Axel,
On Tue, Mar 30, 2021 at 04:30:13PM -0700, Axel Rasmussen wrote:
Yes, a refactor like that is promising. It's hard to say for certain without actually looking at the result - I'll spend some time tomorrow on a few options, and send along the cleanest version I come up with.
Before you move onto a new version... See this commit:
5b51072e97d5 ("userfaultfd: shmem: allocate anonymous memory for MAP_PRIVATE shmem", 2018-11-30)
I found it when I was thinking why not move the whole continue logic directly into mfill_atomic_pte(), if we can have the pte installation helper, because that's all we need.
So previously I got the semantics a bit mixed up: for private shmem mappings, UFFDIO_COPY won't fill in page cache at all, but it's all private. We keep the page cache empty even after UFFDIO_COPY for a private mapping.
UFFDIO_CONTINUE is slightly different, since we _know_ the page cache is there.. So I'm thinking maybe you need to handle the continue request in mfill_atomic_pte() before the VM_SHARED check so as to cover both cases.