Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <4740526.31r3eYUQgx@natalenko.name> <4bd84c983e77486fbc94dfa2a167afaa@AcuMS.aculab.com>
In-Reply-To: <4bd84c983e77486fbc94dfa2a167afaa@AcuMS.aculab.com>
From:   Neal Cardwell <ncardwell@google.com>
Date:   Tue, 24 May 2022 16:05:55 -0400
Message-ID: <CADVnQykt1Lz0m1gfEckHhDLy66xhJvO0F2Z1-yQ=Mgi7gBY5RQ@mail.gmail.com>
Subject: Re: [RFC] tcp_bbr2: use correct 64-bit division
To:     David Laight <David.Laight@aculab.com>
Cc:     Oleksandr Natalenko <oleksandr@natalenko.name>,
        Yuchung Cheng <ycheng@google.com>,
        Yousuk Seung <ysseung@google.com>,
        Soheil Hassas Yeganeh <soheil@google.com>,
        Adithya Abraham Philip <abrahamphilip@google.com>,
        Eric Dumazet <edumazet@google.com>,
        "David S. Miller" <davem@davemloft.net>,
        Hideaki YOSHIFUJI <yoshfuji@linux-ipv6.org>,
        David Ahern <dsahern@kernel.org>,
        Jakub Kicinski <kuba@kernel.org>,
        Paolo Abeni <pabeni@redhat.com>,
        "netdev@vger.kernel.org" <netdev@vger.kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        Konstantin Demin <rockdrilla@gmail.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, May 24, 2022 at 4:01 AM David Laight <David.Laight@aculab.com> wrote:
>
> From: Oleksandr Natalenko
> > Sent: 22 May 2022 23:30
> > To: Neal Cardwell <ncardwell@google.com>
> >
> > Hello Neal.
> >
> > It was reported to me [1] by Konstantin (in Cc) that BBRv2 code suffers from integer division issue on
> > 32 bit systems.
>
> Do any of these divisions ever actually have 64bit operands?
> Even on x86-64 64bit divide is significantly slower than 32bit divide.
>
> It is quite clear that x * 8 / 1000 is the same as x / (1000 / 8).
> So promoting to 64bit cannot be needed.
>
>         David

The sk->sk_pacing_rate can definitely be bigger than 32 bits if the
network path can support more than 34 Gbit/sec  (a pacing rate of 2^32
bytes per sec is roughly 34 Gibt/sec). This definitely happens.

So  this one seems reasonable to me (and is only in debug code, so the
performance is probably fine):
-                (u64)sk->sk_pacing_rate * 8 / 1000,
+                div_u64((u64)sk->sk_pacing_rate * 8, 1000),

For the other two I agree we should rework them to avoid the 64-bit
divide, since we don't need it.

There is similar logic in mainline Linux in tcp_tso_autosize(), which
is currently using "unsigned long" for bytes.

Eric, what do you advise?

thanks,
neal