MX51: add FIQ support for TZIC

[phorton@bitbox.co.uk (Peter Horton)]

On 25/11/2010 20:38, Uwe Kleine-K?nig wrote:
>
> On Thu, Nov 25, 2010 at 03:19:09PM +0000, Peter Horton wrote:
>> Add support for FIQ on MX51 TZIC.
>>
>> (patch is against Sascha's "imx-for-2.6.38" branch)
> <nitpick>The comment about which tree this applies to isn't of interest
> anymore when the patch is applied, so better write it after the triple
> (eww, double?) dash below.  Then it won't make it into git.</nitpick>
>
>> Signed-off-by: Peter Horton<phorton@bitbox.co.uk>
>>
>> --
>>
> Back some time I tried to achieve something similar.  IIRC my patch was
> easier.  I definitly didn't patch entry-macro.S.  But OTOH I had some
> problems with it so this might have been my problem.  I cannot find it
> at the moment, will talk to Sascha tomorrow about it.
>

There is a tiny window where the interrupt entry code reads the HIPND 
register and can see a pending FIQ. This can cause the interrupt code to 
pick up the FIQ interrupt number rather than the requesting interrupt 
number which results in the kernel's interrupt handler disabling the FIQ 
(as an interrupt without a handler).

When playing audio using the FIQ this happens after a few seconds 
without the attached entry macro fix.

>> diff --git a/arch/arm/plat-mxc/include/mach/entry-macro.S b/arch/arm/plat-mxc/include/mach/entry-macro.S
>> index aeb0869..c7dd4a9 100644
>> --- a/arch/arm/plat-mxc/include/mach/entry-macro.S
>> +++ b/arch/arm/plat-mxc/include/mach/entry-macro.S
>> @@ -54,8 +54,18 @@
>>   #elif defined CONFIG_MXC_TZIC
>>   	@ Load offset&  priority of the highest priority
>>   	@ interrupt pending.
>> +	@ 0x080 is INTSEC0 register
>>   	@ 0xD80 is HIPND0 register
>>   	mov     \irqnr, #0
>> +#ifdef CONFIG_FIQ
>> +1000:
>> +	add	\irqstat, \base, \irqnr, lsr #3
>> +	ldr	\tmp, [\irqstat, #0xd80]
>> +	ldr	\irqstat, [\irqstat, #0x080]
>> +	ands	\tmp, \tmp, \irqstat
>> +	bne	1001f
>> +	add	\irqnr, \irqnr, #32
>> +#else
>>   	mov     \irqstat, #0x0D80
>>   1000:
>>   	ldr     \tmp,   [\irqstat, \base]
>> @@ -63,6 +73,7 @@
>>   	bne     1001f
>>   	addeq   \irqnr, \irqnr, #32
>>   	addeq   \irqstat, \irqstat, #4
>> +#endif
>>   	cmp     \irqnr, #128
>>   	blo     1000b
>>   	b       2001f
>> diff --git a/arch/arm/plat-mxc/tzic.c b/arch/arm/plat-mxc/tzic.c
>> index 3703ab2..b648a76 100644
>> --- a/arch/arm/plat-mxc/tzic.c
>> +++ b/arch/arm/plat-mxc/tzic.c
>> @@ -47,6 +47,33 @@
>>
>>   void __iomem *tzic_base; /* Used as irq controller base in entry-macro.S */
>>
>> +#ifdef CONFIG_FIQ
>> +
>> +/*
>> + * switch interrupt between IRQ and FIQ mode.
>> + *
>> + * 'type' is true for FIQ mode and false for IRQ mode.
>> + */
>> +int mxc_set_irq_fiq(unsigned int irq, unsigned int type)
>> +{
>> +	unsigned int index, mask, value;
>> +
>> +	index = irq>>  5;
>> +	if (unlikely(index>= 4))
>> +		return -EBUSY;
>> +	mask = 1U<<  (irq&  0x1F);
>> +
>> +	value = __raw_readl(tzic_base + TZIC_INTSEC0(index));
>> +	value = (value | mask) ^ (type ? mask : 0);
>> +	__raw_writel(value, tzic_base + TZIC_INTSEC0(index));
> Does this result in better code than the straigt forward
>
> 	value = __raw_readl(...)&  ~mask;
> 	if (type)
> 		vale |= mask
> 	__raw_writel(...);
>
> If not, what is the idea to write it this way?

Write it however you like :)

>> +
>> +	return 0;
>> +}
>> +
>> +EXPORT_SYMBOL(mxc_set_irq_fiq);
> Hmm, I guess this makes it impossible to compile avic.c and tzic.c with
> FIQ=y into a single binary.
>

Fair point.

P.