Re: [PATCH net-next v7 15/16] net: ethtool: ts: Let the active time stamping layer be selectable

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 22 Nov 2023 16:08:50 +0200 Vladimir Oltean wrote:
> My understanding of Jakub's email was that he wants to see the functionality
> offered by SIOCGHWTSTAMP and SIOCSHWTSTAMP converted to netlink. I don't
> think that ethtool is the correct netlink family for that, given that
> these aren't ethtool ioctls to begin with. Maybe the new netdev netlink
> family. The conversion in its basic form would offer exactly the same
> functionality.

Well, ethtool has been the catch all for a lot of random things
for the longest time. The question is whether we want to extend
ETHTOOL_GET_TS_INFO or add a third API somewhere else. And if we
do - do we also duplicate the functionality of ETHTOOL_GET_TS_INFO
(i.e. getting capabilities)?

My vote is that keeping it in ethtool is less bad than 3rd API.

> The _listing_ of hwtstamp providers is what could be done through ethtool
> netlink, similar but not identical to the way in which you are proposing
> today (you are presenting blanket "layers" which correspond to netdev and
> phylib, rather than individual providers).
> 
> The concept of an "active phc_index" would not explicitly exist in the
> UAPI. Thus I'm not sure what's with this TSINFO_SET being floated around.
> The only thing would exist is a configurable rx_filter and tx_type per
> hwtstamp provider (aka "{phc_index, qualifier}"). User space will have
> to learn to select the hwtstamp provider it wants to configure through
> netlink, and use for its class of traffic.

"Active provider" is the one that has TX_ON, rx != FILTER_NONE, right?

> This is why I mentioned by ndo_hwtstamp_set() conversion, because
> suddenly it is a prerequisite for any further progress to be done.
> You can't convert SIOCSHWTSTAMP to netlink if there are some driver
> implementations which still use ndo_eth_ioctl(). They need to be
> UAPI-agnostic.

Right, definitely.

> I'm not sure what's with Richard's mention of the "_2" variants of the
> ioctls. Probably a low-effort suggestion which was a bit out of context.
> His main point, that you cannot extend struct hwtstamp_config as that
> has a fixed binary format, is perfectly valid though. This is why
> netlink is preferable, because if done correctly (meaning not with
> NLA_BINARY attributes), then it is much more extensible because all
> attributes are TLVs. Use NLA_BINARY, and you will run into the exact
> extensibility issues that the ioctl interface has.