Skip to contentSkip to conversation

mfd: intel-lpss: Avoid resuming runtime-suspended lpss unnecessarily

PATCH In review Chen Yu 4 participants 8 messages Sep 29, 2016 – Jan 3, 2017

Subsystems (5)
lore
Share & tools
lore

Chen Yu

author
#0 lore raw

PATCH v3 1/1 mfd: intel-lpss: Avoid resuming runtime-suspended lpss unnecessarily

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
 	/*
+	 * This is safe because:
+	 * 1. The runtime suspend and system suspend
+	 * are of the same hook.
+	 * 2. This device is neither runtime wakeup source
+	 * nor system wakeup source.
+	 */
+	if (pm_runtime_status_suspended(dev))
+		return DPM_DIRECT_COMPLETE;
+	/*
 	 * Resume both child devices before entering system sleep. This
 	 * ensures that they are in proper state before they get suspended.
 	 */
include/linux/pm.h +7 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
 	DPM_ORDER_DEV_LAST,
 };
 
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
Signature · 1 line
2.7.4
#1 lore raw

Hi Lee,

Quoted 17 lines · Chen Yu

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

Quoted 8 lines

If this is fine with you and you'd like to apply it, please feel free to add my ACK to it.

Alternatively, if you'd prefer me to apply it, please let me know.

Quoted 32 lines
---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
        /*
+        * This is safe because:
+        * 1. The runtime suspend and system suspend
+        * are of the same hook.
+        * 2. This device is neither runtime wakeup source
+        * nor system wakeup source.
+        */
+       if (pm_runtime_status_suspended(dev))
+               return DPM_DIRECT_COMPLETE;
+       /*
         * Resume both child devices before entering system sleep. This
         * ensures that they are in proper state before they get suspended.
         */
include/linux/pm.h +0 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
        DPM_ORDER_DEV_LAST,
 };

include/linux/pm.h enum dpm_order {

Quoted 10 lines
include/linux/pm.h +7 -1
 
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
--

Thanks,
Rafael

#2 lore raw
Quoted 29 lines · Rafael J. Wysocki

Hi Lee,

Quoted 17 lines · Chen Yu

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

Quoted 8 lines
Quoted 4 lines

If this is fine with you and you'd like to apply it, please feel free to add my ACK to it.

Alternatively, if you'd prefer me to apply it, please let me know.

You want this in for v3.9?

I just started applying patches for v3.10.

If you're certain there are 0% chance of regressions, I will still apply this for v3.9 with your Ack.

Quoted 45 lines
Quoted 32 lines
---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
        /*
+        * This is safe because:
+        * 1. The runtime suspend and system suspend
+        * are of the same hook.
+        * 2. This device is neither runtime wakeup source
+        * nor system wakeup source.
+        */
+       if (pm_runtime_status_suspended(dev))
+               return DPM_DIRECT_COMPLETE;
+       /*
         * Resume both child devices before entering system sleep. This
         * ensures that they are in proper state before they get suspended.
         */
include/linux/pm.h +0 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
        DPM_ORDER_DEV_LAST,
 };

include/linux/pm.h enum dpm_order {

Quoted 10 lines
include/linux/pm.h +7 -1
 
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
--

Thanks,
Rafael

Signature · 4 lines
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org │ Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog
#3 lore raw
Quoted 31 lines · Lee Jones
Quoted 29 lines · Rafael J. Wysocki

Hi Lee,

Quoted 25 lines · Chen Yu

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

Quoted 6 lines
Quoted 4 lines

If this is fine with you and you'd like to apply it, please feel free to add my ACK to it.

Alternatively, if you'd prefer me to apply it, please let me know.

You want this in for v3.9?

I'd rather queue it up for 4.10 (assuming that the above and below major version numbers are simply off by one by mistake).

Quoted 4 lines

I just started applying patches for v3.10.

If you're certain there are 0% chance of regressions, I will still apply this for v3.9 with your Ack.

4.10 should be fine.

Quoted 33 lines
Quoted 33 lines
Quoted 33 lines
---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
        /*
+        * This is safe because:
+        * 1. The runtime suspend and system suspend
+        * are of the same hook.
+        * 2. This device is neither runtime wakeup source
+        * nor system wakeup source.
+        */
+       if (pm_runtime_status_suspended(dev))
+               return DPM_DIRECT_COMPLETE;
+       /*
         * Resume both child devices before entering system sleep. This
         * ensures that they are in proper state before they get suspended.
         */
include/linux/pm.h +0 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
        DPM_ORDER_DEV_LAST,
 };
 
Quoted 9 lines
Quoted 9 lines

include/linux/pm.h enum dpm_order {

Quoted 9 lines
include/linux/pm.h +7 -1
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
--

Thanks,
Rafael

#4 lore raw
Quoted 35 lines · Rafael J. Wysocki
Quoted 34 lines · Lee Jones
Quoted 32 lines · Rafael J. Wysocki

Hi Lee,

Quoted 17 lines · Chen Yu

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

Quoted 8 lines

If this is fine with you and you'd like to apply it, please feel free to add my ACK to it.

Quoted 6 lines
Quoted 3 lines
Quoted 1 line

Alternatively, if you'd prefer me to apply it, please let me know.

You want this in for v3.9?

I'd rather queue it up for 4.10 (assuming that the above and below major version numbers are simply off by one by mistake).

Yes, of course they are off by one.

This is what happens when you conduct mail duties with jetlag. :)

Quoted 6 lines
Quoted 4 lines

I just started applying patches for v3.10.

If you're certain there are 0% chance of regressions, I will still apply this for v3.9 with your Ack.

4.10 should be fine.

Great.

Quoted 45 lines
Quoted 33 lines
Quoted 33 lines
Quoted 33 lines
---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
        /*
+        * This is safe because:
+        * 1. The runtime suspend and system suspend
+        * are of the same hook.
+        * 2. This device is neither runtime wakeup source
+        * nor system wakeup source.
+        */
+       if (pm_runtime_status_suspended(dev))
+               return DPM_DIRECT_COMPLETE;
+       /*
         * Resume both child devices before entering system sleep. This
         * ensures that they are in proper state before they get suspended.
         */
include/linux/pm.h +0 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
        DPM_ORDER_DEV_LAST,
 };
 
Quoted 9 lines
Quoted 9 lines

include/linux/pm.h enum dpm_order {

Quoted 9 lines
include/linux/pm.h +7 -1
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
--

Thanks,
Rafael

Signature · 4 lines
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org │ Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog
#5 lore raw
In reply to Chen Yu
Quoted 25 lines · Chen Yu

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

Quoted 4 lines
---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)

Applied for v4.10, thanks.

Quoted 36 lines
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
 	/*
+	 * This is safe because:
+	 * 1. The runtime suspend and system suspend
+	 * are of the same hook.
+	 * 2. This device is neither runtime wakeup source
+	 * nor system wakeup source.
+	 */
+	if (pm_runtime_status_suspended(dev))
+		return DPM_DIRECT_COMPLETE;
+	/*
 	 * Resume both child devices before entering system sleep. This
 	 * ensures that they are in proper state before they get suspended.
 	 */
include/linux/pm.h +7 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
 	DPM_ORDER_DEV_LAST,
 };
 
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
Signature · 4 lines
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org │ Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog
#6 lore raw
In reply to Chen Yu

Hi Lee,

you wrote on Oct 4 that you applied the below patch for v4.10. However I don't see the patch on your for-mfd-next branch. Maybe the patch slipped between the cracks, please double-check. I would like to use the DPM_DIRECT_COMPLETE macro introduced by it in a patch of my own.

I also don't see a pull from you in Linus' tree yet. Please note that this merge window will be shorter than usual and pulls should be sent out before coming Friday.

Thanks,

Lukas

Quoted 66 lines · Chen Yu

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
 	/*
+	 * This is safe because:
+	 * 1. The runtime suspend and system suspend
+	 * are of the same hook.
+	 * 2. This device is neither runtime wakeup source
+	 * nor system wakeup source.
+	 */
+	if (pm_runtime_status_suspended(dev))
+		return DPM_DIRECT_COMPLETE;
+	/*
 	 * Resume both child devices before entering system sleep. This
 	 * ensures that they are in proper state before they get suspended.
 	 */
include/linux/pm.h +7 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
 	DPM_ORDER_DEV_LAST,
 };
 
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
#7 lore raw
Quoted 5 lines · Lukas Wunner

you wrote on Oct 4 that you applied the below patch for v4.10. However I don't see the patch on your for-mfd-next branch. Maybe the patch slipped between the cracks, please double-check. I would like to use the DPM_DIRECT_COMPLETE macro introduced by it in a patch of my own.

Chen asked me to remove the patch.

Quoted 3 lines

I also don't see a pull from you in Linus' tree yet. Please note that this merge window will be shorter than usual and pulls should be sent out before coming Friday.

Right, I was on vacation. All sorted now.

Quoted 67 lines · Chen Yu
Quoted 66 lines

We have report that the intel_lpss_prepare() takes too much time during suspend, and this is because we first resume the devices from runtime suspend by resume_lpss_device(), to make sure they are in proper state before system suspend, which takes 100ms for each LPSS devices(PCI power state from D3_cold to D0). And since resume_lpss_device() resumes the devices synchronously, we might get huge latency if we have many LPSS devices.

So first try is to use pm_request_resume() instead, to make the runtime resume process asynchronously. Unfortunately the asynchronous runtime resume relies on pm_wq, which is freezed at early stage. So we choose another method, that is to avoid resuming runtime-suspended devices, if they are already runtime suspended. This is safe because for LPSS driver, the runtime suspend and system suspend are of the same hook - i.e., intel_lpss_suspend(). And moreover, this device is neither runtime wakeup source nor system wakeup source.

---
 drivers/mfd/intel-lpss.c | 9 +++++++++
 include/linux/pm.h       | 7 +++++++
 2 files changed, 16 insertions(+)
drivers/mfd/intel-lpss.c +9 -0
index 41b1138..2583db8 100644
@@ -485,6 +485,15 @@ static int resume_lpss_device(struct device *dev, void *data)
 int intel_lpss_prepare(struct device *dev)
 {
 	/*
+	 * This is safe because:
+	 * 1. The runtime suspend and system suspend
+	 * are of the same hook.
+	 * 2. This device is neither runtime wakeup source
+	 * nor system wakeup source.
+	 */
+	if (pm_runtime_status_suspended(dev))
+		return DPM_DIRECT_COMPLETE;
+	/*
 	 * Resume both child devices before entering system sleep. This
 	 * ensures that they are in proper state before they get suspended.
 	 */
include/linux/pm.h +7 -0
index 06eb353..4a788b4 100644
@@ -786,4 +786,11 @@ enum dpm_order {
 	DPM_ORDER_DEV_LAST,
 };
 
+/*
+ * Return this from system suspend/hibernation ->prepare() callback to
+ * request the core to leave the device runtime-suspended during system
+ * suspend if possible.
+ */
+#define DPM_DIRECT_COMPLETE 1
+
 #endif /* _LINUX_PM_H */
Signature · 4 lines
Lee Jones
Linaro STMicroelectronics Landing Team Lead
Linaro.org │ Open source software for ARM SoCs
Follow Linaro: Facebook | Twitter | Blog