* [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate
@ 2026-09-17 1:52 Eva Kurchatova
2026-09-17 1:52 ` [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count Eva Kurchatova
` (2 more replies)
0 siblings, 3 replies; 6+ messages in thread
From: Eva Kurchatova @ 2026-09-17 1:52 UTC (permalink / raw)
To: Anna-Maria Behnsen, Frederic Weisbecker, Thomas Gleixner,
John Stultz, Stephen Boyd, Miroslav Lichvar, Shuah Khan
Cc: linux-kernel, Eva Kurchatova, linux-kselftest
raw_skew measures the drift between CLOCK_MONOTONIC and
CLOCK_MONOTONIC_RAW and compares it against the adjustment adjtimex()
reports, but for the latter it reads only tx.freq. The kernel takes
the tick value into the adjustment as well, worth a microsecond of the
tick length per unit, a hundred ppm each, and a time synchronisation
daemon does put the coarse part of its correction there. Where it
does, the two numbers cannot meet:
# Estimating clock drift: -51.724(est) 48.277(act) [FAILED]
On the machine measured, chronyd held freq at 48.278 ppm with tick at
9999, so the correction really applied is -51.722 ppm, which is what
the clocks show. The skip for an externally adjusted clock does not
catch this: the clock is steady, the offset is zero, and neither freq
nor tick moves during the run.
Count what tick is worth alongside freq. On the same machine:
# Estimating clock drift: -51.724(est) -51.723(act) [OK]
Signed-off-by: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
---
tools/testing/selftests/timers/raw_skew.c | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/tools/testing/selftests/timers/raw_skew.c b/tools/testing/selftests/timers/raw_skew.c
index 0c87a8fb0d7f..dfca3e2d50d2 100644
--- a/tools/testing/selftests/timers/raw_skew.c
+++ b/tools/testing/selftests/timers/raw_skew.c
@@ -91,6 +91,7 @@ int main(int argc, char **argv)
{
struct timespec mon, raw, start, end;
long long delta1, delta2, interval, eppm, ppm;
+ long tick_nom;
struct timex tx1, tx2;
setbuf(stdout, NULL);
@@ -129,6 +130,17 @@ int main(int argc, char **argv)
/* Avg the two actual freq samples adjtimex gave us */
ppm = (long long)(tx1.freq + tx2.freq) * 1000 / 2;
ppm = shift_right(ppm, 16);
+
+ /*
+ * The tick value holds the coarse part of the adjustment, a
+ * microsecond of the tick length per unit, and time sync daemons
+ * do put part of their correction there. What it is worth has to
+ * be counted in as well, or a clock disciplined through it looks
+ * off by a hundred ppm a unit against what the two clocks show.
+ */
+ tick_nom = USEC_PER_SEC / sysconf(_SC_CLK_TCK);
+ ppm += (long long)(tx1.tick + tx2.tick - 2 * tick_nom) *
+ (USEC_PER_SEC / tick_nom) * 1000 / 2;
printf(" %lld.%i(act)", ppm/1000, abs((int)(ppm%1000)));
if (llabs(eppm - ppm) > 1000) {
--
2.55.0
^ permalink raw reply [flat|nested] 6+ messages in thread* [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count
2026-09-17 1:52 [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate Eva Kurchatova
@ 2026-09-17 1:52 ` Eva Kurchatova
2026-09-24 19:29 ` John Stultz
2026-09-29 20:05 ` [tip: timers/core] selftests: timers: Measure " tip-bot2 for Eva Kurchatova
2026-09-24 19:40 ` [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate John Stultz
2026-09-29 20:05 ` [tip: timers/core] selftests: timers: Count " tip-bot2 for Eva Kurchatova
2 siblings, 2 replies; 6+ messages in thread
From: Eva Kurchatova @ 2026-09-17 1:52 UTC (permalink / raw)
To: Anna-Maria Behnsen, Frederic Weisbecker, Thomas Gleixner,
John Stultz, Stephen Boyd, Miroslav Lichvar, Shuah Khan
Cc: linux-kernel, Eva Kurchatova, linux-kselftest
Commit b00385b8d081 ("selftests/posix_timers: Use CLOCK_THREAD_CPUTIME_ID
for ITIMER_PROF measurements") gave ITIMER_PROF the clock it counts, as
measuring a CPU timer against the wall says nothing on a machine where
the task does not have a CPU to itself. The other CPU timers are still
measured against CLOCK_REALTIME and fail the same way. With four busy
loops on the two CPUs of a test machine:
not ok 2 ITIMER_VIRTUAL
not ok 5 timer_create() per CLOCK_THREAD_CPUTIME_ID
not ok 6 timer_create() per CLOCK_PROCESS_CPUTIME_ID
ITIMER_VIRTUAL counts the time the task spends in userspace, so take
CLOCK_THREAD_CPUTIME_ID for it as well; the loop it runs stays in
userspace, so what that clock adds for system time is far inside the
half a second the check allows. check_timer_create() arms its timer on
the clock it is given, so measure on that one.
