On Sat, Sep 20, 2008 at 10:35:38PM -0700, David Miller wrote:

From: Jarek Poplawski <redacted>
Date: Sun, 21 Sep 2008 01:48:43 +0200

quoted

On Sat, Sep 20, 2008 at 12:21:37AM -0700, David Miller wrote:
...

quoted

Let's look at what actually matters for cpu utilization.  These
__qdisc_run() things are invoked in two situations where we might
block on the hw queue being stopped:

1) When feeding packets into the qdisc in dev_queue_xmit().

 ...

quoted

quoted

2) When waking up a queue.  And here we should schedule the qdisc_run
   _unconditionally_.

 ...

quoted

quoted

The cpu utilization savings exist for case #1 only, and we can
implement the bypass logic _perfectly_ as described above.

For #2 there is nothing to check, just do it and see what comes
out of the qdisc.

Right, unless __netif_schedule() wasn't done when waking up. I've
thought about this because of another thread/patch around this
problem, and got misled by dev_requeue_skb() scheduling. Now, I think
this could be the main reason for this high load. Anyway, if we want
to skip this check for #2 I think something like the patch below is
needed.

Hmmm, looking at your patch....

It's only doing something new when the driver returns NETDEV_TX_BUSY
from ->hard_start_xmit().

That _never_ happens in any sane driver.  That case is for buggy
devices that do not maintain their TX queue state properly.  And
in fact it's a case for which I advocate we just drop the packet
instead of requeueing.  :-)

OK, then let's do it! Why I can't see this in your new patch?

Oh I see, you're concerned about that cases where qdisc_restart() ends
up using the default initialization of the 'ret' variable.

Yes, this is my main concern.

Really, for the case where the driver actually returns NETDEV_TX_BUSY
we _do_ want to unconditionally __netif_schedule(), since the device
doesn't maintain it's queue state in the normal way.

So, do you advocate both to drop the packet and unconditionally
__netif_schedule()?!

Therefore it seems logical that what really needs to happen is that
we simply pick some new local special token value for 'ret' so that
we can handle that case.  "-1" would probably work fine.

So I'm dropping your patch.

I also think the qdisc_run() test needs to be there.  When the TX
queue fills up, we will doing tons of completely useless work going:

1) ->dequeue
2) qdisc unlock
3) TXQ lock
4) test state
5) TXQ unlock
6) qdisc lock
7) ->requeue

for EVERY SINGLE packet that is generated towards that device.

That has to be expensive,

I agree this useless work should be avoided, but only with a reliable
(and not too expensive) test. Your test might be done for the last
packet in the queue, while all the previous packets (and especially
the first one) have a different state of the queue. This should work
well for uniqueue devs and multiqueues with dedicated qdiscs, but is
doubtful for multiqueues with one qdisc, where it actually should be
most needed, because of potentially complex multiclass configs with
this new problem of blocking at the head-of-line (Alexander's main
concern).

BTW, since this problem is strongly conected with the requeuing
policy, I wonder why you seemingly lost interest in this. I tried to
advocate for your simple, one level requeuing, but also Herbert's
peek, and Alexander's early detection, after some polish(!), should
make this initial test meaningless.

and I am still very much convinced that
this was the original regression cause that made me put that TXQ
state test back into qdisc_run().

I doubt this: I've just looked at this Andrew Gallatin's report, and
there is really a lot of net_tx_action, __netif_schedule, and guess
what: pfifo_fast_requeue in this oprofile...

Jarek P.