ip_finish_outpu2 question

ip_finish_outpu2 question

[Prashanth]

hi, In the following function at line 189 , we are getting aligned
length (hh_alen) for the hardware header , but while doing a
skb_push() why are we using the hh->hh_len;for ethernet hh_alen would
be 16 , and hh->hh_len would be 14, since we are doing skb_push() with
14 (i.e hh->hh_len), the skb->data will not be aligned right? Can
someone please shed some light on this. thanks.

    163 static inline int ip_finish_output2(struct sk_buff *skb)
    164 {
    165     struct dst_entry *dst = skb->dst;
    166     struct hh_cache *hh = dst->hh;
    167     struct net_device *dev = dst->dev;
    168     int hh_len = LL_RESERVED_SPACE(dev);
    169
    170     /* Be paranoid, rather than too clever. */
    171     if (unlikely(skb_headroom(skb) < hh_len && dev->hard_header)) {
    172         struct sk_buff *skb2;
    173
    174         skb2 = skb_realloc_headroom(skb, LL_RESERVED_SPACE(dev));
    175         if (skb2 == NULL) {
    176             kfree_skb(skb);
    177             return -ENOMEM;
    178         }
    179         if (skb->sk)
    180             skb_set_owner_w(skb2, skb->sk);
    181         kfree_skb(skb);
    182         skb = skb2;
    183     }
    184
    185     if (hh) {
    186         int hh_alen;
    187
    188         read_lock_bh(&hh->hh_lock);
    189         hh_alen = HH_DATA_ALIGN(hh->hh_len);
    190         memcpy(skb->data - hh_alen, hh->hh_data, hh_alen);
    191         read_unlock_bh(&hh->hh_lock);
    192         skb_push(skb, hh->hh_len);
    193         return hh->hh_output(skb);
    194     } else if (dst->neighbour)
    195         return dst->neighbour->output(skb);
    196
    197     if (net_ratelimit())
    198         printk(KERN_DEBUG "ip_finish_output2: No header cache
and no neighbour!\n");
    199     kfree_skb(skb);
    200     return -EINVAL;
    201 }

Re: ip_finish_outpu2 question

[David Miller]

From: Prashanth <bshanth@gmail.com>
Date: Wed, 14 May 2008 09:12:54 +0530

> hi, In the following function at line 189 , we are getting aligned
> length (hh_alen) for the hardware header , but while doing a
> skb_push() why are we using the hh->hh_len;for ethernet hh_alen would
> be 16 , and hh->hh_len would be 14, since we are doing skb_push() with
> 14 (i.e hh->hh_len), the skb->data will not be aligned right? Can
> someone please shed some light on this. thanks.

We are using 16 for the copy so that we do an aligned
copy.  But the ethernet header size is only 14 bytes.

It's safe to copy those extra 2 bytes at the front
since we always will have some extra slack space there.

Re: ip_finish_outpu2 question

[Prashanth]

So skb->data will be pointing to an unaligned memory address; atleast
for the ethernet case which has 14 Bytes header length.
because of this my chip is not allowing to do DMA for unaligned memory address.

If we take the case of ethernet , The hh->hh_data will be pointing to
the ethernet header by leaving the first 2 bytes.
HH_DATA_OFF() macro is used for this purpose.

By doing skb_push() only for 14 bytes and making skb->data to point to
unaligned memory will have performance hit,
In my case i was not even  able to xmit the packet.So is it not better
to do skb_push() for 16 bytes and make skb->data aligned?

In ip_finish_output2() function  it's better to do skb_push() for
aligned length. And let others access the ethernet header
by adding HH_DATA_OFF() to it. Do we face any problems by doing this.

Please fix me if i'm wrong.

Thanks

On Wed, May 14, 2008 at 10:04 AM, David Miller <davem@davemloft.net> wrote:
> From: Prashanth <bshanth@gmail.com>
> Date: Wed, 14 May 2008 09:12:54 +0530
>
>> hi, In the following function at line 189 , we are getting aligned
>> length (hh_alen) for the hardware header , but while doing a
>> skb_push() why are we using the hh->hh_len;for ethernet hh_alen would
>> be 16 , and hh->hh_len would be 14, since we are doing skb_push() with
>> 14 (i.e hh->hh_len), the skb->data will not be aligned right? Can
>> someone please shed some light on this. thanks.
>
> We are using 16 for the copy so that we do an aligned
> copy.  But the ethernet header size is only 14 bytes.
>
> It's safe to copy those extra 2 bytes at the front
> since we always will have some extra slack space there.
>

Re: ip_finish_outpu2 question

[David Miller]

From: Prashanth <bshanth@gmail.com>
Date: Wed, 14 May 2008 10:53:36 +0530

> So skb->data will be pointing to an unaligned memory address;
> atleast for the ethernet case which has 14 Bytes header length.

That's correct.

> because of this my chip is not allowing to do DMA for unaligned
> memory address.

Most chips have an "packet offset" parameter you can feed into the TX
descriptor, and you align the DMA address that is given to the
chip.  This allows the chip to do an aligned DMA fetch yet still
be agnostic the the alignment of the packet given to it by the
networking stack.

> So is it not better
> to do skb_push() for 16 bytes and make skb->data aligned?

If we did that skb->data would point to the two garbage bytes
in front of the ethernet header.

> In ip_finish_output2() function it's better to do skb_push() for
> aligned length. And let others access the ethernet header by adding
> HH_DATA_OFF() to it. Do we face any problems by doing this.

No, that would punish good hardware which can handle arbitrary
TX packet alignment for the sake of bad hardware.

Re: ip_finish_outpu2 question

[kalash nainwal]

On Wed, May 14, 2008 at 10:04 AM, David Miller <davem@davemloft.net> wrote:
> From: Prashanth <bshanth@gmail.com>
> Date: Wed, 14 May 2008 09:12:54 +0530
>
>> hi, In the following function at line 189 , we are getting aligned
>> length (hh_alen) for the hardware header , but while doing a
>> skb_push() why are we using the hh->hh_len;for ethernet hh_alen would
>> be 16 , and hh->hh_len would be 14, since we are doing skb_push() with
>> 14 (i.e hh->hh_len), the skb->data will not be aligned right? Can
>> someone please shed some light on this. thanks.
>
> We are using 16 for the copy so that we do an aligned
> copy.  But the ethernet header size is only 14 bytes.
>

I thought memcpy in linux is anyway always alignment safe. no?

-Kalash

Re: ip_finish_outpu2 question

[David Miller]

From: "kalash nainwal" <kalash.nainwal@gmail.com>
Date: Wed, 14 May 2008 14:17:58 +0530

> I thought memcpy in linux is anyway always alignment safe. no?

Yes, but an aligned copy is faster.