Re: [PATCH] net: deadlock during net device unregistration - V2

[Daniel Lezcano <dlezcano@fr.ibm.com>]

Jarek Poplawski wrote:
> On Wed, Oct 01, 2008 at 04:14:35PM +0200, Benjamin Thery wrote:
>> This is the second version of a patch aimed at fixing a deadlock that
>> can occur when unregistering net devices. 
>>
>> This new version of the patch ensures the garbage collector 
>> dst_gc_task() is run when waiting in netdev_wait_allrefs(), by calling
>> it in the netdevice notifier dst_dev_event() when receiving a 
>> NETDEV_UNREGISTER event.
>>
>> Thanks to Jarek Poplawski for proposing this fix.
> 
> Not at all! Actually, I'm sorry for this mess... (Read below.)
> 
>> (The previous version proposed to replace in linkwatch_event()
>> the call to rntl_unlock() by __rtnl_lock())
> ...
>> --- net-next-2.6.orig/net/core/dst.c
>> +++ net-next-2.6/net/core/dst.c
>> @@ -328,6 +328,10 @@ static int dst_dev_event(struct notifier
>>  			dst_ifdown(dst, dev, event != NETDEV_DOWN);
>>  		}
>>  		mutex_unlock(&dst_gc_mutex);
>> +
>> +		if (event == NETDEV_UNREGISTER &&
>> +		    cancel_delayed_work(&dst_gc_work))
>> +			dst_gc_task(&dst_gc_work.work);
> 
> Hmm... It seems this shouldn't work yet: cancel_delayed_work() can only
> kill this while on timer, but not when queued and maybe blocked already.

Perhaps, I am misunderstanding but, dst_gc_work is always called with 
schedule_delayed_work. If the task is not running, we can cancel it and 
if the task is running, no need to trigger the gc, no ?

> Probably cancel_delayed_work_sync() should be considered instead, but
> since this needs more checking, and David is waiting for this, I think
> it's safer to use this previous (__rtnl_unlock) patch for now
> (especially for -stable).