Skip to contentSkip to conversation

atmel_tc clocksource/clockevent code

PATCH In review David Brownell 2 participants 7 messages Feb 23, 2008 – Feb 26, 2008

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 2/2 atmel_tc clocksource/clockevent code

Clocksource and clockevent device based on the Atmel TC blocks.

The clockevent device handles both periodic and oneshot modes, so this enables NO_HZ and high res timers on some platforms that previously couldn't use those mechanisms.

This works on both AVR32 and AT91 chips, given relevant patches for tclib support (always) and clockevents (or else this will only look like a higher precision clocksource). It's an updated and modularized version of an AT91-only patch that has circulated for some time now.

---
As with the preceding patch, this depends on platform updates that won't merge before 2.6.26 ... in this case, updates to switch over to clocksources (at91sam9 chips, but not at91rm9200) and clockevents (avr32 and at91sam9), on top of platform_device setup.

 drivers/clocksource/Makefile     |    1 
 drivers/clocksource/tcb_clksrc.c |  303 +++++++++++++++++++++++++++++++++++++++
 drivers/misc/Kconfig             |   25 +++
 3 files changed, 329 insertions(+)
drivers/clocksource/Makefile +1 -0
@@ -1,3 +1,4 @@
+obj-$(CONFIG_ATMEL_TCB_CLKSRC)	+= tcb_clksrc.o
 obj-$(CONFIG_X86_CYCLONE_TIMER)	+= cyclone.o
 obj-$(CONFIG_X86_PM_TIMER)	+= acpi_pm.o
 obj-$(CONFIG_SCx200HR_TIMER)	+= scx200_hrt.o
drivers/clocksource/tcb_clksrc.c new file +303 -0

303 lines not on this page · raw message

drivers/misc/Kconfig +25 -0
@@ -30,6 +30,31 @@ config ATMEL_TCLIB
 	  blocks found on many Atmel processors.  This facilitates using
 	  these blocks by different drivers despite processor differences.
 
+config ATMEL_TCB_CLKSRC
+	bool "TC Block Clocksource"
+	depends on ATMEL_TCLIB && GENERIC_TIME
+	default y
+	help
+	  Select this to get a high precision clocksource based on a
+	  TC block with a 5+ MHz base clock rate.  Two timer channels
+	  are combined to make a single 32-bit timer.
+
+	  When GENERIC_CLOCKEVENTS is defined, the third timer channel
+	  may be used as a clock event device supporting oneshot mode
+	  (delays of up to two seconds) based on the 32 KiHz clock.
+
+config ATMEL_TCB_CLKSRC_BLOCK
+	int
+	depends on ATMEL_TCB_CLKSRC
+	prompt "TC Block" if ARCH_AT91RM9200 || ARCH_AT91SAM9260 || CPU_AT32AP700X
+	default 0
+	range 0 1
+	help
+	  Some chips provide more than one TC block, so you have the
+	  choice of which one to use for the clock framework.  The other
+	  TC can be used for other purposes, such as PWM generation and
+	  interval timing.
+
 config IBM_ASM
 	tristate "Device driver for IBM RSA service processor"
 	depends on X86 && PCI && INPUT && EXPERIMENTAL
#1 lore raw

Again, sorry for the delay...it really sucks that I haven't been able to look at this stuff closely until now. Hopefully a late review is better than no review.

drivers/clocksource/tcb_clksrc.c line 47

Quoted 6 lines · On Fri, 22 Feb 2008 17:28:37 -0800 David Brownell <[email protected]> wrote:
drivers/clocksource/tcb_clksrc.c new file +6 -0
+static cycle_t tc_get_cycles(void)
+{
+	unsigned long	flags;
+	u32		lower, upper;
+
+	raw_local_irq_save(flags);

Why do you need to use the raw version?

