diff mbox

[RFC,2/2] clocksource: dw_apb: allow build for architectures other than arm

Message ID 1369570367-994-2-git-send-email-baruch@tkos.co.il (mailing list archive)
State New, archived
Headers show

Commit Message

Baruch Siach May 26, 2013, 12:12 p.m. UTC
ARM is the only architecture providing sched_clock.h and setup_sched_clock().
Implement sched_clock() for use by other architectures.

Cc: Ingo Molnar <mingo@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Jamie Iles <jamie@jamieiles.com>
Cc: Dinh Nguyen <dinguyen@altera.com>
Signed-off-by: Baruch Siach <baruch@tkos.co.il>
---
 arch/arm/Kconfig                      |    1 +
 drivers/clocksource/Kconfig           |    3 +++
 drivers/clocksource/dw_apb_timer_of.c |   18 ++++++++++++++++++
 3 files changed, 22 insertions(+)

Comments

John Stultz May 28, 2013, 8:24 p.m. UTC | #1
On 05/26/2013 05:12 AM, Baruch Siach wrote:
> ARM is the only architecture providing sched_clock.h and setup_sched_clock().
> Implement sched_clock() for use by other architectures.
>
> Cc: Ingo Molnar <mingo@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: John Stultz <john.stultz@linaro.org>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Jamie Iles <jamie@jamieiles.com>
> Cc: Dinh Nguyen <dinguyen@altera.com>
> Signed-off-by: Baruch Siach <baruch@tkos.co.il>
> ---
>   arch/arm/Kconfig                      |    1 +
>   drivers/clocksource/Kconfig           |    3 +++
>   drivers/clocksource/dw_apb_timer_of.c |   18 ++++++++++++++++++
>   3 files changed, 22 insertions(+)
>
> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
> index 1cacda4..7489604 100644
> --- a/arch/arm/Kconfig
> +++ b/arch/arm/Kconfig
> @@ -47,6 +47,7 @@ config ARM
>   	select HAVE_OPROFILE if (HAVE_PERF_EVENTS)
>   	select HAVE_PERF_EVENTS
>   	select HAVE_REGS_AND_STACK_ACCESS_API
> +	select HAVE_SETUP_SCHED_CLOCK
>   	select HAVE_SYSCALL_TRACEPOINTS
>   	select HAVE_UID16
>   	select KTIME_SCALAR
> diff --git a/drivers/clocksource/Kconfig b/drivers/clocksource/Kconfig
> index e507ab7..0fd6d13 100644
> --- a/drivers/clocksource/Kconfig
> +++ b/drivers/clocksource/Kconfig
> @@ -1,3 +1,6 @@
> +config HAVE_SETUP_SCHED_CLOCK
> +	bool
> +
>   config CLKSRC_OF
>   	bool
>   
> diff --git a/drivers/clocksource/dw_apb_timer_of.c b/drivers/clocksource/dw_apb_timer_of.c
> index 5a61ebc..9fdb4d3 100644
> --- a/drivers/clocksource/dw_apb_timer_of.c
> +++ b/drivers/clocksource/dw_apb_timer_of.c
> @@ -21,7 +21,11 @@
>   #include <linux/of_address.h>
>   #include <linux/of_irq.h>
>   
> +#ifdef CONFIG_HAVE_SETUP_SCHED_CLOCK
>   #include <asm/sched_clock.h>
> +#else
> +#include <linux/sched.h>
> +#endif
>   
>   static void timer_get_base_and_rate(struct device_node *np,
>   				    void __iomem **base, u32 *rate)
> @@ -73,6 +77,9 @@ static void add_clocksource(struct device_node *source_timer)
>   }
>   
>   static void __iomem *sched_io_base;
> +#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> +static u64 sched_clock_mult __read_mostly;
> +#endif
>   
>   static u32 read_sched_clock(void)
>   {
> @@ -97,7 +104,11 @@ static void init_sched_clock(void)
>   	timer_get_base_and_rate(sched_timer, &sched_io_base, &rate);
>   	of_node_put(sched_timer);
>   
> +#ifdef CONFIG_HAVE_SETUP_SCHED_CLOCK
>   	setup_sched_clock(read_sched_clock, 32, rate);
> +#else
> +	sched_clock_mult = NSEC_PER_SEC / rate;
> +#endif
>   }

Can you rework this to not use #ifdefs within the function? They make it 
annoying to read the code.

