Re: [PATCH -tip 1/3] IA64: fix swiotlb alloc_coherent for non DMA_64BIT_MASK devices

[Yasunori Goto <y-goto@jp.fujitsu.com>]

> Before the dma ops unification, IA64 always uses GFP_DMA for
> dma_alloc_coherent like:
> 
> #define dma_alloc_coherent(dev, size, handle, gfp)	\
> 	platform_dma_alloc_coherent(dev, size, handle, (gfp) | GFP_DMA)
> 
> This GFP_DMA enforcement doesn't make sense for IOMMUs since they can
> do address translation to give addresses that devices can access
> to. The IOMMU drivers ignore the zone flag. However, this is still
> necessary for swiotlb since it can't do address translation.
> 
> We don't always need to use GFP_DMA for swiotlb. We need GFP_DMA for
> devices incapable of 64bit DMA.
> 
> This patch is sorta updated version of:
> 
> http://marc.info/?l=linux-kernel&m\x122638215612705&w=2
> 
> Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
> Cc: Yasunori Goto <y-goto@jp.fujitsu.com>
> ---
>  arch/ia64/kernel/pci-swiotlb.c |   10 +++++++++-
>  1 files changed, 9 insertions(+), 1 deletions(-)
> 
> diff --git a/arch/ia64/kernel/pci-swiotlb.c b/arch/ia64/kernel/pci-swiotlb.c
> index 717ad4f..573f02c 100644
> --- a/arch/ia64/kernel/pci-swiotlb.c
> +++ b/arch/ia64/kernel/pci-swiotlb.c
> @@ -13,8 +13,16 @@
>  int swiotlb __read_mostly;
>  EXPORT_SYMBOL(swiotlb);
>  
> +static void *ia64_swiotlb_alloc_coherent(struct device *dev, size_t size,
> +					 dma_addr_t *dma_handle, gfp_t gfp)
> +{
> +	if (dev->coherent_dma_mask != DMA_64BIT_MASK)
> +		gfp |= GFP_DMA;
> +	return swiotlb_alloc_coherent(dev, size, dma_handle, gfp);
> +}
> +
>  struct dma_map_ops swiotlb_dma_ops = {
> -	.alloc_coherent = swiotlb_alloc_coherent,
> +	.alloc_coherent = ia64_swiotlb_alloc_coherent,
>  	.free_coherent = swiotlb_free_coherent,
>  	.map_page = swiotlb_map_page,
>  	.unmap_page = swiotlb_unmap_page,
> -- 
> 1.6.0.6


Though I may be misunderstanding something, this initialization seems to be
called only when CONFIG_IA64_DIG_VTD is on.

CONFIG_IA64_DIG_VTD
  platform_dma_init() -> pci_iommu_alloc() -> pci_swiotlb_init()

CONFIG_IA64_DIG/CONFIG_IA64_GENERIC
  platform_dma_init() -> swiotlb_dma_init()

Is it intended? 
Do you wish GFP_DMA should be always ON when generic config?


Bye.

-- 
Yasunori Goto