sock_udp_send() reports success even though the message was never sent

[benpicco]

Description

When sending the first message to a host with sock_udp_send(), the function will report the number of byes sent, indicating success. But in reality the packet never reached the device driver - it was silently dropped by the IPv6 layer with no error being propagated. (unless you ENABLE_DEBUG is set in gnrc_ipv6.c - then a debug message is printed)

Steps to reproduce the issue

• configure a custom, non-link-local IP address:

    ipv6_addr_t addr;
    gnrc_netif_t *netif = gnrc_netif_iter(NULL);
    gnrc_netif_ipv6_addrs_get(netif, &addr, sizeof(addr));

    addr.u8[0] = 0xde;
    addr.u8[1] = 0x71;
    addr.u8[2] = 0xc1;
    addr.u8[3] = 0x05;

    gnrc_netif_ipv6_addr_add(netif, &addr, 64, 0);

• send a message to the remote

    sock_udp_ep_t remote = { .family = AF_INET6, .port = port };
    ipv6_addr_from_str((ipv6_addr_t *)&remote.addr, "de71:c105::1");
    int res = sock_udp_send(NULL, "Hello", 6, &remote);

Expected results

Either res < 0 and the message failed to be send, or res == 6 and the message is being sent.

Actual results

res == 6, the message is not being sent.

When tracing the packet through the network stack it becomes evident that it is dropped in _send_unicast(): no link-layer address or interface for next hop to de71:c105::1.

The Address is subsequently resolved:

ipv6: waiting for incoming message.
ipv6: GNRC_NETAPI_MSG_TYPE_RCV received
ipv6: Received (src = de71:c105::1, dst = de71:c105::8d1b:3202:ae20:bb66, next header = 58, length = 40)
ipv6: forward nh = 58 to other threads
ipv6: handle ICMPv6 packet (nh = 58)

However, _send_unicast does not return an error and the GNRC_NETAPI_MSG_TYPE_SND case will not report an error to the upper layer.

My current hack around this is to send a dummy message after startup, but this is of course not a very nice solution.

Versions

RIOT 2019.04

[miri64]

Due to the asynchronous nature of GNRC this is deactivated by default. There is however the poorly utilized, poorly tested and poorly documented gnrc_neterr module which could provide such error reporting for sock. At least most releases in IPv6 are using gnrc_pktbuf_release_error() to signal an error.

[PeterKietzmann]

I'm wondering if we should consider this issue as 'solved' or feature request. Personally I've never used the gnrc_neterr. @benpicco did you give it a try?

[benpicco]

At the very least the documentation should be updated.
The way I read the description of sock_udp_send, when size is returned, those bytes have been successfully written to the network/Aether/left the device in the intended way.

For the case that the destination is unknown and the packet was therefore dropped without being written to the network, I would have expected to get -EHOSTUNREACH.

[benpicco]

something like this?

--- a/sys/include/net/sock/udp.h
+++ b/sys/include/net/sock/udp.h
@@ -419,7 +419,10 @@ ssize_t sock_udp_recv(sock_udp_t *sock, void *data, size_t max_len,
  *                      end point of @p sock provides this information.
  *                      sock_udp_ep_t::port may not be 0.
  *
- * @return  The number of bytes sent on success.
+ * @return  The number of bytes that might have been sent.
+ *          A return value > 0 is not an indication of success.
+ *          To make sure the message has been send, check @ref net_gnrc_neterr
+ *          for the absence of errors.
  * @return  -EADDRINUSE, if `sock` has no local end-point or was `NULL` and the
  *          pool of available ephemeral ports is depleted.
  * @return  -EAFNOSUPPORT, if `remote != NULL` and sock_udp_ep_t::family of

[miri64]

The first two sentences are good, the third assumes GNRC as the backend, which in the general doc is a no-go.