Re: [Xen-devel] [PATCH 1/1] XEN: Use correct masking in xen_swiotlb_alloc_coherent.

From: Stefano Panella
Date: Tue Sep 04 2012 - 10:07:37 EST


On 08/31/2012 05:40 PM, Konrad Rzeszutek Wilk wrote:
On Fri, Aug 31, 2012 at 01:47:05PM +0100, David Vrabel wrote:
On 31/08/12 10:57, Stefano Panella wrote:
When running 32-bit pvops-dom0 and a driver tries to allocate a coherent
DMA-memory the xen swiotlb-implementation returned memory beyond 4GB.

This caused for example not working sound on a system with 4 GB and a 64-bit
compatible sound-card with sets the DMA-mask to 64bit.

On bare-metal and the forward-ported xen-dom0 patches from OpenSuse a coherent
DMA-memory is always allocated inside the 32-bit address-range by calling
dma_alloc_coherent_mask.
We should have the same behaviour under Xen as bare metal so:

Acked-By: David Vrabel <david.vrabel@xxxxxxxxxx>

This does limit the DMA mask to 32-bits by passing it through an
unsigned long, which seems a bit sneaky...
so is the issue that we are not casting it from 'u64' to 'u32'
(unsigned long) on 32-bit?

Yes. I do not completely understand why but I think on 32-bit kernel we need to cast dma_mask to u32. This is done automatically using dma_alloc_coherent_mask()


Presumably the sound card is capable of handling 64 bit physical
addresses (or it would break under 64-bit kernels) so it's not clear why
this sound driver requires this restriction.

Is there a bug in the sound driver or sound subsystem where it's
truncating a dma_addr_t by assigning it to an unsigned long or similar?

--- a/drivers/xen/swiotlb-xen.c
+++ b/drivers/xen/swiotlb-xen.c
@@ -232,7 +232,7 @@ xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
return ret;
if (hwdev && hwdev->coherent_dma_mask)
- dma_mask = hwdev->coherent_dma_mask;
+ dma_mask = dma_alloc_coherent_mask(hwdev, flags);
Suggest

if (hwdev)
dma_mask = dma_alloc_coherent_mask(hwdev, flags)

I can change the patch like that if you like.

Isn't that code just doing this:
atic inline unsigned long dma_alloc_coherent_mask(struct device *dev,
gfp_t gfp)
{
unsigned long dma_mask = 0;

dma_mask = dev->coherent_dma_mask;
if (!dma_mask)
dma_mask = (gfp & GFP_DMA) ? DMA_BIT_MASK(24) :
DMA_BIT_MASK(32);

return dma_mask;
}

and in our code, the dma_mask by default is DMA_BIT_MASK(32):

u64 dma_mask = DMA_BIT_MASK(32);

So what I am missing?

I am not sure what you mean with "what am I missing?"

Current code looks like:

void *
xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
dma_addr_t *dma_handle, gfp_t flags,
struct dma_attrs *attrs)
{
void *ret;
int order = get_order(size);
u64 dma_mask = DMA_BIT_MASK(32);
unsigned long vstart;
phys_addr_t phys;
dma_addr_t dev_addr;

/*
* Ignore region specifiers - the kernel's ideas of
* pseudo-phys memory layout has nothing to do with the
* machine physical layout. We can't allocate highmem
* because we can't return a pointer to it.
*/
flags &= ~(__GFP_DMA | __GFP_HIGHMEM);

if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
return ret;

vstart = __get_free_pages(flags, order);
ret = (void *)vstart;

if (!ret)
return ret;

if (hwdev && hwdev->coherent_dma_mask)
dma_mask = hwdev->coherent_dma_mask;


So if hwdev->coherent_dma_mask is set to 0xffffffffffffffff our dma_mask will
be u64 set to 0xffffffffffffffff even if we set it to DMA_BIT_MASK(32) previously.

I hope I am not getting this wrong and let me know if I should send an updated version
of the patch including David V. change.

Regards,

Stefano

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/