]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commit
lwt: Disable BH too in run_lwt_bpf()
authorDongdong Wang <wangdongdong.6@bytedance.com>
Sat, 5 Dec 2020 07:59:45 +0000 (23:59 -0800)
committerAlexei Starovoitov <ast@kernel.org>
Mon, 7 Dec 2020 19:53:39 +0000 (11:53 -0800)
commitd9054a1ff585ba01029584ab730efc794603d68f
tree1423dfd779aad5f398dd1330a84e060306d1fa56
parent12c8a8ca117f3d734babc3fba131fdaa329d2163
lwt: Disable BH too in run_lwt_bpf()

The per-cpu bpf_redirect_info is shared among all skb_do_redirect()
and BPF redirect helpers. Callers on RX path are all in BH context,
disabling preemption is not sufficient to prevent BH interruption.

In production, we observed strange packet drops because of the race
condition between LWT xmit and TC ingress, and we verified this issue
is fixed after we disable BH.

Although this bug was technically introduced from the beginning, that
is commit 3a0af8fd61f9 ("bpf: BPF for lightweight tunnel infrastructure"),
at that time call_rcu() had to be call_rcu_bh() to match the RCU context.
So this patch may not work well before RCU flavor consolidation has been
completed around v5.0.

Update the comments above the code too, as call_rcu() is now BH friendly.

Signed-off-by: Dongdong Wang <wangdongdong.6@bytedance.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Reviewed-by: Cong Wang <cong.wang@bytedance.com>
Link: https://lore.kernel.org/bpf/20201205075946.497763-1-xiyou.wangcong@gmail.com
net/core/lwt_bpf.c