Permalink
8f76c84 Nov 20, 2015
@wowczarek @gvnn3
430 lines (380 sloc) 19.7 KB
2013-10-23 Wojciech Owczarek <wojciech@owczarek.co.uk>
* 2.3.2 release
* Bug fixes / improvements since 2.3.1:
- systemd service file added and RPM spec updated
to handle systemd
- fixed segfault when shutting down NTP time service
- ptpengine:always_respect_utc_offset is now enabled
by default, to protect from clock jumps when GM
stops announcing currentUtcOffsetValid flag
- re-arranged order of InterfaceInfo fields to
fix memory alignment issues (SF bug #70)
- corrected logMessageInterval field values
for hybrid mode (SF bug #69)
- fixed NTPd check failures resulting from signals
interrupting select()
- increased statistics file line buffer size to prevent
truncation and merging of lines
- added Sequence ID to statistics log
- Solaris build fixes - linking lnsl and lsocket when
building without SNMP support
- critical: added minimum POSIX timer interval to prevent from
timers firing to quickly for the process to handle,
resulting in 100% CPU and endless signal queue
- improved processing of unicast destination, domain
and local preference lists
- CPU affinity setting moved to a separate function
- fixed incorrect PTP_UNICAST flag setting for multicast
- fixed missing management or signalling message acceptance
condition (clockid==clockid, portid=0xffff): ISPCS 2015 plugfest,
fourth missing: (clockid=0x(f), portid==portid)
- signaling.c: fixed order of grant processing for PTP messages
(ann,syn,fup,dresp,pdelayresp) instead of numeric loop,
to assist with interop where transmission request order
makes a difference to the master
- signaling.c: do not reset known clock ID when changing GMs,
to allow to keep accepting announce from other masters
- signaling.c: update unicast index when transport address
match only: fix for cache misses for non-best masters
- daemonconfig:c - reworked all config mapping macros
to functions, parseConfig now handless all operations:
parsing, quiet parsing, reload check, help - no hidden
values in dictionary anymore
- closed sf bug #64 - leaving mcast group with 0x0 address
- peer multicast group only joined when needed, not always
- logMessage now keeps hash of the last message body,
preventing from same message being repeated
- null checks getInterfaceInfo set of helper functions
in case if some data returned by getifaddrs() is null:
fixes segfaults on some exotic interface types
- signaling.c: use message type <-> smaller index number
conversion so grant table is 5-element, not 16-element:
significantly reduced memory footprint
- added dummy Makefile to src/dep to allow make commands
from that directory
- added buffer overflow protection for management
TLVs: protects against read past message length,
and against read past buffer space
* New features since 2.3.1:
- support for user variables in configuration:
- variables specified variables:name=val, and
used as @name@
- built-in variables: @pid@, @hostname@
- support for configuration templates:
- built-in
- template files
- unicast port number mask to assist with nodes
using different port IDs for negotiation
- option to accept any GM when unicast negotiation
is enabled
- UDP checksums disabled on Linux by default:
enables interop with TCs that don't correct checksums
- Added QNX support:
- clean build on QNX 6.5 out of the box
- adjfreq emulation using native ClockAdjust
- binding to CPU cores
- Experimental (but yield best results):
* interpolated getTime attached to IRQ0
* RX/TX timestamps in software on send or receive,
bypassing pcap or socket timestamping
- support for systems without getopt_long (like QNX)
- adjfreq-like behaviour on systems with adjtime only:
properly scaled adjtime() pulled into PI servo
(vastly improved sync on systems like OpenBSD and OSX),
support for drift file for those systems
- fixed preserving incorrect master address when
best master lowers priority (GM change as a result
of announce from current master, not foreign)
- PTP MIB enhanced with string values, counter support,
clearing counters and more, added some missing items
from previously supported items
- SNMP trap support for common alarm conditions
- enhanced management message support: enable/disable port,
some missing values, PTPd now properly applies
configuration when receiving SET messages
- added simple SNMP poller script: snmptpq
- SNMP MIB now using PTPd's own Enterprise number
2013-06-10 Wojciech Owczarek <wojciech@owczarek.co.uk>
* 2.3.1 release
* New features since 2.3.0:
- full unicast support both in master and in slave mode:
+ support for multiple unicast destinations for masters
+ feature-complete Telecom Profile support:
* unicast negotiation for masters and slaves
* disabling BMCA for masters (also works for multicast)
* support for slaves with masters in multiple domains
* influencing slave BMCA with local preference
* serving multiple Telecom slaves at different rates
* slaves correctly negotiate message rates over supported
range of intervals until an acceptable interval is granted
+ support for unicast Peer Delay target - exotic use case
- greatly improved network robustness
- added support for slave-only builds (configure --slave-only):
allows building a binary incapable of running as master
- simple failover mechanism - ptpengine:backup_interface
- ptpengine:max_delay_max_rejected setting to unblock ptpd
rejecting all messages when max_delay set
- ptpengine:max_delay_stable to only use max_dely when clock
is stable
- selectable log timestamp format (unix or datetime)
- ptpd now compiles and runs out of he box on NetBSD, OpenBSD, FreeBSD,
OSX and Solaris and its derivatives (OpenIndiana, OmniOS etc.)