drivers/clocksource/tcb_clksrc.c line 52

Quoted 5 lines
drivers/clocksource/tcb_clksrc.c new file +5 -0
+retry:
+	upper = __raw_readl(tcaddr + ATMEL_TC_REG(1, CV));
+	lower = __raw_readl(tcaddr + ATMEL_TC_REG(0, CV));
+	if (upper != __raw_readl(tcaddr + ATMEL_TC_REG(1, CV)))
+		goto retry;

Did you just open-code a do/while loop using goto? ;-)

drivers/clocksource/tcb_clksrc.c line 79

Quoted 13 lines
drivers/clocksource/tcb_clksrc.c new file +13 -0
+#ifdef CONFIG_GENERIC_CLOCKEVENTS
+#define USE_TC_CLKEVT
+#endif
+
+#ifdef CONFIG_ARCH_AT91RM9200
+/* The AT91rm9200 system timer is a oneshot-capable 32k timer that's
+ * always running ... configuring a TC channel to work the same way
+ * would just waste some power.
+ */
+#undef USE_TC_CLKEVT
+#endif
+
+#ifdef USE_TC_CLKEVT

Can't you just use #ifndef CONFIG_ARCH_AT91RM9200 and avoid the whole ifdef/define/undef dance above?

drivers/clocksource/tcb_clksrc.c line 116

Quoted 2 lines
drivers/clocksource/tcb_clksrc.c new file +2 -0
+	case CLOCK_EVT_MODE_ONESHOT:
+		/* slow clock, count up to RC, then irq and stop */

Hmm. Do you really want to stop it? Won't you get better accuracy if you let it run and program the next event as (prev_event + delta)?

drivers/clocksource/tcb_clksrc.c line 166

Quoted 5 lines
drivers/clocksource/tcb_clksrc.c new file +5 -0
+static struct irqaction tc_irqaction = {
+	.name		= "tc_clkevt",
+	.flags		= IRQF_TIMER | IRQF_DISABLED,
+	.handler	= ch2_irq,
+};

I don't think you need to define this statically. You can call request_irq() at arch_initcall() time.

Quoted 6 lines
drivers/clocksource/tcb_clksrc.c new file +6 -0
+static void __init setup_clkevents(struct platform_device *pdev,
+		struct clk *t0_clk, int clk32k_divisor_idx)
+{
+	struct clk	*t2_clk = clk_get(&pdev->dev, "t2_clk");
+	int		irq;
+

drivers/clocksource/tcb_clksrc.c line 183

Quoted 10 lines
drivers/clocksource/tcb_clksrc.c new file +10 -0
+	/* SOCs have either one IRQ and one clock for the whole
+	 * TC block; or one each per channel.  We use channel 2.
+	 */
+	if (IS_ERR(t2_clk)) {
+		t2_clk = t0_clk;
+		irq = platform_get_irq(pdev, 0);
+	} else {
+		irq = platform_get_irq(pdev, 2);
+		clk_enable(t2_clk);
+	}

I don't think it is safe to assume that one clock per channel always means one irq per channel as well...

What's wrong with

	irq = platform_get_irq(pdev, 2);
	if (irq < 0)
		irq = platform_get_irq(pdev, 0);

?

Quoted 9 lines
drivers/misc/Kconfig +9 -0
+config ATMEL_TCB_CLKSRC
+	bool "TC Block Clocksource"
+	depends on ATMEL_TCLIB && GENERIC_TIME
+	default y
+	help
+	  Select this to get a high precision clocksource based on a
+	  TC block with a 5+ MHz base clock rate.  Two timer channels
+	  are combined to make a single 32-bit timer.
+

drivers/misc/Kconfig config ATMEL_TCLIB · line 51

Quoted 10 lines
drivers/misc/Kconfig +10 -0
+	  When GENERIC_CLOCKEVENTS is defined, the third timer channel
+	  may be used as a clock event device supporting oneshot mode
+	  (delays of up to two seconds) based on the 32 KiHz clock.
+
+config ATMEL_TCB_CLKSRC_BLOCK
+	int
+	depends on ATMEL_TCB_CLKSRC
+	prompt "TC Block" if ARCH_AT91RM9200 || ARCH_AT91SAM9260 || CPU_AT32AP700X
+	default 0
+	range 0 1

Hmm...I don't like the direction this is going...

