Re: Folio mapcount

[Peter Xu <peterx@redhat.com>]

On Tue, Feb 07, 2023 at 11:27:17PM +0000, Matthew Wilcox wrote:
> I've been thinking about this one, and I wonder if we can do it
> without taking any pgtable locks.  The locking environment we're in
> is the page fault handler, so we have the mmap_lock for read (for now
> anyway ...).  We also hold the folio lock, so _if_ the folio is mapped,
> those entries can't disappear under us.

Could MADV_DONTNEED do that from another pgtable that we don't hold the
pgtable lock?

> They also can't appear under us.

This seems right.

> We hold the PTL on one PMD, but not necessarily on any other PMD we
> examine.
> 
> I appreciate that PTEs can _change_ under us if we do not hold the PTL,
> but by virtue of holding the folio lock, they can't change from or to
> our PFNs.  I also think the PMD table cannot disappear under us
> since we're holding the mmap_lock for read, and anyone removing page
> tables has to take the mmap_lock for write.

Seems right to me too.

-- 
Peter Xu