- major rework of the ACL code
- outlier filter auto-tune to improve operation in noisy environments(high PDV)
- added support for statistical filters (min/mean/median) for sync
and delay messages - min or median can be used in envifonments with high PDV
- added small step detection (millisecond level)
- added ptpengine:clock_update_timeout setting to reset slave
if no clock updates for a period of time
- when running slave-only in unicast mode, unicast address
does not have to be specified - ptpd will wait for a GM and latch onto it
- automatic Delay Request override for Hybrid and Unicast modes
- added options allowing to step the clock on startup
- improved sequence and re-ordering checks
- added support for "any domain" operation where ptpd will sync with
any available domain
- idle timeout (PTPd hands over clock control)
- clock update timeout (PTPd restarts slave on timeout)
- maximum state machine reset count - full network restart when exceeded:
allows ptpd to survive situations where network interface is removed,
shut down or IP address changes
- support for NIST/IETF leap seconds file both for master and slave
- configurable leap second handling: accept, ignore, step
- support for leap second smear for slave operation
- option to step clock on startup
- TimingDomain API: all NTP failover code removed
in favour of a generic interface for multiple time controllers.
- configurable port number - useful when running multiple instances
- periodic log file updates with slave or master information - separate
from statistics file: allows sending statistics into syslog
- support for software transmit timestamping on Linux
- lookup tables used for matching Sync and FollowUp when
no transport information available
- added periodic status updates to log file (global:periodic_updates)
- status file will always display counters for common error events
if they are non-zero (includes denied unicast grants and ignored
Announce resulting from rejected UTC properties)
- reporting on message receive and transmit rates
in the status file
- restored compatibility with 802.1AS
* Bug fixes / improvements since 2.3.0:
- added CentOS / RHEL 7.x support
- statistics code enabled by default - can be disabled
- NTPd failover code enabled by default - can be disabled
- Improved status file - more error conditions included,
message performance statistics included
- Rewritten timer subsystem allowing for easily implementing
other implementations: included is the old interval timer implementation,
and new (now default), POSIX timer implementation. This allows for
fully standards-compliant message rates. This can be disabled with
a configure flag
- ptpd now usable on systems with no adjtimex() call like OSX,
OpenBSD and LynxOS ()
- moving standard deviation algorithm fix improving the outlier
filters
- major servo.c spaghetti code cleanups
- network code cleanup
- fixed bug #66 - Ethernet interfaces no longer need IP addresses
to run PTP Ethernet transport
- fixed bug / feature request #21: timestamping initialisation on
PHC kernels
- fixed bug where ptpd would cause 100% CPU usage because of a
late TX timestamp
- fixed bug / feature request #22: prevent FMR from overwriting
best master
- improved servo stability detection: servo no longer reporting
stable when discarding most samples or heavily slewing
- ptpd will now join and leave multicast groups only when running
multicast or hybrid: allows running unicast and multicast
independently
- ptpd will now ignore unicast packets in multicast-only mode
and vice versa. In hybrid mode, only unicast Delay Request /
response is accepted; multicast management messages are always
accepted
- further autoconf improvements across all supported OSes
- added missing files to distribution
- ptpd will now recover even from critical network failures,
such as reloading drivers
- remaining memory leaks fixed
- more code cleanup, no more ajtimex() and similar calls in protocol.c
- some refactoring, constified rtOpts wherever possible
- less "dramatic" warnings about drift file
and clock stability detection
- fixed unnecessary TX timestamp attempts in PCAP mode
- fixed issue where drift file would not be correctly
written on exit under certain conditions
- other minor bugs fixed
