[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Wed, 2 Mar 2022 16:47:17 -0800
From: Eric Dumazet <edumazet@...gle.com>
To: Martin KaFai Lau <kafai@...com>
Cc: bpf <bpf@...r.kernel.org>, netdev <netdev@...r.kernel.org>,
Alexei Starovoitov <ast@...nel.org>,
Andrii Nakryiko <andrii@...nel.org>,
Daniel Borkmann <daniel@...earbox.net>,
David Miller <davem@...emloft.net>,
Jakub Kicinski <kuba@...nel.org>,
kernel-team <kernel-team@...com>,
Willem de Bruijn <willemb@...gle.com>
Subject: Re: [PATCH v6 net-next 10/13] net: Postpone skb_clear_delivery_time()
until knowing the skb is delivered locally
On Wed, Mar 2, 2022 at 4:20 PM Martin KaFai Lau <kafai@...com> wrote:
>
> On Wed, Mar 02, 2022 at 03:41:59PM -0800, Eric Dumazet wrote:
> > 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.
Yes, but typically there is often one active timestamp consumer,
enabling the static key.
This is a corner case anyway, not sure if hosts running VM would have
slow ktime_get_real()
Powered by blists - more mailing lists