The test then passes whole on that machine, 19 of 19, busy or idle.
Signed-off-by: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
---
tools/testing/selftests/timers/posix_timers.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/tools/testing/selftests/timers/posix_timers.c b/tools/testing/selftests/timers/posix_timers.c
index a92d4b957747..52b7289e1450 100644
--- a/tools/testing/selftests/timers/posix_timers.c
+++ b/tools/testing/selftests/timers/posix_timers.c
@@ -113,9 +113,10 @@ static void check_itimer(int which, const char *name)
done = 0;
- if (which == ITIMER_VIRTUAL)
+ if (which == ITIMER_VIRTUAL) {
+ clock_id = CLOCK_THREAD_CPUTIME_ID;
signal(SIGVTALRM, sig_handler);
- else if (which == ITIMER_PROF) {
+ } else if (which == ITIMER_PROF) {
clock_id = CLOCK_THREAD_CPUTIME_ID;
signal(SIGPROF, sig_handler);
}
@@ -148,7 +149,7 @@ static void check_timer_create(int which)
struct itimerspec val = {
.it_value.tv_sec = DELAY,
};
- int clock_id = CLOCK_REALTIME;
+ int clock_id = which;
timer_t id;
done = 0;
--
2.55.0
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count
2026-09-17 1:52 ` [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count Eva Kurchatova
@ 2026-09-24 19:29 ` John Stultz
2026-09-29 20:05 ` [tip: timers/core] selftests: timers: Measure " tip-bot2 for Eva Kurchatova
1 sibling, 0 replies; 6+ messages in thread
From: John Stultz @ 2026-09-24 19:29 UTC (permalink / raw)
To: Eva Kurchatova
Cc: Anna-Maria Behnsen, Frederic Weisbecker, Thomas Gleixner,
Stephen Boyd, Miroslav Lichvar, Shuah Khan, linux-kernel,
linux-kselftest
On Wed, Sep 16, 2026 at 6:52 PM Eva Kurchatova
<eva.kurchatova@virtuozzo.com> wrote:
>
> Commit b00385b8d081 ("selftests/posix_timers: Use CLOCK_THREAD_CPUTIME_ID
> for ITIMER_PROF measurements") gave ITIMER_PROF the clock it counts, as
> measuring a CPU timer against the wall says nothing on a machine where
> the task does not have a CPU to itself. The other CPU timers are still
> measured against CLOCK_REALTIME and fail the same way. With four busy
> loops on the two CPUs of a test machine:
>
> not ok 2 ITIMER_VIRTUAL
> not ok 5 timer_create() per CLOCK_THREAD_CPUTIME_ID
> not ok 6 timer_create() per CLOCK_PROCESS_CPUTIME_ID
>
> ITIMER_VIRTUAL counts the time the task spends in userspace, so take
> CLOCK_THREAD_CPUTIME_ID for it as well; the loop it runs stays in
> userspace, so what that clock adds for system time is far inside the
> half a second the check allows. check_timer_create() arms its timer on
> the clock it is given, so measure on that one.
>
> The test then passes whole on that machine, 19 of 19, busy or idle.
>
> Signed-off-by: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
Looks reasonable to me.
Acked-by: John Stultz <jstultz@google.com>
thanks
-john
^ permalink raw reply [flat|nested] 6+ messages in thread* [tip: timers/core] selftests: timers: Measure the CPU timers on the clock they count
2026-09-17 1:52 ` [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count Eva Kurchatova
2026-09-24 19:29 ` John Stultz
@ 2026-09-29 20:05 ` tip-bot2 for Eva Kurchatova
1 sibling, 0 replies; 6+ messages in thread
From: tip-bot2 for Eva Kurchatova @ 2026-09-29 20:05 UTC (permalink / raw)
To: linux-tip-commits
Cc: Eva Kurchatova, Thomas Gleixner, John Stultz, x86, linux-kernel
The following commit has been merged into the timers/core branch of tip:
Commit-ID: b03638013add0609d35aaa8291548dea92cb9484
Gitweb: https://git.kernel.org/tip/b03638013add0609d35aaa8291548dea92cb9484
Author: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
AuthorDate: Thu, 17 Sep 2026 04:52:32 +03:00
Committer: Thomas Gleixner <tglx@kernel.org>
CommitterDate: Tue, 29 Sep 2026 22:03:31 +02:00
selftests: timers: Measure the CPU timers on the clock they count
Commit b00385b8d081 ("selftests/posix_timers: Use CLOCK_THREAD_CPUTIME_ID
for ITIMER_PROF measurements") gave ITIMER_PROF the clock it counts, as
measuring a CPU timer against the wall says nothing on a machine where
the task does not have a CPU to itself. The other CPU timers are still
measured against CLOCK_REALTIME and fail the same way. With four busy
loops on the two CPUs of a test machine:
not ok 2 ITIMER_VIRTUAL
not ok 5 timer_create() per CLOCK_THREAD_CPUTIME_ID
not ok 6 timer_create() per CLOCK_PROCESS_CPUTIME_ID
ITIMER_VIRTUAL counts the time the task spends in userspace, so take
CLOCK_THREAD_CPUTIME_ID for it as well; the loop it runs stays in
userspace, so what that clock adds for system time is far inside the
half a second the check allows. check_timer_create() arms its timer on
the clock it is given, so measure on that one.
