Re: [PATCH] radix-tree: account nodes to memcg only if explicitly requested

[Johannes Weiner]

On Mon, Aug 01, 2016 at 04:13:08PM +0300, Vladimir Davydov wrote:
> Radix trees may be used not only for storing page cache pages, so
> unconditionally accounting radix tree nodes to the current memory cgroup
> is bad: if a radix tree node is used for storing data shared among
> different cgroups we risk pinning dead memory cgroups forever. So let's
> only account radix tree nodes if it was explicitly requested by passing
> __GFP_ACCOUNT to INIT_RADIX_TREE. Currently, we only want to account
> page cache entries, so mark mapping->page_tree so.

Is this a theoretical fix, or did you actually run into problems? I
wouldn't expect any other radix tree node consumer in the kernel to
come anywhere close to the page cache, so I wonder why it matters.

> @@ -351,6 +351,12 @@ static int __radix_tree_preload(gfp_t gfp_mask, int nr)
>  	struct radix_tree_node *node;
>  	int ret = -ENOMEM;
>  
> +	/*
> +	 * Nodes preloaded by one cgroup can be be used by another cgroup, so
> +	 * they should never be accounted to any particular memory cgroup.
> +	 */
> +	gfp_mask &= ~__GFP_ACCOUNT;

But *all* page cache radix tree nodes are allocated from inside the
preload code, since the tree insertions need mapping->tree_lock. So
this would effectively disable accounting of the biggest radix tree
consumer in the kernel, no?