Re: [PATCHv4 net-next] bonding: add software tx timestamping support

[Hangbin Liu <liuhangbin@gmail.com>]

On Fri, Apr 14, 2023 at 06:02:05PM -0700, Jakub Kicinski wrote:
> On Fri, 14 Apr 2023 16:35:26 +0800 Hangbin Liu wrote:
> > v4: add ASSERT_RTNL to make sure bond_ethtool_get_ts_info() called via
> >     RTNL. Only check _TX_SOFTWARE for the slaves.
> 
> > +	ASSERT_RTNL();
> > +
> >  	rcu_read_lock();
> >  	real_dev = bond_option_active_slave_get_rcu(bond);
> >  	dev_hold(real_dev);
> > @@ -5707,10 +5713,36 @@ static int bond_ethtool_get_ts_info(struct net_device *bond_dev,
> >  			ret = ops->get_ts_info(real_dev, info);
> >  			goto out;
> >  		}
> > +	} else {
> > +		/* Check if all slaves support software tx timestamping */
> > +		rcu_read_lock();
> > +		bond_for_each_slave_rcu(bond, slave, iter) {
> 
> > +			ret = -1;
> > +			ops = slave->dev->ethtool_ops;
> > +			phydev = slave->dev->phydev;
> > +
> > +			if (phy_has_tsinfo(phydev))
> > +				ret = phy_ts_info(phydev, &ts_info);
> > +			else if (ops->get_ts_info)
> > +				ret = ops->get_ts_info(slave->dev, &ts_info);
> 
> My comment about this path being under rtnl was to point out that we
> don't need the RCU protection to iterate over the slaves. This is 
> a bit of a guess, I don't know bonding, but can we not use
> bond_for_each_slave() ?

Sorry, I misunderstood your comment in patchv3. I thought you agreed to use
the same use case like rlb_next_rx_slave(). I will post a new fix version.

> 
> As a general rule we should let all driver callbacks sleep. Drivers 
> may need to consult the FW or read something over a slow asynchronous
> bus which requires process / non-atomic context. RCU lock puts us in 
> an atomic context. And ->get_ts_info() is a driver callback.
> 
> It's not a deal breaker if we can't avoid RCU, but if we can - we should
> let the drivers sleep. Sorry if I wasn't very clear previously.

Thanks for the explanation.

Regards
Hangbin