Re: [PATCH] mm: memcg: flush empty per-cpu stock slots on memcg offlining

On Mon, Aug 17, 2026 at 09:58:39PM +0800, Song Hu wrote:
> Hi,Michal
> 
> 在 2026/8/17 21:29, Michal Hocko 写道:
> > On Mon 17-08-26 21:12:21, Song Hu wrote:
> >> On Mon 17-08-26, Michal Hocko wrote:
> >>> Is there any specific reason why the memcg stays in the cache slot
> >>> without any pages?
> >>
> >> consume_stock() doesn't release the slot when nr_pages hits zero.  It
> >> is kept for the next charge of the same task and only gets displaced
> >> by a charge under a different memcg or by CPU hotplug.  The problem is
> >> that the offlining drain skips empty slots, so the css reference they
> >> hold is never dropped unless something unrelated displaces them.
> > 
> > This doesn't answer my question, really, does it? Is there any good
> > reason for this implementation? Why do we need to drop references
> > remotely when we can do so when the last cached charge is consumed?
> > 
> 
> Fair enough.  There is no strong reason.  Keeping the slot
> populated after the last page is consumed only saves a
> css_get()/css_put() pair when the same memcg charges again on that
> CPU - a micro-optimization from the original single-slot
> implementation.
> 
> Dropping the reference in consume_stock() when the slot empties is
> the better place.  Empty slots stop existing, so the offlining drain
> has nothing left to miss and is_memcg_drain_needed() stays as it
> is.  This also makes Joshua's concern about the full-stock drain go
> away entirely.  The cost is one refcount pair per emptied slot, at
> most once per MEMCG_CHARGE_BATCH pages.
> 
> Joshua, this supersedes the css_is_dying gating you suggested and
> that I said I would do - with no empty slots left, the check would
> never fire, and the kill_css_sync() ordering argument becomes moot
> as well.  Since you are reworking this code, I'd appreciate a
> sanity check on releasing from consume_stock().
> 
> If this direction works for you both, I'll rework the patch
> accordingly.

Yes, this seems reasonable. Go ahead with the rework.