The test then passes whole on that machine, 19 of 19, busy or idle.
Signed-off-by: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Acked-by: John Stultz <jstultz@google.com>
Link: https://patch.msgid.link/20260917015237.2462954-2-eva.kurchatova@virtuozzo.com
---
tools/testing/selftests/timers/posix_timers.c | 7 ++++---
1 file changed, 4 insertions(+), 3 deletions(-)
diff --git a/tools/testing/selftests/timers/posix_timers.c b/tools/testing/selftests/timers/posix_timers.c
index a92d4b9..52b7289 100644
--- a/tools/testing/selftests/timers/posix_timers.c
+++ b/tools/testing/selftests/timers/posix_timers.c
@@ -113,9 +113,10 @@ static void check_itimer(int which, const char *name)
done = 0;
- if (which == ITIMER_VIRTUAL)
+ if (which == ITIMER_VIRTUAL) {
+ clock_id = CLOCK_THREAD_CPUTIME_ID;
signal(SIGVTALRM, sig_handler);
- else if (which == ITIMER_PROF) {
+ } else if (which == ITIMER_PROF) {
clock_id = CLOCK_THREAD_CPUTIME_ID;
signal(SIGPROF, sig_handler);
}
@@ -148,7 +149,7 @@ static void check_timer_create(int which)
struct itimerspec val = {
.it_value.tv_sec = DELAY,
};
- int clock_id = CLOCK_REALTIME;
+ int clock_id = which;
timer_t id;
done = 0;
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate
2026-09-17 1:52 [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate Eva Kurchatova
2026-09-17 1:52 ` [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count Eva Kurchatova
@ 2026-09-24 19:40 ` John Stultz
2026-09-29 20:05 ` [tip: timers/core] selftests: timers: Count " tip-bot2 for Eva Kurchatova
2 siblings, 0 replies; 6+ messages in thread
From: John Stultz @ 2026-09-24 19:40 UTC (permalink / raw)
To: Eva Kurchatova
Cc: Anna-Maria Behnsen, Frederic Weisbecker, Thomas Gleixner,
Stephen Boyd, Miroslav Lichvar, Shuah Khan, linux-kernel,
linux-kselftest
On Wed, Sep 16, 2026 at 6:52 PM Eva Kurchatova
<eva.kurchatova@virtuozzo.com> wrote:
>
> raw_skew measures the drift between CLOCK_MONOTONIC and
> CLOCK_MONOTONIC_RAW and compares it against the adjustment adjtimex()
> reports, but for the latter it reads only tx.freq. The kernel takes
> the tick value into the adjustment as well, worth a microsecond of the
> tick length per unit, a hundred ppm each, and a time synchronisation
> daemon does put the coarse part of its correction there. Where it
> does, the two numbers cannot meet:
>
> # Estimating clock drift: -51.724(est) 48.277(act) [FAILED]
>
> On the machine measured, chronyd held freq at 48.278 ppm with tick at
> 9999, so the correction really applied is -51.722 ppm, which is what
> the clocks show. The skip for an externally adjusted clock does not
> catch this: the clock is steady, the offset is zero, and neither freq
> nor tick moves during the run.
>
> Count what tick is worth alongside freq. On the same machine:
>
> # Estimating clock drift: -51.724(est) -51.723(act) [OK]
>
> Signed-off-by: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
I've not closely reviewed the math, but it looks roughly ok to me.
I was going to object that I thought we nulled out the other adjtime
values before testing, but I must be recalling a different test
(adjtick.c I think).
Acked-by: John Stultz <jstultz@google.com>
Thanks for fixing this!
