Re: [PATCH v4] mm: mglru: fix stale batch updates after memcg reparenting

[Qi Zheng <qi.zheng@linux.dev>]

On 7/10/26 9:51 AM, Johannes Weiner wrote:
> On Thu, Jul 02, 2026 at 09:38:41AM +0800, Qi Zheng wrote:
>>
>>
>> On 7/1/26 11:59 PM, Johannes Weiner wrote:
>>> On Thu, Jul 02, 2026 at 12:36:42AM +0900, Harry Yoo wrote:
>>>>
>>>>
>>>> On 7/1/26 11:57 PM, Usama Arif wrote:
>>>>> On Wed,  1 Jul 2026 15:52:51 +0800 Qi Zheng <qi.zheng@linux.dev> wrote:
>>>>>
>>>>>> From: Qi Zheng <zhengqi.arch@bytedance.com>
>>>>>>
>>>>>> The mglru page table walker batches per-generation size deltas in
>>>>>> walk->nr_pages while walking page tables without holding the lruvec lock.
>>>>>> The reset_batch_size() later folds those deltas into walk->lruvec under
>>>>>> the lruvec lock.
>>>>>>
>>>>>> The page table walker can run concurrently with the memcg reparenting path
>>>>>> as follows:
>>>>>>
>>>>>> CPU0                           CPU1
>>>>>> ====                           ====
>>>>>>
>>>>>> walk_mm
>>>>>> --> walk_page_range
>>>>>>       --> update_batch_size
>>>>>>           --> walk->nr_pages += delta
>>>>>>
>>>>>>                                 mem_cgroup_css_offline
>>>>>>                                 --> memcg_reparent_objcgs
>>>>>>                                     --> lock lruvec
>>>>>>                                         lru_gen_reparent_memcg
>>>>>>                                         --> reparent child folios to parent
>>>>>>                                         unlock lruvec
>>>>>>
>>>>>>       lock lruvec
>>>>>>       reset_batch_size
>>>>>>       --> child lrugen->nr_pages += delta
>>>>>>
>>>>>> This will trigger the following warning in lru_gen_exit_memcg():
>>>>>>
>>>>>> 	VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0,
>>>>>> 				   sizeof(lruvec->lrugen.nr_pages)));
>>>>>>
>>>>>> And the user-visible impact of underestimated nr_pages in MGLRU was
>>>>>> premature OOMs because MGLRU does not try to reclaim memory when nr_pages
>>>>>> reaches zero, but there are still more pages.
>>>>>>
>>>>>> To fix it, make reset_batch_size() check CSS_DYING under RCU before
>>>>>> flushing the pending batch. A non-dying memcg keeps the original lruvec
>>>>>> stable against RCU-delayed offlining; a dying memcg redirects the deltas
>>>>>> to the first non-dying ancestor.
>>>>>>
>>>>>> Reported-by: Peiyang He <peiyang_he@smail.nju.edu.cn>
>>>>>> Closes: https://lore.kernel.org/all/5A9E929D82717101+12fcf643-efb8-4b9a-a53a-1e28cc894f0b@smail.nju.edu.cn
>>>>>> Fixes: f304652609ea ("mm: vmscan: prepare for reparenting MGLRU folios")
>>>>>> Cc: <stable@vger.kernel.org>
>>>>>> Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
>>>>>> Reviewed-by: Harry Yoo (Oracle) <harry@kernel.org>
>>>>>> ---
>>>>>> Changes in v4:
>>>>>>    - re-implement lock_batch_lruvec() in a simpler way
>>>>>>      (suggested by Johannes and Harry)
>>>>>>    - collect Reviewed-by
>>>>>>    - rebase onto the next-20260630
>>>>>>
>>>>>> Changes in v3:
>>>>>>    - re-implement lock_batch_lruvec() by checking CSS_DYING under the RCU lock
>>>>>>      (suggested by Harry)
>>>>>>    - update the commit message (suggested by Harry)
>>>>>>    - temporarily drop the previous Reviewed-by tags
>>>>>>      (since the sync method has changed)
>>>>>>    - rebase onto the next-20260624
>>>>>>
>>>>>> Changes in v2:
>>>>>>    - update the commit message (pointed by Barry)
>>>>>>    - collect Reviewed-by
>>>>>>
>>>>>>    mm/vmscan.c | 41 ++++++++++++++++++++++++++++++++++-------
>>>>>>    1 file changed, 34 insertions(+), 7 deletions(-)
>>>>>>
>>>>>> diff --git a/mm/vmscan.c b/mm/vmscan.c
>>>>>> index 35c3bb15ae96..ca1e2a870d51 100644
>>>>>> --- a/mm/vmscan.c
>>>>>> +++ b/mm/vmscan.c
>>>>>> @@ -3262,10 +3262,40 @@ static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio,
>>>>>>    	walk->nr_pages[new_gen][type][zone] += delta;
>>>>>>    }
>>>>>>    
>>>>>> +#ifdef CONFIG_MEMCG
>>>>>> +static struct lruvec *lock_batch_lruvec(struct lruvec *lruvec)
>>>>>> +{
>>>>>> +	struct pglist_data *pgdat = lruvec_pgdat(lruvec);
>>>>>> +	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
>>>>>> +
>>>>>> +	rcu_read_lock();
>>>>>> +
>>>>>> +	/*
>>>>>> +	 * The memcg can be NULL when the memory controller is disabled.
>>>>>> +	 * Otherwise, the caller keeps the memcg owning @lruvec alive.
>>>>>> +	 */
>>>>>> +	while (unlikely(memcg && css_is_dying(&memcg->css))) {
>>>>>> +		memcg = parent_mem_cgroup(memcg);
>>>>>> +		lruvec = mem_cgroup_lruvec(memcg, pgdat);
>>>>>> +	}
>>>>>> +
>>>>>> +	spin_lock_irq(&lruvec->lru_lock);
>>>>>
>>>>> Do we need an rcu_read_unlock() here?
>>>>
>>>> lruvec_unlock_irq() does that.
>>>
>>> Yeah, that tripped me up too. And it makes me think Shakeel was right
>>> after all: this should live next to the other lruvec_lock() primitives.
>>>
>>> Sure, MGLRU is the only user, but it's still much easier to understand
>>> this if the code sits next to the rest of the API (and the unlock!).
>>>
>>> lruvec_live_lock_irq()?
>>
>> But lruvec_lock_irq() grabs the rcu lock too. :(
> 
> Yes, but it's self-explanatory if you put it with those definitions:
> 
> static inline void lruvec_lock_irq(struct lruvec *lruvec)
> {
>          rcu_read_lock();
>          spin_lock_irq(&lruvec->lru_lock);
> }
> 
> static struct lruvec *lruvec_live_lock_irq(struct lruvec *lruvec)
> {
> 	struct mem_cgroup *memcg = lruvec_memcg(lruvec);
> 
> 	rcu_read_lock();
> 	while (unlikely(memcg && css_is_dying(&memcg->css))) {
> 		memcg = parent_mem_cgroup(memcg);
> 		lruvec = mem_cgroup_lruvec(memcg, lruvec_pgdat(lruvec));
> 	}
> 	spin_lock_irq(&lruvec->lru_lock);
> }

All right, should the implementation for !CONFIG_MEMCG be placed here
too?

Will send the v5.

Thanks,
Qi

> 
> static inline void lruvec_unlock_irq(struct lruvec *lruvec)
> {
>          spin_unlock_irq(&lruvec->lru_lock);
>          rcu_read_unlock();
> }
> 
> etc.