On Sun, May 24, 2020 at 3:04 AM Václav Zindulka
[off-list ref] wrote:

On Tue, May 19, 2020 at 7:57 PM Cong Wang [off-list ref] wrote:

quoted

On Tue, May 19, 2020 at 1:04 AM Václav Zindulka
[off-list ref] wrote:

quoted

quoted

Let me think how to fix this properly, I have some ideas and will provide
you some patch(es) to test soon.

Sure, I'll wait. I have plenty of time now with the main problem fixed :-)

Can you help to test the patches below?
https://github.com/congwang/linux/commits/qdisc_reset2

I removed the last patch you previously tested and added two
more patches on top. These two patches should get rid of most
of the unnecessary resets. I tested them with veth pair with 8 TX
queues, but I don't have any real physical NIC to test.

Thanks.

I've tested it and I have to admit that your Kung-fu is even better
than mine :-D My large ruleset with over 13k qdiscs defined got from
22s to 520ms. I've tested downtime of interface during deletion of
root qdisc and it corresponds to the time I measured so it is great.

/usr/bin/time -p tc qdisc del dev enp1s0f0np0 root
real 0.52
user 0.00
sys 0.52

I've even added my patch for only active queues but it didn't make any
difference from your awesome patch :-)

Great! I will send out the patches formally with your Tested-by.

Thank you very much. You helped me really a lot.

Thanks for testing!