Reroot exposes SSH, Telnet, Web UI to WAN

SteveITS

@user1337 That’s a feature? I wonder if that’s something Netgate can disable in the driver/config. You might open a Redmine.pfsense.org about it.

Side note:there’s a long thread here somewhere about (some?) Chelsio cards downloading at half speed in 23.01.

stephenw10

Hmm, so this was eliminated by disabling TOE or rather simply not enabling it?

What pfSense version did you see this in?

You only saw it after rerooting?

This was permanent after the reroot? Not just while the reroot was happening?

Steve

User1337

@steveits TOE is disabled by default and can't be enabled by accident so nothing Netgate needs to worry about. My WAN speed matches my expectations when using typical web-based speed tests, but I only have a ~350/10Mbps plan so it's not enough to expose a large bottleneck.

@stephenw10 Yes, I disabled TOE by commenting out these lines in a custom script in /usr/local/etc/rc.d:

#ifconfig cxl0 toe
#ifconfig cxl1 toe

I enabled TOE on LAN manually via SSH ifconfig cxl1 toe and indeed it works, 1.15 vs .23 TIME+ shown in htop for each respective iperf2 server process. No difference with iperf3. Interestingly enough pfblockerng also works, which means the system firewall is taking effect. This all may be due to this detail from the same calomel link:

TCP TIMING: We noticed that short lived connections of less then 0.6 seconds will NOT use the Chelsio TCP Offload Engine (TOE) even if TOE is allowed universally or through Chelsio Offload Policy (COP). Not sure of the reason.

Next I tried enabling TOE on WAN ifconfig cxl0 toe and noticed SSH and web UI don't become exposed to WAN until after reroot. Version CE 2.6.0 (22.01). Yes it's permanently exposed after reroot. At this point it's not a real issue but may be of help to anyone else that goes down this path.

stephenw10

Yes, that's a nice catch. It should definitely be documented.

Did you try both 2.6 and 22.01 or just noted they are closely equivalent?

User1337

@stephenw10 Just 2.6.0.

jimp

Given the description of how TOE works I would have expected it to always bypass pf and not just at reroot. I'll try to work a working against using that into the docs but I'm not sure where it might fit that users would see it. The most likely place would seem to be https://docs.netgate.com/pfsense/en/latest/hardware/tune.html#chelsio-cxgbe-4-cards

In the future, if you believe you have discovered a security vulnerability, please report it privately as described on https://www.netgate.com/security and not on a public forum post. That way we can investigate it and work on a fix before it is widely known to the public.

jimp

Also I wonder how this works then:

https://redmine.pfsense.org/issues/9091

Maybe it's different on the T4 vs T5 cards? Or maybe you need that module loaded so TOE works as expected.

User1337

@jimp Yes, I also load t4_tom as part of the script mentioned above. As for why those services become exposed after reroot, I'm not sure, perhaps something to do with the startup sequence and that 600ms exception. The real issue is there's no mention of TOE bypassing pf in Chelsio's FreeBSD manual. However obvious it is to most, it's quite the gotcha for someone like me and unfortunately resulted in creating this thread. A lot of things are poorly or not documented at all.

jimp

Yeah that is definitely unexpected, and also very inconsistent.

Were you seeing any actual measurable performance gains from TOE for traffic passing through the firewall?

Usually for connection-based offloading like that (e.g. TSO, LRO) it primarily benefits when acting as an endpoint and can degrade performance when acting as a router.

I'm still trying to figure out how best to warn about this in the docs because I'm wondering if there is any time we should even recommend enabling that, not just warning against it.

User1337

@jimp fetch http://ping.online.net/1000Mo.dat ~1.12s, but after enabling TOE on cxl0 the download speed is severely kneecapped for the firewall itself, from my PC the speed is as expected. Reroot made no difference in regards to speed.

Gertjan

@User1337 said in Reroot exposes SSH, Telnet, Web UI to WAN:

.... on cxl0 the download speed is severely kneecapped for the firewall itself

Totally normal.
A firewall is optimized for packet handling.
Downloading a file from the firewall command line make packets enter into user land, to be stored on some drive afterwards, that takes extra time.

You just discovered why 'speed tests should be done on LAN devices, not the firewall itself'.

@User1337 said in Reroot exposes SSH, Telnet, Web UI to WAN:

from my PC the speed is as expected.

Point proven.

jimp

@User1337 said in Reroot exposes SSH, Telnet, Web UI to WAN:

@jimp fetch http://ping.online.net/1000Mo.dat ~1.12s, but after enabling TOE on cxl0 the download speed is severely kneecapped for the firewall itself, from my PC the speed is as expected. Reroot made no difference in regards to speed.

That's just one part of it. Is it better/worse with or without TOE in each of those cases?

Just one single stream downloading a large file isn't a great test, but not everyone has a setup capable of testing things in multiple ways (different numbers of traffic streams, packet sizes, etc.)

jimp

I went ahead and added a warning in the docs:

https://docs.netgate.com/pfsense/en/latest/hardware/tune.html#chelsio-tcp-offload-engine-toe

I can always refine it from there but I'd say for now it doesn't look like something anyone should be running in a firewall role. That kind of offloading is meant for initiating/terminating connections on the device in question, not for passing through traffic.

Someone might want to do that if they're doing something like using pfSense as a GUI for HAProxy or Squid internally (not on an edge).

User1337

@jimp I gave up, the download speed was around 2Mbps with TOE enabled.