Re: [PATCH v3 5/6] iommu/mediatek: Add mt8173 IOMMU driver

[Will Deacon <will.deacon-5wv7dgnIgG8@public.gmane.org>]

On Wed, Jul 29, 2015 at 06:41:31AM +0100, Yong Wu wrote:
> On Mon, 2015-07-27 at 16:49 +0100, Robin Murphy wrote:
> > On 27/07/15 16:31, Russell King - ARM Linux wrote:
> > > On Mon, Jul 27, 2015 at 02:23:26PM +0100, Robin Murphy wrote:
> > >> On 16/07/15 10:04, Yong Wu wrote:
> > >>> This patch adds support for mediatek m4u (MultiMedia Memory Management
> > >>> Unit).
> > >>>
> > >>> Signed-off-by: Yong Wu <yong.wu-NuS5LvNUpcJWk0Htik3J/w@public.gmane.org>
> > >> [...]
> > >>> +static void mtk_iommu_flush_pgtable(void *ptr, size_t size, void *cookie)
> > >>> +{
> > >>> +       struct mtk_iommu_domain *domain = cookie;
> > >>> +       unsigned long offset = (unsigned long)ptr & ~PAGE_MASK;
> > >>> +
> > >>> +       dma_map_page(domain->data->dev, virt_to_page(ptr), offset,
> > >>> +                    size, DMA_TO_DEVICE);
> > >>
> > >> Nit: this looks like it may as well be dma_map_single.
> > >>
> > >> It would probably be worth following it with a matching unmap too, just to
> > >> avoid any possible leakage bugs (especially if this M4U ever appears in a
> > >> SoC supporting RAM above the 32-bit boundary).
> > >
> > > Why not do the job properly?  Take a look at how I implemented the
> > > streaming DMA API on Tegra SMMU (patch set recently sent out earlier
> > > today).
> > >
> > > There's no need for hacks like dma_map_page() (and discarding it's
> > > return value) or dma_map_page() followed by dma_unmap_page().
> > 
> > Indeed, as it happens I do have a branch where I prototyped that for the 
> > long-descriptor io-pgtable-arm code a while ago; this discussion has 
> > prompted me to dig it up again. Stay tuned, folks...
> 
> Hi Russell, Robin,
> 
>      From I see in arm-smmu-v3.c in v4.2-rc1, 
>      
>      The flush_pgtable seems like this:
> //==========
> 	dma_addr_t dma_addr;
> 
> 	dma_addr = dma_map_page(dev, ptr, size, DMA_TO_DEVICE);
> 
> 	if (dma_mapping_error(dev, dma_addr))
> 		dev_err(dev, "failed to flush pgtable at %p\n", ptr);
> 	else
> 		dma_unmap_page(dev, dma_addr, size, DMA_TO_DEVICE);
> //==========
>        I will change map like this and use dma_map_single instead.
> 
>        Is this also seems to be not proper?
> 
>        Then how to do it?  add this before unmap? :
>        dma_sync_single_for_device(dev, dma_addr, size, DMA_TO_DEVICE);

Robin's proposed a series fixing this in the io-pgtable code:

  http://lists.linuxfoundation.org/pipermail/iommu/2015-July/013821.html

which is currently under review. Please take a look and give comments!
Once merged, the code you cite above will be removed.

Will