] drivers: leds/trigger: system cannot enter suspend

PATCH Superseded Zhang Bo 5 participants 11 messages Jun 7, 2017 – Jun 11, 2017 LED

Share & tools

Zhang Bo

author

2017-06-07 02:50 UTC #0 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

System cannot enter suspend mode because of heartbeat led trigger. In autosleep_wq, try_to_suspend function will try to enter suspend mode in specific period. it will get wakeup_count then call pm_notifier chain callback function and freeze processes. Heartbeat_pm_notifier is called and it call led_trigger_unregister to change the trigger of led device to none. It will send uevent message and the wakeup source count changed.

Add suspend/resume ops to the struct led_trigger, implement these two ops for ledtrig-heartbeat. Add led_trigger_suspend/led_trigger_resume funcitons to iterate through all LED class devices registered on given trigger and call their suspend/resume ops and disable/enable sysfs at the same time.

Call led_trigger_suspend{resuem} API in pm_notifier function to suspend/resume all led devices listed in given trigger.

Signed-off-by Zhang Bo <[email protected]>

---
 drivers/leds/led-triggers.c              | 30 +++++++++++++++++++++++++++
 drivers/leds/trigger/ledtrig-heartbeat.c | 35 ++++++++++++++++++++++++++------
 include/linux/leds.h                     |  6 ++++++
 3 files changed, 65 insertions(+), 6 deletions(-)

drivers/leds/led-triggers.c +30 -0

index 431123b..b239907 100644
@@ -333,6 +333,36 @@ void led_trigger_blink_oneshot(struct led_trigger *trig,
 }
 EXPORT_SYMBOL_GPL(led_trigger_blink_oneshot);
 
+void led_trigger_suspend(struct led_trigger *trig)
+{
+	struct led_classdev *led_cdev;
+
+	if (!trig)
+		return;
+
+	read_lock(&trig->leddev_list_lock);
+	list_for_each_entry(led_cdev, &trig->led_cdevs, trig_list)
+		if (led_cdev->trigger->suspend)
+			led_cdev->trigger->suspend(led_cdev);
+	read_unlock(&trig->leddev_list_lock);
+}
+EXPORT_SYMBOL_GPL(led_trigger_suspend);
+
+void led_trigger_resume(struct led_trigger *trig)
+{
+	struct led_classdev *led_cdev;
+
+	if (!trig)
+		return;
+
+	read_lock(&trig->leddev_list_lock);
+	list_for_each_entry(led_cdev, &trig->led_cdevs, trig_list)
+		if (led_cdev->trigger->resume)
+			led_cdev->trigger->resume(led_cdev);
+	read_unlock(&trig->leddev_list_lock);
+}
+EXPORT_SYMBOL_GPL(led_trigger_resume);
+
 void led_trigger_register_simple(const char *name, struct led_trigger **tp)
 {
 	struct led_trigger *trig;

drivers/leds/trigger/ledtrig-heartbeat.c +29 -6

index afa3b40..8f779dd 100644
@@ -142,6 +142,7 @@ static void heartbeat_trig_activate(struct led_classdev *led_cdev)
 	led_heartbeat_function(heartbeat_data->timer.data);
 	set_bit(LED_BLINK_SW, &led_cdev->work_flags);
 	led_cdev->activated = true;
+	led_cdev->suspended = false;
 }
 
 static void heartbeat_trig_deactivate(struct led_classdev *led_cdev)

@@ -154,6 +155,30 @@ static void heartbeat_trig_deactivate(struct led_classdev *led_cdev)
 		kfree(heartbeat_data);
 		clear_bit(LED_BLINK_SW, &led_cdev->work_flags);
 		led_cdev->activated = false;
+		led_cdev->suspended = false;
+	}
+}
+
+static void heartbeat_trig_resume(struct led_classdev *led_cdev)
+{
+	struct heartbeat_trig_data *heartbeat_data = led_cdev->trigger_data;
+
+	if (led_cdev->suspended) {
+		setup_timer(&heartbeat_data->timer,
+			    led_heartbeat_function, (unsigned long) led_cdev);
+		heartbeat_data->phase = 0;
+		led_heartbeat_function(heartbeat_data->timer.data);
+		led_cdev->suspended = false;
+	}
+}
+
+static void heartbeat_trig_suspend(struct led_classdev *led_cdev)
+{
+	struct heartbeat_trig_data *heartbeat_data = led_cdev->trigger_data;
+
+	if (led_cdev->activated && !led_cdev->suspended) {
+		del_timer_sync(&heartbeat_data->timer);
+		led_cdev->suspended = true;
 	}
 }