How about we make tcb_clksrc_init() global and call it from the platform code with whatever TCB the platform thinks is appropriate?

Perhaps remove the prompt from ATMEL_TCB_CLKSRC and have the platform select it as well? I certainly want to force this stuff on on the AP7000...otherwise we'll just get lots of complaints that the thing is using 4x more power than it's supposed to...

Haavard

#2 lore raw
Quoted 3 lines

Again, sorry for the delay...it really sucks that I haven't been able to look at this stuff closely until now. Hopefully a late review is better than no review.

Yes. The review I've gotten so far has been of the "it works for me, thanks" variety.

(The only issue reported to me is that NO_HZ on AT91 seems to make the DBGU lose characters sometimes ... that one-byte "fifo" makes trouble whenever there's the slightest delay in IRQ handling, and NO_HZ can easily add such delays as part of ensuring that "jiffies" is current before invoking handlers.)

Quoted 11 lines

drivers/clocksource/tcb_clksrc.c line 47

Quoted 6 lines · On Fri, 22 Feb 2008 17:28:37 -0800 David Brownell <[email protected]> wrote:
drivers/clocksource/tcb_clksrc.c new file +6 -0
+static cycle_t tc_get_cycles(void)
+{
+	unsigned long	flags;
+	u32		lower, upper;
+
+	raw_local_irq_save(flags);

Why do you need to use the raw version?

This is part of the system timer code, and it should never be a preemption point. Plus I didn't want to waste any instruction cycles in code that runs before e.g. the DBGU IRQ handler would get called... observably, such extra cycles *do* hurt.

Quoted 7 lines

drivers/clocksource/tcb_clksrc.c line 52

Quoted 5 lines
drivers/clocksource/tcb_clksrc.c new file +5 -0
+retry:
+	upper = __raw_readl(tcaddr + ATMEL_TC_REG(1, CV));
+	lower = __raw_readl(tcaddr + ATMEL_TC_REG(0, CV));
+	if (upper != __raw_readl(tcaddr + ATMEL_TC_REG(1, CV)))
+		goto retry;

Did you just open-code a do/while loop using goto? ;-)

I take that back about late review being better than none. ;)

Quoted 8 lines
Quoted 8 lines
drivers/clocksource/tcb_clksrc.c new file +8 -0
+#ifdef CONFIG_GENERIC_CLOCKEVENTS
+#define USE_TC_CLKEVT
+#endif
+
+#ifdef CONFIG_ARCH_AT91RM9200
+/* The AT91rm9200 system timer is a oneshot-capable 32k timer that's
+ * always running ... configuring a TC channel to work the same way
+ * would just waste some power.
Quoted 8 lines

drivers/clocksource/tcb_clksrc.c line 79

Quoted 5 lines

+ */

drivers/clocksource/tcb_clksrc.c new file +4 -0
+#undef USE_TC_CLKEVT
+#endif
+
+#ifdef USE_TC_CLKEVT

Can't you just use #ifndef CONFIG_ARCH_AT91RM9200 and avoid the whole ifdef/define/undef dance above?

I can't know that some other platform won't have a better system timer solution, and didn't want to assume that only that single venerable chip had such a solution ... it's kind of puzzling to me (software guy) that none of the newest Atmel SOCs have better timer infrastructure than they do. ;)

Quoted 5 lines

drivers/clocksource/tcb_clksrc.c line 116

Quoted 2 lines
drivers/clocksource/tcb_clksrc.c new file +2 -0
+	case CLOCK_EVT_MODE_ONESHOT:
+		/* slow clock, count up to RC, then irq and stop */

Hmm. Do you really want to stop it? Won't you get better accuracy if you let it run and program the next event as (prev_event + delta)?

