[PATCHv3] arm: socfpga: fix fetching cpu1start_addr for SMP

[dinguyen@opensource.altera.com (Dinh Nguyen)]

On 10/14/2014 02:18 PM, Russell King - ARM Linux wrote:
> On Tue, Oct 14, 2014 at 02:04:08PM -0500, dinguyen at opensource.altera.com wrote:
>>  ENTRY(secondary_trampoline)
>> -	movw	r2, #:lower16:cpu1start_addr
>> -	movt  r2, #:upper16:cpu1start_addr
>> -
>> -	/* The socfpga VT cannot handle a 0xC0000000 page offset when loading
>> -		the cpu1start_addr, we bit clear it. Tested on HW and VT. */
>> -	bic	r2, r2, #0x40000000
>> -
>> -	ldr	r0, [r2]
>> -	ldr	r1, [r0]
>> -	bx	r1
>> +	adr	r0, 1f
>> +	ldmia	r0, {r1, r2}
>> +	sub	r2, r2, #PAGE_OFFSET
>> +	ldr	r3, [r2]
>> +	ldr	r4, [r3]
>> +	bx	r4
>>  
>> +	.align
>> +1:	.long	.
>> +	.long	socfpga_cpu1start_addr
> 
> I'm still not happy with this.
> 
> The old code used to take the socfpga_cpu1start_addr address (which will
> probably have both bits 31 and 30 set), and clear bit 30.  That made the
> address in the range of 0x80000000 - 0xbfffffff.

That's correct. And with the our previous settings in the SDRAM
controller, it was working because it appears that the access was just
wrapping around when the controller was accessing memory outside of it's
settings. Now that we've changed the memory controller's setting to
block accesses beyond the max configured memory, simply clearing bit 30
will not work.

> 
> What you're doing above is taking the socfpga_cpu1start_addr address and
> subtracting PAGE_OFFSET from it.  This will give an address in the
> 0x00000000 - 0x3fffffff range.
> 
> PAGE_OFFSET is *not* the offset between physical and virtual addresses.
> It is the offset from _zero_ of the start of the kernel lowmem mapping.
> 
> In other words, if lowmem starts at 0xc0000000, then PAGE_OFFSET will be
> 0xc0000000 whether or not your physical memory starts at 0, 0x10000000,
> 0x40000000, 0x50000000 or whatever.
> 
> So, what this new code implies is that the start of physical memory is
> at zero on this platform.  A comment to that effect in the code would
> be nice so that people understand what's going on here, and also know
> not to copy this code...
> 

Yes, the physical start of CPU1 is will always be zero for this
platform. I will add a comment of something to that affect for V3.

Thanks,
Dinh