@@ -161,25 +186,23 @@ static void heartbeat_trig_deactivate(struct led_classdev *led_cdev)
 	.name     = "heartbeat",
 	.activate = heartbeat_trig_activate,
 	.deactivate = heartbeat_trig_deactivate,
+	.suspend = heartbeat_trig_suspend,
+	.resume = heartbeat_trig_resume,
 };
 
 static int heartbeat_pm_notifier(struct notifier_block *nb,
 				 unsigned long pm_event, void *unused)
 {
-	int rc;
-
 	switch (pm_event) {
 	case PM_SUSPEND_PREPARE:
 	case PM_HIBERNATION_PREPARE:
 	case PM_RESTORE_PREPARE:
-		led_trigger_unregister(&heartbeat_led_trigger);
+		led_trigger_suspend(&heartbeat_led_trigger);
 		break;
 	case PM_POST_SUSPEND:
 	case PM_POST_HIBERNATION:
 	case PM_POST_RESTORE:
-		rc = led_trigger_register(&heartbeat_led_trigger);
-		if (rc)
-			pr_err("could not re-register heartbeat trigger\n");
+		led_trigger_resume(&heartbeat_led_trigger);
 		break;
 	default:
 		break;

include/linux/leds.h +6 -0

index 64c56d4..b792950 100644
@@ -111,6 +111,7 @@ struct led_classdev {
 	void			*trigger_data;
 	/* true if activated - deactivate routine uses it to do cleanup */
 	bool			activated;
+	bool			suspended;
 #endif
 
 #ifdef CONFIG_LEDS_BRIGHTNESS_HW_CHANGED

@@ -254,6 +255,8 @@ struct led_trigger {
 	const char	 *name;
 	void		(*activate)(struct led_classdev *led_cdev);
 	void		(*deactivate)(struct led_classdev *led_cdev);
+	void		(*suspend)(struct led_classdev *led_cdev);
+	void		(*resume)(struct led_classdev *led_cdev);
 
 	/* LEDs under control by this trigger (for simple triggers) */
 	rwlock_t	  leddev_list_lock;

@@ -291,6 +294,9 @@ extern void led_trigger_set(struct led_classdev *led_cdev,
 			struct led_trigger *trigger);
 extern void led_trigger_remove(struct led_classdev *led_cdev);
 
+extern void led_trigger_suspend(struct led_trigger *trig);
+extern void led_trigger_resume(struct led_trigger *trig);
+
 static inline void *led_get_trigger_data(struct led_classdev *led_cdev)
 {
 	return led_cdev->trigger_data;

Signature · 1 line

1.9.1

Pavel Machek

NAK

2017-06-07 07:12 UTC #1 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Quoted 13 lines · Zhang Bo

System cannot enter suspend mode because of heartbeat led trigger. In autosleep_wq, try_to_suspend function will try to enter suspend mode in specific period. it will get wakeup_count then call pm_notifier chain callback function and freeze processes. Heartbeat_pm_notifier is called and it call led_trigger_unregister to change the trigger of led device to none. It will send uevent message and the wakeup source count changed.
Add suspend/resume ops to the struct led_trigger, implement these two ops for ledtrig-heartbeat. Add led_trigger_suspend/led_trigger_resume funcitons to iterate through all LED class devices registered on given trigger and call their suspend/resume ops and disable/enable sysfs at the same time.

Quoted 4 lines

Call led_trigger_suspend{resuem} API in pm_notifier function to suspend/resume all led devices listed in given trigger.
Signed-off-by Zhang Bo <[email protected]>

NAK.

led_trigger should not need special handling over suspend.

Quoted 5 lines

It can also fix my issue by reverting the commit 5ab92a7cb. But this action only does not make
led_set_brightness_nosleep function to set brightness. The heartbeat trigger timer is still
running even though it is not harmful.