No, ONESHOT means "stop after the IRQ I asked for". And when tc_next_event() is asked to trigger that IRQ, it's relative to the instant it's requested, not relative to the last one that was requested (which may not have triggered yet, or may have been quite some time ago).

Quoted 8 lines

drivers/clocksource/tcb_clksrc.c line 166

Quoted 5 lines
drivers/clocksource/tcb_clksrc.c new file +5 -0
+static struct irqaction tc_irqaction = {
+	.name		= "tc_clkevt",
+	.flags		= IRQF_TIMER | IRQF_DISABLED,
+	.handler	= ch2_irq,
+};

I don't think you need to define this statically. You can call request_irq() at arch_initcall() time.

That could be done, yes ... and using request_irq()/free_irq() would also let this be linked as a module, not just statically.

On the other hand, doing it this way doesn't hurt either does it? Unless a modular build is important.

Quoted 11 lines
Quoted 11 lines
drivers/clocksource/tcb_clksrc.c new file +11 -0
+static void __init setup_clkevents(struct platform_device *pdev,
+		struct clk *t0_clk, int clk32k_divisor_idx)
+{
+	struct clk	*t2_clk = clk_get(&pdev->dev, "t2_clk");
+	int		irq;
+
+	/* SOCs have either one IRQ and one clock for the whole
+	 * TC block; or one each per channel.  We use channel 2.
+	 */
+	if (IS_ERR(t2_clk)) {
+		t2_clk = t0_clk;
Quoted 8 lines

drivers/clocksource/tcb_clksrc.c line 183

Quoted 5 lines
drivers/clocksource/tcb_clksrc.c new file +5 -0
+		irq = platform_get_irq(pdev, 0);
+	} else {
+		irq = platform_get_irq(pdev, 2);
+		clk_enable(t2_clk);
+	}

I don't think it is safe to assume that one clock per channel always means one irq per channel as well...

On current chips, that's how it works.

Quoted 5 lines

What's wrong with

	irq = platform_get_irq(pdev, 2);
	if (irq < 0)
		irq = platform_get_irq(pdev, 0);

Nothing much, except that I took the more concise path. Got patch?

Quoted 22 lines
Quoted 9 lines
drivers/misc/Kconfig +9 -0
+config ATMEL_TCB_CLKSRC
+	bool "TC Block Clocksource"
+	depends on ATMEL_TCLIB && GENERIC_TIME
+	default y
+	help
+	  Select this to get a high precision clocksource based on a
+	  TC block with a 5+ MHz base clock rate.  Two timer channels
+	  are combined to make a single 32-bit timer.
+

drivers/misc/Kconfig config ATMEL_TCLIB · line 51

Quoted 10 lines
drivers/misc/Kconfig +10 -0
+	  When GENERIC_CLOCKEVENTS is defined, the third timer channel
+	  may be used as a clock event device supporting oneshot mode
+	  (delays of up to two seconds) based on the 32 KiHz clock.
+
+config ATMEL_TCB_CLKSRC_BLOCK
+	int
+	depends on ATMEL_TCB_CLKSRC
+	prompt "TC Block" if ARCH_AT91RM9200 || ARCH_AT91SAM9260 || CPU_AT32AP700X
+	default 0
+	range 0 1

Hmm...I don't like the direction this is going...

Quoted 2 lines

How about we make tcb_clksrc_init() global and call it from the platform code with whatever TCB the platform thinks is appropriate?

That could work too, but if it goes that way then there's no point in changes to support a modular build (e.g. the irqaction thing you noted above).

Quoted 4 lines

Perhaps remove the prompt from ATMEL_TCB_CLKSRC and have the platform select it as well? I certainly want to force this stuff on on the AP7000...otherwise we'll just get lots of complaints that the thing is using 4x more power than it's supposed to...

Well, "force" seems the wrong approach to me. That should be a board-specific choice. The ap700x power budget may not be the most important system design consideration, so making such a decision at the platform ("ap700x") level seems wrong.

