gianfar: fix headroom expansion code

The code that was added to increase headroom was wrong.
It doesn't handle the case where gfar_add_fcb() changes the skb.
Better to do check at start of transmit (outside of lock), where
error handling is better anyway.

Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Fri, Mar 27, 2009 at 1:08 AM, Stephen Hemminger
<shemminger@vyatta.com> wrote:
> The code that was added to increase headroom was wrong.
> It doesn't handle the case where gfar_add_fcb() changes the skb.

Oops.  I messed up the pointer to pointer manipulating.

> Better to do check at start of transmit (outside of lock), where
> error handling is better anyway.
>
> Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>
>
>
> --- a/drivers/net/gianfar.c     2009-03-26 09:14:39.273669929 -0700
> +++ b/drivers/net/gianfar.c     2009-03-26 09:22:46.477545004 -0700
> @@ -1239,19 +1239,9 @@ static int gfar_enet_open(struct net_dev
>        return err;
>  }
>
> -static inline struct txfcb *gfar_add_fcb(struct sk_buff **skbp)
> +static inline struct txfcb *gfar_add_fcb(struct sk_buff *skb)
>  {
> -       struct txfcb *fcb;
> -       struct sk_buff *skb = *skbp;
> -
> -       if (unlikely(skb_headroom(skb) < GMAC_FCB_LEN)) {
> -               struct sk_buff *old_skb = skb;
> -               skb = skb_realloc_headroom(old_skb, GMAC_FCB_LEN);
> -               if (!skb)
> -                       return NULL;
> -               dev_kfree_skb_any(old_skb);
> -       }
> -       fcb = (struct txfcb *)skb_push(skb, GMAC_FCB_LEN);
> +       struct txfcb *fcb = (struct txfcb *)skb_push(skb, GMAC_FCB_LEN);
>        cacheable_memzero(fcb, GMAC_FCB_LEN);
>
>        return fcb;
> @@ -1320,6 +1310,20 @@ static int gfar_start_xmit(struct sk_buf
>
>        base = priv->tx_bd_base;
>
> +       /* make space for additional header */
> +       if (skb_headroom(skb) < GMAC_FCB_LEN) {
> +               struct sk_buff *skb_new;
> +
> +               skb_new = skb_realloc_headroom(skb, GMAC_FCB_LEN);
> +               if (!skb_new) {
> +                       dev->stats.tx_errors++;
> +                       kfree(skb);
> +                       return NETDEV_TX_OK;
> +               }
> +               kfree_skb(skb);
> +               skb = skb_new;
> +       }
> +

We have legacy devices without the offloading feature.  And we can
even turn off the IP checksum offloading at runtime.  Your code will
cause unnecessary realloc for these cases.

I can propose a new patch to fix the pointer problem and add more
error handling.

>        /* total number of fragments in the SKB */
>        nr_frags = skb_shinfo(skb)->nr_frags;
>
> @@ -1372,20 +1376,18 @@ static int gfar_start_xmit(struct sk_buf
>
>        /* Set up checksumming */
>        if (CHECKSUM_PARTIAL == skb->ip_summed) {
> -               fcb = gfar_add_fcb(&skb);
> -               if (likely(fcb != NULL)) {
> -                       lstatus |= BD_LFLAG(TXBD_TOE);
> -                       gfar_tx_checksum(skb, fcb);
> -               }
> +               fcb = gfar_add_fcb(skb);
> +               lstatus |= BD_LFLAG(TXBD_TOE);
> +               gfar_tx_checksum(skb, fcb);
>        }
>
>        if (priv->vlgrp && vlan_tx_tag_present(skb)) {
> -               if (unlikely(NULL == fcb))
> -                       fcb = gfar_add_fcb(&skb);
> -               if (likely(fcb != NULL)) {
> +               if (unlikely(NULL == fcb)) {
> +                       fcb = gfar_add_fcb(skb);
>                        lstatus |= BD_LFLAG(TXBD_TOE);
> -                       gfar_tx_vlan(skb, fcb);
>                }
> +
> +               gfar_tx_vlan(skb, fcb);
>        }
>
>        /* setup the TxBD length and buffer pointer for the first BD */
> @@ -1433,7 +1435,7 @@ static int gfar_start_xmit(struct sk_buf
>        /* Unlock priv */
>        spin_unlock_irqrestore(&priv->txlock, flags);
>
> -       return 0;
> +       return NETDEV_TX_OK;
>  }
>
>  /* Stops the kernel queue, and halts the controller */
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

From: Stephen Hemminger <shemminger@vyatta.com>
Date: Thu, 26 Mar 2009 10:08:56 -0700

> The code that was added to increase headroom was wrong.
> It doesn't handle the case where gfar_add_fcb() changes the skb.
> Better to do check at start of transmit (outside of lock), where
> error handling is better anyway.
> 
> Signed-off-by: Stephen Hemminger <shemminger@vyatta.com>

Applied, thanks Stephen.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

From: Li Yang <leoli@freescale.com>
Date: Fri, 27 Mar 2009 12:26:33 +0800

> We have legacy devices without the offloading feature.  And we can
> even turn off the IP checksum offloading at runtime.  Your code will
> cause unnecessary realloc for these cases.
> 
> I can propose a new patch to fix the pointer problem and add more
> error handling.

I'm applying Stephen's patch for now, please send any improvements
relative to that.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

On Fri, Mar 27, 2009 at 3:39 PM, David Miller <davem@davemloft.net> wrote:
> From: Li Yang <leoli@freescale.com>
> Date: Fri, 27 Mar 2009 12:26:33 +0800
>
>> We have legacy devices without the offloading feature.  And we can
>> even turn off the IP checksum offloading at runtime.  Your code will
>> cause unnecessary realloc for these cases.
>>
>> I can propose a new patch to fix the pointer problem and add more
>> error handling.
>
> I'm applying Stephen's patch for now, please send any improvements
> relative to that.

Ok.

However is it ok to use kfree() instead of kfree_skb() in Stephen's patch?

+               skb_new = skb_realloc_headroom(skb, GMAC_FCB_LEN);
+               if (!skb_new) {
+                       dev->stats.tx_errors++;
+                       kfree(skb);
+                       return NETDEV_TX_OK;
+               }

- Leo
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

From: Li Yang <leoli@freescale.com>
Date: Fri, 27 Mar 2009 16:06:27 +0800

> On Fri, Mar 27, 2009 at 3:39 PM, David Miller <davem@davemloft.net> wrote:
> > From: Li Yang <leoli@freescale.com>
> > Date: Fri, 27 Mar 2009 12:26:33 +0800
> >
> >> We have legacy devices without the offloading feature.  And we can
> >> even turn off the IP checksum offloading at runtime.  Your code will
> >> cause unnecessary realloc for these cases.
> >>
> >> I can propose a new patch to fix the pointer problem and add more
> >> error handling.
> >
> > I'm applying Stephen's patch for now, please send any improvements
> > relative to that.
> 
> Ok.
> 
> However is it ok to use kfree() instead of kfree_skb() in Stephen's patch?
> 
> +               skb_new = skb_realloc_headroom(skb, GMAC_FCB_LEN);
> +               if (!skb_new) {
> +                       dev->stats.tx_errors++;
> +                       kfree(skb);
> +                       return NETDEV_TX_OK;
> +               }

I'll fix this, thanks for noticing.
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html