[PATCH] ARM: reinsert ARCH_MULTI_V4 Kconfig option

[arnd@arndb.de (Arnd Bergmann)]

On Wednesday 09 April 2014 16:27:11 Russell King - ARM Linux wrote:
> On Wed, Apr 09, 2014 at 05:13:26PM +0200, Uwe Kleine-K?nig wrote:
> > On Wed, Apr 09, 2014 at 04:06:40PM +0100, Russell King - ARM Linux wrote:
> > > On Wed, Apr 09, 2014 at 04:54:16PM +0200, Jonas Jensen wrote:
> > > > On 13 December 2013 12:39, Russell King - ARM Linux <linux@arm.linux.org.uk> wrote:
> > > > diff --git a/arch/arm/kernel/entry-armv.S b/arch/arm/kernel/entry-armv.S
> > > > index 1879e8d..de15bfd 100644
> > > > --- a/arch/arm/kernel/entry-armv.S
> > > > +++ b/arch/arm/kernel/entry-armv.S
> > > > @@ -739,6 +739,18 @@ ENDPROC(__switch_to)
> > > > 
> > > >         .macro  usr_ret, reg
> > > >  #ifdef CONFIG_ARM_THUMB
> > > > +       /*
> > > > +        * Having CONFIG_ARM_THUMB isn't a guarantee that the cpu has support
> > > > +        * for Thumb and so the bx instruction. Use a mov if the address to
> > > > +        * jump to is 32 bit aligned. (Note that this code is compiled in ARM
> > > > +        * mode, so this is the right test.)
> > > > +        */
> > > > +#if defined(CONFIG_CPU_32v4)
> > > > +       tst     \reg, #3
> > > > +       moveq   pc, \reg
> > > > +       b       .
> > > > +#endif
> > > > +
> > > >         bx      \reg
> > > 
> > > What's wrong with:
> > >     tst     \reg, #3
> > >     moveq   pc, \reg
> > >     bx      \reg
> > > 
> > > rather than ending in an infinite loop?
> > The added b . was a test to check if the machine then hangs instead of
> > crashing. (And yes, that was the case, so it was tried to return to a
> > non-aligned address.)
> 
> If it's called from ARM code, then \reg will contain a 4-byte aligned
> address.  If it's called from Thumb code, \reg will contain a 2-byte
> aligned address with bit 0 *always* set.

Right, that is the assumption.

> So, with the code originally quoted above, if the helper is called from
> thumb code, and CONFIG_CPU_32v4 is enabled, then we end up falling past
> the moveq to the "b ." and entering an infinite loop.

As Uwe said, that "b ." was not meant to be in the patch used for
submission, it was to check what goes wrong when running this code
on ARMv4 -- either crash user space or hang in an infinite loop.
I forgot what the result of that experiment was.

The trouble is that this code:

       .macro  usr_ret, reg
#ifdef CONFIG_ARM_THUMB
#ifdef CONFIG_CPU_32v4
	tst     \reg, #3
	moveq   pc, \reg
#endif
        bx      \reg
#else
	mov	pc, \reg
#endif
	.endm

for some reason does the wrong thing running on ARMv4 (fa526) with non-thumb
user space when both CONFIG_CPU_32v4 and CONFIG_ARM_THUMB are enabled:
it still tries to do the 'bx' and triggers an invalid opcode exception.

	Arnd