Re: [PATCH v4] mm, compaction: Skip all non-migratable pages during scan

[John Hubbard]

On 5/26/23 11:50, David Hildenbrand wrote:
> On 26.05.23 20:46, Matthew Wilcox wrote:
> > On Fri, May 26, 2023 at 06:46:15PM +0200, David Hildenbrand wrote:
> > > On 26.05.23 18:44, Matthew Wilcox wrote:
> > > > On Fri, May 26, 2023 at 09:44:34AM -0600, Khalid Aziz wrote:
> > > > > > Oh, I think I found it!  pin_user_pages_remote() is called by
> > > > > > vaddr_get_pfns().  If these are the pages you're concerned about,
> > > > > > then the efficient way to do what you want is simply to call
> > > > > > folio_maybe_dma_pinned().  Far more efficient than the current mess
> > > > > > of total_mapcount().
> > > > >
> > > > > vfio pinned pages triggered this change. Wouldn't checking refcounts against
> > > > > mapcount provide a more generalized way of detecting non-migratable pages?
> > > >
> > > > Well, you changed the comment to say that we were concerned about
> > > > long-term pins.  If we are, than folio_maybe_dma_pinned() is how to test
> > > > for long-term pins.  If we want to skip pages which are short-term pinned,
> > > > then we need to not change the comment, and keep using mapcount/refcount
> > > > differences.
> > >
> > > folio_maybe_dma_pinned() is all about FOLL_PIN, not FOLL_LONGTERM.
> >
> > But according to our documentation, FOLL_LONGTERM implies FOLL_PIN.
>
> Yes. But folio_maybe_dma_pinned() will indicate both, long-term pins and short-term pins. There really is no way to distinguish both, unfortunately.

Not yet, anyway. :)

> > Anyway, right now, the code skips any pages which are merely FOLL_GET,
> > so we'll skip fewer pages if we do only skip the FOLL_PIN ones,
> > regardless if we'd prefer to only skip the FOLL_LONGTERM ones.
> >
> > > folio_maybe_dma_pinned() would skip migrating any page that has more than
> > > 1024 references. (shared libraries?)
> >
> > True, but maybe we should be skipping any page with that many mappings,
> > given how disruptive it is to the rest of the system to unmap a page
> > from >1024 processes.
>
> So any user with 1024 processes can fragment physical memory? :/
>
> Sorry, I'd like to minimize the usage of folio_maybe_dma_pinned().

I was actually thinking that we should minimize any more cases of
fragile mapcount and refcount comparison, which then leads to
Matthew's approach here!


thanks,
--
John Hubbard
NVIDIA