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@xxxxxxxxx> > --- > 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); > >