PM / Sleep: Timer quiesce in freeze state

+++ b/arch/x86/kernel/apic/apic.c

arch/x86/kernel/apic/apic.c +1 -0

@@ -917,6 +917,14 @@ static void local_apic_timer_interrupt(void)
 	 */
 	inc_irq_stat(apic_timer_irqs);
 
+	/*

kernel/power/suspend.c +54 -4

index 4ca9a33..e58d880 100644
@@ -28,16 +28,20 @@
 #include <linux/ftrace.h>
 #include <trace/events/power.h>
 #include <linux/compiler.h>
+#include <linux/stop_machine.h>
+#include <linux/clockchips.h>
+#include <linux/hrtimer.h>
 
 #include "power.h"
+#include "../time/tick-internal.h"
+#include "../time/timekeeping_internal.h"
 
 const char *pm_labels[] = { "mem", "standby", "freeze", NULL };
 const char *pm_states[PM_SUSPEND_MAX];
 
 static const struct platform_suspend_ops *suspend_ops;
 static const struct platform_freeze_ops *freeze_ops;
-static DECLARE_WAIT_QUEUE_HEAD(suspend_freeze_wait_head);
-static bool suspend_freeze_wake;
+static int suspend_freeze_wake;
 
 void freeze_set_ops(const struct platform_freeze_ops *ops)
 {

@@ -48,22 +52,179 @@ void freeze_set_ops(const struct platform_freeze_ops *ops)
 
 static void freeze_begin(void)
 {
-	suspend_freeze_wake = false;
+	suspend_freeze_wake = -1;
 }
 
-static void freeze_enter(void)
+enum freezer_state {
+	FREEZER_NONE,
+	FREEZER_PICK_TK,
+	FREEZER_SUSPEND_CLKEVT,
+	FREEZER_SUSPEND_TK,
+	FREEZER_IDLE,
+	FREEZER_RESUME_TK,
+	FREEZER_RESUME_CLKEVT,
+	FREEZER_EXIT,
+};
+
+struct freezer_data {
+	int			thread_num;
+	atomic_t		thread_ack;
+	enum freezer_state	state;
+};
+
+static void set_state(struct freezer_data *fd, enum freezer_state state)
+{
+	/* set ack counter */
+	atomic_set(&fd->thread_ack, fd->thread_num);
+	/* guarantee the write ordering between ack counter and state */
+	smp_wmb();
+	fd->state = state;
+}
+
+static void ack_state(struct freezer_data *fd)
+{
+	if (atomic_dec_and_test(&fd->thread_ack))
+		set_state(fd, fd->state + 1);
+}
+
+static void freezer_pick_tk(int cpu)
+{
+	if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) {
+		static DEFINE_SPINLOCK(lock);
+
+		spin_lock(&lock);
+		if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
+			tick_do_timer_cpu = cpu;
+		spin_unlock(&lock);
+	}
+}
+
+static void freezer_suspend_clkevt(int cpu)
+{
+	if (tick_do_timer_cpu == cpu)

kernel/power/suspend.c +23 -1

+	timekeeping_suspend();
+
 	cpuidle_use_deepest_state(true);
 	cpuidle_resume();
-	wait_event(suspend_freeze_wait_head, suspend_freeze_wake);
+}
+
+static void freezer_idle(int cpu)
+{
+	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
+	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
+
+	stop_critical_timings();
+
+	while (suspend_freeze_wake == -1) {
+		int next_state;
+
+		/*
+		 * interrupt must be disabled before cpu enters idle
+		 */
+		local_irq_disable();
+
+		next_state = cpuidle_select(drv, dev);
+		if (next_state < 0) {
+			arch_cpu_idle();
+			continue;

kernel/power/suspend.c +6 -0

+	local_irq_disable();
+	hrtimers_resume();
+	local_irq_enable();
+}
+
+typedef void (*freezer_fn)(int);

kernel/power/suspend.c +7 -0

+static int freezer_stopper_fn(void *arg)
+{
+	struct freezer_data *fd = arg;
+	enum freezer_state state = FREEZER_NONE;
+	int cpu = smp_processor_id();
+
+	do {

Quoted 5 lines · Aubrey

+++ b/arch/x86/kernel/apic/apic.c

arch/x86/kernel/apic/apic.c +0 -0

@@ -917,6 +917,14 @@ static void local_apic_timer_interrupt(void)
 	 */
 	inc_irq_stat(apic_timer_irqs);
 

arch/x86/kernel/apic/apic.c static void local_apic_timer_interrupt(void) · line 930

Quoted 11 lines

arch/x86/kernel/apic/apic.c +8 -0

+	/*
+	 * if timekeeping is suspended, the clock event device will be
+	 * suspended as well, so we are not supposed to invoke the event
+	 * handler of clock event device.
+	 */
+	if (unlikely(timekeeping_suspended))
+		return;
+
 	evt->event_handler(evt);
 }
 

How would this even happen? Didn't we just suspend the lapic?

