Skip to contentSkip to conversation

cpuidle: Move cpuidle driver forward before acpi driver in Makefile

PATCH In review Dapeng Mi 3 participants 9 messages Jul 13, 2022 – Jul 26, 2022

lore
Share & tools
lore

Dapeng Mi

author
#0 lore raw

PATCH 1/1 cpuidle: Move cpuidle driver forward before acpi driver in Makefile

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/Makefile +1 -1
index 9a30842b22c5..921ed481b520 100644
@@ -26,6 +26,7 @@ obj-y				+= idle/
 # IPMI must come before ACPI in order to provide IPMI opregion support
 obj-y				+= char/ipmi/
 
+obj-$(CONFIG_CPU_IDLE)		+= cpuidle/
 obj-$(CONFIG_ACPI)		+= acpi/
 
 # PnP must come after ACPI since it will eventually need to check if acpi
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)		+= edac/
 obj-$(CONFIG_EISA)		+= eisa/
 obj-$(CONFIG_PM_OPP)		+= opp/
 obj-$(CONFIG_CPU_FREQ)		+= cpufreq/
-obj-$(CONFIG_CPU_IDLE)		+= cpuidle/
 obj-y				+= mmc/
 obj-y				+= ufs/
 obj-$(CONFIG_MEMSTICK)		+= memstick/
Signature · 1 line
2.30.2
#1 lore raw
Quoted 28 lines · Dapeng Mi

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/Makefile +1 -0
index 9a30842b22c5..921ed481b520 100644
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI opregion support
 obj-y                          += char/ipmi/
 
+obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-$(CONFIG_ACPI)             += acpi/
 
 # PnP must come after ACPI since it will eventually need to check if acpi

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 8 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Also what if haltpoll is enabled, but the user wants ACPI idle?

#2 lore raw
Quoted 40 lines · Rafael J. Wysocki
Quoted 28 lines · Dapeng Mi

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/Makefile +1 -0
index 9a30842b22c5..921ed481b520 100644
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI opregion support
 obj-y                          += char/ipmi/
 
+obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-$(CONFIG_ACPI)             += acpi/
 
 # PnP must come after ACPI since it will eventually need to check if acpi

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 8 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Quoted 1 line

Also what if haltpoll is enabled, but the user wants ACPI idle?

Exactly. For example while on real boxes BIOS might often present broken ACPI idle, on VMs I am guessing if ACPI is present one can assume it's actually correct.

Signature · 1 line
MST
#3 lore raw
In reply to Rafael J. Wysocki
Quoted 50 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Thursday, July 14, 2022 1:53 AM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

Quoted 28 lines · Dapeng Mi

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/Makefile index renamed from drivers/Makefile +0 -0

9a30842b22c5..921ed481b520 100644

drivers/Makefile +1 -0
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI opregion support
 obj-y                          += char/ipmi/
 
+obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-$(CONFIG_ACPI)             += acpi/
 
 # PnP must come after ACPI since it will eventually need to check if acpi

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 8 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Quoted 1 line

Also what if haltpoll is enabled, but the user wants ACPI idle?

Thanks Rafael for reviewing this patch.

acpi_idle driver and cpuidle_haltpoll driver have same initialization level and both are initialized on the level device_initcall. So the building order would decide the loading sequence. Just like the intel_idle driver which also has same initialization level (device_initcall), but as it's built before acpi_idle driver, it would be loaded first before acpi_driver if intel_idle driver is enabled.

There is another method to make cpuidle_haltpoll driver loaded first before acpi_driver, it's change the initialization level to postcore_initcall. I'm not sure which one is better, but it seems current patch is more reasonable.

There is an parameter "force" to manage the haltpoll enabling. If user want to use ACPI idle, it can change this parameter to disable haltpolll driver.

