This commit factorizes some code in lib/conf.c from the two hosted
startup files, but the aim is passing "-f <filename>" on the command
line, to test with unconventional setups without touching the "real"
filesystem.

Now ppsi looks for the configuration file is from threee places, in
this order:
    - command line, if any
    - architecture-specific, if any (e.g. "/wr/etc/ppsi.conf")
    - compile-time default ("/etc/ppsi.conf")
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The next commit makes a little order in startup files, and config file
parsing ends up after all allocations, so we can't reply on
ppg->nlinks to be properly set.

In hosted environments we have no issues with a few kilobytes more, so
allocate the max number of instances even if less are used.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

If we run both Ethernet and UDP services in a 18-port switch, we need
36 logical links. Thus, increase the previous limit of 32 to 64,
knowing we have no serious memory constraints in hosted environments.

The definition is moved to <ppsi/constants.h> as implementation-specific.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

When running a WR device as non-WR slave, we adjust the clock instead
of the phase.  When adjusting the counters, timestamps are reported as
invalid until two increments of the "seconds" register.

So, when T3 is not valid (tx stamp of delay_request), avoid running
the non-WR servo. A later iteration of the slave FSM will work.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

When setting time, set Unix time as well. And when asking for the time,
return Unix time if the WR inter process communication fails.

Currently (and in all past history) the IPC call to get the time is
not implemented, so we are sure we fail. But I keep it in place nonetheless
in case we'll have it working in the future. Such time_get calls are
mainly for timestamping the diagnostic messages anyways.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This function allows a servo to initialize its hardware and return the
current "observed drift" value that is in charge.

In unix-time this is used to return the current value for frequency
correction -- being consistent with current naming and use of values,
which unfortunately is not really correct.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

In my opinion, it makes no sense to make calculations with a
well-filtered OWD and the instantaneous OFM, which is subject to the
same oscillations.  Thus, this commit filters OFM with the same rules
used for OWD.

However, the detection of outliers is more difficult because OFM can
have either sign, and OFM is really changing while we are not yet
synchronized.  So the code does the following:

- if we are at PP_ADJ_FREQ_MAX it means we are bridging a gap as
fast as possible. In that case, don't average OFM.

- if we receive a bigger-than-expected OFM value, trim it, but relax
the expectation so we can track a real change if futher samples confirm
it or push it farther.

- if the averaged OFM changed sign, start averaging from scratch.

The second item works by keeping the average of ofm magnitude, and we
accept a change no bigger than the averaged magnitude.  So, if we are
synced in the millisecond range we accept changes of the order of
milliseconds (this usually applies while syncing, and OFM is expected
to change in that case); if we are in the microsecond range outliers
move us by a few microseconds, but such trimming is then relaxed so we
increase our ability to move exponentially.

The last item prevents oscillations: this OFM is fed to a PI controller,
so if we are late in responding to changes in sign, both P and I
continue pushing in the wrong direction. So, combining the average and
the PI we have an almost inversion in phase of the error, and the
servo oscillates a lot while converging.

Unfortunately, while working on this I realized there's a bug when
dealing with very short and consistent timestamps. For example, a
current average of 20ns, will never reach 40ns even if all further
samples request 40ns, due to integer truncation.  We need to either
move to floating point or count fractional bits as the magnitudes
permit.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit almost ignores long owd values. They happen when there are
software delays; for example I sometimes see 7ms RTT values every now
and then, which destroy synchronization by a few hundred microseconds.

I "almost" ignore them, by using an OWD value twice as the current
average.  This is meant to gradually recover if the real OWD is bigger
than what we initially measured (e.g., initially the slave clock was
too slow during a recovery).

The commit also slightly reorders some calculation to be safer in
dealing with errors in initial values.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit makes the servo engine discard events where "one way
delay" is over one second (this will be documented).  The management
was never tested and likely wrong anyways.

It also ignores events where the trip time appears to be negative.  In
this case the calculation is brought forward with the current value
for one way delay (the current average).  This happens when the slave
is running a faster clock than the master during recovery.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This moves the currenct checks for bad events into a specific function.
Some duplication is removed in that function and log messages are
slightly different, but no technical change is in charge.

