Patrick McHardy <kaber@trash.net> writes:
> On Thu, May 02, 2013 at 04:16:25AM -0400, David Miller wrote:
>> From: Patrick McHardy <kaber@trash.net>
>> Date: Thu, 2 May 2013 09:03:37 +0200
>> 
>> > I'll also have a look at this.
>> 
>> By the mere existence of /sys/devices/${DEV_PATH}/net/${netdev_name}/flags
>> we have to preserve the bit layout.
>> 
>> So Linus was right.
>> 
>> So network manager is probably reading that flags sysfs file and
>> interpreting it.
>
> Right, that seems plausible. 
>
>> I'll fix the layout to how it was before.
>
> I also found one spot in net/core/dev.c which was using an int for the
> features. Patch attached.

And a couple more attached.

I am also wondering about the consequences of the
ETHTOOL_DEV_FEATURE_WORDS calculation in ethtool.c.  Adding the new
netdev features will make it go from 1 to 2:

#define ETHTOOL_DEV_FEATURE_WORDS	((NETDEV_FEATURE_COUNT + 31) / 32)


But this constant seems to be part of the userspace API AFAICS, so it
cannot just change like that:

static int ethtool_set_features(struct net_device *dev, void __user *useraddr)
{
	struct ethtool_sfeatures cmd;
	struct ethtool_set_features_block features[ETHTOOL_DEV_FEATURE_WORDS];
	netdev_features_t wanted = 0, valid = 0;
	int i, ret = 0;

	if (copy_from_user(&cmd, useraddr, sizeof(cmd)))
		return -EFAULT;
	useraddr += sizeof(cmd);

	if (cmd.size != ETHTOOL_DEV_FEATURE_WORDS)
		return -EINVAL;

..


Is this correctly analyzed?  If so, then I have no clue how to fix
that...



Bjørn