recv.2: With pending 0-length datagram read() and recv() with flags == 0 differ

Quoting Tom:

The statement "With a zero flags argument, recv() is equivalent to
read(2)." is not correct. In the case of passing a UDP socket an
empty buffer the two calls differ. read(2) will be a noop (as it
rightly says in its manpage), but recv(2) will discard the
packet.

We ran into this in networkd, as we use FIONREAD to determine the
buffer size (and allocate the right buffer), so in case someone
passed us an empty packet we would end up in a busy loop when we
were using read(2).  Changing to recv(2) fixed the issue
[https://github.com/systemd/systemd/pull/3299].

Reported-by: Tom Gundersen <teg@jklm.no>
Signed-off-by: Michael Kerrisk <mtk.manpages@gmail.com>

man2/recv.2

@@ -78,8 +78,9 @@ With a zero
 .I flags
 argument,
 .BR recv ()
-is equivalent to
-.BR read (2).
+is generally equivalent to
+.BR read (2)
+(but see NOTES).
 Also, the following call
 
     recv(sockfd, buf, len, flags);
@@ -515,6 +516,19 @@ and
 .B MSG_WAITALL
 flags.
 .SH NOTES
+If a zero-length datagram is pending,
+.BR read (2)
+and
+.BR recv ()
+with a
+.I flags
+argument of zero provide different behavior.
+In this circumstance,
+.BR read (2)
+has no effect (the datagram remains pending), while
+.BR recv ()
+consumes the pending datagram.
+
 The
 .I socklen_t
 type was invented by POSIX.