Wed, Nov 01, 2017 at 03:12:50AM CET, alexei.starovoi...@gmail.com wrote:
>On Tue, Oct 31, 2017 at 04:12:22PM +0100, Jiri Pirko wrote:
>> From: Jiri Pirko <j...@mellanox.com>
>>
>> In sch_handle_egress and sch_handle_ingress tp->q is used only in order
>> to update stats. So stats and filter list are the only things that are
>> needed in clsact qdisc fastpath processing. Introduce new mini_Qdisc
>> struct to hold those items. Also, introduce a helper to swap the
>> mini_Qdisc structures in case filter list head changes.
>>
>> This removes need for tp->q usage without added overhead.
>>
>> Signed-off-by: Jiri Pirko <j...@mellanox.com>
>> ---
>> v2->v3:
>> - Using head change callback to replace miniq pointer every time tp head
>> changes. This eliminates one rcu dereference and makes the claim "without
>> added overhead" valid.
>
>you kidding, right?
>It's still two loads.
I'm not.
I replace:
one rcu_dereference_bh(dev->egress_cl_list)
by one rcu_dereference_bh(dev->miniq_egress)
one dereference cl->q
by one dereference miniq->filter_list
What do I miss?
>
>> diff --git a/net/core/dev.c b/net/core/dev.c
>> index 24ac908..1423cf4 100644
>> --- a/net/core/dev.c
>> +++ b/net/core/dev.c
>> @@ -3274,22 +3274,22 @@ EXPORT_SYMBOL(dev_loopback_xmit);
>> static struct sk_buff *
>> sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
>> {
>> - struct tcf_proto *cl = rcu_dereference_bh(dev->egress_cl_list);
>> + struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
>> struct tcf_result cl_res;
>>
>> - if (!cl)
>> + if (!miniq)
>> return skb;
>>
>> /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
>> - qdisc_bstats_cpu_update(cl->q, skb);
>> + mini_qdisc_bstats_cpu_update(miniq, skb);
>>
>> - switch (tcf_classify(skb, cl, &cl_res, false)) {
>> + switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
>
>I don't think it's great, but I don't have any suggestions on
>how to avoid it, so I'm not objecting. Just disappointed that
>you keep adding stuff to tc and messing with sw fast path only to
>make parity with some obscure hw feature.
>If it keeps going like this we'd need to come up with some new fast
>hook for clsbpf in ingress/egress paths. We use it for
>every packet, so extra loads are not great.
>I guess they should be cache hits, but will take extra cache line.
>All of the bugs in tc logic recently are not comforting either.
>
