ARM: 7747/1: pcpu: ensure __my_cpu_offset cannot be re-ordered across barrier()

__my_cpu_offset is non-volatile, since we want its value to be cached
when we access several per-cpu variables in a row with preemption
disabled. This means that we rely on preempt_{en,dis}able to hazard
with the operation via the barrier() macro, so that we can't end up
migrating CPUs without reloading the per-cpu offset.

Unfortunately, GCC doesn't treat a "memory" clobber on a non-volatile
asm block as a side-effect, and will happily re-order it before other
memory clobbers (including those in prempt_disable()) and cache the
value. This has been observed to break the cmpxchg logic in the slub
allocator, leading to livelock in kmem_cache_alloc in mainline kernels.

This patch adds a dummy memory input operand to __my_cpu_offset,
forcing it to be ordered with respect to the barrier() macro.

Cc: <stable@vger.kernel.org>
Cc: Rob Herring <rob.herring@calxeda.com>
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

1 files changed, 9 insertions, 2 deletions

diff --git a/arch/arm/include/asm/percpu.h b/arch/arm/include/asm/percpu.h
index 968c0a14e0a3..209e6504922e 100644
--- a/arch/arm/include/asm/percpu.h
+++ b/arch/arm/include/asm/percpu.h
@@ -30,8 +30,15 @@ static inline void set_my_cpu_offset(unsigned long off)
 static inline unsigned long __my_cpu_offset(void)
 {
         unsigned long off;
-        /* Read TPIDRPRW */
-        asm("mrc p15, 0, %0, c13, c0, 4" : "=r" (off) : : "memory");
+        register unsigned long *sp asm ("sp");
+
+        /*
+         * Read TPIDRPRW.
+         * We want to allow caching the value, so avoid using volatile and
+         * instead use a fake stack read to hazard against barrier().
+         */
+        asm("mrc p15, 0, %0, c13, c0, 4" : "=r" (off) : "Q" (*sp));
+
         return off;
 }
 #define __my_cpu_offset __my_cpu_offset()