Quoted 90 lines

kernel/power/suspend.c +60 -4

index 4ca9a33..e58d880 100644
@@ -28,16 +28,20 @@
 #include <linux/ftrace.h>
 #include <trace/events/power.h>
 #include <linux/compiler.h>
+#include <linux/stop_machine.h>
+#include <linux/clockchips.h>
+#include <linux/hrtimer.h>
 
 #include "power.h"
+#include "../time/tick-internal.h"
+#include "../time/timekeeping_internal.h"
 
 const char *pm_labels[] = { "mem", "standby", "freeze", NULL };
 const char *pm_states[PM_SUSPEND_MAX];
 
 static const struct platform_suspend_ops *suspend_ops;
 static const struct platform_freeze_ops *freeze_ops;
-static DECLARE_WAIT_QUEUE_HEAD(suspend_freeze_wait_head);
-static bool suspend_freeze_wake;
+static int suspend_freeze_wake;
 
 void freeze_set_ops(const struct platform_freeze_ops *ops)
 {

@@ -48,22 +52,179 @@ void freeze_set_ops(const struct platform_freeze_ops *ops)
 
 static void freeze_begin(void)
 {
-	suspend_freeze_wake = false;
+	suspend_freeze_wake = -1;
 }
 
-static void freeze_enter(void)
+enum freezer_state {
+	FREEZER_NONE,
+	FREEZER_PICK_TK,
+	FREEZER_SUSPEND_CLKEVT,
+	FREEZER_SUSPEND_TK,
+	FREEZER_IDLE,
+	FREEZER_RESUME_TK,
+	FREEZER_RESUME_CLKEVT,
+	FREEZER_EXIT,
+};
+
+struct freezer_data {
+	int			thread_num;
+	atomic_t		thread_ack;
+	enum freezer_state	state;
+};
+
+static void set_state(struct freezer_data *fd, enum freezer_state state)
+{
+	/* set ack counter */
+	atomic_set(&fd->thread_ack, fd->thread_num);
+	/* guarantee the write ordering between ack counter and state */
+	smp_wmb();
+	fd->state = state;
+}
+
+static void ack_state(struct freezer_data *fd)
+{
+	if (atomic_dec_and_test(&fd->thread_ack))
+		set_state(fd, fd->state + 1);
+}
+
+static void freezer_pick_tk(int cpu)
+{
+	if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) {
+		static DEFINE_SPINLOCK(lock);
+
+		spin_lock(&lock);
+		if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
+			tick_do_timer_cpu = cpu;
+		spin_unlock(&lock);
+	}
+}
+
+static void freezer_suspend_clkevt(int cpu)
+{
+	if (tick_do_timer_cpu == cpu)
+		return;
+
+	clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
+}
+
+static void freezer_suspend_tk(int cpu)

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 114

Quoted 4 lines

kernel/power/suspend.c +3 -0

 {
+	if (tick_do_timer_cpu != cpu)
+		return;
+

I had a note here that this might be broken for clocksource drivers that have suspend/resume methods. You seem to have 'lost' that note, is that because you found it isn't a problem?

Quoted 26 lines

kernel/power/suspend.c +23 -1

+	timekeeping_suspend();
+
 	cpuidle_use_deepest_state(true);
 	cpuidle_resume();
-	wait_event(suspend_freeze_wait_head, suspend_freeze_wake);
+}
+
+static void freezer_idle(int cpu)
+{
+	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
+	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
+
+	stop_critical_timings();
+
+	while (suspend_freeze_wake == -1) {
+		int next_state;
+
+		/*
+		 * interrupt must be disabled before cpu enters idle
+		 */
+		local_irq_disable();
+
+		next_state = cpuidle_select(drv, dev);
+		if (next_state < 0) {
+			arch_cpu_idle();
+			continue;

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 149

Quoted 10 lines

kernel/power/suspend.c +10 -0

+		}
+		/*
+		 * cpuidle_enter will return with interrupt enabled
+		 */
+		cpuidle_enter(drv, dev, next_state);
+	}
+
+	if (suspend_freeze_wake == cpu)
+		kick_all_cpus_sync();
+

So I disabled IRQs here

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 160

Quoted 11 lines

kernel/power/suspend.c +9 -0

+	start_critical_timings();
+}
+
+static void freezer_resume_tk(int cpu)
+{
+	if (tick_do_timer_cpu != cpu)
+		return;
+
 	cpuidle_pause();
 	cpuidle_use_deepest_state(false);
+

Such that they would still be disabled here

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 172

Quoted 12 lines

kernel/power/suspend.c +12 -0

+	local_irq_disable();
+	timekeeping_resume();
+	local_irq_enable();
+}
+
+static void freezer_resume_clkevt(int cpu)
+{
+	if (tick_do_timer_cpu == cpu)
+		return;
+
+	touch_softlockup_watchdog();
+	clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);

And here.

Quoted 11 lines

kernel/power/suspend.c +11 -0

+	local_irq_disable();
+	hrtimers_resume();
+	local_irq_enable();
+}
+
+typedef void (*freezer_fn)(int);
+
+static freezer_fn freezer_func[FREEZER_EXIT] = {
+	NULL,
+	freezer_pick_tk,
+	freezer_suspend_clkevt,

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 188

Quoted 5 lines

kernel/power/suspend.c +5 -0

+	freezer_suspend_tk,
+	freezer_idle,
+	freezer_resume_tk,
+	freezer_resume_clkevt,
+};

Because this is a stop_machine callback, which are nominally run with IRQs disabled.

Quoted 15 lines

kernel/power/suspend.c +14 -0

+static int freezer_stopper_fn(void *arg)
+{
+	struct freezer_data *fd = arg;
+	enum freezer_state state = FREEZER_NONE;
+	int cpu = smp_processor_id();
+
+	do {
+		cpu_relax();
+		if (fd->state != state) {
+			state = fd->state;
+			if (freezer_func[state])
+				(*freezer_func[state])(cpu);
+			ack_state(fd);
+		}

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 207

Quoted 3 lines

kernel/power/suspend.c +3 -0

+	} while (fd->state != FREEZER_EXIT);
+	return 0;
+}

Now I suppose the problem is with cpu_pause() which needs IPIs to complete? Do we really need cpuidle_pause() there? cpuidle_uninstall_handlers() seems like a daft thing to call just about there.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quoted 4 lines

Now I suppose the problem is with cpu_pause() which needs IPIs to complete? Do we really need cpuidle_pause() there? cpuidle_uninstall_handlers() seems like a daft thing to call just about there.

