Folks,
While discussing port randomization (in the context of
https://www.ietf.org/archive/id/draft-ietf-ntp-port-randomization-06.txt
), it has been raised to us that some NAT devices do not translate the
source port if the source port is a privileged port (<1024).
Any clues/examples of this type of NATs?
Thanks!
Regards,
I believe all devices will translate a privileged ports, but it won't translate to the same number on the other side. It will translate to an unprivileged port. Is it what you meant or really there are some devices that will not translate at all a privileged port?
What are you trying to achieve?
Jean
Current gen Cisco ASA firewalls have logic so that if the connection from a private host originated from a privileged source port, the NAT translation to public IP also uses an unprivileged source port (not necessarily the same source port though).
I found out that this behavior can cause issues when you have devices on your network that implement older DNS libraries or configs using UDP 53 as a source and destination port for their DNS lookups. Occasionally the source port gets translated to one that ISC BIND servers have in a blocklist (chargen, echo, time, and a few others) and the query is ignored. As I recall, this behavior is hard coded so patching and recompiling BIND is required to work around it.
I forget what the older ASA behavior was. It may have been to leave the source port unchanged through the NAT process (I think this is what you mean by "not translated"). In that case the client doesn't implement source port randomization and the NAT doesn't "upgrade" the connection to a random source port so I don't really see it as an issue. Ideally the client would implement source port randomization itself so it would be using source ports within its ephemeral port range for outgoing connections.
--Blake
For Linux iptables SNAT (used with --to-source), the default is to change the packet as little as possible.
https://linux.die.net/man/8/iptables
"If no port range is specified, then source ports below 512 will be mapped to other ports below 512: those between 512 and 1023 inclusive will be mapped to ports below 1024, and other ports will be mapped to 1024 or above.
Where possible, no port alteration will occur."
So, if there are no “collisions”, the same src port will be used. If there are “collisions” (multiple flows with the same src port and dst IP/port), then another src port within its “range” will be used.
But it can be configured, for example, to use ports 1024-65535, in which case flows with src port < 1024 will endup using ports > 1024 after they are NATed.
https://datatracker.ietf.org/doc/html/rfc6335#section-6 is also a good reference.
Alvaro
Hi, Blake,
Thanks a lot for your comments! In-line....
Current gen Cisco ASA firewalls have logic so that if the connection
from a private host originated from a privileged source port, the
NAT
translation to public IP also uses an unprivileged source port (not
necessarily the same source port though).
Did you actaully mean "...also uses a *privileged port*"?
I found out that this behavior can cause issues when you have devices
on
your network that implement older DNS libraries or configs using UDP
53
as a source and destination port for their DNS lookups. Occasionally
the
source port gets translated to one that ISC BIND servers have in a
blocklist (chargen, echo, time, and a few others) and the query is
ignored. As I recall, this behavior is hard coded so patching and
recompiling BIND is required to work around it.I forget what the older ASA behavior was. It may have been to leave
the
source port unchanged through the NAT process (I think this is what
you
mean by "not translated"). In that case the client doesn't implement
source port randomization and the NAT doesn't "upgrade" the
connection
to a random source port so I don't really see it as an issue.
The issue would be that if the port is not translated, and multiple
systems in the internal real of the NAT try to use the same privileged
port (say, 123) simultaneously, things wouldn't work.
Thanks,
Hi, Jean,
I believe all devices will translate a privileged ports, but it won't
translate to the same number on the other side. It will translate to
an unprivileged port. Is it what you meant or really there are some
devices that will not translate at all a privileged port?
What has been reported to us is that some boxes do not translate the
src port if it's a privileged port.
IN such scenarios, NTP implementations that always use src port=123,
dst port=123 might be in trouble if there are multiple NTP clients
behind the same NAT device....
Thanks!
Regards,
Fernando Gont via NANOG <nanog@nanog.org> writes:
What has been reported to us is that some boxes do not translate the
src port if it's a privileged port.IN such scenarios, NTP implementations that always use src port=123,
dst port=123 might be in trouble if there are multiple NTP clients
behind the same NAT device....
This problem used to be very common for 500/udp. Ref
https://datatracker.ietf.org/doc/html/rfc3715#section-2.3
Bjørn
Hi, Bjørn,
Fernando Gont via NANOG <nanog@nanog.org> writes:
> What has been reported to us is that some boxes do not translate
> the
> src port if it's a privileged port.
>
> IN such scenarios, NTP implementations that always use src
> port=123,
> dst port=123 might be in trouble if there are multiple NTP clients
> behind the same NAT device....This problem used to be very common for 500/udp. Ref
rfc3715
THanks a lot for the link! -- this is indeed a good read. I'm curious
if there exists something similar for UDP/123?
FWIW, we have this IETF I-D on NTP port randomization:
https://datatracker.ietf.org/doc/html/draft-ietf-ntp-port-randomization-06
, which has this section on the same kind of behavior, but for the NTP
port:
---- cut here ----
3.4. Effect on NAT devices
Some NAT devices will not translate the source port of a packet when
a privileged port number is employed. In networks where such NAT
devices are employed, use of the NTP well-known port for the client
port will essentially limit the number of hosts that may successfully
employ NTP client implementations.
In the case of NAT devices that will translate the source port even
when a privileged port is employed, packets reaching the external
realm of the NAT will not employ the NTP well-known port as the local
port, since the local port will normally be translated by the NAT
device possibly, but not necessarily, with a random port.
---- cut here ----
So I'm trying to find some reference that documents such behavior for
the NTP case....
Thanks!
