Re: skb_padto and small fragmented transmits

Re: skb_padto and small fragmented transmits

[Chris Leech]

On Thu, 2003-02-06 at 10:44, David S. Miller wrote:
>    From: Chris Leech <christopher.leech@intel.com>
>    Date: 06 Feb 2003 11:22:51 -0800
>    
>    I fail to see how the statement "skb->len + skb->data_len" has any
>    usable meaning, or how it can be anything other than a bug.
> 
> This equation is the standard way to find the full length
> on any skb.  For linear skbs, data_len is always zero.
> 
> I asked Alan to use this formula so that greps on the source
> tree would always show data_len being taken into account, and
> thus usage would be consistent.

OK, now I'm really getting confused.  Every other example I can find in
the networking code, and every scatter-gather capable driver, uses
skb->len as the full length and skb->len - skb->data_len as the length
of the first or linear portion.

Re: skb_padto and small fragmented transmits

[David S. Miller]

   From: Chris Leech <christopher.leech@intel.com>
   Date: 06 Feb 2003 11:22:08 -0800
   
   OK, now I'm really getting confused.  Every other example I can find in
   the networking code, and every scatter-gather capable driver, uses
   skb->len as the full length and skb->len - skb->data_len as the length
   of the first or linear portion.
   
Indeed, Alan you need to fix the skb_padto stuff to use
skb->len, ignore the skb->data_len as skb->len is the
full length.

Sorry for telling you to do the wrong thing Alan, my bad :)

Re: skb_padto and small fragmented transmits

[Alan Cox]

On Thu, 2003-02-06 at 22:43, David S. Miller wrote:
>    From: Chris Leech <christopher.leech@intel.com>
>    Date: 06 Feb 2003 11:22:08 -0800
>    
>    OK, now I'm really getting confused.  Every other example I can find in
>    the networking code, and every scatter-gather capable driver, uses
>    skb->len as the full length and skb->len - skb->data_len as the length
>    of the first or linear portion.
>    
> Indeed, Alan you need to fix the skb_padto stuff to use
> skb->len, ignore the skb->data_len as skb->len is the
> full length.

Dave just fix it next time you touch the code and push it to Marcelo. It
doesnt affect the 2.2 backport so that will be ok

Re: skb_padto and small fragmented transmits

[David S. Miller]

   From: Alan Cox <alan@lxorguk.ukuu.org.uk>
   Date: 07 Feb 2003 13:33:41 +0000

   On Thu, 2003-02-06 at 22:43, David S. Miller wrote:
   > Indeed, Alan you need to fix the skb_padto stuff to use
   > skb->len, ignore the skb->data_len as skb->len is the
   > full length.
   
   Dave just fix it next time you touch the code and push it to Marcelo. It
   doesnt affect the 2.2 backport so that will be ok
   
Ok, I will take care of this.

Re: skb_padto and small fragmented transmits

[Chris Leech]

On Thu, 2003-02-06 at 03:58, David S. Miller wrote:

> skb_padto() only works on linear skb.

The result is always a linear skb.  Given that skb_padto() takes into
account data_len (incorrectly, but still) and skb_pad() contains a
comment about non-linear skb always having zero tailroom, it certainly
looks like these were written with the attempt to work for non-linear
buffers.

I fail to see how the statement "skb->len + skb->data_len" has any
usable meaning, or how it can be anything other than a bug.

The checksum issue I mentioned is not as clear.  I haven't looked at all
the callers of skb_copy_expand() and copy_skb_header() to see what
effect copying ip_summed in one of those calls might have elsewhere.

> And if you look at all the drivers where it is used, they
> do not enable things like scatter-gather.

So because the problem is not currently exposed, it's acceptable for the
code to be incorrect?

-- Chris


diff -aur a/include/linux/skbuff.h b/include/linux/skbuff.h
--- a/include/linux/skbuff.h	2003-01-13 12:45:20.000000000 -0800
+++ b/include/linux/skbuff.h	2003-02-05 12:25:38.000000000 -0800
@@ -1102,7 +1102,7 @@
  
 static inline struct sk_buff *skb_padto(struct sk_buff *skb, unsigned int len)
 {
-	unsigned int size = skb->len + skb->data_len;
+	unsigned int size = skb->len;
 	if (likely(size >= len))
 		return skb;
 	return skb_pad(skb, len-size);

Re: skb_padto and small fragmented transmits

[David S. Miller]

   From: Chris Leech <christopher.leech@intel.com>
   Date: 06 Feb 2003 11:22:51 -0800
   
   I fail to see how the statement "skb->len + skb->data_len" has any
   usable meaning, or how it can be anything other than a bug.

This equation is the standard way to find the full length
on any skb.  For linear skbs, data_len is always zero.

I asked Alan to use this formula so that greps on the source
tree would always show data_len being taken into account, and
thus usage would be consistent.

Re: skb_padto and small fragmented transmits

[David S. Miller]

skb_padto() only works on linear skb.

And if you look at all the drivers where it is used, they
do not enable things like scatter-gather.

skb_padto and small fragmented transmits

[Chris Leech]

While looking at the new software padding routines, something caught my
eye in skb_padto.  It seemed that the fragmented portion of a packet
would actually be counted twice when checking to see if padding is
needed, as skb->len already includes the count of skb->data_len.

> unsigned int size = skb->len + skb->data_len;

I tested this by modifying e1000 to use skb_padto, disabling TCP
timestamps, and writing a small app to transmit 4 bytes using sendfile. 
The resulting packet had 54 bytes of headers, and 4 bytes of data in a
separate fragment.  Calling skb_padto(skb,60) should have linearized the
skb, and zeroed out the first 2 bytes of tailroom.  Instead the length
was incorrectly calculated as 62 bytes, and the buffer was returned as
is.

Changing skb_padto to simply use size = skb->len fixed the padding, but
then I started seeing incorrect TCP checksums going out.  I found this
comment in skb_copy_expand that seemed to explain things.

> BUG ALERT: ip_summed is not copied. Why does this work? Is it used
> only by netfilter in the cases when checksum is recalculated? --ANK

So after calling skb_copy_expand the checksum is not recalculated in
software, but the checksum offload information is discarded.

-- 
Chris Leech <christopher.leech@intel.com>