Please check the log of 8651f97bd951d0bb1c10fa24e3fa3455193f3548. Rafael should know more this question than me.

Quoted 7 lines · Aubrey

Quoted 4 lines

Now I suppose the problem is with cpu_pause() which needs IPIs to complete? Do we really need cpuidle_pause() there? cpuidle_uninstall_handlers() seems like a daft thing to call just about there.

Please check the log of 8651f97bd951d0bb1c10fa24e3fa3455193f3548. Rafael should know more this question than me.

That changelog explains its complete bollocks to do it here. We _want_ to enter and/or remain in deep idle states.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quoted 6 lines · Peter Zijlstra

Quoted 5 lines · Aubrey

Quoted 4 lines

Now I suppose the problem is with cpu_pause() which needs IPIs to complete? Do we really need cpuidle_pause() there? cpuidle_uninstall_handlers() seems like a daft thing to call just about there.

Quoted 5 lines

Quoted 2 lines

Please check the log of 8651f97bd951d0bb1c10fa24e3fa3455193f3548. Rafael should know more this question than me.

That changelog explains its complete bollocks to do it here. We _want_ to enter and/or remain in deep idle states.

cpuidle_resume() will be called at the end of dpm_resume_noirq(). So we still are able to enter deep idle states after resume.

Quoted 7 lines · Aubrey

Quoted 6 lines · Peter Zijlstra

Quoted 5 lines · Aubrey

Quoted 4 lines

Now I suppose the problem is with cpu_pause() which needs IPIs to complete? Do we really need cpuidle_pause() there? cpuidle_uninstall_handlers() seems like a daft thing to call just about there.

Quoted 8 lines

Quoted 5 lines

Quoted 2 lines

Please check the log of 8651f97bd951d0bb1c10fa24e3fa3455193f3548. Rafael should know more this question than me.

That changelog explains its complete bollocks to do it here. We _want_ to enter and/or remain in deep idle states.

cpuidle_resume() will be called at the end of dpm_resume_noirq(). So we still are able to enter deep idle states after resume.

cpuidle_resume is absolute crap, as is cpuidle_suspend for that matter -- in this case.

The only reason we needed cpuidle_suspend is because some BIOS shat its pants when some CPUs were in higher C states while trying to do the S3 thing. We're not going to use S states or BIOS calls _at_all_, so no point in kicking CPUs out of their deep C states.

I think you can simply remove them altogether, they're nonsense.

Read that changelog you referred me to again.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quoted 17 lines · Peter Zijlstra

Quoted 7 lines · Aubrey

Quoted 6 lines · Peter Zijlstra

Quoted 5 lines · Aubrey

Quoted 4 lines

Now I suppose the problem is with cpu_pause() which needs IPIs to complete? Do we really need cpuidle_pause() there? cpuidle_uninstall_handlers() seems like a daft thing to call just about there.

Quoted 8 lines

Quoted 5 lines

Quoted 2 lines

Please check the log of 8651f97bd951d0bb1c10fa24e3fa3455193f3548. Rafael should know more this question than me.

That changelog explains its complete bollocks to do it here. We _want_ to enter and/or remain in deep idle states.

cpuidle_resume() will be called at the end of dpm_resume_noirq(). So we still are able to enter deep idle states after resume.

Quoted 7 lines

cpuidle_resume is absolute crap, as is cpuidle_suspend for that matter -- in this case.

The only reason we needed cpuidle_suspend is because some BIOS shat its pants when some CPUs were in higher C states while trying to do the S3 thing. We're not going to use S states or BIOS calls _at_all_, so no point in kicking CPUs out of their deep C states.

We already kicked CPUs out of their deep C states in dpm_suspend_noirq().

We pause cpuidle in dpm_suspend_noirq() and resume cpuidle in dpm_resume_noirq(), so currently we can't enter deep c-state during S states. That's an intention for some buggy BIOS.

However, for freeze state, there is another intention that we want always to enter the *deepest* c-state every time we enter freeze. So we need cpuidle_resume() to make sure we have deep cstate in freeze.

So back to your question in another email,

Quoted 1 line

I think you can simply remove them altogether, they're nonsense.

We need them to resume cpuidle in freeze.

Quoted 26 lines · Aubrey

Quoted 17 lines · Peter Zijlstra

Quoted 7 lines · Aubrey

Quoted 6 lines · Peter Zijlstra

Quoted 5 lines · Aubrey

Quoted 4 lines

Now I suppose the problem is with cpu_pause() which needs IPIs to complete? Do we really need cpuidle_pause() there? cpuidle_uninstall_handlers() seems like a daft thing to call just about there.

Quoted 8 lines

Quoted 5 lines

Quoted 2 lines

Please check the log of 8651f97bd951d0bb1c10fa24e3fa3455193f3548. Rafael should know more this question than me.

That changelog explains its complete bollocks to do it here. We _want_ to enter and/or remain in deep idle states.

cpuidle_resume() will be called at the end of dpm_resume_noirq(). So we still are able to enter deep idle states after resume.

Quoted 7 lines

cpuidle_resume is absolute crap, as is cpuidle_suspend for that matter -- in this case.

The only reason we needed cpuidle_suspend is because some BIOS shat its pants when some CPUs were in higher C states while trying to do the S3 thing. We're not going to use S states or BIOS calls _at_all_, so no point in kicking CPUs out of their deep C states.

Quoted 5 lines

We already kicked CPUs out of their deep C states in dpm_suspend_noirq().

We pause cpuidle in dpm_suspend_noirq() and resume cpuidle in dpm_resume_noirq(), so currently we can't enter deep c-state during S states. That's an intention for some buggy BIOS.

And work arounds for buggy crap hardware should not be in generic code. They should be in the platform drivers associated with said crap bugs.

But I think I see what you're saying, we're going through this dpm_ crap even for suspend to idle, which is wrong too.

