From: Chen-Yu Tsai <[email protected]>
max_pfn, as set in arch/arm/mm/init.c:
static void __init find_limits(unsigned long *min,
unsigned long *max_low,
unsigned long *max_high)
{
*max_low = PFN_DOWN(memblock_get_current_limit());
*min = PFN_UP(memblock_start_of_DRAM());
*max_high = PFN_DOWN(memblock_end_of_DRAM());
}
with memblock_end_of_DRAM() pointing to the next byte after DRAM. As
such, max_pfn points to the PFN after the end of DRAM.
Thus when using max_pfn to check DMA masks, we should subtract one
when checking DMA ranges against it.
Commit 8bf1268f48ad ("ARM: dma-api: fix off-by-one error in
__dma_supported()") fixed the same issue, but missed this spot.
This issue was found while working on the sun4i-csi v4l2 driver on the
Allwinner R40 SoC. On Allwinner SoCs, DRAM is offset at 0x40000000,
and we are starting to use of_dma_configure() with the "dma-ranges"
property in the device tree to have the DMA API handle the offset.
In this particular instance, dma-ranges was set to the same range as
the actual available (2 GiB) DRAM. The following error appeared when
the driver attempted to allocate a buffer:
sun4i-csi 1c09000.csi: Coherent DMA mask 0x7fffffff (pfn 0x40000-0xc0000)
covers a smaller range of system memory than the DMA zone pfn 0x0-0xc0001
sun4i-csi 1c09000.csi: dma_alloc_coherent of size 307200 failed
Fixing the off-by-one error makes things work.
Fixes: 11a5aa32562e ("ARM: dma-mapping: check DMA mask against available memory")
Fixes: 9f28cde0bc64 ("ARM: another fix for the DMA mapping checks")
Fixes: ab746573c405 ("ARM: dma-mapping: allow larger DMA mask than supported")
Cc: <[email protected]>
Signed-off-by: Chen-Yu Tsai <[email protected]>
---
arch/arm/mm/dma-mapping.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
index e822af0d9219..f4daafdbac56 100644
--- a/arch/arm/mm/dma-mapping.c
+++ b/arch/arm/mm/dma-mapping.c
@@ -227,12 +227,12 @@ static int __dma_supported(struct device *dev, u64 mask, bool warn)
* Translate the device's DMA mask to a PFN limit. This
* PFN number includes the page which we can DMA to.
*/
- if (dma_to_pfn(dev, mask) < max_dma_pfn) {
+ if (dma_to_pfn(dev, mask) < max_dma_pfn - 1) {
if (warn)
dev_warn(dev, "Coherent DMA mask %#llx (pfn %#lx-%#lx) covers a smaller range of system memory than the DMA zone pfn 0x0-%#lx\n",
mask,
dma_to_pfn(dev, 0), dma_to_pfn(dev, mask) + 1,
- max_dma_pfn + 1);
+ max_dma_pfn);
return 0;
}
--
2.24.0
On 11/12/2019 10:41 am, Chen-Yu Tsai wrote:
> From: Chen-Yu Tsai <[email protected]>
>
> max_pfn, as set in arch/arm/mm/init.c:
>
> static void __init find_limits(unsigned long *min,
> unsigned long *max_low,
> unsigned long *max_high)
> {
> *max_low = PFN_DOWN(memblock_get_current_limit());
> *min = PFN_UP(memblock_start_of_DRAM());
> *max_high = PFN_DOWN(memblock_end_of_DRAM());
> }
>
> with memblock_end_of_DRAM() pointing to the next byte after DRAM. As
> such, max_pfn points to the PFN after the end of DRAM.
>
> Thus when using max_pfn to check DMA masks, we should subtract one
> when checking DMA ranges against it.
>
> Commit 8bf1268f48ad ("ARM: dma-api: fix off-by-one error in
> __dma_supported()") fixed the same issue, but missed this spot.
>
> This issue was found while working on the sun4i-csi v4l2 driver on the
> Allwinner R40 SoC. On Allwinner SoCs, DRAM is offset at 0x40000000,
> and we are starting to use of_dma_configure() with the "dma-ranges"
> property in the device tree to have the DMA API handle the offset.
>
> In this particular instance, dma-ranges was set to the same range as
> the actual available (2 GiB) DRAM. The following error appeared when
> the driver attempted to allocate a buffer:
>
> sun4i-csi 1c09000.csi: Coherent DMA mask 0x7fffffff (pfn 0x40000-0xc0000)
> covers a smaller range of system memory than the DMA zone pfn 0x0-0xc0001
> sun4i-csi 1c09000.csi: dma_alloc_coherent of size 307200 failed
>
> Fixing the off-by-one error makes things work.
>
> Fixes: 11a5aa32562e ("ARM: dma-mapping: check DMA mask against available memory")
> Fixes: 9f28cde0bc64 ("ARM: another fix for the DMA mapping checks")
> Fixes: ab746573c405 ("ARM: dma-mapping: allow larger DMA mask than supported")
> Cc: <[email protected]>
> Signed-off-by: Chen-Yu Tsai <[email protected]>
> ---
> arch/arm/mm/dma-mapping.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
> index e822af0d9219..f4daafdbac56 100644
> --- a/arch/arm/mm/dma-mapping.c
> +++ b/arch/arm/mm/dma-mapping.c
> @@ -227,12 +227,12 @@ static int __dma_supported(struct device *dev, u64 mask, bool warn)
> * Translate the device's DMA mask to a PFN limit. This
> * PFN number includes the page which we can DMA to.