#4 lore raw
Quoted 63 lines
Quoted 11 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Thursday, July 14, 2022 1:53 AM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

On Wed, Jul 13, 2022 at 10:21 AM Dapeng Mi <[email protected]>

Quoted 15 lines · wrote:
Quoted 15 lines

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 6 lines

drivers/Makefile obj-y += idle/ · line 26

Quoted 6 lines
drivers/Makefile index renamed from drivers/Makefile +0 -0

9a30842b22c5..921ed481b520 100644

drivers/Makefile +0 -0
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI opregion

support

Quoted 18 lines
Quoted 7 lines
drivers/Makefile +1 -0
 obj-y                          += char/ipmi/
 
+obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-$(CONFIG_ACPI)             += acpi/
 
 # PnP must come after ACPI since it will eventually need to check if acpi

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 8 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Quoted 1 line

Also what if haltpoll is enabled, but the user wants ACPI idle?

Thanks Rafael for reviewing this patch.

acpi_idle driver and cpuidle_haltpoll driver have same initialization level and both are initialized on the level device_initcall. So the building order would decide the loading sequence. Just like the intel_idle driver which also has same initialization level (device_initcall), but as it's built before acpi_idle driver, it would be loaded first before acpi_driver if intel_idle driver is enabled.

Quoted 6 lines

There is another method to make cpuidle_haltpoll driver loaded first before acpi_driver, it's change the initialization level to postcore_initcall. I'm not sure which one is better, but it seems current patch is more reasonable.

There is an parameter "force" to manage the haltpoll enabling. If user want to use ACPI idle, it can change this parameter to disable haltpolll driver.

Any feedback on this? Thanks.

#5 lore raw
Quoted 64 lines · On Wed, Jul 20, 2022 at 5:00 AM Mi, Dapeng1 <[email protected]> wrote:
Quoted 64 lines
Quoted 12 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Thursday, July 14, 2022 1:53 AM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

On Wed, Jul 13, 2022 at 10:21 AM Dapeng Mi <[email protected]>

Quoted 15 lines · wrote:
Quoted 15 lines

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 6 lines

drivers/Makefile obj-y += idle/ · line 26

Quoted 6 lines
drivers/Makefile index renamed from drivers/Makefile +0 -0

9a30842b22c5..921ed481b520 100644

drivers/Makefile +0 -0
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI opregion

support

Quoted 18 lines
Quoted 7 lines
drivers/Makefile +1 -0
 obj-y                          += char/ipmi/
 
+obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-$(CONFIG_ACPI)             += acpi/
 
 # PnP must come after ACPI since it will eventually need to check if acpi

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 8 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Quoted 1 line

Also what if haltpoll is enabled, but the user wants ACPI idle?

Thanks Rafael for reviewing this patch.

acpi_idle driver and cpuidle_haltpoll driver have same initialization level and both are initialized on the level device_initcall. So the building order would decide the loading sequence. Just like the intel_idle driver which also has same initialization level (device_initcall), but as it's built before acpi_idle driver, it would be loaded first before acpi_driver if intel_idle driver is enabled.

Quoted 6 lines
Quoted 6 lines

There is another method to make cpuidle_haltpoll driver loaded first before acpi_driver, it's change the initialization level to postcore_initcall. I'm not sure which one is better, but it seems current patch is more reasonable.

There is an parameter "force" to manage the haltpoll enabling. If user want to use ACPI idle, it can change this parameter to disable haltpolll driver.

That would require things to be appended to the kernel command line in cases where that's not necessary today and that's not acceptable.

What you really seem to be wanting to do is to use haltpoll instead of ACPI idle. Is that correct?

