On 15.03.25 00:15, Zi Yan wrote:
On 14 Mar 2025, at 17:49, David Hildenbrand wrote:
On 14.03.25 14:36, Matthew Wilcox (Oracle) wrote:
Folios always use memcg_data to refer to the mem_cgroup while pages
allocated with GFP_ACCOUNT have a pointer to the obj_cgroup. Since the
caller already knows what it has, split the function into two and then
we don't need to check.
Move the assignment of split folio memcg_data to the point where we set
up the other parts of the new folio. That leaves folio_split_memcg_refs()
just handling the memcg accounting.
Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>
Acked-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
Acked-by: Zi Yan <ziy@xxxxxxxxxx>
---
include/linux/memcontrol.h | 7 +++++++
mm/huge_memory.c | 16 ++++------------
mm/memcontrol.c | 17 +++++++++++++----
3 files changed, 24 insertions(+), 16 deletions(-)
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index 57664e2a8fb7..d090089c5497 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -1039,6 +1039,8 @@ static inline void memcg_memory_event_mm(struct mm_struct *mm,
}
void split_page_memcg(struct page *head, int old_order, int new_order);
+void folio_split_memcg_refs(struct folio *folio, unsigned old_order,
+ unsigned new_order);
static inline u64 cgroup_id_from_mm(struct mm_struct *mm)
{
@@ -1463,6 +1465,11 @@ static inline void split_page_memcg(struct page *head, int old_order, int new_or
{
}
+static inline void folio_split_memcg_refs(struct folio *folio,
+ unsigned old_order, unsigned new_order)
+{
+}
+
static inline u64 cgroup_id_from_mm(struct mm_struct *mm)
{
return 0;
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 14b1963898a7..3e5ecc8f3d13 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -3394,6 +3394,9 @@ static void __split_folio_to_order(struct folio *folio, int old_order,
folio_set_young(new_folio);
if (folio_test_idle(folio))
folio_set_idle(new_folio);
+#ifdef CONFIG_MEMCG
+ new_folio->memcg_data = folio->memcg_data;
+#endif
folio_xchg_last_cpupid(new_folio, folio_last_cpupid(folio));
}
@@ -3525,18 +3528,7 @@ static int __split_unmapped_folio(struct folio *folio, int new_order,
}
}
- /*
- * Reset any memcg data overlay in the tail pages.
- * folio_nr_pages() is unreliable until prep_compound_page()
- * was called again.
- */
-#ifdef NR_PAGES_IN_LARGE_FOLIO
- folio->_nr_pages = 0;
-#endif
I remember that we could trigger a warning without that, but I don't immediately find where that warning was. IIRC, if we'd split to order-0, page[1] would have indicated that it had a memcg set, and something bailed out.
Maybe Zi Yan recalls where that check fired.
The error I encountered is different. When I rebase my folio_split()
on top of David’s mapcount patchset, my original patch used folio_nr_pages()
after memcg split. Since memcg overlays with _nr_pages, when splitting
to order-0, folio->_nr_page is overwritten with memcg_data, causing
folio_nr_pages() to return a bogus value. With Matthew’s this patch,
memcg_data of page[1] is written inside __split_folio_to_order(),
so in theory __split_folio_to_order() can call folio_nr_pages() like
my original patch.
Ah, my memory comes back. Right, I fixed that by passing "old_order"
instead when rebasing on your series and ...
For folio->_nr_pages = 0, I suppose it is trying to suppress any
page[1]->memcg_data != NULL check in the following code. But I could
not find any.
... that was likely an issue I ran into before rebasing on your code,
where I had to effectively clear page[1]->memcg_data.
So all good, clearing folio->_nr_pages can be dropped.
--
Cheers,
David / dhildenb
