Re: [PATCH net-next v1 5/6] ptp: Support late timestamp determination

[Gerhard Engleder <gerhard@engleder-embedded.com>]

> > > > > diff --git a/drivers/ptp/ptp_clock.c b/drivers/ptp/ptp_clock.c
> > > > > index 54b9f54ac0b2..b7a8cf27c349 100644
> > > > > --- a/drivers/ptp/ptp_clock.c
> > > > > +++ b/drivers/ptp/ptp_clock.c
> > > > > @@ -450,6 +450,33 @@ void ptp_cancel_worker_sync(struct ptp_clock *ptp)
> > > > >  }
> > > > >  EXPORT_SYMBOL(ptp_cancel_worker_sync);
> > > > >
> > > > > +ktime_t ptp_get_timestamp(int index,
> > > > > +                       const struct skb_shared_hwtstamps *hwtstamps,
> > > > > +                       bool cycles)
> > > > > +{
> > > > > +     char name[PTP_CLOCK_NAME_LEN] = "";
> > > > > +     struct ptp_clock *ptp;
> > > > > +     struct device *dev;
> > > > > +     ktime_t ts;
> > > > > +
> > > > > +     snprintf(name, PTP_CLOCK_NAME_LEN, "ptp%d", index);
> > > > > +     dev = class_find_device_by_name(ptp_class, name);
> > > >
> > > > This seems expensive for every single Rx frame in a busy PTP network.
> > > > Can't this be cached in the socket?
> > > I thought that PTP packages are rare and that bloating the socket is
> > > not welcome.
> >
> > Some PTP profiles use insanely high frame rates.  like G.8275.1 with
> > Sync and Delay Req at 16/sec each.  times the number of clients.
>
> times the number of vclocks/Domains.

Getting the physical clock in __sock_recv_timestamp() is the expensive path.
The network device is already available __sock_recv_timestamp(). Timestamp
determination based on address/cookie could be done by the network device
instead of the physical clock. In my opinion, that would be a good fit, because
timestamp generation is already a task of the network device and implementation
would be faster/simpler. What do you think?

Gerhard