Hi Julian,
On Mon, 20 Apr 2026 at 20:02, Justin Iurman justin.iurman@gmail.com wrote:
On 4/17/26 07:54, David Carlier wrote:
gtp_genl_send_echo_req() runs as a generic netlink doit handler in process context with BH not disabled. It calls udp_tunnel_xmit_skb(), which eventually invokes iptunnel_xmit() — that uses __this_cpu_inc/dec on softnet_data.xmit.recursion to track the tunnel xmit recursion level.
Without local_bh_disable(), the task may migrate between dev_xmit_recursion_inc() and dev_xmit_recursion_dec(), breaking the per-CPU counter pairing. The result is stale or negative recursion levels that can later produce false-positive SKB_DROP_REASON_RECURSION_LIMIT drops on either CPU.
The other udp_tunnel_xmit_skb() call sites in gtp.c are unaffected: the data path runs under ndo_start_xmit and the echo response handlers run from the UDP encap rx softirq, both with BH already disabled.
Fix it by disabling BH around the udp_tunnel_xmit_skb() call, mirroring commit 2cd7e6971fc2 ("sctp: disable BH before calling udp_tunnel_xmit_skb()").
Why not fix iptunnel_xmit() directly, rather than fixing all possible callers? Basically, jut like we did for lwtunnel_{output|xmit}(). The advantage would be that we no longer have to worry about BHs in the callers, and BHs would only be disabled when necessary.
Good point — your lwtunnel fix (c03a49f3093a) is a close parallel, and a central fix would avoid chasing callers one by one (sctp was patched last week, gtp is this one, and tipc/wireguard/ovpn genl paths look similar).
Happy to respin as v2 with local_bh_disable/enable moved into iptunnel_xmit() (and ip6tunnel_xmit() for symmetry), and drop the gtp-local hunk. That would also supersede Xin Long's recent sctp commit (2cd7e6971fc2), so I'll make sure to Cc him.
One thing I'd like your take on before I send: iptunnel_xmit() feels like the natural home since it owns the recursion counter, but would you rather see it in udp_tunnel_xmit_skb()? I don't want to pick the wrong spot if you already have a preference.
Cheers !