[PATCH] ACPI: implement acpi_os_get_timer() according the spec

[PATCH] ACPI: implement acpi_os_get_timer() according the spec

[Mika Westerberg]

ACPI Timer() opcode should return monotonically increasing clock with 100ns
granularity. Implement this with the help of ktime_get().

Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
---
 drivers/acpi/osl.c |   14 +-------------
 1 file changed, 1 insertion(+), 13 deletions(-)

diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c
index 586e7e9..2a22170 100644
--- a/drivers/acpi/osl.c
+++ b/drivers/acpi/osl.c
@@ -835,19 +835,7 @@ void acpi_os_stall(u32 us)
  */
 u64 acpi_os_get_timer(void)
 {
-	static u64 t;
-
-#ifdef	CONFIG_HPET
-	/* TBD: use HPET if available */
-#endif
-
-#ifdef	CONFIG_X86_PM_TIMER
-	/* TBD: default to PM timer if HPET was not available */
-#endif
-	if (!t)
-		printk(KERN_ERR PREFIX "acpi_os_get_timer() TBD\n");
-
-	return ++t;
+	return ktime_to_ns(ktime_get()) / 100;
 }
 
 acpi_status acpi_os_read_port(acpi_io_address port, u32 * value, u32 width)
-- 
1.7.10.4

Re: [PATCH] ACPI: implement acpi_os_get_timer() according the spec

[Rafael J. Wysocki]

On Monday, May 13, 2013 01:27:51 PM Mika Westerberg wrote:
> ACPI Timer() opcode should return monotonically increasing clock with 100ns
> granularity. Implement this with the help of ktime_get().
> 
> Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>

That looks reasobable.  Have you tested it?

Rafael


> ---
>  drivers/acpi/osl.c |   14 +-------------
>  1 file changed, 1 insertion(+), 13 deletions(-)
> 
> diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c
> index 586e7e9..2a22170 100644
> --- a/drivers/acpi/osl.c
> +++ b/drivers/acpi/osl.c
> @@ -835,19 +835,7 @@ void acpi_os_stall(u32 us)
>   */
>  u64 acpi_os_get_timer(void)
>  {
> -	static u64 t;
> -
> -#ifdef	CONFIG_HPET
> -	/* TBD: use HPET if available */
> -#endif
> -
> -#ifdef	CONFIG_X86_PM_TIMER
> -	/* TBD: default to PM timer if HPET was not available */
> -#endif
> -	if (!t)
> -		printk(KERN_ERR PREFIX "acpi_os_get_timer() TBD\n");
> -
> -	return ++t;
> +	return ktime_to_ns(ktime_get()) / 100;
>  }
>  
>  acpi_status acpi_os_read_port(acpi_io_address port, u32 * value, u32 width)
> 
-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.

Re: [PATCH] ACPI: implement acpi_os_get_timer() according the spec

[Mika Westerberg]

On Mon, May 13, 2013 at 01:38:46PM +0200, Rafael J. Wysocki wrote:
> On Monday, May 13, 2013 01:27:51 PM Mika Westerberg wrote:
> > ACPI Timer() opcode should return monotonically increasing clock with 100ns
> > granularity. Implement this with the help of ktime_get().
> > 
> > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> 
> That looks reasobable.  Have you tested it?

Very lightly. Basically I added some debug printks() between two
successsive calls of Timer() and it seemed like it returned correct time.

It is certainly better than returning t+1 every time Timer() is called :)

Re: [PATCH] ACPI: implement acpi_os_get_timer() according the spec

[Mika Westerberg]

On Mon, May 13, 2013 at 02:44:32PM +0300, Mika Westerberg wrote:
> On Mon, May 13, 2013 at 01:38:46PM +0200, Rafael J. Wysocki wrote:
> > On Monday, May 13, 2013 01:27:51 PM Mika Westerberg wrote:
> > > ACPI Timer() opcode should return monotonically increasing clock with 100ns
> > > granularity. Implement this with the help of ktime_get().
> > > 
> > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> > 
> > That looks reasobable.  Have you tested it?
> 
> Very lightly. Basically I added some debug printks() between two
> successsive calls of Timer() and it seemed like it returned correct time.
> 
> It is certainly better than returning t+1 every time Timer() is called :)

I did somewhat better test for this. I added following ASL code:

	...
	Store(Timer, Local1)
	Sleep(10)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

	Store(Timer, Local1)
	Sleep(200)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

	Store(Timer, Local1)
	Sleep(1300)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

The second sleep should be pretty close to the first one.