Quoted 4 lines

However, for freeze state, there is another intention that we want always to enter the *deepest* c-state every time we enter freeze. So we need cpuidle_resume() to make sure we have deep cstate in freeze.

Quoted 5 lines

So back to your question in another email,

Quoted 1 line

I think you can simply remove them altogether, they're nonsense.

We need them to resume cpuidle in freeze.

So you can do cpuidle_resume() before we do the stop machine dance,

but ideally it'd all be ripped out from generic code and stuffed into the platform drivers where it belongs. But at the very least it should be isolated to the S3 path, I bet suspend to disk doesn't care either.

Quoted 10 lines

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 114

Quoted 5 lines

kernel/power/suspend.c +4 -0

+static void freezer_suspend_tk(int cpu)
 {
+	if (tick_do_timer_cpu != cpu)
+		return;
+

I had a note here that this might be broken for clocksource drivers that have suspend/resume methods. You seem to have 'lost' that note, is that because you found it isn't a problem?

I don't see it's a problem as long as we do not refer the clock source before it resumes. But I think my testing didn't cover that case that clock source drivers have suspend/resume methods. Can you please give more details what do you worry about?

Quoted 10 lines · Aubrey

Quoted 10 lines

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 114

Quoted 5 lines

kernel/power/suspend.c +4 -0

+static void freezer_suspend_tk(int cpu)
 {
+	if (tick_do_timer_cpu != cpu)
+		return;
+

I had a note here that this might be broken for clocksource drivers that have suspend/resume methods. You seem to have 'lost' that note, is that because you found it isn't a problem?

Quoted 4 lines

I don't see it's a problem as long as we do not refer the clock source before it resumes. But I think my testing didn't cover that case that clock source drivers have suspend/resume methods. Can you please give more details what do you worry about?

I can't seem to recall :/ Maybe I conflated clocksource and clockevents and figured we'd need to run the clocksource suspend callback on each cpu.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Both clocksource and clockevents are not per-cpu device, why do we need to run their suspend callback on *each* cpu?

Quoted 2 lines · Aubrey

Both clocksource and clockevents are not per-cpu device, why do we need to run their suspend callback on *each* cpu?

Uhm, you mean to say we don't use per-cpu timer lists and per-cpu timer hardware for clockevents then?

From OS level, currently tick device is per-cpu implementation while

Quoted 8 lines · Peter Zijlstra

Quoted 2 lines · Aubrey

Both clocksource and clockevents are not per-cpu device, why do we need to run their suspend callback on *each* cpu?

Uhm, you mean to say we don't use per-cpu timer lists and per-cpu timer hardware for clockevents then?

From OS level, currently tick device is per-cpu implementation while clocksource and clockevent devices are global device.

We already stop tick by clockevents_notify(suspend) on each cpu, that addresses per-cpu timer list.

And, we already call clocksource_suspend() and clockevents_suspend() in timekeeping_suspend() on the tick timer CPU. Yes, we didn't suspend per-cpu timer hardware on x86 because x86 does not have lapic timer suspend implementation. If we need to implement this, I think we can do the cross-CPU calls in clocksource/clockevents suspend(), but I didn't see any necessary we need to do this now.

so, I think we are okay now, :)

The patch is based on v3.17, merged with Rafael's pm+acpi-3.18-rc1 tag from linux-pm.git tree.

The patch is based on the patch PeterZ initially wrote.
---
Freeze is a general power saving state that processes are frozen, devices are suspended and CPUs are in idle state. However, when the system enters freeze state, there are a few timers keep ticking and hence consumes more power unnecessarily. The observed timer events in freeze state are:
- tick_sched_timer
- watchdog lockup detector
- realtime scheduler period timer

The system power consumption in freeze state will be reduced significantly if we quiesce these timers.

On Baytrail-T(ASUS_T100) platform, when the system is freezed to low power idle state(S0ix), quiescing these timers saves 29.8% power(94.48mw -> 66.32mw).

The patch is also tested on:
- Sandybrdige-EP system, both RTC alarm and power button are able to wake the system up from freeze state.
- HP laptop EliteBook 8460p, both RTC alarm and power button are able to wake the system up from freeze state.

Signed-off-by Aubrey Li <[email protected]>

Signed-off-by Peter Zijlstra <[email protected]>

Cc Rafael J. Wysocki <[email protected]>

Cc Len Brown <[email protected]>

Cc Alan Cox <[email protected]>

---
 arch/x86/kernel/apic/apic.c        |   8 ++
 drivers/cpuidle/cpuidle.c          |  12 +++
 kernel/power/suspend.c             | 185 +++++++++++++++++++++++++++++++++++--
 kernel/time/timekeeping.c          |   4 +-
 kernel/time/timekeeping_internal.h |   3 +
 5 files changed, 204 insertions(+), 8 deletions(-)

arch/x86/kernel/apic/apic.c +8 -0

index 6776027..f2bb645 100644
@@ -917,6 +917,14 @@ static void local_apic_timer_interrupt(void)
 	 */
 	inc_irq_stat(apic_timer_irqs);
 
+	/*
+	 * if timekeeping is suspended, the clock event device will be
+	 * suspended as well, so we are not supposed to invoke the event
+	 * handler of clock event device.
+	 */
+	if (unlikely(timekeeping_suspended))
+		return;
+
 	evt->event_handler(evt);
 }
 

drivers/cpuidle/cpuidle.c +12 -0

index ee9df5e..8f84f40 100644
@@ -119,6 +119,18 @@ int cpuidle_enter_state(struct cpuidle_device *dev, struct cpuidle_driver *drv,
 	ktime_t time_start, time_end;
 	s64 diff;
 