Yes, timers can and should be active over suspend.

Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.) That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.

Pavel

Signature · 9 lines

(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170607/7171a591/attachment.sig>

Ulf Hansson

2017-06-07 09:27 UTC #2 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Quoted 29 lines · Pavel Machek

Quoted 13 lines · Zhang Bo
System cannot enter suspend mode because of heartbeat led trigger. In autosleep_wq, try_to_suspend function will try to enter suspend mode in specific period. it will get wakeup_count then call pm_notifier chain callback function and freeze processes. Heartbeat_pm_notifier is called and it call led_trigger_unregister to change the trigger of led device to none. It will send uevent message and the wakeup source count changed.
Add suspend/resume ops to the struct led_trigger, implement these two ops for ledtrig-heartbeat. Add led_trigger_suspend/led_trigger_resume funcitons to iterate through all LED class devices registered on given trigger and call their suspend/resume ops and disable/enable sysfs at the same time.
Quoted 4 lines
Call led_trigger_suspend{resuem} API in pm_notifier function to suspend/resume all led devices listed in given trigger.
Signed-off-by Zhang Bo <[email protected]>
NAK.
led_trigger should not need special handling over suspend.
Quoted 5 lines
It can also fix my issue by reverting the commit 5ab92a7cb. But this action only does not make
led_set_brightness_nosleep function to set brightness. The heartbeat trigger timer is still
running even though it is not harmful.

Quoted 8 lines

Yes, timers can and should be active over suspend.
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.) That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.

I fully agree! However, I haven't yet figured out exactly how to solve it yet. :-)

FYI, the same problem exist to for example ledtrig-cpu, which uses a sycore_ops do deal with suspend/resume.
I think that is way too late, especially since it's not uncommon that an i2c transfer needs to be done to change the led. Then as i2c is suspended far earlier than when entering the syscore stages in suspend, it breaks.

Kind regards
Uffe

Linus Walleij

2017-06-07 09:46 UTC #3 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

In reply to Pavel Machek

Quoted 5 lines · Pavel Machek

Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)

OK yeah I guess you're right.

I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.

Quoted 2 lines

That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.

Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?

That is essentially what I was trying to achieve.

The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.

The patch also solves the following phenomenon, sorry for not writing in the commit:

- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware

It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.

That is the tricky part, which is simplified by just removing and adding back the trigger.

Yours,
Linus Walleij

Pavel Machek

2017-06-07 10:31 UTC #4 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Hi!

Quoted 13 lines · Linus Walleij

Quoted 5 lines · Pavel Machek
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)
OK yeah I guess you're right.
I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.

Quoted 7 lines

Quoted 2 lines
That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.
Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?
That is essentially what I was trying to achieve.

I don't think we should be fixing it at trigger level.

If userspace keeps blinking using "brightness" attribute, would we like the LED to be off during suspend? I think so.

Quoted 20 lines

The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.
The patch also solves the following phenomenon, sorry for not writing in the commit:
- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware

Quoted 3 lines

It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.

No, leave the timer alone, and actually leave the trigger alone.

Simply make the driver turn the LED off in .suspend() callback, and then ignore further requests until .resume(). That will get rid of power drain _and_ "failed I2C transaction" messages, right?

(Actually, I tested with the heartbeat trigger, and it is not re-installed after resume. Another reason to revert the patch).

Thanks,

Pavel

Signature · 9 lines

(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170607/2879506a/attachment.sig>

Bruce Zhang

author

2017-06-08 02:57 UTC #5 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Hi,

As suspend/resume procedure is as below:
pm_notifier_call_chain(PM_SUSPEND_PREPARE)->device suspend -> syscore suspend -> enter suspend mode -> syscore resume -> device resume -> pm_notifier_call_chain(PM_POST_SUSPEND). If led brightness control depends on other device/hardware (such as I2C), the LED turn off action should be earlier than device suspend. pm_notifier may be a suitable method to turn off LED.

Best Regards,
Bo

-----Original Message-----
From: Pavel Machek [mailto:pavel at ucw.cz]
Sent: Wednesday, June 07, 2017 6:31 PM
To: Linus Walleij <[email protected]>
Cc: Bruce Zhang <[email protected]>; Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Jacek Anaszewski <[email protected]>; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend

Hi!

