Initcall changes for omaps

We don't need timers right away and initializing them later gives us few nice things like interrupts and kmalloc. As the timers have a dependency to the clock framework, we're better off initializing things later rather than early if things go wrong. And this allows us to make the mux clock driver needed for system timers into early_platform drivers.

Note that smp_prepare_cpus() will get called later on during the init so we just need to local_irq_enable/disable for clocksource_probe().

Cc Felipe Balbi <[email protected]>

Cc Grygorii Strashko <[email protected]>

Cc Paul Walmsley <[email protected]>

Cc Tero Kristo <[email protected]>

Signed-off-by Tony Lindgren <[email protected]>

---
  arch/arm/mach-omap2/timer.c | 46 +++++++++++++++++++++++++++++++++++++++------
  1 file changed, 40 insertions(+), 6 deletions(-)

arch/arm/mach-omap2/timer.c +0 -0

index b18ebbe..68bf482 100644

Quoted 23 lines · Tony Lindgren

We don't need timers right away and initializing them later gives us few nice things like interrupts and kmalloc. As the timers have a dependency to the clock framework, we're better off initializing things later rather than early if things go wrong. And this allows us to make the mux clock driver needed for system timers into early_platform drivers.

Note that smp_prepare_cpus() will get called later on during the init so we just need to local_irq_enable/disable for clocksource_probe().

Cc Felipe Balbi <[email protected]>

Cc Grygorii Strashko <[email protected]>

Cc Paul Walmsley <[email protected]>

Cc Tero Kristo <[email protected]>

Signed-off-by Tony Lindgren <[email protected]>

---
  arch/arm/mach-omap2/timer.c | 46 +++++++++++++++++++++++++++++++++++++++------
  1 file changed, 40 insertions(+), 6 deletions(-)

arch/arm/mach-omap2/timer.c +0 -0

index b18ebbe..68bf482 100644

arch/arm/mach-omap2/timer.c static void __init __omap_sync32k_timer_init(int clkev_nr, const char *clkev_src · line 487

Quoted 10 lines

arch/arm/mach-omap2/timer.c +2 -0

@@ -478,36 +478,56 @@ static void __init __omap_sync32k_timer_init(int clkev_nr, const char *clkev_src
  	omap2_gp_clockevent_init(clkev_nr, clkev_src, clkev_prop);
 
  	/* Enable the use of clocksource="gp_timer" kernel parameter */
+	local_irq_disable();
  	if (use_gptimer_clksrc || gptimer)
  		omap2_gptimer_clocksource_init(clksrc_nr, clksrc_src,
  						clksrc_prop);
  	else
  		omap2_sync32k_clocksource_init();
+	local_irq_enable();

So, this will be called now after sched_clock_postinit() and to W/A warnings you've added local_irq_disable()/local_irq_enable(). Am I right?

Are you sure this is safe?

We want to be able to probe a few selected device drivers before hwmod code populates the clocks in omap_hwmod_setup_all(). This allows us to convert most of the clock drivers into regular device drivers.

We only need a few minimal clock drivers early for the system timers to select between the 32KiHz clock and the high frequency oscillator.

With these changes, initializing the clock drivers can be just done at core_initcall time with something like:

np = of_find_node_by_name(NULL, "plls");
if (np)

	of_platform_populate(np, NULL, NULL, NULL);

And then these clocks will be available for the interconnect code to use.

Having most of the clock drivers being regular device drivers allows us to use the nice things like devm_* functions and dev_err and dev_dbg. As an extra bonus, this also allows us to develop the clock drivers for new SoCs as loadable modules initially for cases where we can boot up the system based on the bootloader configured clocks.

To do this, let's change the core_initcalls to postcore_initcall under mach-omap2.

Quoted 3 lines · Tony Lindgren

We want to be able to probe a few selected device drivers before hwmod code populates the clocks in omap_hwmod_setup_all(). This allows us to convert most of the clock drivers into regular device drivers.

Quoted 19 lines

We only need a few minimal clock drivers early for the system timers to select between the 32KiHz clock and the high frequency oscillator.

With these changes, initializing the clock drivers can be just done at core_initcall time with something like:

np = of_find_node_by_name(NULL, "plls");
if (np)

	of_platform_populate(np, NULL, NULL, NULL);

And then these clocks will be available for the interconnect code to use.

Having most of the clock drivers being regular device drivers allows us to use the nice things like devm_* functions and dev_err and dev_dbg. As an extra bonus, this also allows us to develop the clock drivers for new SoCs as loadable modules initially for cases where we can boot up the system based on the bootloader configured clocks.

Quoted 2 lines

To do this, let's change the core_initcalls to postcore_initcall under mach-omap2.

FYI I'm applying this one into omap-for-v4.5/soc today, the first patch in this series needs more work as discussed.

Quoted 4 lines · Tony Lindgren

Quoted 3 lines · Tony Lindgren

We want to be able to probe a few selected device drivers before hwmod code populates the clocks in omap_hwmod_setup_all(). This allows us to convert most of the clock drivers into regular device drivers.

if understand things right, ti clks now will be populated and initialized from
__omap_sync32k_timer_init
- omap_clk_init()
- ..
- of_clk_init()
- ..
- omap_clk_soc_init()

and __omap_sync32k_timer_init(), in turn, will be called from:
arch/arm/kernel/time.c
- time_init()

	machine_desc->init_time();

(without your patch 1).

So, I don't see real dependency here between clk initialization and hwmods :(

Quoted 15 lines · Grygorii Strashko

Quoted 4 lines · Tony Lindgren

Quoted 3 lines · Tony Lindgren

We want to be able to probe a few selected device drivers before hwmod code populates the clocks in omap_hwmod_setup_all(). This allows us to convert most of the clock drivers into regular device drivers.

if understand things right, ti clks now will be populated and initialized from
__omap_sync32k_timer_init
- omap_clk_init()
- ..
- of_clk_init()
- ..
- omap_clk_soc_init()

Quoted 5 lines

and __omap_sync32k_timer_init(), in turn, will be called from:
arch/arm/kernel/time.c
- time_init()

	machine_desc->init_time();

(without your patch 1).

Yes that's the current approach, but we can do better. We only need the following clocks for system timers at that point:

- mux clocks to select between the 32k and hf oscillator source

- clkctrl driver to gate the ocp clock

All the other clocks can be initialized at core_initcall time with this change.

Quoted 1 line

So, I don't see real dependency here between clk initialization and hwmods :(

You don't because it's only implemented so far for the dm814x ADPLL clock :)

That I posted as "[PATCH 0/2] Clock driver for dm814x ADPLL". Note how it's just a regular device driver that also works as loadable module on boards that have all the necessary clocks enabled already by the bootloader.