[PATCH v5 11/13] xen: introduce xen_alloc/free_coherent_pages

[Catalin Marinas]

On Thu, Sep 12, 2013 at 03:44:50PM +0100, Stefano Stabellini wrote:
> On Thu, 12 Sep 2013, Catalin Marinas wrote:
> > On Wed, Sep 11, 2013 at 06:34:17PM +0100, Stefano Stabellini wrote:
> > > On Wed, 11 Sep 2013, Catalin Marinas wrote:
> > > >  static inline void set_dma_ops(struct device *dev, struct dma_map_ops *ops)
> > > >  {
> > > >  	BUG_ON(!dev);
> > > > @@ -32,7 +39,7 @@ static inline void set_dma_ops(struct device *dev, struct dma_map_ops *ops)
> > > >  
> > > >  static inline int dma_set_mask(struct device *dev, u64 mask)
> > > >  {
> > > > -	return get_dma_ops(dev)->set_dma_mask(dev, mask);
> > > > +	return __get_dma_ops(dev)->set_dma_mask(dev, mask);
> > > >  }
> > > >  
> > > >  #ifdef __arch_page_to_dma
> > > 
> > > I don't understand the reason for this change though: shouldn't
> > > set_dma_mask go via the "default" (whatever that is), like the others?
> > 
> > Under Xen, this would mean xen_dma_ops->set_dma_mask(). I thought you
> > want to set the default dma address.
> > 
> > I can see why you may want set_dma_mask() for xen_dma_ops but in also
> > needs to be passed to __get_dma_ops(dev)->set_dma_ops() otherwise the
> > actual dma_alloc_attrs() will miss the mask (and on arm64 it does
> > bouncing via swiotlb).
>  
> xen_dma_ops->set_dma_mask() is implemented by
> drivers/xen/swiotlb-xen.c:xen_swiotlb_set_dma_mask (provided by this
> patch series). The function takes care of setting:
> 
> *dev->dma_mask = dma_mask;
> 
> I think that should be enough to make dma_alloc_attrs happy.

I think it would be safer to call __get_dma_ops()->set_dma_mask() in
case the driver overrides that and does more than dev->dma_mask setting.
I don't know whether this is the case though.

-- 
Catalin