Re: [PATCH net-next 2/2] net: optimize skb_postpull_rcsum()

From: Vladimir Oltean <olteanv@gmail.com>
To: Eric Dumazet <edumazet@google.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>,
	"David S . Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>, netdev <netdev@vger.kernel.org>,
	David Laight <David.Laight@aculab.com>
Subject: Re: [PATCH net-next 2/2] net: optimize skb_postpull_rcsum()
Date: Thu, 2 Dec 2021 23:07:05 +0200	[thread overview]
Message-ID: <20211202210705.vsqr67svt3sm53h3@skbuf> (raw)
In-Reply-To: <CANn89iJOH0QtRhDfBzJr3fpJPNCQJhbMqT_8sa+vH_6mmZ7xhw@mail.gmail.com>

On Thu, Dec 02, 2021 at 12:37:35PM -0800, Eric Dumazet wrote:
> On Thu, Dec 2, 2021 at 11:32 AM Eric Dumazet <edumazet@google.com> wrote:
> >
> > Thanks Vladimir
> >
> > I think that maybe the issue is that the initial skb->csum is zero,
> > and the csum_parttial(removed_block) is also zero.
> >
> > But the initial skb->csum should not be zero if you have a non " all
> > zero"  frame.
> >
> > Can you double check this in drivers/net/ethernet/freescale/enetc/enetc.c ?
> 
> Yes, I am not sure why the csum is inverted in enetc_get_offloads()

I guess it's inverted because the hardware doesn't provide its one's
complement.

> Perhaps
> 
> diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c
> b/drivers/net/ethernet/freescale/enetc/enetc.c
> index 504e12554079e306e477b9619f272d6e96527377..72524f14cae0093763f8a3f57b1e08e31bc4df1a
> 100644
> --- a/drivers/net/ethernet/freescale/enetc/enetc.c
> +++ b/drivers/net/ethernet/freescale/enetc/enetc.c
> @@ -986,7 +986,7 @@ static void enetc_get_offloads(struct enetc_bdr *rx_ring,
>         if (rx_ring->ndev->features & NETIF_F_RXCSUM) {
>                 u16 inet_csum = le16_to_cpu(rxbd->r.inet_csum);
> 
> -               skb->csum = csum_unfold((__force __sum16)~htons(inet_csum));
> +               skb->csum = csum_unfold((__force __sum16)htons(inet_csum));
>                 skb->ip_summed = CHECKSUM_COMPLETE;
>         }
> 
> If this does not work, then maybe :
> 
> diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c
> b/drivers/net/ethernet/freescale/enetc/enetc.c
> index 504e12554079e306e477b9619f272d6e96527377..d190faa9a8242f9f3f962dd30b9f4409a83ee697
> 100644
> --- a/drivers/net/ethernet/freescale/enetc/enetc.c
> +++ b/drivers/net/ethernet/freescale/enetc/enetc.c
> @@ -987,7 +987,8 @@ static void enetc_get_offloads(struct enetc_bdr *rx_ring,
>                 u16 inet_csum = le16_to_cpu(rxbd->r.inet_csum);
> 
>                 skb->csum = csum_unfold((__force __sum16)~htons(inet_csum));
> -               skb->ip_summed = CHECKSUM_COMPLETE;
> +               if (likely(skb->csum))
> +                       skb->ip_summed = CHECKSUM_COMPLETE;
>         }
> 
>         if (le16_to_cpu(rxbd->r.flags) & ENETC_RXBD_FLAG_VLAN) {

I guess you aren't interested any longer in the result of these changes,
since the csum isn't zero in enetc?