check for unexpected changes in frequency offset (#2150815)

Resolves: #2150815
This commit is contained in:
Miroslav Lichvar 2023-01-05 15:21:23 +01:00
parent 43e8a98327
commit 483030cf38
2 changed files with 211 additions and 0 deletions

208
linuxptp-freqcheck.patch Normal file
View File

@ -0,0 +1,208 @@
commit 513c7457865dba262a43e03fbe9178f4b08ba319
Author: Miroslav Lichvar <mlichvar@redhat.com>
Date: Mon Oct 24 15:25:30 2022 +0200
Drop support for old kernels returning zero frequency.
Kernels before 3.10 had a bug in reading of the system clock frequency,
which was worked around by commit da347d7a36f2 ("ptp4l: Set clock
frequency on start").
Drop this workaround and support for the old kernels to make
clockadj_get_freq() useful.
(Rebased to 3.1.1)
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
diff --git a/clock.c b/clock.c
index f3df220..9079428 100644
--- a/clock.c
+++ b/clock.c
@@ -1147,11 +1147,6 @@ struct clock *clock_create(enum clock_type type, struct config *config,
if (c->clkid != CLOCK_INVALID) {
fadj = (int) clockadj_get_freq(c->clkid);
- /* Due to a bug in older kernels, the reading may silently fail
- and return 0. Set the frequency back to make sure fadj is
- the actual frequency of the clock. */
- clockadj_set_freq(c->clkid, fadj);
-
/* Disable write phase mode if not implemented by driver */
if (c->write_phase_mode && !phc_has_writephase(c->clkid)) {
pr_err("clock does not support write phase mode");
@@ -1746,7 +1741,6 @@ int clock_switch_phc(struct clock *c, int phc_index)
return -1;
}
fadj = (int) clockadj_get_freq(clkid);
- clockadj_set_freq(clkid, fadj);
servo = servo_create(c->config, c->servo_type, -fadj, max_adj, 0);
if (!servo) {
pr_err("Switching PHC, failed to create clock servo");
diff --git a/phc2sys.c b/phc2sys.c
index 5f6fbaa..af948eb 100644
--- a/phc2sys.c
+++ b/phc2sys.c
@@ -147,9 +147,6 @@ static struct servo *servo_add(struct phc2sys_private *priv,
clockadj_init(clock->clkid);
ppb = clockadj_get_freq(clock->clkid);
- /* The reading may silently fail and return 0, reset the frequency to
- make sure ppb is the actual frequency of the clock. */
- clockadj_set_freq(clock->clkid, ppb);
if (clock->clkid == CLOCK_REALTIME) {
sysclk_set_leap(0);
max_ppb = sysclk_max_freq();
diff --git a/ts2phc_slave.c b/ts2phc_slave.c
index 749efe5..5541d91 100644
--- a/ts2phc_slave.c
+++ b/ts2phc_slave.c
@@ -183,10 +183,6 @@ static struct ts2phc_slave *ts2phc_slave_create(struct config *cfg, const char *
pr_debug("PHC slave %s has ptp index %d", device, junk);
fadj = (int) clockadj_get_freq(slave->clk);
- /* Due to a bug in older kernels, the reading may silently fail
- and return 0. Set the frequency back to make sure fadj is
- the actual frequency of the clock. */
- clockadj_set_freq(slave->clk, fadj);
max_adj = phc_max_adj(slave->clk);
commit 80d6875b0072c7ea636a5631e590fa72d169c351
Author: Miroslav Lichvar <mlichvar@redhat.com>
Date: Mon Oct 24 15:25:31 2022 +0200
Don't accept errors in clockadj_get_freq().
Exit if an error is returned from the clock_adjtime() call in
clockadj_get_freq(). No recoverable errors are expected.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
diff --git a/clockadj.c b/clockadj.c
index 957dc57..4c920b9 100644
--- a/clockadj.c
+++ b/clockadj.c
@@ -19,6 +19,7 @@
#include <errno.h>
#include <math.h>
+#include <stdlib.h>
#include <string.h>
#include <unistd.h>
@@ -72,6 +73,7 @@ double clockadj_get_freq(clockid_t clkid)
memset(&tx, 0, sizeof(tx));
if (clock_adjtime(clkid, &tx) < 0) {
pr_err("failed to read out the clock frequency adjustment: %m");
+ exit(1);
} else {
f = tx.freq / 65.536;
if (clkid == CLOCK_REALTIME && realtime_nominal_tick && tx.tick)
commit 211cbfe810ef4082d1af8e9b50f65d4f6fd7246a
Author: Miroslav Lichvar <mlichvar@redhat.com>
Date: Mon Oct 24 15:25:32 2022 +0200
Extend clockcheck to check for changes in frequency.
Before setting the new frequency offset on a clock update, compare the
current frequency returned by the kernel with the value saved from the
previous update. Print a warning message if the difference is larger