Quoted 13 lines · Linus Walleij

Quoted 5 lines · Pavel Machek
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)
OK yeah I guess you're right.
I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.

Quoted 7 lines

Quoted 2 lines
That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.
Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?
That is essentially what I was trying to achieve.

I don't think we should be fixing it at trigger level.

If userspace keeps blinking using "brightness" attribute, would we like the LED to be off during suspend? I think so.

Quoted 19 lines

The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.
The patch also solves the following phenomenon, sorry for not writing in the commit:
- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware

Quoted 3 lines

It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.

No, leave the timer alone, and actually leave the trigger alone.

Simply make the driver turn the LED off in .suspend() callback, and then ignore further requests until .resume(). That will get rid of power drain _and_ "failed I2C transaction" messages, right?

(Actually, I tested with the heartbeat trigger, and it is not re-installed after resume. Another reason to revert the patch).

Thanks,

Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Jacek Anaszewski

2017-06-08 20:27 UTC #6 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Hi Bruce,

Quoted 5 lines · Bruce Zhang

Hi,
As suspend/resume procedure is as below:
pm_notifier_call_chain(PM_SUSPEND_PREPARE)->device suspend -> syscore suspend -> enter suspend mode -> syscore resume -> device resume -> pm_notifier_call_chain(PM_POST_SUSPEND). If led brightness control depends on other device/hardware (such as I2C), the LED turn off action should be earlier than device suspend. pm_notifier may be a suitable method to turn off LED.

OK, so I agree that the most proper fix to the discussed problem would be ignoring brightness setting requests after suspend, as Pavel suggested.

led_classdev_suspend() already turns the LED off and led_classdev_resume() brings the brightness back.

Also led_set_brightness_nosleep() and led_set_brightness_sync() don't propagate brightness change to the driver if LED_SUSPENDED is set.

There are two problems though:
1. __led_set_brightness() and __led_set_brightness_blocking() don't check the flag, and they can be called from set_brightness_delayed(), that was already queued at the time of setting LED_SUSPENDED flag.
2. LED_SUSPENDED flag is accessed in a non-atomic way.

The solution as I see it would be respectively:
1. Return from __led_set_brightness() and __led_set_brightness_blocking() if LED_SUSPENDED is set.
2. Switch to accessing struct led_classdev's flags with use of bitops. Fortunately those flags are mainly set prior LED class device registration and only LED_SYSFS_DISABLE, LED_SUSPENDED and LED_UNREGISTERING are accessed afterwards, and only in the LED core, so only those places would have to be modified.

Best regards,
Jacek Anaszewski

Quoted 71 lines

-----Original Message-----
From: Pavel Machek [mailto:pavel at ucw.cz]
Sent: Wednesday, June 07, 2017 6:31 PM
To: Linus Walleij <[email protected]>
Cc: Bruce Zhang <[email protected]>; Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Jacek Anaszewski <[email protected]>; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi!
Quoted 13 lines · Linus Walleij
Quoted 5 lines · Pavel Machek
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)
OK yeah I guess you're right.
I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.
Quoted 7 lines
Quoted 2 lines
That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.
Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?
That is essentially what I was trying to achieve.
I don't think we should be fixing it at trigger level.
If userspace keeps blinking using "brightness" attribute, would we like the LED to be off during suspend? I think so.
Quoted 19 lines
The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.
The patch also solves the following phenomenon, sorry for not writing in the commit:
- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware
Quoted 3 lines
It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.
No, leave the timer alone, and actually leave the trigger alone.
Simply make the driver turn the LED off in .suspend() callback, and then ignore further requests until .resume(). That will get rid of power drain _and_ "failed I2C transaction" messages, right?
(Actually, I tested with the heartbeat trigger, and it is not re-installed after resume. Another reason to revert the patch).
Thanks,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Bruce Zhang

author

2017-06-09 02:57 UTC #7 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Hi Jacek,

I think that no change need to make for __led_set_brightness() and __led_set_brightness_blocking() function. Because led_classdev_suspend() need to turn off LED with these API. set_brightness_delayed() is the same level API as __led_set_brightness(). No need to add LED_SUSPENDED checking as well.

