frv: Rewrite atomic implementation

Mostly complete rewrite of the FRV atomic implementation, instead of
using assembly files, use inline assembler.

The out-of-line CONFIG option makes a bit of a mess of things, but a
little CPP trickery gets that done too.

FRV already had the atomic logic ops but under a non standard name,
the reimplementation provides the generic names and provides the
intermediate form required for the bitops implementation.

The slightly inconsistent __atomic32_fetch_##op naming is because
__atomic_fetch_##op conlicts with GCC builtin functions.

The 64bit atomic ops use the inline assembly %Ln construct to access
the low word register (r+1), afaik this construct was not previously
used in the kernel and is completely undocumented, but I found it in
the FRV GCC code and it seems to work.

FRV had a non-standard definition of atomic_{clear,set}_mask() which
would work types other than atomic_t, the one user relying on that
(arch/frv/kernel/dma.c) got converted to use the new intermediate
form.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

1 files changed, 172 insertions, 0 deletions

diff --git a/arch/frv/include/asm/atomic_defs.h b/arch/frv/include/asm/atomic_defs.h
new file mode 100644
index 000000000000..36e126d2f801
--- /dev/null
+++ b/arch/frv/include/asm/atomic_defs.h
@@ -0,0 +1,172 @@
+
+#include <asm/spr-regs.h>
+
+#ifdef __ATOMIC_LIB__
+
+#ifdef CONFIG_FRV_OUTOFLINE_ATOMIC_OPS
+
+#define ATOMIC_QUALS
+#define ATOMIC_EXPORT(x)	EXPORT_SYMBOL(x)
+
+#else /* !OUTOFLINE && LIB */
+
+#define ATOMIC_OP_RETURN(op)
+#define ATOMIC_FETCH_OP(op)
+
+#endif /* OUTOFLINE */
+
+#else /* !__ATOMIC_LIB__ */
+
+#define ATOMIC_EXPORT(x)
+
+#ifdef CONFIG_FRV_OUTOFLINE_ATOMIC_OPS
+
+#define ATOMIC_OP_RETURN(op)						\
+extern int __atomic_##op##_return(int i, int *v);			\
+extern long long __atomic64_##op##_return(long long i, long long *v);
+
+#define ATOMIC_FETCH_OP(op)						\
+extern int __atomic32_fetch_##op(int i, int *v);			\
+extern long long __atomic64_fetch_##op(long long i, long long *v);
+
+#else /* !OUTOFLINE && !LIB */
+
+#define ATOMIC_QUALS	static inline
+
+#endif /* OUTOFLINE */
+#endif /* __ATOMIC_LIB__ */
+
+
+/*
+ * Note on the 64 bit inline asm variants...
+ *
+ * CSTD is a conditional instruction and needs a constrained memory reference.
+ * Normally 'U' provides the correct constraints for conditional instructions
+ * and this is used for the 32 bit version, however 'U' does not appear to work
+ * for 64 bit values (gcc-4.9)
+ *
+ * The exact constraint is that conditional instructions cannot deal with an
+ * immediate displacement in the memory reference, so what we do is we read the
+ * address through a volatile cast into a local variable in order to insure we
+ * _have_ to compute the correct address without displacement. This allows us
+ * to use the regular 'm' for the memory address.
+ *
+ * Furthermore, the %Ln operand, which prints the low word register (r+1),
+ * really only works for registers, this means we cannot allow immediate values
+ * for the 64 bit versions -- like we do for the 32 bit ones.
