xsk: Fix l2fwd for copy mode + busy poll combo

While checking AF_XDP copy mode combined with busy poll, strange
results were observed. rxdrop and txonly scenarios worked fine, but
l2fwd broke immediately.

After a deeper look, it turned out that for l2fwd, Tx side was exiting
early due to xsk_no_wakeup() returning true and in the end
xsk_generic_xmit() was never called. Note that AF_XDP Tx in copy mode
is syscall steered, so the current behavior is broken.

Txonly scenario only worked due to the fact that
sk_mark_napi_id_once_xdp() was never called - since Rx side is not in
the picture for this case and mentioned function is called in
xsk_rcv_check(), sk::sk_napi_id was never set, which in turn meant that
xsk_no_wakeup() was returning false (see the sk->sk_napi_id >=
MIN_NAPI_ID check in there).

To fix this, prefer busy poll in xsk_sendmsg() only when zero copy is
enabled on a given AF_XDP socket. By doing so, busy poll in copy mode
would not exit early on Tx side and eventually xsk_generic_xmit() will
be called.

Fixes: a073195 ("xsk: Add busy-poll support for {recv,send}msg()")
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Link: https://lore.kernel.org/bpf/20220406155804.434493-1-maciej.fijalkowski@intel.com

net/xdp/xsk.c

@@ -639,7 +639,7 @@ static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len
 	if (sk_can_busy_loop(sk))
 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 
-	if (xsk_no_wakeup(sk))
+	if (xs->zc && xsk_no_wakeup(sk))
 		return 0;
 
 	pool = xs->pool;