David Dillow [off-list ref] :
[...]

I've not been able to reproduce my lockups under medium testing with
Francois's patch applied, so I'm a little more comfortable with it.

Nice :o)

At the same time, I'm worried that the timing just changed enough to
make it harder to trigger, as was the case when I did the patch IIRC.

It is a legitimate concern.

The kernel's interrupt handling changed in a manner that made it much
easier to hit about that time. The testing I did in May pointed strongly
at us failing to ACK an interrupt source, causing the MSI generation to
stop, so I need to think about things some more.

It can be understood as us claiming to perform some work we didn't too.

In this regard, a "ack everything and perform no work loop in the irq
handler" design would require some work : it races with the - supposedly
fast, register read free - napi handler which does not check that
unprocessed events are acked.

As the current patch was provided with almost no explanation :
- the irq handler and the napi handler are allowed / assumed / expected
  to race
- the napi and irq handlers ack respectively their own events (IntrStatus).
  They do not ack their friend ones.
- everybody acks (IntrStatus) before the work is done
- napi irqs are disabled before napi is (tentatively) scheduled.
  napi irqs are only expected to be disabled most of the time the napi
  handler runs.
- the napi handler enables its irqs, tests new events and conditionaly
  schedules itself.

-- 
Ueimor