On Wed, Mar 02, 2022 at 03:41:59PM -0800, Eric Dumazet wrote:
> On Wed, Mar 2, 2022 at 2:34 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
> >
> > On Wed, Mar 02, 2022 at 12:30:14PM -0800, Eric Dumazet wrote:
> > > On Wed, Mar 2, 2022 at 11:56 AM Martin KaFai Lau <kafai@xxxxxx> wrote:
> > > >
> > > > The previous patches handled the delivery_time in the ingress path
> > > > before the routing decision is made. This patch can postpone clearing
> > > > delivery_time in a skb until knowing it is delivered locally and also
> > > > set the (rcv) timestamp if needed. This patch moves the
> > > > skb_clear_delivery_time() from dev.c to ip_local_deliver_finish()
> > > > and ip6_input_finish().
> > > >
> > > > Signed-off-by: Martin KaFai Lau <kafai@xxxxxx>
> > > > ---
> > > > net/core/dev.c | 8 ++------
> > > > net/ipv4/ip_input.c | 1 +
> > > > net/ipv6/ip6_input.c | 1 +
> > > > 3 files changed, 4 insertions(+), 6 deletions(-)
> > > >
> > > > diff --git a/net/core/dev.c b/net/core/dev.c
> > > > index 0fc02cf32476..3ff686cc8c84 100644
> > > > --- a/net/core/dev.c
> > > > +++ b/net/core/dev.c
> > > > @@ -5193,10 +5193,8 @@ static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
> > > > goto out;
> > > > }
> > > >
> > > > - if (skb_skip_tc_classify(skb)) {
> > > > - skb_clear_delivery_time(skb);
> > > > + if (skb_skip_tc_classify(skb))
> > > > goto skip_classify;
> > > > - }
> > > >
> > > > if (pfmemalloc)
> > > > goto skip_taps;
> > > > @@ -5225,14 +5223,12 @@ static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
> > > > goto another_round;
> > > > if (!skb)
> > > > goto out;
> > > > - skb_clear_delivery_time(skb);
> > > >
> > > > nf_skip_egress(skb, false);
> > > > if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
> > > > goto out;
> > > > - } else
> > > > + }
> > > > #endif
> > > > - skb_clear_delivery_time(skb);
> > > > skb_reset_redirect(skb);
> > > > skip_classify:
> > > > if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
> > > > diff --git a/net/ipv4/ip_input.c b/net/ipv4/ip_input.c
> > > > index d94f9f7e60c3..95f7bb052784 100644
> > > > --- a/net/ipv4/ip_input.c
> > > > +++ b/net/ipv4/ip_input.c
> > > > @@ -226,6 +226,7 @@ void ip_protocol_deliver_rcu(struct net *net, struct sk_buff *skb, int protocol)
> > > >
> > > > static int ip_local_deliver_finish(struct net *net, struct sock *sk, struct sk_buff *skb)
> > > > {
> > > > + skb_clear_delivery_time(skb);
> > > > __skb_pull(skb, skb_network_header_len(skb));
> > > >
> > > > rcu_read_lock();
> > > > diff --git a/net/ipv6/ip6_input.c b/net/ipv6/ip6_input.c
> > > > index d4b1e2c5aa76..5b5ea35635f9 100644
> > > > --- a/net/ipv6/ip6_input.c
> > > > +++ b/net/ipv6/ip6_input.c
> > > > @@ -459,6 +459,7 @@ void ip6_protocol_deliver_rcu(struct net *net, struct sk_buff *skb, int nexthdr,
> > > >
> > > > static int ip6_input_finish(struct net *net, struct sock *sk, struct sk_buff *skb)
> > > > {
> > > > + skb_clear_delivery_time(skb);
> > > > rcu_read_lock();
> > > > ip6_protocol_deliver_rcu(net, skb, 0, false);
> > > > rcu_read_unlock();
> > > > --
> > > > 2.30.2
> > > >
> > >
> > > It is not clear to me why we need to clear tstamp if packet is locally
> > > delivered ?
> > It does not clear the rx tstamp in skb->tstamp.
> >
> > It only clears the EDT in skb->tstamp when the skb
> > is transmitted out of a local tcp_sock and then loop back from egress
> > to ingress through virtual interface like veth.
> >
> > >
> > > TCP stack is using tstamp for incoming packets (look for
> > > TCP_SKB_CB(skb)->has_rxtstamp)
> > skb_clear_delivery_time() will put ktime_get_real() back to skb->tstamp
> > so that the receiving tcp_sock can use it.
>
>
> Oh... I had to look at
>
> +static inline void skb_clear_delivery_time(struct sk_buff *skb)
> +{
> + if (skb->mono_delivery_time) {
> + skb->mono_delivery_time = 0;
> + if (static_branch_unlikely(&netstamp_needed_key))
> + skb->tstamp = ktime_get_real();
> + else
> + skb->tstamp = 0;
> + }
> +}
>
> Name was a bit confusing :)
A few names were attempted in the early version and
then concluded on delivery_time. :p
>
> And it seems you have a big opportunity to not call ktime_get_real()
> when skb->sk is known at this point (early demux)
> because few sockets actually enable timestamping ?
iiuc, you are suggesting to also check the skb->sk (if early demux)
and check for SK_FLAGS_TIMESTAMP.
Without checking skb->sk here, it should not be worse than the
current ktime_get_real() done in dev.c where it also does not have sk
available? netstamp_needed_key should have been enabled as
long as there is one sk asked for it.
