Skip to contentSkip to conversation

OMAP: timekeeping: time should not stop during suspend

PATCH Superseded Kevin Hilman 2 participants 4 messages Sep 24, 2009 – Oct 22, 2009

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 OMAP: timekeeping: time should not stop during suspend

During suspend, the kernel timekeeping subsystem is shut down. Before suspend and upon resume, it uses a weak function read_persistent_clock() to determine the amount of time that elapsed during suspend.

This function was not implemented on OMAP, so from the timekeeping subsystem perspective (and thus userspace as well) it appeared that no time elapsed during suspend.

This patch uses the 32k sync timer as a the persistent clock the 32k sync timer value converted to seconds.

NOTE: This does *NOT* handle wrapping of the 32k sync timer, so
wrapping of the 32k sync timer during suspend may cause problems. Also, there are not interrupts when the 32k sync timer wraps, so something else has to be done.

---
Tested on OMAP3 using PM branch.
If no issues, I will queue for 2.6.32-rc fixes

 arch/arm/plat-omap/common.c |   15 +++++++++++++++
 1 files changed, 15 insertions(+), 0 deletions(-)
arch/arm/plat-omap/common.c +15 -0
index b3f70e6..3e4325b 100644
@@ -178,6 +178,21 @@ unsigned long long sched_clock(void)
 	return ret;
 }
 
+/**
+ * read_persistent_clock -  Return time in seconds from the persistent clock.
+ */
+unsigned long read_persistent_clock(void)
+{
+	unsigned long long ret;
+	cycle_t cycles;
+
+	cycles = clocksource_32k.read(&clocksource_32k);
+	ret = (cycles * clocksource_32k.mult_orig) >> clocksource_32k.shift;
+	do_div(ret, NSEC_PER_SEC);
+
+	return (unsigned long)ret;
+}
+
 static int __init omap_init_clocksource_32k(void)
 {
 	static char err[] __initdata = KERN_ERR
Signature · 1 line
1.6.4.3
#1 lore raw

Kevin,

Quoted 9 lines · On Friday 25 September 2009 01:35:48 Kevin Hilman wrote:

During suspend, the kernel timekeeping subsystem is shut down. Before suspend and upon resume, it uses a weak function read_persistent_clock() to determine the amount of time that elapsed during suspend.

This function was not implemented on OMAP, so from the timekeeping subsystem perspective (and thus userspace as well) it appeared that no time elapsed during suspend.

Quoted 7 lines

This patch uses the 32k sync timer as a the persistent clock the 32k sync timer value converted to seconds.

NOTE: This does *NOT* handle wrapping of the 32k sync timer, so
wrapping of the 32k sync timer during suspend may cause problems. Also, there are not interrupts when the 32k sync timer wraps, so something else has to be done.

I think we should read the 32k timer value before entering the sleep and reading it again at resume time. Doing so up to one overflow could be handled, which means a maximum sleep period of 36.4 hours.

Now the time wrap problem imposes the system to wake-up before 36.4 hours. We have a few options:
1) always use the wake-up timer with a wake-up time < 36.4 hours
2) rely on the user space to set the wake-up timer correctly, depending on the applications needs.

I am for option 1.

Any thoughts?

Regards,
Jean

Quoted 36 lines

---
Tested on OMAP3 using PM branch.
If no issues, I will queue for 2.6.32-rc fixes

 arch/arm/plat-omap/common.c |   15 +++++++++++++++
 1 files changed, 15 insertions(+), 0 deletions(-)
arch/arm/plat-omap/common.c +2 -0
index b3f70e6..3e4325b 100644
@@ -178,6 +178,21 @@ unsigned long long sched_clock(void)
 	return ret;
 }
 
+/**
+ * read_persistent_clock -  Return time in seconds from the persistent

clock. + */

arch/arm/plat-omap/common.c +12 -0
+unsigned long read_persistent_clock(void)
+{
+	unsigned long long ret;
+	cycle_t cycles;
+
+	cycles = clocksource_32k.read(&clocksource_32k);
+	ret = (cycles * clocksource_32k.mult_orig) >> clocksource_32k.shift;
+	do_div(ret, NSEC_PER_SEC);
+
+	return (unsigned long)ret;
+}
+
 static int __init omap_init_clocksource_32k(void)
 {
 	static char err[] __initdata = KERN_ERR
#2 lore raw
Quoted 29 lines · Jean Pihet

Kevin,

Quoted 9 lines · On Friday 25 September 2009 01:35:48 Kevin Hilman wrote:

During suspend, the kernel timekeeping subsystem is shut down. Before suspend and upon resume, it uses a weak function read_persistent_clock() to determine the amount of time that elapsed during suspend.

