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:
>>  
>>> 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().

Wolfgang.