On Tue, Jan 20, 2026 at 4:41 PM Eric Dumazet [off-list ref] wrote:

On Tue, Jan 20, 2026 at 4:30 PM Jakub Kicinski [off-list ref] wrote:

quoted

On Sun, 18 Jan 2026 17:52:12 +0000 Eric Dumazet wrote:

quoted

On some platforms, GRO stack is too deep and causes cpu stalls.

Decreasing call depths by one shows a 1.5 % gain on Zen2 cpus.
(32 RX queues, 100Gbit NIC, RFS enabled, tcp_rr with 128 threads and 10,000 flows)

We can go further by inlining ipv6_gro_{receive,complete}
and take care of IPv4 if there is interest.

Note: two temporary __always_inline will be replaced with
      inline_for_performance when available.

v2: dealt with udp6_gro_receive()/udp6_gro_complete()
    missing declarations (kernel test robot [off-list ref])
    for CONFIG_MITIGATION_RETPOLINE=n

Still not good?

net/ipv6/udp_offload.c:136:17: error: static declaration of ‘udp6_gro_receive’ follows non-static declaration
  136 | struct sk_buff *udp6_gro_receive(struct list_head *head, struct sk_buff *skb)
      |                 ^~~~~~~~~~~~~~~~
In file included from net/ipv6/udp_offload.c:16:
./include/net/gro.h:408:17: note: previous declaration of ‘udp6_gro_receive’ with type ‘struct sk_buff *(struct list_head *, struct sk_buff *)’
  408 | struct sk_buff *udp6_gro_receive(struct list_head *, struct sk_buff *);
      |                 ^~~~~~~~~~~~~~~~
net/ipv6/udp_offload.c:168:29: error: static declaration of ‘udp6_gro_complete’ follows non-static declaration
  168 | INDIRECT_CALLABLE_SCOPE int udp6_gro_complete(struct sk_buff *skb, int nhoff)
      |                             ^~~~~~~~~~~~~~~~~
./include/net/gro.h:409:5: note: previous declaration of ‘udp6_gro_complete’ with type ‘int(struct sk_buff *, int)’
  409 | int udp6_gro_complete(struct sk_buff *, int);
      |     ^~~~~~~~~~~~~~~~~

Oh well, I thought I tested this stuff.

Interesting... clang (our default compiler for kernel) does not complain at all.