Suppose someone has to build an AVR32 based system that uses both TCB modules to hook up to external hardware, so that neither one is available for use as a "system timer"?

- Dave

#3 lore raw
Quoted 12 lines · On Sun, 24 Feb 2008 15:42:51 -0800 David Brownell <[email protected]> wrote:
Quoted 11 lines

drivers/clocksource/tcb_clksrc.c line 47

Quoted 6 lines · On Fri, 22 Feb 2008 17:28:37 -0800 David Brownell <[email protected]> wrote:
drivers/clocksource/tcb_clksrc.c new file +6 -0
+static cycle_t tc_get_cycles(void)
+{
+	unsigned long	flags;
+	u32		lower, upper;
+
+	raw_local_irq_save(flags);

Why do you need to use the raw version?

Quoted 4 lines

This is part of the system timer code, and it should never be a preemption point. Plus I didn't want to waste any instruction cycles in code that runs before e.g. the DBGU IRQ handler would get called... observably, such extra cycles *do* hurt.

I don't understand what you mean by preemption point, but I guess the non-raw version may consume some extra cycles when lockdep is enabled.

Quoted 17 lines
Quoted 8 lines
Quoted 8 lines
drivers/clocksource/tcb_clksrc.c new file +8 -0
+#ifdef CONFIG_GENERIC_CLOCKEVENTS
+#define USE_TC_CLKEVT
+#endif
+
+#ifdef CONFIG_ARCH_AT91RM9200
+/* The AT91rm9200 system timer is a oneshot-capable 32k timer that's
+ * always running ... configuring a TC channel to work the same way
+ * would just waste some power.
Quoted 8 lines

drivers/clocksource/tcb_clksrc.c line 79

Quoted 5 lines

+ */

drivers/clocksource/tcb_clksrc.c new file +4 -0
+#undef USE_TC_CLKEVT
+#endif
+
+#ifdef USE_TC_CLKEVT

Can't you just use #ifndef CONFIG_ARCH_AT91RM9200 and avoid the whole ifdef/define/undef dance above?

Quoted 5 lines

I can't know that some other platform won't have a better system timer solution, and didn't want to assume that only that single venerable chip had such a solution ... it's kind of puzzling to me (software guy) that none of the newest Atmel SOCs have better timer infrastructure than they do. ;)

Heh.

If we really expect using TC as a clocksource but not as a clockevent is going to be a common usage, perhaps we should move the decision into Kconfig?

Besides, I don't really see the difference between having a big #ifdef expression around the whole clockevent block and having a big #ifdef expression around the definition of USE_TC_CLKEVT except that the latter is more code...

Quoted 11 lines
Quoted 5 lines

drivers/clocksource/tcb_clksrc.c line 116

Quoted 2 lines
drivers/clocksource/tcb_clksrc.c new file +2 -0
+	case CLOCK_EVT_MODE_ONESHOT:
+		/* slow clock, count up to RC, then irq and stop */

Hmm. Do you really want to stop it? Won't you get better accuracy if you let it run and program the next event as (prev_event + delta)?

No, ONESHOT means "stop after the IRQ I asked for". And when tc_next_event() is asked to trigger that IRQ, it's relative to the instant it's requested, not relative to the last one that was requested (which may not have triggered yet, or may have been quite some time ago).

Right. Sounds like it might introduce some inaccuracy, but I guess it's the clocksource's job to keep track of the actual time at the time of the event.

Quoted 9 lines
Quoted 8 lines

drivers/clocksource/tcb_clksrc.c line 166

Quoted 5 lines
drivers/clocksource/tcb_clksrc.c new file +5 -0
+static struct irqaction tc_irqaction = {
+	.name		= "tc_clkevt",
+	.flags		= IRQF_TIMER | IRQF_DISABLED,
+	.handler	= ch2_irq,
+};

I don't think you need to define this statically. You can call request_irq() at arch_initcall() time.

Quoted 5 lines

That could be done, yes ... and using request_irq()/free_irq() would also let this be linked as a module, not just statically.

