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

[David Vrabel <david.vrabel@citrix.com>]

On 04/09/12 16:12, Stefano Panella wrote:
> On 09/04/2012 03:55 PM, David Vrabel wrote:
>> On 04/09/12 15:37, Konrad Rzeszutek Wilk wrote:
>>> On Tue, Sep 04, 2012 at 03:07:42PM +0100, Stefano Panella wrote:
>>>> 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.
>>> That is what I was missing. Let me include that in the git commit and
>>> also
>>> put this patch on the stable tree.
>> Note that this appears to be a work around for a bug in the sound system
>> or Intel HDA device driver which is incorrectly truncating a dma_addr_t
>> to a u32.  So by ensuring a DMA_BIT_MASK(32) when the dma_addr_t is
>> truncated it still works.
>>
>> David
> Sorry David, I am not completely sure I understand your argument in
> favour of a bug in the
> sound system or Intel HDA device driver. Could you please elaborate more
> in detail about this?

The HDA driver claims to support 64-bit DMA addresses, so it should be
perfectly fine for the Xen DMA mapping/allocation functions to return
DMA addresses > 4 GiB.  For most devices this seems to work just fine.

I think that somewhere between the buffer being mapped or allocated and
the address being programmed into the hardware, the 64-bit dma_addr_t
value is being truncated to 32 bits giving an invalid DMA address.

The patch would avoid this by setting the DMA mask to 32-bit and only
returning DMA address < 4 GiB which can quite happily be stuffed into a
u32 without losing bits.

I think it would be useful (without the patch applied) to print the DMA
addresses returned by the allocate/mapping functions and the address
programmed into the hardware.  It will be easily to spot if it got
truncated or not.

David