[PATCH v2 1/2] iommu/dma: Add support for DMA_ATTR_FORCE_CONTIGUOUS

[Robin Murphy]

Hi Geert,

On 27/01/17 15:34, Geert Uytterhoeven wrote:
> Add helpers for allocating physically contiguous DMA buffers to the
> generic IOMMU DMA code.  This can be useful when two or more devices
> with different memory requirements are involved in buffer sharing.
> 
> The iommu_dma_{alloc,free}_contiguous() functions complement the existing
> iommu_dma_{alloc,free}() functions, and allow architecture-specific code
> to implement support for the DMA_ATTR_FORCE_CONTIGUOUS attribute on
> systems with an IOMMU.  As this uses the CMA allocator, setting this
> attribute has a runtime-dependency on CONFIG_DMA_CMA.
> 
> Note that unlike the existing iommu_dma_alloc() helper,
> iommu_dma_alloc_contiguous() has no callback to flush pages.
> Ensuring the returned region is made visible to a non-coherent device is
> the responsibility of the caller.
> 
> Signed-off-by: Geert Uytterhoeven <geert+renesas at glider.be>
> ---
> v2:
>   - Provide standalone iommu_dma_{alloc,free}_contiguous() functions, as
>     requested by Robin Murphy,
>   - Simplify operations by getting rid of the page array/scatterlist
>     dance, as the buffer is contiguous,
>   - Move CPU cache magement into the caller, which is much simpler with
>     a single contiguous buffer.

Thanks for the rework, that's a lot easier to make sense of! Now, please
don't hate me, but...

> ---
>  drivers/iommu/dma-iommu.c | 72 +++++++++++++++++++++++++++++++++++++++++++++++
>  include/linux/dma-iommu.h |  4 +++
>  2 files changed, 76 insertions(+)
> 
> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
> index 2db0d641cf4505b5..8f8ed4426f9a3a12 100644
> --- a/drivers/iommu/dma-iommu.c
> +++ b/drivers/iommu/dma-iommu.c
> @@ -30,6 +30,7 @@
>  #include <linux/pci.h>
>  #include <linux/scatterlist.h>
>  #include <linux/vmalloc.h>
> +#include <linux/dma-contiguous.h>
>  
>  struct iommu_dma_msi_page {
>  	struct list_head	list;
> @@ -408,6 +409,77 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
>  }
>  
>  /**
> + * iommu_dma_free_contiguous - Free a buffer allocated by
> + *			       iommu_dma_alloc_contiguous()
> + * @dev: Device which owns this buffer
> + * @page: Buffer page pointer as returned by iommu_dma_alloc_contiguous()
> + * @size: Size of buffer in bytes
> + * @handle: DMA address of buffer
> + *
> + * Frees the pages associated with the buffer.
> + */
> +void iommu_dma_free_contiguous(struct device *dev, struct page *page,
> +		size_t size, dma_addr_t *handle)
> +{
> +	__iommu_dma_unmap(iommu_get_domain_for_dev(dev), *handle);
> +	dma_release_from_contiguous(dev, page, PAGE_ALIGN(size) >> PAGE_SHIFT);
> +	*handle = DMA_ERROR_CODE;
> +}
> +
> +/**
> + * iommu_dma_alloc_contiguous - Allocate and map a buffer contiguous in IOVA
> + *				and physical space
> + * @dev: Device to allocate memory for. Must be a real device attached to an
> + *	 iommu_dma_domain
> + * @size: Size of buffer in bytes
> + * @prot: IOMMU mapping flags
> + * @handle: Out argument for allocated DMA handle
> + *
> + * Return: Buffer page pointer, or NULL on failure.
> + *
> + * Note that unlike iommu_dma_alloc(), it's the caller's responsibility to
> + * ensure the returned region is made visible to the given non-coherent device.
> + */
> +struct page *iommu_dma_alloc_contiguous(struct device *dev, size_t size,
> +		int prot, dma_addr_t *handle)
> +{
> +	struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
> +	struct iova_domain *iovad = cookie_iovad(domain);
> +	dma_addr_t dma_addr;
> +	unsigned int count;
> +	struct page *page;
> +	struct iova *iova;
> +	int ret;
> +
> +	*handle = DMA_ERROR_CODE;
> +
> +	size = PAGE_ALIGN(size);
> +	count = size >> PAGE_SHIFT;
> +	page = dma_alloc_from_contiguous(dev, count, get_order(size));
> +	if (!page)
> +		return NULL;
> +
> +	iova = __alloc_iova(domain, size, dev->coherent_dma_mask);
> +	if (!iova)
> +		goto out_free_pages;
> +
> +	size = iova_align(iovad, size);
> +	dma_addr = iova_dma_addr(iovad, iova);
> +	ret = iommu_map(domain, dma_addr, page_to_phys(page), size, prot);
> +	if (ret < 0)
> +		goto out_free_iova;
> +
> +	*handle = dma_addr;
> +	return page;
> +
> +out_free_iova:
> +	__free_iova(iovad, iova);
> +out_free_pages:
> +	dma_release_from_contiguous(dev, page, count);
> +	return NULL;
> +}

...now that I can see it clearly, isn't this more or less just:

	page = dma_alloc_from_contiguous(dev, ...);
	if (page)
		dma_addr = iommu_dma_map_page(dev, page, ...);
?

Would it not be even simpler to just make those two calls directly from
the arm64 code?

Robin.

> +
> +/**
>   * iommu_dma_mmap - Map a buffer into provided user VMA
>   * @pages: Array representing buffer from iommu_dma_alloc()
>   * @size: Size of buffer in bytes
> diff --git a/include/linux/dma-iommu.h b/include/linux/dma-iommu.h
> index 7f7e9a7e3839966c..7eee62c2b0e752f7 100644
> --- a/include/linux/dma-iommu.h
> +++ b/include/linux/dma-iommu.h
> @@ -45,6 +45,10 @@ struct page **iommu_dma_alloc(struct device *dev, size_t size, gfp_t gfp,
>  		void (*flush_page)(struct device *, const void *, phys_addr_t));
>  void iommu_dma_free(struct device *dev, struct page **pages, size_t size,
>  		dma_addr_t *handle);
> +struct page *iommu_dma_alloc_contiguous(struct device *dev, size_t size,
> +		int prot, dma_addr_t *handle);
> +void iommu_dma_free_contiguous(struct device *dev, struct page *page,
> +		size_t size, dma_addr_t *handle);
>  
>  int iommu_dma_mmap(struct page **pages, size_t size, struct vm_area_struct *vma);
>  
>