diff mbox series

[17/18] dma-iommu: implement ->alloc_noncoherent

Message ID 20200915155122.1768241-18-hch@lst.de (mailing list archive)
State New, archived
Headers show
Series [01/18] media/v4l2: remove V4L2-FLAG-MEMORY-NON-CONSISTENT flag | expand

Commit Message

Christoph Hellwig Sept. 15, 2020, 3:51 p.m. UTC
Implement the alloc_noncoherent method to provide memory that is neither
coherent not contiguous.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 drivers/iommu/dma-iommu.c | 41 +++++++++++++++++++++++++++++++++++----
 1 file changed, 37 insertions(+), 4 deletions(-)

Comments

Tomasz Figa Sept. 25, 2020, 6:46 p.m. UTC | #1
Hi Christoph,

On Tue, Sep 15, 2020 at 05:51:21PM +0200, Christoph Hellwig wrote:
> Implement the alloc_noncoherent method to provide memory that is neither
> coherent not contiguous.
> 
> Signed-off-by: Christoph Hellwig <hch@lst.de>
> ---
>  drivers/iommu/dma-iommu.c | 41 +++++++++++++++++++++++++++++++++++----
>  1 file changed, 37 insertions(+), 4 deletions(-)
> 

Sorry for being late to the party and thanks a lot for the patch. Please see my
comments inline.

[snip]
> @@ -1052,6 +1055,34 @@ static void *iommu_dma_alloc(struct device *dev, size_t size,
>  	return cpu_addr;
>  }
>  
> +#ifdef CONFIG_DMA_REMAP
> +static void *iommu_dma_alloc_noncoherent(struct device *dev, size_t size,
> +		dma_addr_t *handle, enum dma_data_direction dir, gfp_t gfp)
> +{
> +	if (!gfpflags_allow_blocking(gfp)) {
> +		struct page *page;
> +
> +		page = dma_common_alloc_pages(dev, size, handle, dir, gfp);
> +		if (!page)
> +			return NULL;
> +		return page_address(page);
> +	}
> +
> +	return iommu_dma_alloc_remap(dev, size, handle, gfp | __GFP_ZERO,
> +				     PAGE_KERNEL, 0);

iommu_dma_alloc_remap() makes use of the DMA_ATTR_ALLOC_SINGLE_PAGES attribute
to optimize the allocations for devices which don't care about how contiguous
the backing memory is. Do you think we could add an attrs argument to this
function and pass it there?

As ARM is being moved to the common iommu-dma layer as well, we'll probably
make use of the argument to support the DMA_ATTR_NO_KERNEL_MAPPING attribute to
conserve the vmalloc area.

Best regards,
Tomasz
Christoph Hellwig Sept. 26, 2020, 2:14 p.m. UTC | #2
On Fri, Sep 25, 2020 at 06:46:22PM +0000, Tomasz Figa wrote:
> > +static void *iommu_dma_alloc_noncoherent(struct device *dev, size_t size,
> > +		dma_addr_t *handle, enum dma_data_direction dir, gfp_t gfp)
> > +{
> > +	if (!gfpflags_allow_blocking(gfp)) {
> > +		struct page *page;
> > +
> > +		page = dma_common_alloc_pages(dev, size, handle, dir, gfp);
> > +		if (!page)
> > +			return NULL;
> > +		return page_address(page);
> > +	}
> > +
> > +	return iommu_dma_alloc_remap(dev, size, handle, gfp | __GFP_ZERO,
> > +				     PAGE_KERNEL, 0);
> 
> iommu_dma_alloc_remap() makes use of the DMA_ATTR_ALLOC_SINGLE_PAGES attribute
> to optimize the allocations for devices which don't care about how contiguous
> the backing memory is. Do you think we could add an attrs argument to this
> function and pass it there?
> 
> As ARM is being moved to the common iommu-dma layer as well, we'll probably
> make use of the argument to support the DMA_ATTR_NO_KERNEL_MAPPING attribute to
> conserve the vmalloc area.

We could probably at it.  However I wonder why this is something the
drivers should care about.  Isn't this really something that should
be a kernel-wide policy for a given system?
Tomasz Figa Sept. 26, 2020, 3:25 p.m. UTC | #3
On Sat, Sep 26, 2020 at 4:14 PM Christoph Hellwig <hch@lst.de> wrote:
>
> On Fri, Sep 25, 2020 at 06:46:22PM +0000, Tomasz Figa wrote:
> > > +static void *iommu_dma_alloc_noncoherent(struct device *dev, size_t size,
> > > +           dma_addr_t *handle, enum dma_data_direction dir, gfp_t gfp)
> > > +{
> > > +   if (!gfpflags_allow_blocking(gfp)) {
> > > +           struct page *page;
> > > +
> > > +           page = dma_common_alloc_pages(dev, size, handle, dir, gfp);
> > > +           if (!page)
> > > +                   return NULL;
> > > +           return page_address(page);
> > > +   }
> > > +
> > > +   return iommu_dma_alloc_remap(dev, size, handle, gfp | __GFP_ZERO,
> > > +                                PAGE_KERNEL, 0);
> >
> > iommu_dma_alloc_remap() makes use of the DMA_ATTR_ALLOC_SINGLE_PAGES attribute
> > to optimize the allocations for devices which don't care about how contiguous
> > the backing memory is. Do you think we could add an attrs argument to this
> > function and pass it there?
> >
> > As ARM is being moved to the common iommu-dma layer as well, we'll probably
> > make use of the argument to support the DMA_ATTR_NO_KERNEL_MAPPING attribute to
> > conserve the vmalloc area.
>
> We could probably at it.  However I wonder why this is something the
> drivers should care about.  Isn't this really something that should
> be a kernel-wide policy for a given system?

