Re: [PATCH net-next] bonding: add software timestamping support

[Hangbin Liu <liuhangbin@gmail.com>]

On Wed, Mar 29, 2023 at 09:12:44PM -0700, Jay Vosburgh wrote:
> >> Would it make sense to check if all devices in the bond support
> >> SOF_TIMESTAMPING_TX_SOFTWARE before returning it for the bond?
> >> Applications might expect that a SW TX timestamp will be always
> >> provided if the capability is reported.
> >
> >In my understanding this is a software feature, no need for hardware support.
> >In __sock_tx_timestamp() it will set skb tx_flags when we have
> >SOF_TIMESTAMPING_TX_SOFTWARE flag. Do I understand wrong?
> 
> 	Right, but the network device driver is required to call
> skb_tx_timestamp() in order to record the actual timestamp for the
> software timestamping case.
> 
> 	Do all drivers that may be members of a bond return
> SOF_TIMESTAMPING_TX_SOFTWARE to .get_ts_info and properly call
> skb_tx_timestamp()?  I.e., is this something that needs to be checked,
> or is it safe to assume it's always true?
> 
> 	If I'm reading things correctly, the answer is no, as one
> exception appears to be IPOIB, which doesn't define .get_ts_info that I
> CAN Find, and does not call skb_tx_timestamp() in ipoib_start_xmit().

Oh.. I thought it's a software timestamp and all driver's should support it.
I didn't expect that Infiniband doesn't support it. Based on this, it seems
we can't even assume that all Ethernet drivers will support it, since a
private driver may also not call skb_tx_timestamp() during transmit. Even if
we check the slaves during ioctl call, we can't expect a later-joined slave
to have SW TX timestamp support. It seems that we'll have to drop this feature."

Thanks
Hangbin