On Thu, 2018-07-26 at 11:17 +0200, Christoph Hellwig wrote:
> On Tue, Jul 24, 2018 at 01:10:00PM +0300, Eugeniy Paltsev wrote:
> > Refactoring, no functional change intended.
> >
[snip]
> >
> > *dma_handle = paddr;
> >
> > + /*
> > + * - A coherent buffer needs MMU mapping to enforce non-cachability
> > + * - A highmem page needs a virtual handle (hence MMU mapping)
> > + * independent of cachability.
> > + * kvaddr is kernel Virtual address (0x7000_0000 based)
> > + */
> > + if (PageHighMem(page) || need_coh) {
>
> dma_alloc_attrs clears __GFP_HIGHMEM from the passed in gfp mask, so
> you'll never get a highmem page here.
>
Nice catch, thanks.
Will remove check for highmem page in next patch version.
> That also means you can merge this conditional with the one for the cache
> writeback and invalidation and kill the need_coh flag entirely.
>
> > kvaddr = ioremap_nocache(paddr, size);
> > if (kvaddr == NULL) {
> > __free_pages(page, order);
> > @@ -81,11 +75,9 @@ void arch_dma_free(struct device *dev, size_t size, void *vaddr,
> > {
> > phys_addr_t paddr = dma_handle;
> > struct page *page = virt_to_page(paddr);
> > - int is_non_coh = 1;
> > -
> > - is_non_coh = (attrs & DMA_ATTR_NON_CONSISTENT);
> > + bool is_coh = !(attrs & DMA_ATTR_NON_CONSISTENT);
> >
> > - if (PageHighMem(page) || !is_non_coh)
> > + if (PageHighMem(page) || is_coh)
> > iounmap((void __force __iomem *)vaddr);
> >
>
> Same here.
>
> Also if you clean this up it would be great to take the per-device pfn offset
> into account, even if that isn't used anywhere on arc yet, that is call
> phys_to_dma and dma_to_phys to convert to an from the dma address.
Ok, I'll look at it.
Probably I'll implement it as a separate patch as it is irrelevant to this
patch series topic.
--
Eugeniy Paltsev
