Re: [PATCH v7 6/8] xen/arm/arm64: merge xen/mm32.c into xen/mm.c

[Stefano Stabellini]

On Fri, 7 Nov 2014, Catalin Marinas wrote:
> On Mon, Oct 27, 2014 at 03:09:26PM +0000, Stefano Stabellini wrote:
> > Merge xen/mm32.c into xen/mm.c.
> > As a consequence the code gets compiled on arm64 too: introduce a few
> > compat functions to actually be able to compile it.
> > 
> > Signed-off-by: Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>
> > Acked-by: Ian Campbell <ian.campbell@xxxxxxxxxx>
> 
> Since I missed the commit introducing mm32.c (340720be32d4 xen/arm:
> reimplement xen_dma_unmap_page & friends), I'll add a retrospective NAK ;).
> 
> The main reason is the asymmetry between dma map and unmap. With host
> swiotlb somehow getting !dma_capable(dev), you even risk leaking dom0
> swiotlb bounce buffers (on arm64).
> 
> > --- a/arch/arm/xen/mm.c
> > +++ b/arch/arm/xen/mm.c
> [...]
> > +/* functions called by SWIOTLB */
> > +
> > +static void dma_cache_maint(dma_addr_t handle, unsigned long offset,
> > +       size_t size, enum dma_data_direction dir,
> > +       void (*op)(const void *, size_t, int))
> > +{
> > +       unsigned long pfn;
> > +       size_t left = size;
> > +
> > +       pfn = (handle >> PAGE_SHIFT) + offset / PAGE_SIZE;
> > +       offset %= PAGE_SIZE;
> > +
> > +       do {
> > +               size_t len = left;
> > +               void *vaddr;
> > +
> > +               if (!pfn_valid(pfn))
> 
> Is this the pfn or the mfn? As you said in the previous email, there is
> no mfn_to_pfn() conversion, so that's actually in another address space
> where dom0 pfn_valid() would not make sense.

That is actually the mfn. The check works because dom0 is mapped 1:1, so
if the mfn is a valid pfn, then it means that it is a local page.


> Do you use this as a check for foreign pages? If yes, is the mfn for
> such pages guaranteed to be different from any valid dom0 pfn?

Yes and yes


> > +               {
> > +                       /* TODO: cache flush */
> 
> What does this TODO mean here? Which cases are not covered yet?

We are going to fill in the cache flush implementation for foreign pages
later, in patch 8/8.


> > +               } else {
> > +                       struct page *page = pfn_to_page(pfn);
> 
> If the pfn here is correct, can you not just have something like:
> 
> void xen_dma_unmap_page(struct device *hwdev, dma_addr_t handle,
> 		size_t size, enum dma_data_direction dir,
> 		struct dma_attrs *attrs)
> 
> {
> 	unsigned long pfn = handle >> PAGE_SHIFT;	/* could use some macros */
> 	if (!pfn_valid(pfn)) {
> 		/* FIXME */
> 		return;
> 	}
> 	__generic_dma_ops(hwdev)->unmap_page(hwdev, handle, size, dir, attrs);
> }

Yes, this is possible. Then in patch 8/8 I could remove the FIXME and
add a call to a function that issues the new GNTTABOP_cache_flush
hypercall. It would also remove the asymmetry you mentioned before
because we could do the same for map_page.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/