Re: [PATCH net-next 4/4] gro: improve flow distribution across GRO buckets in dev_gro_receive()

From: Alexander Lobakin <alobakin@pm.me>
To: Eric Dumazet <edumazet@google.com>
Cc: Alexander Lobakin <alobakin@pm.me>,
	"David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Andrii Nakryiko <andriin@fb.com>, Wei Wang <weiwan@google.com>,
	Cong Wang <cong.wang@bytedance.com>,
	Taehee Yoo <ap420073@gmail.com>, netdev <netdev@vger.kernel.org>,
	LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH net-next 4/4] gro: improve flow distribution across GRO buckets in dev_gro_receive()
Date: Fri, 12 Mar 2021 18:28:12 +0000	[thread overview]
Message-ID: <20210312182754.241807-1-alobakin@pm.me> (raw)
In-Reply-To: <CANn89i+T-r=i3GBv-9EWBjpR_NhgZ=vP08BwTGXc8Kw3nO+OEQ@mail.gmail.com>

From: Eric Dumazet <edumazet@google.com>
Date: Fri, 12 Mar 2021 17:33:53 +0100

> On Fri, Mar 12, 2021 at 5:22 PM Alexander Lobakin <alobakin@pm.me> wrote:
> >
> > Most of the functions that "convert" hash value into an index
> > (when RPS is configured / XPS is not configured / etc.) set
> > reciprocal_scale() on it. Its logics is simple, but fair enough and
> > accounts the entire input value.
> > On the opposite side, 'hash & (GRO_HASH_BUCKETS - 1)' expression uses
> > only 3 least significant bits of the value, which is far from
> > optimal (especially for XOR RSS hashers, where the hashes of two
> > different flows may differ only by 1 bit somewhere in the middle).
> >
> > Use reciprocal_scale() here too to take the entire hash value into
> > account and improve flow dispersion between GRO hash buckets.
> >
> > Signed-off-by: Alexander Lobakin <alobakin@pm.me>
> > ---
> >  net/core/dev.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/net/core/dev.c b/net/core/dev.c
> > index 65d9e7d9d1e8..bd7c9ba54623 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -5952,7 +5952,7 @@ static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
> >
> >  static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
> >  {
> > -       u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
> > +       u32 bucket = reciprocal_scale(skb_get_hash_raw(skb), GRO_HASH_BUCKETS);
>
> This is going to use 3 high order bits instead of 3 low-order bits.

We-e-ell, seems like it.

> Now, had you use hash_32(skb_get_hash_raw(skb), 3), you could have
> claimed to use "more bits"

Nice suggestion, I'll try. If there won't be any visible improvements,
I'll just drop this one.

> Toeplitz already shuffles stuff.

As well as CRC and others, but I feel like we shouldn't rely only on
the hardware.

> Adding a multiply here seems not needed.
>
> Please provide experimental results, because this looks unnecessary to me.

Thanks,
Al