Without this patch I get:

[   11.488100] ACPI: acpi_os_get_timer() TBD
[   11.492150] ACPI: Sleep(10)
[   11.502993] ACPI: Sleep(0)
[   11.506315] ACPI: Sleep(200)
[   11.706237] ACPI: Sleep(0)
[   11.709550] ACPI: Sleep(1300)
[   13.008929] ACPI: Sleep(0)

With the patch applied I get:

[   11.486786] ACPI: Sleep(10)
[   11.499029] ACPI: Sleep(12)
[   11.512350] ACPI: Sleep(200)
[   11.712282] ACPI: Sleep(200)
[   11.912170] ACPI: Sleep(1300)
[   13.211577] ACPI: Sleep(1300)

The above looks much more correct to me.

Re: [PATCH] ACPI: implement acpi_os_get_timer() according the spec

[Rafael J. Wysocki]

On Monday, May 20, 2013 01:25:30 PM Mika Westerberg wrote:
> On Mon, May 13, 2013 at 02:44:32PM +0300, Mika Westerberg wrote:
> > On Mon, May 13, 2013 at 01:38:46PM +0200, Rafael J. Wysocki wrote:
> > > On Monday, May 13, 2013 01:27:51 PM Mika Westerberg wrote:
> > > > ACPI Timer() opcode should return monotonically increasing clock with 100ns
> > > > granularity. Implement this with the help of ktime_get().
> > > > 
> > > > Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
> > > 
> > > That looks reasobable.  Have you tested it?
> > 
> > Very lightly. Basically I added some debug printks() between two
> > successsive calls of Timer() and it seemed like it returned correct time.
> > 
> > It is certainly better than returning t+1 every time Timer() is called :)
> 
> I did somewhat better test for this. I added following ASL code:
> 
> 	...
> 	Store(Timer, Local1)
> 	Sleep(10)
> 	Divide(Subtract(Timer, Local1), 10000,, Local1)
> 	Sleep(Local1)
> 
> 	Store(Timer, Local1)
> 	Sleep(200)
> 	Divide(Subtract(Timer, Local1), 10000,, Local1)
> 	Sleep(Local1)
> 
> 	Store(Timer, Local1)
> 	Sleep(1300)
> 	Divide(Subtract(Timer, Local1), 10000,, Local1)
> 	Sleep(Local1)
> 
> The second sleep should be pretty close to the first one.
> 
> Without this patch I get:
> 
> [   11.488100] ACPI: acpi_os_get_timer() TBD
> [   11.492150] ACPI: Sleep(10)
> [   11.502993] ACPI: Sleep(0)
> [   11.506315] ACPI: Sleep(200)
> [   11.706237] ACPI: Sleep(0)
> [   11.709550] ACPI: Sleep(1300)
> [   13.008929] ACPI: Sleep(0)
> 
> With the patch applied I get:
> 
> [   11.486786] ACPI: Sleep(10)
> [   11.499029] ACPI: Sleep(12)
> [   11.512350] ACPI: Sleep(200)
> [   11.712282] ACPI: Sleep(200)
> [   11.912170] ACPI: Sleep(1300)
> [   13.211577] ACPI: Sleep(1300)
> 
> The above looks much more correct to me.

That demostrates the problem nicely.  Can you please add the above info to the
patch changelog?

Rafael


-- 
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.

[PATCH v2] ACPI: implement acpi_os_get_timer() according the spec

[Mika Westerberg]

ACPI Timer() opcode should return monotonically increasing clock with 100ns
granularity according the ACPI 5.0 spec.

Testing the current Timer() implementation with following ASL code (and an
additional debug print in acpi_os_sleep() to get the sleep times dumped out
to dmesg):

	// Test: 10ms
	Store(Timer, Local1)
	Sleep(10)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

	// Test: 200ms
	Store(Timer, Local1)
	Sleep(200)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

	// Test 1300ms
	Store(Timer, Local1)
	Sleep(1300)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

The second sleep value is calculated using Timer(). If the implementation
is good enough we should be able to get the second value pretty close to
the first.

