Re: [PATCH v2 4/7] powerpc/mm: handle highmem pages in arch_dma_prep_coherent()

From: IBM

Date: Thu Nov 19 2026 - 09:36:24 EST


Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> writes:

> On 19.11.2026 13:07, Ritesh Harjani (IBM) wrote:
>> Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> writes:
>>
>>> arch_dma_prep_coherent() performs the cache maintenance on the address
>>> returned by page_address(). On 32-bit powerpc with CONFIG_HIGHMEM that
>>> is NULL for a highmem page which is not kmapped, so the flush would
>>> operate on the wrong virtual range and leave the dirty lines of the
>>> buffer in place.
>>>
>>> The generic DMA code may pass highmem pages to this hook:
>>> dma_direct_alloc() handles a highmem page returned by
>>> dma_alloc_contiguous() by forcing the remap path, which calls
>>> arch_dma_prep_coherent() on that page before creating the uncached
>>> mapping. This cannot happen on powerpc today only because it does not
>>> select HAVE_DMA_CONTIGUOUS, so dma-direct never gets a highmem page
>>> there, but nothing in the hook's contract rules it out.
>>>
>>> Use __dma_sync_page() with DMA_BIDIRECTIONAL, which the
>>> arch_sync_dma_for_*() hooks already use and which kmaps highmem pages
>>> one at a time. Without CONFIG_HIGHMEM it does the same
>>> flush_dcache_range() on the linear mapping as before.
>>>
>> Does that mean, this change can add additional overhead for
>> CONFIG_HIGHMEM=y kernels, where the buffer received still belongs to
>> lowmem? (Because of extra per-page book keeping like
>> local_irq_save/restore, kmap_xxx). Thoughts?
>>
>> Although functionally the change looks correct to me. But maybe
>> __dma_sync_page() on powerpc side could use some cleanups (outside of
>> this patch series perhaps).
> Indeed, __dma_sync_page() could be improved by calling __dma_sync_page_highmem()
> only for highmem pages. Something like:
>
>
> diff --git a/arch/powerpc/mm/dma-noncoherent.c b/arch/powerpc/mm/dma-noncoherent.c
> index e646256365d5..199ce0467b58 100644
> --- a/arch/powerpc/mm/dma-noncoherent.c
> +++ b/arch/powerpc/mm/dma-noncoherent.c
> @@ -97,11 +97,12 @@ static void __dma_sync_page(phys_addr_t paddr, size_t size, int dir)
>         unsigned offset = paddr & ~PAGE_MASK;
>

Should we return early if size=0 is passed. Although I couldn't find any
caller where this could happen, but since we are doing "paddr + size - 1",
hence wanted to flag it.


>  #ifdef CONFIG_HIGHMEM
> -       __dma_sync_page_highmem(page, offset, size, dir);
> -#else
> -       unsigned long start = (unsigned long)page_address(page) + offset;
> -       __dma_sync((void *)start, size, dir);
> +       if (PageHighMem(pfn_to_page(PHYS_PFN(paddr + size - 1)))) {
> +               __dma_sync_page_highmem(page, offset, size, dir);
> +               return;
> +       }
>  #endif
> +       __dma_sync(page_address(page) + offset, size, dir);
>  }
>
>  void arch_sync_dma_for_device(phys_addr_t paddr, size_t size,
>
>
> Do You want me to add it as another patch for this series?

I think this should work. I guess we should add this patch first and
then patch-4 on top of it.

Thanks for taking care of it. This at least will avoid the extra work
which we would have done otherwise for highmem kernels, where the buffer
passed to this still comes from lowmem.

-ritesh