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: <74c6f54cd3869258f4c83b46d9e5b95f7f0dab4b.1656878516.git.cdleonard@gmail.com>
 <248071bc915140d8c58669b288c15c731407fa76.camel@redhat.com> <324c9844-1ecb-60c0-c976-16627dff1815@gmail.com>
In-Reply-To: <324c9844-1ecb-60c0-c976-16627dff1815@gmail.com>
From:   Soheil Hassas Yeganeh <soheil@google.com>
Date:   Tue, 5 Jul 2022 18:26:42 -0400
Message-ID: <CACSApvZoi1sMHOfJ+a9fen53NJ=O7GNayeWWUCghAkNiT1YUHA@mail.gmail.com>
Subject: Re: [PATCH] net: Shrink sock.sk_err sk_err_soft to u16 from int
To:     Leonard Crestez <cdleonard@gmail.com>
Cc:     Paolo Abeni <pabeni@redhat.com>,
        "David S. Miller" <davem@davemloft.net>,
        Eric Dumazet <edumazet@google.com>,
        Jakub Kicinski <kuba@kernel.org>, Wei Wang <weiwan@google.com>,
        Joanne Koong <joannelkoong@gmail.com>,
        netdev <netdev@vger.kernel.org>,
        linux-kernel <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Jul 5, 2022 at 9:01 AM Leonard Crestez <cdleonard@gmail.com> wrote:
>
> On 7/5/22 13:31, Paolo Abeni wrote:
> > On Sun, 2022-07-03 at 23:06 +0300, Leonard Crestez wrote:
> >> These fields hold positive errno values which are limited by
> >> ERRNO_MAX=4095 so 16 bits is more than enough.
> >>
> >> They are also always positive; setting them to a negative errno value
> >> can result in falsely reporting a successful read/write of incorrect
> >> size.
> >>
> >> Signed-off-by: Leonard Crestez <cdleonard@gmail.com>
> >> ---
> >>   include/net/sock.h | 2 +-
> >>   1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> I ran some relatively complex tests without noticing issues but some corner
> >> case where this breaks might exist.
> >
> > Could you please explain in length the rationale behind this change?
> >
> > Note that this additionally changes the struct sock binary layout,
> > which in turn in quite relevant for high speed data transfer.
>
> The rationale is that shrinking structs is almost always better. I know
> that due to various roundings it likely won't actually impact memory
> consumption unless accumulated with other size reductions.
>
> These sk_err fields don't seem to be in a particularly "hot" area so I
> don't think it will impact performance.
>
> My expectation is that after a socket error is reported the socket will
> likely be closed so that there will be very few writes to this field.

Since you're packing sk_err and sk_err_soft into a DWORD, I'd suggest
adding another patch on top to move both fields right before sk_filter
where we have a 4-byte hole. As far as I can tell, this should save
one QWORD from "struct sock".

Eric, I believe these fields are read-mostly and that wouldn't infer
with your previous layout optimizations. Is my understanding correct?

Thanks,
Soheil

> --
> Regards,
> Leonard