Re: [PATCH v2 3/7] [PATCH 3/8] can: CAN Network device driver and Netlink interface

[Wolfgang Grandegger <wg@grandegger.com>]

Oliver Hartkopp wrote:
> Wolfgang Grandegger wrote:
>> Oliver Hartkopp wrote:
>>  
>>> Wolfgang Grandegger wrote:
>>>    
>>>> Oliver Hartkopp wrote:
>>>>  
>>>>      
>>>>> Andrew Morton wrote:
>>>>>           
>>>>>> On Tue, 12 May 2009 11:28:00 +0200 Wolfgang Grandegger
>>>>>> <wg@grandegger.com> wrote:
>>>>>>
>>>>>>  
>>>>>>               
>>>>>>> +int can_restart_now(struct net_device *dev)
>>>>>>> +{
>>>>>>> +    struct can_priv *priv = netdev_priv(dev);
>>>>>>> +    struct net_device_stats *stats = &dev->stats;
>>>>>>> +    struct sk_buff *skb;
>>>>>>> +    struct can_frame *cf;
>>>>>>> +    int err;
>>>>>>> +
>>>>>>> +    /* Synchronize with dev->hard_start_xmit() */
>>>>>>> +    netif_tx_lock(dev);
>>>>>>> +
>>>>>>> +    /* Ensure that no more messages can go out */
>>>>>>> +    if (netif_carrier_ok(dev))
>>>>>>> +        netif_carrier_off(dev);
>>>>>>> +
>>>>>>> +    /* Ensure that no more messages can come in */
>>>>>>> +    err = priv->do_set_mode(dev, CAN_MODE_STOP);
>>>>>>> +    if (err)
>>>>>>> +        return err;
>>>>>>> +
>>>>>>> +    /*  Now it's save to clean up */
>>>>>>> +    del_timer_sync(&priv->restart_timer);
>>>>>>>                         
>>>>>> This is deadlockable.
>>>>>>
>>>>>> It calls del_timer_sync() while holding netif_tx_lock().  But the
>>>>>> timer
>>>>>> handler (can_restart_now()) also takes netif_tx_lock().  So if the
>>>>>> timer handler is presently running, it's sitting there spinning in
>>>>>> netif_tx_lock().  And del_timer_sync() is sitting there waiting
>>>>>> for the
>>>>>> timer handler to complete.
>>>>>>
>>>>>>
>>>>>>                   
>>>>> Hi Wolfgang,
>>>>>
>>>>> would it be an appropriate solution, just to invoke
>>>>>
>>>>> netif_stop_queue() in can_bus_off()
>>>>>
>>>>> and invoke
>>>>>
>>>>> netif_wake_queue() in can_restart_now()
>>>>>
>>>>> ???
>>>>>
>>>>> In a BUSOFF condition we're not able to send CAN frames anyway, so  we
>>>>> can disable the device queue and the we won't  need any
>>>>> netif_tx_lock()
>>>>> right?
>>>>>
>>>>> AFAIK this was the original implementation before some of the latest
>>>>> improvement with the netif_carrier stuff.
>>>>>
>>>>> What do you think?
>>>>>             
>>>> The problem is the "manual" restart triggered via the netlink
>>>> interface,
>>>> which can occur in the middle of ndo_start_xmit().
>>>>
>>>>         
>>> Ah, i see.
>>>
>>> What if the manual restart via netlink would also stop the queue and
>>> start the timer?
>>>     
>>
>> It will not help if the restart is triggered in the middle of
>> ndo_start_xmit().
>>
>>   
> 
> Hi Wolfgang,
> 
> i think, i found a solution that removes the locking problem completely:
> 
> When a bus-off occurs in the controller, the communication on the CAN
> bus can be treated as unusable for this controller (let's say it is dead).
> E.g. the SJA1000 set's its reset bit for that reason and waits to be
> initialized by the CPU again.
> 
> So IMO restarting the CAN controller while in operational state is not a
> valid use case.
>
> When a bus-off (interrupt) occurs, we should
> 
> - invoke netif_carrier_off(dev)
> - invoke netif_stop_queue(dev)
> - set the state to CAN_STATE_BUS_OFF
> 
> and of course create the error message, clear the interrupts(!) and then
> leave the irq service function.
> 
> That's it.

It's already like that.

> When the automatic CAN controller restart is enabled: start the timer.
> 
> For the manual netlink function: Test for CAN_STATE_BUS_OFF (!) and
> invoke the current can_restart_now(dev) or start the timer with e.g. 1ms
> ...
> 
> This approach should make it and fulfills the bus-off intention of the
> CAN controllers ("disabled and wait for re-initialisation").
> 
> And there's no locking of the tx_queue needed anymore as the tx_queue is
> already stopped, when the restart is performed.
> 
> What do you think about this approach?

That's a solution and would comply with the can spec, I believe, but it
should be discussed on the Socket-CAN mailing list. We should not change
policy just to avoid locking or simplify the code.

Wolfgang.