[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20161118.134304.1574614508508170507.davem@davemloft.net>
Date: Fri, 18 Nov 2016 13:43:04 -0500 (EST)
From: David Miller <davem@...emloft.net>
To: fw@...len.de
Cc: netdev@...r.kernel.org, edumazet@...gle.com, ycheng@...gle.com,
ncardwell@...gle.com
Subject: Re: [PATCH -next] tcp: make undo_cwnd mandatory for congestion
modules
From: Florian Westphal <fw@...len.de>
Date: Thu, 17 Nov 2016 13:56:51 +0100
> The undo_cwnd fallback in the stack doubles cwnd based on ssthresh,
> which un-does reno halving behaviour.
>
> It seems more appropriate to let congctl algorithms pair .ssthresh
> and .undo_cwnd properly. Add a 'tcp_reno_undo_cwnd' function and wire it
> up for all congestion algorithms that used to rely on the fallback.
>
> highspeed, illinois, scalable, veno and yeah use 'reno undo' while their
> .ssthresh implementation doesn't halve the slowstart threshold, this
> might point to similar issue as the one fixed for dctcp in
> ce6dd23329b1e ("dctcp: avoid bogus doubling of cwnd after loss").
>
> Cc: Eric Dumazet <edumazet@...gle.com>
> Cc: Yuchung Cheng <ycheng@...gle.com>
> Cc: Neal Cardwell <ncardwell@...gle.com>
> Signed-off-by: Florian Westphal <fw@...len.de>
If you really suspect that highspeed et al. need to implement their own
undo_cwnd instead of using the default reno fallback, I would really
rather that this gets either fixed or explicitly marked as likely wrong
(in an "XXX" comment or similar).
Otherwise nobody is going to remember this down the road.
Powered by blists - more mailing lists