[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20231214100154.2298fc8d@kernel.org>
Date: Thu, 14 Dec 2023 10:01:54 -0800
From: Jakub Kicinski <kuba@...nel.org>
To: Jamal Hadi Salim <jhs@...atatu.com>
Cc: Jamal Hadi Salim <hadi@...atatu.com>, Eric Dumazet
<edumazet@...gle.com>, Victor Nogueira <victor@...atatu.com>,
xiyou.wangcong@...il.com, jiri@...nulli.us, davem@...emloft.net,
pabeni@...hat.com, daniel@...earbox.net, dcaratti@...hat.com,
netdev@...r.kernel.org, kernel@...atatu.com, Taehee Yoo
<ap420073@...il.com>
Subject: Re: [PATCH net-next v3 2/3] net: sched: Make tc-related drop reason
more flexible for remaining qdiscs
On Thu, 14 Dec 2023 10:31:24 -0500 Jamal Hadi Salim wrote:
> > I'm still a bit confused about the cb, tho.
> >
> > struct qdisc_skb_cb is the state struct.
>
> Per packet state within tc though, no? Once it leaves tc whatever sits
> in that space cant be trusted to be valid.
> To answer your earlier question tcf_result is not available at the
> qdisc level (when we call free_skb_list() but cb is and thats why we
> used it)
>
> > But we put the drop reason in struct tc_skb_cb.
> > How does that work. Qdiscs will assume they own all of
> > qdisc_skb_cb::data ?
> >
>
> Short answer, yes. Anyone can scribble over that. And multiple
> consumers have a food fight going on - but it is expected behavior:
> ebpf's skb->cb, cake, fq_codel etc - all use qdisc_skb_cb::data.
> Out of the 48B in skb->cb qdisc_skb_cb redefined the first 28B and
> left in qdisc_skb_cb::data as free-for-all space. I think,
> unfortunately, that is now cast in stone.
> Which still leaves us 20 bytes which is now being squatered by
> tc_skb_cb where the drop reason was placed.
Okay I got it now, for some reason I thought the new fields in
struct tc_skb_cb overlay the data[] in struct qdisc_skb_cb...
Powered by blists - more mailing lists