On Mon, Mar 16, 2020 at 07:13:24PM +0100, Christoph Hellwig wrote:

On Mon, Mar 16, 2020 at 03:07:13PM -0300, Jason Gunthorpe wrote:

...

I chose this to be simple without having to goto unwind it.

So, instead like this:

As ѕaid, and per the previous discussion: I think just removing the pgmap lookup is the right thing to do here. Something like this patch:

OK. I think I get it now.

We don't even signal that the pfn is a pgmap to the caller, so the caller must assume the pfn is CPU memory and can be dma mapped. At that point it doesn't matter what kind of pgmap it is.

Races here are resolved by notifiers as we can't destroy the pgmap without triggering invalidation of the pte

So removing is the right thing to do, and the fixing for the device_private case is closer to the hunk I just sent.

Jason