in xen scenario. 
So this means we have to invoke pci-assign framework completely if we 
try to force to use this. So actually this would be equal to merge kvm 
and xen.

Or if you already figure out to avoid this happened let me know.

Thanks
Tiejun