I’m curious what people are seeing these days on the UDP/123 policers in their networks.
I know while I was at NTT we rolled some out, and there are a number of variants that have occurred over the past 6-7 years. I’ve heard from people at the NTP Pool as well as having observed some issues with NTP at Akamai and time sync from time to time.
Are you still seeing a lot of NTP attacks in your flows these days?
Should we be looking to remove these, similar to how we did for SQL/Slammer after a time?
Yes, we still see lots of UDP amplification attacks using NTP monlist. We use a filter to block UDP src 123 packets of 468 bytes in length (monlist reply with the max 6 IPs).
Yes, we still see lots of UDP amplification attacks using NTP monlist. We use a filter to block UDP src 123 packets of 468 bytes in length (monlist reply with the max 6 IPs).
-Rich
+1 , still see, still have policers
Fyi, ipv6 ntp / udp tends to have a much higher success rate getting through cgn / policers / …
The various NTP filters (rate limits, packet size limits) are negatively affecting the NTP Pool, the new secure NTP protocol (Network Time Security) and other clients. NTP filters were deployed several years ago to solve serious DDoS issues, I’m not second guessing those decisions. Changing the filters to instead block NTP mode 7, which cover monlist and other diagnostics, would improve NTP usability.
The various NTP filters (rate limits, packet size limits) are negatively affecting the NTP Pool, the new secure NTP protocol (Network Time Security) and other clients. NTP filters were deployed several years ago to solve serious DDoS issues, I’m not second guessing those decisions. Changing the filters to instead block NTP mode 7, which cover monlist and other diagnostics, would improve NTP usability.
On many edge routers from Juniper, Nokia and Cisco you can create
offset based bit-matches. I'm NTP illiterate, but isn't NTP mode in
fixed offset after UDP header? So it should be possible to do this in
hardware at linerate just the same as matching UDP port on typical
edge router.
I’ve advocated a throttle (not a hard block) on udp/123 packets with 468 Bytes/packet (the size of a full monlist response). In your paper you mention NTS extensions can be 200+ bytes. How large do those packets typically get, in practice? And how significant is packet loss for them (if there’s high packet loss during the occasional attack, does that pose a problem)?
The various NTP filters (rate limits, packet size limits) are
negatively affecting the NTP Pool, the new secure NTP protocol
(Network Time Security) and other clients. NTP filters were
deployed several years ago to solve serious DDoS issues, I'm not
second guessing those decisions. Changing the filters to instead
block NTP mode 7, which cover monlist and other diagnostics, would
improve NTP usability.
I've advocated a throttle (not a hard block) on udp/123 packets with 468
Bytes/packet (the size of a full monlist response). In your paper you
mention NTS extensions can be 200+ bytes. How large do those packets
typically get, in practice? And how significant is packet loss for them
(if there's high packet loss during the occasional attack, does that
pose a problem)?
I expect to see NTP UDP packets that would approach the MTU limit, in
some cases.
If a packet is "too big" for some pathway, then are we talking about a
fractional packet loss or are we talking about 100% packet loss (dropped
mid-way due to size)?
NTS is initialized using a relatively expensive, but short lived, TCP TLS session. NTP loss due to rate limiting will require more frequent TCP initializations.
The NTP size-blocks I’ve observed have been hard, not rate limits. Martin Langer provided a table showing sizes between 228 and 1468 bytes depending on the encryption algorithm and number of cookies. I’ve asked the NTS authors about potential workarounds, but suspect it would be difficult. The authors have also confirmed that size blocks have been detected during NTS tests.
The secure time transfer of NTS was designed to avoid amplification attacks. Impairment of UDP port 123 could require use of alternate UDP port(s), which might be unpopular.
I implement it as a throttle… but that effectively means “0% packet loss most of the time, and significant packet loss when there’s a large attack”. Doing it as a complete block (which Sommars showed some networks do) is unnecessary.
So my question is whether occasional bursts of hard-block are a problem. If you only need large packets during the initialization phase, but the ongoing communication is done with smaller packets, then my approach may be acceptable, and we just need to convince other network operators to throttle rather than hard-block the large packets. But if you need large packets all the time, and occasional breakage of large packets causes problems, then I’ll need to re-think my approach.