I'm going to use pp_servo_bad_event() in later commits.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The "Use the average between current value and previous one" in the
"one way delay" calculation was actually a running-average of two
samples. This happened before the real (and properly implemented)
running average of the value.  By removing the first one we better
respect the configured width of the filter.

As a side effect the default (64 samples) is now effective, while earlier
it behaved like a 128-sample running average. It converges faster.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The "offset from master" filter was "a simple two-sample average" as
documented in the header, and the result of such average is used only
in the PI controller.

Thus, by averaging we have no effect at all in the I component and
only a spread over two samples of the P component for each sample.  In
practice, it has no effect at all.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This trivial change makes the code more readable, and more
symmetric in the two filtering sections.

Unrelated, it also moves a temporary variable to the only block where
it is used.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This rename is for consistency: the filter is called owd_fltr but the
data it acts upon were called a meaningless "mean path delay".
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This reports more information about the servo, to help me trace some
misbehaviour. At level 2 it also reports T1..T4 and at level 3 it
reports the correction fields (currently always 0 in my setup).

The big hunk in this diff is just a reindentation to remove a goto
introduced a few commits ago; it is now an else stanza. Try "git show -pw"
if in doubt.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The code was using "mean path delay" before calculating it. Which means
that is used wrong values (likely 0) the first time and the previous
value at each successive iteration.

This commit moves calculation of "mpd" before code that uses it.  The
calculation itself is not changed.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This moves the first check for "max_delay" into the calculation
function. As a side effect we got rid of the field "t1_t2_valid" in
the servo structure.

While nothing is changed in the calculations (and I see the same
(mis)behaviour in my network) we now have a sane servo.c, where
pp_servo_got_sync only takes care of the correction factor for the
sync/follow-up message and then everything is done in
pp_servo_got_resp.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This simplifies slightly the code. and has no technical effects.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

When we only have t1/t2 we can't do all the corrections, as we miss t3/t4.

This moves all the calculations from the first servo entry point
(sync/follow-up) to the second one (delay-response).  Diagnostic
prints are moved at the end as well.

Nothing else is changed by this commit: the code is just moved.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

"delay_ms" and "m_to_s_delay" were duplicates. This removes one of them
(and the same for s_to_m) and a long-standing "TODO check".

The original code was unclear about correction factor, that sometimes
was removed and sometimes not. Since the value is an overhead figure
added by any transparent clocks in the path, this commit counts it
immediately, to forget about it as soon as possible.  This also makes
the code more symmetric (t1,t2 code is similar to t3,t4 code).
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This renames "ofst_first_updated" (wrongly a runtime option) into
"t1_t2_valid" as a servo field. This makes thing more readable and
is a fix in that it allows more than one servo to coexist in the future.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit is another step in simplifying the servo. It makes the following
changes:

- format_TimeInternal is just moved (no effect)

- pp_update_offset() and __pp_update_clock() are squashed together.

They were called in sequence anyways, as two halves of a single thing.
This changes the error return paths, because the whole procedure is
aborted in case of error.  I've never seen such errors, but I think if
something is "aborted" it should not proceed.

One non-error return from the first half is turned into a goto, to avoid
changing indentation (that appears in the commit diff).

- the meaning of "ofst_first_updated" is slightly changed.

If we get an error in the sync phase the bit is cleared so delay-resp
won't use the tuple.  The field is then renamed in the next commit.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The servo implementation we brought forward from ptpd was strange
and difficult to understand. This commit is the first step in trying
to make some order, but it doesn't actually change anything but names.

The entry points for the servo are now called pp_servo_init(),
pp_servo_got_sync() and pp_servo_got_resp().  A two-step protocol
calls got_sync after getting the follow-up message.

Also, the hook for response messages is called handle_resp with a
name similar to the hooks for sync and follow-up.

The servo itself is not changed, but I reordered the functions so
the calculations appear in the real order in which they happen.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Servo logs show strings like "0.-00071881". We already had an unused
display_TimeInternal, which is correct. This turns it into an sprintf
and uses it as a static function in servo.c (only printer of
TimeInternals, it seems).

Meanwhile, this also removes the last pp_Vprintf.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

In case a port is inactive, its fd is not valid, so do not handle it
in select.
Signed-off-by: Aurelio Colosimo <aurelio@aureliocolosimo.it>