On Mon, Nov 10, 2025 at 8:44 AM Jakub Kicinski [off-list ref] wrote:

On Mon, 10 Nov 2025 09:44:55 +0000 Eric Dumazet wrote:

quoted

Avoid up to two cache line misses in qdisc dequeue() to fetch
skb_shinfo(skb)->gso_segs/gso_size while qdisc spinlock is held.

Idea is to cache gso_segs at enqueue time before spinlock is
acquired, in the first skb cache line, where we already
have qdisc_skb_cb(skb)->pkt_len.

This series gives a 8 % improvement in a TX intensive workload.

(120 Mpps -> 130 Mpps on a Turin host, IDPF with 32 TX queues)

According to CI this breaks a bunch of tests.

https://netdev.bots.linux.dev/contest.html?branch=net-next-2025-11-10--12-00

I think they all hit:

[   20.682474][  T231] WARNING: CPU: 3 PID: 231 at ./include/net/sch_generic.h:843 __dev_xmit_skb+0x786/0x1550

Oh well, I will add this in V2, thank you !

diff --git a/include/net/sch_generic.h b/include/net/sch_generic.h
index b76436ec3f4aa412bac1be3371f5c7c6245cc362..79501499dafba56271b9ebd97a8f379ffdc83cac

100644

--- a/include/net/sch_generic.h
+++ b/include/net/sch_generic.h

@@ -841,7 +841,7 @@ static inline unsigned int qdisc_pkt_segs(const

struct sk_buff *skb)
        u32 pkt_segs = qdisc_skb_cb(skb)->pkt_segs;

        DEBUG_NET_WARN_ON_ONCE(pkt_segs !=
-                              skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1);
+                       (skb_is_gso(skb) ? skb_shinfo(skb)->gso_segs : 1));
        return pkt_segs;
 }