Hi stable team,
Could you please backport [1] to linux-5.10.y?
I noticed a regression caused by [2], which was merged to linux-5.10.y since v5.10.80.
After sock_map_unhash() helper was removed in [2], sock elems added to the bpf sock map
via sock_hash_update_common() cannot be removed if they are in the icsk_accept_queue
of the listener sock. Since they have not been accept()ed, they cannot be removed via
sock_map_close()->sock_map_remove_links() either.
It can be reproduced in network test with short-lived connections. If the server is
stopped during the test, there is a probability that some sock elems will remain in
the bpf sock map.
And with [1], the sock_map_destroy() helper is introduced to invoke sock_map_remove_links()
when inet_csk_listen_stop()->inet_child_forget()->inet_csk_destroy_sock(), to remove the
sock elems from the bpf sock map in such situation.
[1] d8616ee2affc ("bpf, sockmap: Fix sk->sk_forward_alloc warn_on in sk_stream_kill_queues")
(link: https://lore.kernel.org/all/20220524075311.649153-1-wangyufen@xxxxxxxxxx/)
[2] 8b5c98a67c1b ("bpf, sockmap: Remove unhash handler for BPF sockmap usage")
(link: https://lore.kernel.org/all/20211103204736.248403-3-john.fastabend@xxxxxxxxx/)
Thanks!
Wen Gu
