question about UFO behavior for bridge device

question about UFO behavior for bridge device

[Shan Wei]

UDP protocol creates a big packet(skb) in which the data length is greater than MTU.

If device(eth0, lo) does not supports UFO, This big skb will be fragmented to
many fragments in IP protocol, and fragments are sent one by one by device.

But, if device(eth0, lo) supports UFO, IP protocol don't fragment it, and
device directly sends it out.

For bridge device which enable UFO, and Ethernet device(eth0) which not 
support UFO, IP protocol also doesn't fragment it, and bridge forwards original
skb to eth0. For this UFO disabled eth0, kernel needs to perform segmentation
on so big skb in dev_gso_segment(), and link segmented skbs to next field
of original skb. Then device sends segmented skbs out, but not original skb.

But, actually, i saw original big skb in eth0's tcpdump file, but not segmented skbs.
The behavior is right or what we want?
Is there anything missed about my analysis?

-- 
Best Regards
-----
Shan Wei

Re: question about UFO behavior for bridge device

[Ben Hutchings]

On Fri, 2011-05-13 at 17:21 +0800, Shan Wei wrote:
> UDP protocol creates a big packet(skb) in which the data length is greater than MTU.
> 
> If device(eth0, lo) does not supports UFO, This big skb will be fragmented to
> many fragments in IP protocol, and fragments are sent one by one by device.
> 
> But, if device(eth0, lo) supports UFO, IP protocol don't fragment it, and
> device directly sends it out.
> 
> For bridge device which enable UFO, and Ethernet device(eth0) which not 
> support UFO, IP protocol also doesn't fragment it, and bridge forwards original
> skb to eth0. For this UFO disabled eth0, kernel needs to perform segmentation
> on so big skb in dev_gso_segment(), and link segmented skbs to next field
> of original skb. Then device sends segmented skbs out, but not original skb.

Right, it does that just before calling the driver's ndo_start_xmit().

> But, actually, i saw original big skb in eth0's tcpdump file, but not segmented skbs.
> The behavior is right or what we want?
> Is there anything missed about my analysis?

I assume that packet capturing is handled earlier in the transmit path.

Ben.

-- 
Ben Hutchings, Senior Software Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.

Re: question about UFO behavior for bridge device

[Herbert Xu]

On Fri, May 13, 2011 at 02:06:52PM +0100, Ben Hutchings wrote:
>
> > But, actually, i saw original big skb in eth0's tcpdump file, but not segmented skbs.
> > The behavior is right or what we want?
> > Is there anything missed about my analysis?
> 
> I assume that packet capturing is handled earlier in the transmit path.

Yes it is.  It's the same with TSO.

Cheers,
-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: question about UFO behavior for bridge device

[Shan Wei]

Herbert Xu wrote, at 05/13/2011 10:36 PM:
> On Fri, May 13, 2011 at 02:06:52PM +0100, Ben Hutchings wrote:
>>
>>> But, actually, i saw original big skb in eth0's tcpdump file, but not segmented skbs.
>>> The behavior is right or what we want?
>>> Is there anything missed about my analysis?
>>
>> I assume that packet capturing is handled earlier in the transmit path.
> 
> Yes it is.  It's the same with TSO.

Be cheated by tcpdump. :-(
dev_queue_xmit_nit(skb, dev) which dumps packet to user is called before handling GSO segments.

Thanks very much.

-- 
Best Regards
-----
Shan Wei