On Mon, 2018-08-06 at 23:27 -0700, Christoph Hellwig wrote:

On Tue, Aug 07, 2018 at 08:13:56AM +1000, Benjamin Herrenschmidt wrote:

quoted

It would be indeed ideal if all we had to do was setup some kind of
bus_dma_mask on all PCI devices and have virtio automagically insert
swiotlb when necessary.

For 4.20 I plan to remove the swiotlb ops and instead do the bounce
buffering in the common code, including a direct call to the direct
ops to avoid retpoline overhead.  For that you still need a flag
in virtio that instead of blindly working physical addresses it needs
to be treated like a real device in terms of DMA.

But you will still call the swiotlb infrastructure, right ? IE, I sitll
need to control where/how the swiotlb "pool" is allocated.

And for powerpc to make use of that I need to get the dma series I
posted last week reviewed and included, otherwise powerpc will have
to be excepted (like arm, where rmk didn't like the way the code
was factored, everything else has already been taken care of).

https://lists.linuxfoundation.org/pipermail/iommu/2018-July/028989.html

Yes, I saw your series. I'm just back from a week of travel, I plan to
start reviewing it this week if Michael doesn't beat me to it.

Cheers,
Ben.