On the other hand, doing it this way doesn't hurt either does it? Unless a modular build is important.

No, it doesn't hurt. Maybe we should keep it static so that we have the option of initializing it earlier if we need to.

Quoted 4 lines
Quoted 2 lines

I don't think it is safe to assume that one clock per channel always means one irq per channel as well...

On current chips, that's how it works.

Indeed. I just don't see any fundamental reason why it has to work that way, which is why I don't think we should depend on it.

Quoted 7 lines
Quoted 5 lines

What's wrong with

	irq = platform_get_irq(pdev, 2);
	if (irq < 0)
		irq = platform_get_irq(pdev, 0);

Nothing much, except that I took the more concise path. Got patch?

Not until we reach a conclusion about the tclib core.

Quoted 6 lines
Quoted 2 lines

How about we make tcb_clksrc_init() global and call it from the platform code with whatever TCB the platform thinks is appropriate?

That could work too, but if it goes that way then there's no point in changes to support a modular build (e.g. the irqaction thing you noted above).

True.

Quoted 5 lines
Quoted 4 lines

Perhaps remove the prompt from ATMEL_TCB_CLKSRC and have the platform select it as well? I certainly want to force this stuff on on the AP7000...otherwise we'll just get lots of complaints that the thing is using 4x more power than it's supposed to...

Quoted 8 lines

Well, "force" seems the wrong approach to me. That should be a board-specific choice. The ap700x power budget may not be the most important system design consideration, so making such a decision at the platform ("ap700x") level seems wrong.

Suppose someone has to build an AVR32 based system that uses both TCB modules to hook up to external hardware, so that neither one is available for use as a "system timer"?

Good point. Let's keep it as it is and let the board "select" it through its defconfig.

Haavard

#4 lore raw
Quoted 9 lines
Quoted 9 lines
Quoted 8 lines

drivers/clocksource/tcb_clksrc.c line 47

Quoted 6 lines
drivers/clocksource/tcb_clksrc.c new file +6 -0
+static cycle_t tc_get_cycles(void)
+{
+	unsigned long	flags;
+	u32		lower, upper;
+
+	raw_local_irq_save(flags);

Why do you need to use the raw version?

Quoted 7 lines
Quoted 4 lines

This is part of the system timer code, and it should never be a preemption point. Plus I didn't want to waste any instruction cycles in code that runs before e.g. the DBGU IRQ handler would get called... observably, such extra cycles *do* hurt.

I don't understand what you mean by preemption point, but I guess the non-raw version may consume some extra cycles when lockdep is enabled.

A preemption point is where CONFIG_PREEMPT kicks in task switching logic; lockdep is different.

Quoted 3 lines

If we really expect using TC as a clocksource but not as a clockevent is going to be a common usage, perhaps we should move the decision into Kconfig?

Maybe, but I already spent lots more time on this than I wanted. :(

Another way to address that rm9200 issue would be to just rate the TC clockevent source lower than the one based on the system timer, so it's set up but never enabled ... and remember "t2_clk", calling clk_enable() only when that clockevent device is active.

That would address one half of the suspend/resume equation too, ensuring that clock is disabled during suspend...

The other half being: how to clk_disable(t0_clk) during system suspend? (And t1_clk on some systems.) There's currently no clocksource.set_mode() call; evidently there's an assumption that such counters cost no power, so they can always be left running.

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

I don't think it is safe to assume that one clock per channel always means one irq per channel as well...

On current chips, that's how it works.

Indeed. I just don't see any fundamental reason why it has to work that way, which is why I don't think we should depend on it.

AT91 chips share identifiers between clocks and interrupts; that's fundamental, yes?

If some future chip works differently, that's a good time to change things. Otherwise I see little point in caring about such issues.

