Re: [PATCH net-next v2 3/3] tcp_cubic: fix divide by zero and endless loop on bad module params

[Neal Cardwell]

On Wed, Oct 7, 2026 at 2:49 AM Jiayuan Chen <jiayuan.chen@xxxxxxxxx> wrote:
>
>
> On 10/6/26 8:12 AM, Jakub Kicinski wrote:
> > On Thu,  1 Oct 2026 19:29:43 +0800 Jiayuan Chen wrote:
> >> Subject: [PATCH net-next v2 3/3] tcp_cubic: fix divide by zero and endless loop on bad module params
> > if the LLM Medium severity complaint is true maybe also update
> > the subject? I thought there's an actual div by zero somewhere
> > but the commit msg seems to describe the inf loop
>
>
>
> Hi Kuba, sorry for the delay.
>
> The LLM is talking about (cwnd * beta_scale) wrapping, which needs
> cwnd >= 33M packets, way past the 1M limit this code documents. So
> the loop fix is fine.
>
> The div by zero is real: beta == 1024 in "/ (BICTCP_BETA_SCALE - beta)"
> and bic_scale == 0 in do_div(), both at init. The new range check in
> this patch rejects both, I just did not say "divide by zero" in the
> commit message.
>
> Want me to spell this out in the commit message? I can also add the
> READ_ONCE() Eric mentioned in the same respin.

Yes, please spell that out in the commit message, and please also add
the READ_ONCE() Eric mentioned in the same respin.

Thanks,
neal