Skip to contentSkip to conversation

timekeeping: Add tracepoints for xtime changes - v2

PATCH In review David Ahern 2 participants 5 messages Mar 19, 2013 – Apr 1, 2013 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 timekeeping: Add tracepoints for xtime changes - v2

One component of tracking perf_clock timestamps to time-of-day is updates to xtime by a user or ntpd. To that end this patch adds tracepoints to the timekeeping code as suggested by Thomas:
https://lkml.org/lkml/2011/3/2/186

v2: use timekeeping_get_ns to retrieve nsec component

Cc Thomas Gleixner <[email protected]>
Cc John Stultz <[email protected]>
Cc Steven Rostedt <[email protected]>
---
 include/trace/events/timekeeping.h |   51 ++++++++++++++++++++++++++++++++++++
 kernel/time/timekeeping.c          |   11 ++++++++
 2 files changed, 62 insertions(+)
 create mode 100644 include/trace/events/timekeeping.h
include/trace/events/timekeeping.h new file +51 -0
new file mode 100644
index 0000000..ff7a631
@@ -0,0 +1,51 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM timekeeping
+
+#if !defined(_TRACE_TIMEKEEP_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_TIMEKEEP_H
+
+#include <linux/tracepoint.h>
+#include <linux/time.h>
+
+DECLARE_EVENT_CLASS(xtime_template,
+
+	TP_PROTO(u64 sec, u64 nsec),
+
+	TP_ARGS(sec, nsec),
+
+	TP_STRUCT__entry(
+		__field( u64,	sec)
+		__field( u64,	nsec)
+	),
+
+	TP_fast_assign(
+		__entry->sec  = sec;
+		__entry->nsec = nsec;
+	),
+
+	TP_printk("sec=%Lu nsec=%Lu", __entry->sec, __entry->nsec)
+);
+
+DEFINE_EVENT(xtime_template, do_settimeofday,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, timekeeping_inject_offset,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, timekeeping_inject_sleeptime,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, timekeeping_resume,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, change_clocksource,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+#endif /*  _TRACE_TIMEKEEP_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>
kernel/time/timekeeping.c +11 -0
index 9a0bc98..0ce45ec 100644
@@ -23,6 +23,8 @@
 #include <linux/stop_machine.h>
 #include <linux/pvclock_gtod.h>
 
+#define CREATE_TRACE_POINTS
+#include <trace/events/timekeeping.h>
 
 static struct timekeeper timekeeper;
 
@@ -462,6 +464,8 @@ int do_settimeofday(const struct timespec *tv)
 
 	timekeeping_update(tk, true);
 
+	trace_do_settimeofday(tk->xtime_sec, timekeeping_get_ns(tk));
+
 	write_sequnlock_irqrestore(&tk->lock, flags);
 
 	/* signal hrtimers about time change */
@@ -504,6 +508,8 @@ int timekeeping_inject_offset(struct timespec *ts)
 error: /* even if we error out, we forwarded the time, so call update */
 	timekeeping_update(tk, true);
 
+	trace_timekeeping_inject_offset(tk->xtime_sec, timekeeping_get_ns(tk));
+
 	write_sequnlock_irqrestore(&tk->lock, flags);
 
 	/* signal hrtimers about time change */
@@ -536,6 +542,7 @@ static int change_clocksource(void *data)
 			old->disable(old);
 	}
 	timekeeping_update(tk, true);
+	trace_change_clocksource(tk->xtime_sec, timekeeping_get_ns(tk));
 
 	write_sequnlock_irqrestore(&tk->lock, flags);
 
@@ -772,6 +779,9 @@ void timekeeping_inject_sleeptime(struct timespec *delta)
 
 	timekeeping_update(tk, true);
 
+	trace_timekeeping_inject_sleeptime(tk->xtime_sec,
+					   timekeeping_get_ns(tk));
+
 	write_sequnlock_irqrestore(&tk->lock, flags);
 
 	/* signal hrtimers about time change */
@@ -807,6 +817,7 @@ static void timekeeping_resume(void)
 	tk->ntp_error = 0;
 	timekeeping_suspended = 0;
 	timekeeping_update(tk, false);
+	trace_timekeeping_resume(tk->xtime_sec, timekeeping_get_ns(tk));
 	write_sequnlock_irqrestore(&tk->lock, flags);
 
 	touch_softlockup_watchdog();
Signature · 1 line
1.7.10.1
#1 lore raw

ping... any comments or opinions on the tracepoints?

