Re: [RFC PATCH] PCI MMCONFIG: add validation against ACPI motherboard resources

[Jesse Barnes <jbarnes@virtuousgeek.org>]

On Wednesday, May 23, 2007 4:04 pm Robert Hancock wrote:
> Jesse Barnes wrote:
> > On Tuesday, May 22, 2007 6:06 pm Robert Hancock wrote:
> >> There was a big discussion about this back in 2002, in which Linus
> >> wasn't overly enthused about disabling the decode during probing
> >> due to risk of causing problems with some devices:
> >>
> >> http://lkml.org/lkml/2002/12/19/145
> >>
> >> In this particular case (64-bit BAR) we might be able to avoid the
> >> problem by changing the order in which we probe the two halves of
> >> the address, i.e. change the top half to 0xffffffff before messing
> >> with the bottom half and then change it back last. That way, we
> >> end up mapping it way to the top of 64-bit address space, which
> >> hopefully is less likely to conflict..
> >
> > Fixed it (finally).  I don't think moving the 64 bit probing around
> > would make a difference, since we'd restore its original value
> > anyway before moving on to the 32 bit probe which is where I think
> > the problem is.
>
> You couldn't just reorder the code the way it is now, you'd have to
> rearrange the way we do things for 64-bit BARs:
>
> -write FFFFFFFF to high part of 64-bit address (we end up moving the
> BAR to 0xFFFFFFFFC0000000 for example)
> -If any bits stick, we know what the size is now (more than 4GB of
> decode), so just change it back, we're done
> -If not, we need to check the low part, so write FFFFFFFF to low part
> of 64-bit address (BAR moves to 0xFFFFFFFFFFFFFFFF)
> -Check which bits stick and calculate the address
> -Change the low part of the address back (BAR moves to
> 0xFFFFFFFFC000000) -Change the high part of the address back (BAR
> moves to the original 0xC0000000 address)
>
> This means that at no point do we map the BAR anywhere near the top
> of 32-bit memory, so we should avoid this issue in this particular
> case. I don't think this strategy is too likely to break anything,
> surely less likely than disabling command bits. Jesse, you might want
> to try hacking up something like this and see what happens.

Ah yeah, that would probably work in this particular case, but doesn't 
seem very general.  I think just using type 1 accesses for non-extended 
config space is a bit more solid.

Jesse