Re: [RFC] mt76: fix tx hung regression on MT7630E

[Lorenzo Bianconi <lorenzo.bianconi@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 5246 bytes --]

> On Wed, Jul 31, 2019 at 11:09:27AM +0200, Lorenzo Bianconi wrote:
> > > On Wed, Jul 31, 2019 at 10:19:58AM +0200, Stanislaw Gruszka wrote:
> > > > On Tue, Jul 30, 2019 at 04:55:31PM +0200, Lorenzo Bianconi wrote:
> > > > > > > > diff --git a/drivers/net/wireless/mediatek/mt76/mt76x02_mmio.c b/drivers/net/wireless/mediatek/mt76/mt76x02_mmio.c
> > > > > > > > index 467b28379870..622251faa415 100644
> > > > > > > > --- a/drivers/net/wireless/mediatek/mt76/mt76x02_mmio.c
> > > > > > > > +++ b/drivers/net/wireless/mediatek/mt76/mt76x02_mmio.c
> > > > > > > > @@ -170,7 +170,7 @@ static int mt76x02_poll_tx(struct napi_struct *napi, int budget)
> > > > > > > >  					       mt76.tx_napi);
> > > > > > > >  	int i;
> > > > > > > >  
> > > > > > > > -	mt76x02_mac_poll_tx_status(dev, false);
> > > > > > > > +	mt76x02_mac_poll_tx_status(dev, true);
> > > > > > > 
> > > > > > > I am not sure if we really need mt76x02_mac_poll_tx_status() here since we run
> > > > > > > it in mt76x02_tx_complete_skb() and in mt76x02_tx_tasklet(). Anyway the only
> > > > > > > difference doing so is we do not run mt76x02_send_tx_status().
> > > > > > 
> > > > > > I thought this is the problem, but it was my mistake during testing.
> > > > > > I tested the above change together with mt76_txq_schedule(dev, txq->ac)
> > > > > > change and get wrong impression it fixes the issue. But above change
> > > > > > alone does not help.
> > > > > > 
> > > > > > I tried to add some locking to avoid parallel execution of mt76x02_poll_tx()
> > > > > > and mt76x02_tx_tasklet(), but it didn't help either. So far only patch
> > > > > > originally posted here make the problem gone.
> > > > > 
> > > > > so, in order to be on the same page, if you comment out mt76x02_mac_poll_tx_status()
> > > > > in mt76x02_poll_tx() the issue will still occur. The only to 'fix' it is to run
> > > > > mt76_txq_schedule_all() in mt76x02_poll_tx(), right?
> > > > 
> > > > Yes.
> > > 
> > > Err, no, I should read more cerfully. It is partiall revert of 
> > > 41634aa8d6db ("mt76: only schedule txqs from the tx tasklet") : 
> > > 
> > > diff --git a/drivers/net/wireless/mediatek/mt76/tx.c b/drivers/net/wireless/mediatek/mt76/tx.c
> > > index 5397827668b9..fefe0ee52584 100644
> > > --- a/drivers/net/wireless/mediatek/mt76/tx.c
> > > +++ b/drivers/net/wireless/mediatek/mt76/tx.c
> > > @@ -598,7 +598,7 @@ void mt76_wake_tx_queue(struct ieee80211_hw *hw, struct ieee80211_txq *txq)
> > >         if (!test_bit(MT76_STATE_RUNNING, &dev->state))
> > >                 return;
> > >  
> > > -       tasklet_schedule(&dev->tx_tasklet);
> > > +       mt76_txq_schedule(dev, txq->ac);
> > >  }
> > >  EXPORT_SYMBOL_GPL(mt76_wake_tx_queue);
> > 
> > reviewing the code I think:
> > 
> > - we should not run mt76u_tx_tasklet() from mt76_wake_tx_queue() since we do
> >   not have tx_napi for usb and it will unnecessary go through tx queue checks.
> >   We should probably do in mt76_wake_tx_queue() something like:
> > 
> >   if (is_mmio())
> 
> Adding '&& !is_mt7630()' will solve the problem for MT7630E as well ...
> 
> > 	  tasklet_schedule(&dev->tx_tasklet);
> >   else
> > 	  mt76_txq_schedule(dev, txq->ac);
> > 
> >   Another solution would be add a status_tasklet that just goes through the tx
> >   queues receiving the usb tx completion and it schedules the tx_tasklet
> >   What do you think?
> > 
> > - I guess it does not fix the 76x0e issue but we should just schedule tx queues in
> >   mt76x02_tx_tasklet() (like it is done for mt7603 and mt7615) and move status
> >   processing in mt76x02_poll_tx()
> 
> ... but I think we have bug when do mt76_txq_schedule_all() in
> tx_tasklet, because we can schedule on queues that are stoped.
> So reverting 41634aa8d6db and then optimize by removing tx_tasklet
> for mmio and remove not needed mt76_txq_schedule_all() calls looks
> more reasoneble to me.

schedule a stopped queue seems not harmful at a first glance since we do not
copy pending skbs if we have not enough room in the dma ring. Maybe we can be
more conservative doing something like:

diff --git a/drivers/net/wireless/mediatek/mt76/dma.c b/drivers/net/wireless/mediatek/mt76/dma.c
index d8f61e540bfd..c6482155e5e4 100644
--- a/drivers/net/wireless/mediatek/mt76/dma.c
+++ b/drivers/net/wireless/mediatek/mt76/dma.c
@@ -346,6 +346,11 @@ mt76_dma_tx_queue_skb(struct mt76_dev *dev, enum mt76_txq_id qid,
 		goto unmap;
 
 	if (q->queued + (tx_info.nbuf + 1) / 2 >= q->ndesc - 1) {
+		if (!q->stopped) {
+			ieee80211_stop_queue(dev->hw,
+					     skb_get_queue_mapping(skb));
+			q->stopped = true;
+		}
 		ret = -ENOMEM;
 		goto unmap;
 	}
diff --git a/drivers/net/wireless/mediatek/mt76/tx.c b/drivers/net/wireless/mediatek/mt76/tx.c
index 5397827668b9..bd2d34c4f326 100644
--- a/drivers/net/wireless/mediatek/mt76/tx.c
+++ b/drivers/net/wireless/mediatek/mt76/tx.c
@@ -495,6 +495,9 @@ mt76_txq_schedule_list(struct mt76_dev *dev, enum mt76_txq_id qid)
 	while (1) {
 		bool empty = false;
 
+		if (hwq->stopped)
+			break;
+
 		if (sq->swq_queued >= 4)
 			break;

Does it fix the issue you are facing?

Regards,
Lorenzo

>  
> Stanislaw 

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]