+	/*
+	 * under the scenario of use deepest idle state, the timekeeping
+	 * could be suspended as well as the clock source device, so we
+	 * bypass the idle counter update for this case
+	 */
+	if (unlikely(use_deepest_state)) {
+		entered_state = target_state->enter(dev, drv, index);
+		if (!cpuidle_state_is_coupled(dev, drv, entered_state))
+			local_irq_enable();
+		return entered_state;
+	}
+
 	trace_cpu_idle_rcuidle(index, dev->cpu);
 	time_start = ktime_get();
 

kernel/power/suspend.c +179 -6

index 4ca9a33..660fd15 100644
@@ -28,16 +28,20 @@
 #include <linux/ftrace.h>
 #include <trace/events/power.h>
 #include <linux/compiler.h>
+#include <linux/stop_machine.h>
+#include <linux/clockchips.h>
+#include <linux/hrtimer.h>
 
 #include "power.h"
+#include "../time/tick-internal.h"
+#include "../time/timekeeping_internal.h"
 
 const char *pm_labels[] = { "mem", "standby", "freeze", NULL };
 const char *pm_states[PM_SUSPEND_MAX];
 
 static const struct platform_suspend_ops *suspend_ops;
 static const struct platform_freeze_ops *freeze_ops;
-static DECLARE_WAIT_QUEUE_HEAD(suspend_freeze_wait_head);
-static bool suspend_freeze_wake;
+static int suspend_freeze_wake;
 
 void freeze_set_ops(const struct platform_freeze_ops *ops)
 {

@@ -48,22 +52,191 @@ void freeze_set_ops(const struct platform_freeze_ops *ops)
 
 static void freeze_begin(void)
 {
-	suspend_freeze_wake = false;
+	suspend_freeze_wake = -1;
+}
+
+enum freezer_state {
+	FREEZER_NONE,
+	FREEZER_PICK_TK,
+	FREEZER_SUSPEND_CLKEVT,
+	FREEZER_SUSPEND_TK,
+	FREEZER_IDLE,
+	FREEZER_RESUME_TK,
+	FREEZER_RESUME_CLKEVT,
+	FREEZER_EXIT,
+};
+
+struct freezer_data {
+	int			thread_num;
+	atomic_t		thread_ack;
+	enum freezer_state	state;
+};
+
+static void set_state(struct freezer_data *fd, enum freezer_state state)
+{
+	/* set ack counter */
+	atomic_set(&fd->thread_ack, fd->thread_num);
+	/* guarantee the write ordering between ack counter and state */
+	smp_wmb();
+	fd->state = state;
+}
+
+static void ack_state(struct freezer_data *fd)
+{
+	if (atomic_dec_and_test(&fd->thread_ack))
+		set_state(fd, fd->state + 1);
+}
+
+static void freezer_pick_tk(int cpu)
+{
+	if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) {
+		static DEFINE_SPINLOCK(lock);
+
+		spin_lock(&lock);
+		if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
+			tick_do_timer_cpu = cpu;
+		spin_unlock(&lock);
+	}
+}
+
+static void freezer_suspend_clkevt(int cpu)
+{
+	if (tick_do_timer_cpu == cpu)
+		return;
+
+	clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);
+}
+
+static void freezer_suspend_tk(int cpu)
+{
+	if (tick_do_timer_cpu != cpu)
+		return;
+
+	timekeeping_suspend();
+
+}
+
+static void freezer_idle(int cpu)
+{
+	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
+	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
+
+	stop_critical_timings();
+
+	while (suspend_freeze_wake == -1) {
+		int next_state;
+
+		/*
+		 * interrupt must be disabled before cpu enters idle
+		 */
+		local_irq_disable();
+
+		next_state = cpuidle_select(drv, dev);
+		if (next_state < 0) {
+			arch_cpu_idle();
+			continue;
+		}
+		/*
+		 * cpuidle_enter will return with interrupt enabled
+		 */
+		cpuidle_enter(drv, dev, next_state);
+	}
+
+	if (suspend_freeze_wake == cpu)
+		kick_all_cpus_sync();
+
+	/*
+	 * We disable interrupt here for the rest of resume operations
+	 */
+	local_irq_disable();
+	start_critical_timings();
+}
+
+static void freezer_resume_tk(int cpu)
+{
+	if (tick_do_timer_cpu != cpu)
+		return;
+
+	timekeeping_resume();
+}
+
+static void freezer_resume_clkevt(int cpu)
+{
+	if (tick_do_timer_cpu == cpu) {
+		/*
+		 * Turn on the interrupt on the tick timer CPU as freezer
+		 * tasks are finished.
+		 */
+		local_irq_enable();
+		return;
+	}
+
+	touch_softlockup_watchdog();
+	clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);
+	hrtimers_resume();
+	/*
+	 * Turn on the interrupt on the non-tick-timer CPUs as freezer
+	 * tasks are finished
+	 */
+	local_irq_enable();
+}
+
+typedef void (*freezer_fn)(int);
+
+static freezer_fn freezer_func[FREEZER_EXIT] = {
+	NULL,
+	freezer_pick_tk,
+	freezer_suspend_clkevt,
+	freezer_suspend_tk,
+	freezer_idle,
+	freezer_resume_tk,
+	freezer_resume_clkevt,
+};
+
+static int freezer_stopper_fn(void *arg)
+{
+	struct freezer_data *fd = arg;
+	enum freezer_state state = FREEZER_NONE;
+	int cpu = smp_processor_id();
+
+	do {
+		cpu_relax();
+		if (fd->state != state) {
+			state = fd->state;
+			if (freezer_func[state])
+				(*freezer_func[state])(cpu);
+			ack_state(fd);
+		}
+	} while (fd->state != FREEZER_EXIT);
+
+	return 0;
 }
 
