Re: netconsole hangs w/ alt-sysrq-t

[Jeff Moyer <jmoyer@redhat.com>]

==> Regarding Re: netconsole hangs w/ alt-sysrq-t; Matt Mackall <mpm@selenic.com> adds:

[snip]

mpm> Huh. I'm not sure if that covers everything.

mpm> We want to:

mpm> a) push stuff through netpoll_rx() iff rx is enabled b) drop packets
mpm> when netpoll_rx() says to c) drop packets when we're trapped for
mpm> debugging/netdump d) drop packets when we manually call dev->poll e)
mpm> keep as little overhead as possible

mpm> The above breaks (a) when we're not trapped (consider breaking in with
mpm> debugger) and (b) in any case.

mpm> My current thought is we want to get this down to a single test in the
mpm> fastpath by replacing netpoll_rx in skb->dev with a flag that is equal
mpm> to netpoll_trapped() | dev->netpoll_rx, which we manipulate when call
dev-> poll within netpoll.

mpm> How's that sound?

That sounds fine.  So, in the netif_receive_skb function, we'd now have
something like:

	if (skb->dev->netpoll_rx && netpoll_rx(skb)) {
		kfree_skb(skb);
		return NET_RX_DROP;
	}

I removed the test for skb->dev->poll, since this is the NAPI code path, I
think it should always be present, right?

And then at the top of the netpoll_rx routine:

	if (!(skb->dev->netpoll_rx & NETPOLL_RX_ENABLED))
		goto out;

The routine, as before, returns the value of trapped.  And now,
netpoll_poll would do this:

	if(np->dev->poll && test_bit(__LINK_STATE_RX_SCHED, &np->dev->state)) {
		np->dev->netpoll_rx |= NETPOLL_RX_TRAPPED;
		if (trapped)
			np->dev->poll(np->dev, &budget);
		else {
			trapped = 1;
			np->dev->poll(np->dev, &budget);
			trapped = 0;
		}
		np->dev->netpoll_rx &= ~NETPOLL_RX_TRAPPED;
	}

Is this more in line with what you had in mind?  One thing I would consider
changing, here, is having netpoll_set_trap take a struct netpoll argument.
That way, we could have netpoll_set_trap manipulate the NETPOLL_RX_TRAPPED
bit.  This would make the trap specific to each interface, but I think that
may be desirable.

-Jeff