Quoted 132 lines · David Ahern

One component of tracking perf_clock timestamps to time-of-day is updates to xtime by a user or ntpd. To that end this patch adds tracepoints to the timekeeping code as suggested by Thomas:
https://lkml.org/lkml/2011/3/2/186

v2: use timekeeping_get_ns to retrieve nsec component

Cc Thomas Gleixner <[email protected]>
Cc John Stultz <[email protected]>
Cc Steven Rostedt <[email protected]>
---
  include/trace/events/timekeeping.h |   51 ++++++++++++++++++++++++++++++++++++
  kernel/time/timekeeping.c          |   11 ++++++++
  2 files changed, 62 insertions(+)

create mode 100644 include/trace/events/timekeeping.h

include/trace/events/timekeeping.h new file +51 -0
new file mode 100644
index 0000000..ff7a631
@@ -0,0 +1,51 @@
+#undef TRACE_SYSTEM
+#define TRACE_SYSTEM timekeeping
+
+#if !defined(_TRACE_TIMEKEEP_H) || defined(TRACE_HEADER_MULTI_READ)
+#define _TRACE_TIMEKEEP_H
+
+#include <linux/tracepoint.h>
+#include <linux/time.h>
+
+DECLARE_EVENT_CLASS(xtime_template,
+
+	TP_PROTO(u64 sec, u64 nsec),
+
+	TP_ARGS(sec, nsec),
+
+	TP_STRUCT__entry(
+		__field( u64,	sec)
+		__field( u64,	nsec)
+	),
+
+	TP_fast_assign(
+		__entry->sec  = sec;
+		__entry->nsec = nsec;
+	),
+
+	TP_printk("sec=%Lu nsec=%Lu", __entry->sec, __entry->nsec)
+);
+
+DEFINE_EVENT(xtime_template, do_settimeofday,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, timekeeping_inject_offset,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, timekeeping_inject_sleeptime,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, timekeeping_resume,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+
+DEFINE_EVENT(xtime_template, change_clocksource,
+	TP_PROTO(u64 sec, u64 nsec),
+	TP_ARGS(sec, nsec));
+#endif /*  _TRACE_TIMEKEEP_H */
+
+/* This part must be outside protection */
+#include <trace/define_trace.h>
kernel/time/timekeeping.c +11 -0
index 9a0bc98..0ce45ec 100644
@@ -23,6 +23,8 @@
  #include <linux/stop_machine.h>
  #include <linux/pvclock_gtod.h>
 
+#define CREATE_TRACE_POINTS
+#include <trace/events/timekeeping.h>
 
  static struct timekeeper timekeeper;
 
@@ -462,6 +464,8 @@ int do_settimeofday(const struct timespec *tv)
 
  	timekeeping_update(tk, true);
 
+	trace_do_settimeofday(tk->xtime_sec, timekeeping_get_ns(tk));
+
  	write_sequnlock_irqrestore(&tk->lock, flags);
 
  	/* signal hrtimers about time change */
@@ -504,6 +508,8 @@ int timekeeping_inject_offset(struct timespec *ts)
  error: /* even if we error out, we forwarded the time, so call update */
  	timekeeping_update(tk, true);
 
+	trace_timekeeping_inject_offset(tk->xtime_sec, timekeeping_get_ns(tk));
+
  	write_sequnlock_irqrestore(&tk->lock, flags);
 
  	/* signal hrtimers about time change */
@@ -536,6 +542,7 @@ static int change_clocksource(void *data)
  			old->disable(old);
  	}
  	timekeeping_update(tk, true);
+	trace_change_clocksource(tk->xtime_sec, timekeeping_get_ns(tk));
 
  	write_sequnlock_irqrestore(&tk->lock, flags);
 
@@ -772,6 +779,9 @@ void timekeeping_inject_sleeptime(struct timespec *delta)
 
  	timekeeping_update(tk, true);
 
+	trace_timekeeping_inject_sleeptime(tk->xtime_sec,
+					   timekeeping_get_ns(tk));
+
  	write_sequnlock_irqrestore(&tk->lock, flags);
 
  	/* signal hrtimers about time change */
@@ -807,6 +817,7 @@ static void timekeeping_resume(void)
  	tk->ntp_error = 0;
  	timekeeping_suspended = 0;
  	timekeeping_update(tk, false);
+	trace_timekeeping_resume(tk->xtime_sec, timekeeping_get_ns(tk));
  	write_sequnlock_irqrestore(&tk->lock, flags);
 
  	touch_softlockup_watchdog();

A message that is not in the archive (sent off-list or not yet received): CANcMJZCJK_8r6y0MbhhgyTkk=7dMOFZO-KkeVy4MLE5PZxEZqg@mail.gmail.com

#3 lore raw
Quoted 2 lines · John Stultz

Sorry I missed this, I no longer receive emails at that address.

