Re: [PATCH 1/1] mm/madvise: enhance lazyfreeing with mTHP in madvise_free

[David Hildenbrand]

On 26.02.24 13:57, Ryan Roberts wrote:
> On 26/02/2024 08:35, Lance Yang wrote:
> > Hey Fengwei,
> >
> > Thanks for taking time to review!
> >
> > > On Mon, Feb 26, 2024 at 10:38 AM Yin Fengwei <fengwei.yin@xxxxxxxxx> wrote:
> > > > On Sun, Feb 25, 2024 at 8:32 PM Lance Yang <ioworker0@xxxxxxxxx> wrote:
> > [...]
> > > > --- a/mm/madvise.c
> > > > +++ b/mm/madvise.c
> > > > @@ -676,11 +676,43 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
> > > >                 */
> > > >                if (folio_test_large(folio)) {
> > > >                        int err;
> > > > +                     unsigned long next_addr, align;
> > > >
> > > > -                     if (folio_estimated_sharers(folio) != 1)
> > > > -                             break;
> > > > -                     if (!folio_trylock(folio))
> > > > -                             break;
> > > > +                     if (folio_estimated_sharers(folio) != 1 ||
> > > > +                         !folio_trylock(folio))
> > > > +                             goto skip_large_folio;
> > > > +
> > > > +                     align = folio_nr_pages(folio) * PAGE_SIZE;
> > > > +                     next_addr = ALIGN_DOWN(addr + align, align);
> > > There is a possible corner case:
> > > If there is a cow folio associated with this folio and the cow folio
> > > has smaller size than this folio for whatever reason, this change can't
> > > handle it correctly.
> >
> > Thanks for pointing that out; it's very helpful to me!
> > I made some changes. Could you please check if this corner case is now resolved?
> >
> > As a diff against this patch.
> >
> > diff --git a/mm/madvise.c b/mm/madvise.c
> > index bcbf56595a2e..c7aacc9f9536 100644
> > --- a/mm/madvise.c
> > +++ b/mm/madvise.c
> > @@ -686,10 +686,12 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
> >   			next_addr = ALIGN_DOWN(addr + align, align);
> >   
> >   			/*
> > -			 * If we mark only the subpages as lazyfree,
> > -			 * split the large folio.
> > +			 * If we mark only the subpages as lazyfree, or
> > +			 * if there is a cow folio associated with this folio,
> > +			 * then split the large folio.
> >   			 */
> > -			if (next_addr > end || next_addr - addr != align)
> > +			if (next_addr > end || next_addr - addr != align ||
> > +			    folio_total_mapcount(folio) != folio_nr_pages(folio))
>
> I still don't think this is correct. I think you were previously assuming that
> if you see a page from a large folio then the whole large folio should be
> contiguously mapped? This new check doesn't validate that assumption reliably;
> you need to iterate through every pte to generate a batch, like David does in
> folio_pte_batch() for this to be safe.
>
> An example of when this check is insufficient; let's say you have a 4 page anon
> folio mapped contiguously in a process (total_mapcount=4). The process is forked
> (total_mapcount=8). Then each process munmaps the second 2 pages
> (total_mapcount=4). In place of the munmapped 2 pages, 2 new pages are mapped.
> Then call madvise. It's probably even easier to trigger for file-backed memory
> (I think this code path is used for both file and anon?)

What would work here is using folio_pte_batch() to get how many PTEs are 
mapped *here*, then comparing the the batch size to folio_nr_pages(). If 
both match, we are mapping all subpages.

--
Cheers,

David / dhildenb