 static void freeze_enter(void)
 {
+	struct freezer_data fd;
+
 	cpuidle_use_deepest_state(true);
 	cpuidle_resume();
-	wait_event(suspend_freeze_wait_head, suspend_freeze_wake);
+
+	get_online_cpus();
+
+	fd.thread_num = num_online_cpus();
+	set_state(&fd, FREEZER_PICK_TK);
+
+	__stop_machine(freezer_stopper_fn, &fd, cpu_online_mask);
+
+	put_online_cpus();
+
 	cpuidle_pause();
 	cpuidle_use_deepest_state(false);
 }
 
 void freeze_wake(void)
 {
-	suspend_freeze_wake = true;
-	wake_up(&suspend_freeze_wait_head);
+	if (suspend_freeze_wake != -1)
+		return;
+	suspend_freeze_wake = smp_processor_id();
 }
 EXPORT_SYMBOL_GPL(freeze_wake);
 

kernel/time/timekeeping.c +2 -2

index ec1791f..23d8feb 100644
@@ -1114,7 +1114,7 @@ void timekeeping_inject_sleeptime(struct timespec *delta)
  * xtime/wall_to_monotonic/jiffies/etc are
  * still managed by arch specific suspend/resume code.
  */
-static void timekeeping_resume(void)
+void timekeeping_resume(void)
 {
 	struct timekeeper *tk = &tk_core.timekeeper;
 	struct clocksource *clock = tk->tkr.clock;

@@ -1195,7 +1195,7 @@ static void timekeeping_resume(void)
 	hrtimers_resume();
 }
 
-static int timekeeping_suspend(void)
+int timekeeping_suspend(void)
 {
 	struct timekeeper *tk = &tk_core.timekeeper;
 	unsigned long flags;

kernel/time/timekeeping_internal.h +3 -0

index 4ea005a..ed7a574 100644
@@ -26,4 +26,7 @@ static inline cycle_t clocksource_delta(cycle_t now, cycle_t last, cycle_t mask)
 }
 #endif
 
+extern int timekeeping_suspend(void);
+extern void timekeeping_resume(void);
+
 #endif /* _TIMEKEEPING_INTERNAL_H */

Peter, Thomas, any comments here?

Freeze is a general power saving state that processes are frozen, devices are suspended and CPUs are in idle state. However, when the system enters freeze state, there are a few timers keep ticking and hence consumes more power unnecessarily. The observed timer events in freeze state are:
- tick_sched_timer
- watchdog lockup detector
- realtime scheduler period timer

The system power consumption in freeze state will be reduced significantly if we quiesce these timers.

arch/x86/kernel/apic/apic.c +0 -0

index 6776027..f2bb645 100644

kernel/power/suspend.c +0 -0

index 4ca9a33..660fd15 100644

kernel/power/suspend.c +17 -0

+static void freezer_pick_tk(int cpu)
+{
+	if (tick_do_timer_cpu == TICK_DO_TIMER_NONE) {
+		static DEFINE_SPINLOCK(lock);
+
+		spin_lock(&lock);
+		if (tick_do_timer_cpu == TICK_DO_TIMER_NONE)
+			tick_do_timer_cpu = cpu;
+		spin_unlock(&lock);
+	}
+}
+static void freezer_suspend_clkevt(int cpu)
+{
+	if (tick_do_timer_cpu == cpu)
+		return;
+
+	clockevents_notify(CLOCK_EVT_NOTIFY_SUSPEND, NULL);

Quoted 8 lines · Aubrey

Freeze is a general power saving state that processes are frozen, devices are suspended and CPUs are in idle state. However, when the system enters freeze state, there are a few timers keep ticking and hence consumes more power unnecessarily. The observed timer events in freeze state are:
- tick_sched_timer
- watchdog lockup detector
- realtime scheduler period timer

Quoted 2 lines

The system power consumption in freeze state will be reduced significantly if we quiesce these timers.

So the obvious question is why dont we quiesce these timers by telling the subsystems which manage these timers to shut them down?

I really want a proper answer for this in the first place, but let me look at the proposed "solution" as well.

But before we do that we want a proper explanation why the interrupt fires at all. The lack of explanation cleary documents that this is a 'hacked it into submission' approach.

From what I remember its the waking interrupt that ends up in the

kernel/power/suspend.c line 37

Quoted 2 lines

kernel/power/suspend.c +2 -0

+#include "../time/tick-internal.h"
+#include "../time/timekeeping_internal.h"

Eew.

So you export the world and some more from timekeeping and the tick code and fiddle with it randomly just to do:

1) Suspend clock event devices
2) Suspend timekeeping
3) Resume timekeeping
4) Resume clock event devices

stomp_machine() is in 99% of all use cases a clear indicator for a complete design failure.

So the generic idle task needs a check like this:

   if (idle_should_freeze())
      	frozen_idle();

Quoted 10 lines · Thomas Gleixner

Quoted 8 lines · Aubrey

Freeze is a general power saving state that processes are frozen, devices are suspended and CPUs are in idle state. However, when the system enters freeze state, there are a few timers keep ticking and hence consumes more power unnecessarily. The observed timer events in freeze state are:
- tick_sched_timer
- watchdog lockup detector
- realtime scheduler period timer

Quoted 8 lines

Quoted 2 lines

The system power consumption in freeze state will be reduced significantly if we quiesce these timers.

So the obvious question is why dont we quiesce these timers by telling the subsystems which manage these timers to shut them down?

I really want a proper answer for this in the first place, but let me look at the proposed "solution" as well.

Two arguments here:

1) the current suspend modes don't care, so if this suspend mode starts to care, its likely to 'break' in the future simply because people never cared about timers.

2) there could be userland timers, userland is frozen but they'll still have their timers set and those can and will fire.