This function was not implemented on OMAP, so from the timekeeping subsystem perspective (and thus userspace as well) it appeared that no time elapsed during suspend.

Quoted 7 lines

This patch uses the 32k sync timer as a the persistent clock the 32k sync timer value converted to seconds.

NOTE: This does *NOT* handle wrapping of the 32k sync timer, so
wrapping of the 32k sync timer during suspend may cause problems. Also, there are not interrupts when the 32k sync timer wraps, so something else has to be done.

I think we should read the 32k timer value before entering the sleep and reading it again at resume time. Doing so up to one overflow could be handled, which means a maximum sleep period of 36.4 hours.

Now the time wrap problem imposes the system to wake-up before 36.4 hours. We have a few options:
1) always use the wake-up timer with a wake-up time < 36.4 hours
2) rely on the user space to set the wake-up timer correctly, depending on the applications needs.

Quoted 3 lines

I am for option 1.

Any thoughts?

I agree with option 1.

Kevin

Quoted 36 lines
Quoted 36 lines

---
Tested on OMAP3 using PM branch.
If no issues, I will queue for 2.6.32-rc fixes

 arch/arm/plat-omap/common.c |   15 +++++++++++++++
 1 files changed, 15 insertions(+), 0 deletions(-)
arch/arm/plat-omap/common.c +2 -0
index b3f70e6..3e4325b 100644
@@ -178,6 +178,21 @@ unsigned long long sched_clock(void)
 	return ret;
 }
 
+/**
+ * read_persistent_clock -  Return time in seconds from the persistent

clock. + */

arch/arm/plat-omap/common.c +12 -0
+unsigned long read_persistent_clock(void)
+{
+	unsigned long long ret;
+	cycle_t cycles;
+
+	cycles = clocksource_32k.read(&clocksource_32k);
+	ret = (cycles * clocksource_32k.mult_orig) >> clocksource_32k.shift;
+	do_div(ret, NSEC_PER_SEC);
+
+	return (unsigned long)ret;
+}
+
 static int __init omap_init_clocksource_32k(void)
 {
 	static char err[] __initdata = KERN_ERR
#3 lore raw
In reply to Kevin Hilman
Quoted 17 lines · Kevin Hilman

During suspend, the kernel timekeeping subsystem is shut down. Before suspend and upon resume, it uses a weak function read_persistent_clock() to determine the amount of time that elapsed during suspend.

This function was not implemented on OMAP, so from the timekeeping subsystem perspective (and thus userspace as well) it appeared that no time elapsed during suspend.

This patch uses the 32k sync timer as a the persistent clock the 32k sync timer value converted to seconds.

NOTE: This does *NOT* handle wrapping of the 32k sync timer, so
wrapping of the 32k sync timer during suspend may cause problems. Also, there are not interrupts when the 32k sync timer wraps, so something else has to be done.

Quoted 5 lines

---
Tested on OMAP3 using PM branch.
If no issues, I will queue for 2.6.32-rc fixes

oops, forgot about this one. Will queue in PM branch fixes.

Kevin

Quoted 31 lines
 arch/arm/plat-omap/common.c |   15 +++++++++++++++
 1 files changed, 15 insertions(+), 0 deletions(-)
arch/arm/plat-omap/common.c +15 -0
index b3f70e6..3e4325b 100644
@@ -178,6 +178,21 @@ unsigned long long sched_clock(void)
 	return ret;
 }
 
+/**
+ * read_persistent_clock -  Return time in seconds from the persistent clock.
+ */
+unsigned long read_persistent_clock(void)
+{
+	unsigned long long ret;
+	cycle_t cycles;
+
+	cycles = clocksource_32k.read(&clocksource_32k);
+	ret = (cycles * clocksource_32k.mult_orig) >> clocksource_32k.shift;
+	do_div(ret, NSEC_PER_SEC);
+
+	return (unsigned long)ret;
+}
+
 static int __init omap_init_clocksource_32k(void)
 {
 	static char err[] __initdata = KERN_ERR
Signature · 1 line
1.6.4.3