This modifies all transports to use a new common address type, struct
address. This address is stored in a ptp_message for all received messages.
For sending, the "default" address is used with the default sending
functions, transport_send and transport_peer. The default address depends on
the transport; it's supposed to be the multicast address assigned by the
transport specification.
Later, a new transport_sendto function will be implemented that sends to the
address contained in the passed ptp_message.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
The callers of those functions are all using ptp_message. As we're going to
return more information (the address), let those functions just fill in the
ptp_message fields directly.
Some minor reshuffling needed to prevent circular header dependencies.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
In order to be able to convert to a generic address struct, separate source
and destination address into separate fields.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
When less bytes than the header size is read, do not indicate to the caller
that the read was successful, as the caller would read uninitialized memory.
To achieve that, subtract the header size unconditionally (unless an error
was returned by sk_receive).
In addition, do not check for Ethernet type when full Ethernet header was
not read. This again may lead to reading of uninitialized memory.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
The task of preparing the message for transmission and sending it appears
at many places. Unify them into a new function.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
Currently, it is assumed that the management TLV data of management COMMAND
messages is always empty. This is not true for the INITIALIZE command and
also for a custom command we'll be introducing.
Move the check to msg_post_recv and let it check only the TLVs defined by
the standard.
Signed-off-by: Jiri Benc <jbenc@redhat.com>
Most PHC drivers implement stepping (ADJ_SETOFFSET) by reading the
clock, adjusting the value by the offset and writing it back. This is
not perfectly accurate and if the operation is slow (e.g. due to PCIe
latencies), the error can be in microseconds.
Increase the default first step threshold from 100 nanoseconds to 20
microseconds to step only when the initial offset is larger than
the error in the step.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
With the new linreg servo the frequency offset and time offset are
controlled separately. The ratio between master's frequency and the
current frequency of the local clock is known and can be used when
calculating delay or peer delay to improve their accuracy.
This greatly improves the stability of the delay when the servo is
correcting a large offset.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
This servo uses linear regression to estimate current time and
frequency error. The number of points used in the regression is
variable (from 4 to 64 in powers of 2) and is selected by a long-term
statistic of the prediction error.
Future improvements could include tracking of sudden frequency changes
(e.g. due to temperature variations), better stability of the error
statistic when a large offset is corrected, options to set the speed of
the adaptation, minimum and maximum number of points, or an option to
prefer frequency accuracy over time accuracy.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
These settings will be useful for all implemented servos, so move them
to the common servo code to avoid duplication. The configuration options
are renamed, but the they can be still set by their old names.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
The initial check if the clock should be stepped always passed after the
servo was reset (e.g. from clock sanity check). Fix the condition to
step only when the threshold is exceeded.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
Setting GRANDMASTER_SETTINGS_NP is allowed only from UDS, but
PORT_DATA_SET can be set from any port. Move the port check to
clock_manage(), so all clock and port SET actions are restricted
to the UDS port.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
This fixes cumulativeScaledRateOffset reported in TIME_STATUS_NP before
the rate ratio is updated from port.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
When peer delay is < min_neighbor_prop_delay the port is flagged
as non 802.1AS capable. min_neighbor_prop_delay defaults to -20ms.
Signed-off-by: Delio Brignoli <dbrignoli@audioscience.com>
Previouly the peer delay was not taking into account the
frequency offset between the local clock and the peer's clock.
Reset neighborRateRatio to 1.0 in port_nrate_initialize().
Signed-off-by: Delio Brignoli <dbrignoli@audioscience.com>
Commit e425da2f inadvertently enabled the announce timer on the UDS port,
causing it to continually reopen the socket when in slave mode. This patch
fixes the issue by passing zero in the 'span' field of the new function,
set_tmo_random, which disables the timer again.
Signed-off-by: Richard Cochran <richardcochran@gmail.com>
Reported-by: Rohrer Hansjoerg <hj.rohrer@mobatime.com>
The 'struct interface' for the UDS port is allocated on the stack, and
when this gets passed to port_open, the port instance takes a pointer to
the string in the 'name' field. Currently this is harmless for the UDS
port, but it is not good form. This patch fixes the issue by making the
interface part of the clock instance.
Signed-off-by: Richard Cochran <richardcochran@gmail.com>
Reported-by: Miroslav Lichvar <mlichvar@redhat.com>
Store the current frequency and return it when the servo is in unlocked
state instead of zero. This fixes values printed in log messages.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
Negative path delay measurements are expected with E2E, the user doesn't
need to know when that happens.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
When the drift value is adjusted by the newly measured frequency offset,
multiply the frequencies instead of adding the measured offset to the
old value to get accurate result even when updating a large drift.
[ RC: use a simpler form of the frequency update calculation. ]
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
Signed-off-by: Richard Cochran <richardcochran@gmail.com>
Although it does not need to, the UDS port tries to allocate a filter,
just as any normal port would. Since the given length is zero, the
filter tries to allocate storage of size zero. When running with
uClibc, calloc(1, 0) returns NULL, but glibc is apparently returning
"a unique pointer value that can later be successfully passed to
free()." Both behaviors are allowed (see MALLOC(3)).
This patch works around the issue by letting the UDS port have a
filter of length one.
Signed-off-by: Richard Cochran <richardcochran@gmail.com>
This patch lets the clock release allocated memory when shutting down. The
main benefit is just to exhibit good form and to make valgrind happier.
Signed-off-by: Richard Cochran <richardcochran@gmail.com>
According to 802.1AS, ports are always expected to transmit announce
messages, even if they never want to become the grand master. Instead
of using a slave only BMC state machine as in 1588, 802.1AS offers a
"grand master capable" flag which allows clocks to not send sync
messages.
This patch keeps a port from transmitting sync (but not announce)
messages when there is no other master.
Signed-off-by: Richard Cochran <richardcochran@gmail.com>
[ RC: the point is that a port may not be considered capable until
enough messages to compute the ratio have been received. ]
Signed-off-by: Delio Brignoli <dbrignoli@audioscience.com>
Reviewed-by: Richard Cochran <richardcochran@gmail.com>
Sync rx timeout should be set only after receiving the first sync, see
section 10.2.7, figure 10-4 PortSyncSyncReceive state machine in 802.1AS
Signed-off-by: Delio Brignoli <dbrignoli@audioscience.com>
Add new options delay_filter and delay_filter_length to select the
filter and its length. They set both the clock delay filter and the port
peer delay filter. The default is now moving median with 10 samples.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
Median filter has an advantage over moving average that it is much less
sensitive to outliers.
For instance, it allows much faster recovery from an external clock
time step which happened between receiving sync message and sending
delay_req message. The measured delay includes a large error, but the
median is still a good estimate of the delay and the first step
correction applied by the servo is right.
In this implementation the median update has linear time complexity.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
Similarly to the servo interface, allow multiple filters to be
used for delay filtering. Convert mave to the new interface.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
This patch cleans up the error message for the Tx timestamp polling, and adds a
line which indicates where the bug might have originated. This should help a
user more easily diagnose what went wrong.
Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Include also nanoseconds from the current time in the srandom() call.
This should significantly decrease the chance of two machines using the
same random sequence.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
Instead of maintaining a table of precalculated values, use the
newly added set_tmo_random() function to set the delay request timeout.
It saves some memory and improves the timeout granularity, but has a
higher computational cost. It follows the requirements from section
9.5.11.2 of the spec.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
Jiri Benc
Miroslav Lichvar
Delio Brignoli
Richard Cochran
Stephan Gatzka
Jacob Keller