Hi,
I have a peculiar problem which I hope someone can shed some
light on.
I have the following setup:
Linux box running 2.6.8.1 kernel + my own multicast listener
program
waiting for packets.
I encounter packet loss in the order of about 10 packets per
1000
messages sent. (Remote node sending 1000 byte packets at every
100ms).
However, when I run TCPDUMP 3.9.8 on the listener box as well, I
get no
packet loss. This is repeatable.
running tcpdump on module that is sink for mcast packets seems
to cause
no loss of packets.
{
tcpdump dst host 231.10.11.1
tcpdump: verbose output suppressed, use -v or -vv for full
protocol
decode
listening on eth0, link-type EN10MB (Ethernet), capture size 68
bytes
20:16:08.417072 IP p0e0.localdomain > 231.10.11.1: igmp v2
report
231.10.11.1
20:16:08.500188 IP 192.168.57.11.12441 > 231.10.11.1.12441:
UDP, length
1112
//For 200 packets sent:
202 packets captured
202 packets received by filter
0 packets dropped by kernel
}
running tcpdump -p dst host 231.10.11.1 (promiscuous mode off)
cause 13
packets lost (987 received by tcpdump)
I have looked at the /proc/net/dev, and the muticast packet
count ties
up to the number of packets I sent from the remote node, but the
listener reports packets are lost when tcpdump is not running.
Hence my
assumption is that the eth driver is reading all the packets,
but a
higher level kernel process drops some packets when the
processor is
under load.
What does tcpdump do that fixes my problem of packets being
lost? i.e
What do I need to do to my application (setsockopt()? or
otherwise) that
is being done by tcpdump?
I setup my listener socket as O_NONBLOCK, with SO_RCVBUF set to
512000.
I then poll the socket with select() to receive data.
I tried to set the promiscuous mode option in my multicast
listener
program, but I still get packet loss.
Any help would be much appreciated.
Thanks
Kind Regards
A. Kadwa
-
This is the tcpdump-workers list.
Visit https://cod.sandelman.ca/ to unsubscribe.
