Updating to pfSense+ 24.3 breaks routing - kernel routes now gone
-
@marcosm Thanks. 9.0.3 is still fine. Straight after boot up:
[23.09.1-RELEASE][root@GTpfsense01.<<hidden>>]/root: vtysh Hello, this is FRRouting (version 9.0.3). Copyright 1996-2005 Kunihiro Ishiguro, et al. GTpfsense01.<<hidden>># show ip route Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF, I - IS-IS, B - BGP, E - EIGRP, T - Table, v - VNC, V - VNC-Direct, A - Babel, f - OpenFabric, > - selected route, * - FIB route, q - queued, r - rejected, b - backup t - trapped, o - offload failure K>* 0.0.0.0/0 [0/0] via <<hidden>>, vmx2, 00:00:53 C>* 10.27.10.0/24 [0/1] is directly connected, vmx1.10, 00:00:53 C>* 10.27.194.0/24 [0/1] is directly connected, ovpns1, 00:00:53 C>* 10.30.20.0/24 [0/1] is directly connected, vmx1.20, 00:00:53 C>* 10.254.40.0/28 [0/1] is directly connected, vmx1.40, 00:00:53 C>* 10.254.100.0/24 [0/1] is directly connected, vmx1.100, 00:00:53 C>* 10.255.195.2/32 [0/1] is directly connected, ovpns2, 00:00:53 C>* 10.255.196.2/32 [0/1] is directly connected, ovpns3, 00:00:53 C>* 10.255.197.2/32 [0/1] is directly connected, ovpns4, 00:00:53 K>* <<hidden>>/32 [0/0] via <<hidden>>, vmx2, 00:00:53 C>* <<hidden>>/29 [0/1] is directly connected, vmx2, 00:00:53 C>* <<hidden>>/22 [0/1] is directly connected, vmx3, 00:00:53 C>* 172.16.27.1/32 [0/1] is directly connected, lo0, 00:00:53 C>* 192.168.57.0/24 [0/1] is directly connected, vmx0, 00:00:53 K>* 203.12.160.35/32 [0/0] via <<hidden>>, vmx2, 00:00:53 K>* 203.12.160.36/32 [0/0] via <<hidden>>, vmx2, 00:00:53 GTpfsense01.<<hidden>>#This was about a minute after boot up. No issues with K routes. All 4 expected ones - most crucially the default - are all there.
So something in the 9.1.x series broke K routes on FreeBSD14/15 when starting the FRR service. If you "tickle" things when the service is running - like a remote interface shutdown and unshut, or config change and change-back. The K routes can be coaxed out, but obviously - this is not workable/practical.
https://frrouting.org/release/9.1/
"FRR 9.1 brings a long list of enhancements and fixes with 941 commits from 73 developers."I scanned the CI
https://ci1.netdef.org/browse/FRR-FRR121/Then scanned the tests for FreeBSD
https://ci1.netdef.org/browse/FRR-FRR121-FBSD14AMD-101/testThat all seems to be just BGP specific. There doesn't seem to be any CI tests specifically for FreeBSD and this functionality of K routes. No wonder regressions come in like this - if no one is testing for it. Geez, what a nightmare - trying to find out which of the 941 commits to 9.1 broke it on FreeBSD.
Maybe Alexander Skorichenko askorichenko@netgate.com can provide some input, as he signed off on one of the changes backported to 9.1 https://ci1.netdef.org/browse/FRR-FRR121-54
Do you think it could be related to my NIC type? I am using VMware vmxnet3 for both production and lab. I can rebuild in my GNS3 lab as igbX NICs to see if that changes anything. Mind you in production I am tied to vmxnet3 as anything other than the paravirtualized vmxnet3 NICs give comparatively poor performance (the alternative being e1000e but that does not perform well at scale). So it would only be for information gathering. vmxnet3 works wonderfully for FRR 9.0.x shouldn't shouldn't have to change virtualized NIC types because someone broke vmxnet3. But I'll test anyway and see what I get.
-
Thank you for testing. Let's try our luck with frr10.1 then. We can determine the next steps after that.
-
@marcosm Thank you so much for your package builds.
9.0.3 is on the left, and 10.1 is on the right:
In 10.1, these L routes (local) show up. First time I've seen those in FRR. They are some interface IPs. A few OSPF routes show up in that as well. Unfortunately no K routes.
The "tickle techniques" I have mentioned previously can still be used to coax the K routes back.
