On Mon Jun 29, 2020 at 3:07 PM CEST, David Miller wrote:
> From: Tobias Waldekranz <tob...@waldekranz.com>
> Date: Mon, 29 Jun 2020 21:16:01 +0200
>
> > In the ISR, we poll the event register for the queues in need of
> > service and then enter polled mode. After this point, the event
> > register will never be read again until we exit polled mode.
> >
> > In a scenario where a UDP flow is routed back out through the same
> > interface, i.e. "router-on-a-stick" we'll typically only see an rx
> > queue event initially. Once we start to process the incoming flow
> > we'll be locked polled mode, but we'll never clean the tx rings since
> > that event is never caught.
> >
> > Eventually the netdev watchdog will trip, causing all buffers to be
> > dropped and then the process starts over again.
> >
> > By adding a poll of the active events at each NAPI call, we avoid the
> > starvation.
> >
> > Fixes: 4d494cdc92b3 ("net: fec: change data structure to support
> > multiqueue")
> > Signed-off-by: Tobias Waldekranz <tob...@waldekranz.com>
>
> I don't see how this can happen since you process the TX queue
> unconditionally every NAPI pass, regardless of what bits you see
> set in the IEVENT register.
>
> Or don't you? Oh, I see, you don't:
>
> for_each_set_bit(queue_id, &fep->work_tx, FEC_ENET_MAX_TX_QS) {
>
> That's the problem. Just unconditionally process the TX work regardless
> of what is in IEVENT. That whole ->tx_work member and the code that
> uses it can just be deleted. fec_enet_collect_events() can just return
> a boolean saying whether there is any RX or TX work at all.
Maybe Andy could chime in here, but I think the ->tx_work construction
is load bearing. It seems to me like that is the only thing stopping
us from trying to process non-existing queues on older versions of the
silicon which only has a single queue.
