Re: dumpcap: timestamp is years ahead when in pcapng format

[Isaac Boukris <iboukris@gmail.com>]

On Thu, Sep 21, 2023 at 5:09 AM Stephen Hemminger
<stephen@networkplumber.org> wrote:
>
> On Wed, 20 Sep 2023 22:55:21 +0300
> Isaac Boukris <iboukris@gmail.com> wrote:
>
> > I found a way to get a better resolution; at init we set
> > 'pcapng_time.tsc_hz=rte_get_tsc_hz()/NSEC_PER_SEC' this way we keep
> > the number of cycles in a nano-second, then at run time we just need
> > to divide delta by this number (with no need to multiply by
> > NSEC_PER_SEC).
> >
> > The problem is I guess, that on slow systems we'll end up with
> > tsc_hz=0? Perhaps we'd need to drop to ms resolution in such a case.
> >
> > With the attach patch I get:
> >
> > 2023-09-20 10:22:13.579219 IP Rocky8 > A: ICMP echo request, id 13,
> > seq 63, length 64
> > 2023-09-20 10:22:13.580582 IP A > Rocky8: ICMP echo reply, id 13, seq
> > 63, length 64                                         3
> > 2023-09-20 10:22:14.745176 IP Rocky8 > A: ICMP echo request, id 13,
> > seq 64, length 64
> > 2023-09-20 10:22:14.746206 IP ...
> >
> > On Wed, Sep 20, 2023 at 9:53 PM Isaac Boukris <iboukris@gmail.com> wrote:
> > >
> > > I figured the first packet bug, fixed with:
> > > -       if (!pcapng_time.tsc_hz)
> > > +       if (!pcapng_time.tsc_hz) {
> > >                 pcapng_init();
> > > +               return pcapng_time.ns;
> > > +       }
> > >
> > > However I noticed a caveat with my proposed fix as it seem we only get
> > > a time resolution of one sec:
> > >
> > > 2023-09-20 09:40:20.727638 IP Rocky8 > A: ICMP echo request, id 11,
> > > seq 81, length 64
> > > 2023-09-20 09:40:20.727638 IP A > Rocky8: ICMP echo reply, id 11, seq
> > > 81, length 64
> > > 2023-09-20 09:40:21.727638 IP ...
> > >
> > > On Wed, Sep 20, 2023 at 8:59 PM Isaac Boukris <iboukris@gmail.com> wrote:
> > > >
> > > > On Tue, Sep 19, 2023 at 9:00 PM Stephen Hemminger
> > > > <stephen@networkplumber.org> wrote:
> > > > >
> > > > > On Tue, 19 Sep 2023 19:35:55 +0300
> > > > > Isaac Boukris <iboukris@gmail.com> wrote:
> > > > >
> > > > > > Looking with git log, i found the original line was:
> > > > > > return pcapng_time.ns + (delta * NSEC_PER_SEC) / rte_get_tsc_hz();
> > > > > >
> > > > > > Testing that does show a wrapping issue, e.g. (it stays around 08:05).
> > > > > >
> > > > > > 2023-09-19 08:05:24.372037 IP _gateway.domain > Rocky8.38358: 31975
> > > > > > NXDomain 0/0/0 (46)                                              10
> > > > > > 2023-09-19 08:05:21.577497 ARP, Request who-has _gateway tell Rocky8,
> > > > > > length 46
> > > > > > 2023-09-19 08:05:21.577599 ARP, Reply _gateway is-at 00:50:56:f8:92:76
> > > > > > (oui Unknown), length 46                                     13
> > > > > > 2023-09-19 08:05:22.833897 IP 192.168.202.1.50886 >
> > > > > > 239.255.255.250.ssdp: UDP, length 174
> > > > > >
> > > > > > However with my change it looks fine and always increments. I dropped
> > > > > > all the parenthesis:
> > > > > > return pcapng_time.ns + delta / pcapng_time.tsc_hz * NSEC_PER_SEC;
> > > > >
> > > > > The issue is that timestamping is in the fast path and that 64 bit divide is slow.
> > > > > Looking at other alternatives.
> > > >
> > > > Then perhaps we can keep the division optimization and just get rid of
> > > > the overflow check, relying on the change to multiply by NSEC_PER_SEC
> > > > after the division.
> > > >
> > > > With the below change only the first packet is from 2257 while all
> > > > subsequent packets are fine. But if I keep the overflow check and only
> > > > change to multiply after the division, then all packets are shown from
> > > > 2257.
> > > >
> > > > [admin@Rocky8 dpdk]$ git diff lib/pcapng/rte_pcapng.c
> > > > diff --git a/lib/pcapng/rte_pcapng.c b/lib/pcapng/rte_pcapng.c
> > > > index 80d08e1..fa545cd 100644
> > > > --- a/lib/pcapng/rte_pcapng.c
> > > > +++ b/lib/pcapng/rte_pcapng.c
> > > > @@ -79,7 +79,7 @@ static uint64_t pcapng_tsc_to_ns(uint64_t cycles)
> > > >          * Currently all TSCs operate below 5GHz.
> > > >          */
> > > >         delta = cycles - pcapng_time.cycles;
> > > > -       if (unlikely(delta >= pcapng_time.tsc_hz)) {
> > > > +       if (0 && unlikely(delta >= pcapng_time.tsc_hz)) {
> > > >                 if (likely(delta < pcapng_time.tsc_hz * 2)) {
> > > >                         delta -= pcapng_time.tsc_hz;
> > > >                         pcapng_time.cycles += pcapng_time.tsc_hz;
> > > > @@ -92,8 +92,9 @@ static uint64_t pcapng_tsc_to_ns(uint64_t cycles)
> > > >                 }
> > > >         }
> > > >
> > > > -       return pcapng_time.ns + rte_reciprocal_divide_u64(delta * NSEC_PER_SEC,
> > > > -
> > > > &pcapng_time.tsc_hz_inverse);
> > > > +       return pcapng_time.ns + rte_reciprocal_divide_u64(delta,
> > > > +
> > > > &pcapng_time.tsc_hz_inverse) * NSEC_PER_SEC;
> > > >  }
>
> This is less accurate. The TSC (CPU clock frequency) is not necessarily
> an even multiple of nanoseconds.
>
> If you want to send patches please follow the contributing guidelines
> and run checkpatch on them.

Yeah, I realized that and tried to improve on it by dividing by less
at init and multiplying by less at run time. However I noticed another
problem with my patches, that there is a time gap that keeps growing
for some reason, and I can't figure out what's wrong. I'll try some
more and will gladly test anything proposed.

Thanks!