[PATCH 2/5] ARM: mvebu: Simplify memory init order

[Sebastian Hesselbarth]

On 09/17/2014 08:45 AM, Sascha Hauer wrote:
> On Tue, Sep 16, 2014 at 10:05:44PM +0200, Sebastian Hesselbarth wrote:
>> On 09/15/2014 09:41 AM, Sascha Hauer wrote:
>>> The initialisation of the memory nodes on mvebu is a bit
>>> compilcated:
>>>
>>> pure_initcall(mvebu_memory_fixup_register)
>>> 	of_register_fixup(mvebu_memory_of_fixup, NULL)
>>> core_initcall(kirkwood_init_soc)
>>> 	mvebu_set_memory()
>>> core_initcall(of_arm_init)
>>> 	of_fix_tree()
>>> 		mvebu_memory_of_fixup()
>>>
>>> First a mvebu common of_fixup function is registered, then the SoC
>>> calls mvebu_set_memory which stores the memory base and size in global
>>> variables. Afterwards the of_fixup is executed which fixes the memory
>>> nodes according to the global variables.
>>>
>>> Instead register a SoC specific fixup which directly calls mvebu_set_memory
>>> with the memory base and size as arguments:
>>>
>>> pure_initcall(kirkwood_register_soc_fixup);
>>> 	of_register_fixup(kirkwood_init_soc, NULL);
>>> core_initcall(of_arm_init)
>>> 	of_fix_tree()
>>> 		kirkwood_init_soc()
>>> 			mvebu_set_memory(phys_base, phys_size);
>>>
>>> Signed-off-by: Sascha Hauer <s.hauer at pengutronix.de>
>>
>> Hmm, this breaks Armada 370 and most likely also Armada XP. Actually,
>> it breaks any SoC that has a DTB with internal regs set to 0xd0000000.
>>
>>> ---
>>>   arch/arm/mach-mvebu/armada-370-xp.c       |  9 ++++++--
>>>   arch/arm/mach-mvebu/common.c              | 34 ++++++++++---------------------
>>>   arch/arm/mach-mvebu/dove.c                |  9 ++++++--
>>>   arch/arm/mach-mvebu/include/mach/common.h |  2 +-
>>>   arch/arm/mach-mvebu/kirkwood.c            |  9 ++++++--
>>>   5 files changed, 33 insertions(+), 30 deletions(-)
>>>
>>> diff --git a/arch/arm/mach-mvebu/armada-370-xp.c b/arch/arm/mach-mvebu/armada-370-xp.c
>>> index 6251100..5c8499b 100644
>>> --- a/arch/arm/mach-mvebu/armada-370-xp.c
>>> +++ b/arch/arm/mach-mvebu/armada-370-xp.c
>>> @@ -52,7 +52,7 @@ static void __noreturn armada_370_xp_reset_cpu(unsigned long addr)
>>>   		;
>>>   }
>>>
>>> -static int armada_370_xp_init_soc(void)
>>> +static int armada_370_xp_init_soc(struct device_node *root, void *context)
>>>   {
>>>   	unsigned long phys_base, phys_size;
>>>   	u32 reg;
>>> @@ -74,4 +74,9 @@ static int armada_370_xp_init_soc(void)
>>
>> Because armada_370_xp_init_soc() does
>>
>> 	mvebu_mbus_add_range(0xf0, 0x01, MVEBU_REMAP_INT_REG_BASE);
>>
>> right above, which will add the range(s) required for internal register
>> of_fixup. Since this patch moved armada_370_xp_init_soc to the
>> of_fixups, we don't fix this up for the initial DT tree.
>>
>>>
>>>   	return 0;
>>>   }
>>> -core_initcall(armada_370_xp_init_soc);
>>> +
>>> +static int armada_370_register_soc_fixup(void)
>>> +{
>>
>> I guess moving mvebu_mbus_add_range() in here does not work, because
>> it will add the armada_370_xp range also for dove and kirkwood.
>
> Do we need a separate of_fixup registered for fixing up the mbus? Can't
> we just call mvebu_mbus_of_fixup() directly from mvebu_mbus_add_range()?
>
> This won't be very efficient for dove which calls mvebu_mbus_add_range()
> twice so it has to iterate over the device tree twice, but it should
> work, right?

It is not the number of iterations I am concerned about. In a Multi-SoC
environment, we add ranges for MBUS_ID(0xf0, 0x01) three times anyway
plus Dove's MBUS_ID(0xf0, 0x02).

Currently, we are lucky that MBUS_ID(0xf0, 0x01) is the same for all
SoCs and MBUS_ID(0xf0, 0x02) is exclusive for Dove. If we cross our
fingers and hope it will never change, we can just do:

static int dove_register_soc_fixup(void)
{
	mvebu_mbus_add_range(0xf0, 0x01, MVEBU_REMAP_INT_REG_BASE);
	mvebu_mbus_add_range(0xf0, 0x02, DOVE_REMAP_MC_REGS);
	return of_register_fixup(dove_init_soc, NULL);
}
pure_initcall(dove_register_soc_fixup);

and

static int armada_370_xp_register_soc_fixup(void)
{
	mvebu_mbus_add_range(0xf0, 0x01, MVEBU_REMAP_INT_REG_BASE);
	return of_register_fixup(armada_370_xp_init_soc, NULL);
}
pure_initcall(armada_370_xp_register_soc_fixup);

and the same for Kirkwood.

If we want to make it safe for any future SoC that might break the
MBUS_ID() assumptions above, we can add the machine_id compatible
to mvebu_mbus_add_range(), i.e.

static int armada_370_xp_register_soc_fixup(void)
{
	mvebu_mbus_add_range("marvell,armada-370-xp", 0xf0, 0x01, 
MVEBU_REMAP_INT_REG_BASE);
	return of_register_fixup(armada_370_xp_init_soc, NULL);
}
pure_initcall(armada_370_xp_register_soc_fixup);

If you put your current series into some branch I can pull, I can
add the required patches later today.

Sebastian