Re: [Qemu-devel] [PATCH] pci: Length-align config space accesses

[Michael S. Tsirkin]

On Wed, Jul 20, 2011 at 07:10:57PM +0200, Jan Kiszka wrote:
> On 2011-07-20 18:45, Michael S. Tsirkin wrote:
> > On Wed, Jul 20, 2011 at 06:18:43PM +0200, Jan Kiszka wrote:
> >> On 2011-07-20 18:17, Isaku Yamahata wrote:
> >>> On Wed, Jul 20, 2011 at 04:27:08PM +0200, Jan Kiszka wrote:
> >>>> On 2011-07-20 14:15, Jan Kiszka wrote:
> >>>>> On 2011-07-20 14:00, Isaku Yamahata wrote:
> >>>>>> Hi. This clean up looks good basically.
> >>>>>
> >>>>> Oops, forgot to cc you. Sorry.
> >>>>>
> >>>>>> But when conventional pci device is accessed via MMCONFIG area,
> >>>>>> addr &= addr_mask doesn't work as expected.
> >>>>>> The config area of [256, 4K) of conventional pci should have no effect.
> >>>>>
> >>>>> Mmh, I see. Looks like we need to split accesses at this boundary and
> >>>>> executed them separately.
> >>>>
> >>>> Nope, no such issue: we already automatically split up accesses that
> >>>> span the legacy/extended boundary. Just like so far, legacy config space
> >>>> handlers have to filter out requests that address regions >= 256.
> >>>
> >>> For example, when accessing to offset 257 of conventional pci device,
> >>> the access is routed to offset 1 due to the masking.
> >>> Such overwrapping isn't correct.
> >>
> >> No, it isn't routed like that. The mask used via mmio is 0xfff.
> >>
> >> Jan
> > 
> > I thought about this some more, I'd like to see how devices
> > are going to benefit. Any examples?
> 
> No in-tree device currently gets beyond the "write default, check range"
> pattern.
> 
> > If not, is it easier to simply make this logic
> > part of dev assignment?
> 
> That's a question of clean interfaces. Not checking at the core means
> exposing invalid accesses to the callbacks.

That's the point: at least the express spec seems to explicitly allow
any accesses within a dword, and disallow accesses crossing dwords.
And there's no way to create such on the PCI bus either.

So it makes sense to handle cross-dword accesses in a special way
(possibly disallow them altogether, need to check non-pc systems
for this).

But non-aligned access within a dword is legal according to the spec.

> The logic is required anyway, so better put it in a central place. E.g.
> if the Xen people push their device assignment as well, we will suddenly
> have more than one user.
> 
> Jan

That's the question: why is the logic required?

-- 
MST