2013-11-22 George Neville-Neil <gnn@neville-neil.com>, Wojciech Owczarek <wojciech@owczarek.co.uk>
* 2.3.0 release
* New Features:
- IEEE 802.3 Support
- Support for reading from the Berkeley Packet Filter, resulting in lower jitter
- Configuration file support, multiple previously unsupported options
- Rewritten command line parameter handling, new parameters
- Command line options to print default configuration, print lock file, check configuration and more,
- Full help (-H) for all configuration options,
- Configuration reload support with SIGHUP,
- Rewritten logging subsystem,
- Separate log file and statistics file, SIGHUP reprints the headers
- Support for mode and interface specific lock files
- Network interface checks before startup and during config reload
- Added ./configure --sigusr2=counters - will dump PTP engine countersto current log target when SIGUSR2 received
- Support for presets - groups of PTP protocol options defining ptpd behaviour
presets available: slaveonly, masteronly, masterslave
- Support for tick adjustment as well as frequency
maximum frequency offset is now configurable and can be above 512 ppm - tick is used above 512 ppm,
- Added a compatibility option to always honour the UTC offset announced by a GM when in slave state
- Added an announce receipt timeout "grace period" for slave state
slave can wait n times timeout without fully resetting
allows a seamless failover without resetting delay values etc.,
- Added support for DELAY_DISABLED delay mode - syntonisation only
- Initial support for informing the clock subsystem about the sync status
STA_UNSYNC is unset when adjusting the clock (allows syncing hardware RTC clocks with the system clock)
also maxerror and esterror are set,
- Support for RTC clock control on Linux
- Support for legacy command line switches from previous versions
- Configurable log level for normal logs, not only debug logs.
- Generic log file handler mechanism with a simple built-in logrotate implementation
for each log file container you can specify max size and max number of files
- Extensive support for realtime statistics
online long term statistics and moving statistics
mean and std dev.
Some new features included depend on statistics support, enabled with the --enable-statistics
configure switch. Statistics rely on double precision so may be disabled for simple embedded systems
- Outlier filters based on Peirce's criterion
blocks / filters spikes from delayMS and delaySM
- Generic PI controler model
- P and I components changed from attenuations to gains
- Double precision servo allows for tighter clock control
- Clock stabilisation detection based on the standard dev of observed drift
- Clock "calibration delay"
allows a configurable delay between seeing a GM for the first time and enabling clock control
- Realtime statistics - mean and std dev of one-way delay, ofm and observed drift
- Status file: one-screen dump of information about the current state of ptpd, updated in configurable intervals.
- Support for "panic mode"
if ptpd sees offset from master above 1 second, it will pause clock updates for a given amount of time
- NTPd integration and failover
ptpd can now connect to local ntpd using mode 7 packets (code derived from ntpdc)
and using authenticated requests can enable/disable ntpd clock control.
Enable with configure --enable-ntpdc
- Dump counters + clear counters using a SIGUSR2 handler to dump all counters to the log file
configure --enable-sigusr2=counters
- Support for ACLs for management and timing messages
- Basic support for SO_TIMESTAMPING (RX, TX) on Linux
- Support for non-Ethernet interfaces on Linux (Infiniband, GRE etc.)
- Support for offset / delay asymmetry correction
- Ptpd can now prefer or require GMs with valid UTC time
- Support for setting IP DSCP values
- Periodic IGMP joins also in master mode - improved robustness against network port failures
- Support for binding PTPd to a CPU core
- Control over tick rate on Linux, allowing faster clock slewing if needed
* Bug fixes:
- TTL no longer defaults to 64 in P2P mode - in P2P it's set to 1,
- Fixed an issue whereby observed drift would be reset to 0 if it
could not be read from the drift file (empty or nonexistent drift file),
now the kernel value is used if read from file fails.