Best Regards,
Bo

-----Original Message-----
From: Jacek Anaszewski [mailto:jacek.anaszewski at gmail.com]
Sent: Friday, June 09, 2017 4:28 AM
To: Bruce Zhang <[email protected]>; Pavel Machek <[email protected]>; Linus Walleij <[email protected]>
Cc: Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend

Hi Bruce,

Quoted 5 lines · Bruce Zhang

Hi,
As suspend/resume procedure is as below:
pm_notifier_call_chain(PM_SUSPEND_PREPARE)->device suspend -> syscore suspend -> enter suspend mode -> syscore resume -> device resume -> pm_notifier_call_chain(PM_POST_SUSPEND). If led brightness control depends on other device/hardware (such as I2C), the LED turn off action should be earlier than device suspend. pm_notifier may be a suitable method to turn off LED.

OK, so I agree that the most proper fix to the discussed problem would be ignoring brightness setting requests after suspend, as Pavel suggested.

led_classdev_suspend() already turns the LED off and led_classdev_resume() brings the brightness back.

Also led_set_brightness_nosleep() and led_set_brightness_sync() don't propagate brightness change to the driver if LED_SUSPENDED is set.

Best regards,
Jacek Anaszewski

Quoted 76 lines

-----Original Message-----
From: Pavel Machek [mailto:pavel at ucw.cz]
Sent: Wednesday, June 07, 2017 6:31 PM
To: Linus Walleij <[email protected]>
Cc: Bruce Zhang <[email protected]>; Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Jacek Anaszewski <[email protected]>; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi!
Quoted 13 lines · Linus Walleij
Quoted 5 lines · Pavel Machek
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)
OK yeah I guess you're right.
I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.
Quoted 7 lines
Quoted 2 lines
That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.
Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?
That is essentially what I was trying to achieve.
I don't think we should be fixing it at trigger level.
If userspace keeps blinking using "brightness" attribute, would we like the LED to be off during suspend? I think so.
Quoted 19 lines
The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.
The patch also solves the following phenomenon, sorry for not writing in the commit:
- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware
Quoted 3 lines
It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.
No, leave the timer alone, and actually leave the trigger alone.
Simply make the driver turn the LED off in .suspend() callback, and then ignore further requests until .resume(). That will get rid of power drain _and_ "failed I2C transaction" messages, right?
(Actually, I tested with the heartbeat trigger, and it is not re-installed after resume. Another reason to revert the patch).
Thanks,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures)
http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Jacek Anaszewski

2017-06-09 20:47 UTC #8 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Hi Bruce,

Quoted 3 lines · Bruce Zhang

Hi Jacek,
I think that no change need to make for __led_set_brightness() and __led_set_brightness_blocking() function. Because led_classdev_suspend() need to turn off LED with these API. set_brightness_delayed() is the same level API as __led_set_brightness(). No need to add LED_SUSPENDED checking as well.

Indeed, it would be even harmful with the current sequence of instructions in the led_classdev_suspend().

So, the only action to take is to revert the commit 5ab92a7cb8.

Would you mind submitting the patch so that everyone interested could give their ack?

