[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20230530221043.5ae05030@kernel.org>
Date: Tue, 30 May 2023 22:10:43 -0700
From: Jakub Kicinski <kuba@...nel.org>
To: Eric Dumazet <edumazet@...gle.com>
Cc: Kuniyuki Iwashima <kuniyu@...zon.com>, davem@...emloft.net,
dsahern@...nel.org, kuni1840@...il.com, netdev@...r.kernel.org,
pabeni@...hat.com, willemdebruijn.kernel@...il.com
Subject: Re: [PATCH v1 net-next 00/14] udp: Farewell to UDP-Lite.
On Wed, 31 May 2023 06:25:33 +0200 Eric Dumazet wrote:
> > Yes, if it's ok, it would be better to add a WARN_ONCE() to stable.
> >
> > If we added it only in net-next, no one might notice it and we could
> > remove UDP-Lite before the warning is available in the next LTS stable
> > tree.
>
> WARN_ONCE() will fire a syzbot report.
>
> Honestly I do not think UDP-Lite is a significant burden.
>
> What about instead adding a CONFIG_UDPLITE and default it to
> "CONFIG_UDPLITE is not set" ?
>
> And add a static key, with /proc/sys/net/core/udplite_enable to
> eventually save some cycles in various fast paths
> and let the user opt-in, in case it is using a distro kernel. with
> CONFIG_UDPLITE=y
oohm, fair point user-reachable WARN() is a liability.
CONFIG_UDPLITE sounds like the best available option :(
With an appropriately discouraging config text.
That way syzbot can prevent bitrot but distros will hopefully drop it.
> DCCP is more interesting because removing it would allow for a better
> organisation of tcp fields to reduce
> number of cache lines hit in the fast path.
Powered by blists - more mailing lists