[PATCH 2/4]: powerpc/cell spidernet low watermark patch.

[Benjamin Herrenschmidt]

On Fri, 2006-08-18 at 15:51 -0700, David Miller wrote:
> From: linas at austin.ibm.com (Linas Vepstas)
> Date: Fri, 18 Aug 2006 17:46:18 -0500
> 
> > > We're not saying to use the RX interrupt as the trigger for
> > > RX and TX work.  Rather, either of RX or TX interrupt will
> > > schedule the NAPI poll.
> > 
> > And, for a lark, this is exactly what I did. Just to see.
> > Because there are so few ack packets, there are very few 
> > RX interrupts -- not enough to get NAPI to actually keep
> > the device busy.
> 
> You're misreading me.  TX interrupts are intended to be "enabled" and
> trigger NAPI polls.  TX IRQ enabled, enabled :-)

Maybe be because you actually typed "disabled" in your previous
message ? :)

>> The idea is to use NAPI polling with TX interrupts disabled.

> If you want to eliminate them if the kernel keeps hopping into
> the ->hard_start_xmit() via hw interrupt mitigation or whatever,
> that's fine.  But if you do need to do TX interrupt processing,
> do it in NAPI ->poll().

Well, we do need to harvest descriptors of course, though I suppose that
can be done in hard_xmit as well. I'm not sure if there is any real
benefit in batching those.

> > I'm somewhat disoriened from this conversation. Its presumably
> > clear that low-watermark mechanisms are superior to NAPI. 
> > >From what I gather, NAPI was invented to deal with cheap 
> > or low-function hardware; it adds nothing to this particular
> > situation. Why are we talking about this?
> 
> NAPI is meant to give fairness to all devices receiving packets
> in the system, particularly in times of high load or overload.
> 
> And equally importantly, it allows you to run the majority of your
> interrupt handler in software IRQ context.

That is the most important point imho for the specific case of spidernet
on cell.

> This allows not only your
> locking to be simpler, but it also allows things like oprofile to
> monitor almost your entire IRQ processing path even with just timer
> interrupt based oprofile profiling.
> 
> I see you moving TX reclaim into tasklets and stuff.  I've vehemently
> against that because you wouldn't need it in order to move TX
> processing into software interrupts if you did it all in NAPI
> ->poll().

Ben.