My current suspicions are that it's something to do with something changing from 9.0.x to 9.1 in relation to vmxnet3 NICs. I tried rebuilding to e1000e "emX" NICs but that didn't go so well - I remapped and that took a long time to process - about half an hour - and then after reboot, things seem to get stuck on loading VLAN interfaces.
I wouldn't want to have to change NIC types in production. Especially when vmxnet3 should work fine, and is the recommended type for VMware/ESXi https://docs.netgate.com/pfsense/en/latest/recipes/virtualize-esxi.html
I have always had Hardware TSO and LRO disabled. I tried also disabling hardware checksum offload but that didn't help.
EDIT: In the same lab I have a pfSense 2.7.2 CE box emulating Intel 82545EM NIC (similar to e1000e) and I upgraded that to FRR 10.1, and the kernel routes were still there afterwards. I am suspecting that FRR from 9.1 onwards doesn't play well with vmxnet3 paravirtualized NICs on FreeBSD14/15.
-
I wouldn't think the interface driver would matter in this case, but I suppose it's possible. Anyway, we'll try to get FRR10 in for 24.08. Given that testing shows this is more likely to be an issue with the package, I suggest looking through the upstream issues and opening an issue report there (and link it on the redmine).
For now, I will attach add frr9.0.3 to the redmine to serve as a temporary workaround for those that need it.
-
K keyser referenced this topic on
-
@marcosm I did some more testing, and you were right in thinking that the interface driver wouldn't matter in this case. I wanted confirmation though, so I made a completely separate GNS3 lab so that I wouldn't have to keep blacking out my WAN addresses and FQDN's. This is the lab:
The clouds are just so I can get HTTP access to the devices, and cloud1 also provides Internet to the lab, for the initial FRR package installs.
Both firewalls are pfSense CE 2.7.2 fresh installs.
They only differ in hardware for NIC types.
Firewall 1 emulates Intel 82576 (igb#)
Firewall 2 emulates VMware vmxnet3 (vmx#)I checked the route tables in FRR, when running the default FRR 9.0.2. and there was a single K route for both. I would assume that FRR 9.0.3 also has this, given my previous testing:
[2.7.2-RELEASE][root@firewall1.home.arpa]/root: vtysh Hello, this is FRRouting (version 9.0.2). Copyright 1996-2005 Kunihiro Ishiguro, et al. firewall1.home.arpa# show ip route Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF, I - IS-IS, B - BGP, E - EIGRP, T - Table, v - VNC, V - VNC-Direct, A - Babel, f - OpenFabric, > - selected route, * - FIB route, q - queued, r - rejected, b - backup t - trapped, o - offload failure K>* 0.0.0.0/0 [0/0] via 192.0.2.1, igb0, 00:00:21 C>* 192.0.2.0/29 [0/1] is directly connected, igb0, 00:00:21 C>* 192.168.41.0/24 [0/1] is directly connected, igb1, 00:00:21 C>* 192.168.57.0/24 [0/1] is directly connected, igb5, 00:00:21 firewall1.home.arpa# [2.7.2-RELEASE][root@firewall2.home.arpa]/root: vtysh Hello, this is FRRouting (version 9.0.2). Copyright 1996-2005 Kunihiro Ishiguro, et al. firewall2.home.arpa# show ip route Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF, I - IS-IS, B - BGP, E - EIGRP, T - Table, v - VNC, V - VNC-Direct, A - Babel, f - OpenFabric, > - selected route, * - FIB route, q - queued, r - rejected, b - backup t - trapped, o - offload failure K>* 0.0.0.0/0 [0/0] via 192.0.2.9, vmx0, 00:00:36 C>* 192.0.2.8/29 [0/1] is directly connected, vmx0, 00:00:36 C>* 192.168.42.0/24 [0/1] is directly connected, vmx1, 00:00:36 C>* 192.168.57.0/24 [0/1] is directly connected, vmx5, 00:00:36 firewall2.home.arpa#I then stopped FRR, and installed FRR 9.1.1. Results:
