Re: [PATCH] KVM: Fix page-crossing MMIO

[Avi Kivity <avi@redhat.com>]

On 04/19/2012 11:59 AM, Gleb Natapov wrote:
> On Wed, Apr 18, 2012 at 07:22:47PM +0300, Avi Kivity wrote:
> > MMIO that are split across a page boundary are currently broken - the
> > code does not expect to be aborted by the exit to userspace for the
> > first MMIO fragment.
> > 
> > This patch fixes the problem by generalizing the current code for handling
> > 16-byte MMIOs to handle a number of "fragments", and changes the MMIO
> > code to create those fragments.
> > 
> For 16 bit IO userspace will see two 8bit writes. Is this OK? 

I believe so.  The Pentium bus was 64-bit wide with no address lines
A0-A2, instead it had byte enables BE0-BE7.  So a write that crosses a
quadword boundary would be split.  Similarly the 32-bit PCI bus has 4
byte enables, so anything that crosses a dword boundary is split. 
LOCKed ops are implemented by asserting a signal during the two
operations.  No idea how it's implemented with modern busless
processors, but the same semantics have been kept, probably.

Even more interesting is that a 32-bit write is seen as an 8 bit write
followed by a 24-bit write.

Note also that the two accesses can target different devices.

> Is there
> real code that does this kind of IO, or the patch is for correctness
> only?

I should have mentioned it in the changelog - it's Windows 95.  There is
also a Red Hat Linux version that does something like this, I remember
adding some code in qemu-kvm to handle the 3-byte write.  So I guess
this fixes a regression.

(not sure - in the Windows 95 case we have a write to VGA that crosses a
page boundary (but the pages are physically contiguous, nothing strange
is going on; the RHL case was writing partially to RAM and partially to
unallocated memory, so it might work with the current code).

-- 
error compiling committee.c: too many arguments to function