On Mon, Mar 15, 2021 at 6:54 AM Jie Deng [off-list ref] wrote:

On 2021/3/15 11:13, Jason Wang wrote:

quoted

On 2021/3/15 9:14 上午, Jie Deng wrote:

quoted

On 2021/3/12 16:58, Arnd Bergmann wrote:

Then do you think it is necessary to mark the virtio bufs with
____cacheline_aligned ?

I think so, yes.

I haven't seen any virtio interface being marked yet. If this is a
problem, I believe it should  be common for all virtio devices, right ?

Yes, but it's not a problem if the buffers are allocated separately
because kmalloc provinces a cachelinen aligned buffer on architectures
that need it.

It's only a problem here because there is a single allocation for three
objects that have different ownership states during the DMA (device
owned to-device, cpu-owned, device owned to-cpu).

       Arnd