Re: [PATCH] dropmon: add ability to detect when hardware dropsrxpackets

[Stephen Hemminger <shemminger@vyatta.com>]

On Fri, 15 May 2009 20:11:57 +0200
Eric Dumazet <dada1@cosmosbay.com> wrote:

> Neil Horman a écrit :
> > +static int dropmon_net_event(struct notifier_block *ev_block,
> > +			unsigned long event, void *ptr)
> > +{
> > +	struct net_device *dev = ptr;
> > +	struct dm_hw_stat_delta *new_stat = NULL;
> > +	int found = 0;
> > +
> > +	switch (event) {
> > +	case NETDEV_REGISTER:
> > +		new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
> > +
> > +		if (!new_stat)
> > +			goto out;
> > +
> > +		new_stat->dev = dev;
> > +		INIT_RCU_HEAD(&new_stat->rcu);
> > +		spin_lock(&trace_state_lock);
> > +		list_add_rcu(&new_stat->list, &hw_stats_list);
> > +		spin_unlock(&trace_state_lock);
> > +		break;
> > +	case NETDEV_UNREGISTER:
> > +		rcu_read_lock();
> > +		list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
> > +			if (new_stat->dev == dev)
> > +				new_stat->dev = NULL;
> > +				found = 1;
> > +				break;
> > +		}
> > +		rcu_read_unlock();
> 
> This is racy, unless caller already owns a lock.
> 
> If caller aleady owns a lock, you dont need :
> 
> rcu_read_lock()
> list_for_each_entry_rcu()
> rcu_read_unlock();

RTNL mutex is always held on notification call backs.
Actually why is trace_state_lock needed at all? Why not
just use the RTNL mutex?



--