| Message ID | 1680086855-7989-1-git-send-email-zhaoyang.huang@unisoc.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | mm: mark folio as workingset in lru_deactivate_fn | expand |
On Wed, Mar 29, 2023 at 06:47:35PM +0800, zhaoyang.huang wrote: > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > folio will skip of being set as workingset in lru_deactivate_fn. Can you please elaborate why that's undesirable? What's the problem you're fixing?
On Wed, Mar 29, 2023 at 10:55 PM Johannes Weiner <hannes@cmpxchg.org> wrote: > > On Wed, Mar 29, 2023 at 06:47:35PM +0800, zhaoyang.huang wrote: > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > folio will skip of being set as workingset in lru_deactivate_fn. > > Can you please elaborate why that's undesirable? What's the problem > you're fixing? If I am correct, folio will skip being set as workingset when moving from active lru to inactive lru, which is performed on every folio in shrink_active_list during normal reclaim.
On Thu, Mar 30, 2023 at 09:38:48AM +0800, Zhaoyang Huang wrote: > On Wed, Mar 29, 2023 at 10:55 PM Johannes Weiner <hannes@cmpxchg.org> wrote: > > > > On Wed, Mar 29, 2023 at 06:47:35PM +0800, zhaoyang.huang wrote: > > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > > > folio will skip of being set as workingset in lru_deactivate_fn. > > > > Can you please elaborate why that's undesirable? What's the problem > > you're fixing? > If I am correct, folio will skip being set as workingset when moving > from active lru to inactive lru, which is performed on every folio in > shrink_active_list during normal reclaim. shrink_active_list directly calls folio_set_workingset(). The function you're editing is used for things like MADV_COLD and truncate(). It sounds like there is just a misunderstanding of the code, not an actual problem.
On Thu, Mar 30, 2023 at 5:32 PM Johannes Weiner <hannes@cmpxchg.org> wrote: > > On Thu, Mar 30, 2023 at 09:38:48AM +0800, Zhaoyang Huang wrote: > > On Wed, Mar 29, 2023 at 10:55 PM Johannes Weiner <hannes@cmpxchg.org> wrote: > > > > > > On Wed, Mar 29, 2023 at 06:47:35PM +0800, zhaoyang.huang wrote: > > > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > > > > > folio will skip of being set as workingset in lru_deactivate_fn. > > > > > > Can you please elaborate why that's undesirable? What's the problem > > > you're fixing? > > If I am correct, folio will skip being set as workingset when moving > > from active lru to inactive lru, which is performed on every folio in > > shrink_active_list during normal reclaim. > > shrink_active_list directly calls folio_set_workingset(). The function > you're editing is used for things like MADV_COLD and truncate(). Yes. > > It sounds like there is just a misunderstanding of the code, not an > actual problem. Isn't that a problem? As my understanding, MADV_COLD could be deemed as a stimulation of normal reclaiming which turbo the folio towards eviction, while the page moving by it should be also delt in the same way(PG_active has been cleaned)
On Thu, Mar 30, 2023 at 5:41 PM Zhaoyang Huang <huangzhaoyang@gmail.com> wrote: > > On Thu, Mar 30, 2023 at 5:32 PM Johannes Weiner <hannes@cmpxchg.org> wrote: > > > > On Thu, Mar 30, 2023 at 09:38:48AM +0800, Zhaoyang Huang wrote: > > > On Wed, Mar 29, 2023 at 10:55 PM Johannes Weiner <hannes@cmpxchg.org> wrote: > > > > > > > > On Wed, Mar 29, 2023 at 06:47:35PM +0800, zhaoyang.huang wrote: > > > > > From: Zhaoyang Huang <zhaoyang.huang@unisoc.com> > > > > > > > > > > folio will skip of being set as workingset in lru_deactivate_fn. > > > > > > > > Can you please elaborate why that's undesirable? What's the problem > > > > you're fixing? > > > If I am correct, folio will skip being set as workingset when moving > > > from active lru to inactive lru, which is performed on every folio in > > > shrink_active_list during normal reclaim. > > > > shrink_active_list directly calls folio_set_workingset(). The function > > you're editing is used for things like MADV_COLD and truncate(). > Yes. > > > > It sounds like there is just a misunderstanding of the code, not an > > actual problem. > Isn't that a problem? As my understanding, MADV_COLD could be deemed > as a stimulation of normal reclaiming which turbo the folio towards > eviction, while the page moving by it should be also delt in the same > way(PG_active has been cleaned) Sorry, I am still confused. Does it mean the pages deactivated via MADV_COLD like methods should NOT be deemed as workingset pages?
diff --git a/mm/swap.c b/mm/swap.c index 70e2063..4d1c14f 100644 --- a/mm/swap.c +++ b/mm/swap.c @@ -603,6 +603,7 @@ static void lru_deactivate_file_fn(struct lruvec *lruvec, struct folio *folio) lruvec_del_folio(lruvec, folio); folio_clear_active(folio); folio_clear_referenced(folio); + folio_set_workingset(folio); if (folio_test_writeback(folio) || folio_test_dirty(folio)) { /* @@ -637,6 +638,7 @@ static void lru_deactivate_fn(struct lruvec *lruvec, struct folio *folio) lruvec_del_folio(lruvec, folio); folio_clear_active(folio); folio_clear_referenced(folio); + folio_set_workingset(folio); lruvec_add_folio(lruvec, folio); __count_vm_events(PGDEACTIVATE, nr_pages);