#6 lore raw
Quoted 83 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Wednesday, July 20, 2022 6:24 PM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>; Zhenyu Wang <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

Quoted 6 lines · On Wed, Jul 20, 2022 at 5:00 AM Mi, Dapeng1 <[email protected]> wrote:
Quoted 6 lines
Quoted 6 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Thursday, July 14, 2022 1:53 AM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van

Assche

Quoted 29 lines
Quoted 29 lines
Quoted 6 lines

<[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

On Wed, Jul 13, 2022 at 10:21 AM Dapeng Mi <[email protected]>

Quoted 15 lines · wrote:
Quoted 15 lines

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 6 lines

drivers/Makefile obj-y += idle/ · line 26

Quoted 6 lines
drivers/Makefile index renamed from drivers/Makefile +0 -0

9a30842b22c5..921ed481b520 100644

drivers/Makefile +0 -0
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI opregion

support

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

drivers/Makefile obj-y += idle/

Quoted 6 lines
drivers/Makefile +1 -0
 obj-y                          += char/ipmi/
 
+obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-$(CONFIG_ACPI)             += acpi/
 
 # PnP must come after ACPI since it will eventually need to check if

acpi

Quoted 14 lines
Quoted 14 lines
Quoted 13 lines

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 9 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Also what if haltpoll is enabled, but the user wants ACPI idle?

Quoted 8 lines
Quoted 8 lines

Thanks Rafael for reviewing this patch.

acpi_idle driver and cpuidle_haltpoll driver have same initialization level and both are initialized on the level device_initcall. So the building order would decide the loading sequence. Just like the intel_idle driver which also has same initialization level (device_initcall), but as it's built before acpi_idle driver, it would be loaded first before acpi_driver if intel_idle

driver is enabled.

Quoted 4 lines
Quoted 4 lines

There is another method to make cpuidle_haltpoll driver loaded first before acpi_driver, it's change the initialization level to postcore_initcall. I'm not sure which one is better, but it seems current

Quoted 9 lines

patch is more reasonable.

Quoted 3 lines
Quoted 3 lines

There is an parameter "force" to manage the haltpoll enabling. If user want to use ACPI idle, it can change this parameter to disable

haltpolll driver.

That would require things to be appended to the kernel command line in cases where that's not necessary today and that's not acceptable.

The haltpoll driver's "force" parameter is false by default, we don't need to add extra command line options in most cases except we want to enable the haltpolling driver.

Quoted 2 lines

What you really seem to be wanting to do is to use haltpoll instead of ACPI idle. Is that correct?

Yes, I'm trying to enable guest halt polling to improve the performance of some Virtual Machine. But I found the halt poll driver can't be enabled as long as acpi idle driver is enabled. I tried to disable the acpi idle first, but I found there is no parameter to enable/disable acpi idle driver except disabling the Kconfig "CONFIG_ACPI_PROCESSOR_IDLE", and but unfortunately Kconfig "ACPI_PROCESSOR" would enable "ACPI_PROCESSOR_IDLE" by default. If I want to disable acpi_idle driver, I have to disable the "ACPI_PROCESSOR", but this would cause acpi throttling and acpi thermal are also disabled. That's not what I want. That's why I do this change to make halt poll driver has a chance to run without disable the entire acpi processor functionality .

#7 lore raw
Quoted 115 lines

From: Mi, Dapeng1
Sent: Thursday, July 21, 2022 11:09 AM
To: Rafael J. Wysocki <[email protected]>
Cc: Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>; Zhenyu Wang <[email protected]>
Subject: RE: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

Quoted 90 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Wednesday, July 20, 2022 6:24 PM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>; Zhenyu Wang <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

Quoted 6 lines · On Wed, Jul 20, 2022 at 5:00 AM Mi, Dapeng1 <[email protected]> wrote:
Quoted 6 lines
Quoted 6 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Thursday, July 14, 2022 1:53 AM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van

Assche

Quoted 31 lines
Quoted 31 lines
Quoted 7 lines

<[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

On Wed, Jul 13, 2022 at 10:21 AM Dapeng Mi
<[email protected]>

Quoted 15 lines · wrote:
Quoted 15 lines

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 7 lines
Quoted 7 lines
drivers/Makefile index renamed from drivers/Makefile +0 -0

9a30842b22c5..921ed481b520 100644

drivers/Makefile +0 -0
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI

opregion

support

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

obj-y += char/ipmi/

+obj-$(CONFIG_CPU_IDLE) += cpuidle/
obj-$(CONFIG_ACPI) += acpi/

# PnP must come after ACPI since it will eventually need to
check if

acpi

Quoted 16 lines
Quoted 16 lines
Quoted 12 lines

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 9 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Quoted 1 line

Also what if haltpoll is enabled, but the user wants ACPI idle?

