On Thursday, 11 February 2021 6:55:10 PM AEDT Christoph Hellwig wrote:

On Wed, Feb 10, 2021 at 01:59:13PM -0400, Jason Gunthorpe wrote:

...

Really what you want to do here is leave the CPU page in the VMA and the page tables where it started and deny CPU access to the page. Then all the proper machinery will continue to work.

IMHO "migration" is the wrong idea if the data isn't actually moving.

Agreed.

I chose "migration" because device private pages seemed like a good way of reusing existing code to do what was required (a callback on CPU access).

However I have been reworking this to use mmu notifiers as the callback and it seems to simplify some things nicely so think I also agree. It removes the requirement for the pin as well which is nice, I'll post it as a v2 soon.

- Alistair