+ *
+ */
+
+#ifndef ATOMIC_OP_RETURN
+#define ATOMIC_OP_RETURN(op)						\
+ATOMIC_QUALS int __atomic_##op##_return(int i, int *v)			\
+{									\
+	int val;							\
+									\
+	asm volatile(							\
+	    "0:						\n"		\
+	    "	orcc		gr0,gr0,gr0,icc3	\n"		\
+	    "	ckeq		icc3,cc7		\n"		\
+	    "	ld.p		%M0,%1			\n"		\
+	    "	orcr		cc7,cc7,cc3		\n"		\
+	    "   "#op"%I2	%1,%2,%1		\n"		\
+	    "	cst.p		%1,%M0		,cc3,#1	\n"		\
+	    "	corcc		gr29,gr29,gr0	,cc3,#1	\n"		\
+	    "	beq		icc3,#0,0b		\n"		\
+	    : "+U"(*v), "=&r"(val)					\
+	    : "NPr"(i)							\
+	    : "memory", "cc7", "cc3", "icc3"				\
+	    );								\
+									\
+	return val;							\
+}									\
+ATOMIC_EXPORT(__atomic_##op##_return);					\
+									\
+ATOMIC_QUALS long long __atomic64_##op##_return(long long i, long long *v)	\
+{									\
+	long long *__v = READ_ONCE(v);					\
+	long long val;							\
+									\
+	asm volatile(							\
+	    "0:						\n"		\
+	    "	orcc		gr0,gr0,gr0,icc3	\n"		\
+	    "	ckeq		icc3,cc7		\n"		\
+	    "	ldd.p		%M0,%1			\n"		\
+	    "	orcr		cc7,cc7,cc3		\n"		\
+	    "   "#op"cc		%L1,%L2,%L1,icc0	\n"		\
+	    "   "#op"x		%1,%2,%1,icc0		\n"		\
+	    "	cstd.p		%1,%M0		,cc3,#1	\n"		\
+	    "	corcc		gr29,gr29,gr0	,cc3,#1	\n"		\
+	    "	beq		icc3,#0,0b		\n"		\
+	    : "+m"(*__v), "=&e"(val)					\
+	    : "e"(i)							\
+	    : "memory", "cc7", "cc3", "icc0", "icc3"			\
+	    );								\
+									\
+	return val;							\
+}									\
+ATOMIC_EXPORT(__atomic64_##op##_return);
+#endif
+
+#ifndef ATOMIC_FETCH_OP
+#define ATOMIC_FETCH_OP(op)						\
+ATOMIC_QUALS int __atomic32_fetch_##op(int i, int *v)			\
+{									\
+	int old, tmp;							\
+									\
+	asm volatile(							\
+		"0:						\n"	\
+		"	orcc		gr0,gr0,gr0,icc3	\n"	\
+		"	ckeq		icc3,cc7		\n"	\
+		"	ld.p		%M0,%1			\n"	\
+		"	orcr		cc7,cc7,cc3		\n"	\
+		"	"#op"%I3	%1,%3,%2		\n"	\
+		"	cst.p		%2,%M0		,cc3,#1	\n"	\
+		"	corcc		gr29,gr29,gr0	,cc3,#1	\n"	\
+		"	beq		icc3,#0,0b		\n"	\
+		: "+U"(*v), "=&r"(old), "=r"(tmp)			\
+		: "NPr"(i)						\
+		: "memory", "cc7", "cc3", "icc3"			\
+		);							\
+									\
+	return old;							\
+}									\
+ATOMIC_EXPORT(__atomic32_fetch_##op);					\
+									\
+ATOMIC_QUALS long long __atomic64_fetch_##op(long long i, long long *v)	\
+{									\
+	long long *__v = READ_ONCE(v);					\
+	long long old, tmp;						\
+									\
+	asm volatile(							\
+		"0:						\n"	\
+		"	orcc		gr0,gr0,gr0,icc3	\n"	\
+		"	ckeq		icc3,cc7		\n"	\
+		"	ldd.p		%M0,%1			\n"	\
+		"	orcr		cc7,cc7,cc3		\n"	\
+		"	"#op"		%L1,%L3,%L2		\n"	\
+		"	"#op"		%1,%3,%2		\n"	\
+		"	cstd.p		%2,%M0		,cc3,#1	\n"	\
+		"	corcc		gr29,gr29,gr0	,cc3,#1	\n"	\
+		"	beq		icc3,#0,0b		\n"	\
+		: "+m"(*__v), "=&e"(old), "=e"(tmp)			\
+		: "e"(i)						\
+		: "memory", "cc7", "cc3", "icc3"			\
+		);							\
+									\
+	return old;							\
+}									\
+ATOMIC_EXPORT(__atomic64_fetch_##op);
+#endif
+
+ATOMIC_FETCH_OP(or)
+ATOMIC_FETCH_OP(and)
+ATOMIC_FETCH_OP(xor)
+
+ATOMIC_OP_RETURN(add)
+ATOMIC_OP_RETURN(sub)
+
+#undef ATOMIC_FETCH_OP
+#undef ATOMIC_OP_RETURN
+#undef ATOMIC_QUALS
+#undef ATOMIC_EXPORT