udp: remove busylock and add per NUMA queues

busylock was protecting UDP sockets against packet floods,
but unfortunately was not protecting the host itself.

Under stress, many cpus could spin while acquiring the busylock,
and NIC had to drop packets. Or packets would be dropped
in cpu backlog if RPS/RFS were in place.

This patch replaces the busylock by intermediate
lockless queues. (One queue per NUMA node).

This means that fewer number of cpus have to acquire
the UDP receive queue lock.

Most of the cpus can either:
- immediately drop the packet.
- or queue it in their NUMA aware lockless queue.

Then one of the cpu is chosen to process this lockless queue
in a batch.

The batch only contains packets that were cooked on the same
NUMA node, thus with very limited latency impact.

Tested:

DDOS targeting a victim UDP socket, on a platform with 6 NUMA nodes
(Intel(R) Xeon(R) 6985P-C)

Before:

nstat -n ; sleep 1 ; nstat | grep Udp
Udp6InDatagrams                 1004179            0.0
Udp6InErrors                    3117               0.0
Udp6RcvbufErrors                3117               0.0

After:
nstat -n ; sleep 1 ; nstat | grep Udp
Udp6InDatagrams                 1116633            0.0
Udp6InErrors                    14197275           0.0
Udp6RcvbufErrors                14197275           0.0

We can see this host can now proces 14.2 M more packets per second
while under attack, and the victim socket can receive 11 % more
packets.

I used a small bpftrace program measuring time (in us) spent in
__udp_enqueue_schedule_skb().

Before:

@udp_enqueue_us[398]:
[0]                24901 |@@@                                                 |
[1]                63512 |@@@@@@@@@                                           |
[2, 4)            344827 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[4, 8)            244673 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@                |
[8, 16)            54022 |@@@@@@@@                                            |
[16, 32)          222134 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@                   |
[32, 64)          232042 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@                  |
[64, 128)           4219 |                                                    |
[128, 256)           188 |                                                    |

After:

@udp_enqueue_us[398]:
[0]              5608855 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[1]              1111277 |@@@@@@@@@@                                          |
[2, 4)            501439 |@@@@                                                |
[4, 8)            102921 |                                                    |
[8, 16)            29895 |                                                    |
[16, 32)           43500 |                                                    |
[32, 64)           31552 |                                                    |
[64, 128)            979 |                                                    |
[128, 256)            13 |                                                    |

Note that the remaining bottleneck for this platform is in
udp_drops_inc() because we limited struct numa_drop_counters
to only two nodes so far.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250922104240.2182559-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

1 files changed, 8 insertions, 1 deletions

diff --git a/include/linux/udp.h b/include/linux/udp.h
index e554890c4415..58795688a186 100644
--- a/include/linux/udp.h
+++ b/include/linux/udp.h
@@ -44,6 +44,12 @@ enum {
 	UDP_FLAGS_UDPLITE_RECV_CC, /* set via udplite setsockopt */
 };
 
+/* per NUMA structure for lockless producer usage. */
+struct udp_prod_queue {
+	struct llist_head	ll_root ____cacheline_aligned_in_smp;
+	atomic_t		rmem_alloc;
+};
+
 struct udp_sock {
 	/* inet_sock has to be the first member */
 	struct inet_sock inet;
@@ -90,6 +96,8 @@ struct udp_sock {
 						struct sk_buff *skb,
 						int nhoff);
 
+	struct udp_prod_queue *udp_prod_queue;
+
 	/* udp_recvmsg try to use this before splicing sk_receive_queue */
 	struct sk_buff_head	reader_queue ____cacheline_aligned_in_smp;
 
@@ -109,7 +117,6 @@ struct udp_sock {
 	 */
 	struct hlist_node	tunnel_list;
 	struct numa_drop_counters drop_counters;
-	spinlock_t		busylock ____cacheline_aligned_in_smp;
 };
 
 #define udp_test_bit(nr, sk)			\