We used to use generic implementation of dma_map_ops.mmap which is
dma_common_mmap() but that only worked for simpler cached mappings when
vaddr = paddr.
If a driver requests uncached DMA buffer kernel maps it to virtual
address so that MMU gets involved and page uncached status takes into
account. In that case usage of dma_common_mmap() lead to mapping of
vaddr to vaddr for user-space which is obviously wrong. For more detals
please refer to verbose explanation here [1].
So here we implement our own version of mmap() which always deals
with dma_addr and maps underlying memory to user-space properly
(note that DMA buffer mapped to user-space is always uncached
because there's no way to properly manage cache from user-space).
[1] https://lkml.org/lkml/2016/10/26/973
Signed-off-by: Alexey Brodkin <[email protected]>
Reviewed-by: Catalin Marinas <[email protected]>
Cc: Marek Szyprowski <[email protected]>
Cc: Vineet Gupta <[email protected]>
Cc: <[email protected]>
---
Changes v1 -> v2:
* Added plat_dma_to_phys wrapper around dma_addr
arch/arc/mm/dma.c | 26 ++++++++++++++++++++++++++
1 file changed, 26 insertions(+)
diff --git a/arch/arc/mm/dma.c b/arch/arc/mm/dma.c
index 20afc65e22dc..9288851d43a0 100644
--- a/arch/arc/mm/dma.c
+++ b/arch/arc/mm/dma.c
@@ -105,6 +105,31 @@ static void arc_dma_free(struct device *dev, size_t size, void *vaddr,
__free_pages(page, get_order(size));
}
+static int arc_dma_mmap(struct device *dev, struct vm_area_struct *vma,
+ void *cpu_addr, dma_addr_t dma_addr, size_t size,
+ unsigned long attrs)
+{
+ unsigned long user_count = vma_pages(vma);
+ unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
+ unsigned long pfn = __phys_to_pfn(plat_dma_to_phys(dev, dma_addr));
+ unsigned long off = vma->vm_pgoff;
+ int ret = -ENXIO;
+
+ vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
+
+ if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
+ return ret;
+
+ if (off < count && user_count <= (count - off)) {
+ ret = remap_pfn_range(vma, vma->vm_start,
+ pfn + off,
+ user_count << PAGE_SHIFT,
+ vma->vm_page_prot);
+ }
+
+ return ret;
+}
+
/*
* streaming DMA Mapping API...
* CPU accesses page via normal paddr, thus needs to explicitly made
@@ -193,6 +218,7 @@ static int arc_dma_supported(struct device *dev, u64 dma_mask)
struct dma_map_ops arc_dma_ops = {
.alloc = arc_dma_alloc,
.free = arc_dma_free,
+ .mmap = arc_dma_mmap,
.map_page = arc_dma_map_page,
.map_sg = arc_dma_map_sg,
.sync_single_for_device = arc_dma_sync_single_for_device,
--
2.7.4
On 11/03/2016 08:06 AM, Alexey Brodkin wrote:
> We used to use generic implementation of dma_map_ops.mmap which is
> dma_common_mmap() but that only worked for simpler cached mappings when
> vaddr = paddr.
>
> If a driver requests uncached DMA buffer kernel maps it to virtual
> address so that MMU gets involved and page uncached status takes into
> account. In that case usage of dma_common_mmap() lead to mapping of
> vaddr to vaddr for user-space which is obviously wrong. For more detals
> please refer to verbose explanation here [1].
>
> So here we implement our own version of mmap() which always deals
> with dma_addr and maps underlying memory to user-space properly
> (note that DMA buffer mapped to user-space is always uncached
> because there's no way to properly manage cache from user-space).
>
> [1] https://lkml.org/lkml/2016/10/26/973
>
> Signed-off-by: Alexey Brodkin <[email protected]>
> Reviewed-by: Catalin Marinas <[email protected]>
> Cc: Marek Szyprowski <[email protected]>
> Cc: Vineet Gupta <[email protected]>
> Cc: <[email protected]>
I've added a stable 4.5+, since ARC didn't use dma ops until 4.5-rc1.
Pushed to for-curr !