However, the current Timer() gives pretty bad sleep times:

	[   11.488100] ACPI: acpi_os_get_timer() TBD
	[   11.492150] ACPI: Sleep(10)
	[   11.502993] ACPI: Sleep(0)
	[   11.506315] ACPI: Sleep(200)
	[   11.706237] ACPI: Sleep(0)
	[   11.709550] ACPI: Sleep(1300)
	[   13.008929] ACPI: Sleep(0)

Fix this with the help of ktime_get(). Once the fix is applied and run
against the same ASL code we get:

	[   11.486786] ACPI: Sleep(10)
	[   11.499029] ACPI: Sleep(12)
	[   11.512350] ACPI: Sleep(200)
	[   11.712282] ACPI: Sleep(200)
	[   11.912170] ACPI: Sleep(1300)
	[   13.211577] ACPI: Sleep(1300)

That is much more closer to the values we expected.

Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
---
 drivers/acpi/osl.c |   14 +-------------
 1 file changed, 1 insertion(+), 13 deletions(-)

diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c
index e721863..cdda7cf 100644
--- a/drivers/acpi/osl.c
+++ b/drivers/acpi/osl.c
@@ -835,19 +835,7 @@ void acpi_os_stall(u32 us)
  */
 u64 acpi_os_get_timer(void)
 {
-	static u64 t;
-
-#ifdef	CONFIG_HPET
-	/* TBD: use HPET if available */
-#endif
-
-#ifdef	CONFIG_X86_PM_TIMER
-	/* TBD: default to PM timer if HPET was not available */
-#endif
-	if (!t)
-		printk(KERN_ERR PREFIX "acpi_os_get_timer() TBD\n");
-
-	return ++t;
+	return ktime_to_ns(ktime_get()) / 100;
 }
 
 acpi_status acpi_os_read_port(acpi_io_address port, u32 * value, u32 width)
-- 
1.7.10.4

[PATCH v3] ACPI: implement acpi_os_get_timer() according the spec

[Mika Westerberg]

ACPI Timer() opcode should return monotonically increasing clock with 100ns
granularity according the ACPI 5.0 spec.

Testing the current Timer() implementation with following ASL code (and an
additional debug print in acpi_os_sleep() to get the sleep times dumped out
to dmesg):

	// Test: 10ms
	Store(Timer, Local1)
	Sleep(10)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

	// Test: 200ms
	Store(Timer, Local1)
	Sleep(200)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

	// Test 1300ms
	Store(Timer, Local1)
	Sleep(1300)
	Divide(Subtract(Timer, Local1), 10000,, Local1)
	Sleep(Local1)

The second sleep value is calculated using Timer(). If the implementation
is good enough we should be able to get the second value pretty close to
the first.

However, the current Timer() gives pretty bad sleep times:

	[   11.488100] ACPI: acpi_os_get_timer() TBD
	[   11.492150] ACPI: Sleep(10)
	[   11.502993] ACPI: Sleep(0)
	[   11.506315] ACPI: Sleep(200)
	[   11.706237] ACPI: Sleep(0)
	[   11.709550] ACPI: Sleep(1300)
	[   13.008929] ACPI: Sleep(0)

Fix this with the help of ktime_get(). Once the fix is applied and run
against the same ASL code we get:

	[   11.486786] ACPI: Sleep(10)
	[   11.499029] ACPI: Sleep(12)
	[   11.512350] ACPI: Sleep(200)
	[   11.712282] ACPI: Sleep(200)
	[   11.912170] ACPI: Sleep(1300)
	[   13.211577] ACPI: Sleep(1300)

That is much more closer to the values we expected.

Signed-off-by: Mika Westerberg <mika.westerberg@linux.intel.com>
---
Difference to the previous version is that now we use do_div() instead so
that i386 build won't break. This was found by kbuild test robot.

In addition I tested this on both i386 and x86_64.

 drivers/acpi/osl.c |   16 +++-------------
 1 file changed, 3 insertions(+), 13 deletions(-)

diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c
index e721863..c290769 100644
--- a/drivers/acpi/osl.c
+++ b/drivers/acpi/osl.c
@@ -835,19 +835,9 @@ void acpi_os_stall(u32 us)
  */
 u64 acpi_os_get_timer(void)
 {
-	static u64 t;
-
-#ifdef	CONFIG_HPET
-	/* TBD: use HPET if available */
-#endif
-
-#ifdef	CONFIG_X86_PM_TIMER
-	/* TBD: default to PM timer if HPET was not available */
-#endif
-	if (!t)
-		printk(KERN_ERR PREFIX "acpi_os_get_timer() TBD\n");
-
-	return ++t;
+	u64 time_ns = ktime_to_ns(ktime_get());
+	do_div(time_ns, 100);
+	return time_ns;
 }
 
 acpi_status acpi_os_read_port(acpi_io_address port, u32 * value, u32 width)
-- 
1.7.10.4