Instead maybe have a local setup_sched_clock() function that sets the 
mult value for the !CONFIG_HAVE_SETUP_SCHED_CLOCK case?

>   
>   static const struct of_device_id osctimer_ids[] __initconst = {
> @@ -124,3 +135,10 @@ void __init dw_apb_timer_init(void)
>   
>   	init_sched_clock();
>   }
> +
> +#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> +unsigned long long notrace sched_clock()
> +{
> +	return read_sched_clock() * sched_clock_mult;
> +}
> +#endif

Also, can you try to condense the number of #ifndef 
CONFIG_HAVE_SETUP_SCHED_CLOCK checks to one, and consolidate the needed 
functions all in that one conditional?

thanks
-john
Baruch Siach May 30, 2013, 5:32 a.m. UTC | #2
Hi John,

On Tue, May 28, 2013 at 01:24:17PM -0700, John Stultz wrote:
> On 05/26/2013 05:12 AM, Baruch Siach wrote:
> >ARM is the only architecture providing sched_clock.h and setup_sched_clock().
> >Implement sched_clock() for use by other architectures.
> >
> >Cc: Ingo Molnar <mingo@redhat.com>
> >Cc: Peter Zijlstra <peterz@infradead.org>
> >Cc: John Stultz <john.stultz@linaro.org>
> >Cc: Thomas Gleixner <tglx@linutronix.de>
> >Cc: Jamie Iles <jamie@jamieiles.com>
> >Cc: Dinh Nguyen <dinguyen@altera.com>
> >Signed-off-by: Baruch Siach <baruch@tkos.co.il>
> >---

[snip]

> >@@ -73,6 +77,9 @@ static void add_clocksource(struct device_node 
> >  *source_timer)
> >  }
> >  static void __iomem *sched_io_base;
> >+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> >+static u64 sched_clock_mult __read_mostly;
> >+#endif
> >  static u32 read_sched_clock(void)
> >  {
> >@@ -97,7 +104,11 @@ static void init_sched_clock(void)
> >  	timer_get_base_and_rate(sched_timer, &sched_io_base, &rate);
> >  	of_node_put(sched_timer);
> >+#ifdef CONFIG_HAVE_SETUP_SCHED_CLOCK
> >  	setup_sched_clock(read_sched_clock, 32, rate);
> >+#else
> >+	sched_clock_mult = NSEC_PER_SEC / rate;
> >+#endif
> >  }
> 
> Can you rework this to not use #ifdefs within the function? They
> make it annoying to read the code.
> 
> Instead maybe have a local setup_sched_clock() function that sets
> the mult value for the !CONFIG_HAVE_SETUP_SCHED_CLOCK case?
> 
> >  static const struct of_device_id osctimer_ids[] __initconst = {
> >@@ -124,3 +135,10 @@ void __init dw_apb_timer_init(void)
> >  	init_sched_clock();
> >  }
> >+
> >+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> >+unsigned long long notrace sched_clock()
> >+{
> >+	return read_sched_clock() * sched_clock_mult;
> >+}
> >+#endif
> 
> Also, can you try to condense the number of #ifndef
> CONFIG_HAVE_SETUP_SCHED_CLOCK checks to one, and consolidate the
> needed functions all in that one conditional?

Thanks for your comments. I'll rework the patch and resubmit.

