Re: [PATCH net] net/sched: ets: use old 'nbands' while purging unused classes

On Mon, Aug 11, 2025 at 10:52:08AM -0300, Victor Nogueira wrote:

On 8/11/25 06:53, Victor Nogueira wrote:

quoted

On 8/11/25 04:49, Davide Caratti wrote:

quoted

Maybe it's better to extend sch_ets.sh from net/forwarding instead?
If so, I can follow-up on net-next with a patch that adds a new
test-case that includes the 3-lines in [1] - while this patch can go
as-is in 'net' (and eventually in stable). In alternative, I can
investigate on TDC adding "sch_plug" to the qdisc tree in a way
that DWRR never deplete, and the crash would then happen with
"verifyCmd".

WDYT?

That works for me as well.

Sorry, should've been more specific.
I meant that the net/forwarding approach you suggested
seems ok. The tdc approach would be a lot of work and
I don't believe it's worth it.

I was more of the idea of avoiding a non-deterministic kselftest, because
with mausezahn running in the background we have to be "lucky" enough to
see the tc qdisc change command executed while the packet socket is
still emitting packets. And the sch_plug approach I mentioned this morning
looks doable: I just reproduced the NULL dereference on unpatched kernel
using something like:

 # ip link add name ddd0 type dummy
 # tc qdisc add dev ddd0 root handle 1: ets bands 4 strict 2 priomap 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
 # tc qdisc add dev ddd0 handle 10: parent 1:4 plug 
 # ip link set dev ddd0 up
 # tc qdisc change dev ddd0 handle 10: plug limit 5
 # mausezahn ddd0 -A 10.10.10.1 -B 10.10.10.2 -c 0 -a own -c 5 00:c1:a0:c1:a0:00 -t udp
 # printf "press enter to crash..."
 # read -r _
 # tc qdisc change dev ddd0 handle 1: ets bands 2 strict 0

so, including "plug" children in the tree should make kselftest feasible either with 'net/forwarding'
or with TDC + scapy plugin superpowers.

@Victor + @Jakub, can we apply this patch to 'net', so that regression is fixed ASAP, and then I post
the kselftest in a separate submission for net-next?

-- 
davide