Re: [PATCH net-next 1/2] net/sched: sch_ingress: Support clsact egress mini-Qdisc option

On Tue, Aug 3, 2021 at 1:08 AM Daniel Borkmann [off-list ref] wrote:

On 8/3/21 2:08 AM, Cong Wang wrote:

quoted

On Mon, Aug 2, 2021 at 2:11 PM Daniel Borkmann [off-list ref] wrote:

quoted

NAK, just use clsact qdisc in the first place which has both ingress and egress
support instead of adding such hack. You already need to change your scripts for
clsact-on, so just swap 'tc qdisc add dev eth0 ingress' to 'tc qdisc add dev eth0
clsact' w/o needing to change kernel.

If we were able to change the "script" as easily as you described,
you would not even see such a patch. The fact is it is not under
our control, the most we can do is change the qdisc after it is
created by the "script", ideally without interfering its traffic,
hence we have such a patch.

(BTW, it is actually not a script, it is a cloud platform.)

Sigh, so you're trying to solve a non-technical issue with one cloud provider by
taking a detour for unnecessarily extending the kernel instead with functionality
that already exists in another qdisc (and potentially waiting few years until they
eventually upgrade). I presume Bytedance should be a big enough entity to make a
case for that provider to change it. After all swapping ingress with clsact for
such script is completely transparent and there is nothing that would break. (Fwiw,
from all the major cloud providers we have never seen such issue in our deployments.)

Well, it is both non-technical and technical at the same time.

The non-technical part is that it is really hard to convince people from
other team to restart their services just for a kernel change, people are just
not happy to take risks.

The technical part is the bad design of clsact. It is too late to complain,
but it should not create two _conceptual_  qdiscs (actually just one struct
Qdisc) at the same time. If it only created just egress, we would
not even bother changing ingress at all. Sigh.

Thanks.