for tx stamps, when we added the check about the frame content
(to avoid problems with high data load, see 4179fcdc), it got broken
for UDP. That's because the cmesg we get back includes a raw frame
anyways, even if we transmitted UDP or other protocols.

So now check for the trailing part of the frame we get back, to
avoid any extra leading header whatever its lenght.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

During track-phase we used to find the setpoint and then track changes in
the master-to-slave delay, thus not zeroing any initial error that may
happen (e.g. the initial sample was more jittery than average).

Now we use the clock offset as correction source. Setpoint is
adjusted by 1/4th of the current offset, to smooth a little the jitter.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This is likely the result of my cleanup of wr-servo, where I forgot
some pieces.

Greg collected some interesting logs in wrpc where the setpoint
was calculated wrongly. This fixes the thing and removes some redundancy.

I "git diff" order:

- avoid WR_UNINITIALIZED, really unused. Do what's needed at
  wr_servo_init() time

- adjust phase to 0 at init time, where s->cur_setpoint is set (we
  really should set and use in a single unified place)

- set delta_ms_prev before entering TRACK_PHASE, not earlier where it
  doesn't fit

- avoid extra checks of non-zero offset.seconds and
offset.nanoseconds, because they are checked at the beginning anyways.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This adds -Wstrict-prototypes -Wmissing-prototypes and takes actions
accordingly.

Some of the solutions are ugly, especially in the bare architectures,
but I prefer to be warned if I take dangerous paths wrt prototypes while
writing further code.

I'm upset by those two not being included in -Wall, as they used to.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The existing rule to check whether WRPCSW_ROOT is properly set (by
environment or default in Kconfig) forced a dependency on all object
file.  Thus, all files were rebuilt every time.

We now force the check only as a dependency for the first object file,
which means ony fsm.o (and then ppsi.o) is built if you run "make" on
a just-built source tree.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This reverts commit 1d5a461d.

actually, the error count must be zeroed every time, not only after
printing: we want to print if the error persists for 5 iterations, not
every 5th time.

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

This is a temporary workaround. We are going to fix the tx/rx functions
to received the header and payload together, but currently they are
separate so the offset must be 0 for raw ethernet.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Make max offset range from -500ps to +500ps, instead of -1000ps to +1000ps
Signed-off-by: Adam Wujek <adam.wujek@cern.ch>

Fix problem with hanging send in PPSI

Even under load, with queues full, the wrs delays frames less than one ms.

Thus, if we didn't get a timestamp in 1ms, we won't ever get it. Add
some software delay and clearly we don't need 100ms as in current
code.  Please note that all state machines are not running during this
wait, as we are single-process single-thread.

So now we wait for 6ms at most (3 times 2ms at most, but the first
time we may get an older timestamp, then discarded).
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This turns the while(1) with internal retry++ to a loop based on retry.
With two changes:

1- when recvmsg() gets EAGAIN (or other errors), we loop over instead
of returning. Maybe this doesn't happen anymore thanks to the change
from POLLIN to POLLERR (previous commit), but this is philosophically
correct anyways.

2- we don't print anything when poll() returns 0, as these messages
were redundant, especially when there were many in a row. And if
nothing is there we expect to get 0. So we removed these:

   poll_tx_timestamp: poll() = 0 (Success)

Still, knowing the timestamp was late is useful, but a single message
suffices, and it's there now.  If you wonder why we loop instead of
using a longer timeout, the reason is we must get rid of timestamps
for older frames (the ones that timeout at the previous iteration).
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

It may happen, though rarely, that frames are not sent (or not freed
at kernel level), and we build up an output queue, so at some point we
get -EAGAIN on sending.

With blocking writes, this situation was stalling the whole process.
This patch helped us diagnosing a problem in our own network driver,
which had a memory leakage, but using non-blocking sockets and
reporting erors is the right thing to do anyways.

This patch reports the error using pp_diag with a priority 0; thus
the message is always printed, but by virtue of pp_diag we have the
name of the port automatically added.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

While the manual pages say nothing, https://lwn.net/Articles/291793/
hints that we should use POLLERR instead of POLLIN.

I hoped to get rid of some of the poll messages, but nothing changes
in simple tests. I hope we get rid of the EAGAIN in the later recvmsg()
(likely with POLLIN we got the socket as readable, but nothing was
in the error queue -- we had a new incoming frame instead).
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

commit 3b985b02 added the new vlan-aware issue function, but didn't
remove the old call. Thus, wrs is sending out announce messages twice
in a row (with two proper sequence numbers).
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

The old code used to discard every other tuple, because when "wait for
hardware was set", the code cleared the flag and did nothing. This was
exposed by one of the clean-up passes, but I chose not to change the
behaviour.

