[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20230330130828.4aa7f911@kernel.org>
Date: Thu, 30 Mar 2023 13:08:28 -0700
From: Jakub Kicinski <kuba@...nel.org>
To: Eric Dumazet <edumazet@...gle.com>
Cc: "David S . Miller" <davem@...emloft.net>,
Paolo Abeni <pabeni@...hat.com>, netdev@...r.kernel.org,
eric.dumazet@...il.com
Subject: Re: [PATCH net-next] tcp: tsq: avoid using a tasklet if possible
On Thu, 30 Mar 2023 19:25:34 +0000 Eric Dumazet wrote:
> Instead of always defer TCP Small Queue handling to a tasklet
> we can try to lock the socket and immediately call tcp_tsq_handler()
>
> In my experiments on a 100Gbit NIC with FQ qdisc, number of times
> tcp_tasklet_func() is fired per second goes from ~70,000 to ~1,000.
>
> This reduces number of ksoftirqd wakeups and thus extra cpu
> costs and latencies.
Oh, you posted already. LMK what you think about other locks,
I reckon we should apply this patch... just maybe around -rc1
to give ourselves plenty of time to find the broken drivers?
Powered by blists - more mailing lists