Re: [PATCH 09/10] xen/arm: introduce phys/dma translations in xen_dma_sync_for_*

[Julien Grall]

On Thu, 21 May 2020 at 21:08, Stefano Stabellini <sstabellini@xxxxxxxxxx> wrote:
>
> On Thu, 21 May 2020, Julien Grall wrote:
> > > @@ -97,8 +98,7 @@ bool xen_arch_need_swiotlb(struct device *dev,
> > >                        phys_addr_t phys,
> > >                        dma_addr_t dev_addr)
> > >   {
> > > -   unsigned int xen_pfn = XEN_PFN_DOWN(phys);
> > > -   unsigned int bfn = XEN_PFN_DOWN(dev_addr);
> > > +   unsigned int bfn = XEN_PFN_DOWN(dma_to_phys(dev, dev_addr));
> > >             /*
> > >      * The swiotlb buffer should be used if
> > > @@ -115,7 +115,7 @@ bool xen_arch_need_swiotlb(struct device *dev,
> > >      * require a bounce buffer because the device doesn't support coherent
> > >      * memory and we are not able to flush the cache.
> > >      */
> > > -   return (!hypercall_cflush && (xen_pfn != bfn) &&
> > > +   return (!hypercall_cflush && !pfn_valid(bfn) &&
> >
> > I believe this change is incorrect. The bfn is a frame based on Xen page
> > granularity (always 4K) while pfn_valid() is expecting a frame based on the
> > Kernel page granularity.
>
> Given that kernel granularity >= xen granularity it looks like it would
> be safe to use PFN_DOWN instead of XEN_PFN_DOWN:
>
>   unsigned int bfn = PFN_DOWN(dma_to_phys(dev, dev_addr));

Yes. But is the change worth it though? pfn_valid() is definitely
going to be more expensive than the current check.

Cheers,