But sure, we can add suspend notifiers to stuff to shut down timers; I should have a patch for at least one of the offenders somewhere. But I really think that we should not be looking at the individual timers for this, none of the other suspend modes care about active timers.

Quoted 5 lines

But before we do that we want a proper explanation why the interrupt fires at all. The lack of explanation cleary documents that this is a 'hacked it into submission' approach.

From what I remember its the waking interrupt that ends up in the

timekeeping code, Li should have a backtrace somwhere.

Quoted 5 lines

kernel/power/suspend.c line 37

Quoted 3 lines

kernel/power/suspend.c +2 -0

+#include "../time/tick-internal.h"
+#include "../time/timekeeping_internal.h"

Eew.

I knew you'd love that :-)

Quoted 7 lines

So you export the world and some more from timekeeping and the tick code and fiddle with it randomly just to do:

1) Suspend clock event devices
2) Suspend timekeeping
3) Resume timekeeping
4) Resume clock event devices

Sure, we can add some exports and clean that up, but..

Quoted 7 lines

stomp_machine() is in 99% of all use cases a clear indicator for a complete design failure.

So the generic idle task needs a check like this:

   if (idle_should_freeze())
      	frozen_idle();

So that is adding extra code to fairly common/hot paths just for this one extra special case. I tried to avoid doing that.

But I suppose we can try and merge that with the offline case and guard both special cases with a single variable or so.

Quoted 34 lines · Peter Zijlstra

Quoted 10 lines · Thomas Gleixner

Quoted 8 lines · Aubrey

Freeze is a general power saving state that processes are frozen, devices are suspended and CPUs are in idle state. However, when the system enters freeze state, there are a few timers keep ticking and hence consumes more power unnecessarily. The observed timer events in freeze state are:
- tick_sched_timer
- watchdog lockup detector
- realtime scheduler period timer

Quoted 8 lines

Quoted 2 lines

The system power consumption in freeze state will be reduced significantly if we quiesce these timers.

So the obvious question is why dont we quiesce these timers by telling the subsystems which manage these timers to shut them down?

I really want a proper answer for this in the first place, but let me look at the proposed "solution" as well.

Two arguments here:

1) the current suspend modes don't care, so if this suspend mode starts to care, its likely to 'break' in the future simply because people never cared about timers.

2) there could be userland timers, userland is frozen but they'll still have their timers set and those can and will fire.

But sure, we can add suspend notifiers to stuff to shut down timers; I should have a patch for at least one of the offenders somewhere. But I really think that we should not be looking at the individual timers for this, none of the other suspend modes care about active timers.

Quoted 6 lines

Quoted 5 lines

But before we do that we want a proper explanation why the interrupt fires at all. The lack of explanation cleary documents that this is a 'hacked it into submission' approach.

From what I remember its the waking interrupt that ends up in the

timekeeping code, Li should have a backtrace somwhere.

There are two race conditions:

The first one occurs after the interrupt is disabled and before we suspend lapic. In this time slot, if apic timer interrupt occurs, the interrupt is pending there because the interrupt is disabled. Then we suspend timekeeping, and then we enter idle and exit idle with interrupt re-enabled, the timer interrupt is handled with timekeeping is suspended.

The other occurs after timekeeping_suspended = 1 and before we suspend lapic. In this time slot, if apic timer interrupt occurs, we invoke the timer interrupt while timekeeping is suspended as above.

Quoted 51 lines · Aubrey

Quoted 34 lines · Peter Zijlstra

Quoted 10 lines · Thomas Gleixner

Quoted 8 lines · Aubrey

Freeze is a general power saving state that processes are frozen, devices are suspended and CPUs are in idle state. However, when the system enters freeze state, there are a few timers keep ticking and hence consumes more power unnecessarily. The observed timer events in freeze state are:
- tick_sched_timer
- watchdog lockup detector
- realtime scheduler period timer

Quoted 8 lines

Quoted 2 lines

The system power consumption in freeze state will be reduced significantly if we quiesce these timers.

So the obvious question is why dont we quiesce these timers by telling the subsystems which manage these timers to shut them down?

I really want a proper answer for this in the first place, but let me look at the proposed "solution" as well.

Two arguments here:

1) the current suspend modes don't care, so if this suspend mode starts to care, its likely to 'break' in the future simply because people never cared about timers.

2) there could be userland timers, userland is frozen but they'll still have their timers set and those can and will fire.

But sure, we can add suspend notifiers to stuff to shut down timers; I should have a patch for at least one of the offenders somewhere. But I really think that we should not be looking at the individual timers for this, none of the other suspend modes care about active timers.

Quoted 6 lines

Quoted 5 lines

But before we do that we want a proper explanation why the interrupt fires at all. The lack of explanation cleary documents that this is a 'hacked it into submission' approach.

From what I remember its the waking interrupt that ends up in the

timekeeping code, Li should have a backtrace somwhere.

There are two race conditions:

The first one occurs after the interrupt is disabled and before we suspend lapic. In this time slot, if apic timer interrupt occurs, the interrupt is pending there because the interrupt is disabled. Then we suspend timekeeping, and then we enter idle and exit idle with interrupt re-enabled, the timer interrupt is handled with timekeeping is suspended.

Quoted 3 lines

The other occurs after timekeeping_suspended = 1 and before we suspend lapic. In this time slot, if apic timer interrupt occurs, we invoke the timer interrupt while timekeeping is suspended as above.

And that race exists for every implementation and is not at all apic timer specific. So we fix it at the core and not at some random place in the architecture code.

On Wed, Nov 12, 2014 at 10:09:47PM +0100, Thomas Gleixner wrote:
But sure, we can add suspend notifiers to stuff to shut down timers; I should have a patch for at least one of the offenders somewhere. But I really think that we should not be looking at the individual timers for this, none of the other suspend modes care about active timers.

