On Thu, 2017-11-30 at 23:14 -0800, Cong Wang wrote: > On Wed, Nov 29, 2017 at 6:25 AM, Paolo Abeni <pab...@redhat.com> wrote: > > Currently deleting qdisc with a large number of children and filters > > can take a lot of time: > > > > tc qdisc add dev lo root htb > > for I in `seq 1 1000`; do > > tc class add dev lo parent 1: classid 1:$I htb rate 100kbit > > tc qdisc add dev lo parent 1:$I handle $((I + 1)): htb > > for J in `seq 1 10`; do > > tc filter add dev lo parent $((I + 1)): u32 match ip src > > 1.1.1.$J > > done > > done > > time tc qdisc del dev root > > > > real 0m54.764s > > user 0m0.023s > > sys 0m0.000s > > > > This is due to the multiple rcu_barrier() calls, one for each tcf_block > > freed, invoked with the rtnl lock held. Most other network related > > tasks will block in this timespan. > > Yeah, Eric pointed out this too and I already had an idea to cure > this. > > As I already mentioned before, my idea is to refcnt the tcf block > so that we don't need to worry about which is the last. Something > like the attached patch below, note it is PoC _only_, not even > compiled yet. And I am not 100% sure it works either, I will look > deeper tomorrow.
Thank you for the feedback. I tested your patch and in the above scenario I measure: real 0m0.017s user 0m0.000s sys 0m0.017s so it apparently works well for this case. We could still have a storm of rtnl lock/unlock operations while deleting a large tc tree with lot of filters, and I think we can reduce them with bulk free, evenutally applying it to filters, too. That will also reduce the pressure on the rtnl lock when e.g. OVS H/W offload pushes a lot of rules/sec. WDYT? Cheers, Paolo
