Re: [PATCH 6/7] dma-iommu: implement ->alloc_noncontiguous

From: Tomasz Figa
Date: Tue Feb 16 2021 - 03:17:26 EST


Hi Christoph


On Tue, Feb 2, 2021 at 6:51 PM Christoph Hellwig <hch@xxxxxx> wrote:
>
> Implement support for allocating a non-contiguous DMA region.
>
> Signed-off-by: Christoph Hellwig <hch@xxxxxx>
> ---
> drivers/iommu/dma-iommu.c | 35 +++++++++++++++++++++++++++++++++++
> 1 file changed, 35 insertions(+)
>
> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
> index 85cb004d7a44c6..4e0b170d38d57a 100644
> --- a/drivers/iommu/dma-iommu.c
> +++ b/drivers/iommu/dma-iommu.c
> @@ -718,6 +718,7 @@ static struct page **__iommu_dma_alloc_noncontiguous(struct device *dev,
> goto out_free_sg;
>
> sgt->sgl->dma_address = iova;
> + sgt->sgl->dma_length = size;
> return pages;
>
> out_free_sg:
> @@ -755,6 +756,36 @@ static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
> return NULL;
> }
>
> +#ifdef CONFIG_DMA_REMAP
> +static struct sg_table *iommu_dma_alloc_noncontiguous(struct device *dev,
> + size_t size, enum dma_data_direction dir, gfp_t gfp)
> +{
> + struct dma_sgt_handle *sh;
> +
> + sh = kmalloc(sizeof(*sh), gfp);
> + if (!sh)
> + return NULL;
> +
> + sh->pages = __iommu_dma_alloc_noncontiguous(dev, size, &sh->sgt, gfp,
> + PAGE_KERNEL, 0);

When working on the videobuf2 integration with Sergey I noticed that
we always pass 0 as DMA attrs here, which removes the ability for
drivers to use DMA_ATTR_ALLOC_SINGLE_PAGES.

It's quite important from a system stability point of view, because by
default the iommu_dma allocator would prefer big order allocations for
TLB locality reasons. For many devices, though, it doesn't really
affect the performance, because of random access patterns, so single
pages are good enough and reduce the risk of allocation failures or
latency due to fragmentation.

Do you think we could add the attrs parameter to the
dma_alloc_noncontiguous() API?

Best regards,
Tomasz