[PATCH net-next v1 4/4] net: lan966x: generate software timestamp just before the doorbell

[Vladimir Oltean]

Horatiu,

On Thu, May 08, 2025 at 11:41:56AM +0200, Horatiu Vultur wrote:
> > > > -       skb_tx_timestamp(skb);
> > >
> > > Changing this will break the PHY timestamping because the frame gets
> > > modified in the next line, meaning that the classify function will
> > > always return PTP_CLASS_NONE.
> > 
> > Sorry that I'm not that familiar with the details. I will remove it
> > from this series, but still trying to figure out what cases could be.
> > 
> > Do you mean it can break when bpf prog is loaded because
> > 'skb_push(skb, IFH_LEN_BYTES);' expands the skb->data area?
> 
> Well, the bpf program will check if it is a PTP frame that needs to be
> timestamp when it runs ptp_classify_raw, and as we push some data in
> front of the frame, the bpf will run from that point meaning that it
> would failed to detect the PTP frames.
> 
> > May I ask
> > how the modified data of skb breaks the PHY timestamping feature?
> 
> If it fails to detect that it is a PTP frame, then the frame will not be
> passed to the PHY using the callback txtstamp. So the PHY will timestamp the
> frame but it doesn't have the frame to attach the timestamp.

While I was further discussing this in private with Jason, a thought
popped up in my head.

Shouldn't skb_tx_timestamp(skb); be done _before_ this section?

	/* skb processing */
	needed_headroom = max_t(int, IFH_LEN_BYTES - skb_headroom(skb), 0);
	needed_tailroom = max_t(int, ETH_FCS_LEN - skb_tailroom(skb), 0);
	if (needed_headroom || needed_tailroom || skb_header_cloned(skb)) {
		err = pskb_expand_head(skb, needed_headroom, needed_tailroom,
				       GFP_ATOMIC);
		if (unlikely(err)) {
			dev->stats.tx_dropped++;
			err = NETDEV_TX_OK;
			goto release;
		}
	}

The idea is that skb_tx_timestamp() calls skb_clone_tx_timestamp(), and
that should require skb_unshare() before you make any further
modification like insert an IFH here, so that the IFH is not visible to
clones (and thus to user space, on the socket error queue).

I think pskb_expand_head() serves the role of skb_unshare(), because I
see skb_header_cloned() is one of the conditions on which it is called.

But the problem is that skb_header_cloned() may have been false, then
skb_tx_timestamp() makes skb_header_cloned() true, but pskb_expand_head()
has already run. So the IFH is shared with the clone made for TX
timestamping purposes, I guess.

Am I completely off?

Also, I believe you can set dev->needed_headroom = IFH_LEN_BYTES,
dev->needed_tailroom = ETH_FCS_LEN, and then just call
skb_ensure_writable_head_tail().