udp: fix busy polling
authorEric Dumazet <edumazet@google.com>
Thu, 18 Jan 2024 20:17:49 +0000 (20:17 +0000)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Thu, 1 Feb 2024 00:19:01 +0000 (16:19 -0800)
commitef8ad3071dface0c878d6de4cd8fa298a48b028c
tree4df034e5571d1e8d21b06eec8d94241cf5f749a7
parentf1f34a515fb1e25e85dee94f781e7869ae351fb8
udp: fix busy polling

[ Upstream commit a54d51fb2dfb846aedf3751af501e9688db447f5 ]

Generic sk_busy_loop_end() only looks at sk->sk_receive_queue
for presence of packets.

Problem is that for UDP sockets after blamed commit, some packets
could be present in another queue: udp_sk(sk)->reader_queue

In some cases, a busy poller could spin until timeout expiration,
even if some packets are available in udp_sk(sk)->reader_queue.

v3: - make sk_busy_loop_end() nicer (Willem)

v2: - add a READ_ONCE(sk->sk_family) in sk_is_inet() to avoid KCSAN splats.
    - add a sk_is_inet() check in sk_is_udp() (Willem feedback)
    - add a sk_is_inet() check in sk_is_tcp().

Fixes: 2276f58ac589 ("udp: use a separate rx queue for packet reception")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
include/linux/skmsg.h
include/net/inet_sock.h
include/net/sock.h
net/core/sock.c