Thx,
-Vineet
> ---
>
> Changes v1 -> v2:
> * Added plat_dma_to_phys wrapper around dma_addr
>
> arch/arc/mm/dma.c | 26 ++++++++++++++++++++++++++
> 1 file changed, 26 insertions(+)
>
> diff --git a/arch/arc/mm/dma.c b/arch/arc/mm/dma.c
> index 20afc65e22dc..9288851d43a0 100644
> --- a/arch/arc/mm/dma.c
> +++ b/arch/arc/mm/dma.c
> @@ -105,6 +105,31 @@ static void arc_dma_free(struct device *dev, size_t size, void *vaddr,
> __free_pages(page, get_order(size));
> }
>
> +static int arc_dma_mmap(struct device *dev, struct vm_area_struct *vma,
> + void *cpu_addr, dma_addr_t dma_addr, size_t size,
> + unsigned long attrs)
> +{
> + unsigned long user_count = vma_pages(vma);
> + unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT;
> + unsigned long pfn = __phys_to_pfn(plat_dma_to_phys(dev, dma_addr));
> + unsigned long off = vma->vm_pgoff;
> + int ret = -ENXIO;
> +
> + vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
> +
> + if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
> + return ret;
> +
> + if (off < count && user_count <= (count - off)) {
> + ret = remap_pfn_range(vma, vma->vm_start,
> + pfn + off,
> + user_count << PAGE_SHIFT,
> + vma->vm_page_prot);
> + }
> +
> + return ret;
> +}
> +
> /*
> * streaming DMA Mapping API...
> * CPU accesses page via normal paddr, thus needs to explicitly made
> @@ -193,6 +218,7 @@ static int arc_dma_supported(struct device *dev, u64 dma_mask)
> struct dma_map_ops arc_dma_ops = {
> .alloc = arc_dma_alloc,
> .free = arc_dma_free,
> + .mmap = arc_dma_mmap,
> .map_page = arc_dma_map_page,
> .map_sg = arc_dma_map_sg,
> .sync_single_for_device = arc_dma_sync_single_for_device,
>
Hi Vineet,
> -----Original Message-----
> From: Vineet Gupta [mailto:[email protected]]
> Sent: Thursday, November 03, 2016 8:04 PM
> To: Alexey Brodkin <[email protected]>; [email protected]
> Cc: [email protected]; [email protected]; Vineet Gupta <[email protected]>; Marek Szyprowski
> <[email protected]>; [email protected]
> Subject: Re: [PATCH v2] arc: Implement arch-specific dma_map_ops.mmap
>
> On 11/03/2016 08:06 AM, Alexey Brodkin wrote:
> > We used to use generic implementation of dma_map_ops.mmap which is
> > dma_common_mmap() but that only worked for simpler cached mappings when
> > vaddr = paddr.
> >
> > If a driver requests uncached DMA buffer kernel maps it to virtual
> > address so that MMU gets involved and page uncached status takes into
> > account. In that case usage of dma_common_mmap() lead to mapping of
> > vaddr to vaddr for user-space which is obviously wrong. For more detals
> > please refer to verbose explanation here [1].
> >
> > So here we implement our own version of mmap() which always deals
> > with dma_addr and maps underlying memory to user-space properly
> > (note that DMA buffer mapped to user-space is always uncached
> > because there's no way to properly manage cache from user-space).
> >
> > [1] https://lkml.org/lkml/2016/10/26/973
> >
> > Signed-off-by: Alexey Brodkin <[email protected]>
> > Reviewed-by: Catalin Marinas <[email protected]>
> > Cc: Marek Szyprowski <[email protected]>
> > Cc: Vineet Gupta <[email protected]>
> > Cc: <[email protected]>
>
> I've added a stable 4.5+, since ARC didn't use dma ops until 4.5-rc1.
Again I was hitting a strange problem when sending patch via "git send-email" to
address "<[email protected]> # 3.3.x". Mail server complains on wrong email.
Thus I settled to just mention [email protected].
Anyways, thanks for doing that!
-Alexey
On 11/3/2016 3:06 PM, Alexey Brodkin wrote:
> We used to use generic implementation of dma_map_ops.mmap which is
> dma_common_mmap() but that only worked for simpler cached mappings when
> vaddr = paddr.
>
> If a driver requests uncached DMA buffer kernel maps it to virtual
> address so that MMU gets involved and page uncached status takes into
> account. In that case usage of dma_common_mmap() lead to mapping of
> vaddr to vaddr for user-space which is obviously wrong. For more detals
> please refer to verbose explanation here [1].
>
> So here we implement our own version of mmap() which always deals
> with dma_addr and maps underlying memory to user-space properly
> (note that DMA buffer mapped to user-space is always uncached
> because there's no way to properly manage cache from user-space).
>
> [1] https://lkml.org/lkml/2016/10/26/973
>
> Signed-off-by: Alexey Brodkin <[email protected]>
> Reviewed-by: Catalin Marinas <[email protected]>
> Cc: Marek Szyprowski <[email protected]>
> Cc: Vineet Gupta <[email protected]>
> Cc: <[email protected]>
> ---
Tested-by: Ramiro Oliveira <[email protected]>