There are IOMMUs out there which support huge pages and those can
benefit *some* hardware depending on what kind of accesses they
perform, possibly on a per-buffer basis. At the same time, order > 0
allocations can be expensive, significantly affecting allocation
latency, so for devices which don't care about huge pages anyone would
prefer simple single-page allocations. Currently the drivers know the
best on whether the hardware they drive would care. There are some
decision factors listed in the documentation [1].

I can imagine cases where drivers could not be the best to decide
about this - for example, the workload could vary depending on the
userspace or a product decision regarding the performance vs
allocation latency, but we haven't seen such cases in practice yet.

[1] https://www.kernel.org/doc/html/latest/core-api/dma-attributes.html?highlight=dma_attr_alloc_single_pages#dma-attr-alloc-single-pages

Best regards,
Tomasz
diff mbox series

Patch

diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
index 00a5b49248e334..c12c1dc43d312e 100644
--- a/drivers/iommu/dma-iommu.c
+++ b/drivers/iommu/dma-iommu.c
@@ -572,6 +572,7 @@  static struct page **__iommu_dma_alloc_pages(struct device *dev,
  * @size: Size of buffer in bytes
  * @dma_handle: Out argument for allocated DMA handle
  * @gfp: Allocation flags
+ * @prot: pgprot_t to use for the remapped mapping
  * @attrs: DMA attributes for this allocation
  *
  * If @size is less than PAGE_SIZE, then a full CPU page will be allocated,
@@ -580,14 +581,14 @@  static struct page **__iommu_dma_alloc_pages(struct device *dev,
  * Return: Mapped virtual address, or NULL on failure.
  */
 static void *iommu_dma_alloc_remap(struct device *dev, size_t size,
-		dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs)
+		dma_addr_t *dma_handle, gfp_t gfp, pgprot_t prot,
+		unsigned long attrs)
 {
 	struct iommu_domain *domain = iommu_get_dma_domain(dev);
 	struct iommu_dma_cookie *cookie = domain->iova_cookie;
 	struct iova_domain *iovad = &cookie->iovad;
 	bool coherent = dev_is_dma_coherent(dev);
 	int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
-	pgprot_t prot = dma_pgprot(dev, PAGE_KERNEL, attrs);
 	unsigned int count, min_size, alloc_sizes = domain->pgsize_bitmap;
 	struct page **pages;
 	struct sg_table sgt;
@@ -1030,8 +1031,10 @@  static void *iommu_dma_alloc(struct device *dev, size_t size,
 	gfp |= __GFP_ZERO;
 
 	if (IS_ENABLED(CONFIG_DMA_REMAP) && gfpflags_allow_blocking(gfp) &&
-	    !(attrs & DMA_ATTR_FORCE_CONTIGUOUS))
-		return iommu_dma_alloc_remap(dev, size, handle, gfp, attrs);
+	    !(attrs & DMA_ATTR_FORCE_CONTIGUOUS)) {
+		return iommu_dma_alloc_remap(dev, size, handle, gfp,
+				dma_pgprot(dev, PAGE_KERNEL, attrs), attrs);
+	}
 
 	if (IS_ENABLED(CONFIG_DMA_DIRECT_REMAP) &&
 	    !gfpflags_allow_blocking(gfp) && !coherent)
@@ -1052,6 +1055,34 @@  static void *iommu_dma_alloc(struct device *dev, size_t size,
 	return cpu_addr;
 }
 
+#ifdef CONFIG_DMA_REMAP
+static void *iommu_dma_alloc_noncoherent(struct device *dev, size_t size,
+		dma_addr_t *handle, enum dma_data_direction dir, gfp_t gfp)
+{
+	if (!gfpflags_allow_blocking(gfp)) {
+		struct page *page;
+
+		page = dma_common_alloc_pages(dev, size, handle, dir, gfp);
+		if (!page)
+			return NULL;
+		return page_address(page);
+	}
+
+	return iommu_dma_alloc_remap(dev, size, handle, gfp | __GFP_ZERO,
+				     PAGE_KERNEL, 0);
+}
+
+static void iommu_dma_free_noncoherent(struct device *dev, size_t size,
+		void *cpu_addr, dma_addr_t handle, enum dma_data_direction dir)
+{
+	__iommu_dma_unmap(dev, handle, size);
+	__iommu_dma_free(dev, size, cpu_addr);
+}
+#else
+#define iommu_dma_alloc_noncoherent		NULL
+#define iommu_dma_free_noncoherent		NULL
+#endif /* CONFIG_DMA_REMAP */
+
 static int iommu_dma_mmap(struct device *dev, struct vm_area_struct *vma,
 		void *cpu_addr, dma_addr_t dma_addr, size_t size,
 		unsigned long attrs)
@@ -1122,6 +1153,8 @@  static const struct dma_map_ops iommu_dma_ops = {
 	.free			= iommu_dma_free,
 	.alloc_pages		= dma_common_alloc_pages,
 	.free_pages		= dma_common_free_pages,
+	.alloc_noncoherent	= iommu_dma_alloc_noncoherent,
+	.free_noncoherent	= iommu_dma_free_noncoherent,
 	.mmap			= iommu_dma_mmap,
 	.get_sgtable		= iommu_dma_get_sgtable,
 	.map_page		= iommu_dma_map_page,