Re: [Qemu-devel] [PATCH 1/9] Add stub functions for PCI device models to do PCI DMA

[Michael S. Tsirkin]

On Wed, Oct 12, 2011 at 02:11:37PM +1100, David Gibson wrote:
> On Sun, Oct 02, 2011 at 12:52:39PM +0200, Michael S. Tsirkin wrote:
> > On Sun, Oct 02, 2011 at 12:29:08PM +0200, Avi Kivity wrote:
> > > On 10/02/2011 12:25 PM, Michael S. Tsirkin wrote:
> > > >On Mon, Sep 05, 2011 at 02:34:56PM +1000, David Gibson wrote:
> > > >>  This patch adds functions to pci.[ch] to perform PCI DMA operations.  
> > > >> At
> > > >>  present, these are just stubs which perform directly cpu physical 
> > > >> memory
> > > >>  accesses.
> > > >>
> > > >>  Using these stubs, however, distinguishes PCI device DMA transactions 
> > > >> from
> > > >>  other accesses to physical memory, which will allow PCI IOMMU support 
> > > >> to
> > > >>  be added in one place, rather than updating every PCI driver at that 
> > > >> time.
> > > >>
> > > >>  That is, it allows us to update individual PCI drivers to support an 
> > > >> IOMMU
> > > >>  without having yet determined the details of how the IOMMU emulation 
> > > >> will
> > > >>  operate.  This will let us remove the most bitrot-sensitive part of an
> > > >>  IOMMU patch in advance.
> > > >>
> > > >>  Signed-off-by: David Gibson<address@hidden>
> > > >
> > > >So something I just thought about:
> > > >
> > > >all wrappers now go through cpu_physical_memory_rw.
> > > >This is a problem as e.g. virtio assumes that
> > > >accesses such as stw are atomic. cpu_physical_memory_rw
> > > >is a memcpy which makes no such guarantees.
> > > >
> > > 
> > > Let's change cpu_physical_memory_rw() to provide that guarantee for
> > > aligned two and four byte accesses.  Having separate paths just for
> > > that is not maintainable.
> > 
> > Well, we also have stX_phys convert to target native endian-ness
> > (nop for KVM but not necessarily for qemu).
> 
> Yes.. as do the stX_pci_dma() helpers.  They assume LE, rather than
> having two variants, because PCI is an LE spec, and all normal PCI
> devices work in LE.

IMO, not really. PCI devices do DMA any way they like.  LE is
probably more common because both ARM and x86 processors are LE.

>  If we need to model some perverse BE PCI device,
> it can reswap itself.

An explicit API for this would be cleaner.

-- 
MST