On Sun, May 17, 2026 at 02:01:29PM -0700, Kuniyuki Iwashima wrote: > On Sun, May 17, 2026 at 12:53 PM Stefano Brivio <[email protected]> wrote: > > > > On Sun, 17 May 2026 12:05:45 -0700 > > Kuniyuki Iwashima <[email protected]> wrote: > > > > > On Sun, May 17, 2026 at 11:41 AM Stefano Brivio <[email protected]> > > > wrote: > > > > > > > > Once a socket enters repair mode (TCP_REPAIR socket option with > > > > TCP_REPAIR_ON value), it's possible to dump the receive sequence > > > > number (TCP_QUEUE_SEQ) and the contents of the receive queue itself > > > > (using TCP_REPAIR_QUEUE to select it). > > > > > > > > If we receive data after the application fetched the sequence number > > > > or saved the contents of the queue, though, the application will now > > > > have outdated information, which defeats the whole functionality, > > > > because this leads to gaps in sequence and data once they're restored > > > > by the target instance of the application, resulting in a hanging or > > > > otherwise non-functional TCP connection. > > > > > > > > This type of race condition was discovered in the KubeVirt integration > > > > of passt(1), using a remote iperf3 client connected to an iperf3 > > > > server running in the guest which is being migrated. The setup allows > > > > traffic to reach the origin node hosting the guest during the > > > > migration. > > > > > > > > If passt dumps sequence number and contents of the queue *before* > > > > further data is received and acknowledged to the peer by the kernel, > > > > once the TCP data connection is migrated to the target node, the > > > > remote client becomes unable to continue sending, because a portion > > > > of the data it sent *and received an acknowledgement for* is now lost. > > > > > > > > Schematically: > > > > > > > > 1. client --seq 1:100--> origin host --> passt --> guest --> server > > > > > > > > 2. client <--ACK: 100-- origin host > > > > > > > > 3. migration starts, > > > > > > Here, a netfilter rule or bpf prog must be installed to > > > drop packets temporarily until migration completes. > > > > Thanks for the review. > > > > I have to say it's rather unexpected to have to work around obvious > > kernel issues in userspace: TCP_REPAIR implies that queues are frozen, > > and this is handled correctly on the sending side (see > > tcp_write_xmit()), but was clearly forgotten on the receiving side. > > > > TCP_REPAIR also allows to dump queues, not just sequence numbers, so > > this really is a bad race condition making the whole functionality > > unreliable. > > > > But anyway, even looking for a practical workaround of the kind you > > suggested, I see two issues with it: > > > > 1. we would still have a race condition, because userspace doesn't have > > a way to synchronise application of nftables rules (or even a BPF > > program) with the effects of TCP_REPAIR. > > Note that setsockopt(TCP_REPAIR) is under lock_sock(), so the > backlog is always cleared before returning to userspace, and the > following getsockopt() will have stable view. > > > > We could apply nftables > > rules "a while before" just to be sure, but this is severely going to > > impact migration downtime > > So it's "just before", not "a while before". > > > > > > 2. passt(1) runs unprivileged and uses a very simple helper to set > > TCP_REPAIR on the socket. > > I guess it uses userns ?
It does, but it doesn't help.
> setsockopt(TCP_REPAIR) requires CAP_NET_ADMIN in the
> userns tied to the socket's netns (in tcp_can_repair_sock()),
> so you should be able to use netfilter anyway.
The sockets we use TCP_REPAIR on are *outside* the netns/userns - the
whole point is that we're bridging between the outside world and the
ns. So we have no privileges there to use netfilter (and even if we
did, we couldn't safely use it, since we don't know how else netfilter
is being used on the host system).
--
David Gibson (he or they) | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you, not the other way
| around.
http://www.ozlabs.org/~dgibson
signature.asc
Description: PGP signature
