Re: [PATCH bpf] flow_dissector: Drop BPF flow dissector prog ref on netns cleanup

[sdf@google.com]

On 05/20, Jakub Sitnicki wrote:
> When attaching a flow dissector program to a network namespace with
> bpf(BPF_PROG_ATTACH, ...) we grab a reference to bpf_prog.

> If netns gets destroyed while a flow dissector is still attached, and  
> there
> are no other references to the prog, we leak the reference and the program
> remains loaded.

> Leak can be reproduced by running flow dissector tests from selftests/bpf:

>    # bpftool prog list
>    # ./test_flow_dissector.sh
>    ...
>    selftests: test_flow_dissector [PASS]
>    # bpftool prog list
>    4: flow_dissector  name _dissect  tag e314084d332a5338  gpl
>            loaded_at 2020-05-20T18:50:53+0200  uid 0
>            xlated 552B  jited 355B  memlock 4096B  map_ids 3,4
>            btf_id 4
>    #

> Fix it by detaching the flow dissector program when netns is going away.

> Fixes: d58e468b1112 ("flow_dissector: implements flow dissector BPF hook")
> Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
> ---

> Discovered while working on bpf_link support for netns-attached progs.
> Looks like bpf tree material so pushing it out separately.
Oh, good catch!

> -jkbs

>   net/core/flow_dissector.c | 29 ++++++++++++++++++++++++++++-
>   1 file changed, 28 insertions(+), 1 deletion(-)

> diff --git a/net/core/flow_dissector.c b/net/core/flow_dissector.c
> index 3eff84824c8b..b6179cd20158 100644
> --- a/net/core/flow_dissector.c
> +++ b/net/core/flow_dissector.c
> @@ -179,6 +179,27 @@ int skb_flow_dissector_bpf_prog_detach(const union  
> bpf_attr *attr)
>   	return 0;
>   }

> +static void __net_exit flow_dissector_pernet_pre_exit(struct net *net)
> +{
> +	struct bpf_prog *attached;
> +
> +	/* We don't lock the update-side because there are no
> +	 * references left to this netns when we get called. Hence
> +	 * there can be no attach/detach in progress.
> +	 */
> +	rcu_read_lock();
> +	attached = rcu_dereference(net->flow_dissector_prog);
> +	if (attached) {
> +		RCU_INIT_POINTER(net->flow_dissector_prog, NULL);
> +		bpf_prog_put(attached);
> +	}
> +	rcu_read_unlock();
> +}
I wonder, should we instead refactor existing
skb_flow_dissector_bpf_prog_detach to accept netns (instead of attr)
can call that here? Instead of reimplementing it (I don't think we
care about mutex lock/unlock efficiency here?). Thoughts?