Thanks Rafael for reviewing this patch.

Quoted 7 lines
Quoted 7 lines

acpi_idle driver and cpuidle_haltpoll driver have same initialization level and both are initialized on the level device_initcall. So the building order would decide the loading sequence. Just like the intel_idle driver which also has same initialization level (device_initcall), but as it's built before acpi_idle driver, it would be loaded first before acpi_driver if intel_idle

driver is enabled.

Quoted 5 lines
Quoted 5 lines

There is another method to make cpuidle_haltpoll driver loaded first before acpi_driver, it's change the initialization level to postcore_initcall. I'm not sure which one is better, but it seems current

patch is more reasonable.

Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

There is an parameter "force" to manage the haltpoll enabling. If user want to use ACPI idle, it can change this parameter to disable

haltpolll driver.

That would require things to be appended to the kernel command line in cases where that's not necessary today and that's not acceptable.

The haltpoll driver's "force" parameter is false by default, we don't need to add extra command line options in most cases except we want to enable the haltpolling driver.

Quoted 2 lines

What you really seem to be wanting to do is to use haltpoll instead of ACPI idle. Is that correct?

Quoted 12 lines

Yes, I'm trying to enable guest halt polling to improve the performance of some Virtual Machine. But I found the halt poll driver can't be enabled as long as acpi idle driver is enabled. I tried to disable the acpi idle first, but I found there is no parameter to enable/disable acpi idle driver except disabling the Kconfig "CONFIG_ACPI_PROCESSOR_IDLE", and but unfortunately Kconfig "ACPI_PROCESSOR" would enable "ACPI_PROCESSOR_IDLE" by default. If I want to disable acpi_idle driver, I have to disable the "ACPI_PROCESSOR", but this would cause acpi throttling and acpi thermal are also disabled. That's not what I want. That's why I do this change to make halt poll driver has a chance to run without disable the entire acpi processor functionality .

Any feedback? Thanks.

#8 lore raw
Quoted 129 lines · On Tue, Jul 26, 2022 at 3:46 AM Mi, Dapeng1 <[email protected]> wrote:
Quoted 116 lines

From: Mi, Dapeng1
Sent: Thursday, July 21, 2022 11:09 AM
To: Rafael J. Wysocki <[email protected]>
Cc: Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>; Zhenyu Wang <[email protected]>
Subject: RE: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

Quoted 90 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Wednesday, July 20, 2022 6:24 PM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van Assche <[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>; Zhenyu Wang <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

Quoted 6 lines · On Wed, Jul 20, 2022 at 5:00 AM Mi, Dapeng1 <[email protected]> wrote:
Quoted 6 lines
Quoted 6 lines

From: Rafael J. Wysocki <[email protected]>
Sent: Thursday, July 14, 2022 1:53 AM
To: Mi, Dapeng1 <[email protected]>
Cc: Rafael J. Wysocki <[email protected]>; Michael S. Tsirkin <[email protected]>; Arnd Bergmann <[email protected]>; Bart Van

Assche

Quoted 31 lines
Quoted 31 lines
Quoted 7 lines

<[email protected]>; Linux Kernel Mailing List <linux- [email protected]>; Linux PM <[email protected]>
Subject: Re: [PATCH] cpuidle: Move cpuidle driver forward before acpi driver in Makefile

On Wed, Jul 13, 2022 at 10:21 AM Dapeng Mi
<[email protected]>

Quoted 15 lines · wrote:
Quoted 15 lines

As long as Kconfig ACPI_PROCESSOR is enabled, ACPI_PROCESSOR would select ACPI_PROCESSOR_IDLE and acpi_idle driver is enabled. But in current driver loading order acpi_idle driver is always loaded before cpuidle_haltpoll driver. This leads to cpuidle_hatpoll driver has no chance to be loaded when it's enabled.