[2.7.2-RELEASE][root@firewall1.home.arpa]/tmp: vtysh Hello, this is FRRouting (version 9.1.1). Copyright 1996-2005 Kunihiro Ishiguro, et al. firewall1.home.arpa# show ip route Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF, I - IS-IS, B - BGP, E - EIGRP, T - Table, v - VNC, V - VNC-Direct, A - Babel, f - OpenFabric, > - selected route, * - FIB route, q - queued, r - rejected, b - backup t - trapped, o - offload failure C>* 192.0.2.0/29 [0/1] is directly connected, igb0, 00:00:10 C>* 192.168.41.0/24 [0/1] is directly connected, igb1, 00:00:10 C>* 192.168.57.0/24 [0/1] is directly connected, igb5, 00:00:10 firewall1.home.arpa# [2.7.2-RELEASE][root@firewall2.home.arpa]/tmp: vtysh Hello, this is FRRouting (version 9.1.1). Copyright 1996-2005 Kunihiro Ishiguro, et al. firewall2.home.arpa# show ip route Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF, I - IS-IS, B - BGP, E - EIGRP, T - Table, v - VNC, V - VNC-Direct, A - Babel, f - OpenFabric, > - selected route, * - FIB route, q - queued, r - rejected, b - backup t - trapped, o - offload failure C>* 192.0.2.8/29 [0/1] is directly connected, vmx0, 00:01:19 C>* 192.168.42.0/24 [0/1] is directly connected, vmx1, 00:01:19 C>* 192.168.57.0/24 [0/1] is directly connected, vmx5, 00:01:19 firewall2.home.arpa#So yes - NIC type doesn't seem to matter. Well there's over 20 NIC types to choose from in GNS3... but I'll just stop at 2 for now. We know it is at least not limited to vmxnet3. Also not related to OSPF as I am not running it in this lab. I am not running any routing protocols. I also checked the routing tables in FRR 9.1.1 about 20 minutes afterwards, and still no K routes. I wasn't even able to coax the K routes to come out by rebooting ISP1 and ISP2. In this lab, I only have a single WAN on each firewall, unlike my production lab, where I have dual WAN links.
Thanks for the link to the upstream issues. I'll go through them and see if it's known. If not then I'll log the issue.
-
If you're not running any dynamic routing on that lab and the default route is still missing - where is it supposed to come from? Is it just defined as a static route in Zebra?
-
@marcosm I have updated 15623 with my findings. 9.0.3 works without issue for displaying 0/0 route
-
@marcosm I don't know why the default route still works when it goes missing in FRR. Perhaps it falls back to the system/kernel routing table for any destinations not found in the FRR RIB/FIB?
[2.7.2-RELEASE][root@firewall1.home.arpa]/root: vtysh Hello, this is FRRouting (version 9.1.1). Copyright 1996-2005 Kunihiro Ishiguro, et al. firewall1.home.arpa# show ip route Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF, I - IS-IS, B - BGP, E - EIGRP, T - Table, v - VNC, V - VNC-Direct, A - Babel, f - OpenFabric, > - selected route, * - FIB route, q - queued, r - rejected, b - backup t - trapped, o - offload failure C>* 192.0.2.0/29 [0/1] is directly connected, igb0, 00:12:30 C>* 192.168.41.0/24 [0/1] is directly connected, igb1, 00:12:30 C>* 192.168.57.0/24 [0/1] is directly connected, igb5, 00:12:30 firewall1.home.arpa# show ip fib Codes: K - kernel route, C - connected, S - static, R - RIP, O - OSPF, I - IS-IS, B - BGP, E - EIGRP, T - Table, v - VNC, V - VNC-Direct, A - Babel, f - OpenFabric, > - selected route, * - FIB route, q - queued, r - rejected, b - backup t - trapped, o - offload failure C>* 192.0.2.0/29 [0/1] is directly connected, igb0, 00:12:33 C>* 192.168.41.0/24 [0/1] is directly connected, igb1, 00:12:33 C>* 192.168.57.0/24 [0/1] is directly connected, igb5, 00:12:33 firewall1.home.arpa# exit [2.7.2-RELEASE][root@firewall1.home.arpa]/root: netstat -rn4 Routing tables Internet: Destination Gateway Flags Netif Expire default 192.0.2.1 UGS igb0 127.0.0.1 link#8 UH lo0 192.0.2.0/29 link#1 U igb0 192.0.2.2 link#8 UHS lo0 192.168.41.0/24 link#2 U igb1 192.168.41.1 link#8 UHS lo0 192.168.57.0/24 link#6 U igb5 192.168.57.204 link#8 UHS lo0 [2.7.2-RELEASE][root@firewall1.home.arpa]/root: traceroute 8.8.8.8 traceroute to 8.8.8.8 (8.8.8.8), 64 hops max, 40 byte packets 1 192.0.2.1 (192.0.2.1) 1.736 ms 1.638 ms 1.182 ms 2 192.168.1.1 (192.168.1.1) 1.958 ms 2.197 ms 1.552 ms 3 192.168.57.1 (192.168.57.1) 2.678 ms 2.682 ms 2.554 ms 4 10.231.48.10 (10.231.48.10) 23.731 ms 22.466 ms 21.816 ms 5 