Re: [PATCH] net: fec: fix rxvlan feature

[Michael Grzeschik <mgr@pengutronix.de>]

On Tue, Mar 10, 2015 at 03:29:35AM +0000, fugang.duan@freescale.com wrote:
> From: David Miller <davem@davemloft.net> Sent: Tuesday, March 10, 2015 10:43 AM
> > To: m.grzeschik@pengutronix.de
> > Cc: Duan Fugang-B38611; Li Frank-B20596; netdev@vger.kernel.org;
> > kernel@pengutronix.de
> > Subject: Re: [PATCH] net: fec: fix rxvlan feature
> > 
> > From: Michael Grzeschik <m.grzeschik@pengutronix.de>
> > Date: Mon,  9 Mar 2015 17:55:53 +0100
> > 
> > > The patch 1b7bde6d659d30f171259cc2dfba8e5dab34e735
> > > "net: fec: implement rx_copybreak to improve rx performance"
> > > changed the code path for the vlan check in fec_enet_rx_queue:
> > >
> > >  @@ -1417,62 +1486,48 @@ fec_enet_rx_queue(struct net_device *ndev, int
> > budget, u16 queue_id)
> > >  		/* If this is a VLAN packet remove the VLAN Tag */
> > >  		vlan_packet_rcvd = false;
> > >  		if ((ndev->features & NETIF_F_HW_VLAN_CTAG_RX) &&
> > >  			fep->bufdesc_ex && (ebdp->cbd_esc & BD_ENET_RX_VLAN)) {
> > >  			/* Push and remove the vlan tag */
> > >                          struct vlan_hdr *vlan_header =
> > >                                          (struct vlan_hdr *) (data +
> > ETH_HLEN);
> > >                          vlan_tag = ntohs(vlan_header->h_vlan_TCI);
> > >  -                       pkt_len -= VLAN_HLEN;
> > >
> > >                          vlan_packet_rcvd = true;  +
> > >  +                       skb_copy_to_linear_data_offset(skb, VLAN_HLEN,
> > >  +                                                      data, (2 *
> > ETH_ALEN));
> > >  +                       skb_pull(skb, VLAN_HLEN);
> > >                  }
> > >
> > > With the call of skb_copy_to_linear_data_offset the code here is doing
> > > more than previously and is breaking the rxvlan feature. This patch
> > > removes this call to fix it.
> > >
> > > Signed-off-by: Michael Grzeschik <m.grzeschik@pengutronix.de>
> > 
> > But don't we want to copy the proper header there in the linear case?
> > 
> > I'd rather hear from the original author why they put that copy there
> > before it just gets blindly removed.
> > 
> > And you'll need to update your commit message with more explanation once
> > that discussion occurs.
> 
> There have no function change comparing with previous commit version.
> If enable hw VLAN support (sw simulate hw), software remove VLAN tag. 
> 
> Below code is previous code base, it also removes the VLAN tag.
> -                       /* Extract the frame data without the VLAN header. */
> -                       skb_copy_to_linear_data(skb, data, (2 * ETH_ALEN));
> -                       if (vlan_packet_rcvd)
> -                               payload_offset = (2 * ETH_ALEN) + VLAN_HLEN;
> -                       skb_copy_to_linear_data_offset(skb, (2 * ETH_ALEN),
> -                                                      data + payload_offset,
> -                                                      pkt_len - 4 - (2 * ETH_ALEN));
> 
> 
> Does there have un-correct points ? I did run some VLAN test pass.

Yes, in the previous version the skb_copy_... functions were working
on two different memory pointers. The code was calling netdev_alloc_skb
to create the new memory.

Currently the code is calling this:

data = skb->data;

which leads the skb_copy_... operations, which are no more then memcpy,
to work on the same memory. This can not be valid.

Michael

-- 
Pengutronix e.K.                           |                             |
Industrial Linux Solutions                 | http://www.pengutronix.de/  |
Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |