ARM: 8550/1: protect idiv patching against undefined gcc behavior

It was reported that a kernel with CONFIG_ARM_PATCH_IDIV=y stopped
booting when compiled with the upcoming gcc 6.  Turns out that turning
a function address into a writable array is undefined and gcc 6 decided
it was OK to omit the store to the first word of the function while
still preserving the store to the second word.

Even though gcc 6 is now fixed to behave more coherently, it is a
mystery that gcc 4 and gcc 5 actually produce wanted code in the kernel.
And in fact the reduced test case to illustrate the issue does indeed
break with gcc < 6 as well.

In any case, let's guard the kernel against undefined compiler behavior
by hiding the nature of the array location as suggested by gcc
developers.

Reference: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70128

Signed-off-by: Nicolas Pitre <nico@linaro.org>
Reported-by: Marcin Juszkiewicz <mjuszkiewicz@redhat.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: stable@vger.kernel.org # v4.5
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

diff --git a/arch/arm/kernel/setup.c b/arch/arm/kernel/setup.c
index 139791ed473d5264682c004ea2ea7af8ddd28f5d..a28fce0bdbbe11b7a950a3a0943f8362d7124b87 100644 (file)
--- a/arch/arm/kernel/setup.c
+++ b/arch/arm/kernel/setup.c
@@ -430,11 +430,13 @@ static void __init patch_aeabi_idiv(void)
        pr_info("CPU: div instructions available: patching division code\n");
 
        fn_addr = ((uintptr_t)&__aeabi_uidiv) & ~1;
+       asm ("" : "+g" (fn_addr));
        ((u32 *)fn_addr)[0] = udiv_instruction();
        ((u32 *)fn_addr)[1] = bx_lr_instruction();
        flush_icache_range(fn_addr, fn_addr + 8);
 
        fn_addr = ((uintptr_t)&__aeabi_idiv) & ~1;
+       asm ("" : "+g" (fn_addr));
        ((u32 *)fn_addr)[0] = sdiv_instruction();
        ((u32 *)fn_addr)[1] = bx_lr_instruction();
        flush_icache_range(fn_addr, fn_addr + 8);