Re: [PATCH v4 net-next 04/10] net/ncsi: Ethtool operation to get NCSI topology

[Gavin Shan <gwshan@linux.vnet.ibm.com>]

On Thu, May 04, 2017 at 02:49:33AM +0200, Andrew Lunn wrote:
>> +void ncsi_ethtool_register_dev(struct net_device *dev)
>> +{
>> +	struct ethtool_ops *ops;
>> +
>> +	ops = (struct ethtool_ops *)(dev->ethtool_ops);
>
>Why do you need the cast here?
>
>Ah, is it because net_device has:
>
>          const struct ethtool_ops *ethtool_ops;
>
>i.e. you are casting off the const.
>
>> +	if (!ops)
>> +		return;
>> +
>> +	ops->get_ncsi_channels = ncsi_get_channels;
>
>and here you modify it. Which is going to blow up, because it will be
>in a read only segment and should throw an opps when you write to it?
>
>You need to export ncsi_get_channels, and let the underlying driver
>add it to its own ethtool_ops.
>

I didn't see oops because of this on the ARM board where I did the
testing. The intention was to make ethtool invisible to drivers, for
simplicity. However, we need to expose ethtool to driver when the HW
and SW statistics are conveyed by ETHTOOL_GSTATS as agree'd in another
thread.

Cheers,
Gavin