Best regards,
Jacek Anaszewski

Quoted 119 lines

-----Original Message-----
From: Jacek Anaszewski [mailto:jacek.anaszewski at gmail.com]
Sent: Friday, June 09, 2017 4:28 AM
To: Bruce Zhang <[email protected]>; Pavel Machek <[email protected]>; Linus Walleij <[email protected]>
Cc: Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi Bruce,
Quoted 5 lines · Bruce Zhang
Hi,
As suspend/resume procedure is as below:
pm_notifier_call_chain(PM_SUSPEND_PREPARE)->device suspend -> syscore suspend -> enter suspend mode -> syscore resume -> device resume -> pm_notifier_call_chain(PM_POST_SUSPEND). If led brightness control depends on other device/hardware (such as I2C), the LED turn off action should be earlier than device suspend. pm_notifier may be a suitable method to turn off LED.
OK, so I agree that the most proper fix to the discussed problem would be ignoring brightness setting requests after suspend, as Pavel suggested.
led_classdev_suspend() already turns the LED off and led_classdev_resume() brings the brightness back.
Also led_set_brightness_nosleep() and led_set_brightness_sync() don't propagate brightness change to the driver if LED_SUSPENDED is set.
There are two problems though:
1. __led_set_brightness() and __led_set_brightness_blocking() don't check the flag, and they can be called from set_brightness_delayed(), that was already queued at the time of setting LED_SUSPENDED flag.
2. LED_SUSPENDED flag is accessed in a non-atomic way.
The solution as I see it would be respectively:
1. Return from __led_set_brightness() and __led_set_brightness_blocking() if LED_SUSPENDED is set.
2. Switch to accessing struct led_classdev's flags with use of bitops. Fortunately those flags are mainly set prior LED class device registration and only LED_SYSFS_DISABLE, LED_SUSPENDED and LED_UNREGISTERING are accessed afterwards, and only in the LED core, so only those places would have to be modified.
Best regards,
Jacek Anaszewski
Quoted 76 lines
-----Original Message-----
From: Pavel Machek [mailto:pavel at ucw.cz]
Sent: Wednesday, June 07, 2017 6:31 PM
To: Linus Walleij <[email protected]>
Cc: Bruce Zhang <[email protected]>; Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Jacek Anaszewski <[email protected]>; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi!
Quoted 13 lines · Linus Walleij
Quoted 5 lines · Pavel Machek
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)
OK yeah I guess you're right.
I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.
Quoted 7 lines
Quoted 2 lines
That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.
Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?
That is essentially what I was trying to achieve.
I don't think we should be fixing it at trigger level.
If userspace keeps blinking using "brightness" attribute, would we like the LED to be off during suspend? I think so.
Quoted 19 lines
The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.
The patch also solves the following phenomenon, sorry for not writing in the commit:
- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware
Quoted 3 lines
It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.
No, leave the timer alone, and actually leave the trigger alone.
Simply make the driver turn the LED off in .suspend() callback, and then ignore further requests until .resume(). That will get rid of power drain _and_ "failed I2C transaction" messages, right?
(Actually, I tested with the heartbeat trigger, and it is not re-installed after resume. Another reason to revert the patch).
Thanks,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures)
http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Jacek Anaszewski

2017-06-09 22:21 UTC #9 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

Quoted 11 lines · Jacek Anaszewski

Hi Bruce,
Quoted 3 lines · Bruce Zhang
Hi Jacek,
I think that no change need to make for __led_set_brightness() and __led_set_brightness_blocking() function. Because led_classdev_suspend() need to turn off LED with these API. set_brightness_delayed() is the same level API as __led_set_brightness(). No need to add LED_SUSPENDED checking as well.
Indeed, it would be even harmful with the current sequence of instructions in the led_classdev_suspend().
So, the only action to take is to revert the commit 5ab92a7cb8.

Ugh, of course it would fix your problem with uevent, but it would also bring back the problem the commit 5ab92a7cb8 was solving.

The question is why the problem existed at all, despite that led_classdev_suspend() seems to do what's needed.

There are three options:
- races between led_heartbeat_function and LED core
- the LED class driver for which the phenomenon was observed doesn't set LED_CORE_SUSPENDRESUME flag
- if the driver uses brightness_set_blocking op, then maybe it is possible that the work queue task scheduled from led_classdev_suspend() that sets brightness to 0 isn't handled before the device enters suspend

We could narrow down the possible options if we knew the name of the LED class driver in question.