- Dave

#5 lore raw
Quoted 14 lines · On Mon, 25 Feb 2008 09:51:16 -0800 David Brownell <[email protected]> wrote:
Quoted 14 lines
Quoted 13 lines
Quoted 8 lines

drivers/clocksource/tcb_clksrc.c line 47

Quoted 6 lines
drivers/clocksource/tcb_clksrc.c new file +6 -0
+static cycle_t tc_get_cycles(void)
+{
+	unsigned long	flags;
+	u32		lower, upper;
+
+	raw_local_irq_save(flags);

Why do you need to use the raw version?

This is part of the system timer code, and it should never be a preemption point. Plus I didn't want to waste any instruction cycles in code that runs before e.g. the DBGU IRQ handler would get called... observably, such extra cycles *do* hurt.

Quoted 5 lines
Quoted 2 lines

I don't understand what you mean by preemption point, but I guess the non-raw version may consume some extra cycles when lockdep is enabled.

A preemption point is where CONFIG_PREEMPT kicks in task switching logic; lockdep is different.

I know, but I dont' see how local_irq_save/restore has anything to do with it, raw or not. There would be absolutely no point checking for preemption on local_irq_restore() since no one would have been able to set the TIF_NEED_RESCHED flag while interrupts were disabled...

raw_local_irq_save/restore is only different from local_irq_save/restore when lockdep is enabled. That's why I don't understand why you're talking about preemption.

Quoted 5 lines
Quoted 3 lines

If we really expect using TC as a clocksource but not as a clockevent is going to be a common usage, perhaps we should move the decision into Kconfig?

Maybe, but I already spent lots more time on this than I wanted. :(

I'm not asking you to do it. I'm asking if it would be a good thing to do. I said that I can take these patches off your back if you want, but I want to make sure I don't do anything with them that you disagree with.

Quoted 7 lines

Another way to address that rm9200 issue would be to just rate the TC clockevent source lower than the one based on the system timer, so it's set up but never enabled ... and remember "t2_clk", calling clk_enable() only when that clockevent device is active.

That would address one half of the suspend/resume equation too, ensuring that clock is disabled during suspend...

Yes, we could probably do that. Which means we can just remove all the ifdeffery?

Quoted 4 lines

The other half being: how to clk_disable(t0_clk) during system suspend? (And t1_clk on some systems.) There's currently no clocksource.set_mode() call; evidently there's an assumption that such counters cost no power, so they can always be left running.

Yes...that would be a clocksource core issue I guess. Better leave that for later...

Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
Quoted 2 lines

I don't think it is safe to assume that one clock per channel always means one irq per channel as well...

On current chips, that's how it works.

Quoted 8 lines
Quoted 2 lines

Indeed. I just don't see any fundamental reason why it has to work that way, which is why I don't think we should depend on it.

AT91 chips share identifiers between clocks and interrupts; that's fundamental, yes?

If some future chip works differently, that's a good time to change things. Otherwise I see little point in caring about such issues.

Agreed.

Haavard

#6 lore raw
Quoted 10 lines · Haavard Skinnemoen
Quoted 7 lines

Another way to address that rm9200 issue would be to just rate the TC clockevent source lower than the one based on the system timer, so it's set up but never enabled ... and remember "t2_clk", calling clk_enable() only when that clockevent device is active.

That would address one half of the suspend/resume equation too, ensuring that clock is disabled during suspend...

Yes, we could probably do that. Which means we can just remove all the ifdeffery?

There'd still be the #ifdef for CONFIG_GENERIC_CLOCKEVENTS, unless all the platforms get updated to support that. Right now it's a "patches available but not merged" situation.

Quoted 7 lines
Quoted 4 lines

The other half being:  how to clk_disable(t0_clk) during system suspend?  (And t1_clk on some systems.)  There's currently no clocksource.set_mode() call; evidently there's an assumption that such counters cost no power, so they can always be left running.

Yes...that would be a clocksource core issue I guess. Better leave that for later...

My thoughts exactly. ;)

Plus a bit of puzzlement why that didn't trigger at least a warning during AT91 suspend testing. There used to be warnings there about clocks which were wrongly left enabled...

- Dave