Flowtables ignore timeout settings in recent kernels

Netfilter ignores the timeout settings for a flowtable

# sysctl -a -r flowtable
net.netfilter.nf_flowtable_tcp_timeout = 30
net.netfilter.nf_flowtable_udp_timeout = 30

Situation. A long udp connection (tunnel) with some data flowing through a router. The connection is sent to a flowtable on the router. It's a few packets per second, more here and there, a pause here and there, and so on over and over. The pauses are minimal and are also limited by the tunnel settings to be no longer than 25 seconds. Everything is satisfying to make the connection last continuously in the flowtable and not reappear in forward. However, the connection keeps dropping out of the flowtable. It stays in the flowtable (offloaded) for a second at most and then it is kicked out, back to forward.

In an attached test script you can see counters that should be zero but are not. If I watch the normal packet flow on a particular router, I can see packets in the conntrack table that should be OFFLOAD as ASSURED.

Tested in kernel 6.5.6. In an old(er) kernel 5.10 it works as expected.

Quoted 16 lines · Vladimir Smelhaus

Netfilter ignores the timeout settings for a flowtable

# sysctl -a -r flowtable
net.netfilter.nf_flowtable_tcp_timeout = 30
net.netfilter.nf_flowtable_udp_timeout = 30

Situation. A long udp connection (tunnel) with some data flowing through a router. The connection is sent to a flowtable on the router. It's a few packets per second, more here and there, a pause here and there, and so on over and over. The pauses are minimal and are also limited by the tunnel settings to be no longer than 25 seconds. Everything is satisfying to make the connection last continuously in the flowtable and not reappear in forward. However, the connection keeps dropping out of the flowtable. It stays in the flowtable (offloaded) for a second at most and then it is kicked out, back to forward.

Quoted 4 lines

In an attached test script you can see counters that should be zero but are not. If I watch the normal packet flow on a particular router, I can see packets in the conntrack table that should be OFFLOAD as ASSURED.

Tested in kernel 6.5.6. In an old(er) kernel 5.10 it works as expected.

Then please perform bisection to find a culprit that introduces your regression (see Documentation/admin-guide/bug-bisect.rst in the kernel sources for reference). Also, it'd been great if you also post the reproducer script inline (within your email) instead, as some MUAs (like mutt that I'm using now) may ignore the attachment.

Anyway, thanks for the regression report. I'm adding it to regzbot:

#regzbot ^introduced: v5.10..v6.5

Quoted 30 lines · Bagas Sanjaya

Quoted 16 lines · Vladimir Smelhaus

Netfilter ignores the timeout settings for a flowtable

# sysctl -a -r flowtable
net.netfilter.nf_flowtable_tcp_timeout = 30
net.netfilter.nf_flowtable_udp_timeout = 30

Situation. A long udp connection (tunnel) with some data flowing through a router. The connection is sent to a flowtable on the router. It's a few packets per second, more here and there, a pause here and there, and so on over and over. The pauses are minimal and are also limited by the tunnel settings to be no longer than 25 seconds. Everything is satisfying to make the connection last continuously in the flowtable and not reappear in forward. However, the connection keeps dropping out of the flowtable. It stays in the flowtable (offloaded) for a second at most and then it is kicked out, back to forward.

Quoted 4 lines

In an attached test script you can see counters that should be zero but are not. If I watch the normal packet flow on a particular router, I can see packets in the conntrack table that should be OFFLOAD as ASSURED.

Tested in kernel 6.5.6. In an old(er) kernel 5.10 it works as expected.

Then please perform bisection to find a culprit that introduces your regression (see Documentation/admin-guide/bug-bisect.rst in the kernel sources for reference). Also, it'd been great if you also post the reproducer script inline (within your email) instead, as some MUAs (like mutt that I'm using now) may ignore the attachment.

Anyway, thanks for the regression report. I'm adding it to regzbot:

Quoted 1 line

#regzbot ^introduced: v5.10..v6.5

Fix here:

https://patchwork.ozlabs.org/project/netfilter-devel/patch/[email protected]/

it is a bug from Jun 2023, regression was introduced in the v6.5 development cycle.