Re: [PATCH] mm: memcontrol: do not miss MEMCG_MAX events for enforced allocations

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri 01-07-22 20:35:21, Roman Gushchin wrote:
> Yafang Shao reported an issue related to the accounting of bpf
> memory: if a bpf map is charged indirectly for memory consumed
> from an interrupt context and allocations are enforced, MEMCG_MAX
> events are not raised.

So I guess this will be a GFP_ATOMIC request failing due to the hard
limit, right? I think it would be easier to understand if the specific
allocation request type was mentioned.

> It's not/less of an issue in a generic case because consequent
> allocations from a process context will trigger the reclaim and
> MEMCG_MAX events. However a bpf map can belong to a dying/abandoned
> memory cgroup, so it might never happen. So the cgroup can
> significantly exceed the memory.max limit without even triggering
> MEMCG_MAX events.

More on that in other reply.

> Fix this by making sure that we never enforce allocations without
> raising a MEMCG_MAX event.
> 
> Reported-by: Yafang Shao <laoar.shao@xxxxxxxxx>
> Signed-off-by: Roman Gushchin <roman.gushchin@xxxxxxxxx>
> Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
> Cc: Michal Hocko <mhocko@xxxxxxxxxx>
> Cc: Shakeel Butt <shakeelb@xxxxxxxxxx>
> Cc: Muchun Song <songmuchun@xxxxxxxxxxxxx>
> Cc: cgroups@xxxxxxxxxxxxxxx
> Cc: linux-mm@xxxxxxxxx
> Cc: bpf@xxxxxxxxxxxxxxx

The patch makes sense to me though even without the weird charge to a
dead memcg aspect. It is true that a very calm memcg can trigger the
even much later after a GFP_ATOMIC charge (or __GFP_HIGH in general)
fails.

Acked-by: Michal Hocko <mhocko@xxxxxxxx>

> ---
>  mm/memcontrol.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 655c09393ad5..eb383695659a 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -2577,6 +2577,7 @@ static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
>  	bool passed_oom = false;
>  	bool may_swap = true;
>  	bool drained = false;
> +	bool raised_max_event = false;
>  	unsigned long pflags;
>  
>  retry:
> @@ -2616,6 +2617,7 @@ static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
>  		goto nomem;
>  
>  	memcg_memory_event(mem_over_limit, MEMCG_MAX);
> +	raised_max_event = true;
>  
>  	psi_memstall_enter(&pflags);
>  	nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
> @@ -2682,6 +2684,13 @@ static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
>  	if (!(gfp_mask & (__GFP_NOFAIL | __GFP_HIGH)))
>  		return -ENOMEM;
>  force:
> +	/*
> +	 * If the allocation has to be enforced, don't forget to raise
> +	 * a MEMCG_MAX event.
> +	 */
> +	if (!raised_max_event)
> +		memcg_memory_event(mem_over_limit, MEMCG_MAX);
> +
>  	/*
>  	 * The allocation either can't fail or will lead to more memory
>  	 * being freed very soon.  Allow memory usage go over the limit
> -- 
> 2.36.1

-- 
Michal Hocko
SUSE Labs




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux