sparc: mark __arch_xchg() as __always_inline

An otherwise correct change to the atomic operations uncovered an
existing bug in the sparc __arch_xchg() function, which is calls
__xchg_called_with_bad_pointer() when its arguments are unknown at
compile time:

ERROR: modpost: "__xchg_called_with_bad_pointer" [lib/atomic64_test.ko] undefined!

This now happens because gcc determines that it's better to not inline the
function. Avoid this by just marking the function as __always_inline
to force the compiler to do the right thing here.

Reported-by: Guenter Roeck <[email protected]>
Link: https://lore.kernel.org/all/[email protected]/
Fixes: d12157e ("locking/atomic: make atomic*_{cmp,}xchg optional")
Signed-off-by: Arnd Bergmann <[email protected]>
Acked-by: Palmer Dabbelt <[email protected]>
Acked-by: Mark Rutland <[email protected]>
Reviewed-by: Sam Ravnborg <[email protected]>
Acked-by: Guenter Roeck <[email protected]>
Acked-by: Andi Shyti <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Kees Cook <[email protected]>

Author: arndb

arch/sparc/include/asm/cmpxchg_32.h

@@ -15,7 +15,7 @@
 unsigned long __xchg_u32(volatile u32 *m, u32 new);
 void __xchg_called_with_bad_pointer(void);
 
-static inline unsigned long __arch_xchg(unsigned long x, __volatile__ void * ptr, int size)
+static __always_inline unsigned long __arch_xchg(unsigned long x, __volatile__ void * ptr, int size)
 {
 	switch (size) {
 	case 4:

arch/sparc/include/asm/cmpxchg_64.h

@@ -87,7 +87,7 @@ xchg16(__volatile__ unsigned short *m, unsigned short val)
 	return (load32 & mask) >> bit_shift;
 }
 
-static inline unsigned long
+static __always_inline unsigned long
 __arch_xchg(unsigned long x, __volatile__ void * ptr, int size)
 {
 	switch (size) {