Le mardi 26 octobre 2010 à 15:53 -0400, Jeremy Jackson a écrit :

quoted

Trying to find source of packet loss on an 8node compute cluster, we
find:
(not in this example, but on the real cluster)

in /proc/sys/net/snmp
Udp: InDatagrams NoPorts InErrors OutDatagrams RcvbufErrors SndbufErrors
Udp: 976460 1750 0 986795 0 0

InErrors *and* RcvbufErrors both go up with full socket buffer, this has
made troubleshooting our application more difficult.  We were chasing
UDP
checksum problems, until we checked linux source code, and found
aliasing.

Is this done for assembly code efficiency?  Any reason ENOMEM (ie socket
buffer full) can't avoid aliasing to UDP checksum failed errors?

in linux-source-2.6.32/net/ipv4/udp.c:__udp_queue_rcv_skb()
....
                /* Note that an ENOMEM error is charged twice */
                if (rc == -ENOMEM) {
                        UDP_INC_STATS_BH(sock_net(sk),
UDP_MIB_RCVBUFERRORS,
                                         is_udplite);
                        atomic_inc(&sk->sk_drops);
                }
                goto drop;
...
drop:
        UDP_INC_STATS_BH(sock_net(sk), UDP_MIB_INERRORS, is_udplite);

In MIBS, there is no counter for UDP checksum errors

So we use the standard UDP_MIB_INERRORS

Yes, this part I understand, but what I don't understand is why ENOMEM
errors *and* checksum errors both use the same counter, while ENOMEM has
it's own already.

udpInErrors OBJECT-TYPE
    SYNTAX     Counter32
    MAX-ACCESS read-only
    STATUS     current
    DESCRIPTION
           "The number of received UDP datagrams that could not be
            delivered for reasons other than the lack of an
            application at the destination port.


We could add a LINUX specific MIB entry, eventually...