Re: [Qemu-devel] [PATCH 05/14] eepro100: Use PCI DMA stub functions

[David Gibson <david@gibson.dropbear.id.au>]

On Thu, Nov 03, 2011 at 02:25:45PM +0200, Michael S. Tsirkin wrote:
> On Thu, Nov 03, 2011 at 04:16:34PM +1100, David Gibson wrote:
> > On Wed, Nov 02, 2011 at 09:16:34AM +0200, Michael S. Tsirkin wrote:
> > > On Mon, Oct 31, 2011 at 05:06:49PM +1100, David Gibson wrote:
> > > > From: Eduard - Gabriel Munteanu <eduard.munteanu@linux360.ro>
> > [snip]
> > > > @@ -744,21 +713,26 @@ static void dump_statistics(EEPRO100State * s)
> > > >       * values which really matter.
> > > >       * Number of data should check configuration!!!
> > > >       */
> > > > -    cpu_physical_memory_write(s->statsaddr, &s->statistics, s->stats_size);
> > > > -    e100_stl_le_phys(s->statsaddr + 0, s->statistics.tx_good_frames);
> > > > -    e100_stl_le_phys(s->statsaddr + 36, s->statistics.rx_good_frames);
> > > > -    e100_stl_le_phys(s->statsaddr + 48, s->statistics.rx_resource_errors);
> > > > -    e100_stl_le_phys(s->statsaddr + 60, s->statistics.rx_short_frame_errors);
> > > > +    pci_dma_write(&s->dev, s->statsaddr,
> > > > +                  (uint8_t *) &s->statistics, s->stats_size);
> > > > +    stl_le_pci_dma(&s->dev, s->statsaddr + 0,
> > > > +                   s->statistics.tx_good_frames);
> > > > +    stl_le_pci_dma(&s->dev, s->statsaddr + 36,
> > > > +                   s->statistics.rx_good_frames);
> > > > +    stl_le_pci_dma(&s->dev, s->statsaddr + 48,
> > > > +                   s->statistics.rx_resource_errors);
> > > > +    stl_le_pci_dma(&s->dev, s->statsaddr + 60,
> > > > +                   s->statistics.rx_short_frame_errors);
> > > 
> > > This might introduce a bug: stlXX APIs assume aligned addresses,
> > > an address in statsaddr is user-controlled so I'm not sure
> > > it's always aligned.
> > > 
> > > Why isn't the patch simply replacing cpu_physical_memory_read
> > > with pci_XXX ? Any cleanups should be done separately.
> > 
> > Because it seemed like a good idea at the time.  When I first wrote
> > this, the possibility of unaligned addresses wasn't obvious to me.
> > So, I'm working on fixing this now.  I can take one of two approaches:
> > 
> >  - Simply revert this part of the change, reinstate the e100_stl
> > functions as calling into dma_write().
> > 
> >  - Alter the stX_dma() functions to work for unaligned addresses (by
> > falling back to dma_rw() in that case).  This is a little more
> > involved but might make device writing safer in future.
> 
> Yes but then we lose the atomicity guarantee. So this might
> still result in subtle emulation bugs.

Sorry, I should have been clearer - I was planning to fall back to
dma_rw() *only* for unaligned addresses.  Aligned addresses would
still have the atomicity guarantee.

> > Anthony, Michael, any preferred direction here?
> 
> For 1.0 I'd go for option 1 as the simplest.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson