Message-ID: <20231002162601.6b71c4d9@xps-13>
Date: Mon, 2 Oct 2023 16:26:01 +0200
From: Miquel Raynal <miquel.raynal@...tlin.com>
To: Marc Kleine-Budde <mkl@...gutronix.de>
Cc: Wolfgang Grandegger <wg@...ndegger.com>, "David S. Miller"
 <davem@...emloft.net>, Jakub Kicinski <kuba@...nel.org>, Paolo Abeni
 <pabeni@...hat.com>, Eric Dumazet <edumazet@...gle.com>,
 netdev@...r.kernel.org, linux-can@...r.kernel.org, Jérémie
 Dautheribes <jeremie.dautheribes@...tlin.com>, Thomas Petazzoni
 <thomas.petazzoni@...tlin.com>, sylvain.girard@...com,
 pascal.eberhard@...com, stable@...r.kernel.org
Subject: Re: [PATCH v2] can: sja1000: Always restart the Tx queue after an
 overrun

Hi Marc,

mkl@...gutronix.de wrote on Thu, 28 Sep 2023 09:53:17 +0200:

> On 27.09.2023 18:44:42, Miquel Raynal wrote:
> > Upstream commit 717c6ec241b5 ("can: sja1000: Prevent overrun stalls with
> > a soft reset on Renesas SoCs") fixes an issue with Renesas own SJA1000
> > CAN controller reception: the Rx buffer is only 5 messages long, so when
> > the bus loaded (eg. a message every 50us), overrun may easily
> > happen. Upon an overrun situation, due to a possible internal crosstalk
> > situation, the controller enters a frozen state which only can be
> > unlocked with a soft reset (experimentally). The solution was to offload
> > a call to sja1000_start() in a threaded handler. This needs to happen in
> > process context as this operation requires to sleep. sja1000_start()
> > basically enters "reset mode", performs a proper software reset and
> > returns back into "normal mode".
> > 
> > Since this fix was introduced, we no longer observe any stalls in
> > reception. However it was sporadically observed that the transmit path
> > would now freeze. Further investigation blamed the fix mentioned above,
> > and especially the reset operation. Reproducing the reset in a loop
> > helped identifying what could possibly go wrong. The sja1000 is a single
> > Tx queue device, which leverages the netdev helpers to process one Tx
> > message at a time. The logic is: the queue is stopped, the message sent
> > to the transceiver, once properly transmitted the controller sets a
> > status bit which triggers an interrupt, in the interrupt handler the
> > transmission status is checked and the queue woken up. Unfortunately, if
> > an overrun happens, we might perform the soft reset precisely between
> > the transmission of the buffer to the transceiver and the advent of the
> > transmission status bit. We would then stop the transmission operation
> > without re-enabling the queue, leading to all further transmissions to
> > be ignored.
> > 
> > The reset interrupt can only happen while the device is "open", and
> > after a reset we anyway want to resume normal operations, no matter if a
> > packet to transmit got dropped in the process, so we shall wake up the
> > queue. Restarting the device and waking-up the queue is exactly what
> > sja1000_set_mode(CAN_MODE_START) does. In order to be consistent about
> > the queue state, we must acquire a lock both in the reset handler and in
> > the transmit path to ensure serialization of both operations. As the
> > reset handler might still be called after the transmission of a frame to
> > the transceiver but before it actually gets transmitted, we must ensure
> > we don't leak the skb, so we free it (the behavior is consistent, no
> > matter if there was an skb on the stack or not).
> > 
> > Fixes: 717c6ec241b5 ("can: sja1000: Prevent overrun stalls with a soft reset on Renesas SoCs")
> > Cc: stable@...r.kernel.org
> > Signed-off-by: Miquel Raynal <miquel.raynal@...tlin.com>
> > ---
> > 
> > Changes in v2:
> > * As Marc sugested, use netif_tx_{,un}lock() instead of our own
> >   spin_lock.
> > 
> >  drivers/net/can/sja1000/sja1000.c | 11 ++++++++++-
> >  1 file changed, 10 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/net/can/sja1000/sja1000.c b/drivers/net/can/sja1000/sja1000.c
> > index ae47fc72aa96..91e3fb3eed20 100644
> > --- a/drivers/net/can/sja1000/sja1000.c
> > +++ b/drivers/net/can/sja1000/sja1000.c
> > @@ -297,6 +297,7 @@ static netdev_tx_t sja1000_start_xmit(struct sk_buff *skb,
> >  	if (can_dropped_invalid_skb(dev, skb))
> >  		return NETDEV_TX_OK;
> >  
> > +	netif_tx_lock(dev);
> >  	netif_stop_queue(dev);
> >  
> >  	fi = dlc = cf->can_dlc;
> > @@ -335,6 +336,8 @@ static netdev_tx_t sja1000_start_xmit(struct sk_buff *skb,
> >  
> >  	sja1000_write_cmdreg(priv, cmd_reg_val);
> >  
> > +	netif_tx_unlock(dev);
> > +  
> 
> I think netif_tx_lock() should be used in a different way. As far as I
> understand it, you should call it only in the sja1000_reset_interrupt(),
> where you want to tx path to interfere.

I believe you meant "don't want"? And yes you're right current use
can't properly handle my problem.

> Please test the new code with lockdep enabled.

I will fix the current implementation and test again by manually
producing overruns.

Thanks,
Miquèl

Powered by blists - more mailing lists