sparc32: unbreak arch_write_unlock()
authorMikael Pettersson <mikpe@it.uu.se>
Mon, 15 Aug 2011 10:11:50 +0000 (10:11 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 15 Aug 2011 21:35:19 +0000 (14:35 -0700)
commit3f6aa0b113846a8628baa649af422cfc6fb1d786
tree5b3d1f9b112a779dab37db5f685c24e5e857df96
parenta0fba3eb059e73fed2d376a901f8117734c12f1f
sparc32: unbreak arch_write_unlock()

The sparc32 version of arch_write_unlock() is just a plain assignment.
Unfortunately this allows the compiler to schedule side-effects in a
protected region to occur after the HW-level unlock, which is broken.
E.g., the following trivial test case gets miscompiled:

#include <linux/spinlock.h>
rwlock_t lock;
int counter;
void foo(void) { write_lock(&lock); ++counter; write_unlock(&lock); }

Fixed by adding a compiler memory barrier to arch_write_unlock().  The
sparc64 version combines the barrier and assignment into a single asm(),
and implements the operation as a static inline, so that's what I did too.

Compile-tested with sparc32_defconfig + CONFIG_SMP=y.

Signed-off-by: Mikael Pettersson <mikpe@it.uu.se>
Signed-off-by: David S. Miller <davem@davemloft.net>
arch/sparc/include/asm/spinlock_32.h