ae10.chw-ice301.sydney.telstra.net (203.50.61.65) 22.058 ms 22.454 ms ae10.ken-ice301.sydney.telstra.net (203.50.61.81) 24.173 ms 6 bundle-ether25.hay-core30.sydney.telstra.net (203.50.61.80) 22.530 ms bundle-ether25.stl-core30.sydney.telstra.net (203.50.61.64) 24.026 ms bundle-ether25.hay-core30.sydney.telstra.net (203.50.61.80) 23.712 ms 7 bundle-ether1.chw-edge903.sydney.telstra.net (203.50.11.177) 22.088 ms 22.215 ms 22.276 ms 8 goo2503144.lnk.telstra.net (58.163.91.202) 23.689 ms goo2503069.lnk.telstra.net (58.163.91.194) 22.960 ms 72.14.212.22 (72.14.212.22) 23.076 ms 9 192.178.97.87 (192.178.97.87) 24.997 ms 192.178.98.33 (192.178.98.33) 23.710 ms 192.178.98.21 (192.178.98.21) 23.375 ms 10 142.251.64.177 (142.251.64.177) 23.976 ms 142.251.64.179 (142.251.64.179) 23.778 ms 216.239.56.69 (216.239.56.69) 23.841 ms 11 dns.google (8.8.8.8) 27.750 ms 23.449 ms 23.632 msIn my basic lab here, I just have the gateway configured like so:
Just to recap the issue in my production network: Since that default gateway / (default route) doesn't show up as a kernel route in FRR (from FRR 9.1 and onwards), the situation is that when the pfSense firewall with this affliction learns a default route via OSPF from another router device behind it (on the firewall "LAN"), the whole network Internet traffic still arrives at the pfSense firewall, because it is still advertising the default route into the network and has the lowest cost, but then the pfSense then decides to send the traffic back to the LAN router - back to where it came from. It crazily thinks that some high-cost OSPF route is a better option than its directly connected default gateway, and it shouldn't. It was working fine before that. Fine in 9.0.2 / 9.0.3.
I guess not too many people run OSPF on their network with a competing default route? Otherwise people would be screaming about this issue all over the place. It would be broken for everyone right now on the new code, but the breakage only affects certain topologies.
The secondary default route I have is across a 700Mbps microwave link to a remote site that also has an pfSense firewall with a fairly low-speed Internet link. If the main site's internet goes down, then the customer traffic ends up going to the LAN router (Mikrotik Cloud Core LAN router) as normal and then takes the high-cost OSPF route across the microwave link to the remote site and still have Internet (vs the low-cost OSPF route to the local pfSense router).
I fixed this issue on the day by turning off the remote firewall's advertisement of OSPF (put in a temporary static route where needed). Then afterhours I used the Netgate Installer to reinstall the previous version of pfSense Plus, to get the earlier FRR back. Now both are advertising their default routes and redundancy is fine. Just want to get this issue solved so I can upgrade the main site to pfSense Plus 24.3... 24.8 etc.
-
To clarify, is the default route is missing from both Zebra and the kernel, or just Zebra?
It crazily thinks that some high-cost OSPF route is a better option than its directly connected default gateway, and it shouldn't.
Is this happening while the lower-cost route exists in the kernel? If so, is that happening with newly established traffic as well (as in not traffic for which states already exist)?
-
Please test this patched frr 9.1 version and let us know if the issue persists.
-
@marcosm I tested that in my production lab by upgrading the lab PfSense Plus 23.x to 24.x and seeing the breakage (K routes disappearing), and then I stopped FRR and applied that patched version, and started it again - Kernel routes showing up. Rebooted - still have the K routes.
Is there a bug reference ID you can link to? I'm really curious! I've spent days on this and would love to find out.
Would you recommend I use this in production? Maybe I am best waiting for 24.8 - where perhaps an updated FRR build will have more testing? Then I can skip 24.3 altogether and just go straight to 24.8.
