Received-SPF: pass (google.com: domain of linux-kernel+bounces-139634-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:45d1:ec00::1 as permitted sender) client-ip=2604:1380:45d1:ec00::1;
Date: Wed, 10 Apr 2024 21:36:25 -0400
From: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
To: Rahul Rameshbabu <rrameshbabu@nvidia.com>, 
 Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Cc: Jakub Kicinski <kuba@kernel.org>, 
 John Fraker <jfraker@google.com>, 
 netdev@vger.kernel.org, 
 Praveen  Kaligineedi <pkaligineedi@google.com>, 
 Harshitha Ramamurthy <hramamurthy@google.com>, 
 Shailend Chand <shailend@google.com>, 
 Willem de  Bruijn <willemb@google.com>, 
 "David S. Miller" <davem@davemloft.net>, 
 Junfeng Guo <junfeng.guo@intel.com>, 
 Ziwei Xiao <ziweixiao@google.com>, 
 Jeroen de Borst <jeroendb@google.com>, 
 linux-kernel@vger.kernel.org, 
 kory.maincent@bootlin.com, 
 andrew@lunn.ch, 
 richardcochran@gmail.com
Message-ID: <66173e9972287_2d6bc6294d1@willemb.c.googlers.com.notmuch>
In-Reply-To: <874jc8ww68.fsf@nvidia.com>
References: <20240408180918.2773238-1-jfraker@google.com>
 <661550e348224_23a2b2294f7@willemb.c.googlers.com.notmuch>
 <20240409172838.247738f3@kernel.org>
 <87jzl5akh5.fsf@nvidia.com>
 <20240410061928.712ff9a3@kernel.org>
 <6616e92cbcca_2bfabf294c5@willemb.c.googlers.com.notmuch>
 <874jc8ww68.fsf@nvidia.com>
Subject: Re: [PATCH net-next] gve: Correctly report software timestamping
 capabilities
Precedence: bulk
Mime-Version: 1.0
Content-Type: text/plain;
 charset=utf-8
Content-Transfer-Encoding: 7bit

Rahul Rameshbabu wrote:
> 
> On Wed, 10 Apr, 2024 15:31:56 -0400 Willem de Bruijn <willemdebruijn.kernel@gmail.com> wrote:
> > Jakub Kicinski wrote:
> >> On Tue, 09 Apr 2024 21:40:46 -0700 Rahul Rameshbabu wrote:
> >> > > My gut tells me we force drivers to set the ethtool op because
> >> > > while at it they will probably also implement tx stamping.  
> >> > 
> >> > I think the logic should be the other way (in terms of the
> >> > relationship). A call to skb_tx_timestamp should throw a warning if the
> >> > driver does not advertise its timestamping capabilities. This way, a
> >> > naive netdev driver for some lightweight device does not need to worry
> >> > about this. I agree that anyone implementing tx timestamping should have
> >> > this operation defined. An skb does not contain any mechanism to
> >> > reference the driver's ethtool callback. Maybe the right choice is to
> >> > have a ts capability function registered for each netdev that can be
> >> > used by the core stack and that powers the ethtool operation as well
> >> > instead of the existing callback for ethtool?
> >> 
> >> Adding a check which only need to runs once in the lifetime of
> >> the driver to the fastpath may be a little awkward. Another option
> >> would be a sufficiently intelligent grep, which would understand
> >> which files constitute a driver. At which point grepping for 
> >> the ethtool op and skb_tx_timestamp would be trivial?
> >
> > Many may not define the flags themselves, but defer this to
> > ethtool_op_get_ts_info.
> >
> > A not so much intelligent, but sufficiently ugly, grep indicates
> > not a a massive amount of many missing entries among ethernet
> > drivers. But this first attempt is definitely lossy.
> >
> > $ for symbol in skb_tx_timestamp get_ts_info SOF_TIMESTAMPING_TX_SOFTWARE ethtool_op_get_ts_info "(SOF_TIMESTAMPING_TX_SOFTWARE|ethtool_op_get_ts_info)"; do
> >     echo -n "$symbol: ";
> >     for i in `grep -nrIE "$symbol" drivers/net/ethernet/ | awk '{print $1}' | xargs dirname | uniq`; do echo $i; done | wc -l;
> >   done
> >
> > skb_tx_timestamp: 69
> > get_ts_info: 66
> > SOF_TIMESTAMPING_TX_SOFTWARE: 33
> > ethtool_op_get_ts_info: 40
> > (SOF_TIMESTAMPING_TX_SOFTWARE|ethtool_op_get_ts_info): 59
> >
> > This does not add up, but that's because some drivers share prefixes,
> > and some drivers have different paths where one open codes and the
> > other calls ethtool_op_get_ts_info. Marvell is a good example of both:
> >
> > $ grep -nrIE '(SOF_TIMESTAMPING_TX_SOFTWARE|ethtool_op_get_ts_info)' drivers/net/ethernet
> > /marvell
> > drivers/net/ethernet/marvell/pxa168_eth.c:1367: .get_ts_info    = ethtool_op_get_ts_info,
> > drivers/net/ethernet/marvell/mv643xx_eth.c:1756:        .get_ts_info            = ethtool_op_get_ts_info,
> > drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c:5266:   info->so_timestamping = SOF_TIMESTAMPING_TX_SOFTWARE |
> > drivers/net/ethernet/marvell/octeontx2/nic/otx2_ethtool.c:962:          return ethtool_op_get_ts_info(netdev, info);
> > drivers/net/ethernet/marvell/octeontx2/nic/otx2_ethtool.c:964:  info->so_timestamping = SOF_TIMESTAMPING_TX_SOFTWARE |
> 
> If there is a desire to enforce all drivers need to implement
> .get_is_info, would the following make sense?

The only reason to enforce this is if we want to enforce them to also
implement tx software timestamping. Generally, these features are opt
in.

> My biggest objection to
> this idea was mainly my concern that the drivers would miss setting
> info->so_timestamping with SOF_TIMESTAMPING_RX_SOFTWARE |
> SOF_TIMESTAMPING_SOFTWARE, which I do not think should be a
> responsibility of the driver author since this is happening in the core
> stack.
> 
> So maybe something like this (taking Willem's proposal for
> __ethtool_get_ts_info and modifying it a bit)?
> 
>         int err = 0;
> 
>         ...
> 
>         info->phc_index = -1;
> 
>         if (phy_has_tsinfo(phydev))
>                 err = phy_ts_info(phydev, info);
>         else
>                 err = ops->get_ts_info(dev, info);
> 
>         info->so_timestamping |= SOF_TIMESTAMPING_RX_SOFTWARE |
>                                  SOF_TIMESTAMPING_SOFTWARE;
> 
>         return err;

Yes, this is what I meant as well. (the code I showed was just copied
verbatim from net-next as context, not a suggested change.)

> >
> > One more aside, no driver should have to advertise
> > SOF_TIMESTAMPING_SOFTWARE or SOF_TIMESTAMPING_RAW_HARDWARE. Per
> > Documentation/networking/timestamping.rst these are reporting flags,
> > not recording flags. Devices only optionall record a timestamp.
> 
> I think this view aligns with my opinion above (though good point about
> timestamping reporting bits in general should be deduced based on the
> timestamp generation bits set rather than needing to be set as well).