[PATCH v3 RESEND 05/17] ARM: LPAE: support 64-bit virt_to_phys patching

[Dave Martin]

On Fri, Sep 21, 2012 at 11:56:03AM -0400, Cyril Chemparathy wrote:
> This patch adds support for 64-bit physical addresses in virt_to_phys()
> patching.  This does not do real 64-bit add/sub, but instead patches in the
> upper 32-bits of the phys_offset directly into the output of virt_to_phys.
> 
> There is no corresponding change on the phys_to_virt() side, because
> computations on the upper 32-bits would be discarded anyway.
> 
> Signed-off-by: Cyril Chemparathy <cyril at ti.com>
> ---
>  arch/arm/include/asm/memory.h |   38 ++++++++++++++++++++++++++++++++++++--
>  arch/arm/kernel/head.S        |    4 ++++
>  arch/arm/kernel/setup.c       |    2 +-
>  3 files changed, 41 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/arm/include/asm/memory.h b/arch/arm/include/asm/memory.h
> index 88ca206..f3e8f88 100644
> --- a/arch/arm/include/asm/memory.h
> +++ b/arch/arm/include/asm/memory.h
> @@ -154,13 +154,47 @@
>  #ifdef CONFIG_ARM_PATCH_PHYS_VIRT
>  
>  extern unsigned long	__pv_offset;
> -extern unsigned long	__pv_phys_offset;
> +extern phys_addr_t	__pv_phys_offset;
>  #define PHYS_OFFSET	__virt_to_phys(PAGE_OFFSET)
>  
>  static inline phys_addr_t __virt_to_phys(unsigned long x)
>  {
> -	unsigned long t;
> +	phys_addr_t t;
> +
> +#ifndef CONFIG_ARM_LPAE
>  	early_patch_imm8("add", t, x, __pv_offset, 0);
> +#else
> +	unsigned long __tmp;
> +
> +#ifndef __ARMEB__
> +#define PV_PHYS_HIGH	"(__pv_phys_offset + 4)"
> +#else
> +#define PV_PHYS_HIGH	"__pv_phys_offset"
> +#endif
> +
> +	early_patch_stub(
> +	/* type */		PATCH_IMM8,
> +	/* code */
> +		"ldr		%[tmp], =__pv_offset\n"
> +		"ldr		%[tmp], [%[tmp]]\n"
> +		"add		%Q[to], %[from], %[tmp]\n"
> +		"ldr		%[tmp], =" PV_PHYS_HIGH "\n"
> +		"ldr		%[tmp], [%[tmp]]\n"
> +		"mov		%R[to], %[tmp]\n",
> +	/* pad */		4,
> +	/* patch_data */
> +		".long		__pv_offset\n"
> +		"add		%Q[to], %[from], %[imm]\n"
> +		".long	"	PV_PHYS_HIGH "\n"
> +		"mov		%R[to], %[imm]\n",
> +	/* operands */
> +		: [to]	 "=r"	(t),
> +		  [tmp]	 "=&r"	(__tmp)
> +		: [from] "r"	(x),
> +		  [imm]	 "I"	(__IMM8),
> +			 "i"	(&__pv_offset),
> +			 "i"	(&__pv_phys_offset));

So, the actual offset we can apply is:

__pv_phys_offset + __pv_offset

where:

 * the high 32 bits of the address being fixed up are assumed to be 0
   (true, because the kernel is initially always fixed up to an address
   range <4GB)

 * the low 32 bits of __pv_phys_offset are assumed to be 0 (?)

 * the full offset is of the form

        ([..0..]XX[..0..] << 32) | [..0..]YY[..0..] 

Is this intentional?  It seems like a rather weird constraint...  but
it may be sensible.  PAGE_OFFSET is probably 0xc0000000 or 0x80000000,
(so YY can handle that) and the actual RAM above 4GB will likely be
huge and aligned on some enormous boundary in such situations (so that
XX can handle that).

So long as the low RAM alias is not misaligned relative to the high alias
on a finer granularity than 16MB (so that YY = (PAGE_OFFSET +/- the
misalignment) is still a legal immediate), I guess there should not be a
problem.

[...]

Cheers
---Dave