Re: memcg memory accounting in vmalloc is broken

On Thu 07-10-21 16:00:21, Vlastimil Babka wrote:

On 10/7/21 10:50, Vasily Averin wrote:

quoted

On 10/7/21 11:16 AM, Michal Hocko wrote:

quoted

Cc Mel and Uladzislau

On Thu 07-10-21 10:13:23, Michal Hocko wrote:

quoted

On Thu 07-10-21 11:04:40, Vasily Averin wrote:

quoted

vmalloc was switched to __alloc_pages_bulk but it does not account the memory to memcg.

Is it known issue perhaps?

No, I think this was just overlooked. Definitely doesn't look
intentional to me.

I use following patch as a quick fix,
it helps though it is far from ideal and can be optimized.

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index b37435c274cf..e6abe2cac159 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c

@@ -5290,6 +5290,12 @@ unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid,
 
 		page = __rmqueue_pcplist(zone, 0, ac.migratetype, alloc_flags,
 								pcp, pcp_list);
+
+		if (memcg_kmem_enabled() && (gfp & __GFP_ACCOUNT) && page &&
+		    unlikely(__memcg_kmem_charge_page(page, gfp, 0) != 0)) {
+			__free_pages(page, 0);

It's too early for this here, we didn't go through prep_new_page() yet,
minimally the post_alloc_hook() -> set_page_refcounted() part, required for
put_page_testzero() in __free_pages() to work properly, and probably other
stuff.
Either do the full prep+free or a minimal reversion of  __rmqueue_pcplist -
something to just put the page back to pcplist. Probably the former so we
don't introduce subtle errors.

Very well spotted!
-- 
Michal Hocko
SUSE Labs