Best regards,
Jacek Anaszewski

Quoted 119 lines

Quoted 119 lines
-----Original Message-----
From: Jacek Anaszewski [mailto:jacek.anaszewski at gmail.com]
Sent: Friday, June 09, 2017 4:28 AM
To: Bruce Zhang <[email protected]>; Pavel Machek <[email protected]>; Linus Walleij <[email protected]>
Cc: Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi Bruce,
Quoted 5 lines · Bruce Zhang
Hi,
As suspend/resume procedure is as below:
pm_notifier_call_chain(PM_SUSPEND_PREPARE)->device suspend -> syscore suspend -> enter suspend mode -> syscore resume -> device resume -> pm_notifier_call_chain(PM_POST_SUSPEND). If led brightness control depends on other device/hardware (such as I2C), the LED turn off action should be earlier than device suspend. pm_notifier may be a suitable method to turn off LED.
OK, so I agree that the most proper fix to the discussed problem would be ignoring brightness setting requests after suspend, as Pavel suggested.
led_classdev_suspend() already turns the LED off and led_classdev_resume() brings the brightness back.
Also led_set_brightness_nosleep() and led_set_brightness_sync() don't propagate brightness change to the driver if LED_SUSPENDED is set.
There are two problems though:
1. __led_set_brightness() and __led_set_brightness_blocking() don't check the flag, and they can be called from set_brightness_delayed(), that was already queued at the time of setting LED_SUSPENDED flag.
2. LED_SUSPENDED flag is accessed in a non-atomic way.
The solution as I see it would be respectively:
1. Return from __led_set_brightness() and __led_set_brightness_blocking() if LED_SUSPENDED is set.
2. Switch to accessing struct led_classdev's flags with use of bitops. Fortunately those flags are mainly set prior LED class device registration and only LED_SYSFS_DISABLE, LED_SUSPENDED and LED_UNREGISTERING are accessed afterwards, and only in the LED core, so only those places would have to be modified.
Best regards,
Jacek Anaszewski
Quoted 76 lines
-----Original Message-----
From: Pavel Machek [mailto:pavel at ucw.cz]
Sent: Wednesday, June 07, 2017 6:31 PM
To: Linus Walleij <[email protected]>
Cc: Bruce Zhang <[email protected]>; Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Jacek Anaszewski <[email protected]>; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi!
Quoted 13 lines · Linus Walleij
Quoted 5 lines · Pavel Machek
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)
OK yeah I guess you're right.
I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.
Quoted 7 lines
Quoted 2 lines
That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.
Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?
That is essentially what I was trying to achieve.
I don't think we should be fixing it at trigger level.
If userspace keeps blinking using "brightness" attribute, would we like the LED to be off during suspend? I think so.
Quoted 19 lines
The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.
The patch also solves the following phenomenon, sorry for not writing in the commit:
- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware
Quoted 3 lines
It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.
No, leave the timer alone, and actually leave the trigger alone.
Simply make the driver turn the LED off in .suspend() callback, and then ignore further requests until .resume(). That will get rid of power drain _and_ "failed I2C transaction" messages, right?
(Actually, I tested with the heartbeat trigger, and it is not re-installed after resume. Another reason to revert the patch).
Thanks,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures)
http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html

Bruce Zhang

author

2017-06-11 04:43 UTC #10 lore raw

PATCH v2 1/1 ] drivers: leds/trigger: system cannot enter suspend

In reply to Jacek Anaszewski

Hi Jacek,

Of course I can submit the patch. But we need to confirm that it does not bring back the problem the commit 5ab92a7cb8 was solving. Maybe we can use reference counter to record if some device used by other device. When some device suspend, it should decrease the reference count of the other device that it depends on. Only when reference count of some device decrease to zero, it can enter suspend mode. By the way, we should avoid devices call loop.

Best Regards,
Bo

-----Original Message-----
From: Jacek Anaszewski [mailto:jacek.anaszewski at gmail.com]
Sent: Saturday, June 10, 2017 4:47 AM
To: Bruce Zhang <[email protected]>; Pavel Machek <[email protected]>; Linus Walleij <[email protected]>
Cc: Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend

Hi Bruce,

Quoted 3 lines · Bruce Zhang

Hi Jacek,
I think that no change need to make for __led_set_brightness() and __led_set_brightness_blocking() function. Because led_classdev_suspend() need to turn off LED with these API. set_brightness_delayed() is the same level API as __led_set_brightness(). No need to add LED_SUSPENDED checking as well.

Indeed, it would be even harmful with the current sequence of instructions in the led_classdev_suspend().

So, the only action to take is to revert the commit 5ab92a7cb8.

Would you mind submitting the patch so that everyone interested could give their ack?

