[Kernel-packages] [Bug 1959702] Re: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb
It looks to be 'an interesting mystery' we're chasing. This system is in production, so the results below are with the whole 'snooping engine' off as without it the whole thing dies. As such, I don't think the contents of the fdb and mdb tables mean much. The setups below are unchanged, they fail if snooping is on, work when it's off. root@noc1:~# ip -s -d link show vmbridge 6: vmbridge: mtu 1500 qdisc noqueue state UP mode DEFAULT group default qlen 1000 link/ether 52:54:33:bf:a8:78 brd ff:ff:ff:ff:ff:ff promiscuity 0 minmtu 68 maxmtu 65535 bridge forward_delay 1500 hello_time 200 max_age 2000 ageing_time 3 stp_state 0 priority 21 vlan_filtering 1 vlan_protocol 802.1Q bridge_id 0015.52:54:33:bf:a8:78 designated_root 0015.52:54:33:bf:a8:78 root_port 0 root_path_cost 0 topology_change 0 topology_change_detected 0 hello_timer 0.00 tcn_timer0.00 topology_change_timer0.00 gc_timer 137.32 vlan_default_pvid 0 vlan_stats_enabled 0 vlan_stats_per_port 0 group_fwd_mask 0 group_address 01:80:c2:00:00:00 mcast_snooping 0 mcast_router 1 mcast_query_use_ifaddr 0 mcast_querier 0 mcast_hash_elasticity 16 mcast_hash_max 4096 mcast_last_member_count 2 mcast_startup_query_count 2 mcast_last_member_interval 100 mcast_membership_interval 26000 mcast_querier_interval 25500 mcast_query_interval 12500 mcast_query_response_interval 1000 mcast_startup_query_interval 3124 mcast_stats_enabled 0 mcast_igmp_version 2 mcast_mld_version 1 nf_call_iptables 0 nf_call_ip6tables 0 nf_call_arptables 0 addrgenmode none numtxqueues 1 numrxqueues 1 gso_max_size 32000 gso_max_segs 24 RX: bytes packets errors dropped missed mcast 2013504987 17838769 0 0 0 5037107 TX: bytes packets errors dropped carrier collsns 84611 0 0 0 0 The bridge itself has no address of any kind other than a mac. Here's a detail from one of many identical vm setups: ip -s -d link show dbl 16: dbl: mtu 1500 qdisc noqueue master vmbridge state UNKNOWN mode DEFAULT group default qlen 1000 link/ether fe:54:00:60:c5:db brd ff:ff:ff:ff:ff:ff promiscuity 1 minmtu 68 maxmtu 65521 tun type tap pi off vnet_hdr off persist off bridge_slave state forwarding priority 32 cost 100 hairpin off guard off root_block off fastleave off learning on flood on port_id 0x800a port_no 0xa designated_port 32778 designated_cost 0 designated_bridge 0015.52:54:33:bf:a8:78 designated_root 0015.52:54:33:bf:a8:78 hold_timer 0.00 message_age_timer0.00 forward_delay_timer0.00 topology_change_ack 0 config_pending 0 proxy_arp off proxy_arp_wifi off mcast_router 1 mcast_fast_leave off mcast_flood on mcast_to_unicast off neigh_suppress off group_fwd_mask 0 group_fwd_mask_str 0x0 vlan_tunnel off isolated off addrgenmode eui64 numtxqueues 1 numrxqueues 1 gso_max_size 65536 gso_max_segs 65535 RX: bytes packets errors dropped missed mcast 18895018565 174327707 0 25070 0 0 TX: bytes packets errors dropped carrier collsns 30861506740 198028258 0 14270 0 There's some confidential stuff, so I've deleted repetitive vlan entries below. vlan 121 in this sandbox carries 'local lan' traffic. The bridge has an interface on that. vlan 100 is an only-in-rack server-server backbone for ceph. 122 is a public internet ingress on this mixed-use box. #bridge vlan show port vlan-id enp5s2122 PVID Egress Untagged enp4s0f0 103 105 ... 121 PVID Egress Untagged 122 133 enp4s0f1 100 PVID Egress Untagged 101 102 104 lan0noc0port 121 PVID Egress Untagged <-- this has a v4/v6 address for the server's use. ... gate 100 101 ... registry 121 PVID Egress Untagged 127 131 dbl 101 PVID Egress Untagged 121 ... -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1959702 Title: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb Status in linux package in Ubuntu: Confirmed Bug description: Starting at the end: I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. A recent kernel upgrade has broken ipv6/ip6 ndp in a host/kvm setup using a bridge on the host and vlans for some guests. I've tracked the problem to a failure of the mcast code to add entries to the host's mdb table. Manually adding the entries
[Kernel-packages] [Bug 1959702] Re: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb
