Path MTU daemon - broadcast lost ICMP packets on ECMP networks
C Makefile Python
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
deps nflog: dependencies Feb 19, 2015
src Only capture incoming packets Feb 2, 2016
tests
.clang-format
.gitignore Cosmetic: ignore debs during make Feb 19, 2015
.gitmodules nflog: dependencies Feb 19, 2015
LICENSE-BSD-CloudFlare Release v0.1 Feb 4, 2015
Makefile
README.md nflog: make sure the messages aren't buffered for default 1 second Feb 19, 2015
RELEASE_NOTES

README.md

Path MTU daemon

With ECMP enabled the ICMP messages are routed mostly to wrong server. To fix that let's broadcast the ICMP messages that we think are worth it to every machine in colo. Some reading:

$ ./pmtud --help

Usage:

    pmtud [options]

Path MTU Daemon is captures and broadcasts ICMP messages related to
MTU detection. It listens on an interface, waiting for ICMP messages
(IPv4 type 3 code 4 or IPv6 type 2 code 0) and it forwards them
verbatim to the broadcast ethernet address.

Options:

  --iface              Network interface to listen on
  --src-rate           Pps limit from single source (default=1.0 pss)
  --iface-rate         Pps limit to send on a single interface (default=10.0 pps)
  --verbose            Print forwarded packets on screen
  --dry-run            Don't inject packets, just dry run
  --cpu                Pin to particular cpu
  --ports              Forward only ICMP packets with payload
                       containing L4 source port on this list
                       (comma separated)
  --help               Print this message

Example:

    pmtud --iface=eth2 --src-rate=1.0 --iface-rate=10.0

Once again, it listens waiting for packets matching:

((icmp and icmp[0] == 3 and icmp[1] == 4) or
  (icmp6 and ip6[40+0] == 2 and ip6[40+1] == 0)) and
 (ether dst not ff:ff:ff:ff:ff:ff)

And having appropriate length, and forwards them to ethernet broadcast ff:ff:ff:ff:ff:ff.

To debug use tcpdump:

sudo tcpdump -s0 -e -ni eth0 '((icmp and icmp[0] == 3 and icmp[1] == 4) or
                               (icmp6 and ip6[40+0] == 2 and ip6[40+1] == 0))'

To build type:

git submodule update --init --recursive
make

To test run it in dry-run and verbose mode:

sudo ./pmtud --iface=eth0 --dry-run -v -v -v

If you want to use NFLOG interface:

iptables -I INPUT -i lo -p icmp -m icmp --icmp-type 3/4 --j NFLOG --nflog-group 33
ip6tables -I INPUT -i lo -p icmpv6 -m icmpv6 --icmpv6-type 2/0 -j NFLOG --nflog-group 33

You can add -m pkttype ! --pkt-type broadcast to be even more specific. Then to use the NFLOG api run:

sudo ./pmtud --iface=eth0 --dry-run -v -v -v --nflog 33

This will cause pmtud to listen to packets from NFLOG and use eth0 to brodcast them if neccesary. Debug by listing this /proc file:

cat /proc/net/netfilter/nfnetlink_log
33  32781     0 2 65535      0  1

Where columns read:

  • nflog group number of a given queue (16 bits)
  • peer portid: most likely the pid of process
  • number of messages buffered on the kernel side
  • copy mode: 2 for full packet copy
  • copy range: max packet size
  • flush timeout in 1/100th of a second
  • use count