Re: [PATCH net-next 1/2] tcp/dccp: Move initialisation of refcounted

From:   Eric Dumazet <[email protected]>
Date:   Mon, 23 Mar 2020 11:45:48 -0700
> On Mon, Mar 23, 2020 at 11:18 AM Kuniyuki Iwashima <[email protected]> wrote:
> >
> > The refcounted is initialised at most three times.
> >
> >   - __inet_lookup_skb() sets it true.
> >   - skb_steal_sock() is false and __inet_lookup() sets it true.
> >   - __inet_lookup_established() is false and __inet_lookup() sets it false.
> >
> > We do not need to initialise refcounted again and again, so we should do
> > it just before return.
> >
> > Signed-off-by: Kuniyuki Iwashima <[email protected]>
> > ---
> 
> Well, I do not believe this patch (and the following one) makes things
> more readable.
> 
> I doubt setting a boolean in a register or a stack variable has any cost,
> I prefer letting the compiler optimize this.
> 
> The ehash lookup cost is at least 2 or 3 cache lines, this is the
> major contribution.

I confirmed that GCC (v7.3.1) does not make any change with this patch,
but the original code is bit confusing. I think this patch makes the code
bit more readable so that someone (like me...) will not doubt about
complier optimization and write patches.

I will respin only this patch because the changelog can be incorrect by
complier optimization.

Thanks.