Noted. I'll update the address on future versions (and dropped the IBM one from this thread).

Quoted 18 lines
    diff --git a/include/trace/events/timekeeping.h
    b/include/trace/events/timekeeping.h
    new file mode 100644
    index 0000000..ff7a631
    --- /dev/null
    +++ b/include/trace/events/timekeeping.h
    @@ -0,0 +1,51 @@
    +#undef TRACE_SYSTEM
    +#define TRACE_SYSTEM timekeeping
    +
    +#if !defined(_TRACE_TIMEKEEP_H) || defined(TRACE_HEADER_MULTI_READ)
    +#define _TRACE_TIMEKEEP_H
    +
    +#include <linux/tracepoint.h>
    +#include <linux/time.h>
    +
    +DECLARE_EVENT_CLASS(xtime_template,
Quoted 2 lines

So xtime is becoming more of a historical reference in the code. Should this maybe be timespec_template? Or timeupdate_template?

What I want is to track updates (user, ntpd, other) to the timespec returned for gettimeofday (or, per the comment in that function, getnstimeofday). The event will have the new value for that timekeeper and the perf_clock timestamp as well so I can update the correlation.

I'm fine with calling it timespec_template. No religion on the names, only the feature.

--8<--

Quoted 2 lines

This all looks reasonable. Though do we need to be more explicit in what we're tracing here? ie: CLOCK_REALTIME timestamps?

The tracepoints don't care about the what and the tp names follow the convention of trace_<function_name> so you know where it is triggering.

Quoted 6 lines

I'd someday eventually like to rework the timekeeping core to be mostly ktime_t based, building time in a more logical method up from CLOCK_MONOTONIC rather then using CLOCK_REALTIME as our base and subtracting time from that. I'm just worried about what sort of constraints these tracepoints may put on a larger rework in the future.

Understood. And my comment above is not going to help -- ie., telling perf specific tracepoints which include function names. Should I consolidate this into a single trace_tod_update() that gets invoked in various places? The locations can move without affecting perf. I just want the tod update; where it happens should not matter.

David

David

#4 lore raw
Quoted 8 lines · David Ahern
Quoted 2 lines · John Stultz

This all looks reasonable. Though do we need to be more explicit in what we're tracing here? ie: CLOCK_REALTIME timestamps?

The tracepoints don't care about the what and the tp names follow the convention of trace_<function_name> so you know where it is triggering.

Quoted 1 line
Quoted 8 lines
Quoted 5 lines

I'd someday eventually like to rework the timekeeping core to be mostly ktime_t based, building time in a more logical method up from CLOCK_MONOTONIC rather then using CLOCK_REALTIME as our base and subtracting time from that. I'm just worried about what sort of constraints these tracepoints may put on a larger rework in the future.

Understood. And my comment above is not going to help -- ie., telling perf specific tracepoints which include function names. Should I consolidate this into a single trace_tod_update() that gets invoked in various places? The locations can move without affecting perf. I just want the tod update; where it happens should not matter.

I guess what I'm getting at is: What ABI are we creating here? Can these tracepoints come and go without any consequence? Or would changing them in the future cause application breakage?

I'm somewhat worried even trace_tod_update() is maybe too vague (again not that the name specifically is critical, but that the semantics we're specifying are clear). In other words, I think you're wanting a tracepoint at any time CLOCK_REALTIME is updated by anything other then the normal progression of time? Is that right?

You may want to also include the leapsecond modification in the tracing as well.

thanks
-john

#5 lore raw
Quoted 9 lines · John Stultz

I guess what I'm getting at is: What ABI are we creating here? Can these tracepoints come and go without any consequence? Or would changing them in the future cause application breakage?

I'm somewhat worried even trace_tod_update() is maybe too vague (again not that the name specifically is critical, but that the semantics we're specifying are clear). In other words, I think you're wanting a tracepoint at any time CLOCK_REALTIME is updated by anything other then the normal progression of time? Is that right?

yes. essentially everywhere timekeeping_update() is called except update_wall_time().

Quoted 3 lines

You may want to also include the leapsecond modification in the tracing as well.

I thought those were covered as well. hmm... maybe not if it triggers in update_wall_time.

David