Skip to content

A connection dropped by a cut, handed back when its address leaves the list. Measured, 2026-11-17

A cut drops a connection on the coprocessor by its verdict, and keeps it there while its sender goes on sending (measured 2026-10-23). Until this, nothing took it back by address when the cut was over - its address taken out of the list, the list unchosen, the plugin that cut it turned off: the connection stayed dropped until its sender stopped, or until it was handed back by hand on XGS: Offloaded Flows (found by the review of os-blocklists' phase 1).

hwdrop.release_addresses hands back to the host every connection with an end in a set of addresses, by its index and revision, through each driver's own hand-back - the write flows.py release makes. os-blocklists' cut.py calls it for the entries no watched list holds any more. Handed back, the connection meets pf again: a pass rule makes it a new state, and it goes through.

How it was measured

On the second XGS 3300 (OPNsense 26.7.9, os-xgs-npu main, octep.ko 10eee77b), with os-blocklists' cut.py of its phase 1 branch:

  • UDP at 200 Mbit/s for 60 s from the bench machine on F2 to an address of the XGS 136 that the 3300 reaches out of Port8; the XGS 136 counts what arrives on its link from Port8 every 5 seconds.
  • At 12 s, that address goes into blocklists_feodo and cut.py runs: the pf cut, and the drop on the coprocessor. At 32 s it leaves the table and cut.py runs again.
  • Once with hwdrop.py as on main, and once with release_addresses in it. Block Lists' settings and hwdrop.py were put back as found after.

What came of it

MB arriving per 5 s before the cut after the cut after the address left the list
hwdrop.py as on main 102, 128 0 0, to the end of the stream (30 s)
with release_addresses 69, 128 0, 0, 0 47 within 5 s, then 128 each 5 s to the end

Both times the cut dropped the stream's two connections on the coprocessor ("2 dropped in hardware now"). With release_addresses, the run after the address left the list said "1 handed back from hardware now", and the system log "1 connection handed back from hardware: no list holds their address now". hwdrop.py as on main has no release_addresses, and cut.py hands nothing back.

hwdrop.py as on main: the stream dropped by the cut stays dropped after its address leaves the list

With release_addresses: handed back at the next run of cut.py, and on again at 128 MB each 5 s

Not measured

  • The XGS 136's driver, whose hand-back flow.release takes the same write: nothing in the lab forwards a connection through that unit. tools/test-hwdrop.py holds both drivers' writes and the words they answer with.
  • A list unchosen or the plugin turned off: the same call, for every entry kept for the list; os-blocklists' tools/test-hwdrop.py holds them.

Discussion

The discussion of this page is held in the project's GitHub Discussions. Showing it here loads giscus from giscus.app and GitHub, and writing in it needs a GitHub account - what that means.