> */
> - if (dma_to_pfn(dev, mask) < max_dma_pfn) {
> + if (dma_to_pfn(dev, mask) < max_dma_pfn - 1) {
I think this correction actually wants to happen a couple of lines up in
the definition:
unsigned long max_dma_pfn = min(max_pfn, arm_dma_pfn_limit);
max_pfn is indeed an exclusive limit, but AFAICS arm_dma_pfn_limit is
inclusive, so none of these "+1"s and "-1"s can be entirely right for
both cases.
Robin.
> if (warn)
> dev_warn(dev, "Coherent DMA mask %#llx (pfn %#lx-%#lx) covers a smaller range of system memory than the DMA zone pfn 0x0-%#lx\n",
> mask,
> dma_to_pfn(dev, 0), dma_to_pfn(dev, mask) + 1,
> - max_dma_pfn + 1);
> + max_dma_pfn);
> return 0;
> }
>
>
On Wed, Dec 11, 2019 at 9:41 PM Robin Murphy <[email protected]> wrote:
>
> On 11/12/2019 10:41 am, Chen-Yu Tsai wrote:
> > From: Chen-Yu Tsai <[email protected]>
> >
> > max_pfn, as set in arch/arm/mm/init.c:
> >
> > static void __init find_limits(unsigned long *min,
> > unsigned long *max_low,
> > unsigned long *max_high)
> > {
> > *max_low = PFN_DOWN(memblock_get_current_limit());
> > *min = PFN_UP(memblock_start_of_DRAM());
> > *max_high = PFN_DOWN(memblock_end_of_DRAM());
> > }
> >
> > with memblock_end_of_DRAM() pointing to the next byte after DRAM. As
> > such, max_pfn points to the PFN after the end of DRAM.
> >
> > Thus when using max_pfn to check DMA masks, we should subtract one
> > when checking DMA ranges against it.
> >
> > Commit 8bf1268f48ad ("ARM: dma-api: fix off-by-one error in
> > __dma_supported()") fixed the same issue, but missed this spot.
> >
> > This issue was found while working on the sun4i-csi v4l2 driver on the
> > Allwinner R40 SoC. On Allwinner SoCs, DRAM is offset at 0x40000000,
> > and we are starting to use of_dma_configure() with the "dma-ranges"
> > property in the device tree to have the DMA API handle the offset.
> >
> > In this particular instance, dma-ranges was set to the same range as
> > the actual available (2 GiB) DRAM. The following error appeared when
> > the driver attempted to allocate a buffer:
> >
> > sun4i-csi 1c09000.csi: Coherent DMA mask 0x7fffffff (pfn 0x40000-0xc0000)
> > covers a smaller range of system memory than the DMA zone pfn 0x0-0xc0001
> > sun4i-csi 1c09000.csi: dma_alloc_coherent of size 307200 failed
> >
> > Fixing the off-by-one error makes things work.
> >
> > Fixes: 11a5aa32562e ("ARM: dma-mapping: check DMA mask against available memory")
> > Fixes: 9f28cde0bc64 ("ARM: another fix for the DMA mapping checks")
> > Fixes: ab746573c405 ("ARM: dma-mapping: allow larger DMA mask than supported")
> > Cc: <[email protected]>
> > Signed-off-by: Chen-Yu Tsai <[email protected]>
> > ---
> > arch/arm/mm/dma-mapping.c | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c
> > index e822af0d9219..f4daafdbac56 100644
> > --- a/arch/arm/mm/dma-mapping.c
> > +++ b/arch/arm/mm/dma-mapping.c
> > @@ -227,12 +227,12 @@ static int __dma_supported(struct device *dev, u64 mask, bool warn)
> > * Translate the device's DMA mask to a PFN limit. This
> > * PFN number includes the page which we can DMA to.
> > */
> > - if (dma_to_pfn(dev, mask) < max_dma_pfn) {
> > + if (dma_to_pfn(dev, mask) < max_dma_pfn - 1) {
>
> I think this correction actually wants to happen a couple of lines up in
> the definition:
>
> unsigned long max_dma_pfn = min(max_pfn, arm_dma_pfn_limit);
>
> max_pfn is indeed an exclusive limit, but AFAICS arm_dma_pfn_limit is
> inclusive, so none of these "+1"s and "-1"s can be entirely right for
> both cases.
You're absolutely right. I'll fix it and send a v2 out.
Thanks
ChenYu
> Robin.
>
> > if (warn)
> > dev_warn(dev, "Coherent DMA mask %#llx (pfn %#lx-%#lx) covers a smaller range of system memory than the DMA zone pfn 0x0-%#lx\n",
> > mask,
> > dma_to_pfn(dev, 0), dma_to_pfn(dev, mask) + 1,
> > - max_dma_pfn + 1);
> > + max_dma_pfn);
> > return 0;
> > }
> >
> >