[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <CANn89iKgX=Ci93U1xpRm3P_3kjPsV_AnL_we6FM6mo8B+kTw9Q@mail.gmail.com>
Date: Wed, 7 Feb 2024 14:37:14 +0100
From: Eric Dumazet <edumazet@...gle.com>
To: Jason Xing <kerneljasonxing@...il.com>
Cc: davem@...emloft.net, kuba@...nel.org, pabeni@...hat.com,
dsahern@...nel.org, netdev@...r.kernel.org,
Jason Xing <kernelxing@...cent.com>
Subject: Re: [PATCH net-next 0/2] add more drop reasons in tcp receive path
On Wed, Feb 7, 2024 at 2:25 PM Jason Xing <kerneljasonxing@...il.com> wrote:
>
> Indeed.
>
> It took me some while to consider whether I should add more reasons
> into the fast path.
>
> But for now the NOT_SPECIFIED fake reason does not work if we really
> want to know some useful hints.
> What do you think? Should I give up this patch series or come up with
> other better ideas?
Perhaps find a way to reuse return values from functions to carry a drop_reason.
>
> >
> > Please make sure not to slow down the TCP fast path, while we work
> > hard in the opposite direction.
>
> I tested some times by using netperf, it's not that easy to observe
> the obvious differences before/after this patch applied.
Sure, the difference is only noticeable on moderate load, when a cpu
receives one packet in a while.
icache pressure, something hard to measure with synthetic benchmarks,
but visible in real workloads in the long term.
At Google, we definitely see an increase of network cpu costs releases
after releases.
Powered by blists - more mailing lists