On Wed, May 12, 2021 at 10:13 AM Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
> From: Arnd Bergmann <arnd@xxxxxxxx>
>
> When building the kernel wtih gcc-10 or higher using the
> CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE=y flag, the compiler picks a slightly
> different set of registers for the inline assembly in cpu_init() that
> subsequently results in a corrupt kernel stack as well as remaining in
> FIQ mode. If a banked register is used for the last argument, the wrong
> version of that register gets loaded into CPSR_c. When building in Arm
> mode, the arguments are passed as immediate values and the bug cannot
> happen.
>
> This got introduced when Daniel reworked the FIQ handling and was
> technically always broken, but happened to work with both clang and gcc
> before gcc-10 as long as they picked one of the lower registers.
> This is probably an indication that still very few people build the
> kernel in Thumb2 mode.
>
> Marek pointed out the problem on IRC, Arnd narrowed it down to this
> inline assembly and Russell pinpointed the exact bug.
>
> Change the constraints to force the final mode switch to use a non-banked
> register for the argument to ensure that the correct constant gets loaded.
> Another alternative would be to always use registers for the constant
> arguments to avoid the #ifdef that has now become more complex.
>
> Cc: <stable@xxxxxxxxxxxxxxx> # v3.18+
> Cc: Daniel Thompson <daniel.thompson@xxxxxxxxxx>
> Reported-by: Marek Vasut <marek.vasut@xxxxxxxxx>
> Fixes: c0e7f7ee717e ("ARM: 8150/3: fiq: Replace default FIQ handler")
> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
Wow. Nice bug hunt here, hats off!
Reviewed-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
Yours,
Linus Walleij