This patched version has this:
configured with: '--enable-user=frr' '--enable-group=frr' '--enable-vty-group=frrvty' '--enable-vtysh' '--disable-doc-html' '--sysconfdir=/var/etc/frr' '--localstatedir=/var/run/frr' '--disable-nhrpd' '--disable-pathd' '--disable-ospfclient' '--disable-pimd' '--disable-pbrd' '--with-vtysh-pager=cat' '--enable-backtrace' '--disable-config-rollbacks' '--disable-datacenter' '--enable-fpm' '--disable-ldpd' '--disable-doc' '--without-libpam' '--enable-rpki' '--disable-sharpd' '--disable-shell-access' '--enable-snmp' '--disable-tcmalloc' '--prefix=/usr/local' '--mandir=/usr/local/man' '--disable-silent-rules' '--infodir=/usr/local/share/info/' '--build=amd64-portbld-freebsd15.0' 'build_alias=amd64-portbld-freebsd15.0' 'PKG_CONFIG=pkgconf' 'PKG_CONFIG_LIBDIR=/wrkdirs/usr/ports/net/frr9/work/.pkgconfig:/usr/local/libdata/pkgconfig:/usr/local/share/pkgconfig:/usr/libdata/pkgconfig' 'CC=cc' 'CFLAGS=-O2 -pipe -fstack-protector-strong -fno-strict-aliasing ' 'LDFLAGS= -L/usr/local/lib -L/usr/local/lib -fstack-protector-strong ' 'LIBS=' 'CPPFLAGS=-I/usr/local/include -I/usr/local/include' 'CPP=cpp' 'CXX=c++' 'CXXFLAGS=-O2 -pipe -fstack-protector-strong -fno-strict-aliasing ' 'PYTHON=/usr/local/bin/python3.11'I compared it to other builds and nothing stands out. SNMP was off in one of the builds (for CE) and one of the other builds had "--mandir=/usr/local/share/man" instead of "--mandir=/usr/local/man" so am thinking that the fix was more than just build config.
In case this info is still required.... even though the root cause seems to have been identified/fixed....
To clarify, is the default route is missing from both Zebra and the kernel, or just Zebra?
The default route was missing just from Zebra. It was in the kernel.
Is this happening while the lower-cost route exists in the kernel?
Yes that's right.
If so, is that happening with newly established traffic as well (as in not traffic for which states already exist)?
Yes. Internet web browsing to new websites was broken. Traffic would go from a workstation to the Mikrotik cloud core LAN router. The Mikrotik could see the default route to the local pfsense, and also a default route over the microwave link to the other site's pfsense. The microwave link has a high OSPF cost, so the LAN router would correctly send the Internet traffic to the local pfSense. But then the local pfSense had an OSPF-learned route to the remote site over the microwave link and no K route for the local connected gateway, and bounced the traffic back to the LAN router, which then sent it back to the local pfsense . Can see that with traceroutes - traffic oscillating between firewall and LAN router until TTL timeout.
I don't know how it all works, but my experience suggests that if a route exists in Zebra and is subsequently added to the Zebra FIB, then this is the forwarding that gets used. If Zebra has no RIB/FIB entry, then it falls back to the system RIB/FIB (as given by "netstat -rn") before failing. This layering would make sense so that Zebra can start and stop with the least amount of impact. It's a massive danger though when the kernel routes don't get pushed from system/kernel to Zebra, because an incomplete view can lead to extremely poor routing decisions.
-
We found what looks to be the root cause - info has been posted to the Redmine report.
The route redistribution issue still needs testing with the patched version, any help with that would be appreciated.
I suggest waiting until we pick back the fix to 24.03 for your production systems.
-
@marcosm said in Updating to pfSense+ 24.3 breaks routing - kernel routes now gone:
Please test this patched frr 9.1 version and let us know if the issue persists.
How do you install this? Sorry pretty new. Can I just scp this to my netgate 7100 and use some sort of package manager to install? Any particular process that won't break further releases?
-
@mAineAc See the previous comment.
-
@marcosm Yeah, after installing no change. rebooted no change. I don't see the default route in FRR and it is not redistributing the default route.
-
-
@marcosm I just tested in my production simulation lab and all looks good. I'll update the actual production firewall this weekend. This is a great result - thanks so much for your efforts - it's really appreciated.
-
@marcosm Will this be coming to 24.08.a.20240702.0600? I am running this and the package listed does not seem to work and i am still having the same issue. I have not seen any updated packages.
-
@mAineAc No - you'd have to build/install it manually for the public dev build. I'm not aware of any official bug report for the issue you're experiencing. My suggestion is to treat it like any other bug report: provide steps to reproduce it, and determine if it's a regression by finding the version(s) of the related software when it last worked.