Re: [Xen-devel] [rfc 00/18] ioemu: use devfn instead of slots as the unit for passthrough

[Yuji Shimada <shimada-yxb@xxxxxxxxxxxxxxx>]

On Fri, 20 Feb 2009 18:07:00 +1100
Simon Horman <horms@xxxxxxxxxxxx> wrote:

> On Thu, Feb 19, 2009 at 09:38:24AM +0000, Keir Fraser wrote:
> > On 19/02/2009 09:21, "Yuji Shimada" <shimada-yxb@xxxxxxxxxxxxxxx> wrote:
> > 
> > >> To be honest I am a little confused about what the above maping
> > >> is supposed to achive.
> > > 
> > > Please find the attached figure which shows the interrupt routing in
> > > xen hypervisor.
> > 
> > The point being to deliberately permute the mapping to try to avoid
> > accidental GSI sharing even if there are patterns in DEV:INTX usage (e.g.,
> > all devs use INTA).
> 
> Thanks for the information, especially the diagram. It is very useful.
> 
> Armed with this new kowledge I have a few questions.
> 
> 1. Shimada-san stated that shared GSI are not permitted for
>    pass-through devices. Is it permitted for a GSI to be shared
>    between a pass-through device and a non-pass-through device?

Yes, it is permitted. But guest software will receive spurious
interrupt. So it is not good.

>    The current scheme seems to leave scope for this as
> 
>    gsi 6 A = gsi 13 D = gsi 21 C = gsi 29 B
>    gsi 7 A = gsi 14 D = gsi 22 C = gsi 30 B

Do you mean this?

     Dev 6 INTA = Dev 13 INTD = Dev 21 INTC = Dev 29 INTB -> GSI 40
     Dev 7 INTA = Dev 14 INTD = Dev 22 INTC = Dev 30 INTB -> GSI 44

> 2. In several places in ioemu:io/passthrough.c e_intx is set to 0,
>    corresponding to INTA. Is this because it is virtual and
>    using INTA is convenient? Or is it because it is assumed
>    that the physical device being passed-through is a 0 function
>    (and 0 functions always use INTA) ?

INTx is virtualized, because the single function device normally use
INTA.

When we make multi-function cards appear in guests as multi-function
cards, it is good that virtual INTx reflects the physical INTx. The
reason is one of functions of a device may share INTx of the other
function. In my environment, UHCI(00:1d.0) and EHCI(00:1d.7) share the
same INTA. If physical functions share physical INTx, virtual
functions should share virtual INTx. To achieve this, virtual INTx
needs to reflect the physical INTx.

>    The latter assumption is not valid because even without my pacthes
>    it is possible to pass-through non-0 functions, its just that
>    they end up as the 0th function of the virtual slot in the guest.
> 
> I am now pretty sure that my change leads to incorrect usage of
> hvm_pci_intx_gsi(). Answers to the questions above will help me to
> understand how trivial to fix this is.
> 
> The most difficult cases seem to be 1) sharing of gsi between
> pass-through and non-pass-through devices is not permitted or 2)
> intx used inside ioemu:io/passthrough.c should reflect the physical
> intx. In either case I wonder if a reasonable solution would be to
> just allocate allocate GSI in a non-colliding manner. Say, GSI 16 for
> the first device to ask, 17 for the next one and so on. Or perhaps
> the existing hash + overflow to the next GSI on collision.

The another solution is expanding GSI to 127. I don't sure it is
possible, but sharing virtual GSI will not occur.

Thanks,
--
Yuji Shimada


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel