Re: [PATCH v3 1/4] mm: memcontrol: drop kmemcg_id and use mem_cgroup_id() for list_lru indexing

[Qinyun Tan <qinyuntan@linux.alibaba.com>]

On 9/27/26 4:29 AM, Kairui Song wrote:
> 
> Hi Qinyun
> 
> With this commit, my arm64 VM hangs at every boot with a
> 100%-reproducible infinite loop in xas_find(), driven by
> the dcache shrinker during the first remount of the root filesystem.
> Reverting this fixes the issue:
> 
>    CPU: 2 UID: 0 PID: 579 Comm: mount Not tainted
> 7.3.0-rc4.orig-00725-g09d9672a5d4f #33 PREEMPT(full)
>    pc : xas_find+0x184/0x1c8
>    lr : xas_find+0x6c/0x1c8
>    Call trace:
>     xas_find+0x184/0x1c8
>     xa_find_after+0x88/0x120
>     list_lru_walk_node+0xc0/0x2a0
>     shrink_dcache_sb+0x80/0x130
>     reconfigure_super+0xc0/0x1f8
>     vfs_fsconfig_locked+0xa8/0x120
>     __arm64_sys_fsconfig+0x280/0x33c
> 
> Just for reference, this also fixed it:
> 
> diff --git a/mm/list_lru.c b/mm/list_lru.c
> index 7dbb6125cb1d..42c48c3b9235 100644
> --- a/mm/list_lru.c
> +++ b/mm/list_lru.c
> @@ -427,13 +427,13 @@ unsigned long list_lru_walk_node(struct list_lru
> *lru, int nid,
>   unsigned long index;
> 
>   xa_for_each(&lru->xa, index, mlru) {
> - rcu_read_lock();
> - memcg = mem_cgroup_from_private_id(index);
> - if (!memcg || !mem_cgroup_tryget(memcg)) {
> - rcu_read_unlock();
> + memcg = mem_cgroup_get_from_id(index);
> + if (!memcg)
>   continue;
> - }
> - rcu_read_unlock();
>   isolated += __list_lru_walk_one(lru, nid, memcg,
>   isolate, cb_arg,
>   nr_to_walk, false);
> 
> ===
> 
> And BTW cgroup ID lookups seem much heavier, not sure about the
> performance impact.

Hi Kairui,

Thanks for tracking this down. I missed the reverse lookup in
list_lru_walk_node() when switching to cgroup IDs. Your change fixes
that mismatch.

There is a catch with mem_cgroup_get_from_id(), though: it calls
cgroup_get_from_id(), which only looks up cgroups visible in the
current task's cgroup namespace. Here we need to walk all memcgs
on the LRU, including those outside that namespace, so we need
a lookup without that restriction.

Thanks,
Qinyun Tan