Re: [PATCH net-next V1 00/11] Mellanox 100G extending mlx5 ethtool support

[Saeed Mahameed <saeedm@dev.mellanox.co.il>]

On Tue, Apr 26, 2016 at 11:03 PM, David Miller <davem@davemloft.net> wrote:
>
> Series applied.
>
> Be careful about what feature update semantics you implement.  If you get a request
> to change multiple feature bits, and even one of them cannot be done, the entire
> update operation should fail and the bits should revert to what they were before
> the request.

It will be a nightmare to rollback in such case.  What if the rollback failed ?

>
> Is that how the code behaves now?
>

Not exactly.  If the set/clear of Bit(I) is performed successfully we
set/clear that feature in dev->features ourselves,
and if one fails we return a negative error code so
__netdev_update_features will notify on the failure, but we will be
left with the features that we did success to modify.

__netdev_update_features@net/core/dev.c

if (unlikely(err < 0)) {
        netdev_err(dev,
                        "set_features() failed (%d); wanted %pNF, left %pNF\n",
                        err, &features, &dev->features);
        /* return non-0 since some features might have changed and
         * it's better to fire a spurious notification than miss it
         */
         return -1;
}

>From this I understand it is ok to update the dev->features inside the
device's set_features NDO and return a negative value in case of
failure,
and a notification on changed features will be fired in anycase.

Is this a correct assumption ?

Anyhow if you think this behavior should be fixed, we will fix it.
Failures at set_features should be rare and should be debugged rather
than ignored.

Saeed.