Re: [PATCH] mm, memcg: fix wrong mem cgroup protection

[Roman Gushchin <guro@fb.com>]

On Thu, Apr 23, 2020 at 02:16:29AM -0400, Yafang Shao wrote:
> This patch is an improvement of a previous version[1], as the previous
> version is not easy to understand.
> This issue persists in the newest kernel, I have to resend the fix. As
> the implementation is changed, I drop Roman's ack from the previous
> version.
> 
> Here's the explanation of this issue.
> memory.{low,min} won't take effect if the to-be-reclaimed memcg is the
> sc->target_mem_cgroup, that can also be proved by the implementation in
> mem_cgroup_protected(), see bellow,
> 	mem_cgroup_protected
> 		if (memcg == root) [2]
> 			return MEMCG_PROT_NONE;
> 
> But this rule is ignored in mem_cgroup_protection(), which will read
> memory.{emin, elow} as the protection whatever the memcg is.
> 
> How would this issue happen?
> Because in mem_cgroup_protected() we forget to clear the
> memory.{emin, elow} if the memcg is target_mem_cgroup [2].
> 
> An example to illustrate this issue.
>    root_mem_cgroup
>          /
>         A   memory.max: 1024M
>             memory.min: 512M
>             memory.current: 800M ('current' must be greater than 'min')
> Once kswapd starts to reclaim memcg A, it assigns 512M to memory.emin of A.
> Then kswapd stops.
> As a result of it, the memory values of A will be,
>    root_mem_cgroup
>          /
>         A   memory.max: 1024M
>             memory.min: 512M
>             memory.current: 512M (approximately)
>             memory.emin: 512M
> 
> Then a new workload starts to run in memcg A, and it will trigger memcg
> relcaim in A soon. As memcg A is the target_mem_cgroup of this
> reclaimer, so it return directly without touching memory.{emin, elow}.[2]
> The memory values of A will be,
>    root_mem_cgroup
>          /
>         A   memory.max: 1024M
>             memory.min: 512M
>             memory.current: 1024M (approximately)
>             memory.emin: 512M
> Then this memory.emin will be used in mem_cgroup_protection() to get the
> scan count, which is obvoiusly a wrong scan count.
> 
> [1]. https://lore.kernel.org/linux-mm/20200216145249.6900-1-laoar.shao@gmail.com/
> 
> Fixes: 9783aa9917f8 ("mm, memcg: proportional memory.{low,min} reclaim")
> Cc: Chris Down <chris@chrisdown.name>
> Cc: Roman Gushchin <guro@fb.com>
> Cc: stable@vger.kernel.org
> Signed-off-by: Yafang Shao <laoar.shao@gmail.com>
> ---
>  include/linux/memcontrol.h | 13 +++++++++++--
>  mm/vmscan.c                |  4 ++--
>  2 files changed, 13 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index d275c72c4f8e..114cfe06bf60 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -344,12 +344,20 @@ static inline bool mem_cgroup_disabled(void)
>  	return !cgroup_subsys_enabled(memory_cgrp_subsys);
>  }
>  
> -static inline unsigned long mem_cgroup_protection(struct mem_cgroup *memcg,
> +static inline unsigned long mem_cgroup_protection(struct mem_cgroup *root,
> +						  struct mem_cgroup *memcg,
>  						  bool in_low_reclaim)

I'd rename "root" to "target", maybe it will make the whole thing more clear.

I'll think a bit more about it, but at the first glance the patch looks sane to me.

Thanks!