Now I do the normal work when wait_for_hw gets cleared.  However, as a
side effect, wr_mon was always saying "wait for hardware", because any
action would set the flag.

Thus, the solution is not setting the flag in TRACK_PHASE, but only
in the initial SYNC_SEC and SYNC_NSEC states. The code then checks
if the hardware is busy, irrespective of any flag.  The result in wr_mon
is matching reality, and the code works twice as fast as it used to.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The previous code to set system time from WR time is not really
working.

It is implemented in wrs_time_ops->set and it relies on ppsi being
non-wr slave before being wr slave, but this should not happen.  Such
code, with its "weird" message, remains in place to be used when we
run in non-wr mode (as slaves).

This new code is the right thing to do: when we are done syncing the
"big steps" in WR, and start dealing with sub-tick, we jump system
time.

There is no correction for TAI/UTC because the time operations for
arch-unux (set/get) already do the conversion.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Since commit f3d427ee, we lost the concept of "next_state" (to make code
simple), but this means that we should report to wr_mon the state string
we acted on, not the next. Otherwise, we see SYNC_PHASE all the time,
when we are actually waiting for SYNC_SEC or SYNC_NSEC to complete the
adjustment.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

Earlier we only called it in from slave state, but it's correct to do
it in any state (thus also from st_com_master_handle_announce, used by
listening and master mode.

This makes WR syncing much faster.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

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

Sometimes we get error in the recv(MSGQUEUE), with EAGAIN. Likely
we got a real frame (so poll for reading is ok) but no stamp yet.
This unsynced the transmitted frames and the error queue.

This fix, i.e. retry if getting nothing, seems to fix the (rare) problem.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

If we were offset by a whole second, we'd remain in SYNC_SEC (SYNC_TAI)
without moving to SYNC_PHASE. It happens when you change the second counter
by hand, but it's very unlikely in the initial setup.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Signed-off-by: Adam Wujek <adam.wujek@cern.ch>

What happened, sometimes, is that the phase setpoint take a meaningless
(and high) value. This is because we were into SYNC_NSEC and moved
to SYNC_PHASE without checking. Maybe because of other problems too.
That situation led to a lockup of the softpll, which tries to
reach an unreachable setpoint.

Now we force to stay in SYNC_NSEC if the offset is more than one cycle
(and SYNC_SEC if more than one second).

As a side effect, the phase setpoint is always 0..16ns (while earlier
I got setpoints of 1.5 of 17.5 etc, up to 64.5 for the same setup and
same temperature).

The problem here is: what is going to happen if, by thermal effect,
the phase setpoint crosses the 0..16n interval?
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

Tune SNMP_MAX_OFFSET_PS and SNMP_MAX_DELTA_RTT_PS
Signed-off-by: Adam Wujek <adam.wujek@cern.ch>

Remove the n_err_rxtx_deltas from the struct wr_servo_state since it is not used
anymore. Checking deltas was moved directly to SNMPD.
Signed-off-by: Adam Wujek <adam.wujek@cern.ch>

Checking whether deltas are zero is moved to SNMP.
Signed-off-by: Adam Wujek <adam.wujek@cern.ch>

When changing the state of the WR servo we can call a hook (function
pointer) to a function defined by a user in the upper layer (wrpc-sw).
The purpose of the hook is to add new functionalities when we reach a
specific state (Alarms, IRIGB, etc...) without needing to change the
code of PPSI.

[This is a rework of the original patch by Benoit. Mainly, I added
the extra integer argument to the hook (so I call it twice: both
entering and leaving the servo), and I removed the function to
set the hook from wrpc: the main function can set the pointer directly]
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

The switch is padding short frames sent from the CPU in the nic core.
When the frame is then being sent to an "access" port, the tag is
removed, and thus the frame gets shorter.  The net effect is that
sync/f-up/delay-rep were sent short, and discarded by the recipient.

Greg could move padding to the endpoint, but this extra work is not
really needed, as nobody else is expected to get frames from the CPU
over the SFP ports. So we chose to fix transmission in software and
avoid to change the gateware.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The way vlan frames are received is different on wrs than in normal
Linux. Likely my own driver (for wrs hardware) is not doing the right
thing.

So, when we receive tagged frames, they reach us as-is (not as the
payload only, with "aux" data on a side). This commit considers this
case.

The commit also fixes a pair of other details -- mostly, the amount of
"short packet of 0 bytes, ignored", by returning -2 to mean "drop with
no error".
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

SQUASH fix for ARCH_WRS
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

When we support multiple vlans, we must issue announce to all
supported vlans. Same applies to sync/follow-up. This changes
state-master.c to do the sending several times.

Initially I changed msg.c (for the announce message only), this saved
the repeated call to msg_pack_announce() but by doing it at
state-master level I have simpler code and layer consistency (doing
sync/fup at msg.c level would be a mess).
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>