Re: [Question] devm_kmalloc() for DMA ?

[Lars-Peter Clausen]

On 03/08/2017 10:19 PM, Russell King - ARM Linux wrote:
> On Wed, Mar 08, 2017 at 09:44:17PM +0100, Lars-Peter Clausen wrote:
>> On 03/08/2017 08:59 PM, Russell King - ARM Linux wrote:
>>> On Wed, Mar 08, 2017 at 08:48:31PM +0100, Lars-Peter Clausen wrote:
>>>> When the DMA memory is mapped for reading from the device the associated
>>>> cachelines are invalidated without writeback. There is no guarantee that
>>>> the changes made to the devres_node have made it to main memory yet, or
>>>> is there?
>>>
>>> That is incorrect.
>>>
>>> Overlapping cache lines are always written back on transitions from CPU
>>> to device ownership of the buffer (eg, dma_map_*().)
>>
>> On ARM. But my understanding is that this is not a universal requirement
>> according to DMA-API.txt. It says that mappings must be cache line aligned
>> and otherwise behavior is undefined.
> 
> There is no use of the term "undefined" in the document you refer to.
> 
> There is the recommendation that regions are cache line aligned, but
> there is quite a bit of history in the kernel where DMA has been to
> regions that are not cache line aligned, and where the DMA region
> overlaps with data that has recent accesses made to it.

I says: "Warnings:  Memory coherency operates at a granularity called the
cache line width.  In order for memory mapped by this API to operate
correctly, the mapped region must begin exactly on a cache line
boundary and end exactly on one." That doesn't sound like a recommendation
to me. "should" usually implies a recommendation while "must" indicates a
hard requirement.

I believe e.g. MIPS will align the address by masking the lower bits off,
without any flushes. Wouldn't be surprised if other architectures do the same.

> 
> The situation is improving (in that DMA buffers are being allocated
> separately, rather than being part of some other structure) but that
> doesn't mean that it's safe to assume that overlapping cache lines can
> be invalidated.
> 
> In any case, DMA with devm allocated buffers really is not a good idea.

I very much agree with that part.