svn commit: r325506 - in head: sbin/ifconfig sys/net sys/sys

[Konstantin Belousov]

On Tue, Nov 07, 2017 at 11:00:33AM -0800, John Baldwin wrote:
> On Tuesday, November 07, 2017 07:44:31 PM Konstantin Belousov wrote:
> > On Tue, Nov 07, 2017 at 07:39:26PM +0200, Konstantin Belousov wrote:
> > > On Tue, Nov 07, 2017 at 09:06:52AM -0800, John Baldwin wrote:
> > > > On Tuesday, November 07, 2017 09:29:15 AM Konstantin Belousov wrote:
> > > > > Author: kib
> > > > > Date: Tue Nov  7 09:29:14 2017
> > > > > New Revision: 325506
> > > > > URL: https://svnweb.freebsd.org/changeset/base/325506
> > > > > 
> > > > > Log:
> > > > >   Add a place for a driver to report rx timestamps in nanoseconds from
> > > > >   boot for the received packets.
> > > > >   
> > > > >   The rcv_tstmp field overlaps the place of Ln header length indicators,
> > > > >   not used by received packets.  The basic pkthdr rearrangement change
> > > > >   in sys/mbuf.h was provided by gallatin.
> > > > >   
> > > > >   There are two accompanying M_ flags: M_TSTMP means that there is the
> > > > >   timestamp (and it was generated by hardware).
> > > > >   
> > > > >   Another flag M_TSTMP_HPREC indicates that the timestamp is
> > > > >   high-precision.  Practically M_TSTMP_HPREC means that hardware
> > > > >   provided additional precision comparing with the stamps when the flag
> > > > >   is not set.  E.g., for ConnectX all packets are stamped by hardware
> > > > >   when PCIe transaction to write out the completion descriptor is
> > > > >   performed, but PTP packet are stamped on port.  For Intel cards, when
> > > > >   PTP assist is enabled, only PTP packets are stamped in the limited
> > > > >   number of registers, so if Intel cards ever start support this
> > > > >   mechanism, they would always set M_TSTMP | M_TSTMP_HPREC if hardware
> > > > >   timestamp is present for the given packet.
> > > > >   
> > > > >   Add IFCAP_HWRXTSTMP interface capability to indicate the support for
> > > > >   hardware rx timestamping, and ifconfig(8) command to toggle it.
> > > > 
> > > > Hmm, other NICs (Chelsio T4 and later for example) support timestamps that
> > > > aren't in nanoseconds but some other frequency (which are themselves useful).
> > > > It would be nice to have a more flexible interface that supports not only ns
> > > > timestamps.  Perhaps a way to expose a direct hardware timestamp as a
> > > > "number" without a specific frequency?
> > > 
> > > ConnectX does not provide ns-clocked counter either.  It is some internal
> > > clock driven by a cristal with > 100MHz frequency.
> > > 
> > > There is no much space in the pkthdr, and the request to provide the
> > > timestamp was in the context where the wall clock or some closely related
> > > timer is needed.  Of course, I can put raw hardware timestamp into the
> > > packet header, but only instead of the reduced value.  Then the consumer
> > > of the timestamp would need to find the interface which received the
> > > packet and call its method to convert ?  We have only one consumer in
> > > tree (SO_TIMESTAMP) and perhaps one possible another consumer (TCP) for
> > > this data, both of which require wall clock, so would need to call into
> > > the method.
> > > 
> > > Also please see the discussion in the referenced review about accuracy of
> > > the convertion.
> > > 
> > > Important example are Intel cards where is only limited number of
> > > latched registers, and only PtP packets are stamped.  This (and some
> > > quirk in ConnectX) explains the high-precision flag.
> > > 
> > 
> > And another consideration which was one of the strong argument for me
> > when I thought about this stuff:  the convertion of the hardware timestamp
> > to the useful clock stamp depends on the clock calibraton data which might
> > not be available long time after the packet receive.  In other words, when
> > the consumer would call into the interface method to convert raw timestamp,
> > it might be already not convertable (in kern_tc.c terms, timehands were
> > switched by tc_windup()).
> 
> When using the timestamps from Chelsio NICs previously, the codebase in
> question performed conversions in userland rather than in the kernel using
> other application-specific data in the received packets to aid in mapping
> the timestamp values to wall times.  Reliably converting a random timestamp
> from the hardware to a wall time is indeed non-trivial.  I had imagined
> having a variant of SO_TIMESTAMP (or perhaps a new option) so that userland
> could request the raw timestamp.
> 
> To be clear, however, is the MLX5 timestamp steered by an on-card PTP
> implementation or is it just a free-running timer?  (The Chelsio timer is
> a free-running timer.)

It is free-running timer for non-PtP packets.

PtP packets are handled specially, and besides latching timestamp on
port (for normal packets it is on PCIe write-out) they might be also
corrected by card if parsed as PtP and internal card PtP engine synced
already. But for the later I am not sure, and was not able to confirm.