-john
^ permalink raw reply [flat|nested] 6+ messages in thread
* [tip: timers/core] selftests: timers: Count what tick is worth in the drift estimate
2026-09-17 1:52 [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate Eva Kurchatova
2026-09-17 1:52 ` [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count Eva Kurchatova
2026-09-24 19:40 ` [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate John Stultz
@ 2026-09-29 20:05 ` tip-bot2 for Eva Kurchatova
2 siblings, 0 replies; 6+ messages in thread
From: tip-bot2 for Eva Kurchatova @ 2026-09-29 20:05 UTC (permalink / raw)
To: linux-tip-commits
Cc: Eva Kurchatova, Thomas Gleixner, John Stultz, x86, linux-kernel
The following commit has been merged into the timers/core branch of tip:
Commit-ID: a252cb93e1568d9f0754b9b2ac4829d896daa14e
Gitweb: https://git.kernel.org/tip/a252cb93e1568d9f0754b9b2ac4829d896daa14e
Author: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
AuthorDate: Thu, 17 Sep 2026 04:52:31 +03:00
Committer: Thomas Gleixner <tglx@kernel.org>
CommitterDate: Tue, 29 Sep 2026 22:03:31 +02:00
selftests: timers: Count what tick is worth in the drift estimate
raw_skew measures the drift between CLOCK_MONOTONIC and
CLOCK_MONOTONIC_RAW and compares it against the adjustment adjtimex()
reports, but for the latter it reads only tx.freq. The kernel takes
the tick value into the adjustment as well, worth a microsecond of the
tick length per unit, a hundred ppm each, and a time synchronisation
daemon does put the coarse part of its correction there. Where it
does, the two numbers cannot meet:
# Estimating clock drift: -51.724(est) 48.277(act) [FAILED]
On the machine measured, chronyd held freq at 48.278 ppm with tick at
9999, so the correction really applied is -51.722 ppm, which is what
the clocks show. The skip for an externally adjusted clock does not
catch this: the clock is steady, the offset is zero, and neither freq
nor tick moves during the run.
Count what tick is worth alongside freq. On the same machine:
# Estimating clock drift: -51.724(est) -51.723(act) [OK]
Signed-off-by: Eva Kurchatova <eva.kurchatova@virtuozzo.com>
Signed-off-by: Thomas Gleixner <tglx@kernel.org>
Acked-by: John Stultz <jstultz@google.com>
Link: https://patch.msgid.link/20260917015237.2462954-1-eva.kurchatova@virtuozzo.com
---
tools/testing/selftests/timers/raw_skew.c | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/tools/testing/selftests/timers/raw_skew.c b/tools/testing/selftests/timers/raw_skew.c
index 0c87a8f..dfca3e2 100644
--- a/tools/testing/selftests/timers/raw_skew.c
+++ b/tools/testing/selftests/timers/raw_skew.c
@@ -91,6 +91,7 @@ int main(int argc, char **argv)
{
struct timespec mon, raw, start, end;
long long delta1, delta2, interval, eppm, ppm;
+ long tick_nom;
struct timex tx1, tx2;
setbuf(stdout, NULL);
@@ -129,6 +130,17 @@ int main(int argc, char **argv)
/* Avg the two actual freq samples adjtimex gave us */
ppm = (long long)(tx1.freq + tx2.freq) * 1000 / 2;
ppm = shift_right(ppm, 16);
+
+ /*
+ * The tick value holds the coarse part of the adjustment, a
+ * microsecond of the tick length per unit, and time sync daemons
+ * do put part of their correction there. What it is worth has to
+ * be counted in as well, or a clock disciplined through it looks
+ * off by a hundred ppm a unit against what the two clocks show.
+ */
+ tick_nom = USEC_PER_SEC / sysconf(_SC_CLK_TCK);
+ ppm += (long long)(tx1.tick + tx2.tick - 2 * tick_nom) *
+ (USEC_PER_SEC / tick_nom) * 1000 / 2;
printf(" %lld.%i(act)", ppm/1000, abs((int)(ppm%1000)));
if (llabs(eppm - ppm) > 1000) {
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2026-09-29 20:05 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-09-17 1:52 [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate Eva Kurchatova
2026-09-17 1:52 ` [PATCH 2/2] selftests: timers: measure the CPU timers on the clock they count Eva Kurchatova
2026-09-24 19:29 ` John Stultz
2026-09-29 20:05 ` [tip: timers/core] selftests: timers: Measure " tip-bot2 for Eva Kurchatova
2026-09-24 19:40 ` [PATCH 1/2] selftests: timers: count what tick is worth in the drift estimate John Stultz
2026-09-29 20:05 ` [tip: timers/core] selftests: timers: Count " tip-bot2 for Eva Kurchatova
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®