xtensa: fix futex_atomic_cmpxchg_inatomic

Return 0 if the operation was successful, not the userspace memory
value. Check that userspace value equals passed oldval, not itself.
Don't update *uval if the value wasn't read from userspace memory.

This fixes process hang due to infinite loop in futex_lock_pi.
It also fixes a bunch of glibc tests nptl/tst-mutexpi*.

Cc: stable@vger.kernel.org
Signed-off-by: NMax Filippov <jcmvbkbc@gmail.com>

arch/xtensa/include/asm/futex.h

@@ -92,7 +92,6 @@ futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
 			      u32 oldval, u32 newval)
 {
 	int ret = 0;
-	u32 prev;
 
 	if (!access_ok(VERIFY_WRITE, uaddr, sizeof(u32)))
 		return -EFAULT;
@@ -103,26 +102,24 @@ futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
 
 	__asm__ __volatile__ (
 	"	# futex_atomic_cmpxchg_inatomic\n"
-	"1:	l32i	%1, %3, 0\n"
-	"	mov	%0, %5\n"
-	"	wsr	%1, scompare1\n"
-	"2:	s32c1i	%0, %3, 0\n"
-	"3:\n"
+	"	wsr	%5, scompare1\n"
+	"1:	s32c1i	%1, %4, 0\n"
+	"	s32i	%1, %6, 0\n"
+	"2:\n"
 	"	.section .fixup,\"ax\"\n"
 	"	.align 4\n"
-	"4:	.long	3b\n"
-	"5:	l32r	%1, 4b\n"
-	"	movi	%0, %6\n"
+	"3:	.long	2b\n"
+	"4:	l32r	%1, 3b\n"
+	"	movi	%0, %7\n"
 	"	jx	%1\n"
 	"	.previous\n"
 	"	.section __ex_table,\"a\"\n"
-	"	.long 1b,5b,2b,5b\n"
+	"	.long 1b,4b\n"
 	"	.previous\n"
-	: "+r" (ret), "=&r" (prev), "+m" (*uaddr)
-	: "r" (uaddr), "r" (oldval), "r" (newval), "I" (-EFAULT)
+	: "+r" (ret), "+r" (newval), "+m" (*uaddr), "+m" (*uval)
+	: "r" (uaddr), "r" (oldval), "r" (uval), "I" (-EFAULT)
 	: "memory");
 
-	*uval = prev;
 	return ret;
 }