Re: [PATCH v2 4/4] mm: thp: reparent the split queue during memcg offline

[Shakeel Butt <shakeel.butt@linux.dev>]

On Thu, Sep 25, 2025 at 03:15:26PM -0700, Shakeel Butt wrote:
> On Thu, Sep 25, 2025 at 03:49:52PM -0400, Zi Yan wrote:
> > On 25 Sep 2025, at 15:35, David Hildenbrand wrote:
> > 
> > > On 25.09.25 08:11, Qi Zheng wrote:
> > >> Hi David,
> > >
> > > Hi :)
> > >
> > > [...]
> > >
> > >>>> +++ b/include/linux/mmzone.h
> > >>>> @@ -1346,6 +1346,7 @@ struct deferred_split {
> > >>>>        spinlock_t split_queue_lock;
> > >>>>        struct list_head split_queue;
> > >>>>        unsigned long split_queue_len;
> > >>>> +    bool is_dying;
> > >>>
> > >>> It's a bit weird to query whether the "struct deferred_split" is dying.
> > >>> Shouldn't this be a memcg property? (and in particular, not exist for
> > >>
> > >> There is indeed a CSS_DYING flag. But we must modify 'is_dying' under
> > >> the protection of the split_queue_lock, otherwise the folio may be added
> > >> back to the deferred_split of child memcg.
> > >
> > > Is there no way to reuse the existing mechanisms, and find a way to have the shrinker / queue locking sync against that?
> > >
> > > There is also the offline_css() function where we clear CSS_ONLINE. But it happens after calling ss->css_offline(css);
> > 
> > I see CSS_DYING will be set by kill_css() before offline_css() is called.
> > Probably the code can check CSS_DYING instead.
> > 
> > >
> > > Being able to query "is the memcg going offline" and having a way to sync against that would be probably cleanest.
> > 
> > So basically, something like:
> > 1. at folio_split_queue_lock*() time, get folio’s memcg or
> >    its parent memcg until there is no CSS_DYING set or CSS_ONLINE is set.
> > 2. return the associated deferred_split_queue.
> > 
> 
> Yes, css_is_dying() can be used but please note that there is a rcu
> grace period between setting CSS_DYING and clearing CSS_ONLINE (i.e.
> reparenting deferred split queue) and during that period the deferred
> split THPs of the dying memcg will be hidden from shrinkers (which
> might be fine).

BTW if this period is not acceptable and we don't want to add is_dying
to struct deferred_split, we can use something similar to what list_lru
does in the similar situation i.e. set a special value (LONG_MIN) in its
nr_items variable. That is make split_queue_len a long and set it to
LONG_MIN during memcg offlining/reparenting.