than 1 ppb, allowing for rounding errors in conversion to and from
double. The kernel caches the value set by clock_adjtime() in shifted
ppm, it doesn't request it from the driver, which can have a lower
resulution.
This should detect misconfigurations where multiple processes are trying
to control the clock (e.g. another ptp4l/phc2sys instance or an NTP
client), even when they don't step the clock.
Signed-off-by: Miroslav Lichvar <mlichvar@redhat.com>
diff --git a/clock.c b/clock.c
index 9079428..eea7983 100644
--- a/clock.c
+++ b/clock.c
@@ -1760,6 +1760,9 @@ int clock_switch_phc(struct clock *c, int phc_index)
static void clock_synchronize_locked(struct clock *c, double adj)
{
+ if (c->sanity_check) {
+ clockcheck_freq(c->sanity_check, clockadj_get_freq(c->clkid));
+ }
clockadj_set_freq(c->clkid, -adj);
if (c->clkid == CLOCK_REALTIME) {
sysclk_set_sync();
diff --git a/clockcheck.c b/clockcheck.c
index f0141be..b5a69cc 100644
--- a/clockcheck.c
+++ b/clockcheck.c
@@ -123,6 +123,16 @@ void clockcheck_set_freq(struct clockcheck *cc, int freq)
cc->freq_known = 1;
}
+int clockcheck_freq(struct clockcheck *cc, int freq)
+{
+ /* Allow difference of 1 ppb due to conversion to/from double */
+ if (cc->freq_known && abs(cc->current_freq - freq) > 1) {
+ pr_warning("clockcheck: clock frequency changed unexpectedly!");
+ return 1;
+ }
+ return 0;
+}
+
void clockcheck_step(struct clockcheck *cc, int64_t step)
{
if (cc->last_ts)
diff --git a/clockcheck.h b/clockcheck.h
index 1ff86eb..4b09b98 100644
--- a/clockcheck.h
+++ b/clockcheck.h
@@ -54,6 +54,14 @@ int clockcheck_sample(struct clockcheck *cc, uint64_t ts);
*/
void clockcheck_set_freq(struct clockcheck *cc, int freq);
+/**
+ * Check whether the frequency correction did not change unexpectedly.
+ * @param cc Pointer to a clock check obtained via @ref clockcheck_create().
+ * @param freq Current reading of the frequency correction in ppb.
+ * @return Zero if the frequency did not change, non-zero otherwise.
+ */
+int clockcheck_freq(struct clockcheck *cc, int freq);
+
/**
* Inform clock check that the clock was stepped.
* @param cc Pointer to a clock check obtained via @ref clockcheck_create().
diff --git a/phc2sys.c b/phc2sys.c
index af948eb..4daa1bf 100644
--- a/phc2sys.c
+++ b/phc2sys.c
@@ -566,6 +566,9 @@ static void update_clock(struct phc2sys_private *priv, struct clock *clock,
/* Fall through. */
case SERVO_LOCKED:
case SERVO_LOCKED_STABLE:
+ if (clock->sanity_check)
+ clockcheck_freq(clock->sanity_check,
+ clockadj_get_freq(clock->clkid));
clockadj_set_freq(clock->clkid, -ppb);
if (clock->clkid == CLOCK_REALTIME)
sysclk_set_sync();
diff --git a/ptp4l.8 b/ptp4l.8
index 4917240..f760e2b 100644
--- a/ptp4l.8
+++ b/ptp4l.8
@@ -605,8 +605,10 @@ This option used to be called
The maximum allowed frequency offset between uncorrected clock and the system
monotonic clock in parts per billion (ppb). This is used as a sanity check of
the synchronized clock. When a larger offset is measured, a warning message
-will be printed and the servo will be reset. When set to 0, the sanity check is
-disabled. The default is 200000000 (20%).
+will be printed and the servo will be reset. If the frequency correction set by
+ptp4l changes unexpectedly between updates of the clock (e.g. due to another
+process trying to control the clock), a warning message will be printed. When
+set to 0, the sanity check is disabled. The default is 200000000 (20%).
.TP
.B initial_delay
The initial path delay of the clock in nanoseconds used for synchronization of

View File

@ -49,6 +49,8 @@ Patch13: linuxptp-phcerr.patch
Patch14: linuxptp-vlanbond.patch
# handle EINTR when waiting for transmit timestamp
Patch15: linuxptp-eintr.patch
# check for unexpected changes in frequency offset
Patch16: linuxptp-freqcheck.patch
BuildRequires: gcc gcc-c++ make systemd
@ -77,6 +79,7 @@ Supporting legacy APIs and other platforms is not a goal.
%patch13 -p1 -b .phcerr
%patch14 -p1 -b .vlanbond
%patch15 -p1 -b .eintr
%patch16 -p1 -b .freqcheck
mv linuxptp-testsuite-%{testsuite_ver}* testsuite
mv clknetsim-%{clknetsim_ver}* testsuite/clknetsim