2003-07-31 04:30:49

by Nagendra Singh Tomar

[permalink] [raw]
Subject: Avoiding re-ordering in netif_rx()

I am modifying linux kernel 2.4.18 to add support for our TCP offload
card. The problem is:
The packets that I get from the card are fully TCP processed and
in-order.
Now I feed these packets to netif_rx(). marking a flag in the skbuff
which
says that the full TCP/IP processing is done on this packet and the
higher
layers can just bypass the packet protocol processing. On an SMP m/c
different consequtive in-order packets received from the card can be
queued to different per-cpu queues and it might so happen that the later

received packet is added to the socket receive queue first ( bcos the
softirq on the later CPU got a chance to execute first). This
maliciously
reorders the data.

My question is, what is an elegant way of avoiding this. I have a couple

of choices, but I want to know what people think. f.e one way is to
queue
these packets to a single queue and not a per-cpu queue. In this case
order will be honoured.

brilliant suggestions are very welcome !!

Thanx
tomar