I've just noticed that I have a bigger problem. read_sched_clock() returns 
u32, not u64. This means that in a rate of, say, 100MHz it will wrap around 
after a little more than 40 seconds. Would it make sense to put ARM's 32 bin 
sched_clock extension code (arch/arm/kernel/sched_clock.c) is a common place 
(maybe drivers/clocksource), and use that? There seems to be nothing ARM 
specific in this code, after all.

baruch
John Stultz May 30, 2013, 4:11 p.m. UTC | #3
On 05/29/2013 10:32 PM, Baruch Siach wrote:
> Hi John,
>
> On Tue, May 28, 2013 at 01:24:17PM -0700, John Stultz wrote:
>> On 05/26/2013 05:12 AM, Baruch Siach wrote:
>>>   static const struct of_device_id osctimer_ids[] __initconst = {
>>> @@ -124,3 +135,10 @@ void __init dw_apb_timer_init(void)
>>>   	init_sched_clock();
>>>   }
>>> +
>>> +#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
>>> +unsigned long long notrace sched_clock()
>>> +{
>>> +	return read_sched_clock() * sched_clock_mult;
>>> +}
>>> +#endif
>> Also, can you try to condense the number of #ifndef
>> CONFIG_HAVE_SETUP_SCHED_CLOCK checks to one, and consolidate the
>> needed functions all in that one conditional?
> Thanks for your comments. I'll rework the patch and resubmit.
>
> I've just noticed that I have a bigger problem. read_sched_clock() returns
> u32, not u64. This means that in a rate of, say, 100MHz it will wrap around
> after a little more than 40 seconds. Would it make sense to put ARM's 32 bin
> sched_clock extension code (arch/arm/kernel/sched_clock.c) is a common place
> (maybe drivers/clocksource), and use that? There seems to be nothing ARM
> specific in this code, after all.

Yea, working out an actual generic sched_clock implementation is 
something I'd like to see done.

Though I'd really rather we not toss yet another chunk of infrastructure 
in the drivers/clocksource directory. Instead we should probably have a 
kernel/time/sched_clock.c.

Then its just the issue of tying that and the clocksource code together 
so you just register existing capable clocksources with a SCHED_CLOCK 
flag or via a different registration hook.

I don't think it will be an easy job, but if you want to give it a shot, 
I'd be quite interested in reviewing the patches!

thanks
-john
Baruch Siach May 31, 2013, 5:32 a.m. UTC | #4
Hi John,

On Thu, May 30, 2013 at 09:11:17AM -0700, John Stultz wrote:
> On 05/29/2013 10:32 PM, Baruch Siach wrote:
> >On Tue, May 28, 2013 at 01:24:17PM -0700, John Stultz wrote:
> >>On 05/26/2013 05:12 AM, Baruch Siach wrote:
> >>>  static const struct of_device_id osctimer_ids[] __initconst = {
> >>>@@ -124,3 +135,10 @@ void __init dw_apb_timer_init(void)
> >>>  	init_sched_clock();
> >>>  }
> >>>+
> >>>+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> >>>+unsigned long long notrace sched_clock()
> >>>+{
> >>>+	return read_sched_clock() * sched_clock_mult;
> >>>+}
> >>>+#endif
> >>Also, can you try to condense the number of #ifndef
> >>CONFIG_HAVE_SETUP_SCHED_CLOCK checks to one, and consolidate the
> >>needed functions all in that one conditional?
> >Thanks for your comments. I'll rework the patch and resubmit.
> >
> >I've just noticed that I have a bigger problem. read_sched_clock() returns
> >u32, not u64. This means that in a rate of, say, 100MHz it will wrap around
> >after a little more than 40 seconds. Would it make sense to put ARM's 32 bin
> >sched_clock extension code (arch/arm/kernel/sched_clock.c) is a common place
> >(maybe drivers/clocksource), and use that? There seems to be nothing ARM
> >specific in this code, after all.
> 
> Yea, working out an actual generic sched_clock implementation is
> something I'd like to see done.
> 
> Though I'd really rather we not toss yet another chunk of
> infrastructure in the drivers/clocksource directory. Instead we
> should probably have a kernel/time/sched_clock.c.
> 
> Then its just the issue of tying that and the clocksource code
> together so you just register existing capable clocksources with a
> SCHED_CLOCK flag or via a different registration hook.
> 
> I don't think it will be an easy job, but if you want to give it a
> shot, I'd be quite interested in reviewing the patches!

Although I'd very much like to contribute in this area, I'm afraid I don't 
have time for this now. My problem is much narrower in scope, though, and I 
think it can be solved without a generic sched_clock implementation at first 
step. All I want is to use a 32 bit counter for sched_clock. The code enabling 
this is currently only available for ARM (arch/arm/kernel/sched_clock.c). 
Should this part of the code move to kernel/time/sched_clock.c?

baruch
Pavel Machek June 20, 2013, 2:44 p.m. UTC | #5
Hi!

