Re: [Qemu-devel] [PATCH RFC 1/1] s390x/pci: Extend pci representation by new zpci device

[Frank Blaschka <blaschka@linux.vnet.ibm.com>]

On Tue, Mar 03, 2015 at 10:33:05AM +0100, Alexander Graf wrote:
> 
> 
> 
> > Am 03.03.2015 um 09:06 schrieb Frank Blaschka <blaschka@linux.vnet.ibm.com>:
> > 
> >> On Thu, Feb 26, 2015 at 04:34:06PM +0100, Alexander Graf wrote:
> >> 
> >> 
> >>> On 26.02.15 16:27, Frank Blaschka wrote:
> >>>> On Thu, Feb 26, 2015 at 03:39:15PM +0100, Alexander Graf wrote:
> >>>> 
> >>>> 
> >>>>> On 26.02.15 12:59, Frank Blaschka wrote:
> >>>>> This patch extends the current s390 pci implementation to
> >>>>> provide more flexibility in configuration of s390 specific
> >>>>> device handling. For this we had to introduce a new facility
> >>>>> (and bus) to hold devices representing information actually
> >>>>> provided by s390 firmware and I/O configuration.
> >>>>> 
> >>>>> On s390 the physical structure of the pci system (bridge, bus, slot)
> >>>>> in not shown to the OS. For this the pci bridge and bus created
> >>>>> in qemu can also not be shown to the guest. The new zpci device class
> >>>>> represents this abstract view on the bare pci function and allows to
> >>>>> provide s390 specific configuration attributes for it.
> >>>>> 
> >>>>> Sample qemu configuration:
> >>>>> -device e1000,id=zpci1
> >>>>> -device ne2k_pci,id=zpci2
> >>>>> -device zpci,fid=2,uid=1248,pci_id=zpci1
> >>>>> -device zpci,fid=17,uid=2244,pci_id=zpci2
> >>>>> 
> >>>>> A zpci device references the corresponding PCI device via device id.
> >>>>> The new design allows to define multiple host bridges and support more
> >>>>> pci devices.
> >>>> 
> >>>> Isn't this reverse? Shouldn't it rather be
> >>>> 
> >>>>  -device zpci,...,id=zpci1
> >>>>  -device e1000,bus=zpci1.0
> >>>> 
> >>>> with a limit on each virtual zpci bus to only support one device?
> >>> 
> >>> Do you mean something like having multiple host bridges (providing a pci bus
> >>> each) and limit the bus to just one device?
> >>> 
> >>> -device s390-pcihost,fid=16,uid=1234
> >>> -device s390-pcihost,fid=17,uid=5678
> >>> -device e1000,bus=pci.0
> >>> -device ne2k_pci,bus=pci.1
> >>> 
> >>> We also discussed this option but we don't like the idea to put attributes
> >>> belong to the pci device to the host bridge.
> >> 
> >> I guess I'm not grasping something obvious here :). What exactly are the
> >> attributes again?
> > Sorry for the late response, I was on vacation the last couple days.
> > 
> > The fid and uid values are provided by microcode/io layer on the real hardware.
> 
> So they are arbitrary numbers? What uniqueness constraints do we have on them?
fid and uid must be unique within the same qemu. At a first look the numbers are
arbitrary but our configuration folks want explicitly define a particular fid and uid
to better support migration and pass-through scenarios.
 
> IIUC you can only have a single pcie device behind a virtual "bus" anyway, so what if we just calculate uid and fid from the bus id?
I think this similar to the current implementation. We use the slot (idea for the future was
bus + slot) to generate uid and fid. But this is not flexible enough. As I said, our
configuration folks want to be able to specify fid and uid for the device.

> 
> Alex
> 
> > You can read them out via s390 specific device attributes e.g.
> > 
> > # cat /sys/bus/pci/devices/0000\:00\:00.0/function_id
> > 0x00000016
> > # cat /sys/bus/pci/devices/0000\:00\:00.0/uid
> > 0x25
> > 
> > Since there is no regular pci address (as explained earlier) I think this is a
> > mechanism to unique identify a pci function.
> > 
> > We discussed both options how to model this in qemu, but maybe you have another
> > even better idea how to bring this additional attributes to a qemu pci device.
> > 
> > Thx for any help and new ideas ...
> > 
> > Frank 
> >> 
> >> Alex
> > 
>