- Support for updating utmp/wtmp when stepping the clock (FR #38)
- Mean Path Delay filtering now done in float (patch #49)
- Multiple other SourceForge bugs closed: Bugs closed:
#25, #34, #26, #42, #39, #33, #46, #45, #28,
#27, #48, #54, #47, #40, #37, #53, #50, #52, #59, #60
* Cleanups and other changes:
- Major startup.c cleanup, removed the parallel daemon checks,
in favour of better lock handling and the automatic lock file
name support,
- Some minor refactoring,
- Added a TimePropertiesDS structure and removed its fields
from PtpClock
- Improved log file format with full date and interface name,
removed the "===" indents
- Delay Request interval no longer mandatory in hybrid mode,
but a warning will be issued if it's not explicitly defined -
default of 0 is used,
- Removed the experimental status of hybrid mode,
- Manual stepping of the clock no longer just resets the observed
drift: resets it or restores it based on the drift handling setting,
and sets the time immediately before writing any logs - this allows
to re-sync and stabilise time almost instantly
- Up to date man pages written
2012-06-20 George Neville-Neil <gnn@neville-neil.com>
* 2.2.2 release
* Leap second fix
2012-01-05 George Neville-Neil <gnn@neville-neil.com>
* 2.2.0 release
* Patches: 3134556, 3296405
* Added support for Mac OS X (tested on Snow Leopard and Lion)
* Protocol implementation Fixes:
* Bugs Fixed
- Client now correctly accepts the Master DelayReq rate
- DelayMS and delaySM now correctly show the sign when negative
- Sanity Flags: client now requires -g or -G for slave/master operation.
- Client can print the MAC address of its master (instead of EEUI)
- master now sends ARB timescale only with utc_offset of zero
- slave now only uses the last UTC_Offset if UTC_Valid bit is on.
- passive masters no longer become active every 12s,
- first delayreq is postponed until we receive the first sync
- -G (master with ntpd) now announces a better clock class of 13
- delayReq period is now uniformly randomized from range
- updated to the PTPv2 messages rates (sync / delayreq / announce )
- operator is warned once when the we slew the clock at maximum speed
- and several others too minor too mention
* System fixes and new features
- Frequency adjustment is now calculated in double precision
- Kernel timestamps are now in nanoseconds precision
- Timer system overhead was reduced to 16 alarms per second (was 1000)
- each reset now generates an IGMP leave/join operation to
the multicast group
- Log file is now appended, with the right permissions
- Debug messages show a timestamp
- Signals are now processed synchronously (to avoid race conditions)
- Configurable amount of logging (to avoid filling up /var/log)
- client now checks own filelock, $0 and well-known daemons.
- unicast messages can use DNS
- syslog support (-S)
- quality file can be generated with received syncs (-R)
- messages can be dumped for debug (-P)
- gnore packets that result in large deltas (-M)
- SIGUSR1 now steps the clock to the current PTP offset
- SIGUSR2 now cycles the domain number (useful for testing)
- reverted R135 timer change from integer back to floating point
- rand() is now seeded with last digits of our own mac address
- IGMP_refresh waits 100ms between drop() and add()
- checked to run without leaks inside valgrind 3.5.0
- last message received is identified by a column on the statistics log
- messages are sent to Syslog by default. reversed -S flag
- statistcs file now display /etc/ethers names (besides mac address)
- option -C is console mode with full verbosity
- startup warnings are also duplicated in stdout
- startup: lockfile is checked twice: once at init, to return
correct errorlevel, and a second time after deaemon()
- check for root uid()
- improvements in parallel daemons checking
- command line parameters are dumped at init
- Set the unicast flag when sending unicast packets
(experimental, hybrid mode only).
- Reimplemented integer64_to_internalTime not to use doubles
- Replaced divTime by div2Time
- Replaced all time.seconds = time.nanoseconds = 0 by clearTime(&time)
- Replaced all hex values by named flags
- Optimized comparison of clockIdentity in bmc.c
- Resolved issue of comparison of offsetScaledLogVariance
- Optimized bmcStateDecision not to call bmcDataSetComparison so often with the same parameters
- displayStats now uses getTime instead of gettimeofday
2011-02-01 George Neville-Neil <gnn@neville-neil.com>
* Add support for DNS lookup of timeserver for unicast.
* Add support for unicasting delay requests.
* Add code to dump packets on demand via the -P flag as well as in
response to updates that violate either the -M or -O flags.
2010-10-12 George Neville-Neil <gnn@neville-neil.com>
* 2.1.0 First main line release of PTPv2 code base
(IEEE-1588-2008)
* Add code to limit how much of an offset or delay the client is
willing to tolerate.
* Add support for BINTIME on FreeBSD which gives more accurate
packet timestamps.
* Add quality file support
* Fix significant bugs that prevented correct operation in
End-to-End mode.
* Add support for syslog.
* Add support for user configurable TTL.
* Clean up code formatting, headers, comments etc.