Harry, I'm still working to reproduce this, without success. I have set the .autoconf sysctl to 0 (which controls creation of local addresses in response to received Router Advertisements), as well as setting .addr_gen_mode to 1 (to disable SLAAC (fe80::) addresses). In any event, .autoconf=0 and .addr_gen_mode=1 still fails to reproduce the issue on my test system. I find that if I disable mcast_flood on the relevant bridge ports (i.e., bridge link set dev vnet1 mcast_flood off) I do see the behavior you describe, but in that case no variant that I've tried (no vid, and all vids in use) of "bridge mdb add ... grp ff02::1:ff00:2" appears to permit ND traffic to pass to the VM destination. Can you provide more specifics of how exactly the bridge and ports are configured? Ideally, both the method to set it up, as well as the configuration details when failing (i.e., "ip -s -d link show" for the bridge and relevant bridge ports, "bridge vlan show", "bridge mdb show", "bridge fdb show br [bridgename]") Also, to answer a question from your original report, the default setting in the kernel for multicast_snooping (enabled, i.e., 1) hasn't changed recently (and quite possibly ever). -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1959702 Title: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb Status in linux package in Ubuntu: Confirmed Bug description: Starting at the end: I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. A recent kernel upgrade has broken ipv6/ip6 ndp in a host/kvm setup using a bridge on the host and vlans for some guests. I've tracked the problem to a failure of the mcast code to add entries to the host's mdb table. Manually adding the entries to the mdb on the bridge corrects the problem. It's very easy to demonstrate the bug in an all ubuntu setup. 1. On an ubuntu host, create two vms, I used libvirt, as set up below. 2. On the host, create a bridge and vlan with two ports, each with the chosen vlan as PVID and egress untagged. Assign those ports one each to the guests as the interface, use e1000. Be sure to NOT autoconfigure the host side of the bridge ports with any ip4 or ip6 address (including fe80::), it's just an avoidable security risk. We don't want to allow the host any sort of ip access / exposure to the vlan. In other words, treat the host's bridge ports as if a 'real off-host switch' without expectation of making each bridge's port being ip6 addressable on the bridge itself. (FWIW: Worth checking if the vlan is left tagged and not pvid, and the vlan is decoded in the guest as a separate interface, does the problem go away? It imposes the burden of vlan management awareness to the guest and so is not acceptable as a solution.) 3. On the host, assign a physical NIC to the bridge and the vlan to the nic. The egress is tagged for the chosen vlan and not PVID. Optionally set up an off-host gateway for the vlan, but it isn't necessary to show the bug. 4. On each guest, manually assign a unique ip4 and ip6 address on the same subnet (you'll see though dhcp4 could work if there was an off- host router providing related services, the bug prevents dhcp6 from working). 5. On one vm, ping the other. Notice ip4 pings work, ip6 pings do not. 6. Manually add the fe02::ffxx: entries for each vm to the vlan to the host bridge's multicast table. Use 'temp' if you're quick enough, otherwise perm. 7. Notice pings between the guests now work on ip6 and ipv4. Using tcpdump and watching icmp6 traffic, you'll notice the packets making it across the various bridge ports the moment you manually add the appropriate fe02::ff... multicast address to the mdb table. Beware a false sense of security: Once the ndp completes and the link addresses are in the fdb, it can 'seem like' everything is fine until the fdb times out and the required mdb entry again must be used to allow ndp to refresh the address. Setting mcast_querier doesn't help. Perhaps previous kernels turned off the multicast snooping by default and just flooded all the bridge ports with all multicast traffic so this bug was avoided. It's my hunch the reason there hasn't been more complaint about this is it takes an extra step to not autoconfigure the vm ports with fe80:: link local addresses on the host. I believe the existence of the fe80 address on the host ports engages ndp code on the host to load the mdb as if preparing for the host's side of the