> > >@@ -73,6 +77,9 @@ static void add_clocksource(struct device_node 
> > >  *source_timer)
> > >  }
> > >  static void __iomem *sched_io_base;
> > >+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> > >+static u64 sched_clock_mult __read_mostly;
> > >+#endif
> > >  static u32 read_sched_clock(void)
> > >  {
> > >@@ -97,7 +104,11 @@ static void init_sched_clock(void)
> > >  	timer_get_base_and_rate(sched_timer, &sched_io_base, &rate);
> > >  	of_node_put(sched_timer);
> > >+#ifdef CONFIG_HAVE_SETUP_SCHED_CLOCK
> > >  	setup_sched_clock(read_sched_clock, 32, rate);
> > >+#else
> > >+	sched_clock_mult = NSEC_PER_SEC / rate;
> > >+#endif
> > >  }
> > 
> > Can you rework this to not use #ifdefs within the function? They
> > make it annoying to read the code.
> > 
> > Instead maybe have a local setup_sched_clock() function that sets
> > the mult value for the !CONFIG_HAVE_SETUP_SCHED_CLOCK case?
> > 
> > >  static const struct of_device_id osctimer_ids[] __initconst = {
> > >@@ -124,3 +135,10 @@ void __init dw_apb_timer_init(void)
> > >  	init_sched_clock();
> > >  }
> > >+
> > >+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> > >+unsigned long long notrace sched_clock()
> > >+{
> > >+	return read_sched_clock() * sched_clock_mult;
> > >+}
> > >+#endif
> > 
> > Also, can you try to condense the number of #ifndef
> > CONFIG_HAVE_SETUP_SCHED_CLOCK checks to one, and consolidate the
> > needed functions all in that one conditional?
> 
> Thanks for your comments. I'll rework the patch and resubmit.
> 
> I've just noticed that I have a bigger problem. read_sched_clock() returns 
> u32, not u64. This means that in a rate of, say, 100MHz it will wrap around 
> after a little more than 40 seconds. Would it make sense to put ARM's 32 bin 
> sched_clock extension code (arch/arm/kernel/sched_clock.c) is a common place 
> (maybe drivers/clocksource), and use that? There seems to be nothing ARM 
> specific in this code, after all.

Also note that there are two conflicting changes to this area
pending. It seems one in soc-next arm tree will prevail.

Please take a look...
									Pavel
Baruch Siach June 20, 2013, 3:03 p.m. UTC | #6
Hi Pavel,

On Thu, Jun 20, 2013 at 04:44:33PM +0200, Pavel Machek wrote:
> > > >@@ -73,6 +77,9 @@ static void add_clocksource(struct device_node 
> > > >  *source_timer)
> > > >  }
> > > >  static void __iomem *sched_io_base;
> > > >+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> > > >+static u64 sched_clock_mult __read_mostly;
> > > >+#endif
> > > >  static u32 read_sched_clock(void)
> > > >  {
> > > >@@ -97,7 +104,11 @@ static void init_sched_clock(void)
> > > >  	timer_get_base_and_rate(sched_timer, &sched_io_base, &rate);
> > > >  	of_node_put(sched_timer);
> > > >+#ifdef CONFIG_HAVE_SETUP_SCHED_CLOCK
> > > >  	setup_sched_clock(read_sched_clock, 32, rate);
> > > >+#else
> > > >+	sched_clock_mult = NSEC_PER_SEC / rate;
> > > >+#endif
> > > >  }
> > > 
> > > Can you rework this to not use #ifdefs within the function? They
> > > make it annoying to read the code.
> > > 
> > > Instead maybe have a local setup_sched_clock() function that sets
> > > the mult value for the !CONFIG_HAVE_SETUP_SCHED_CLOCK case?
> > > 
> > > >  static const struct of_device_id osctimer_ids[] __initconst = {
> > > >@@ -124,3 +135,10 @@ void __init dw_apb_timer_init(void)
> > > >  	init_sched_clock();
> > > >  }
> > > >+
> > > >+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
> > > >+unsigned long long notrace sched_clock()
> > > >+{
> > > >+	return read_sched_clock() * sched_clock_mult;
> > > >+}
> > > >+#endif
> > > 
> > > Also, can you try to condense the number of #ifndef
> > > CONFIG_HAVE_SETUP_SCHED_CLOCK checks to one, and consolidate the
> > > needed functions all in that one conditional?
> > 
> > Thanks for your comments. I'll rework the patch and resubmit.
> > 
> > I've just noticed that I have a bigger problem. read_sched_clock() returns 
> > u32, not u64. This means that in a rate of, say, 100MHz it will wrap around 
> > after a little more than 40 seconds. Would it make sense to put ARM's 32 bin 
> > sched_clock extension code (arch/arm/kernel/sched_clock.c) is a common place 
> > (maybe drivers/clocksource), and use that? There seems to be nothing ARM 
> > specific in this code, after all.
> 
> Also note that there are two conflicting changes to this area
> pending. It seems one in soc-next arm tree will prevail.
> 
> Please take a look...

Now that generic sched_clock implementation is available to all architectures 
(38ff87f77 "sched_clock: Make ARM's sched_clock generic for all architectures" 
in the timers/core branch of the tip tree) this patch is not needed anymore.

Thanks for the head up.

baruch
diff mbox

Patch

diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index 1cacda4..7489604 100644
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig
@@ -47,6 +47,7 @@  config ARM
 	select HAVE_OPROFILE if (HAVE_PERF_EVENTS)
 	select HAVE_PERF_EVENTS
 	select HAVE_REGS_AND_STACK_ACCESS_API
+	select HAVE_SETUP_SCHED_CLOCK
 	select HAVE_SYSCALL_TRACEPOINTS
 	select HAVE_UID16
 	select KTIME_SCALAR
diff --git a/drivers/clocksource/Kconfig b/drivers/clocksource/Kconfig
index e507ab7..0fd6d13 100644
--- a/drivers/clocksource/Kconfig
+++ b/drivers/clocksource/Kconfig
@@ -1,3 +1,6 @@ 
+config HAVE_SETUP_SCHED_CLOCK
+	bool
+
 config CLKSRC_OF
 	bool
 
diff --git a/drivers/clocksource/dw_apb_timer_of.c b/drivers/clocksource/dw_apb_timer_of.c
index 5a61ebc..9fdb4d3 100644
--- a/drivers/clocksource/dw_apb_timer_of.c
+++ b/drivers/clocksource/dw_apb_timer_of.c
@@ -21,7 +21,11 @@ 
 #include <linux/of_address.h>
 #include <linux/of_irq.h>
 
+#ifdef CONFIG_HAVE_SETUP_SCHED_CLOCK
 #include <asm/sched_clock.h>
+#else
+#include <linux/sched.h>
+#endif
 
 static void timer_get_base_and_rate(struct device_node *np,
 				    void __iomem **base, u32 *rate)
@@ -73,6 +77,9 @@  static void add_clocksource(struct device_node *source_timer)
 }
 
 static void __iomem *sched_io_base;
+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
+static u64 sched_clock_mult __read_mostly;
+#endif
 
 static u32 read_sched_clock(void)
 {
@@ -97,7 +104,11 @@  static void init_sched_clock(void)
 	timer_get_base_and_rate(sched_timer, &sched_io_base, &rate);
 	of_node_put(sched_timer);
 
+#ifdef CONFIG_HAVE_SETUP_SCHED_CLOCK
 	setup_sched_clock(read_sched_clock, 32, rate);
+#else
+	sched_clock_mult = NSEC_PER_SEC / rate;
+#endif
 }
 
 static const struct of_device_id osctimer_ids[] __initconst = {
@@ -124,3 +135,10 @@  void __init dw_apb_timer_init(void)
 
 	init_sched_clock();
 }
+
+#ifndef CONFIG_HAVE_SETUP_SCHED_CLOCK
+unsigned long long notrace sched_clock()
+{
+	return read_sched_clock() * sched_clock_mult;
+}
+#endif