diff --git a/pkg/xdp/xdp.go b/pkg/xdp/xdp.go
index 3afc2eb6f..f3c7fc81e 100644
--- a/pkg/xdp/xdp.go
+++ b/pkg/xdp/xdp.go
@@ -202,11 +202,15 @@ func ReadOnlyFromSocket(sockfd int, ifaceIdx, queueID uint32, opts ReadOnlySocke
 	rxQueue.init(off, opts)
 
 	addr := unix.SockaddrXDP{
+		// XDP_USE_NEED_WAKEUP lets the driver sleep if there is no
+		// work to do. It will need to be woken by poll. It is expected
+		// that this improves performance by preventing the driver from
+		// burning cycles.
+		//
 		// By not setting either XDP_COPY or XDP_ZEROCOPY, we instruct
 		// the kernel to use zerocopy if available and then fallback to
 		// copy mode.
-		// TODO(b/240191988): Look into whether to use XDP_USE_NEED_WAKEUP.
-		Flags:   0,
+		Flags:   unix.XDP_USE_NEED_WAKEUP,
 		Ifindex: ifaceIdx,
 		// AF_XDP sockets are per device RX queue, although multiple
 		// sockets on multiple queues (or devices) can share a single
diff --git a/tools/xdp/README.md b/tools/xdp/README.md
index 85dabd1e8..1eb8ed3b6 100644
--- a/tools/xdp/README.md
+++ b/tools/xdp/README.md
@@ -31,8 +31,11 @@ must also look at [libbpf itself][libbpf] to understand what's really going on.
 
 ## TODO
 
-Kernel version < 5.4 has some weird offsets behavior. Just don't run on those
-machines.
+-   Kernel version < 5.4 has some weird offsets behavior. Just don't run on
+    those machines.
+-   Implement SHARED, although it looks like we usually run with only 1
+    dispatcher.
+-   Add a -redirect $fromdev $todev option in order to test fast path.
 
 [af_xdp_tutorial]: https://github.com/xdp-project/xdp-tutorial/tree/master/advanced03-AF_XDP
 [libbpf]: https://github.com/torvalds/linux/tree/master/tools/testing/selftests/bpf/xsk.c