[Kernel-packages] [Bug 1959702] Re: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb
P.S. The reason this is a security issue is-- there is now an address on the host that the guest also 'knows' and it sits on the bridge giving access to all the other guests on the bridge. Most admins will not 'just know' they need rules to block fe80 traffic generated by host interfaces-- because of course at least one host interface needs an fe80 address for ipv6 neighbor protocols to work. It's a really big 'soft hole'. -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1959702 Title: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb Status in linux package in Ubuntu: Confirmed Bug description: Starting at the end: I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. A recent kernel upgrade has broken ipv6/ip6 ndp in a host/kvm setup using a bridge on the host and vlans for some guests. I've tracked the problem to a failure of the mcast code to add entries to the host's mdb table. Manually adding the entries to the mdb on the bridge corrects the problem. It's very easy to demonstrate the bug in an all ubuntu setup. 1. On an ubuntu host, create two vms, I used libvirt, as set up below. 2. On the host, create a bridge and vlan with two ports, each with the chosen vlan as PVID and egress untagged. Assign those ports one each to the guests as the interface, use e1000. Be sure to NOT autoconfigure the host side of the bridge ports with any ip4 or ip6 address (including fe80::), it's just an avoidable security risk. We don't want to allow the host any sort of ip access / exposure to the vlan. In other words, treat the host's bridge ports as if a 'real off-host switch' without expectation of making each bridge's port being ip6 addressable on the bridge itself. (FWIW: Worth checking if the vlan is left tagged and not pvid, and the vlan is decoded in the guest as a separate interface, does the problem go away? It imposes the burden of vlan management awareness to the guest and so is not acceptable as a solution.) 3. On the host, assign a physical NIC to the bridge and the vlan to the nic. The egress is tagged for the chosen vlan and not PVID. Optionally set up an off-host gateway for the vlan, but it isn't necessary to show the bug. 4. On each guest, manually assign a unique ip4 and ip6 address on the same subnet (you'll see though dhcp4 could work if there was an off- host router providing related services, the bug prevents dhcp6 from working). 5. On one vm, ping the other. Notice ip4 pings work, ip6 pings do not. 6. Manually add the fe02::ffxx: entries for each vm to the vlan to the host bridge's multicast table. Use 'temp' if you're quick enough, otherwise perm. 7. Notice pings between the guests now work on ip6 and ipv4. Using tcpdump and watching icmp6 traffic, you'll notice the packets making it across the various bridge ports the moment you manually add the appropriate fe02::ff... multicast address to the mdb table. Beware a false sense of security: Once the ndp completes and the link addresses are in the fdb, it can 'seem like' everything is fine until the fdb times out and the required mdb entry again must be used to allow ndp to refresh the address. Setting mcast_querier doesn't help. Perhaps previous kernels turned off the multicast snooping by default and just flooded all the bridge ports with all multicast traffic so this bug was avoided. It's my hunch the reason there hasn't been more complaint about this is it takes an extra step to not autoconfigure the vm ports with fe80:: link local addresses on the host. I believe the existence of the fe80 address on the host ports engages ndp code on the host to load the mdb as if preparing for the host's side of the bridge to participate in ip4 and ip6 higher layer traffic, but that's a 'bad hack that happens to work' -- it shouldn't be a requirement that each host vlan port have an ip6 address, after all it didn't need an IP4 address I've attached a linux-bug for you, but it's probably mostly unrelated info. I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1959702/+subscriptions -- Mailing list:
[Kernel-packages] [Bug 1959702] Re: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb
I need to repeat: in sysctl.d put this line in a file, then reboot, then your test setup will show the failure: net.ipv6.conf.all.autoconf = 0 Otherwise, in your test setup the tables are populated, then you delete the addresses, but the L3/4 code engaged by even a little time with the fe80:... address at least until there is a timeout in the tables (longer than you wait for your test) kills the platform. Putting the above in then rebooting will avoid that, so you will see the bug when at no time was there an fe80 address on the interface. -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1959702 Title: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb Status in linux package in Ubuntu: Confirmed Bug description: Starting at the end: I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. A recent kernel upgrade has broken ipv6/ip6 ndp in a host/kvm setup using a bridge on the host and vlans for some guests. I've tracked the problem to a failure of the mcast code to add entries to the host's mdb table. Manually adding the entries to the mdb on the bridge corrects the problem. It's very easy to demonstrate the bug in an all ubuntu setup. 1. On an ubuntu host, create two vms, I used libvirt, as set up below. 2. On the host, create a bridge and vlan with two ports, each with the chosen vlan as PVID and egress untagged. Assign those ports one each to the guests as the interface, use e1000. Be sure to NOT autoconfigure the host side of the bridge ports with any ip4 or ip6 address (including fe80::), it's just an avoidable security risk. We don't want to allow the host any sort of ip access / exposure to the vlan. In other words, treat the host's bridge ports as if a 'real off-host switch' without expectation of making each bridge's port being ip6 addressable on the bridge itself. (FWIW: Worth checking if the vlan is left tagged and not pvid, and the vlan is decoded in the guest as a separate interface, does the problem go away? It imposes the burden of vlan management awareness to the guest and so is not acceptable as a solution.) 3. On the host, assign a physical NIC to the bridge and the vlan to the nic. The egress is tagged for the chosen vlan and not PVID. Optionally set up an off-host gateway for the vlan, but it isn't necessary to show the bug. 4. On each guest, manually assign a unique ip4 and ip6 address on the same subnet (you'll see though dhcp4 could work if there was an off- host router providing related services, the bug prevents dhcp6 from working). 5. On one vm, ping the other. Notice ip4 pings work, ip6 pings do not. 6. Manually add the fe02::ffxx: entries for each vm to the vlan to the host bridge's multicast table. Use 'temp' if you're quick enough, otherwise perm. 7. Notice pings between the guests now work on ip6 and ipv4. Using tcpdump and watching icmp6 traffic, you'll notice the packets making it across the various bridge ports the moment you manually add the appropriate fe02::ff... multicast address to the mdb table. Beware a false sense of security: Once the ndp completes and the link addresses are in the fdb, it can 'seem like' everything is fine until the fdb times out and the required mdb entry again must be used to allow ndp to refresh the address. Setting mcast_querier doesn't help. Perhaps previous kernels turned off the multicast snooping by default and just flooded all the bridge ports with all multicast traffic so this bug was avoided. It's my hunch the reason there hasn't been more complaint about this is it takes an extra step to not autoconfigure the vm ports with fe80:: link local addresses on the host. I believe the existence of the fe80 address on the host ports engages ndp code on the host to load the mdb as if preparing for the host's side of the bridge to participate in ip4 and ip6 higher layer traffic, but that's a 'bad hack that happens to work' -- it shouldn't be a requirement that each host vlan port have an ip6 address, after all it didn't need an IP4 address I've attached a linux-bug for you, but it's probably mostly unrelated info. I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. To manage notifications about this bug go
[Kernel-packages] [Bug 1959702] Re: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb
** Changed in: linux (Ubuntu) Status: Incomplete => Confirmed -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1959702 Title: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb Status in linux package in Ubuntu: Confirmed Bug description: Starting at the end: I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. A recent kernel upgrade has broken ipv6/ip6 ndp in a host/kvm setup using a bridge on the host and vlans for some guests. I've tracked the problem to a failure of the mcast code to add entries to the host's mdb table. Manually adding the entries to the mdb on the bridge corrects the problem. It's very easy to demonstrate the bug in an all ubuntu setup. 1. On an ubuntu host, create two vms, I used libvirt, as set up below. 2. On the host, create a bridge and vlan with two ports, each with the chosen vlan as PVID and egress untagged. Assign those ports one each to the guests as the interface, use e1000. Be sure to NOT autoconfigure the host side of the bridge ports with any ip4 or ip6 address (including fe80::), it's just an avoidable security risk. We don't want to allow the host any sort of ip access / exposure to the vlan. In other words, treat the host's bridge ports as if a 'real off-host switch' without expectation of making each bridge's port being ip6 addressable on the bridge itself. (FWIW: Worth checking if the vlan is left tagged and not pvid, and the vlan is decoded in the guest as a separate interface, does the problem go away? It imposes the burden of vlan management awareness to the guest and so is not acceptable as a solution.) 3. On the host, assign a physical NIC to the bridge and the vlan to the nic. The egress is tagged for the chosen vlan and not PVID. Optionally set up an off-host gateway for the vlan, but it isn't necessary to show the bug. 4. On each guest, manually assign a unique ip4 and ip6 address on the same subnet (you'll see though dhcp4 could work if there was an off- host router providing related services, the bug prevents dhcp6 from working). 5. On one vm, ping the other. Notice ip4 pings work, ip6 pings do not. 6. Manually add the fe02::ffxx: entries for each vm to the vlan to the host bridge's multicast table. Use 'temp' if you're quick enough, otherwise perm. 7. Notice pings between the guests now work on ip6 and ipv4. Using tcpdump and watching icmp6 traffic, you'll notice the packets making it across the various bridge ports the moment you manually add the appropriate fe02::ff... multicast address to the mdb table. Beware a false sense of security: Once the ndp completes and the link addresses are in the fdb, it can 'seem like' everything is fine until the fdb times out and the required mdb entry again must be used to allow ndp to refresh the address. Setting mcast_querier doesn't help. Perhaps previous kernels turned off the multicast snooping by default and just flooded all the bridge ports with all multicast traffic so this bug was avoided. It's my hunch the reason there hasn't been more complaint about this is it takes an extra step to not autoconfigure the vm ports with fe80:: link local addresses on the host. I believe the existence of the fe80 address on the host ports engages ndp code on the host to load the mdb as if preparing for the host's side of the bridge to participate in ip4 and ip6 higher layer traffic, but that's a 'bad hack that happens to work' -- it shouldn't be a requirement that each host vlan port have an ip6 address, after all it didn't need an IP4 address I've attached a linux-bug for you, but it's probably mostly unrelated info. I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1959702/+subscriptions -- Mailing list: https://launchpad.net/~kernel-packages Post to : kernel-packages@lists.launchpad.net Unsubscribe : https://launchpad.net/~kernel-packages More help : https://help.launchpad.net/ListHelp
Re: [Kernel-packages] [Bug 1959702] Re: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb
Yup, those failures were to do with an old radeon chipset on an ancient server. On 2/1/22 17:33, Seth Arnold wrote: > Sounds good, thanks: > > [0.00] Linux version 5.11.0-49-generic (buildd@lcy02-amd64-054) > (gcc (Ubuntu 10.3.0-1ubuntu1) 10.3.0, GNU ld (GNU Binutils for Ubuntu) > 2.36.1) #55-Ubuntu SMP Wed Jan 12 17:36:34 UTC 2022 (Ubuntu > 5.11.0-49.55-generic 5.11.22) > > btw, there were a bunch of memory allocation failures in the dmesg, in > case they weren't obvious. > > ** Attachment removed: "apport.linux-image-5.11.0-49-generic.rw5aqx7x.apport" > > https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1959702/+attachment/5558647/+files/apport.linux-image-5.11.0-49-generic.rw5aqx7x.apport > > ** Information type changed from Private Security to Public Security > -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1959702 Title: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb Status in linux package in Ubuntu: New Bug description: Starting at the end: I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. A recent kernel upgrade has broken ipv6/ip6 ndp in a host/kvm setup using a bridge on the host and vlans for some guests. I've tracked the problem to a failure of the mcast code to add entries to the host's mdb table. Manually adding the entries to the mdb on the bridge corrects the problem. It's very easy to demonstrate the bug in an all ubuntu setup. 1. On an ubuntu host, create two vms, I used libvirt, as set up below. 2. On the host, create a bridge and vlan with two ports, each with the chosen vlan as PVID and egress untagged. Assign those ports one each to the guests as the interface, use e1000. Be sure to NOT autoconfigure the host side of the bridge ports with any ip4 or ip6 address (including fe80::), it's just an avoidable security risk. We don't want to allow the host any sort of ip access / exposure to the vlan. In other words, treat the host's bridge ports as if a 'real off-host switch' without expectation of making each bridge's port being ip6 addressable on the bridge itself. (FWIW: Worth checking if the vlan is left tagged and not pvid, and the vlan is decoded in the guest as a separate interface, does the problem go away? It imposes the burden of vlan management awareness to the guest and so is not acceptable as a solution.) 3. On the host, assign a physical NIC to the bridge and the vlan to the nic. The egress is tagged for the chosen vlan and not PVID. Optionally set up an off-host gateway for the vlan, but it isn't necessary to show the bug. 4. On each guest, manually assign a unique ip4 and ip6 address on the same subnet (you'll see though dhcp4 could work if there was an off- host router providing related services, the bug prevents dhcp6 from working). 5. On one vm, ping the other. Notice ip4 pings work, ip6 pings do not. 6. Manually add the fe02::ffxx: entries for each vm to the vlan to the host bridge's multicast table. Use 'temp' if you're quick enough, otherwise perm. 7. Notice pings between the guests now work on ip6 and ipv4. Using tcpdump and watching icmp6 traffic, you'll notice the packets making it across the various bridge ports the moment you manually add the appropriate fe02::ff... multicast address to the mdb table. Beware a false sense of security: Once the ndp completes and the link addresses are in the fdb, it can 'seem like' everything is fine until the fdb times out and the required mdb entry again must be used to allow ndp to refresh the address. Setting mcast_querier doesn't help. Perhaps previous kernels turned off the multicast snooping by default and just flooded all the bridge ports with all multicast traffic so this bug was avoided. It's my hunch the reason there hasn't been more complaint about this is it takes an extra step to not autoconfigure the vm ports with fe80:: link local addresses on the host. I believe the existence of the fe80 address on the host ports engages ndp code on the host to load the mdb as if preparing for the host's side of the bridge to participate in ip4 and ip6 higher layer traffic, but that's a 'bad hack that happens to work' -- it shouldn't be a requirement that each host vlan port have an ip6 address, after all it didn't need an IP4 address I've attached a linux-bug for you, but it's probably mostly unrelated info. I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the
[Kernel-packages] [Bug 1959702] Re: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb
Sounds good, thanks: [0.00] Linux version 5.11.0-49-generic (buildd@lcy02-amd64-054) (gcc (Ubuntu 10.3.0-1ubuntu1) 10.3.0, GNU ld (GNU Binutils for Ubuntu) 2.36.1) #55-Ubuntu SMP Wed Jan 12 17:36:34 UTC 2022 (Ubuntu 5.11.0-49.55-generic 5.11.22) btw, there were a bunch of memory allocation failures in the dmesg, in case they weren't obvious. ** Attachment removed: "apport.linux-image-5.11.0-49-generic.rw5aqx7x.apport" https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1959702/+attachment/5558647/+files/apport.linux-image-5.11.0-49-generic.rw5aqx7x.apport ** Information type changed from Private Security to Public Security -- You received this bug notification because you are a member of Kernel Packages, which is subscribed to linux in Ubuntu. https://bugs.launchpad.net/bugs/1959702 Title: Regression: ip6 ndp broken, host bridge doesn't add vlan guest entry to mdb Status in linux package in Ubuntu: New Bug description: Starting at the end: I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what you might call a 'passive security vulnerability'. A recent kernel upgrade has broken ipv6/ip6 ndp in a host/kvm setup using a bridge on the host and vlans for some guests. I've tracked the problem to a failure of the mcast code to add entries to the host's mdb table. Manually adding the entries to the mdb on the bridge corrects the problem. It's very easy to demonstrate the bug in an all ubuntu setup. 1. On an ubuntu host, create two vms, I used libvirt, as set up below. 2. On the host, create a bridge and vlan with two ports, each with the chosen vlan as PVID and egress untagged. Assign those ports one each to the guests as the interface, use e1000. Be sure to NOT autoconfigure the host side of the bridge ports with any ip4 or ip6 address (including fe80::), it's just an avoidable security risk. We don't want to allow the host any sort of ip access / exposure to the vlan. In other words, treat the host's bridge ports as if a 'real off-host switch' without expectation of making each bridge's port being ip6 addressable on the bridge itself. (FWIW: Worth checking if the vlan is left tagged and not pvid, and the vlan is decoded in the guest as a separate interface, does the problem go away? It imposes the burden of vlan management awareness to the guest and so is not acceptable as a solution.) 3. On the host, assign a physical NIC to the bridge and the vlan to the nic. The egress is tagged for the chosen vlan and not PVID. Optionally set up an off-host gateway for the vlan, but it isn't necessary to show the bug. 4. On each guest, manually assign a unique ip4 and ip6 address on the same subnet (you'll see though dhcp4 could work if there was an off- host router providing related services, the bug prevents dhcp6 from working). 5. On one vm, ping the other. Notice ip4 pings work, ip6 pings do not. 6. Manually add the fe02::ffxx: entries for each vm to the vlan to the host bridge's multicast table. Use 'temp' if you're quick enough, otherwise perm. 7. Notice pings between the guests now work on ip6 and ipv4. Using tcpdump and watching icmp6 traffic, you'll notice the packets making it across the various bridge ports the moment you manually add the appropriate fe02::ff... multicast address to the mdb table. Beware a false sense of security: Once the ndp completes and the link addresses are in the fdb, it can 'seem like' everything is fine until the fdb times out and the required mdb entry again must be used to allow ndp to refresh the address. Setting mcast_querier doesn't help. Perhaps previous kernels turned off the multicast snooping by default and just flooded all the bridge ports with all multicast traffic so this bug was avoided. It's my hunch the reason there hasn't been more complaint about this is it takes an extra step to not autoconfigure the vm ports with fe80:: link local addresses on the host. I believe the existence of the fe80 address on the host ports engages ndp code on the host to load the mdb as if preparing for the host's side of the bridge to participate in ip4 and ip6 higher layer traffic, but that's a 'bad hack that happens to work' -- it shouldn't be a requirement that each host vlan port have an ip6 address, after all it didn't need an IP4 address I've attached a linux-bug for you, but it's probably mostly unrelated info. I believe as the bug presently requires each of the host's bridge ports to be ipv6 addressable to enable ipv6 to function in the guest, and most admins won't think to add special entries into their host's nftables.conf to allow for it 'because who knew?' it represents what