This patch allows the use of CMA for DMA coherent memory allocation.
At the moment if the input parameter "is_coherent" is set to true
the allocation is not made using the CMA, which I think is not the
desired behaviour.
Signed-off-by: Lorenzo Nava <lorenx4@xxxxxxxx>
---
Changes in v2:
correct __arm_dma_free() according to __dma_alloc() allocation
---
Changes in v3:
now __dma_alloc(), if 'is_coherent' is true, returns memory from CMA
if there is no need for atomic allocation. If CMA is not available
the function returns the result of __alloc_simple_buffer().
__arm_dma_free() frees memory according to the new alloc function
avoiding __dma_free_remap() for coherent DMA if CMA is not enable.
arm_dma_alloc() mark pages as cacheable if attrs are set by default
to NULL. If attrs is not NULL, attributes are preserved in the allocation.
Coherent allocation tested on Xilinx Zynq processor.
---
Changes in v4:
back to "if..else" code style for __dma_alloc()
avoided unnecessary __free_from_pool() call in __arm_dma_free()
---
arch/arm/mm/dma-mapping.c | 25 +++++++++++++++++--------
1 file changed, 17 insertions(+), 8 deletions(-)
diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index 7e7583d..7d4bcf5 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -645,14 +645,18 @@ static void *__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
size = PAGE_ALIGN(size);
want_vaddr = !dma_get_attr(DMA_ATTR_NO_KERNEL_MAPPING, attrs);
- if (is_coherent || nommu())
+ if (nommu())
+ addr = __alloc_simple_buffer(dev, size, gfp, &page);
+ else if (dev_get_cma_area(dev) && (gfp & __GFP_WAIT))
+ addr = __alloc_from_contiguous(dev, size, prot, &page,
+ caller, want_vaddr);
+ else if (is_coherent)
addr = __alloc_simple_buffer(dev, size, gfp, &page);
else if (!(gfp & __GFP_WAIT))
addr = __alloc_from_pool(size, &page);
- else if (!dev_get_cma_area(dev))
- addr = __alloc_remap_buffer(dev, size, gfp, prot, &page, caller, want_vaddr);
else
- addr = __alloc_from_contiguous(dev, size, prot, &page, caller, want_vaddr);
+ addr = __alloc_remap_buffer(dev, size, gfp, prot, &page,
+ caller, want_vaddr);
if (page)
*handle = pfn_to_dma(dev, page_to_pfn(page));
@@ -680,9 +684,14 @@ void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
dma_addr_t *handle, gfp_t gfp, struct dma_attrs *attrs)
{
- pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
+ pgprot_t prot;
void *memory;
+ if (attrs == NULL)
+ prot = PAGE_KERNEL;
+ else
+ prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
+
if (dma_alloc_from_coherent(dev, size, handle, &memory))
return memory;
@@ -735,12 +744,12 @@ static void __arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
size = PAGE_ALIGN(size);
- if (is_coherent || nommu()) {
+ if (nommu()) {
__dma_free_buffer(page, size);
- } else if (__free_from_pool(cpu_addr, size)) {
+ } else if (!is_coherent && __free_from_pool(cpu_addr, size)) {
return;
} else if (!dev_get_cma_area(dev)) {
- if (want_vaddr)
+ if (want_vaddr && !is_coherent)
__dma_free_remap(cpu_addr, size);
__dma_free_buffer(page, size);
} else {
--
1.7.10.4
On Mon, Jun 29, 2015 at 07:10:45PM +0200, Lorenzo Nava wrote:
> This patch allows the use of CMA for DMA coherent memory allocation.
> At the moment if the input parameter "is_coherent" is set to true
> the allocation is not made using the CMA, which I think is not the
> desired behaviour.
>
> Signed-off-by: Lorenzo Nava <lorenx4@xxxxxxxx>
It's nice to have a proper email address here.
> @@ -680,9 +684,14 @@ void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
> static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
> dma_addr_t *handle, gfp_t gfp, struct dma_attrs *attrs)
> {
> - pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
> + pgprot_t prot;
> void *memory;
>
> + if (attrs == NULL)
> + prot = PAGE_KERNEL;
> + else
> + prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
> +
> if (dma_alloc_from_coherent(dev, size, handle, &memory))
> return memory;
We have a slight inconsistency with commit 55af8a91640d (ARM: 8387/1:
arm/mm/dma-mapping.c: Add arm_coherent_dma_mmap) where the attrs are not
honoured for the coherent ops. That is, DMA_ATTRS_WRITE_COMBINE
doesn't have any effect for mmap but with your patch it still has an
effect on alloc.
I think we could just ignore it, the WRITE_COMBINE attr is a hint that
the arch may or may not implement and I haven't seen it used with
coherent mappings. So in this case just pass PAGE_KERNEL to
__dma_alloc() directly without calling __get_dma_pgprot().
(it also looks like commit 55af8a91640d left an unused attrs argument
for __arm_dma_mmap that can be removed as well)
Apart from the above, you can add:
Reviewed-by: Catalin Marinas <[email protected]>
On Tue, Jun 30, 2015 at 1:05 PM, Catalin Marinas
<[email protected]> wrote:
>
> On Mon, Jun 29, 2015 at 07:10:45PM +0200, Lorenzo Nava wrote:
> > This patch allows the use of CMA for DMA coherent memory allocation.
> > At the moment if the input parameter "is_coherent" is set to true
> > the allocation is not made using the CMA, which I think is not the
> > desired behaviour.
> >
> > Signed-off-by: Lorenzo Nava <lorenx4@xxxxxxxx>
>
> It's nice to have a proper email address here.
oops, sorry I thought it was made to avoid spam. This is an easy fix :-)
>
> > @@ -680,9 +684,14 @@ void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
> > static void *arm_coherent_dma_alloc(struct device *dev, size_t size,
> > dma_addr_t *handle, gfp_t gfp, struct dma_attrs *attrs)
> > {
> > - pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
> > + pgprot_t prot;
> > void *memory;
> >
> > + if (attrs == NULL)
> > + prot = PAGE_KERNEL;
> > + else
> > + prot = __get_dma_pgprot(attrs, PAGE_KERNEL);
> > +
> > if (dma_alloc_from_coherent(dev, size, handle, &memory))
> > return memory;
>
> We have a slight inconsistency with commit 55af8a91640d (ARM: 8387/1:
> arm/mm/dma-mapping.c: Add arm_coherent_dma_mmap) where the attrs are not
> honoured for the coherent ops. That is, DMA_ATTRS_WRITE_COMBINE
> doesn't have any effect for mmap but with your patch it still has an
> effect on alloc.
>
> I think we could just ignore it, the WRITE_COMBINE attr is a hint that
> the arch may or may not implement and I haven't seen it used with
> coherent mappings. So in this case just pass PAGE_KERNEL to
> __dma_alloc() directly without calling __get_dma_pgprot().
>
> (it also looks like commit 55af8a91640d left an unused attrs argument
> for __arm_dma_mmap that can be removed as well)
>
Right: ok I just pass PAGE_KERNEL to the alloc function, so we are
coherent with the mmap.
> Apart from the above, you can add:
>
> Reviewed-by: Catalin Marinas <[email protected]>
Thanks.