loopback interface and soft irqs

[John Heffner <jheffner@xxxxxxx>]

I'm using 2.4.20, and I noticed that when doing an all-out tcp transfer
over the loopback interface, the machine was actually spending lots of
time in the idle loop.  I looked at the device code and after
loopback_xmit() calls netif_rx(), the softirq does not get run.  I added a
do_softirq() call at the end of loopback_xmit() and CPU utilization went
up to 100% throughput significantly increased (about 2x).

Is there a good reason for not running the soft irq here?

Thanks,
  -John