Re: [PATCH] mm: memcg: fix NULL pointer in mem_cgroup_track_foreign_dirty()

From: Kefeng Wang
Date: Sun Jan 29 2023 - 20:16:46 EST




On 2023/1/30 5:48, Andrew Morton wrote:
On Sun, 29 Jan 2023 10:44:51 +0800 Kefeng Wang <wangkefeng.wang@xxxxxxxxxx> wrote:

As commit 18365225f044 ("hwpoison, memcg: forcibly uncharge LRU pages"),

Merged in 2017.

hwpoison will forcibly uncharg a LRU hwpoisoned page, the folio_memcg
could be NULl, then, mem_cgroup_track_foreign_dirty_slowpath() could
occurs a NULL pointer dereference, let's do not record the foreign
writebacks for folio memcg is null in mem_cgroup_track_foreign() to
fix it.

Reported-by: Ma Wupeng <mawupeng1@xxxxxxxxxx>
Fixes: 97b27821b485 ("writeback, memcg: Implement foreign dirty flushing")

Merged in 2019.

--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -1688,10 +1688,13 @@ void mem_cgroup_track_foreign_dirty_slowpath(struct folio *folio,
static inline void mem_cgroup_track_foreign_dirty(struct folio *folio,
struct bdi_writeback *wb)
{
+ struct mem_cgroup *memcg;
+
if (mem_cgroup_disabled())
return;
- if (unlikely(&folio_memcg(folio)->css != wb->memcg_css))
+ memcg = folio_memcg(folio);
+ if (unlikely(memcg && &memcg->css != wb->memcg_css))
mem_cgroup_track_foreign_dirty_slowpath(folio, wb);
}

Has this null deref actually been observed, or is this from code
inspection? (This is why it's nice to include the Link: after a
Reported-by!)

It does occurs in our internal test and report by wupeng(based v5.10),

BUG: KASAN: user-memory-access in mem_cgroup_track_foreign_dirty_slowpath+0xc0/0x480 mm/memcontrol.c:4708
Read of size 8 at addr 0000000000001000 by task syz-executor.2/28325

CPU: 2 PID: 28325 Comm: syz-executor.2 Tainted: G W 5.10.0-03333-g48e46a146cbc-dirty #1
Hardware name: linux,dummy-virt (DT)
Call trace:
...
mem_cgroup_track_foreign_dirty_slowpath+0xc0/0x480 mm/memcontrol.c:4708
mem_cgroup_track_foreign_dirty include/linux/memcontrol.h:1880 [inline]
account_page_dirtied+0x9a0/0xa90 mm/page-writeback.c:2436
__set_page_dirty+0x1f8/0x4b0 fs/buffer.c:608
__set_page_dirty_buffers+0x3d0/0x550 fs/buffer.c:668
set_page_dirty+0x158/0x500 mm/page-writeback.c:2575
filemap_page_mkwrite+0x3dc/0x490 mm/filemap.c:3224
do_page_mkwrite+0xc4/0x3d0 mm/memory.c:2786
wp_page_shared+0x14c/0x980 mm/memory.c:3118
do_wp_page+0x930/0xbc0 mm/memory.c:3219
handle_pte_fault+0x5e0/0x630 mm/memory.c:4570
__handle_mm_fault+0x41c/0x910 mm/memory.c:4690
handle_mm_fault+0x25c/0x484 mm/memory.c:4788
__do_page_fault arch/arm64/mm/fault.c:440 [inline]
do_page_fault+0x3ac/0x9d4 arch/arm64/mm/fault.c:539


Do we have any theories why this took so many years to surface?

After google, I found a similar issue[1], maybe hwpoison/mem_cgroup_track_foreign_dirty concurrency is uncommon.

[1] https://syzkaller.appspot.com/bug?extid=0c84bf23aed8ee0d8399


I'm confused about the mention of 18365225f044, but the Fixes: target
is a different commit. Please explain this?

18365225f044 said that it will uncharge it manually from its memcg in hwpison handler, but when the new feature "writeback, memcg: Implement foreign dirty flushing" is introduced, we don't consider this, when
folio's memcg is cleared by hwpison handler, the issue occurs.


Do you think the fix should be backported into earlier -stable kernels?

it's better to send stable kernel.

If so, it will need some rework due to the subsequent folio
conversion.

When the patch is merged, I could refresh and send to stable maillist.