Re: [PATCH rfc 3/4] mm: filemap: move __lruvec_stat_mod_folio() out of filemap_set_pte_range()

[David Hildenbrand <david@xxxxxxxxxx>]

On 08.05.24 13:15, Kefeng Wang wrote:
> On 2024/5/8 17:33, David Hildenbrand wrote:
> > On 07.05.24 15:12, Kefeng Wang wrote:
> > > On 2024/5/7 19:11, David Hildenbrand wrote:
> > > > On 29.04.24 09:24, Kefeng Wang wrote:
> > > > > Adding __folio_add_file_rmap_ptes() which don't update lruvec stat, it
> > > > > is used in filemap_set_pte_range(), with it, lruvec stat updating is
> > > > > moved into the caller, no functional changes.
> > > > >
> > > > > Signed-off-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
> > > > > ---
> > > > >     include/linux/rmap.h |  2 ++
> > > > >     mm/filemap.c         | 27 ++++++++++++++++++---------
> > > > >     mm/rmap.c            | 16 ++++++++++++++++
> > > > >     3 files changed, 36 insertions(+), 9 deletions(-)
> > > > >
> > > > > diff --git a/include/linux/rmap.h b/include/linux/rmap.h
> > > > > index 7229b9baf20d..43014ddd06f9 100644
> > > > > --- a/include/linux/rmap.h
> > > > > +++ b/include/linux/rmap.h
> > > > > @@ -242,6 +242,8 @@ void folio_add_anon_rmap_pmd(struct folio *,
> > > > > struct page *,
> > > > >             struct vm_area_struct *, unsigned long address, rmap_t
> > > > > flags);
> > > > >     void folio_add_new_anon_rmap(struct folio *, struct
> > > > > vm_area_struct *,
> > > > >             unsigned long address);
> > > > > +int __folio_add_file_rmap_ptes(struct folio *, struct page *, int
> > > > > nr_pages,
> > > > > +        struct vm_area_struct *);
> > > > >     void folio_add_file_rmap_ptes(struct folio *, struct page *, int
> > > > > nr_pages,
> > > > >             struct vm_area_struct *);
> > > > >     #define folio_add_file_rmap_pte(folio, page, vma) \
> > > > > diff --git a/mm/filemap.c b/mm/filemap.c
> > > > > index 7019692daddd..3966b6616d02 100644
> > > > > --- a/mm/filemap.c
> > > > > +++ b/mm/filemap.c
> > > > > @@ -3501,14 +3501,15 @@ static struct folio
> > > > > *next_uptodate_folio(struct xa_state *xas,
> > > > >     static void filemap_set_pte_range(struct vm_fault *vmf, struct folio
> > > > > *folio,
> > > > >                 struct page *page, unsigned int nr, unsigned long addr,
> > > > > -            unsigned long *rss)
> > > > > +            unsigned long *rss, int *nr_mapped)
> > > > >     {
> > > > >         struct vm_area_struct *vma = vmf->vma;
> > > > >         pte_t entry;
> > > > >         entry = prepare_range_pte_entry(vmf, false, folio, page, nr,
> > > > > addr);
> > > > > -    folio_add_file_rmap_ptes(folio, page, nr, vma);
> > > > > +    *nr_mapped += __folio_add_file_rmap_ptes(folio, page, nr, vma);
> > > > > +
> > > > >         set_ptes(vma->vm_mm, addr, vmf->pte, entry, nr);
> > > > >         /* no need to invalidate: a not-present page won't be cached */
> > > > > @@ -3525,7 +3526,8 @@ static void filemap_set_pte_range(struct
> > > > > vm_fault *vmf, struct folio *folio,
> > > > >     static vm_fault_t filemap_map_folio_range(struct vm_fault *vmf,
> > > > >                 struct folio *folio, unsigned long start,
> > > > >                 unsigned long addr, unsigned int nr_pages,
> > > > > -            unsigned long *rss, unsigned int *mmap_miss)
> > > > > +            unsigned long *rss, int *nr_mapped,
> > > > > +            unsigned int *mmap_miss)
> > > > >     {
> > > > >         vm_fault_t ret = 0;
> > > > >         struct page *page = folio_page(folio, start);
> > > > > @@ -3558,7 +3560,8 @@ static vm_fault_t filemap_map_folio_range(struct
> > > > > vm_fault *vmf,
> > > > >             continue;
> > > > >     skip:
> > > > >             if (count) {
> > > > > -            filemap_set_pte_range(vmf, folio, page, count, addr, rss);
> > > > > +            filemap_set_pte_range(vmf, folio, page, count, addr,
> > > > > +                          rss, nr_mapped);
> > > > >                 if (in_range(vmf->address, addr, count * PAGE_SIZE))
> > > > >                     ret = VM_FAULT_NOPAGE;
> > > > >             }
> > > > > @@ -3571,7 +3574,8 @@ static vm_fault_t filemap_map_folio_range(struct
> > > > > vm_fault *vmf,
> > > > >         } while (--nr_pages > 0);
> > > > >         if (count) {
> > > > > -        filemap_set_pte_range(vmf, folio, page, count, addr, rss);
> > > > > +        filemap_set_pte_range(vmf, folio, page, count, addr, rss,
> > > > > +                      nr_mapped);
> > > > >             if (in_range(vmf->address, addr, count * PAGE_SIZE))
> > > > >                 ret = VM_FAULT_NOPAGE;
> > > > >         }
> > > > > @@ -3583,7 +3587,7 @@ static vm_fault_t filemap_map_folio_range(struct
> > > > > vm_fault *vmf,
> > > > >     static vm_fault_t filemap_map_order0_folio(struct vm_fault *vmf,
> > > > >             struct folio *folio, unsigned long addr,
> > > > > -        unsigned long *rss, unsigned int *mmap_miss)
> > > > > +        unsigned long *rss, int *nr_mapped, unsigned int *mmap_miss)
> > > > >     {
> > > > >         vm_fault_t ret = 0;
> > > > >         struct page *page = &folio->page;
> > > > > @@ -3606,7 +3610,7 @@ static vm_fault_t
> > > > > filemap_map_order0_folio(struct vm_fault *vmf,
> > > > >         if (vmf->address == addr)
> > > > >             ret = VM_FAULT_NOPAGE;
> > > > > -    filemap_set_pte_range(vmf, folio, page, 1, addr, rss);
> > > > > +    filemap_set_pte_range(vmf, folio, page, 1, addr, rss, nr_mapped);
> > > > >         return ret;
> > > > >     }
> > > > > @@ -3646,6 +3650,7 @@ vm_fault_t filemap_map_pages(struct vm_fault
> > > > > *vmf,
> > > > >         folio_type = mm_counter_file(folio);
> > > > >         do {
> > > > >             unsigned long end;
> > > > > +        int nr_mapped = 0;
> > > > >             addr += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
> > > > >             vmf->pte += xas.xa_index - last_pgoff;
> > > > > @@ -3655,11 +3660,15 @@ vm_fault_t filemap_map_pages(struct vm_fault
> > > > > *vmf,
> > > > >             if (!folio_test_large(folio))
> > > > >                 ret |= filemap_map_order0_folio(vmf,
> > > > > -                    folio, addr, &rss, &mmap_miss);
> > > > > +                    folio, addr, &rss, &nr_mapped,
> > > > > +                    &mmap_miss);
> > > > >             else
> > > > >                 ret |= filemap_map_folio_range(vmf, folio,
> > > > >                         xas.xa_index - folio->index, addr,
> > > > > -                    nr_pages, &rss, &mmap_miss);
> > > > > +                    nr_pages, &rss, &nr_mapped,
> > > > > +                    &mmap_miss);
> > > > > +
> > > > > +        __lruvec_stat_mod_folio(folio, NR_FILE_MAPPED, nr_mapped);
> > > > >             folio_unlock(folio);
> > > > >             folio_put(folio);
> > > > > diff --git a/mm/rmap.c b/mm/rmap.c
> > > > > index 2608c40dffad..55face4024f2 100644
> > > > > --- a/mm/rmap.c
> > > > > +++ b/mm/rmap.c
> > > > > @@ -1452,6 +1452,22 @@ static __always_inline void
> > > > > __folio_add_file_rmap(struct folio *folio,
> > > > >             mlock_vma_folio(folio, vma);
> > > > >     }
> > > > > +int __folio_add_file_rmap_ptes(struct folio *folio, struct page *page,
> > > > > +        int nr_pages, struct vm_area_struct *vma)
> > > > > +{
> > > > > +    int nr, nr_pmdmapped = 0;
> > > > > +
> > > > > +    VM_WARN_ON_FOLIO(folio_test_anon(folio), folio);
> > > > > +
> > > > > +    nr = __folio_add_rmap(folio, page, nr_pages, RMAP_LEVEL_PTE,
> > > > > +                  &nr_pmdmapped);
> > > > > +
> > > > > +    /* See comments in folio_add_anon_rmap_*() */
> > > > > +    if (!folio_test_large(folio))
> > > > > +        mlock_vma_folio(folio, vma);
> > > > > +
> > > > > +    return nr;
> > > > > +}
> > > >
> > > > I'm not really a fan :/ It does make the code more complicated, and it
> > > > will be harder to extend if we decide to ever account differently (e.g.,
> > > > NR_SHMEM_MAPPED, additional tracking for mTHP etc).
> > >
> > > If more different accounts, this may lead to bad scalability.
> >
> > We already do it for PMD mappings.
> >
> > > > With large folios we'll be naturally batching already here, and I do
> > >
> > > Yes, it is batched with large folios，but our fs is ext4/tmpfs, there
> > > are not support large folio or still upstreaming.
> >
> > Okay, so that will be sorted out sooner or later.
> >
> > > > wonder, if this is really worth for performance, or if we could find
> > > > another way of batching (let the caller activate batching and drain
> > > > afterwards) without exposing these details to the caller.
> > >
> > > It does reduce latency when batch lruvec stat updating without large
> > > folio, but I can't find better way, or let's wait for the large folio
> > > support on ext4/tmpfs, I also Cced memcg maintainers in patch4 to see if
> > > there are any other ideas.
> >
> > I'm not convinced this benefit here is worth making the code more
> > complicated.
> >
> > Maybe we can find another way to optimize this batching in rmap code
> > without having to leak these details to the callers.
> >
> > For example, we could pass an optional batching structure to all rmap
> > add/rel functions that would collect these stat updates. Then we could
> > have one function to flush it and update the counters combined.
> >
> > Such batching could be beneficial also for page unmapping/zapping where
> > we might unmap various different folios in one go.
>
> It sounds better and clearer, I will try it and see the results, thanks
> for your advise!

To batch across multiple folios, it might be sufficient to remember in 
the batching structure for now:

* folio_memcg(folio)
* folio_pgdat(folio)
* NR_ANON_MAPPED diff
* NR_FILE_MAPPED diff

If the memcg of pgdat would change, we simply flush. Otherwise we batch 
and the caller flushes.

Likely we mapping/unmapping multiple folios they belong to the same 
pgdat+memcg.

The only tricky bit is the rcu_read_lock() around folio_memcg(); that 
might require some thought.

Hm ....

--
Cheers,

David / dhildenb