On Wed, Jan 22, 2014 at 04:20:36PM +0100, Hannes Frederic Sowa wrote:

On Wed, Jan 22, 2014 at 07:11:20AM -0800, Dan Ballard wrote:

quoted

diff --git a/net/core/sock.c b/net/core/sock.c
index 5393b4b..1ff69d1 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c

@@ -915,6 +915,10 @@ set_rcvbuf:
                                         sk->sk_max_pacing_rate);
                break;

+       case SO_MAX_DGRAM_QLEN:
+               sk->sk_max_ack_backlog = val;
+               break;
+

Shouldn't the backlog be capped for unprivileged users to some configurable
value? I even think that max_dgram_qlen should be the upper bound.

I guess it is not that serious as socket read accounting does account all
packets which sit in the backlog queue.

Just a follow-up:

sk_max_ack_backlog is also responsible for limiting the af_unix
dgram queues.  Currently there is no socket accounting for the read
side of those unix dgram sockets. I tried to fix this once here,
http://patchwork.ozlabs.org/patch/231032/, but until that is done we
depend on max_dgram_qlen to limit those queues at all.

I hope I can get back to this patch anytime soon, as it solves the problem
that a bidirectional protocol ping-ponging with a dgram server socket
and not fetching its messages from the backlog queue can bring a server
to halt because it doesn't have any send space on the socket anymore.

Greetings,

  Hannes