Re: [Xen-devel] Does a Virtual PCI Device can have MSI's

[Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>]

On Wed, 13 Aug 2014, manish jaggi wrote:
> On 13 August 2014 23:21, Stefano Stabellini
> <stefano.stabellini@xxxxxxxxxxxxx> wrote:
> > On Wed, 13 Aug 2014, manish jaggi wrote:
> >> On 13 August 2014 19:13, Stefano Stabellini
> >> <stefano.stabellini@xxxxxxxxxxxxx> wrote:
> >> > On Wed, 13 Aug 2014, manish jaggi wrote:
> >> >> On 13 August 2014 16:20, Stefano Stabellini
> >> >> <stefano.stabellini@xxxxxxxxxxxxx> wrote:
> >> >> > On Wed, 13 Aug 2014, manish jaggi wrote:
> >> >> >> On 13 August 2014 15:40, Stefano Stabellini
> >> >> >> <stefano.stabellini@xxxxxxxxxxxxx> wrote:
> >> >> >> > On Wed, 13 Aug 2014, manish jaggi wrote:
> >> >> >> >> Hi,
> >> >> >> >>
> >> >> >> >> I think it should be possible, but confirming it that this 
> >> >> >> >> feature is
> >> >> >> >> enabled in xen. I don't know how to test it.
> >> >> >> >>
> >> >> >> >> Does any virtual PCI device in DomU (I don't mean a virtual 
> >> >> >> >> function)
> >> >> >> >> have MSI interrupts ?
> >> >> >> >
> >> >> >> > Yes, they do.
> >> >> >> >
> >> >> >> >
> >> >> >> >> If yes then how is that MSI handled in Xen
> >> >> >> >
> >> >> >> > PV guests on x86 don't receive MSIs or legacy interrupts as is.  
> >> >> >> > They
> >> >> >> > map them into "pirqs" instead, that are a kind of event channels, 
> >> >> >> > Xen
> >> >> >> > specific software interrupts. For each MSI on the PCI device 
> >> >> >> > assigned to
> >> >> >> > the guest, the guest kernel would ask for a pirq, see:
> >> >> >> >
> >> >> >> > arch/x86/pci/xen.c:xen_pcifront_enable_irq
> >> >> >> > arch/x86/pci/xen.c:xen_setup_msi_irqs
> >> >> >> >
> >> >> >> > In the specific case of MSIs and MSI-X, pcifront issues an 
> >> >> >> > hypercall in
> >> >> >> > order to enable them, see:
> >> >> >> >
> >> >> >> > drivers/pci/xen-pcifront.c:pci_frontend_enable_msi
> >> >> >> >
> >> >> >> > and the backend returns the pirq number:
> >> >> >> >
> >> >> >> > drivers/xen/xen-pciback/pciback_ops.c:xen_pcibk_enable_msi
> >> >> >> >
> >> >> >> >
> >> >> >> > On ARM I think it would be best if we delivered MSIs as MSIs to the
> >> >> >> > guest, rather than mapping them into pirqs, to take better 
> >> >> >> > advantage of
> >> >> >> > the hardware. But it would be up to you to change the 
> >> >> >> > pcifront/pciback
> >> >> >> > code to do it.
> >> >> >> > In first instance it would be fine if we end up using pirqs.
> >> >> >>
> >> >> >> I am considering 2 cases here
> >> >> >> a) physical PCI passthrough devices / functions assigned to domU
> >> >> >
> >> >> > Are you sure you mean DomU here, or maybe you mean Dom0?
> >> >> >
> >> >> Yes DomU only.
> >> >> >
> >> >> >> b) emulated (virtual) PCI devices assigned to domU
> >> >> >
> >> >> > We need to clarify the terminology here: what do you mean by (b)?
> >> >> > Emulating an entire PCI device and exposing it to domU? Why do you 
> >> >> > want
> >> >> > to do that? It is not a feature I am keen on having on Xen on ARM.
> >> >> > Otherwise if you are thinking of a virtual function of an SR-IOV card,
> >> >> > that is still (a) from the Xen point of view.
> >> >>
> >> >> I was trying to understand that does Xen support some device like a
> >> >> sata device which is a virtual one emulated using qemu on a PV domU
> >> >> and its interrupts are MSIs
> >> >
> >> > I wouldn't want to support this use case at all, unless strictly
> >> > necessary: emulation is slower and less secure (larger surface of
> >> > attack) than PV interfaces.
> >> >
> >> Thats why I asked this question, if a virtual device is assigned to a
> >> domU how would the MSIs be configured for it, using front-back
> >> communication or using the Linux ITS driver which gets trapped into
> >> Xens ITS driver.
> >
> > If it is an emulated device, by definition there is no front-back
> > communication.
> >
> So if it is not an emulated device and uses PV interfaces
> a) Can MSIs be assigned to them

No. Only in case of PCI passthrough.


> b) If they are,  How MSI are configured for them when their driver
> calls (pci_enable_msi), Using front back communication ?
> >From another mail thread in which we are discussing the possible
> removal of front-back comm for configuring the MSI and rather using
> platform ITS driver, In that case how would it work.
> 
> >
> >> >> >> For (a) it is straight to configure and inject the MSI into guest
> >> >> >
> >> >> > Yep, that is what I was trying to say.
> >> >> >
> >> >> >
> >> >> >> For (b) how does the configuring and injection should work,
> >> >> >> - PCI Front driver using backops requests to enable msi
> >> >> >> - At a later stage xen using dom0 (somehow) inject an virtual LPI 
> >> >> >> into domU.
> >> >> >>
> >> >> >> What are your thoughts on this?
> >> >> >
> >> >> > I am not sure I understand what you mean by (b) anymore. In fact
> >> >> > pcifront is used to deal with PCI passthrough to DomUs, that would be
> >> >> > (a) by your description.
> >> >>
> >>
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel