Re: [patch 5/5]thp: split huge page if head page is isolated

[Shaohua Li <shaohua.li@intel.com>]

On Mon, 2011-10-31 at 16:23 +0800, Minchan Kim wrote:
> On Mon, Oct 31, 2011 at 09:21:28AM +0800, Shaohua Li wrote:
> > On Sat, 2011-10-29 at 08:06 +0800, Minchan Kim wrote:
> > > On Tue, Oct 25, 2011 at 10:59:40AM +0800, Shaohua Li wrote:
> > > > With current logic, if page reclaim finds a huge page, it will just reclaim
> > > > the head page and leave tail pages reclaimed later. Let's take an example,
> > > > lru list has page A and B, page A is huge page:
> > > > 1. page A is isolated
> > > > 2. page B is isolated
> > > > 3. shrink_page_list() adds page A to swap page cache. so page A is split.
> > > > page A+1, page A+2, ... are added to lru list.
> > > > 4. shrink_page_list() adds page B to swap page cache.
> > > > 5. page A and B is written out and reclaimed.
> > > > 6. page A+1, A+2 ... is isolated and reclaimed later.
> > > > So the reclaim order is A, B, ...(maybe other pages), A+1, A+2 ...
> > > >
> > > > We expected the whole huge page A is reclaimed in the meantime, so
> > > > the order is A, A+1, ... A+HPAGE_PMD_NR-1, B, ....
> > > >
> > > > With this patch, we do huge page split just after the head page is isolated
> > > > for inactive lru list, so the tail pages will be reclaimed immediately.
> > > >
> > > > In a test, a range of anonymous memory is written and will trigger swap.
> > > > Without the patch:
> > > > #cat /proc/vmstat|grep thp
> > > > thp_fault_alloc 451
> > > > thp_fault_fallback 0
> > > > thp_collapse_alloc 0
> > > > thp_collapse_alloc_failed 0
> > > > thp_split 238
> > > >
> > > > With the patch:
> > > > #cat /proc/vmstat|grep thp
> > > > thp_fault_alloc 450
> > > > thp_fault_fallback 1
> > > > thp_collapse_alloc 0
> > > > thp_collapse_alloc_failed 0
> > > > thp_split 103
> > > >
> > > > So the thp_split number is reduced a lot, though there is one extra
> > > > thp_fault_fallback.
> > > >
> > > > Signed-off-by: Shaohua Li <shaohua.li@intel.com>
> > > > ---
> > > >  include/linux/memcontrol.h |    3 +-
> > > >  mm/memcontrol.c            |   12 +++++++++--
> > > >  mm/vmscan.c                |   49 ++++++++++++++++++++++++++++++++++-----------
> > > >  3 files changed, 50 insertions(+), 14 deletions(-)
> > > >
> > > > Index: linux/mm/vmscan.c
> > > > ===================================================================
> > > > --- linux.orig/mm/vmscan.c    2011-10-25 08:36:08.000000000 +0800
> > > > +++ linux/mm/vmscan.c 2011-10-25 09:51:44.000000000 +0800
> > > > @@ -1076,7 +1076,8 @@ int __isolate_lru_page(struct page *page
> > > >   */
> > > >  static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
> > > >               struct list_head *src, struct list_head *dst,
> > > > -             unsigned long *scanned, int order, int mode, int file)
> > > > +             unsigned long *scanned, int order, int mode, int file,
> > > > +             struct page **split_page)
> > > >  {
> > > >       unsigned long nr_taken = 0;
> > > >       unsigned long nr_lumpy_taken = 0;
> > > > @@ -1100,7 +1101,12 @@ static unsigned long isolate_lru_pages(u
> > > >               case 0:
> > > >                       list_move(&page->lru, dst);
> > > >                       mem_cgroup_del_lru(page);
> > > > -                     nr_taken += hpage_nr_pages(page);
> > > > +                     if (PageTransHuge(page) && split_page) {
> > > > +                             nr_taken++;
> > > > +                             *split_page = page;
> > > > +                             goto out;
> > > > +                     } else
> > > > +                             nr_taken += hpage_nr_pages(page);
> > > >                       break;
> > > >
> > > >               case -EBUSY:
> > > > @@ -1158,11 +1164,16 @@ static unsigned long isolate_lru_pages(u
> > > >                       if (__isolate_lru_page(cursor_page, mode, file) == 0) {
> > > >                               list_move(&cursor_page->lru, dst);
> > > >                               mem_cgroup_del_lru(cursor_page);
> > > > -                             nr_taken += hpage_nr_pages(page);
> > > >                               nr_lumpy_taken++;
> > > >                               if (PageDirty(cursor_page))
> > > >                                       nr_lumpy_dirty++;
> > > >                               scan++;
> > > > +                             if (PageTransHuge(page) && split_page) {
> > > > +                                     nr_taken++;
> > > > +                                     *split_page = page;
> > > > +                                     goto out;
> > > > +                             } else
> > > > +                                     nr_taken += hpage_nr_pages(page);
> > > >                       } else {
> > > >                               /*
> > > >                                * Check if the page is freed already.
> > > > @@ -1188,6 +1199,7 @@ static unsigned long isolate_lru_pages(u
> > > >                       nr_lumpy_failed++;
> > > >       }
> > > >
> > > > +out:
> > > >       *scanned = scan;
> > > >
> > > >       trace_mm_vmscan_lru_isolate(order,
> > > > @@ -1202,7 +1214,8 @@ static unsigned long isolate_pages_globa
> > > >                                       struct list_head *dst,
> > > >                                       unsigned long *scanned, int order,
> > > >                                       int mode, struct zone *z,
> > > > -                                     int active, int file)
> > > > +                                     int active, int file,
> > > > +                                     struct page **split_page)
> > > >  {
> > > >       int lru = LRU_BASE;
> > > >       if (active)
> > > > @@ -1210,7 +1223,7 @@ static unsigned long isolate_pages_globa
> > > >       if (file)
> > > >               lru += LRU_FILE;
> > > >       return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
> > > > -                                                             mode, file);
> > > > +                                                     mode, file, split_page);
> > > >  }
> > > >
> > > >  /*
> > > > @@ -1444,10 +1457,12 @@ shrink_inactive_list(unsigned long nr_to
> > > >  {
> > > >       LIST_HEAD(page_list);
> > > >       unsigned long nr_scanned;
> > > > +     unsigned long total_scanned = 0;
> > > >       unsigned long nr_reclaimed = 0;
> > > >       unsigned long nr_taken;
> > > >       unsigned long nr_anon;
> > > >       unsigned long nr_file;
> > > > +     struct page *split_page;
> > > >
> > > >       while (unlikely(too_many_isolated(zone, file, sc))) {
> > > >               congestion_wait(BLK_RW_ASYNC, HZ/10);
> > > > @@ -1458,16 +1473,19 @@ shrink_inactive_list(unsigned long nr_to
> > > >       }
> > > >
> > > >       set_reclaim_mode(priority, sc, false);
> > > > +again:
> > > >       lru_add_drain();
> > > > +     split_page = NULL;
> > > >       spin_lock_irq(&zone->lru_lock);
> > > >
> > > >       if (scanning_global_lru(sc)) {
> > > > -             nr_taken = isolate_pages_global(nr_to_scan,
> > > > +             nr_taken = isolate_pages_global(nr_to_scan - total_scanned,
> > > >                       &page_list, &nr_scanned, sc->order,
> > > >                       sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
> > > >                                       ISOLATE_BOTH : ISOLATE_INACTIVE,
> > > > -                     zone, 0, file);
> > > > +                     zone, 0, file, &split_page);
> > > >               zone->pages_scanned += nr_scanned;
> > > > +             total_scanned += nr_scanned;
> > > >               if (current_is_kswapd())
> > > >                       __count_zone_vm_events(PGSCAN_KSWAPD, zone,
> > > >                                              nr_scanned);
> > > > @@ -1475,12 +1493,13 @@ shrink_inactive_list(unsigned long nr_to
> > > >                       __count_zone_vm_events(PGSCAN_DIRECT, zone,
> > > >                                              nr_scanned);
> > > >       } else {
> > > > -             nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
> > > > +             nr_taken = mem_cgroup_isolate_pages(nr_to_scan - total_scanned,
> > > >                       &page_list, &nr_scanned, sc->order,
> > > >                       sc->reclaim_mode & RECLAIM_MODE_LUMPYRECLAIM ?
> > > >                                       ISOLATE_BOTH : ISOLATE_INACTIVE,
> > > >                       zone, sc->mem_cgroup,
> > > > -                     0, file);
> > > > +                     0, file, &split_page);
> > > > +             total_scanned += nr_scanned;
> > > >               /*
> > > >                * mem_cgroup_isolate_pages() keeps track of
> > > >                * scanned pages on its own.
> > > > @@ -1491,11 +1510,19 @@ shrink_inactive_list(unsigned long nr_to
> > > >               spin_unlock_irq(&zone->lru_lock);
> > > >               return 0;
> > > >       }
> > > > +     if (split_page && total_scanned < nr_to_scan) {
> > > > +             spin_unlock_irq(&zone->lru_lock);
> > > > +             split_huge_page(split_page);
> > > > +             goto again;
> > > > +     }
> > > >
> > > >       update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
> > > >
> > > >       spin_unlock_irq(&zone->lru_lock);
> > > >
> > > > +     if (split_page)
> > > > +             split_huge_page(split_page);
> > > > +
> > > >       nr_reclaimed = shrink_page_list(&page_list, zone, sc);
> > > >
> > > >       /* Check if we should syncronously wait for writeback */
> > > > @@ -1589,13 +1616,13 @@ static void shrink_active_list(unsigned
> > > >               nr_taken = isolate_pages_global(nr_pages, &l_hold,
> > > >                                               &pgscanned, sc->order,
> > > >                                               ISOLATE_ACTIVE, zone,
> > > > -                                             1, file);
> > > > +                                             1, file, NULL);
> > > >               zone->pages_scanned += pgscanned;
> > > >       } else {
> > > >               nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
> > > >                                               &pgscanned, sc->order,
> > > >                                               ISOLATE_ACTIVE, zone,
> > > > -                                             sc->mem_cgroup, 1, file);
> > > > +                                             sc->mem_cgroup, 1, file, NULL);
> > > >               /*
> > > >                * mem_cgroup_isolate_pages() keeps track of
> > > >                * scanned pages on its own.
> > > > Index: linux/mm/memcontrol.c
> > > > ===================================================================
> > > > --- linux.orig/mm/memcontrol.c        2011-10-25 08:36:08.000000000 +0800
> > > > +++ linux/mm/memcontrol.c     2011-10-25 09:33:51.000000000 +0800
> > > > @@ -1187,7 +1187,8 @@ unsigned long mem_cgroup_isolate_pages(u
> > > >                                       unsigned long *scanned, int order,
> > > >                                       int mode, struct zone *z,
> > > >                                       struct mem_cgroup *mem_cont,
> > > > -                                     int active, int file)
> > > > +                                     int active, int file,
> > > > +                                     struct page **split_page)
> > > >  {
> > > >       unsigned long nr_taken = 0;
> > > >       struct page *page;
> > > > @@ -1224,7 +1225,13 @@ unsigned long mem_cgroup_isolate_pages(u
> > > >               case 0:
> > > >                       list_move(&page->lru, dst);
> > > >                       mem_cgroup_del_lru(page);
> > > > -                     nr_taken += hpage_nr_pages(page);
> > > > +                     if (PageTransHuge(page) && split_page) {
> > > > +                             nr_taken++;
> > > > +                             *split_page = page;
> > > > +                             goto out;
> > > > +                     } else
> > > > +                             nr_taken += hpage_nr_pages(page);
> > > > +
> > > >                       break;
> > > >               case -EBUSY:
> > > >                       /* we don't affect global LRU but rotate in our LRU */
> > > > @@ -1235,6 +1242,7 @@ unsigned long mem_cgroup_isolate_pages(u
> > > >               }
> > > >       }
> > > >
> > > > +out:
> > > >       *scanned = scan;
> > > >
> > > >       trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
> > > > Index: linux/include/linux/memcontrol.h
> > > > ===================================================================
> > > > --- linux.orig/include/linux/memcontrol.h     2011-10-25 08:36:08.000000000 +0800
> > > > +++ linux/include/linux/memcontrol.h  2011-10-25 09:33:51.000000000 +0800
> > > > @@ -37,7 +37,8 @@ extern unsigned long mem_cgroup_isolate_
> > > >                                       unsigned long *scanned, int order,
> > > >                                       int mode, struct zone *z,
> > > >                                       struct mem_cgroup *mem_cont,
> > > > -                                     int active, int file);
> > > > +                                     int active, int file,
> > > > +                                     struct page **split_page);
> > > >
> > > >  #ifdef CONFIG_CGROUP_MEM_RES_CTLR
> > > >  /*
> > > >
> > > >
> > >
> > > I saw the code. my concern is your patch could make unnecessary split of THP.
> > >
> > > When we isolates page, we can't know whether it's working set or not.
> > > So split should happen after we judge it's working set page.
> > yes, but since memory is big currently, it's unlikely the isolated page
> > get accessed in the window. And I only did the split in
> 
> We don't check page_reference when isolate happens.
> Window which between isolation time and reclaim?
> No. Window is from inactive's head to tail and it's the basic concept of
> our LRU.
> 
> > shrink_inactive_list, not in active list.
> 
> But inactive list's size could be still big and
> page reference heuristic is very important for reclaim algorithm.
I mean pages aren't referenced. but ok, I can't take such assumption.

> > And THP has mechanism to collapse small pages to huge page later.
> 
> You mean "merge" instead of "collapse"?
> 
> >
> > > If you really want to merge this patch, I suggest that
> > > we can handle it in shrink_page_list step, not isolation step.
> > >
> > > My totally untested code which is just to show the concept is as follows,
> > I did consider this option before. It has its problem too. The isolation
> > can isolate several huge page one time. And then later shrink_page_list
> > can swap several huge page one time, which is unfortunate. I'm pretty
> > sure this method can't reduce the thp_split count in my test. It could
> 
> I understand your point but approach isn't good to me.
> Maybe we can check whether we are going on or not before other THP page split happens
> in shrink_page_list. If we split THP page successfully, maybe we can skip another THP split.
> Another idea is we can avoid split of THP unless high order reclaim happens or low order
> high priority pressure happens.
I agreed the split better be done at shrink_page_list, but we must avoid
isolate too many pages. I'll check if I can have a better solution for
next post.

Thanks,
Shaohua