assembly of packets

Craig Jackson cjackson at
Tue Jun 4 09:54:25 EDT 2013

> -----Original Message-----
> From: Varun Sharma [mailto:vsdssd at]
> Sent: Tuesday, June 04, 2013 2:45 AM
> To: Craig Jackson; kernelnewbies at;
> Valdis.Kletnieks at
> Subject: Re: assembly of packets
> On receving side upto tcp layer skb->len=1480 now where is packet
> assembled on tcp layer ?
> or it is directly copies into user space buffer.
> I am sending len = 10000 in send system call.
>> On Mon, Jun 3, 2013 at 9:45 PM, Craig Jackson <cjackson at>
>> wrote:
>> In my experience, the second paragraph isn't quite true. What you see
>> with TSO is the pre-segmentation "packet", up to 65k. (By this I mean
>> the set of data which is given to the offload hardware to segment.) So
>> you need to make sure that your "-s" value is big enough to see
>> everything.
>> (Speaking as someone who was bitten by bugs in the early versions of
>> TSO.)
>> Craig
>>> -----Original Message-----
>>> From: at
>>> [ at]
>>> On Behalf Of Valdis.Kletnieks at
>>> Sent: Monday, June 03, 2013 12:03 PM
>>> To: Varun Sharma
>>> Cc: kernelnewbies at
>>> Subject: Re: assembly of packets
>>> On Mon, 03 Jun 2013 14:48:41 +0530, Varun Sharma said:
>>> > If TSO(tcp segmentation offload ) is enabled then nic card handle
>>> > segmentation then where is assembly of packets happens ? Is it
>>> > tcp_rcv_established function ?
>>> The whole *point* of TSO is so the NIC does all the segmentation
>>> reassembly and DMA, and wake the kernel up when all the data is already
>>> stashed in buffers fully processed.
>>> Incidentally, this is why if you run tcpdump on an interface that has
>>> TSO enabled, you'll only see the first 3 handshaking packets and the
>>> final FIN packets - the other packets wake up the TCP stack at a point
>>> after where tcpdump's tap would have seen the packet.

Sorry for top-posting before. It's incredibly hard to get Outlook 2007 to do things right.

Varun, I now see we may not have answered the question you have. TSO (TCP Segementation Offload)
is associated with sending packets, not receiving them. They are re-assembled at the receiving
end in the standard manner.

Are you actually thinking of GRO (Generic Receive Offload)?


More information about the Kernelnewbies mailing list