Hi,
On Mon, 2019-04-08 at 14:17 -0700, Eric Dumazet wrote:
> On 04/08/2019 09:35 AM, Paolo Abeni wrote:
> > This revert commit 46b1c18f9deb ("net: sched: put back q.qlen
> > into a single location").
> > After the previous patch nobody accesses directly qlen for a child
> > qdisc when such qdisc does per CPU stats accounting.
> > In the control path nobody uses directly qlen since commit
> > 677f1bc207c ("net: sched: introduce and use qdisc tree flush/purge
> > helpers"), so we can remove the contented atomic ops from the
> > datapath.
> >
>
> Have you tested HTB with a pfifo_fast on a throttled class ?
>
> I do not see any changes in HTB in your patch series, so it is not
> clear why your patch series is not adding back the issue.
Thank you for the feedback.
I tested this series enslaving pfifo_fast to each classful qdiscs -
including HTB - sending traffic through the pfifo_fast qdisc, and
checking correct accounting.
When pfifo_fast is enslaved to HTB, the NOLOCK flag is cleared - by
qdisc_graft(), as HTB is a lock qdisc. As per patch 4/5, TCQ_F_CPUSTATS
is cleared, too, so pfifo_fast switches to global accounting, under
root lock protection.
In HTB context, <child pfifo_fast>->q.qlen should be always valid, no
changes required there, nor to any other classful qdisc - until we will
have lockless classful qdiscs: they should not access <child>->q.qlen
directly.
Please let me know if the above is somewhat clear. Have I missed
something?
Thanks,
Paolo
