Re: [PATCH net] net/sched: act_gact, act_police: range check the fallback control action

[Hyunjung Ko]

On Wed, Aug 5, 2026 at 1:59 PM Jamal Hadi Salim <jhs@xxxxxxxxxxxx> wrote:
>
> 1) We test almost _everything_, so to get a review - even if it as
> trivial as this: Always, always send a test case to reproduce even if
> it seems as obvious as this. Preferable will be tdc. But you can send
> or point to an AI generated poc as well if you cant ask it to create a
> tdc test. If the issue is sensitive - send the poc to the tc/netdev
> maintainers in a separate email.
>
> 2) If you got assistance from an ai - please add assisted-by tag.
>
> Same goes for your other patch...

Thanks for the review. Both points addressed for both patches; v2 of
each follows shortly.

The act_ct patch now comes with a tdc case (2/2). It uses the scapy
plugin to inject the malformed IPv6 frame and matches on the clsact
drop counter, which turns out to be a clean discriminator: before the
fix act_ct returns TC_ACT_CONSUMED, so tc_run() never reaches its
TC_ACT_SHOT arm and the counter stays at zero while the skbs leak;
after the fix it reads "dropped 10".

To be straight about how far I verified that: I do not have a
scapy-capable tdc setup, so I have not run tdc.py over the case
itself. I ran the equivalent by hand under qemu on both an unpatched
and a patched kernel - same topology, same ten frames, same tc -s
qdisc show - and got "dropped 0" vs "dropped 10". The JSON is modelled
on the existing scapy cases in the same file (3992, 9c2a). Noted below
the --- line of 2/2 as well.

This patch I could not express in tdc. iproute2 only parses symbolic
control-action names, so tc(8) rejects the bad value before it ever
reaches the kernel:

  $ tc actions add action gact drop random determ ok 2
  RTNETLINK answers: Operation not permitted   <- parsed fine
  $ tc actions add action gact drop random determ 9 2
  Bad action type 9                            <- rejected by iproute2

The fallback has to be planted over raw netlink, so I have inlined a
self-contained C reproducer below the --- line of v2 instead. It sets
up the clsact chain, plants TCA_GACT_PROB.paction = 9 and then
TCA_POLICE_RESULT = 9, and reports skbuff_head_cache growth per
injected packet. If you would rather have this as a tdc case anyway I
can write a plugin that does the raw netlink setup, but that looked
like more machinery than a one-line range check warrants - happy to do
it if you disagree.

The AI assistance tag is on both patches now.

Thanks,
Hyunjung