Best regards,
Jacek Anaszewski

Quoted 123 lines

-----Original Message-----
From: Jacek Anaszewski [mailto:jacek.anaszewski at gmail.com]
Sent: Friday, June 09, 2017 4:28 AM
To: Bruce Zhang <[email protected]>; Pavel Machek <[email protected]>; Linus Walleij <[email protected]>
Cc: Ulf Hansson <[email protected]>;
linux-arm-kernel at lists.infradead.org; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi Bruce,
Quoted 5 lines · Bruce Zhang
Hi,
As suspend/resume procedure is as below:
pm_notifier_call_chain(PM_SUSPEND_PREPARE)->device suspend -> syscore suspend -> enter suspend mode -> syscore resume -> device resume -> pm_notifier_call_chain(PM_POST_SUSPEND). If led brightness control depends on other device/hardware (such as I2C), the LED turn off action should be earlier than device suspend. pm_notifier may be a suitable method to turn off LED.
OK, so I agree that the most proper fix to the discussed problem would be ignoring brightness setting requests after suspend, as Pavel suggested.
led_classdev_suspend() already turns the LED off and led_classdev_resume() brings the brightness back.
Also led_set_brightness_nosleep() and led_set_brightness_sync() don't propagate brightness change to the driver if LED_SUSPENDED is set.
There are two problems though:
1. __led_set_brightness() and __led_set_brightness_blocking() don't check the flag, and they can be called from set_brightness_delayed(), that was already queued at the time of setting LED_SUSPENDED flag.
2. LED_SUSPENDED flag is accessed in a non-atomic way.
The solution as I see it would be respectively:
1. Return from __led_set_brightness() and __led_set_brightness_blocking() if LED_SUSPENDED is set.
2. Switch to accessing struct led_classdev's flags with use of bitops. Fortunately those flags are mainly set prior LED class device registration and only LED_SYSFS_DISABLE, LED_SUSPENDED and LED_UNREGISTERING are accessed afterwards, and only in the LED core, so only those places would have to be modified.
Best regards,
Jacek Anaszewski
Quoted 76 lines
-----Original Message-----
From: Pavel Machek [mailto:pavel at ucw.cz]
Sent: Wednesday, June 07, 2017 6:31 PM
To: Linus Walleij <[email protected]>
Cc: Bruce Zhang <[email protected]>; Ulf Hansson <[email protected]>; linux-arm-kernel at lists.infradead.org; Jacek Anaszewski <[email protected]>; Richard Purdie <[email protected]>; linux-leds at vger.kernel.org
Subject: Re: [[PATCH v2]] drivers: leds/trigger: system cannot enter suspend
Hi!
Quoted 13 lines · Linus Walleij
Quoted 5 lines · Pavel Machek
Core reason is that 5ab92a7cb is wrong. (Very wrong; it papers over the issue for one trigger, but we have more than one... Even if special handling for heartbeat is warranted, that handling would be "turn the led off" not "unregister the trigger", which is userland-visible action.)
OK yeah I guess you're right.
I just couldn't think about anything better and didn't get any review at the time, so mistakes were made.
Quoted 7 lines
Quoted 2 lines
That one should be reverted, then maybe the driver should be fixed to turn the led off during suspend.
Do you mean that the heartbeat trigger driver should be fixed to turn off the LED during sleep?
That is essentially what I was trying to achieve.
I don't think we should be fixing it at trigger level.
If userspace keeps blinking using "brightness" attribute, would we like the LED to be off during suspend? I think so.
Quoted 19 lines
The reason it is done as it is, is that the trigger sets up a timer to do its job, and the timer may trigger between the point you turn off the LED and the system really goes to suspend, again maybe turning the LED on and again leaving the system with a glowing LED at suspend.
The patch also solves the following phenomenon, sorry for not writing in the commit:
- Turn off LED
- Suspend I2C hardware
- Timer trigger
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Actual suspen happens
- System comes online
- Trying to blink the LED using I2C
- Crap in the console about the failed I2C transaction
- Resume I2C hardware
Quoted 3 lines
It's just very fragile this trigger, turning off the LED from the PM notifier is obviously not enough, we also need to disable the timer, and then take it back online after resume.
No, leave the timer alone, and actually leave the trigger alone.
Simply make the driver turn the LED off in .suspend() callback, and then ignore further requests until .resume(). That will get rid of power drain _and_ "failed I2C transaction" messages, right?
(Actually, I tested with the heartbeat trigger, and it is not re-installed after resume. Another reason to revert the patch).
Thanks,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures)
http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html