Thus, move cpuidle driver forward before acpi driver and make cpuidle-hatpoll driver has a chance to be run when it's enabled.

---
 drivers/Makefile | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 7 lines
Quoted 7 lines
drivers/Makefile index renamed from drivers/Makefile +0 -0

9a30842b22c5..921ed481b520 100644

drivers/Makefile +0 -0
@@ -26,6 +26,7 @@ obj-y                         += idle/
 # IPMI must come before ACPI in order to provide IPMI

opregion

support

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

obj-y += char/ipmi/

+obj-$(CONFIG_CPU_IDLE) += cpuidle/
obj-$(CONFIG_ACPI) += acpi/

# PnP must come after ACPI since it will eventually need to
check if

acpi

Quoted 16 lines
Quoted 16 lines
Quoted 12 lines

drivers/Makefile obj-$(CONFIG_EDAC) += edac/

Quoted 9 lines
drivers/Makefile +0 -2
@@ -126,7 +127,6 @@ obj-$(CONFIG_EDAC)          += edac/
 obj-$(CONFIG_EISA)             += eisa/
 obj-$(CONFIG_PM_OPP)           += opp/
 obj-$(CONFIG_CPU_FREQ)         += cpufreq/
-obj-$(CONFIG_CPU_IDLE)         += cpuidle/
 obj-y                          += mmc/
 obj-y                          += ufs/
 obj-$(CONFIG_MEMSTICK)         += memstick/
--

Well, this change doesn't guarantee loading haltpoll before ACPI idle.

Quoted 1 line

Also what if haltpoll is enabled, but the user wants ACPI idle?

Thanks Rafael for reviewing this patch.

Quoted 7 lines
Quoted 7 lines

acpi_idle driver and cpuidle_haltpoll driver have same initialization level and both are initialized on the level device_initcall. So the building order would decide the loading sequence. Just like the intel_idle driver which also has same initialization level (device_initcall), but as it's built before acpi_idle driver, it would be loaded first before acpi_driver if intel_idle

driver is enabled.

Quoted 5 lines
Quoted 5 lines

There is another method to make cpuidle_haltpoll driver loaded first before acpi_driver, it's change the initialization level to postcore_initcall. I'm not sure which one is better, but it seems current

patch is more reasonable.

Quoted 1 line
Quoted 1 line
Quoted 8 lines
Quoted 3 lines
Quoted 3 lines

There is an parameter "force" to manage the haltpoll enabling. If user want to use ACPI idle, it can change this parameter to disable

haltpolll driver.

That would require things to be appended to the kernel command line in cases where that's not necessary today and that's not acceptable.

The haltpoll driver's "force" parameter is false by default, we don't need to add extra command line options in most cases except we want to enable the haltpolling driver.

Quoted 2 lines

What you really seem to be wanting to do is to use haltpoll instead of ACPI idle. Is that correct?

Quoted 12 lines

Yes, I'm trying to enable guest halt polling to improve the performance of some Virtual Machine. But I found the halt poll driver can't be enabled as long as acpi idle driver is enabled. I tried to disable the acpi idle first, but I found there is no parameter to enable/disable acpi idle driver except disabling the Kconfig "CONFIG_ACPI_PROCESSOR_IDLE", and but unfortunately Kconfig "ACPI_PROCESSOR" would enable "ACPI_PROCESSOR_IDLE" by default. If I want to disable acpi_idle driver, I have to disable the "ACPI_PROCESSOR", but this would cause acpi throttling and acpi thermal are also disabled. That's not what I want. That's why I do this change to make halt poll driver has a chance to run without disable the entire acpi processor functionality .

Quoted 1 line

Any feedback? Thanks.

I've already said that messing up with Makefiles is not the way to go here.

You have to enforce a specific ordering of initialization in the code.