[PATCH 3/3] ARM: tegra: make sure the pointer on 4 byte align when THUMB2_KERNEL enabled

[Dave Martin]

On Wed, Apr 17, 2013 at 10:22:02AM -0600, Stephen Warren wrote:
> On 04/17/2013 04:30 AM, Dave Martin wrote:
> > On Tue, Apr 16, 2013 at 10:20:11AM -0600, Stephen Warren wrote:
> >> On 04/16/2013 08:13 AM, Dave Martin wrote:
> >>> On Mon, Apr 15, 2013 at 04:50:55PM -0600, Stephen Warren wrote:
> >>>> From: Joseph Lo <josephl at nvidia.com>
> >>>>
> >>>> When building kernel with CONFIG_THUMB2_KERNEL, the data pointer in the
> >>>> assembly may not on the 4 byte alignment. Then causing a data abort when
> >>>> accessing the pointer. This patch add a ".align" flag in the head of the
> >>>> pointer. And always using 32-bit ADR Thumb instruction to make sure it
> >>>> won't build failure.
> >>>>
> >>>> Signed-off-by: Joseph Lo <josephl at nvidia.com>
> >>>> Signed-off-by: Stephen Warren <swarren at nvidia.com>
> >>>> ---
> >>>>  arch/arm/mach-tegra/reset-handler.S |    1 +
> >>>>  arch/arm/mach-tegra/sleep.h         |    3 ++-
> >>>>  2 files changed, 3 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/arch/arm/mach-tegra/reset-handler.S b/arch/arm/mach-tegra/reset-handler.S
> >>>> index e6de88a..519a8c5 100644
> >>>> --- a/arch/arm/mach-tegra/reset-handler.S
> >>>> +++ b/arch/arm/mach-tegra/reset-handler.S
> >>>> @@ -83,6 +83,7 @@ ENDPROC(tegra_resume)
> >>>>  
> >>>>  #ifdef CONFIG_CACHE_L2X0
> >>>>  	.globl	l2x0_saved_regs_addr
> >>>> +	.align
> >>>>  l2x0_saved_regs_addr:
> >>>>  	.long	0
> >>>>  #endif
> >>>> diff --git a/arch/arm/mach-tegra/sleep.h b/arch/arm/mach-tegra/sleep.h
> >>>> index 4ffae54..8e9b6af 100644
> >>>> --- a/arch/arm/mach-tegra/sleep.h
> >>>> +++ b/arch/arm/mach-tegra/sleep.h
> >>>> @@ -92,7 +92,8 @@
> >>>>  
> >>>>  #ifdef CONFIG_CACHE_L2X0
> >>>>  .macro l2_cache_resume, tmp1, tmp2, tmp3, phys_l2x0_saved_regs
> >>>> -	adr	\tmp1, \phys_l2x0_saved_regs
> >>>> + ARM(	adr	\tmp1, \phys_l2x0_saved_regs )
> >>>> + THUMB(	adr.w	\tmp1, \phys_l2x0_saved_regs )
> >>>
> >>> Can you give an example of the assembler error you get without this?
> >>
> >> arch/arm/mach-tegra/reset-handler.S: Assembler messages:
> >> arch/arm/mach-tegra/reset-handler.S:78: Error: invalid immediate for
> >> address calculation (value = 0x00000004)
> > 
> > The immediate in the 16-bit form of ADR is alignment-sensitive, and
> > can't address anything that's not on a word boundary.  What if you just
> > have the .align, without adr.w?
> > 
> >> This is with gcc-4.5.3, as:
> >> GNU assembler (crosstool-NG hg_unknown at 20110628.165246) 2.20.1.20100303
> >>
> >>> The target symbol is local and the assembler can see where it is,
> >>> so it should choose the correct variant for the adr instruction
> >>> with no need for the ".w" suffix.
> >>>
> >>> If not, it could mean that there is a bug in the version of the
> >>> assembler you're using.
> >>
> >> True. Switching to the Ubuntu-packaged ARM cross-compiler that ships
> >> with Ubuntu 12.10, I do not see this problem.
> >>
> >> Still, many people probably still use older compilers.
> >>
> >>> If it's definitely needed, you can append a Thumb-only .w suffix with
> >>>
> >>> 	W(adr)	\tmp1, \phys_l2x0_saved_regs
> >>>
> >>> ...which is slightly neater.
> >>
> >> Yes, that certainly works too, even without the .align change in
> >> reset-handler.S.
> > 
> > Sure, although the unaligned access is not such a good idea anyway.
> > 
> > Keeping the adr fix is harmless, I just wanted to understand why it's
> > needed.
> 
> With just the following, and no adr.w:
> 
> > diff --git a/arch/arm/mach-tegra/reset-handler.S b/arch/arm/mach-tegra/reset-handler.S
> > index e6de88a..519a8c5 100644
> > --- a/arch/arm/mach-tegra/reset-handler.S
> > +++ b/arch/arm/mach-tegra/reset-handler.S
> > @@ -83,6 +83,7 @@ ENDPROC(tegra_resume)
> >  
> >  #ifdef CONFIG_CACHE_L2X0
> >         .globl  l2x0_saved_regs_addr
> > +       .align
> >  l2x0_saved_regs_addr:
> >         .long   0
> >  #endif
> 
> I still get:
> 
> > arch/arm/mach-tegra/reset-handler.S: Assembler messages:
> > arch/arm/mach-tegra/reset-handler.S:78: Error: invalid immediate for address calculation (value = 0x00000004)
> 
> That's with GNU assembler (crosstool-NG hg_unknown at 20110628.165246)
> 2.20.1.20100303.
> 
> BTW, I'm on vacation for the next 2 weeks starting this afternoon.
> Hopefully Joseph can keep this thread alive while I'm gone if needed.

OK.  I'm happy to go with the W() so long definitely fixes something :)

Cheers
---Dave