Re: [linus:master] [mm] 01b9da291c: stress-ng.switch.ops_per_sec 67.7% regression

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

On Fri, May 15, 2026 at 03:37:22PM +0800, Qi Zheng wrote:
> Hi Shakeel,
> 
> On 5/14/26 9:40 PM, Shakeel Butt wrote:
> > May 14, 2026 at 12:46 AM, "Qi Zheng" <qi.zheng@linux.dev mailto:qi.zheng@linux.dev?to=%22Qi%20Zheng%22%20%3Cqi.zheng%40linux.dev%3E > wrote:
> > 
> > 
> > > 
> > > On 5/13/26 10:27 PM, Shakeel Butt wrote:
> > > 
> > > > 
> > > > On Wed, May 13, 2026 at 06:49:45AM -0700, Shakeel Butt wrote:
> > > > 
> > > > > 
> > > > > On Wed, May 13, 2026 at 10:10:34AM +0800, Qi Zheng wrote:
> > > > > 
> > > >   On 5/13/26 12:03 AM, Shakeel Butt wrote:
> > > >   On Tue, May 12, 2026 at 08:56:52PM +0800, kernel test robot wrote:
> > > > 
> > > >   Hello,
> > > > 
> > > >   kernel test robot noticed a 67.7% regression of stress-ng.switch.ops_per_sec on:
> > > > 
> > > >   commit: 01b9da291c4969354807b52956f4aae1f41b4924 ("mm: memcontrol: convert objcg to be per-memcg per-node type")
> > > >   https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git master
> > > > 
> > > >   This is most probably due to shuffling of struct mem_cgroup and struct
> > > >   mem_cgroup_per_node members.
> > > > 
> > > >   Another possibility is that after objcg was split into per-node, the
> > > >   slab accounting fast path is still designed assuming only one current
> > > >   objcg per CPU:
> > > > 
> > > >   struct obj_stock_pcp {
> > > >   struct obj_cgroup *cached_objcg;
> > > >   };
> > > > 
> > > >   So it's may cause the following thrashing:
> > > > 
> > > >   CPU stock cached = memcg/node0 objcg
> > > >   free object tagged = memcg/node1 objcg
> > > >   => __refill_obj_stock --> objcg mismatch
> > > >   => drain_obj_stock()
> > > >   => cache switches to node1 objcg
> > > > 
> > > >   next local allocation tagged = node0 objcg
> > > >   => mismatch again
> > > >   => drain_obj_stock()
> > > > 
> > > > > 
> > > > > Actually I think this is the issue, we have ping pong threads running on
> > > > >   different nodes where though theu are in same cgroup but their current->obcg is
> > > > >   for local node and thus this ping pong is thrashing the per-cpu objcg stock.
> > > > > 
> > > > >   The easier fix would be to compare objcg->memcg instead of just objcg during
> > > > >   draining and caching. In addition we can add support for multiple objcg per-cpu
> > > > >   stock caching.
> > > > > 
> > > >   Something like the following:
> > > >   From d756abe831a905d6fe32bad9a984fc619dafb7e0 Mon Sep 17 00:00:00 2001
> > > >   From: Shakeel Butt <shakeel.butt@linux.dev>
> > > >   Date: Wed, 13 May 2026 07:24:55 -0700
> > > >   Subject: [PATCH] mm/memcontrol: skip obj_stock drain when refilled objcg
> > > >   shares memcg
> > > >   Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>
> > > >   ---
> > > >   mm/memcontrol.c | 14 +++++++++++++-
> > > >   1 file changed, 13 insertions(+), 1 deletion(-)
> > > >   diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > >   index d978e18b9b2d..01ed7a8e18ac 100644
> > > >   --- a/mm/memcontrol.c
> > > >   +++ b/mm/memcontrol.c
> > > >   @@ -3318,6 +3318,7 @@ static void __refill_obj_stock(struct obj_cgroup *objcg,
> > > >   unsigned int nr_bytes,
> > > >   bool allow_uncharge)
> > > >   {
> > > >   + struct obj_cgroup *cached;
> > > >   unsigned int nr_pages = 0;
> > > >   > if (!stock) {
> > > >   @@ -3327,7 +3328,18 @@ static void __refill_obj_stock(struct obj_cgroup *objcg,
> > > >   goto out;
> > > >   }
> > > >   > - if (READ_ONCE(stock->cached_objcg) != objcg) { /* reset if necessary */
> > > >   + cached = READ_ONCE(stock->cached_objcg);
> > > >   + if (cached != objcg &&
> > > >   + (!cached || obj_cgroup_memcg(cached) != obj_cgroup_memcg(objcg))) {
> > > >   drain_obj_stock(stock);
> > > >   obj_cgroup_get(objcg);
> > > >   stock->nr_bytes = atomic_read(&objcg->nr_charged_bytes)
> > > > 
> > > This change looks like it should be able to fix the ping-pong issue, but
> > > I stiil haven't reproduced the performance regression locally. I'll
> > > continue testing it.
> > 
> > Same here, couldn't reproduce locally. It seems like we had to craft a scenario
> > where the pair pingpong threads get their current->objcg from different nodes.
> > I will try that.
> 
> I still haven't been able to reproduce the LKP results locally, but I
> used an AI bot to generate a pingpong test case (pasted at the end) and
> automatically ran the test on a physical machine. The results are as
> follows:
> 
>   parent: 8285917d6f
>   bad:    01b9da291c
>   fix:    01b9da291c + stock patch
> 
>   | kernel | mq_ops/sec mean | vs parent | drain_obj_stock / round |
>   |--------|-----------------|-----------|-------------------------|
>   | parent |     9.743M      |  baseline |          ~0             |
>   | bad    |     7.821M      |  -19.73%  |          ~11.16M        |
>   | fix    |     9.274M      |  -4.81%   |          ~0             |
> 
> Probing the drain_obj_stock() calls confirms that the fix restores the
> frequency to the parent's baseline.
> 
> And it seems that besides __refill_obj_stock(), we should also modify
> __consume_obj_stock()?
> 

Thanks a lot Qi. I will send the formal patch and will add your Debugged-by if
you don't mind.