Linux 2.4.22 loopback question

Linux 2.4.22 loopback question

[Cheng Jin]

Hi,

I am running into a strange problem with timestamping packets on loopback 
connections.

I calculate the RTT for each outgoing TCP packet by timestamping each skb 
after tcp_transmit_skb returns successfully, and timestamping the 
corresponding ack in tcp_ack (actually netif_rx timestamps incoming skbs 
so I use that whenever it is available).  However, I have noticed 
that sometimes, the timestamp for an ack comes before the timestamp for 
the trasnmitted skb that is being acked.

I have instrumented the code in net/core/dev.c, and it seems that indeed
the corresponding ack would sometimes be generated by the receiver, 
received and timestamped by netif_rx before tcp_transmit_skb returns.

I have a dual Xeon 2.66Ghz machine with Redhat 7.3 base install + 2.4.22
kernel (SMP enabled) from kernel.org.  The network driver is e1000 without 
NAPI.

Any ideas on why this happens?

Many thanks,

Cheng

Re: Linux 2.4.22 loopback question

[Andi Kleen]

On Wed, May 05, 2004 at 02:53:41PM -0700, Cheng Jin wrote:
> I have instrumented the code in net/core/dev.c, and it seems that indeed
> the corresponding ack would sometimes be generated by the receiver, 
> received and timestamped by netif_rx before tcp_transmit_skb returns.

Most likely because the network softirq runs in parallel in interrupt
context and picks up all the already queued packets. e.g. this
can be triggered when you receive a packet from ethernet.

-Adni