Quoted 5 lines

But before we do that we want a proper explanation why the interrupt fires at all. The lack of explanation cleary documents that this is a 'hacked it into submission' approach.

From what I remember its the waking interrupt that ends up in the

timekeeping code, Li should have a backtrace somwhere.

Quoted 7 lines

stomp_machine() is in 99% of all use cases a clear indicator for a complete design failure.

So the generic idle task needs a check like this:

   if (idle_should_freeze())
      	frozen_idle();

So that is adding extra code to fairly common/hot paths just for this one extra special case. I tried to avoid doing that.

Quoted 5 lines · Peter Zijlstra

On Wed, Nov 12, 2014 at 10:09:47PM +0100, Thomas Gleixner wrote:
But sure, we can add suspend notifiers to stuff to shut down timers; I should have a patch for at least one of the offenders somewhere. But I really think that we should not be looking at the individual timers for this, none of the other suspend modes care about active timers.

Fair enough.

Quoted 7 lines · Thomas Gleixner

Quoted 5 lines · Peter Zijlstra

On Wed, Nov 12, 2014 at 10:09:47PM +0100, Thomas Gleixner wrote:
But sure, we can add suspend notifiers to stuff to shut down timers; I should have a patch for at least one of the offenders somewhere. But I really think that we should not be looking at the individual timers for this, none of the other suspend modes care about active timers.

Quoted 2 lines

Fair enough.

If you are okay with the current method to suspend timekeeping entirely, then we can go further to fix the rest concerns.

Quoted 8 lines · Aubrey

Quoted 7 lines · Thomas Gleixner

Quoted 5 lines · Peter Zijlstra

On Wed, Nov 12, 2014 at 10:09:47PM +0100, Thomas Gleixner wrote:
But sure, we can add suspend notifiers to stuff to shut down timers; I should have a patch for at least one of the offenders somewhere. But I really think that we should not be looking at the individual timers for this, none of the other suspend modes care about active timers.

Quoted 5 lines

Quoted 2 lines

Fair enough.

If you are okay with the current method to suspend timekeeping entirely, then we can go further to fix the rest concerns.

I'm fine with that when it's done proper :)

Thanks,

tglx

Quoted 20 lines

kernel/power/suspend.c +20 -0

+
+static void freezer_idle(int cpu)
+{
+	struct cpuidle_device *dev = __this_cpu_read(cpuidle_devices);
+	struct cpuidle_driver *drv = cpuidle_get_cpu_driver(dev);
+
+	stop_critical_timings();
+
+	while (suspend_freeze_wake == -1) {
+		int next_state;
+
+		/*
+		 * interrupt must be disabled before cpu enters idle
+		 */
+		local_irq_disable();
+
+		next_state = cpuidle_select(drv, dev);
+		if (next_state < 0) {
+			arch_cpu_idle();
+			continue;

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 149

Quoted 10 lines

kernel/power/suspend.c +10 -0

+		}
+		/*
+		 * cpuidle_enter will return with interrupt enabled
+		 */
+		cpuidle_enter(drv, dev, next_state);
+	}
+
+	if (suspend_freeze_wake == cpu)
+		kick_all_cpus_sync();
+

So I disabled IRQs here

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 160

Quoted 11 lines

kernel/power/suspend.c +9 -0

+	start_critical_timings();
+}
+
+static void freezer_resume_tk(int cpu)
+{
+	if (tick_do_timer_cpu != cpu)
+		return;
+
 	cpuidle_pause();
 	cpuidle_use_deepest_state(false);
+

Such that they would still be disabled here

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 172

Quoted 12 lines

kernel/power/suspend.c +12 -0

+	local_irq_disable();
+	timekeeping_resume();
+	local_irq_enable();
+}
+
+static void freezer_resume_clkevt(int cpu)
+{
+	if (tick_do_timer_cpu == cpu)
+		return;
+
+	touch_softlockup_watchdog();
+	clockevents_notify(CLOCK_EVT_NOTIFY_RESUME, NULL);

And here.

Quoted 6 lines

kernel/power/suspend.c +6 -0

+	local_irq_disable();
+	hrtimers_resume();
+	local_irq_enable();
+}
+
+typedef void (*freezer_fn)(int);

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 188

Quoted 10 lines

kernel/power/suspend.c +10 -0

+
+static freezer_fn freezer_func[FREEZER_EXIT] = {
+	NULL,
+	freezer_pick_tk,
+	freezer_suspend_clkevt,
+	freezer_suspend_tk,
+	freezer_idle,
+	freezer_resume_tk,
+	freezer_resume_clkevt,
+};

Because this is a stop_machine callback, which are nominally run with IRQs disabled.

Quoted 12 lines

kernel/power/suspend.c +12 -0

+static int freezer_stopper_fn(void *arg)
+{
+	struct freezer_data *fd = arg;
+	enum freezer_state state = FREEZER_NONE;
+	int cpu = smp_processor_id();
+
+	do {
+		cpu_relax();
+		if (fd->state != state) {
+			state = fd->state;
+			if (freezer_func[state])
+				(*freezer_func[state])(cpu);

kernel/power/suspend.c void freeze_set_ops(const struct platform_freeze_ops *ops) · line 207

Quoted 5 lines

kernel/power/suspend.c +5 -0

+			ack_state(fd);
+		}
+	} while (fd->state != FREEZER_EXIT);
+	return 0;
+}

Now I suppose the problem is with cpu_pause() which needs IPIs to complete?

Quoted 2 lines

Now I suppose the problem is with cpu_pause() which needs IPIs to complete?

Yes, cpu_pause() will invoke smp IPI functions which need interrupt is enabled. So I changed irq ops like above. Actually, I have an early version to move cpuidle_pause()/cpuidle_resume() out of stop_machine(), that might be a better solution?