Really, this is an extrac complexity with no advantage (did we ever
use it?)  If users really want a small system they can choose the NONE
implementation of pp_printf.  The extra half a kilobyte saved by diag-no
is really not worth the effort:

% make clean; make -s ARCH=bare-i386 CONFIG_PRINTF_NONE=y
% size ppsi.o
   text    data     bss     dec     hex filename
  10323     208     264   10795    2a2b ppsi.o

% make clean; make -s ARCH=bare-i386 CONFIG_PRINTF_NONE=y PPSI_NO_DIAG=y
% size ppsi.o
   text    data     bss     dec     hex filename
   9792     208     264   10264    2818 ppsi.o
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

fsm.c corrently decodes the received frame, and error management is
built in (even if code in msg.c still needs to be audited to be more
careful about rogue frames).  However, when a received frame is in
error, fsm.c cannot return to the caller because it doesn't know the
current delay. Thus, zero the frame and call the current status, so
we have back a recalculated ppi->next_delay.

Meanwhile, mask with pp_verbose_frames the diagnostic message for
received frames -- and be more careful with accessing it's data.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

A number of states were setting ppi->next_delay before re-setting
it to a different value, or setting it on a state change (which now
occurs immediately)
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

fsm states where we are waitiing for a frame or for a timeout,
now correctly set next_delay to the right number of milliseconds.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The new function reports how many ms in the future is a timeout event,
to be able to correctly time when we should re-entering the state machine
and correctly spread our frames according to the specification.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The function is called by several states, but only state-slave
is checking the delay request timeout. Thus, this commit removes
the associated lines from the common function and moves them back
to state-slave, the only state that issues delay_request frames.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

All archs allocated one foreign master, but the default was 5 and
the command line could even change it.  Hardwire the number to 5
and skip the allocation by folding the array in pp_instance itself.

Anyways, the add_foreign() function needs serious review, later on.
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>

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

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

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

the pp instance is bound to a network interface, so avoid a separate
allocation.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

By post-incrementing sent_seq we had a value that was not
really the "last sent" as the comment claims.  Thus, pre-increment
it instead, so readers find the "last sent" value and don't need
to subtract 1.
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 commits removes all the additions and subtractions of
the protocol offset. Now all recv and send functions, receive
the pointer to the actual data being sent, be it the ptp payload
or the transport header.

This simplification (though based on the two internal pointers within
ppsi -- se previous commit) fixed a long standing bug with bare
architectures, that didn't receive the frames due to some related
problem in offset calculation.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This changes some things to make code more readable (I really didn't
know what the tmp_header was: now it is the "received_ptp_header".
The commit simplifies allocation, because the tx and rx buffers are
now part of ppi itself.

We now have tx_ptp and tx_frame, both pointing within tx_buffer (and
the same for rx). The pointers are the protocol side (aligned) and the
argument passed to send/recv (which may be not be 4-aligned).

At this point nothing changes, i.e. both gnu-linux flavours work and
bare-i386 is not able to received.

Ethernet mode used to reset the time to the epoch.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit has no technical effect: what works still works, and what
doesn't work still doesn't work.  Still, the inlines will do
type-checking at a later time when we have a structure for the ptp
header.

Moreover the renames allow me to trace all uses in order to change
them; it also generates build errors in wrpc-sw, forcing me to fix the new
use in there as well, when I update ppsi as submodule.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

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

All fields are zero at startup, unless explicitly inizialized
by the caller.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit moves the time operations inside the instance.  It
allows to have different time operations for each instance, which will be
useful in the future. For example, one ethernet card may White Rabbit,
and another may not be.

This commit, like moving network operations inside ppi, requires a
change in wrpc-sw.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This is needed for the next commit, where time operations become local
to the specific pp instance.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This commit moves the network operations inside the instance.  It
allows to have different operations for each instance, which will be
useful in the future. For example, one ethernet card may have hardware
timestamping and another may not.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This unifies all bare-linux common code. It's an out-of-context commit
(as I'm not adding bare-arm today) but I prefer to avoid doing the same
change twice in things that are going to change these days.

BTW: this commit introduces no changes in the binary files generated,
with the exception of a small difference between i386 and x86-64 in
the receive buffer.  This keeps the x86-64 version but i386 was not
working anyways (and still doesn't work).  That bug fix is in my queue,
but I've other things to do first with ppsi.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

pp_timed_printf is a great resource, but we need to temporarily
silence diagnostic printing in  time_ops.get, in order to be able to
tell stamping-related calls to just printf-related ones.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

I'm going to split diagnostics into groups based on flags. And
with different diagnostics for each pp instance, to be ready to
debug multi-link situations.  This commit adds the flags in the proper
places and the first flag value, used in the next commit.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Too many messages are confusing: let's make a state change more
visible than a re-enter of the same state.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

By moving calc_timeout into the time functions, we simplify
the split-out of ./time-<arch> as a separate subdirectory.

As a side effect, calc_timeout disappears from arch-wrpc, and
wrpc-sw needs to be adapted to this commit, to provide it
inside its own time_ops.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

The structure was separately allocated with a size of 16.  This avoids
the alloc and makes the array the right size from the enum of message
types. We thus a remove a number of magic "16" in the code.
wrpc-sw must be updated too, in order to use this commit.

There's still a problem: the code claims the number to be "last sent",
but actually the value is incremented after sending. I'll need to pass
again on this thing.

ps: I changed name from "sent_seq_id" to "sent_seq" to the compiler
complains for previous users, otherwise undetectable.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

We are not even trying to build the complete executable here, so
remove crt0.S and any references to it.

I forgot these pieces in commit
   arch-spec: removed unused source files
which happened 101 commits ago.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

While some parameters are still there, for future use when we'll put
back the pdelay mechanism, the default mechanism is clearly not
pdelay, and incorrectly-used parameters are abetter removed.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

Actually, there's no need to build with three pp_printf any more: we
know pp_printf is sound. So let's just print once with "xint"
pp_printf and once with all messages activated, so see the worst size
we get to.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

This fixes a bug introduced in 44deab24 (97 commits ago), where I
split WR_DSPOR(ppi) out of DSPOR(ppi), but not all diagnostics has
ever been built. Next commit fixes this, and allowed me to make this
very fix.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>

If the state just executed switched to a new state, we should enter
that new stat immediately, and ignore ppi->next_delay.
Signed-off-by: Alessandro Rubini <rubini@gnudd.com>