Regards,
Hi Fernando,
NTP sounds simple but it could be very complex when you dig deep down and/or get lost in details.
Here are 2 things to consider:
1. NTP clients can query NTP servers by using SRC UDP ports > 1024.
2. NTP servers cannot query/sync/communicate to another NTP server when using SRC UDP port > 1024.
In short, server to server wants SRC and DST UDP 123. The query and the response will be fully 123 when server to server.
In your case, it sounds like you want to achieve NTP server to NTP server, but you mention NTP clients behind NAT devices.
Because multiple clients behind the same NAT devices should work. Multiple NTP servers behind the same NAT and wanting to use other NTP server *should* not work.
Can you give us more details on what kind of communication you need here? From what I understand client to server should work just fine with any NAT devices.
Maybe you meant multiple NTP servers behind the same NAT to external NTP server?
Thanks
Jean
Hi, Jean,
Hi Fernando,
NTP sounds simple but it could be very complex when you dig deep down
and/or get lost in details.
Here are 2 things to consider:1. NTP clients can query NTP servers by using SRC UDP ports > 1024.
This is indeed the case we're addressing. The NTP spec mandates srt
port=123, even for client-to-server cases.
In your case, it sounds like you want to achieve NTP server to NTP
server, but you mention NTP clients behind NAT devices.
Nope. We simply recommend to randomize the source port for client-to-
server cases.
So in the quoted section we make the case that requiring src port=123
clients doesnt really make sense:
1) if the NAT translates the port, the server won-t see src 123 anyway
2) if the NAT doesn't translate the port, you won't be able to ahve
multiple NTP clients behind the same firewall.
Can you give us more details on what kind of communication you need
here? From what I understand client to server should work just fine
with any NAT devices.Maybe you meant multiple NTP servers behind the same NAT to external
NTP servers
Please let me know if what I wrote above clarifies our intent.
Thanks!
Regards,
Let's start with this example. When I click sync my clock in windows, this happened.
On the inside or Private side
08:15:07.434344 IP 192.168.254.205.123 > 13.86.101.172.123: NTPv3, Client, length 48
08:15:07.473681 IP 13.86.101.172.123 > 192.168.254.205.123: NTPv3, Server, length 48
You are indeed right that the client must use UDP port 123. Is the RFC saying must or should on the client SRC port? I'm not sure.
But, on the Public, this happened.
08:15:07.434381 IP 192.2XX.XXX.58291 > 13.86.101.172.123: NTPv3, Client, length 48
08:15:07.473656 IP 13.86.101.172.123 > 192.2XX.XXX.58291: NTPv3, Server, length 48
// Public ip obfuscated. I know, it indeed starts with 192.2. It's EBOX in Canada.
What we see on the public side, is that a network device did a NAT translation of the SRC UDP port to 58921. My clock synced perfectly.
So your goal is to find the devices that don't follow this behaviour, right?
Jean
Yes I did. Thanks. Not quite. If multiple devices behind a NAT use SRC=123 & DST = 123 for connections, their connections will still work most of the time. First, if the connections are to different destinations there would be no NAT conflict. Second, if the connections occur at different times, there would be no NAT conflict. Third, if there was a NAT conflict (meaning connections using the same SRC port, DST port, and DST host at the same time) the NAT device would see this and would either adjust the translation to use a different SRC port (in which case the connection succeeds) or may drop/reject the connection (in which case the client would eventually retry). So saying “things wouldn’t work” or “you won’t be able to have multiple NTP clients behind the same firewall” is not true. Saying that “the number of simultaneous connections is limited” when behind a NAT would be more accurate. . Most NTP clients utilize multiple NTP servers, often from a pool of available servers, and initiate connections rather infrequently so I do not expect this to be a problem in practice unless there are thousands of NTP clients behind a single NAT accessing a common NTP server (and that NAT does not do a good job of dealing with collisions). I do, however, agree that clients should probably use ephemeral ports when making any outbound connections as this provides more entropy for NAT as well as for connection security. This extends to NTP.
Hi, Jean,
Let's start with this example. When I click sync my clock in windows,
this happened.On the inside or Private side
08:15:07.434344 IP 192.168.254.205.123 > 13.86.101.172.123: NTPv3,
Client, length 48
08:15:07.473681 IP 13.86.101.172.123 > 192.168.254.205.123: NTPv3,
Server, length 48You are indeed right that the client must use UDP port 123. Is the
RFC saying must or should on the client SRC port? I'm not sure.
Section 9.1 ("Peer Process Variables") of [RFC5905] SAYS:
dstport: UDP port number of the client, ordinarily the NTP port
number PORT (123) assigned by the IANA. This becomes the source
port number in packets sent from this association.
That said, as noted in
https://datatracker.ietf.org/doc/html/draft-ietf-ntp-port-randomization-06#section-5
, other client implementations don't bind the local port to 123, and
hence assign an ephemeral port.
But, on the Public, this happened.
08:15:07.434381 IP 192.2XX.XXX.58291 > 13.86.101.172.123: NTPv3,
Client, length 48
08:15:07.473656 IP 13.86.101.172.123 > 192.2XX.XXX.58291: NTPv3,
Server, length 48// Public ip obfuscated. I know, it indeed starts with 192.2. It's
EBOX in Canada.What we see on the public side, is that a network device did a NAT
translation of the SRC UDP port to 58921. My clock synced perfectly.So your goal is to find the devices that don't follow this behaviour,
right?
No. The goal of our I-D is that NTP clients randomize their source
port -- there's no need for clients to use port 123, and using that
port on the client side has negative security implications.
Thanks,