[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <20250722033310.geuc34ln2ie55zqq@gmail.com>
Date: Mon, 21 Jul 2025 20:33:10 -0700
From: John Fastabend <john.fastabend@...il.com>
To: Jakub Kicinski <kuba@...nel.org>
Cc: Vincent Whitchurch via B4 Relay <devnull+vincent.whitchurch.datadoghq.com@...nel.org>,
vincent.whitchurch@...adoghq.com,
Jakub Sitnicki <jakub@...udflare.com>,
Kuniyuki Iwashima <kuniyu@...gle.com>, netdev@...r.kernel.org,
bpf@...r.kernel.org
Subject: Re: [PATCH bpf-next v3 0/5] sockmap: Fix reading with splice(2)
On 2025-07-11 16:09:31, Jakub Kicinski wrote:
> On Wed, 09 Jul 2025 14:47:56 +0200 Vincent Whitchurch via B4 Relay
> wrote:
> > I noticed that if the verdict callback returns SK_PASS, using splice(2)
> > to read from a socket in a sockmap does not work since it never sees the
> > data queued on to it. As far as I can see, this is not a regression but
> > just something that has never worked, but it does make sockmap unusable
> > if you can't guarantee that the programs using the socket will not use
> > splice(2).
>
> On v2 you should you can't replace ops for passively opened
> connections. Can that not be addressed instead of adding
> an indirect call on the data path?
I guess I missed above? We can create the psock and add the socket
to a map on the accept()?
It would be good to get some fix for this.
Thanks,
John
Powered by blists - more mailing lists